fs/btrfs/extent-tree.c

   1 /*
   2  * Copyright (C) 2007 Oracle.  All rights reserved.
   3  *
   4  * This program is free software; you can redistribute it and/or
   5  * modify it under the terms of the GNU General Public
   6  * License v2 as published by the Free Software Foundation.
   7  *
   8  * This program is distributed in the hope that it will be useful,
   9  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  10  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  11  * General Public License for more details.
  12  *
  13  * You should have received a copy of the GNU General Public
  14  * License along with this program; if not, write to the
  15  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
  16  * Boston, MA 021110-1307, USA.
  17  */
  18 #include <linux/sched.h>
  19 #include <linux/pagemap.h>
  20 #include <linux/writeback.h>
  21 #include <linux/blkdev.h>
  22 #include <linux/sort.h>
  23 #include <linux/rcupdate.h>
  24 #include <linux/kthread.h>
  25 #include <linux/slab.h>
  26 #include <linux/ratelimit.h>
  27 #include "compat.h"
  28 #include "hash.h"
  29 #include "ctree.h"
  30 #include "disk-io.h"
  31 #include "print-tree.h"
  32 #include "transaction.h"
  33 #include "volumes.h"
  34 #include "raid56.h"
  35 #include "locking.h"
  36 #include "free-space-cache.h"
  37 #include "math.h"
  38
  39 #undef SCRAMBLE_DELAYED_REFS
  40
  41 /*
  42  * control flags for do_chunk_alloc's force field
  43  * CHUNK_ALLOC_NO_FORCE means to only allocate a chunk
  44  * if we really need one.
  45  *
  46  * CHUNK_ALLOC_LIMITED means to only try and allocate one
  47  * if we have very few chunks already allocated.  This is
  48  * used as part of the clustering code to help make sure
  49  * we have a good pool of storage to cluster in, without
  50  * filling the FS with empty chunks
  51  *
  52  * CHUNK_ALLOC_FORCE means it must try to allocate one
  53  *
  54  */
  55 enum {
  56         CHUNK_ALLOC_NO_FORCE = 0,
  57         CHUNK_ALLOC_LIMITED = 1,
  58         CHUNK_ALLOC_FORCE = 2,
  59 };
  60
  61 /*
  62  * Control how reservations are dealt with.
  63  *
  64  * RESERVE_FREE - freeing a reservation.
  65  * RESERVE_ALLOC - allocating space and we need to update bytes_may_use for
  66  *   ENOSPC accounting
  67  * RESERVE_ALLOC_NO_ACCOUNT - allocating space and we should not update
  68  *   bytes_may_use as the ENOSPC accounting is done elsewhere
  69  */
  70 enum {
  71         RESERVE_FREE = 0,
  72         RESERVE_ALLOC = 1,
  73         RESERVE_ALLOC_NO_ACCOUNT = 2,
  74 };
  75
  76 static int update_block_group(struct btrfs_root *root,
  77                               u64 bytenr, u64 num_bytes, int alloc);
  78 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
  79                                 struct btrfs_root *root,
  80                                 u64 bytenr, u64 num_bytes, u64 parent,
  81                                 u64 root_objectid, u64 owner_objectid,
  82                                 u64 owner_offset, int refs_to_drop,
  83                                 struct btrfs_delayed_extent_op *extra_op);
  84 static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op,
  85                                     struct extent_buffer *leaf,
  86                                     struct btrfs_extent_item *ei);
  87 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
  88                                       struct btrfs_root *root,
  89                                       u64 parent, u64 root_objectid,
  90                                       u64 flags, u64 owner, u64 offset,
  91                                       struct btrfs_key *ins, int ref_mod);
  92 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
  93                                      struct btrfs_root *root,
  94                                      u64 parent, u64 root_objectid,
  95                                      u64 flags, struct btrfs_disk_key *key,
  96                                      int level, struct btrfs_key *ins);
  97 static int do_chunk_alloc(struct btrfs_trans_handle *trans,
  98                           struct btrfs_root *extent_root, u64 flags,
  99                           int force);
 100 static int find_next_key(struct btrfs_path *path, int level,
 101                          struct btrfs_key *key);
 102 static void dump_space_info(struct btrfs_space_info *info, u64 bytes,
 103                             int dump_block_groups);
 104 static int btrfs_update_reserved_bytes(struct btrfs_block_group_cache *cache,
 105                                        u64 num_bytes, int reserve);
 106 static int block_rsv_use_bytes(struct btrfs_block_rsv *block_rsv,
 107                                u64 num_bytes);
 108
 109 static noinline int
 110 block_group_cache_done(struct btrfs_block_group_cache *cache)
 111 {
 112         smp_mb();
 113         return cache->cached == BTRFS_CACHE_FINISHED;
 114 }
 115
 116 static int block_group_bits(struct btrfs_block_group_cache *cache, u64 bits)
 117 {
 118         return (cache->flags & bits) == bits;
 119 }
 120
 121 static void btrfs_get_block_group(struct btrfs_block_group_cache *cache)
 122 {
 123         atomic_inc(&cache->count);
 124 }
 125
 126 void btrfs_put_block_group(struct btrfs_block_group_cache *cache)
 127 {
 128         if (atomic_dec_and_test(&cache->count)) {
 129                 WARN_ON(cache->pinned > 0);
 130                 WARN_ON(cache->reserved > 0);
 131                 kfree(cache->free_space_ctl);
 132                 kfree(cache);
 133         }
 134 }
 135
 136 /*
 137  * this adds the block group to the fs_info rb tree for the block group
 138  * cache
 139  */
 140 static int btrfs_add_block_group_cache(struct btrfs_fs_info *info,
 141                                 struct btrfs_block_group_cache *block_group)
 142 {
 143         struct rb_node **p;
 144         struct rb_node *parent = NULL;
 145         struct btrfs_block_group_cache *cache;
 146
 147         spin_lock(&info->block_group_cache_lock);
 148         p = &info->block_group_cache_tree.rb_node;
 149
 150         while (*p) {
 151                 parent = *p;
 152                 cache = rb_entry(parent, struct btrfs_block_group_cache,
 153                                  cache_node);
 154                 if (block_group->key.objectid < cache->key.objectid) {
 155                         p = &(*p)->rb_left;
 156                 } else if (block_group->key.objectid > cache->key.objectid) {
 157                         p = &(*p)->rb_right;
 158                 } else {
 159                         spin_unlock(&info->block_group_cache_lock);
 160                         return -EEXIST;
 161                 }
 162         }
 163
 164         rb_link_node(&block_group->cache_node, parent, p);
 165         rb_insert_color(&block_group->cache_node,
 166                         &info->block_group_cache_tree);
 167
 168         if (info->first_logical_byte > block_group->key.objectid)
 169                 info->first_logical_byte = block_group->key.objectid;
 170
 171         spin_unlock(&info->block_group_cache_lock);
 172
 173         return 0;
 174 }
 175
 176 /*
 177  * This will return the block group at or after bytenr if contains is 0, else
 178  * it will return the block group that contains the bytenr
 179  */
 180 static struct btrfs_block_group_cache *
 181 block_group_cache_tree_search(struct btrfs_fs_info *info, u64 bytenr,
 182                               int contains)
 183 {
 184         struct btrfs_block_group_cache *cache, *ret = NULL;
 185         struct rb_node *n;
 186         u64 end, start;
 187
 188         spin_lock(&info->block_group_cache_lock);
 189         n = info->block_group_cache_tree.rb_node;
 190
 191         while (n) {
 192                 cache = rb_entry(n, struct btrfs_block_group_cache,
 193                                  cache_node);
 194                 end = cache->key.objectid + cache->key.offset - 1;
 195                 start = cache->key.objectid;
 196
 197                 if (bytenr < start) {
 198                         if (!contains && (!ret || start < ret->key.objectid))
 199                                 ret = cache;
 200                         n = n->rb_left;
 201                 } else if (bytenr > start) {
 202                         if (contains && bytenr <= end) {
 203                                 ret = cache;
 204                                 break;
 205                         }
 206                         n = n->rb_right;
 207                 } else {
 208                         ret = cache;
 209                         break;
 210                 }
 211         }
 212         if (ret) {
 213                 btrfs_get_block_group(ret);
 214                 if (bytenr == 0 && info->first_logical_byte > ret->key.objectid)
 215                         info->first_logical_byte = ret->key.objectid;
 216         }
 217         spin_unlock(&info->block_group_cache_lock);
 218
 219         return ret;
 220 }
 221
 222 static int add_excluded_extent(struct btrfs_root *root,
 223                                u64 start, u64 num_bytes)
 224 {
 225         u64 end = start + num_bytes - 1;
 226         set_extent_bits(&root->fs_info->freed_extents[0],
 227                         start, end, EXTENT_UPTODATE, GFP_NOFS);
 228         set_extent_bits(&root->fs_info->freed_extents[1],
 229                         start, end, EXTENT_UPTODATE, GFP_NOFS);
 230         return 0;
 231 }
 232
 233 static void free_excluded_extents(struct btrfs_root *root,
 234                                   struct btrfs_block_group_cache *cache)
 235 {
 236         u64 start, end;
 237
 238         start = cache->key.objectid;
 239         end = start + cache->key.offset - 1;
 240
 241         clear_extent_bits(&root->fs_info->freed_extents[0],
 242                           start, end, EXTENT_UPTODATE, GFP_NOFS);
 243         clear_extent_bits(&root->fs_info->freed_extents[1],
 244                           start, end, EXTENT_UPTODATE, GFP_NOFS);
 245 }
 246
 247 static int exclude_super_stripes(struct btrfs_root *root,
 248                                  struct btrfs_block_group_cache *cache)
 249 {
 250         u64 bytenr;
 251         u64 *logical;
 252         int stripe_len;
 253         int i, nr, ret;
 254
 255         if (cache->key.objectid < BTRFS_SUPER_INFO_OFFSET) {
 256                 stripe_len = BTRFS_SUPER_INFO_OFFSET - cache->key.objectid;
 257                 cache->bytes_super += stripe_len;
 258                 ret = add_excluded_extent(root, cache->key.objectid,
 259                                           stripe_len);
 260                 BUG_ON(ret); /* -ENOMEM */
 261         }
 262
 263         for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
 264                 bytenr = btrfs_sb_offset(i);
 265                 ret = btrfs_rmap_block(&root->fs_info->mapping_tree,
 266                                        cache->key.objectid, bytenr,
 267                                        0, &logical, &nr, &stripe_len);
 268                 BUG_ON(ret); /* -ENOMEM */
 269
 270                 while (nr--) {
 271                         cache->bytes_super += stripe_len;
 272                         ret = add_excluded_extent(root, logical[nr],
 273                                                   stripe_len);
 274                         BUG_ON(ret); /* -ENOMEM */
 275                 }
 276
 277                 kfree(logical);
 278         }
 279         return 0;
 280 }
 281
 282 static struct btrfs_caching_control *
 283 get_caching_control(struct btrfs_block_group_cache *cache)
 284 {
 285         struct btrfs_caching_control *ctl;
 286
 287         spin_lock(&cache->lock);
 288         if (cache->cached != BTRFS_CACHE_STARTED) {
 289                 spin_unlock(&cache->lock);
 290                 return NULL;
 291         }
 292
 293         /* We're loading it the fast way, so we don't have a caching_ctl. */
 294         if (!cache->caching_ctl) {
 295                 spin_unlock(&cache->lock);
 296                 return NULL;
 297         }
 298
 299         ctl = cache->caching_ctl;
 300         atomic_inc(&ctl->count);
 301         spin_unlock(&cache->lock);
 302         return ctl;
 303 }
 304
 305 static void put_caching_control(struct btrfs_caching_control *ctl)
 306 {
 307         if (atomic_dec_and_test(&ctl->count))
 308                 kfree(ctl);
 309 }
 310
 311 /*
 312  * this is only called by cache_block_group, since we could have freed extents
 313  * we need to check the pinned_extents for any extents that can't be used yet
 314  * since their free space will be released as soon as the transaction commits.
 315  */
 316 static u64 add_new_free_space(struct btrfs_block_group_cache *block_group,
 317                               struct btrfs_fs_info *info, u64 start, u64 end)
 318 {
 319         u64 extent_start, extent_end, size, total_added = 0;
 320         int ret;
 321
 322         while (start < end) {
 323                 ret = find_first_extent_bit(info->pinned_extents, start,
 324                                             &extent_start, &extent_end,
 325                                             EXTENT_DIRTY | EXTENT_UPTODATE,
 326                                             NULL);
 327                 if (ret)
 328                         break;
 329
 330                 if (extent_start <= start) {
 331                         start = extent_end + 1;
 332                 } else if (extent_start > start && extent_start < end) {
 333                         size = extent_start - start;
 334                         total_added += size;
 335                         ret = btrfs_add_free_space(block_group, start,
 336                                                    size);
 337                         BUG_ON(ret); /* -ENOMEM or logic error */
 338                         start = extent_end + 1;
 339                 } else {
 340                         break;
 341                 }
 342         }
 343
 344         if (start < end) {
 345                 size = end - start;
 346                 total_added += size;
 347                 ret = btrfs_add_free_space(block_group, start, size);
 348                 BUG_ON(ret); /* -ENOMEM or logic error */
 349         }
 350
 351         return total_added;
 352 }
 353
 354 static noinline void caching_thread(struct btrfs_work *work)
 355 {
 356         struct btrfs_block_group_cache *block_group;
 357         struct btrfs_fs_info *fs_info;
 358         struct btrfs_caching_control *caching_ctl;
 359         struct btrfs_root *extent_root;
 360         struct btrfs_path *path;
 361         struct extent_buffer *leaf;
 362         struct btrfs_key key;
 363         u64 total_found = 0;
 364         u64 last = 0;
 365         u32 nritems;
 366         int ret = 0;
 367
 368         caching_ctl = container_of(work, struct btrfs_caching_control, work);
 369         block_group = caching_ctl->block_group;
 370         fs_info = block_group->fs_info;
 371         extent_root = fs_info->extent_root;
 372
 373         path = btrfs_alloc_path();
 374         if (!path)
 375                 goto out;
 376
 377         last = max_t(u64, block_group->key.objectid, BTRFS_SUPER_INFO_OFFSET);
 378
 379         /*
 380          * We don't want to deadlock with somebody trying to allocate a new
 381          * extent for the extent root while also trying to search the extent
 382          * root to add free space.  So we skip locking and search the commit
 383          * root, since its read-only
 384          */
 385         path->skip_locking = 1;
 386         path->search_commit_root = 1;
 387         path->reada = 1;
 388
 389         key.objectid = last;
 390         key.offset = 0;
 391         key.type = BTRFS_EXTENT_ITEM_KEY;
 392 again:
 393         mutex_lock(&caching_ctl->mutex);
 394         /* need to make sure the commit_root doesn't disappear */
 395         down_read(&fs_info->extent_commit_sem);
 396
 397         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
 398         if (ret < 0)
 399                 goto err;
 400
 401         leaf = path->nodes[0];
 402         nritems = btrfs_header_nritems(leaf);
 403
 404         while (1) {
 405                 if (btrfs_fs_closing(fs_info) > 1) {
 406                         last = (u64)-1;
 407                         break;
 408                 }
 409
 410                 if (path->slots[0] < nritems) {
 411                         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
 412                 } else {
 413                         ret = find_next_key(path, 0, &key);
 414                         if (ret)
 415                                 break;
 416
 417                         if (need_resched() ||
 418                             btrfs_next_leaf(extent_root, path)) {
 419                                 caching_ctl->progress = last;
 420                                 btrfs_release_path(path);
 421                                 up_read(&fs_info->extent_commit_sem);
 422                                 mutex_unlock(&caching_ctl->mutex);
 423                                 cond_resched();
 424                                 goto again;
 425                         }
 426                         leaf = path->nodes[0];
 427                         nritems = btrfs_header_nritems(leaf);
 428                         continue;
 429                 }
 430
 431                 if (key.objectid < block_group->key.objectid) {
 432                         path->slots[0]++;
 433                         continue;
 434                 }
 435
 436                 if (key.objectid >= block_group->key.objectid +
 437                     block_group->key.offset)
 438                         break;
 439
 440                 if (key.type == BTRFS_EXTENT_ITEM_KEY) {
 441                         total_found += add_new_free_space(block_group,
 442                                                           fs_info, last,
 443                                                           key.objectid);
 444                         last = key.objectid + key.offset;
 445
 446                         if (total_found > (1024 * 1024 * 2)) {
 447                                 total_found = 0;
 448                                 wake_up(&caching_ctl->wait);
 449                         }
 450                 }
 451                 path->slots[0]++;
 452         }
 453         ret = 0;
 454
 455         total_found += add_new_free_space(block_group, fs_info, last,
 456                                           block_group->key.objectid +
 457                                           block_group->key.offset);
 458         caching_ctl->progress = (u64)-1;
 459
 460         spin_lock(&block_group->lock);
 461         block_group->caching_ctl = NULL;
 462         block_group->cached = BTRFS_CACHE_FINISHED;
 463         spin_unlock(&block_group->lock);
 464
 465 err:
 466         btrfs_free_path(path);
 467         up_read(&fs_info->extent_commit_sem);
 468
 469         free_excluded_extents(extent_root, block_group);
 470
 471         mutex_unlock(&caching_ctl->mutex);
 472 out:
 473         wake_up(&caching_ctl->wait);
 474
 475         put_caching_control(caching_ctl);
 476         btrfs_put_block_group(block_group);
 477 }
 478
 479 static int cache_block_group(struct btrfs_block_group_cache *cache,
 480                              int load_cache_only)
 481 {
 482         DEFINE_WAIT(wait);
 483         struct btrfs_fs_info *fs_info = cache->fs_info;
 484         struct btrfs_caching_control *caching_ctl;
 485         int ret = 0;
 486
 487         caching_ctl = kzalloc(sizeof(*caching_ctl), GFP_NOFS);
 488         if (!caching_ctl)
 489                 return -ENOMEM;
 490
 491         INIT_LIST_HEAD(&caching_ctl->list);
 492         mutex_init(&caching_ctl->mutex);
 493         init_waitqueue_head(&caching_ctl->wait);
 494         caching_ctl->block_group = cache;
 495         caching_ctl->progress = cache->key.objectid;
 496         atomic_set(&caching_ctl->count, 1);
 497         caching_ctl->work.func = caching_thread;
 498
 499         spin_lock(&cache->lock);
 500         /*
 501          * This should be a rare occasion, but this could happen I think in the
 502          * case where one thread starts to load the space cache info, and then
 503          * some other thread starts a transaction commit which tries to do an
 504          * allocation while the other thread is still loading the space cache
 505          * info.  The previous loop should have kept us from choosing this block
 506          * group, but if we've moved to the state where we will wait on caching
 507          * block groups we need to first check if we're doing a fast load here,
 508          * so we can wait for it to finish, otherwise we could end up allocating
 509          * from a block group who's cache gets evicted for one reason or
 510          * another.
 511          */
 512         while (cache->cached == BTRFS_CACHE_FAST) {
 513                 struct btrfs_caching_control *ctl;
 514
 515                 ctl = cache->caching_ctl;
 516                 atomic_inc(&ctl->count);
 517                 prepare_to_wait(&ctl->wait, &wait, TASK_UNINTERRUPTIBLE);
 518                 spin_unlock(&cache->lock);
 519
 520                 schedule();
 521
 522                 finish_wait(&ctl->wait, &wait);
 523                 put_caching_control(ctl);
 524                 spin_lock(&cache->lock);
 525         }
 526
 527         if (cache->cached != BTRFS_CACHE_NO) {
 528                 spin_unlock(&cache->lock);
 529                 kfree(caching_ctl);
 530                 return 0;
 531         }
 532         WARN_ON(cache->caching_ctl);
 533         cache->caching_ctl = caching_ctl;
 534         cache->cached = BTRFS_CACHE_FAST;
 535         spin_unlock(&cache->lock);
 536
 537         if (fs_info->mount_opt & BTRFS_MOUNT_SPACE_CACHE) {
 538                 ret = load_free_space_cache(fs_info, cache);
 539
 540                 spin_lock(&cache->lock);
 541                 if (ret == 1) {
 542                         cache->caching_ctl = NULL;
 543                         cache->cached = BTRFS_CACHE_FINISHED;
 544                         cache->last_byte_to_unpin = (u64)-1;
 545                 } else {
 546                         if (load_cache_only) {
 547                                 cache->caching_ctl = NULL;
 548                                 cache->cached = BTRFS_CACHE_NO;
 549                         } else {
 550                                 cache->cached = BTRFS_CACHE_STARTED;
 551                         }
 552                 }
 553                 spin_unlock(&cache->lock);
 554                 wake_up(&caching_ctl->wait);
 555                 if (ret == 1) {
 556                         put_caching_control(caching_ctl);
 557                         free_excluded_extents(fs_info->extent_root, cache);
 558                         return 0;
 559                 }
 560         } else {
 561                 /*
 562                  * We are not going to do the fast caching, set cached to the
 563                  * appropriate value and wakeup any waiters.
 564                  */
 565                 spin_lock(&cache->lock);
 566                 if (load_cache_only) {
 567                         cache->caching_ctl = NULL;
 568                         cache->cached = BTRFS_CACHE_NO;
 569                 } else {
 570                         cache->cached = BTRFS_CACHE_STARTED;
 571                 }
 572                 spin_unlock(&cache->lock);
 573                 wake_up(&caching_ctl->wait);
 574         }
 575
 576         if (load_cache_only) {
 577                 put_caching_control(caching_ctl);
 578                 return 0;
 579         }
 580
 581         down_write(&fs_info->extent_commit_sem);
 582         atomic_inc(&caching_ctl->count);
 583         list_add_tail(&caching_ctl->list, &fs_info->caching_block_groups);
 584         up_write(&fs_info->extent_commit_sem);
 585
 586         btrfs_get_block_group(cache);
 587
 588         btrfs_queue_worker(&fs_info->caching_workers, &caching_ctl->work);
 589
 590         return ret;
 591 }
 592
 593 /*
 594  * return the block group that starts at or after bytenr
 595  */
 596 static struct btrfs_block_group_cache *
 597 btrfs_lookup_first_block_group(struct btrfs_fs_info *info, u64 bytenr)
 598 {
 599         struct btrfs_block_group_cache *cache;
 600
 601         cache = block_group_cache_tree_search(info, bytenr, 0);
 602
 603         return cache;
 604 }
 605
 606 /*
 607  * return the block group that contains the given bytenr
 608  */
 609 struct btrfs_block_group_cache *btrfs_lookup_block_group(
 610                                                  struct btrfs_fs_info *info,
 611                                                  u64 bytenr)
 612 {
 613         struct btrfs_block_group_cache *cache;
 614
 615         cache = block_group_cache_tree_search(info, bytenr, 1);
 616
 617         return cache;
 618 }
 619
 620 static struct btrfs_space_info *__find_space_info(struct btrfs_fs_info *info,
 621                                                   u64 flags)
 622 {
 623         struct list_head *head = &info->space_info;
 624         struct btrfs_space_info *found;
 625
 626         flags &= BTRFS_BLOCK_GROUP_TYPE_MASK;
 627
 628         rcu_read_lock();
 629         list_for_each_entry_rcu(found, head, list) {
 630                 if (found->flags & flags) {
 631                         rcu_read_unlock();
 632                         return found;
 633                 }
 634         }
 635         rcu_read_unlock();
 636         return NULL;
 637 }
 638
 639 /*
 640  * after adding space to the filesystem, we need to clear the full flags
 641  * on all the space infos.
 642  */
 643 void btrfs_clear_space_info_full(struct btrfs_fs_info *info)
 644 {
 645         struct list_head *head = &info->space_info;
 646         struct btrfs_space_info *found;
 647
 648         rcu_read_lock();
 649         list_for_each_entry_rcu(found, head, list)
 650                 found->full = 0;
 651         rcu_read_unlock();
 652 }
 653
 654 u64 btrfs_find_block_group(struct btrfs_root *root,
 655                            u64 search_start, u64 search_hint, int owner)
 656 {
 657         struct btrfs_block_group_cache *cache;
 658         u64 used;
 659         u64 last = max(search_hint, search_start);
 660         u64 group_start = 0;
 661         int full_search = 0;
 662         int factor = 9;
 663         int wrapped = 0;
 664 again:
 665         while (1) {
 666                 cache = btrfs_lookup_first_block_group(root->fs_info, last);
 667                 if (!cache)
 668                         break;
 669
 670                 spin_lock(&cache->lock);
 671                 last = cache->key.objectid + cache->key.offset;
 672                 used = btrfs_block_group_used(&cache->item);
 673
 674                 if ((full_search || !cache->ro) &&
 675                     block_group_bits(cache, BTRFS_BLOCK_GROUP_METADATA)) {
 676                         if (used + cache->pinned + cache->reserved <
 677                             div_factor(cache->key.offset, factor)) {
 678                                 group_start = cache->key.objectid;
 679                                 spin_unlock(&cache->lock);
 680                                 btrfs_put_block_group(cache);
 681                                 goto found;
 682                         }
 683                 }
 684                 spin_unlock(&cache->lock);
 685                 btrfs_put_block_group(cache);
 686                 cond_resched();
 687         }
 688         if (!wrapped) {
 689                 last = search_start;
 690                 wrapped = 1;
 691                 goto again;
 692         }
 693         if (!full_search && factor < 10) {
 694                 last = search_start;
 695                 full_search = 1;
 696                 factor = 10;
 697                 goto again;
 698         }
 699 found:
 700         return group_start;
 701 }
 702
 703 /* simple helper to search for an existing extent at a given offset */
 704 int btrfs_lookup_extent(struct btrfs_root *root, u64 start, u64 len)
 705 {
 706         int ret;
 707         struct btrfs_key key;
 708         struct btrfs_path *path;
 709
 710         path = btrfs_alloc_path();
 711         if (!path)
 712                 return -ENOMEM;
 713
 714         key.objectid = start;
 715         key.offset = len;
 716         btrfs_set_key_type(&key, BTRFS_EXTENT_ITEM_KEY);
 717         ret = btrfs_search_slot(NULL, root->fs_info->extent_root, &key, path,
 718                                 0, 0);
 719         btrfs_free_path(path);
 720         return ret;
 721 }
 722
 723 /*
 724  * helper function to lookup reference count and flags of extent.
 725  *
 726  * the head node for delayed ref is used to store the sum of all the
 727  * reference count modifications queued up in the rbtree. the head
 728  * node may also store the extent flags to set. This way you can check
 729  * to see what the reference count and extent flags would be if all of
 730  * the delayed refs are not processed.
 731  */
 732 int btrfs_lookup_extent_info(struct btrfs_trans_handle *trans,
 733                              struct btrfs_root *root, u64 bytenr,
 734                              u64 num_bytes, u64 *refs, u64 *flags)
 735 {
 736         struct btrfs_delayed_ref_head *head;
 737         struct btrfs_delayed_ref_root *delayed_refs;
 738         struct btrfs_path *path;
 739         struct btrfs_extent_item *ei;
 740         struct extent_buffer *leaf;
 741         struct btrfs_key key;
 742         u32 item_size;
 743         u64 num_refs;
 744         u64 extent_flags;
 745         int ret;
 746
 747         path = btrfs_alloc_path();
 748         if (!path)
 749                 return -ENOMEM;
 750
 751         key.objectid = bytenr;
 752         key.type = BTRFS_EXTENT_ITEM_KEY;
 753         key.offset = num_bytes;
 754         if (!trans) {
 755                 path->skip_locking = 1;
 756                 path->search_commit_root = 1;
 757         }
 758 again:
 759         ret = btrfs_search_slot(trans, root->fs_info->extent_root,
 760                                 &key, path, 0, 0);
 761         if (ret < 0)
 762                 goto out_free;
 763
 764         if (ret == 0) {
 765                 leaf = path->nodes[0];
 766                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
 767                 if (item_size >= sizeof(*ei)) {
 768                         ei = btrfs_item_ptr(leaf, path->slots[0],
 769                                             struct btrfs_extent_item);
 770                         num_refs = btrfs_extent_refs(leaf, ei);
 771                         extent_flags = btrfs_extent_flags(leaf, ei);
 772                 } else {
 773 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
 774                         struct btrfs_extent_item_v0 *ei0;
 775                         BUG_ON(item_size != sizeof(*ei0));
 776                         ei0 = btrfs_item_ptr(leaf, path->slots[0],
 777                                              struct btrfs_extent_item_v0);
 778                         num_refs = btrfs_extent_refs_v0(leaf, ei0);
 779                         /* FIXME: this isn't correct for data */
 780                         extent_flags = BTRFS_BLOCK_FLAG_FULL_BACKREF;
 781 #else
 782                         BUG();
 783 #endif
 784                 }
 785                 BUG_ON(num_refs == 0);
 786         } else {
 787                 num_refs = 0;
 788                 extent_flags = 0;
 789                 ret = 0;
 790         }
 791
 792         if (!trans)
 793                 goto out;
 794
 795         delayed_refs = &trans->transaction->delayed_refs;
 796         spin_lock(&delayed_refs->lock);
 797         head = btrfs_find_delayed_ref_head(trans, bytenr);
 798         if (head) {
 799                 if (!mutex_trylock(&head->mutex)) {
 800                         atomic_inc(&head->node.refs);
 801                         spin_unlock(&delayed_refs->lock);
 802
 803                         btrfs_release_path(path);
 804
 805                         /*
 806                          * Mutex was contended, block until it's released and try
 807                          * again
 808                          */
 809                         mutex_lock(&head->mutex);
 810                         mutex_unlock(&head->mutex);
 811                         btrfs_put_delayed_ref(&head->node);
 812                         goto again;
 813                 }
 814                 if (head->extent_op && head->extent_op->update_flags)
 815                         extent_flags |= head->extent_op->flags_to_set;
 816                 else
 817                         BUG_ON(num_refs == 0);
 818
 819                 num_refs += head->node.ref_mod;
 820                 mutex_unlock(&head->mutex);
 821         }
 822         spin_unlock(&delayed_refs->lock);
 823 out:
 824         WARN_ON(num_refs == 0);
 825         if (refs)
 826                 *refs = num_refs;
 827         if (flags)
 828                 *flags = extent_flags;
 829 out_free:
 830         btrfs_free_path(path);
 831         return ret;
 832 }
 833
 834 /*
 835  * Back reference rules.  Back refs have three main goals:
 836  *
 837  * 1) differentiate between all holders of references to an extent so that
 838  *    when a reference is dropped we can make sure it was a valid reference
 839  *    before freeing the extent.
 840  *
 841  * 2) Provide enough information to quickly find the holders of an extent
 842  *    if we notice a given block is corrupted or bad.
 843  *
 844  * 3) Make it easy to migrate blocks for FS shrinking or storage pool
 845  *    maintenance.  This is actually the same as #2, but with a slightly
 846  *    different use case.
 847  *
 848  * There are two kinds of back refs. The implicit back refs is optimized
 849  * for pointers in non-shared tree blocks. For a given pointer in a block,
 850  * back refs of this kind provide information about the block's owner tree
 851  * and the pointer's key. These information allow us to find the block by
 852  * b-tree searching. The full back refs is for pointers in tree blocks not
 853  * referenced by their owner trees. The location of tree block is recorded
 854  * in the back refs. Actually the full back refs is generic, and can be
 855  * used in all cases the implicit back refs is used. The major shortcoming
 856  * of the full back refs is its overhead. Every time a tree block gets
 857  * COWed, we have to update back refs entry for all pointers in it.
 858  *
 859  * For a newly allocated tree block, we use implicit back refs for
 860  * pointers in it. This means most tree related operations only involve
 861  * implicit back refs. For a tree block created in old transaction, the
 862  * only way to drop a reference to it is COW it. So we can detect the
 863  * event that tree block loses its owner tree's reference and do the
 864  * back refs conversion.
 865  *
 866  * When a tree block is COW'd through a tree, there are four cases:
 867  *
 868  * The reference count of the block is one and the tree is the block's
 869  * owner tree. Nothing to do in this case.
 870  *
 871  * The reference count of the block is one and the tree is not the
 872  * block's owner tree. In this case, full back refs is used for pointers
 873  * in the block. Remove these full back refs, add implicit back refs for
 874  * every pointers in the new block.
 875  *
 876  * The reference count of the block is greater than one and the tree is
 877  * the block's owner tree. In this case, implicit back refs is used for
 878  * pointers in the block. Add full back refs for every pointers in the
 879  * block, increase lower level extents' reference counts. The original
 880  * implicit back refs are entailed to the new block.
 881  *
 882  * The reference count of the block is greater than one and the tree is
 883  * not the block's owner tree. Add implicit back refs for every pointer in
 884  * the new block, increase lower level extents' reference count.
 885  *
 886  * Back Reference Key composing:
 887  *
 888  * The key objectid corresponds to the first byte in the extent,
 889  * The key type is used to differentiate between types of back refs.
 890  * There are different meanings of the key offset for different types
 891  * of back refs.
 892  *
 893  * File extents can be referenced by:
 894  *
 895  * - multiple snapshots, subvolumes, or different generations in one subvol
 896  * - different files inside a single subvolume
 897  * - different offsets inside a file (bookend extents in file.c)
 898  *
 899  * The extent ref structure for the implicit back refs has fields for:
 900  *
 901  * - Objectid of the subvolume root
 902  * - objectid of the file holding the reference
 903  * - original offset in the file
 904  * - how many bookend extents
 905  *
 906  * The key offset for the implicit back refs is hash of the first
 907  * three fields.
 908  *
 909  * The extent ref structure for the full back refs has field for:
 910  *
 911  * - number of pointers in the tree leaf
 912  *
 913  * The key offset for the implicit back refs is the first byte of
 914  * the tree leaf
 915  *
 916  * When a file extent is allocated, The implicit back refs is used.
 917  * the fields are filled in:
 918  *
 919  *     (root_key.objectid, inode objectid, offset in file, 1)
 920  *
 921  * When a file extent is removed file truncation, we find the
 922  * corresponding implicit back refs and check the following fields:
 923  *
 924  *     (btrfs_header_owner(leaf), inode objectid, offset in file)
 925  *
 926  * Btree extents can be referenced by:
 927  *
 928  * - Different subvolumes
 929  *
 930  * Both the implicit back refs and the full back refs for tree blocks
 931  * only consist of key. The key offset for the implicit back refs is
 932  * objectid of block's owner tree. The key offset for the full back refs
 933  * is the first byte of parent block.
 934  *
 935  * When implicit back refs is used, information about the lowest key and
 936  * level of the tree block are required. These information are stored in
 937  * tree block info structure.
 938  */
 939
 940 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
 941 static int convert_extent_item_v0(struct btrfs_trans_handle *trans,
 942                                   struct btrfs_root *root,
 943                                   struct btrfs_path *path,
 944                                   u64 owner, u32 extra_size)
 945 {
 946         struct btrfs_extent_item *item;
 947         struct btrfs_extent_item_v0 *ei0;
 948         struct btrfs_extent_ref_v0 *ref0;
 949         struct btrfs_tree_block_info *bi;
 950         struct extent_buffer *leaf;
 951         struct btrfs_key key;
 952         struct btrfs_key found_key;
 953         u32 new_size = sizeof(*item);
 954         u64 refs;
 955         int ret;
 956
 957         leaf = path->nodes[0];
 958         BUG_ON(btrfs_item_size_nr(leaf, path->slots[0]) != sizeof(*ei0));
 959
 960         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
 961         ei0 = btrfs_item_ptr(leaf, path->slots[0],
 962                              struct btrfs_extent_item_v0);
 963         refs = btrfs_extent_refs_v0(leaf, ei0);
 964
 965         if (owner == (u64)-1) {
 966                 while (1) {
 967                         if (path->slots[0] >= btrfs_header_nritems(leaf)) {
 968                                 ret = btrfs_next_leaf(root, path);
 969                                 if (ret < 0)
 970                                         return ret;
 971                                 BUG_ON(ret > 0); /* Corruption */
 972                                 leaf = path->nodes[0];
 973                         }
 974                         btrfs_item_key_to_cpu(leaf, &found_key,
 975                                               path->slots[0]);
 976                         BUG_ON(key.objectid != found_key.objectid);
 977                         if (found_key.type != BTRFS_EXTENT_REF_V0_KEY) {
 978                                 path->slots[0]++;
 979                                 continue;
 980                         }
 981                         ref0 = btrfs_item_ptr(leaf, path->slots[0],
 982                                               struct btrfs_extent_ref_v0);
 983                         owner = btrfs_ref_objectid_v0(leaf, ref0);
 984                         break;
 985                 }
 986         }
 987         btrfs_release_path(path);
 988
 989         if (owner < BTRFS_FIRST_FREE_OBJECTID)
 990                 new_size += sizeof(*bi);
 991
 992         new_size -= sizeof(*ei0);
 993         ret = btrfs_search_slot(trans, root, &key, path,
 994                                 new_size + extra_size, 1);
 995         if (ret < 0)
 996                 return ret;
 997         BUG_ON(ret); /* Corruption */
 998
 999         btrfs_extend_item(trans, root, path, new_size);
1000
1001         leaf = path->nodes[0];
1002         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1003         btrfs_set_extent_refs(leaf, item, refs);
1004         /* FIXME: get real generation */
1005         btrfs_set_extent_generation(leaf, item, 0);
1006         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1007                 btrfs_set_extent_flags(leaf, item,
1008                                        BTRFS_EXTENT_FLAG_TREE_BLOCK |
1009                                        BTRFS_BLOCK_FLAG_FULL_BACKREF);
1010                 bi = (struct btrfs_tree_block_info *)(item + 1);
1011                 /* FIXME: get first key of the block */
1012                 memset_extent_buffer(leaf, 0, (unsigned long)bi, sizeof(*bi));
1013                 btrfs_set_tree_block_level(leaf, bi, (int)owner);
1014         } else {
1015                 btrfs_set_extent_flags(leaf, item, BTRFS_EXTENT_FLAG_DATA);
1016         }
1017         btrfs_mark_buffer_dirty(leaf);
1018         return 0;
1019 }
1020 #endif
1021
1022 static u64 hash_extent_data_ref(u64 root_objectid, u64 owner, u64 offset)
1023 {
1024         u32 high_crc = ~(u32)0;
1025         u32 low_crc = ~(u32)0;
1026         __le64 lenum;
1027
1028         lenum = cpu_to_le64(root_objectid);
1029         high_crc = crc32c(high_crc, &lenum, sizeof(lenum));
1030         lenum = cpu_to_le64(owner);
1031         low_crc = crc32c(low_crc, &lenum, sizeof(lenum));
1032         lenum = cpu_to_le64(offset);
1033         low_crc = crc32c(low_crc, &lenum, sizeof(lenum));
1034
1035         return ((u64)high_crc << 31) ^ (u64)low_crc;
1036 }
1037
1038 static u64 hash_extent_data_ref_item(struct extent_buffer *leaf,
1039                                      struct btrfs_extent_data_ref *ref)
1040 {
1041         return hash_extent_data_ref(btrfs_extent_data_ref_root(leaf, ref),
1042                                     btrfs_extent_data_ref_objectid(leaf, ref),
1043                                     btrfs_extent_data_ref_offset(leaf, ref));
1044 }
1045
1046 static int match_extent_data_ref(struct extent_buffer *leaf,
1047                                  struct btrfs_extent_data_ref *ref,
1048                                  u64 root_objectid, u64 owner, u64 offset)
1049 {
1050         if (btrfs_extent_data_ref_root(leaf, ref) != root_objectid ||
1051             btrfs_extent_data_ref_objectid(leaf, ref) != owner ||
1052             btrfs_extent_data_ref_offset(leaf, ref) != offset)
1053                 return 0;
1054         return 1;
1055 }
1056
1057 static noinline int lookup_extent_data_ref(struct btrfs_trans_handle *trans,
1058                                            struct btrfs_root *root,
1059                                            struct btrfs_path *path,
1060                                            u64 bytenr, u64 parent,
1061                                            u64 root_objectid,
1062                                            u64 owner, u64 offset)
1063 {
1064         struct btrfs_key key;
1065         struct btrfs_extent_data_ref *ref;
1066         struct extent_buffer *leaf;
1067         u32 nritems;
1068         int ret;
1069         int recow;
1070         int err = -ENOENT;
1071
1072         key.objectid = bytenr;
1073         if (parent) {
1074                 key.type = BTRFS_SHARED_DATA_REF_KEY;
1075                 key.offset = parent;
1076         } else {
1077                 key.type = BTRFS_EXTENT_DATA_REF_KEY;
1078                 key.offset = hash_extent_data_ref(root_objectid,
1079                                                   owner, offset);
1080         }
1081 again:
1082         recow = 0;
1083         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1084         if (ret < 0) {
1085                 err = ret;
1086                 goto fail;
1087         }
1088
1089         if (parent) {
1090                 if (!ret)
1091                         return 0;
1092 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1093                 key.type = BTRFS_EXTENT_REF_V0_KEY;
1094                 btrfs_release_path(path);
1095                 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1096                 if (ret < 0) {
1097                         err = ret;
1098                         goto fail;
1099                 }
1100                 if (!ret)
1101                         return 0;
1102 #endif
1103                 goto fail;
1104         }
1105
1106         leaf = path->nodes[0];
1107         nritems = btrfs_header_nritems(leaf);
1108         while (1) {
1109                 if (path->slots[0] >= nritems) {
1110                         ret = btrfs_next_leaf(root, path);
1111                         if (ret < 0)
1112                                 err = ret;
1113                         if (ret)
1114                                 goto fail;
1115
1116                         leaf = path->nodes[0];
1117                         nritems = btrfs_header_nritems(leaf);
1118                         recow = 1;
1119                 }
1120
1121                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1122                 if (key.objectid != bytenr ||
1123                     key.type != BTRFS_EXTENT_DATA_REF_KEY)
1124                         goto fail;
1125
1126                 ref = btrfs_item_ptr(leaf, path->slots[0],
1127                                      struct btrfs_extent_data_ref);
1128
1129                 if (match_extent_data_ref(leaf, ref, root_objectid,
1130                                           owner, offset)) {
1131                         if (recow) {
1132                                 btrfs_release_path(path);
1133                                 goto again;
1134                         }
1135                         err = 0;
1136                         break;
1137                 }
1138                 path->slots[0]++;
1139         }
1140 fail:
1141         return err;
1142 }
1143
1144 static noinline int insert_extent_data_ref(struct btrfs_trans_handle *trans,
1145                                            struct btrfs_root *root,
1146                                            struct btrfs_path *path,
1147                                            u64 bytenr, u64 parent,
1148                                            u64 root_objectid, u64 owner,
1149                                            u64 offset, int refs_to_add)
1150 {
1151         struct btrfs_key key;
1152         struct extent_buffer *leaf;
1153         u32 size;
1154         u32 num_refs;
1155         int ret;
1156
1157         key.objectid = bytenr;
1158         if (parent) {
1159                 key.type = BTRFS_SHARED_DATA_REF_KEY;
1160                 key.offset = parent;
1161                 size = sizeof(struct btrfs_shared_data_ref);
1162         } else {
1163                 key.type = BTRFS_EXTENT_DATA_REF_KEY;
1164                 key.offset = hash_extent_data_ref(root_objectid,
1165                                                   owner, offset);
1166                 size = sizeof(struct btrfs_extent_data_ref);
1167         }
1168
1169         ret = btrfs_insert_empty_item(trans, root, path, &key, size);
1170         if (ret && ret != -EEXIST)
1171                 goto fail;
1172
1173         leaf = path->nodes[0];
1174         if (parent) {
1175                 struct btrfs_shared_data_ref *ref;
1176                 ref = btrfs_item_ptr(leaf, path->slots[0],
1177                                      struct btrfs_shared_data_ref);
1178                 if (ret == 0) {
1179                         btrfs_set_shared_data_ref_count(leaf, ref, refs_to_add);
1180                 } else {
1181                         num_refs = btrfs_shared_data_ref_count(leaf, ref);
1182                         num_refs += refs_to_add;
1183                         btrfs_set_shared_data_ref_count(leaf, ref, num_refs);
1184                 }
1185         } else {
1186                 struct btrfs_extent_data_ref *ref;
1187                 while (ret == -EEXIST) {
1188                         ref = btrfs_item_ptr(leaf, path->slots[0],
1189                                              struct btrfs_extent_data_ref);
1190                         if (match_extent_data_ref(leaf, ref, root_objectid,
1191                                                   owner, offset))
1192                                 break;
1193                         btrfs_release_path(path);
1194                         key.offset++;
1195                         ret = btrfs_insert_empty_item(trans, root, path, &key,
1196                                                       size);
1197                         if (ret && ret != -EEXIST)
1198                                 goto fail;
1199
1200                         leaf = path->nodes[0];
1201                 }
1202                 ref = btrfs_item_ptr(leaf, path->slots[0],
1203                                      struct btrfs_extent_data_ref);
1204                 if (ret == 0) {
1205                         btrfs_set_extent_data_ref_root(leaf, ref,
1206                                                        root_objectid);
1207                         btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
1208                         btrfs_set_extent_data_ref_offset(leaf, ref, offset);
1209                         btrfs_set_extent_data_ref_count(leaf, ref, refs_to_add);
1210                 } else {
1211                         num_refs = btrfs_extent_data_ref_count(leaf, ref);
1212                         num_refs += refs_to_add;
1213                         btrfs_set_extent_data_ref_count(leaf, ref, num_refs);
1214                 }
1215         }
1216         btrfs_mark_buffer_dirty(leaf);
1217         ret = 0;
1218 fail:
1219         btrfs_release_path(path);
1220         return ret;
1221 }
1222
1223 static noinline int remove_extent_data_ref(struct btrfs_trans_handle *trans,
1224                                            struct btrfs_root *root,
1225                                            struct btrfs_path *path,
1226                                            int refs_to_drop)
1227 {
1228         struct btrfs_key key;
1229         struct btrfs_extent_data_ref *ref1 = NULL;
1230         struct btrfs_shared_data_ref *ref2 = NULL;
1231         struct extent_buffer *leaf;
1232         u32 num_refs = 0;
1233         int ret = 0;
1234
1235         leaf = path->nodes[0];
1236         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1237
1238         if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
1239                 ref1 = btrfs_item_ptr(leaf, path->slots[0],
1240                                       struct btrfs_extent_data_ref);
1241                 num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1242         } else if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
1243                 ref2 = btrfs_item_ptr(leaf, path->slots[0],
1244                                       struct btrfs_shared_data_ref);
1245                 num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1246 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1247         } else if (key.type == BTRFS_EXTENT_REF_V0_KEY) {
1248                 struct btrfs_extent_ref_v0 *ref0;
1249                 ref0 = btrfs_item_ptr(leaf, path->slots[0],
1250                                       struct btrfs_extent_ref_v0);
1251                 num_refs = btrfs_ref_count_v0(leaf, ref0);
1252 #endif
1253         } else {
1254                 BUG();
1255         }
1256
1257         BUG_ON(num_refs < refs_to_drop);
1258         num_refs -= refs_to_drop;
1259
1260         if (num_refs == 0) {
1261                 ret = btrfs_del_item(trans, root, path);
1262         } else {
1263                 if (key.type == BTRFS_EXTENT_DATA_REF_KEY)
1264                         btrfs_set_extent_data_ref_count(leaf, ref1, num_refs);
1265                 else if (key.type == BTRFS_SHARED_DATA_REF_KEY)
1266                         btrfs_set_shared_data_ref_count(leaf, ref2, num_refs);
1267 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1268                 else {
1269                         struct btrfs_extent_ref_v0 *ref0;
1270                         ref0 = btrfs_item_ptr(leaf, path->slots[0],
1271                                         struct btrfs_extent_ref_v0);
1272                         btrfs_set_ref_count_v0(leaf, ref0, num_refs);
1273                 }
1274 #endif
1275                 btrfs_mark_buffer_dirty(leaf);
1276         }
1277         return ret;
1278 }
1279
1280 static noinline u32 extent_data_ref_count(struct btrfs_root *root,
1281                                           struct btrfs_path *path,
1282                                           struct btrfs_extent_inline_ref *iref)
1283 {
1284         struct btrfs_key key;
1285         struct extent_buffer *leaf;
1286         struct btrfs_extent_data_ref *ref1;
1287         struct btrfs_shared_data_ref *ref2;
1288         u32 num_refs = 0;
1289
1290         leaf = path->nodes[0];
1291         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1292         if (iref) {
1293                 if (btrfs_extent_inline_ref_type(leaf, iref) ==
1294                     BTRFS_EXTENT_DATA_REF_KEY) {
1295                         ref1 = (struct btrfs_extent_data_ref *)(&iref->offset);
1296                         num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1297                 } else {
1298                         ref2 = (struct btrfs_shared_data_ref *)(iref + 1);
1299                         num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1300                 }
1301         } else if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
1302                 ref1 = btrfs_item_ptr(leaf, path->slots[0],
1303                                       struct btrfs_extent_data_ref);
1304                 num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1305         } else if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
1306                 ref2 = btrfs_item_ptr(leaf, path->slots[0],
1307                                       struct btrfs_shared_data_ref);
1308                 num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1309 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1310         } else if (key.type == BTRFS_EXTENT_REF_V0_KEY) {
1311                 struct btrfs_extent_ref_v0 *ref0;
1312                 ref0 = btrfs_item_ptr(leaf, path->slots[0],
1313                                       struct btrfs_extent_ref_v0);
1314                 num_refs = btrfs_ref_count_v0(leaf, ref0);
1315 #endif
1316         } else {
1317                 WARN_ON(1);
1318         }
1319         return num_refs;
1320 }
1321
1322 static noinline int lookup_tree_block_ref(struct btrfs_trans_handle *trans,
1323                                           struct btrfs_root *root,
1324                                           struct btrfs_path *path,
1325                                           u64 bytenr, u64 parent,
1326                                           u64 root_objectid)
1327 {
1328         struct btrfs_key key;
1329         int ret;
1330
1331         key.objectid = bytenr;
1332         if (parent) {
1333                 key.type = BTRFS_SHARED_BLOCK_REF_KEY;
1334                 key.offset = parent;
1335         } else {
1336                 key.type = BTRFS_TREE_BLOCK_REF_KEY;
1337                 key.offset = root_objectid;
1338         }
1339
1340         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1341         if (ret > 0)
1342                 ret = -ENOENT;
1343 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1344         if (ret == -ENOENT && parent) {
1345                 btrfs_release_path(path);
1346                 key.type = BTRFS_EXTENT_REF_V0_KEY;
1347                 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1348                 if (ret > 0)
1349                         ret = -ENOENT;
1350         }
1351 #endif
1352         return ret;
1353 }
1354
1355 static noinline int insert_tree_block_ref(struct btrfs_trans_handle *trans,
1356                                           struct btrfs_root *root,
1357                                           struct btrfs_path *path,
1358                                           u64 bytenr, u64 parent,
1359                                           u64 root_objectid)
1360 {
1361         struct btrfs_key key;
1362         int ret;
1363
1364         key.objectid = bytenr;
1365         if (parent) {
1366                 key.type = BTRFS_SHARED_BLOCK_REF_KEY;
1367                 key.offset = parent;
1368         } else {
1369                 key.type = BTRFS_TREE_BLOCK_REF_KEY;
1370                 key.offset = root_objectid;
1371         }
1372
1373         ret = btrfs_insert_empty_item(trans, root, path, &key, 0);
1374         btrfs_release_path(path);
1375         return ret;
1376 }
1377
1378 static inline int extent_ref_type(u64 parent, u64 owner)
1379 {
1380         int type;
1381         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1382                 if (parent > 0)
1383                         type = BTRFS_SHARED_BLOCK_REF_KEY;
1384                 else
1385                         type = BTRFS_TREE_BLOCK_REF_KEY;
1386         } else {
1387                 if (parent > 0)
1388                         type = BTRFS_SHARED_DATA_REF_KEY;
1389                 else
1390                         type = BTRFS_EXTENT_DATA_REF_KEY;
1391         }
1392         return type;
1393 }
1394
1395 static int find_next_key(struct btrfs_path *path, int level,
1396                          struct btrfs_key *key)
1397
1398 {
1399         for (; level < BTRFS_MAX_LEVEL; level++) {
1400                 if (!path->nodes[level])
1401                         break;
1402                 if (path->slots[level] + 1 >=
1403                     btrfs_header_nritems(path->nodes[level]))
1404                         continue;
1405                 if (level == 0)
1406                         btrfs_item_key_to_cpu(path->nodes[level], key,
1407                                               path->slots[level] + 1);
1408                 else
1409                         btrfs_node_key_to_cpu(path->nodes[level], key,
1410                                               path->slots[level] + 1);
1411                 return 0;
1412         }
1413         return 1;
1414 }
1415
1416 /*
1417  * look for inline back ref. if back ref is found, *ref_ret is set
1418  * to the address of inline back ref, and 0 is returned.
1419  *
1420  * if back ref isn't found, *ref_ret is set to the address where it
1421  * should be inserted, and -ENOENT is returned.
1422  *
1423  * if insert is true and there are too many inline back refs, the path
1424  * points to the extent item, and -EAGAIN is returned.
1425  *
1426  * NOTE: inline back refs are ordered in the same way that back ref
1427  *       items in the tree are ordered.
1428  */
1429 static noinline_for_stack
1430 int lookup_inline_extent_backref(struct btrfs_trans_handle *trans,
1431                                  struct btrfs_root *root,
1432                                  struct btrfs_path *path,
1433                                  struct btrfs_extent_inline_ref **ref_ret,
1434                                  u64 bytenr, u64 num_bytes,
1435                                  u64 parent, u64 root_objectid,
1436                                  u64 owner, u64 offset, int insert)
1437 {
1438         struct btrfs_key key;
1439         struct extent_buffer *leaf;
1440         struct btrfs_extent_item *ei;
1441         struct btrfs_extent_inline_ref *iref;
1442         u64 flags;
1443         u64 item_size;
1444         unsigned long ptr;
1445         unsigned long end;
1446         int extra_size;
1447         int type;
1448         int want;
1449         int ret;
1450         int err = 0;
1451
1452         key.objectid = bytenr;
1453         key.type = BTRFS_EXTENT_ITEM_KEY;
1454         key.offset = num_bytes;
1455
1456         want = extent_ref_type(parent, owner);
1457         if (insert) {
1458                 extra_size = btrfs_extent_inline_ref_size(want);
1459                 path->keep_locks = 1;
1460         } else
1461                 extra_size = -1;
1462         ret = btrfs_search_slot(trans, root, &key, path, extra_size, 1);
1463         if (ret < 0) {
1464                 err = ret;
1465                 goto out;
1466         }
1467         if (ret && !insert) {
1468                 err = -ENOENT;
1469                 goto out;
1470         }
1471         BUG_ON(ret); /* Corruption */
1472
1473         leaf = path->nodes[0];
1474         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1475 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1476         if (item_size < sizeof(*ei)) {
1477                 if (!insert) {
1478                         err = -ENOENT;
1479                         goto out;
1480                 }
1481                 ret = convert_extent_item_v0(trans, root, path, owner,
1482                                              extra_size);
1483                 if (ret < 0) {
1484                         err = ret;
1485                         goto out;
1486                 }
1487                 leaf = path->nodes[0];
1488                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1489         }
1490 #endif
1491         BUG_ON(item_size < sizeof(*ei));
1492
1493         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1494         flags = btrfs_extent_flags(leaf, ei);
1495
1496         ptr = (unsigned long)(ei + 1);
1497         end = (unsigned long)ei + item_size;
1498
1499         if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
1500                 ptr += sizeof(struct btrfs_tree_block_info);
1501                 BUG_ON(ptr > end);
1502         } else {
1503                 BUG_ON(!(flags & BTRFS_EXTENT_FLAG_DATA));
1504         }
1505
1506         err = -ENOENT;
1507         while (1) {
1508                 if (ptr >= end) {
1509                         WARN_ON(ptr > end);
1510                         break;
1511                 }
1512                 iref = (struct btrfs_extent_inline_ref *)ptr;
1513                 type = btrfs_extent_inline_ref_type(leaf, iref);
1514                 if (want < type)
1515                         break;
1516                 if (want > type) {
1517                         ptr += btrfs_extent_inline_ref_size(type);
1518                         continue;
1519                 }
1520
1521                 if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1522                         struct btrfs_extent_data_ref *dref;
1523                         dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1524                         if (match_extent_data_ref(leaf, dref, root_objectid,
1525                                                   owner, offset)) {
1526                                 err = 0;
1527                                 break;
1528                         }
1529                         if (hash_extent_data_ref_item(leaf, dref) <
1530                             hash_extent_data_ref(root_objectid, owner, offset))
1531                                 break;
1532                 } else {
1533                         u64 ref_offset;
1534                         ref_offset = btrfs_extent_inline_ref_offset(leaf, iref);
1535                         if (parent > 0) {
1536                                 if (parent == ref_offset) {
1537                                         err = 0;
1538                                         break;
1539                                 }
1540                                 if (ref_offset < parent)
1541                                         break;
1542                         } else {
1543                                 if (root_objectid == ref_offset) {
1544                                         err = 0;
1545                                         break;
1546                                 }
1547                                 if (ref_offset < root_objectid)
1548                                         break;
1549                         }
1550                 }
1551                 ptr += btrfs_extent_inline_ref_size(type);
1552         }
1553         if (err == -ENOENT && insert) {
1554                 if (item_size + extra_size >=
1555                     BTRFS_MAX_EXTENT_ITEM_SIZE(root)) {
1556                         err = -EAGAIN;
1557                         goto out;
1558                 }
1559                 /*
1560                  * To add new inline back ref, we have to make sure
1561                  * there is no corresponding back ref item.
1562                  * For simplicity, we just do not add new inline back
1563                  * ref if there is any kind of item for this block
1564                  */
1565                 if (find_next_key(path, 0, &key) == 0 &&
1566                     key.objectid == bytenr &&
1567                     key.type < BTRFS_BLOCK_GROUP_ITEM_KEY) {
1568                         err = -EAGAIN;
1569                         goto out;
1570                 }
1571         }
1572         *ref_ret = (struct btrfs_extent_inline_ref *)ptr;
1573 out:
1574         if (insert) {
1575                 path->keep_locks = 0;
1576                 btrfs_unlock_up_safe(path, 1);
1577         }
1578         return err;
1579 }
1580
1581 /*
1582  * helper to add new inline back ref
1583  */
1584 static noinline_for_stack
1585 void setup_inline_extent_backref(struct btrfs_trans_handle *trans,
1586                                  struct btrfs_root *root,
1587                                  struct btrfs_path *path,
1588                                  struct btrfs_extent_inline_ref *iref,
1589                                  u64 parent, u64 root_objectid,
1590                                  u64 owner, u64 offset, int refs_to_add,
1591                                  struct btrfs_delayed_extent_op *extent_op)
1592 {
1593         struct extent_buffer *leaf;
1594         struct btrfs_extent_item *ei;
1595         unsigned long ptr;
1596         unsigned long end;
1597         unsigned long item_offset;
1598         u64 refs;
1599         int size;
1600         int type;
1601
1602         leaf = path->nodes[0];
1603         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1604         item_offset = (unsigned long)iref - (unsigned long)ei;
1605
1606         type = extent_ref_type(parent, owner);
1607         size = btrfs_extent_inline_ref_size(type);
1608
1609         btrfs_extend_item(trans, root, path, size);
1610
1611         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1612         refs = btrfs_extent_refs(leaf, ei);
1613         refs += refs_to_add;
1614         btrfs_set_extent_refs(leaf, ei, refs);
1615         if (extent_op)
1616                 __run_delayed_extent_op(extent_op, leaf, ei);
1617
1618         ptr = (unsigned long)ei + item_offset;
1619         end = (unsigned long)ei + btrfs_item_size_nr(leaf, path->slots[0]);
1620         if (ptr < end - size)
1621                 memmove_extent_buffer(leaf, ptr + size, ptr,
1622                                       end - size - ptr);
1623
1624         iref = (struct btrfs_extent_inline_ref *)ptr;
1625         btrfs_set_extent_inline_ref_type(leaf, iref, type);
1626         if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1627                 struct btrfs_extent_data_ref *dref;
1628                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1629                 btrfs_set_extent_data_ref_root(leaf, dref, root_objectid);
1630                 btrfs_set_extent_data_ref_objectid(leaf, dref, owner);
1631                 btrfs_set_extent_data_ref_offset(leaf, dref, offset);
1632                 btrfs_set_extent_data_ref_count(leaf, dref, refs_to_add);
1633         } else if (type == BTRFS_SHARED_DATA_REF_KEY) {
1634                 struct btrfs_shared_data_ref *sref;
1635                 sref = (struct btrfs_shared_data_ref *)(iref + 1);
1636                 btrfs_set_shared_data_ref_count(leaf, sref, refs_to_add);
1637                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
1638         } else if (type == BTRFS_SHARED_BLOCK_REF_KEY) {
1639                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
1640         } else {
1641                 btrfs_set_extent_inline_ref_offset(leaf, iref, root_objectid);
1642         }
1643         btrfs_mark_buffer_dirty(leaf);
1644 }
1645
1646 static int lookup_extent_backref(struct btrfs_trans_handle *trans,
1647                                  struct btrfs_root *root,
1648                                  struct btrfs_path *path,
1649                                  struct btrfs_extent_inline_ref **ref_ret,
1650                                  u64 bytenr, u64 num_bytes, u64 parent,
1651                                  u64 root_objectid, u64 owner, u64 offset)
1652 {
1653         int ret;
1654
1655         ret = lookup_inline_extent_backref(trans, root, path, ref_ret,
1656                                            bytenr, num_bytes, parent,
1657                                            root_objectid, owner, offset, 0);
1658         if (ret != -ENOENT)
1659                 return ret;
1660
1661         btrfs_release_path(path);
1662         *ref_ret = NULL;
1663
1664         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1665                 ret = lookup_tree_block_ref(trans, root, path, bytenr, parent,
1666                                             root_objectid);
1667         } else {
1668                 ret = lookup_extent_data_ref(trans, root, path, bytenr, parent,
1669                                              root_objectid, owner, offset);
1670         }
1671         return ret;
1672 }
1673
1674 /*
1675  * helper to update/remove inline back ref
1676  */
1677 static noinline_for_stack
1678 void update_inline_extent_backref(struct btrfs_trans_handle *trans,
1679                                   struct btrfs_root *root,
1680                                   struct btrfs_path *path,
1681                                   struct btrfs_extent_inline_ref *iref,
1682                                   int refs_to_mod,
1683                                   struct btrfs_delayed_extent_op *extent_op)
1684 {
1685         struct extent_buffer *leaf;
1686         struct btrfs_extent_item *ei;
1687         struct btrfs_extent_data_ref *dref = NULL;
1688         struct btrfs_shared_data_ref *sref = NULL;
1689         unsigned long ptr;
1690         unsigned long end;
1691         u32 item_size;
1692         int size;
1693         int type;
1694         u64 refs;
1695
1696         leaf = path->nodes[0];
1697         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1698         refs = btrfs_extent_refs(leaf, ei);
1699         WARN_ON(refs_to_mod < 0 && refs + refs_to_mod <= 0);
1700         refs += refs_to_mod;
1701         btrfs_set_extent_refs(leaf, ei, refs);
1702         if (extent_op)
1703                 __run_delayed_extent_op(extent_op, leaf, ei);
1704
1705         type = btrfs_extent_inline_ref_type(leaf, iref);
1706
1707         if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1708                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1709                 refs = btrfs_extent_data_ref_count(leaf, dref);
1710         } else if (type == BTRFS_SHARED_DATA_REF_KEY) {
1711                 sref = (struct btrfs_shared_data_ref *)(iref + 1);
1712                 refs = btrfs_shared_data_ref_count(leaf, sref);
1713         } else {
1714                 refs = 1;
1715                 BUG_ON(refs_to_mod != -1);
1716         }
1717
1718         BUG_ON(refs_to_mod < 0 && refs < -refs_to_mod);
1719         refs += refs_to_mod;
1720
1721         if (refs > 0) {
1722                 if (type == BTRFS_EXTENT_DATA_REF_KEY)
1723                         btrfs_set_extent_data_ref_count(leaf, dref, refs);
1724                 else
1725                         btrfs_set_shared_data_ref_count(leaf, sref, refs);
1726         } else {
1727                 size =  btrfs_extent_inline_ref_size(type);
1728                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1729                 ptr = (unsigned long)iref;
1730                 end = (unsigned long)ei + item_size;
1731                 if (ptr + size < end)
1732                         memmove_extent_buffer(leaf, ptr, ptr + size,
1733                                               end - ptr - size);
1734                 item_size -= size;
1735                 btrfs_truncate_item(trans, root, path, item_size, 1);
1736         }
1737         btrfs_mark_buffer_dirty(leaf);
1738 }
1739
1740 static noinline_for_stack
1741 int insert_inline_extent_backref(struct btrfs_trans_handle *trans,
1742                                  struct btrfs_root *root,
1743                                  struct btrfs_path *path,
1744                                  u64 bytenr, u64 num_bytes, u64 parent,
1745                                  u64 root_objectid, u64 owner,
1746                                  u64 offset, int refs_to_add,
1747                                  struct btrfs_delayed_extent_op *extent_op)
1748 {
1749         struct btrfs_extent_inline_ref *iref;
1750         int ret;
1751
1752         ret = lookup_inline_extent_backref(trans, root, path, &iref,
1753                                            bytenr, num_bytes, parent,
1754                                            root_objectid, owner, offset, 1);
1755         if (ret == 0) {
1756                 BUG_ON(owner < BTRFS_FIRST_FREE_OBJECTID);
1757                 update_inline_extent_backref(trans, root, path, iref,
1758                                              refs_to_add, extent_op);
1759         } else if (ret == -ENOENT) {
1760                 setup_inline_extent_backref(trans, root, path, iref, parent,
1761                                             root_objectid, owner, offset,
1762                                             refs_to_add, extent_op);
1763                 ret = 0;
1764         }
1765         return ret;
1766 }
1767
1768 static int insert_extent_backref(struct btrfs_trans_handle *trans,
1769                                  struct btrfs_root *root,
1770                                  struct btrfs_path *path,
1771                                  u64 bytenr, u64 parent, u64 root_objectid,
1772                                  u64 owner, u64 offset, int refs_to_add)
1773 {
1774         int ret;
1775         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1776                 BUG_ON(refs_to_add != 1);
1777                 ret = insert_tree_block_ref(trans, root, path, bytenr,
1778                                             parent, root_objectid);
1779         } else {
1780                 ret = insert_extent_data_ref(trans, root, path, bytenr,
1781                                              parent, root_objectid,
1782                                              owner, offset, refs_to_add);
1783         }
1784         return ret;
1785 }
1786
1787 static int remove_extent_backref(struct btrfs_trans_handle *trans,
1788                                  struct btrfs_root *root,
1789                                  struct btrfs_path *path,
1790                                  struct btrfs_extent_inline_ref *iref,
1791                                  int refs_to_drop, int is_data)
1792 {
1793         int ret = 0;
1794
1795         BUG_ON(!is_data && refs_to_drop != 1);
1796         if (iref) {
1797                 update_inline_extent_backref(trans, root, path, iref,
1798                                              -refs_to_drop, NULL);
1799         } else if (is_data) {
1800                 ret = remove_extent_data_ref(trans, root, path, refs_to_drop);
1801         } else {
1802                 ret = btrfs_del_item(trans, root, path);
1803         }
1804         return ret;
1805 }
1806
1807 static int btrfs_issue_discard(struct block_device *bdev,
1808                                 u64 start, u64 len)
1809 {
1810         return blkdev_issue_discard(bdev, start >> 9, len >> 9, GFP_NOFS, 0);
1811 }
1812
1813 static int btrfs_discard_extent(struct btrfs_root *root, u64 bytenr,
1814                                 u64 num_bytes, u64 *actual_bytes)
1815 {
1816         int ret;
1817         u64 discarded_bytes = 0;
1818         struct btrfs_bio *bbio = NULL;
1819
1820
1821         /* Tell the block device(s) that the sectors can be discarded */
1822         ret = btrfs_map_block(root->fs_info, REQ_DISCARD,
1823                               bytenr, &num_bytes, &bbio, 0);
1824         /* Error condition is -ENOMEM */
1825         if (!ret) {
1826                 struct btrfs_bio_stripe *stripe = bbio->stripes;
1827                 int i;
1828
1829
1830                 for (i = 0; i < bbio->num_stripes; i++, stripe++) {
1831                         if (!stripe->dev->can_discard)
1832                                 continue;
1833
1834                         ret = btrfs_issue_discard(stripe->dev->bdev,
1835                                                   stripe->physical,
1836                                                   stripe->length);
1837                         if (!ret)
1838                                 discarded_bytes += stripe->length;
1839                         else if (ret != -EOPNOTSUPP)
1840                                 break; /* Logic errors or -ENOMEM, or -EIO but I don't know how that could happen JDM */
1841
1842                         /*
1843                          * Just in case we get back EOPNOTSUPP for some reason,
1844                          * just ignore the return value so we don't screw up
1845                          * people calling discard_extent.
1846                          */
1847                         ret = 0;
1848                 }
1849                 kfree(bbio);
1850         }
1851
1852         if (actual_bytes)
1853                 *actual_bytes = discarded_bytes;
1854
1855
1856         if (ret == -EOPNOTSUPP)
1857                 ret = 0;
1858         return ret;
1859 }
1860
1861 /* Can return -ENOMEM */
1862 int btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
1863                          struct btrfs_root *root,
1864                          u64 bytenr, u64 num_bytes, u64 parent,
1865                          u64 root_objectid, u64 owner, u64 offset, int for_cow)
1866 {
1867         int ret;
1868         struct btrfs_fs_info *fs_info = root->fs_info;
1869
1870         BUG_ON(owner < BTRFS_FIRST_FREE_OBJECTID &&
1871                root_objectid == BTRFS_TREE_LOG_OBJECTID);
1872
1873         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1874                 ret = btrfs_add_delayed_tree_ref(fs_info, trans, bytenr,
1875                                         num_bytes,
1876                                         parent, root_objectid, (int)owner,
1877                                         BTRFS_ADD_DELAYED_REF, NULL, for_cow);
1878         } else {
1879                 ret = btrfs_add_delayed_data_ref(fs_info, trans, bytenr,
1880                                         num_bytes,
1881                                         parent, root_objectid, owner, offset,
1882                                         BTRFS_ADD_DELAYED_REF, NULL, for_cow);
1883         }
1884         return ret;
1885 }
1886
1887 static int __btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
1888                                   struct btrfs_root *root,
1889                                   u64 bytenr, u64 num_bytes,
1890                                   u64 parent, u64 root_objectid,
1891                                   u64 owner, u64 offset, int refs_to_add,
1892                                   struct btrfs_delayed_extent_op *extent_op)
1893 {
1894         struct btrfs_path *path;
1895         struct extent_buffer *leaf;
1896         struct btrfs_extent_item *item;
1897         u64 refs;
1898         int ret;
1899         int err = 0;
1900
1901         path = btrfs_alloc_path();
1902         if (!path)
1903                 return -ENOMEM;
1904
1905         path->reada = 1;
1906         path->leave_spinning = 1;
1907         /* this will setup the path even if it fails to insert the back ref */
1908         ret = insert_inline_extent_backref(trans, root->fs_info->extent_root,
1909                                            path, bytenr, num_bytes, parent,
1910                                            root_objectid, owner, offset,
1911                                            refs_to_add, extent_op);
1912         if (ret == 0)
1913                 goto out;
1914
1915         if (ret != -EAGAIN) {
1916                 err = ret;
1917                 goto out;
1918         }
1919
1920         leaf = path->nodes[0];
1921         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1922         refs = btrfs_extent_refs(leaf, item);
1923         btrfs_set_extent_refs(leaf, item, refs + refs_to_add);
1924         if (extent_op)
1925                 __run_delayed_extent_op(extent_op, leaf, item);
1926
1927         btrfs_mark_buffer_dirty(leaf);
1928         btrfs_release_path(path);
1929
1930         path->reada = 1;
1931         path->leave_spinning = 1;
1932
1933         /* now insert the actual backref */
1934         ret = insert_extent_backref(trans, root->fs_info->extent_root,
1935                                     path, bytenr, parent, root_objectid,
1936                                     owner, offset, refs_to_add);
1937         if (ret)
1938                 btrfs_abort_transaction(trans, root, ret);
1939 out:
1940         btrfs_free_path(path);
1941         return err;
1942 }
1943
1944 static int run_delayed_data_ref(struct btrfs_trans_handle *trans,
1945                                 struct btrfs_root *root,
1946                                 struct btrfs_delayed_ref_node *node,
1947                                 struct btrfs_delayed_extent_op *extent_op,
1948                                 int insert_reserved)
1949 {
1950         int ret = 0;
1951         struct btrfs_delayed_data_ref *ref;
1952         struct btrfs_key ins;
1953         u64 parent = 0;
1954         u64 ref_root = 0;
1955         u64 flags = 0;
1956
1957         ins.objectid = node->bytenr;
1958         ins.offset = node->num_bytes;
1959         ins.type = BTRFS_EXTENT_ITEM_KEY;
1960
1961         ref = btrfs_delayed_node_to_data_ref(node);
1962         if (node->type == BTRFS_SHARED_DATA_REF_KEY)
1963                 parent = ref->parent;
1964         else
1965                 ref_root = ref->root;
1966
1967         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
1968                 if (extent_op) {
1969                         BUG_ON(extent_op->update_key);
1970                         flags |= extent_op->flags_to_set;
1971                 }
1972                 ret = alloc_reserved_file_extent(trans, root,
1973                                                  parent, ref_root, flags,
1974                                                  ref->objectid, ref->offset,
1975                                                  &ins, node->ref_mod);
1976         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
1977                 ret = __btrfs_inc_extent_ref(trans, root, node->bytenr,
1978                                              node->num_bytes, parent,
1979                                              ref_root, ref->objectid,
1980                                              ref->offset, node->ref_mod,
1981                                              extent_op);
1982         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
1983                 ret = __btrfs_free_extent(trans, root, node->bytenr,
1984                                           node->num_bytes, parent,
1985                                           ref_root, ref->objectid,
1986                                           ref->offset, node->ref_mod,
1987                                           extent_op);
1988         } else {
1989                 BUG();
1990         }
1991         return ret;
1992 }
1993
1994 static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op,
1995                                     struct extent_buffer *leaf,
1996                                     struct btrfs_extent_item *ei)
1997 {
1998         u64 flags = btrfs_extent_flags(leaf, ei);
1999         if (extent_op->update_flags) {
2000                 flags |= extent_op->flags_to_set;
2001                 btrfs_set_extent_flags(leaf, ei, flags);
2002         }
2003
2004         if (extent_op->update_key) {
2005                 struct btrfs_tree_block_info *bi;
2006                 BUG_ON(!(flags & BTRFS_EXTENT_FLAG_TREE_BLOCK));
2007                 bi = (struct btrfs_tree_block_info *)(ei + 1);
2008                 btrfs_set_tree_block_key(leaf, bi, &extent_op->key);
2009         }
2010 }
2011
2012 static int run_delayed_extent_op(struct btrfs_trans_handle *trans,
2013                                  struct btrfs_root *root,
2014                                  struct btrfs_delayed_ref_node *node,
2015                                  struct btrfs_delayed_extent_op *extent_op)
2016 {
2017         struct btrfs_key key;
2018         struct btrfs_path *path;
2019         struct btrfs_extent_item *ei;
2020         struct extent_buffer *leaf;
2021         u32 item_size;
2022         int ret;
2023         int err = 0;
2024
2025         if (trans->aborted)
2026                 return 0;
2027
2028         path = btrfs_alloc_path();
2029         if (!path)
2030                 return -ENOMEM;
2031
2032         key.objectid = node->bytenr;
2033         key.type = BTRFS_EXTENT_ITEM_KEY;
2034         key.offset = node->num_bytes;
2035
2036         path->reada = 1;
2037         path->leave_spinning = 1;
2038         ret = btrfs_search_slot(trans, root->fs_info->extent_root, &key,
2039                                 path, 0, 1);
2040         if (ret < 0) {
2041                 err = ret;
2042                 goto out;
2043         }
2044         if (ret > 0) {
2045                 err = -EIO;
2046                 goto out;
2047         }
2048
2049         leaf = path->nodes[0];
2050         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2051 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
2052         if (item_size < sizeof(*ei)) {
2053                 ret = convert_extent_item_v0(trans, root->fs_info->extent_root,
2054                                              path, (u64)-1, 0);
2055                 if (ret < 0) {
2056                         err = ret;
2057                         goto out;
2058                 }
2059                 leaf = path->nodes[0];
2060                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2061         }
2062 #endif
2063         BUG_ON(item_size < sizeof(*ei));
2064         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2065         __run_delayed_extent_op(extent_op, leaf, ei);
2066
2067         btrfs_mark_buffer_dirty(leaf);
2068 out:
2069         btrfs_free_path(path);
2070         return err;
2071 }
2072
2073 static int run_delayed_tree_ref(struct btrfs_trans_handle *trans,
2074                                 struct btrfs_root *root,
2075                                 struct btrfs_delayed_ref_node *node,
2076                                 struct btrfs_delayed_extent_op *extent_op,
2077                                 int insert_reserved)
2078 {
2079         int ret = 0;
2080         struct btrfs_delayed_tree_ref *ref;
2081         struct btrfs_key ins;
2082         u64 parent = 0;
2083         u64 ref_root = 0;
2084
2085         ins.objectid = node->bytenr;
2086         ins.offset = node->num_bytes;
2087         ins.type = BTRFS_EXTENT_ITEM_KEY;
2088
2089         ref = btrfs_delayed_node_to_tree_ref(node);
2090         if (node->type == BTRFS_SHARED_BLOCK_REF_KEY)
2091                 parent = ref->parent;
2092         else
2093                 ref_root = ref->root;
2094
2095         BUG_ON(node->ref_mod != 1);
2096         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
2097                 BUG_ON(!extent_op || !extent_op->update_flags ||
2098                        !extent_op->update_key);
2099                 ret = alloc_reserved_tree_block(trans, root,
2100                                                 parent, ref_root,
2101                                                 extent_op->flags_to_set,
2102                                                 &extent_op->key,
2103                                                 ref->level, &ins);
2104         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
2105                 ret = __btrfs_inc_extent_ref(trans, root, node->bytenr,
2106                                              node->num_bytes, parent, ref_root,
2107                                              ref->level, 0, 1, extent_op);
2108         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
2109                 ret = __btrfs_free_extent(trans, root, node->bytenr,
2110                                           node->num_bytes, parent, ref_root,
2111                                           ref->level, 0, 1, extent_op);
2112         } else {
2113                 BUG();
2114         }
2115         return ret;
2116 }
2117
2118 /* helper function to actually process a single delayed ref entry */
2119 static int run_one_delayed_ref(struct btrfs_trans_handle *trans,
2120                                struct btrfs_root *root,
2121                                struct btrfs_delayed_ref_node *node,
2122                                struct btrfs_delayed_extent_op *extent_op,
2123                                int insert_reserved)
2124 {
2125         int ret = 0;
2126
2127         if (trans->aborted)
2128                 return 0;
2129
2130         if (btrfs_delayed_ref_is_head(node)) {
2131                 struct btrfs_delayed_ref_head *head;
2132                 /*
2133                  * we've hit the end of the chain and we were supposed
2134                  * to insert this extent into the tree.  But, it got
2135                  * deleted before we ever needed to insert it, so all
2136                  * we have to do is clean up the accounting
2137                  */
2138                 BUG_ON(extent_op);
2139                 head = btrfs_delayed_node_to_head(node);
2140                 if (insert_reserved) {
2141                         btrfs_pin_extent(root, node->bytenr,
2142                                          node->num_bytes, 1);
2143                         if (head->is_data) {
2144                                 ret = btrfs_del_csums(trans, root,
2145                                                       node->bytenr,
2146                                                       node->num_bytes);
2147                         }
2148                 }
2149                 return ret;
2150         }
2151
2152         if (node->type == BTRFS_TREE_BLOCK_REF_KEY ||
2153             node->type == BTRFS_SHARED_BLOCK_REF_KEY)
2154                 ret = run_delayed_tree_ref(trans, root, node, extent_op,
2155                                            insert_reserved);
2156         else if (node->type == BTRFS_EXTENT_DATA_REF_KEY ||
2157                  node->type == BTRFS_SHARED_DATA_REF_KEY)
2158                 ret = run_delayed_data_ref(trans, root, node, extent_op,
2159                                            insert_reserved);
2160         else
2161                 BUG();
2162         return ret;
2163 }
2164
2165 static noinline struct btrfs_delayed_ref_node *
2166 select_delayed_ref(struct btrfs_delayed_ref_head *head)
2167 {
2168         struct rb_node *node;
2169         struct btrfs_delayed_ref_node *ref;
2170         int action = BTRFS_ADD_DELAYED_REF;
2171 again:
2172         /*
2173          * select delayed ref of type BTRFS_ADD_DELAYED_REF first.
2174          * this prevents ref count from going down to zero when
2175          * there still are pending delayed ref.
2176          */
2177         node = rb_prev(&head->node.rb_node);
2178         while (1) {
2179                 if (!node)
2180                         break;
2181                 ref = rb_entry(node, struct btrfs_delayed_ref_node,
2182                                 rb_node);
2183                 if (ref->bytenr != head->node.bytenr)
2184                         break;
2185                 if (ref->action == action)
2186                         return ref;
2187                 node = rb_prev(node);
2188         }
2189         if (action == BTRFS_ADD_DELAYED_REF) {
2190                 action = BTRFS_DROP_DELAYED_REF;
2191                 goto again;
2192         }
2193         return NULL;
2194 }
2195
2196 /*
2197  * Returns 0 on success or if called with an already aborted transaction.
2198  * Returns -ENOMEM or -EIO on failure and will abort the transaction.
2199  */
2200 static noinline int run_clustered_refs(struct btrfs_trans_handle *trans,
2201                                        struct btrfs_root *root,
2202                                        struct list_head *cluster)
2203 {
2204         struct btrfs_delayed_ref_root *delayed_refs;
2205         struct btrfs_delayed_ref_node *ref;
2206         struct btrfs_delayed_ref_head *locked_ref = NULL;
2207         struct btrfs_delayed_extent_op *extent_op;
2208         struct btrfs_fs_info *fs_info = root->fs_info;
2209         int ret;
2210         int count = 0;
2211         int must_insert_reserved = 0;
2212
2213         delayed_refs = &trans->transaction->delayed_refs;
2214         while (1) {
2215                 if (!locked_ref) {
2216                         /* pick a new head ref from the cluster list */
2217                         if (list_empty(cluster))
2218                                 break;
2219
2220                         locked_ref = list_entry(cluster->next,
2221                                      struct btrfs_delayed_ref_head, cluster);
2222
2223                         /* grab the lock that says we are going to process
2224                          * all the refs for this head */
2225                         ret = btrfs_delayed_ref_lock(trans, locked_ref);
2226
2227                         /*
2228                          * we may have dropped the spin lock to get the head
2229                          * mutex lock, and that might have given someone else
2230                          * time to free the head.  If that's true, it has been
2231                          * removed from our list and we can move on.
2232                          */
2233                         if (ret == -EAGAIN) {
2234                                 locked_ref = NULL;
2235                                 count++;
2236                                 continue;
2237                         }
2238                 }
2239
2240                 /*
2241                  * We need to try and merge add/drops of the same ref since we
2242                  * can run into issues with relocate dropping the implicit ref
2243                  * and then it being added back again before the drop can
2244                  * finish.  If we merged anything we need to re-loop so we can
2245                  * get a good ref.
2246                  */
2247                 btrfs_merge_delayed_refs(trans, fs_info, delayed_refs,
2248                                          locked_ref);
2249
2250                 /*
2251                  * locked_ref is the head node, so we have to go one
2252                  * node back for any delayed ref updates
2253                  */
2254                 ref = select_delayed_ref(locked_ref);
2255
2256                 if (ref && ref->seq &&
2257                     btrfs_check_delayed_seq(fs_info, delayed_refs, ref->seq)) {
2258                         /*
2259                          * there are still refs with lower seq numbers in the
2260                          * process of being added. Don't run this ref yet.
2261                          */
2262                         list_del_init(&locked_ref->cluster);
2263                         btrfs_delayed_ref_unlock(locked_ref);
2264                         locked_ref = NULL;
2265                         delayed_refs->num_heads_ready++;
2266                         spin_unlock(&delayed_refs->lock);
2267                         cond_resched();
2268                         spin_lock(&delayed_refs->lock);
2269                         continue;
2270                 }
2271
2272                 /*
2273                  * record the must insert reserved flag before we
2274                  * drop the spin lock.
2275                  */
2276                 must_insert_reserved = locked_ref->must_insert_reserved;
2277                 locked_ref->must_insert_reserved = 0;
2278
2279                 extent_op = locked_ref->extent_op;
2280                 locked_ref->extent_op = NULL;
2281
2282                 if (!ref) {
2283                         /* All delayed refs have been processed, Go ahead
2284                          * and send the head node to run_one_delayed_ref,
2285                          * so that any accounting fixes can happen
2286                          */
2287                         ref = &locked_ref->node;
2288
2289                         if (extent_op && must_insert_reserved) {
2290                                 btrfs_free_delayed_extent_op(extent_op);
2291                                 extent_op = NULL;
2292                         }
2293
2294                         if (extent_op) {
2295                                 spin_unlock(&delayed_refs->lock);
2296
2297                                 ret = run_delayed_extent_op(trans, root,
2298                                                             ref, extent_op);
2299                                 btrfs_free_delayed_extent_op(extent_op);
2300
2301                                 if (ret) {
2302                                         printk(KERN_DEBUG
2303                                                "btrfs: run_delayed_extent_op "
2304                                                "returned %d\n", ret);
2305                                         spin_lock(&delayed_refs->lock);
2306                                         btrfs_delayed_ref_unlock(locked_ref);
2307                                         return ret;
2308                                 }
2309
2310                                 goto next;
2311                         }
2312                 }
2313
2314                 ref->in_tree = 0;
2315                 rb_erase(&ref->rb_node, &delayed_refs->root);
2316                 delayed_refs->num_entries--;
2317                 if (!btrfs_delayed_ref_is_head(ref)) {
2318                         /*
2319                          * when we play the delayed ref, also correct the
2320                          * ref_mod on head
2321                          */
2322                         switch (ref->action) {
2323                         case BTRFS_ADD_DELAYED_REF:
2324                         case BTRFS_ADD_DELAYED_EXTENT:
2325                                 locked_ref->node.ref_mod -= ref->ref_mod;
2326                                 break;
2327                         case BTRFS_DROP_DELAYED_REF:
2328                                 locked_ref->node.ref_mod += ref->ref_mod;
2329                                 break;
2330                         default:
2331                                 WARN_ON(1);
2332                         }
2333                 }
2334                 spin_unlock(&delayed_refs->lock);
2335
2336                 ret = run_one_delayed_ref(trans, root, ref, extent_op,
2337                                           must_insert_reserved);
2338
2339                 btrfs_free_delayed_extent_op(extent_op);
2340                 if (ret) {
2341                         btrfs_delayed_ref_unlock(locked_ref);
2342                         btrfs_put_delayed_ref(ref);
2343                         printk(KERN_DEBUG
2344                                "btrfs: run_one_delayed_ref returned %d\n", ret);
2345                         spin_lock(&delayed_refs->lock);
2346                         return ret;
2347                 }
2348
2349                 /*
2350                  * If this node is a head, that means all the refs in this head
2351                  * have been dealt with, and we will pick the next head to deal
2352                  * with, so we must unlock the head and drop it from the cluster
2353                  * list before we release it.
2354                  */
2355                 if (btrfs_delayed_ref_is_head(ref)) {
2356                         list_del_init(&locked_ref->cluster);
2357                         btrfs_delayed_ref_unlock(locked_ref);
2358                         locked_ref = NULL;
2359                 }
2360                 btrfs_put_delayed_ref(ref);
2361                 count++;
2362 next:
2363                 cond_resched();
2364                 spin_lock(&delayed_refs->lock);
2365         }
2366         return count;
2367 }
2368
2369 #ifdef SCRAMBLE_DELAYED_REFS
2370 /*
2371  * Normally delayed refs get processed in ascending bytenr order. This
2372  * correlates in most cases to the order added. To expose dependencies on this
2373  * order, we start to process the tree in the middle instead of the beginning
2374  */
2375 static u64 find_middle(struct rb_root *root)
2376 {
2377         struct rb_node *n = root->rb_node;
2378         struct btrfs_delayed_ref_node *entry;
2379         int alt = 1;
2380         u64 middle;
2381         u64 first = 0, last = 0;
2382
2383         n = rb_first(root);
2384         if (n) {
2385                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2386                 first = entry->bytenr;
2387         }
2388         n = rb_last(root);
2389         if (n) {
2390                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2391                 last = entry->bytenr;
2392         }
2393         n = root->rb_node;
2394
2395         while (n) {
2396                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2397                 WARN_ON(!entry->in_tree);
2398
2399                 middle = entry->bytenr;
2400
2401                 if (alt)
2402                         n = n->rb_left;
2403                 else
2404                         n = n->rb_right;
2405
2406                 alt = 1 - alt;
2407         }
2408         return middle;
2409 }
2410 #endif
2411
2412 int btrfs_delayed_refs_qgroup_accounting(struct btrfs_trans_handle *trans,
2413                                          struct btrfs_fs_info *fs_info)
2414 {
2415         struct qgroup_update *qgroup_update;
2416         int ret = 0;
2417
2418         if (list_empty(&trans->qgroup_ref_list) !=
2419             !trans->delayed_ref_elem.seq) {
2420                 /* list without seq or seq without list */
2421                 printk(KERN_ERR "btrfs: qgroup accounting update error, list is%s empty, seq is %llu\n",
2422                         list_empty(&trans->qgroup_ref_list) ? "" : " not",
2423                         trans->delayed_ref_elem.seq);
2424                 BUG();
2425         }
2426
2427         if (!trans->delayed_ref_elem.seq)
2428                 return 0;
2429
2430         while (!list_empty(&trans->qgroup_ref_list)) {
2431                 qgroup_update = list_first_entry(&trans->qgroup_ref_list,
2432                                                  struct qgroup_update, list);
2433                 list_del(&qgroup_update->list);
2434                 if (!ret)
2435                         ret = btrfs_qgroup_account_ref(
2436                                         trans, fs_info, qgroup_update->node,
2437                                         qgroup_update->extent_op);
2438                 kfree(qgroup_update);
2439         }
2440
2441         btrfs_put_tree_mod_seq(fs_info, &trans->delayed_ref_elem);
2442
2443         return ret;
2444 }
2445
2446 static int refs_newer(struct btrfs_delayed_ref_root *delayed_refs, int seq,
2447                       int count)
2448 {
2449         int val = atomic_read(&delayed_refs->ref_seq);
2450
2451         if (val < seq || val >= seq + count)
2452                 return 1;
2453         return 0;
2454 }
2455
2456 /*
2457  * this starts processing the delayed reference count updates and
2458  * extent insertions we have queued up so far.  count can be
2459  * 0, which means to process everything in the tree at the start
2460  * of the run (but not newly added entries), or it can be some target
2461  * number you'd like to process.
2462  *
2463  * Returns 0 on success or if called with an aborted transaction
2464  * Returns <0 on error and aborts the transaction
2465  */
2466 int btrfs_run_delayed_refs(struct btrfs_trans_handle *trans,
2467                            struct btrfs_root *root, unsigned long count)
2468 {
2469         struct rb_node *node;
2470         struct btrfs_delayed_ref_root *delayed_refs;
2471         struct btrfs_delayed_ref_node *ref;
2472         struct list_head cluster;
2473         int ret;
2474         u64 delayed_start;
2475         int run_all = count == (unsigned long)-1;
2476         int run_most = 0;
2477         int loops;
2478
2479         /* We'll clean this up in btrfs_cleanup_transaction */
2480         if (trans->aborted)
2481                 return 0;
2482
2483         if (root == root->fs_info->extent_root)
2484                 root = root->fs_info->tree_root;
2485
2486         btrfs_delayed_refs_qgroup_accounting(trans, root->fs_info);
2487
2488         delayed_refs = &trans->transaction->delayed_refs;
2489         INIT_LIST_HEAD(&cluster);
2490         if (count == 0) {
2491                 count = delayed_refs->num_entries * 2;
2492                 run_most = 1;
2493         }
2494
2495         if (!run_all && !run_most) {
2496                 int old;
2497                 int seq = atomic_read(&delayed_refs->ref_seq);
2498
2499 progress:
2500                 old = atomic_cmpxchg(&delayed_refs->procs_running_refs, 0, 1);
2501                 if (old) {
2502                         DEFINE_WAIT(__wait);
2503                         if (delayed_refs->num_entries < 16348)
2504                                 return 0;
2505
2506                         prepare_to_wait(&delayed_refs->wait, &__wait,
2507                                         TASK_UNINTERRUPTIBLE);
2508
2509                         old = atomic_cmpxchg(&delayed_refs->procs_running_refs, 0, 1);
2510                         if (old) {
2511                                 schedule();
2512                                 finish_wait(&delayed_refs->wait, &__wait);
2513
2514                                 if (!refs_newer(delayed_refs, seq, 256))
2515                                         goto progress;
2516                                 else
2517                                         return 0;
2518                         } else {
2519                                 finish_wait(&delayed_refs->wait, &__wait);
2520                                 goto again;
2521                         }
2522                 }
2523
2524         } else {
2525                 atomic_inc(&delayed_refs->procs_running_refs);
2526         }
2527
2528 again:
2529         loops = 0;
2530         spin_lock(&delayed_refs->lock);
2531
2532 #ifdef SCRAMBLE_DELAYED_REFS
2533         delayed_refs->run_delayed_start = find_middle(&delayed_refs->root);
2534 #endif
2535
2536         while (1) {
2537                 if (!(run_all || run_most) &&
2538                     delayed_refs->num_heads_ready < 64)
2539                         break;
2540
2541                 /*
2542                  * go find something we can process in the rbtree.  We start at
2543                  * the beginning of the tree, and then build a cluster
2544                  * of refs to process starting at the first one we are able to
2545                  * lock
2546                  */
2547                 delayed_start = delayed_refs->run_delayed_start;
2548                 ret = btrfs_find_ref_cluster(trans, &cluster,
2549                                              delayed_refs->run_delayed_start);
2550                 if (ret)
2551                         break;
2552
2553                 ret = run_clustered_refs(trans, root, &cluster);
2554                 if (ret < 0) {
2555                         btrfs_release_ref_cluster(&cluster);
2556                         spin_unlock(&delayed_refs->lock);
2557                         btrfs_abort_transaction(trans, root, ret);
2558                         atomic_dec(&delayed_refs->procs_running_refs);
2559                         return ret;
2560                 }
2561
2562                 atomic_add(ret, &delayed_refs->ref_seq);
2563
2564                 count -= min_t(unsigned long, ret, count);
2565
2566                 if (count == 0)
2567                         break;
2568
2569                 if (delayed_start >= delayed_refs->run_delayed_start) {
2570                         if (loops == 0) {
2571                                 /*
2572                                  * btrfs_find_ref_cluster looped. let's do one
2573                                  * more cycle. if we don't run any delayed ref
2574                                  * during that cycle (because we can't because
2575                                  * all of them are blocked), bail out.
2576                                  */
2577                                 loops = 1;
2578                         } else {
2579                                 /*
2580                                  * no runnable refs left, stop trying
2581                                  */
2582                                 BUG_ON(run_all);
2583                                 break;
2584                         }
2585                 }
2586                 if (ret) {
2587                         /* refs were run, let's reset staleness detection */
2588                         loops = 0;
2589                 }
2590         }
2591
2592         if (run_all) {
2593                 if (!list_empty(&trans->new_bgs)) {
2594                         spin_unlock(&delayed_refs->lock);
2595                         btrfs_create_pending_block_groups(trans, root);
2596                         spin_lock(&delayed_refs->lock);
2597                 }
2598
2599                 node = rb_first(&delayed_refs->root);
2600                 if (!node)
2601                         goto out;
2602                 count = (unsigned long)-1;
2603
2604                 while (node) {
2605                         ref = rb_entry(node, struct btrfs_delayed_ref_node,
2606                                        rb_node);
2607                         if (btrfs_delayed_ref_is_head(ref)) {
2608                                 struct btrfs_delayed_ref_head *head;
2609
2610                                 head = btrfs_delayed_node_to_head(ref);
2611                                 atomic_inc(&ref->refs);
2612
2613                                 spin_unlock(&delayed_refs->lock);
2614                                 /*
2615                                  * Mutex was contended, block until it's
2616                                  * released and try again
2617                                  */
2618                                 mutex_lock(&head->mutex);
2619                                 mutex_unlock(&head->mutex);
2620
2621                                 btrfs_put_delayed_ref(ref);
2622                                 cond_resched();
2623                                 goto again;
2624                         }
2625                         node = rb_next(node);
2626                 }
2627                 spin_unlock(&delayed_refs->lock);
2628                 schedule_timeout(1);
2629                 goto again;
2630         }
2631 out:
2632         atomic_dec(&delayed_refs->procs_running_refs);
2633         smp_mb();
2634         if (waitqueue_active(&delayed_refs->wait))
2635                 wake_up(&delayed_refs->wait);
2636
2637         spin_unlock(&delayed_refs->lock);
2638         assert_qgroups_uptodate(trans);
2639         return 0;
2640 }
2641
2642 int btrfs_set_disk_extent_flags(struct btrfs_trans_handle *trans,
2643                                 struct btrfs_root *root,
2644                                 u64 bytenr, u64 num_bytes, u64 flags,
2645                                 int is_data)
2646 {
2647         struct btrfs_delayed_extent_op *extent_op;
2648         int ret;
2649
2650         extent_op = btrfs_alloc_delayed_extent_op();
2651         if (!extent_op)
2652                 return -ENOMEM;
2653
2654         extent_op->flags_to_set = flags;
2655         extent_op->update_flags = 1;
2656         extent_op->update_key = 0;
2657         extent_op->is_data = is_data ? 1 : 0;
2658
2659         ret = btrfs_add_delayed_extent_op(root->fs_info, trans, bytenr,
2660                                           num_bytes, extent_op);
2661         if (ret)
2662                 btrfs_free_delayed_extent_op(extent_op);
2663         return ret;
2664 }
2665
2666 static noinline int check_delayed_ref(struct btrfs_trans_handle *trans,
2667                                       struct btrfs_root *root,
2668                                       struct btrfs_path *path,
2669                                       u64 objectid, u64 offset, u64 bytenr)
2670 {
2671         struct btrfs_delayed_ref_head *head;
2672         struct btrfs_delayed_ref_node *ref;
2673         struct btrfs_delayed_data_ref *data_ref;
2674         struct btrfs_delayed_ref_root *delayed_refs;
2675         struct rb_node *node;
2676         int ret = 0;
2677
2678         ret = -ENOENT;
2679         delayed_refs = &trans->transaction->delayed_refs;
2680         spin_lock(&delayed_refs->lock);
2681         head = btrfs_find_delayed_ref_head(trans, bytenr);
2682         if (!head)
2683                 goto out;
2684
2685         if (!mutex_trylock(&head->mutex)) {
2686                 atomic_inc(&head->node.refs);
2687                 spin_unlock(&delayed_refs->lock);
2688
2689                 btrfs_release_path(path);
2690
2691                 /*
2692                  * Mutex was contended, block until it's released and let
2693                  * caller try again
2694                  */
2695                 mutex_lock(&head->mutex);
2696                 mutex_unlock(&head->mutex);
2697                 btrfs_put_delayed_ref(&head->node);
2698                 return -EAGAIN;
2699         }
2700
2701         node = rb_prev(&head->node.rb_node);
2702         if (!node)
2703                 goto out_unlock;
2704
2705         ref = rb_entry(node, struct btrfs_delayed_ref_node, rb_node);
2706
2707         if (ref->bytenr != bytenr)
2708                 goto out_unlock;
2709
2710         ret = 1;
2711         if (ref->type != BTRFS_EXTENT_DATA_REF_KEY)
2712                 goto out_unlock;
2713
2714         data_ref = btrfs_delayed_node_to_data_ref(ref);
2715
2716         node = rb_prev(node);
2717         if (node) {
2718                 int seq = ref->seq;
2719
2720                 ref = rb_entry(node, struct btrfs_delayed_ref_node, rb_node);
2721                 if (ref->bytenr == bytenr && ref->seq == seq)
2722                         goto out_unlock;
2723         }
2724
2725         if (data_ref->root != root->root_key.objectid ||
2726             data_ref->objectid != objectid || data_ref->offset != offset)
2727                 goto out_unlock;
2728
2729         ret = 0;
2730 out_unlock:
2731         mutex_unlock(&head->mutex);
2732 out:
2733         spin_unlock(&delayed_refs->lock);
2734         return ret;
2735 }
2736
2737 static noinline int check_committed_ref(struct btrfs_trans_handle *trans,
2738                                         struct btrfs_root *root,
2739                                         struct btrfs_path *path,
2740                                         u64 objectid, u64 offset, u64 bytenr)
2741 {
2742         struct btrfs_root *extent_root = root->fs_info->extent_root;
2743         struct extent_buffer *leaf;
2744         struct btrfs_extent_data_ref *ref;
2745         struct btrfs_extent_inline_ref *iref;
2746         struct btrfs_extent_item *ei;
2747         struct btrfs_key key;
2748         u32 item_size;
2749         int ret;
2750
2751         key.objectid = bytenr;
2752         key.offset = (u64)-1;
2753         key.type = BTRFS_EXTENT_ITEM_KEY;
2754
2755         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
2756         if (ret < 0)
2757                 goto out;
2758         BUG_ON(ret == 0); /* Corruption */
2759
2760         ret = -ENOENT;
2761         if (path->slots[0] == 0)
2762                 goto out;
2763
2764         path->slots[0]--;
2765         leaf = path->nodes[0];
2766         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
2767
2768         if (key.objectid != bytenr || key.type != BTRFS_EXTENT_ITEM_KEY)
2769                 goto out;
2770
2771         ret = 1;
2772         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2773 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
2774         if (item_size < sizeof(*ei)) {
2775                 WARN_ON(item_size != sizeof(struct btrfs_extent_item_v0));
2776                 goto out;
2777         }
2778 #endif
2779         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2780
2781         if (item_size != sizeof(*ei) +
2782             btrfs_extent_inline_ref_size(BTRFS_EXTENT_DATA_REF_KEY))
2783                 goto out;
2784
2785         if (btrfs_extent_generation(leaf, ei) <=
2786             btrfs_root_last_snapshot(&root->root_item))
2787                 goto out;
2788
2789         iref = (struct btrfs_extent_inline_ref *)(ei + 1);
2790         if (btrfs_extent_inline_ref_type(leaf, iref) !=
2791             BTRFS_EXTENT_DATA_REF_KEY)
2792                 goto out;
2793
2794         ref = (struct btrfs_extent_data_ref *)(&iref->offset);
2795         if (btrfs_extent_refs(leaf, ei) !=
2796             btrfs_extent_data_ref_count(leaf, ref) ||
2797             btrfs_extent_data_ref_root(leaf, ref) !=
2798             root->root_key.objectid ||
2799             btrfs_extent_data_ref_objectid(leaf, ref) != objectid ||
2800             btrfs_extent_data_ref_offset(leaf, ref) != offset)
2801                 goto out;
2802
2803         ret = 0;
2804 out:
2805         return ret;
2806 }
2807
2808 int btrfs_cross_ref_exist(struct btrfs_trans_handle *trans,
2809                           struct btrfs_root *root,
2810                           u64 objectid, u64 offset, u64 bytenr)
2811 {
2812         struct btrfs_path *path;
2813         int ret;
2814         int ret2;
2815
2816         path = btrfs_alloc_path();
2817         if (!path)
2818                 return -ENOENT;
2819
2820         do {
2821                 ret = check_committed_ref(trans, root, path, objectid,
2822                                           offset, bytenr);
2823                 if (ret && ret != -ENOENT)
2824                         goto out;
2825
2826                 ret2 = check_delayed_ref(trans, root, path, objectid,
2827                                          offset, bytenr);
2828         } while (ret2 == -EAGAIN);
2829
2830         if (ret2 && ret2 != -ENOENT) {
2831                 ret = ret2;
2832                 goto out;
2833         }
2834
2835         if (ret != -ENOENT || ret2 != -ENOENT)
2836                 ret = 0;
2837 out:
2838         btrfs_free_path(path);
2839         if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID)
2840                 WARN_ON(ret > 0);
2841         return ret;
2842 }
2843
2844 static int __btrfs_mod_ref(struct btrfs_trans_handle *trans,
2845                            struct btrfs_root *root,
2846                            struct extent_buffer *buf,
2847                            int full_backref, int inc, int for_cow)
2848 {
2849         u64 bytenr;
2850         u64 num_bytes;
2851         u64 parent;
2852         u64 ref_root;
2853         u32 nritems;
2854         struct btrfs_key key;
2855         struct btrfs_file_extent_item *fi;
2856         int i;
2857         int level;
2858         int ret = 0;
2859         int (*process_func)(struct btrfs_trans_handle *, struct btrfs_root *,
2860                             u64, u64, u64, u64, u64, u64, int);
2861
2862         ref_root = btrfs_header_owner(buf);
2863         nritems = btrfs_header_nritems(buf);
2864         level = btrfs_header_level(buf);
2865
2866         if (!root->ref_cows && level == 0)
2867                 return 0;
2868
2869         if (inc)
2870                 process_func = btrfs_inc_extent_ref;
2871         else
2872                 process_func = btrfs_free_extent;
2873
2874         if (full_backref)
2875                 parent = buf->start;
2876         else
2877                 parent = 0;
2878
2879         for (i = 0; i < nritems; i++) {
2880                 if (level == 0) {
2881                         btrfs_item_key_to_cpu(buf, &key, i);
2882                         if (btrfs_key_type(&key) != BTRFS_EXTENT_DATA_KEY)
2883                                 continue;
2884                         fi = btrfs_item_ptr(buf, i,
2885                                             struct btrfs_file_extent_item);
2886                         if (btrfs_file_extent_type(buf, fi) ==
2887                             BTRFS_FILE_EXTENT_INLINE)
2888                                 continue;
2889                         bytenr = btrfs_file_extent_disk_bytenr(buf, fi);
2890                         if (bytenr == 0)
2891                                 continue;
2892
2893                         num_bytes = btrfs_file_extent_disk_num_bytes(buf, fi);
2894                         key.offset -= btrfs_file_extent_offset(buf, fi);
2895                         ret = process_func(trans, root, bytenr, num_bytes,
2896                                            parent, ref_root, key.objectid,
2897                                            key.offset, for_cow);
2898                         if (ret)
2899                                 goto fail;
2900                 } else {
2901                         bytenr = btrfs_node_blockptr(buf, i);
2902                         num_bytes = btrfs_level_size(root, level - 1);
2903                         ret = process_func(trans, root, bytenr, num_bytes,
2904                                            parent, ref_root, level - 1, 0,
2905                                            for_cow);
2906                         if (ret)
2907                                 goto fail;
2908                 }
2909         }
2910         return 0;
2911 fail:
2912         return ret;
2913 }
2914
2915 int btrfs_inc_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
2916                   struct extent_buffer *buf, int full_backref, int for_cow)
2917 {
2918         return __btrfs_mod_ref(trans, root, buf, full_backref, 1, for_cow);
2919 }
2920
2921 int btrfs_dec_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
2922                   struct extent_buffer *buf, int full_backref, int for_cow)
2923 {
2924         return __btrfs_mod_ref(trans, root, buf, full_backref, 0, for_cow);
2925 }
2926
2927 static int write_one_cache_group(struct btrfs_trans_handle *trans,
2928                                  struct btrfs_root *root,
2929                                  struct btrfs_path *path,
2930                                  struct btrfs_block_group_cache *cache)
2931 {
2932         int ret;
2933         struct btrfs_root *extent_root = root->fs_info->extent_root;
2934         unsigned long bi;
2935         struct extent_buffer *leaf;
2936
2937         ret = btrfs_search_slot(trans, extent_root, &cache->key, path, 0, 1);
2938         if (ret < 0)
2939                 goto fail;
2940         BUG_ON(ret); /* Corruption */
2941
2942         leaf = path->nodes[0];
2943         bi = btrfs_item_ptr_offset(leaf, path->slots[0]);
2944         write_extent_buffer(leaf, &cache->item, bi, sizeof(cache->item));
2945         btrfs_mark_buffer_dirty(leaf);
2946         btrfs_release_path(path);
2947 fail:
2948         if (ret) {
2949                 btrfs_abort_transaction(trans, root, ret);
2950                 return ret;
2951         }
2952         return 0;
2953
2954 }
2955
2956 static struct btrfs_block_group_cache *
2957 next_block_group(struct btrfs_root *root,
2958                  struct btrfs_block_group_cache *cache)
2959 {
2960         struct rb_node *node;
2961         spin_lock(&root->fs_info->block_group_cache_lock);
2962         node = rb_next(&cache->cache_node);
2963         btrfs_put_block_group(cache);
2964         if (node) {
2965                 cache = rb_entry(node, struct btrfs_block_group_cache,
2966                                  cache_node);
2967                 btrfs_get_block_group(cache);
2968         } else
2969                 cache = NULL;
2970         spin_unlock(&root->fs_info->block_group_cache_lock);
2971         return cache;
2972 }
2973
2974 static int cache_save_setup(struct btrfs_block_group_cache *block_group,
2975                             struct btrfs_trans_handle *trans,
2976                             struct btrfs_path *path)
2977 {
2978         struct btrfs_root *root = block_group->fs_info->tree_root;
2979         struct inode *inode = NULL;
2980         u64 alloc_hint = 0;
2981         int dcs = BTRFS_DC_ERROR;
2982         int num_pages = 0;
2983         int retries = 0;
2984         int ret = 0;
2985
2986         /*
2987          * If this block group is smaller than 100 megs don't bother caching the
2988          * block group.
2989          */
2990         if (block_group->key.offset < (100 * 1024 * 1024)) {
2991                 spin_lock(&block_group->lock);
2992                 block_group->disk_cache_state = BTRFS_DC_WRITTEN;
2993                 spin_unlock(&block_group->lock);
2994                 return 0;
2995         }
2996
2997 again:
2998         inode = lookup_free_space_inode(root, block_group, path);
2999         if (IS_ERR(inode) && PTR_ERR(inode) != -ENOENT) {
3000                 ret = PTR_ERR(inode);
3001                 btrfs_release_path(path);
3002                 goto out;
3003         }
3004
3005         if (IS_ERR(inode)) {
3006                 BUG_ON(retries);
3007                 retries++;
3008
3009                 if (block_group->ro)
3010                         goto out_free;
3011
3012                 ret = create_free_space_inode(root, trans, block_group, path);
3013                 if (ret)
3014                         goto out_free;
3015                 goto again;
3016         }
3017
3018         /* We've already setup this transaction, go ahead and exit */
3019         if (block_group->cache_generation == trans->transid &&
3020             i_size_read(inode)) {
3021                 dcs = BTRFS_DC_SETUP;
3022                 goto out_put;
3023         }
3024
3025         /*
3026          * We want to set the generation to 0, that way if anything goes wrong
3027          * from here on out we know not to trust this cache when we load up next
3028          * time.
3029          */
3030         BTRFS_I(inode)->generation = 0;
3031         ret = btrfs_update_inode(trans, root, inode);
3032         WARN_ON(ret);
3033
3034         if (i_size_read(inode) > 0) {
3035                 ret = btrfs_truncate_free_space_cache(root, trans, path,
3036                                                       inode);
3037                 if (ret)
3038                         goto out_put;
3039         }
3040
3041         spin_lock(&block_group->lock);
3042         if (block_group->cached != BTRFS_CACHE_FINISHED ||
3043             !btrfs_test_opt(root, SPACE_CACHE)) {
3044                 /*
3045                  * don't bother trying to write stuff out _if_
3046                  * a) we're not cached,
3047                  * b) we're with nospace_cache mount option.
3048                  */
3049                 dcs = BTRFS_DC_WRITTEN;
3050                 spin_unlock(&block_group->lock);
3051                 goto out_put;
3052         }
3053         spin_unlock(&block_group->lock);
3054
3055         /*
3056          * Try to preallocate enough space based on how big the block group is.
3057          * Keep in mind this has to include any pinned space which could end up
3058          * taking up quite a bit since it's not folded into the other space
3059          * cache.
3060          */
3061         num_pages = (int)div64_u64(block_group->key.offset, 256 * 1024 * 1024);
3062         if (!num_pages)
3063                 num_pages = 1;
3064
3065         num_pages *= 16;
3066         num_pages *= PAGE_CACHE_SIZE;
3067
3068         ret = btrfs_check_data_free_space(inode, num_pages);
3069         if (ret)
3070                 goto out_put;
3071
3072         ret = btrfs_prealloc_file_range_trans(inode, trans, 0, 0, num_pages,
3073                                               num_pages, num_pages,
3074                                               &alloc_hint);
3075         if (!ret)
3076                 dcs = BTRFS_DC_SETUP;
3077         btrfs_free_reserved_data_space(inode, num_pages);
3078
3079 out_put:
3080         iput(inode);
3081 out_free:
3082         btrfs_release_path(path);
3083 out:
3084         spin_lock(&block_group->lock);
3085         if (!ret && dcs == BTRFS_DC_SETUP)
3086                 block_group->cache_generation = trans->transid;
3087         block_group->disk_cache_state = dcs;
3088         spin_unlock(&block_group->lock);
3089
3090         return ret;
3091 }
3092
3093 int btrfs_write_dirty_block_groups(struct btrfs_trans_handle *trans,
3094                                    struct btrfs_root *root)
3095 {
3096         struct btrfs_block_group_cache *cache;
3097         int err = 0;
3098         struct btrfs_path *path;
3099         u64 last = 0;
3100
3101         path = btrfs_alloc_path();
3102         if (!path)
3103                 return -ENOMEM;
3104
3105 again:
3106         while (1) {
3107                 cache = btrfs_lookup_first_block_group(root->fs_info, last);
3108                 while (cache) {
3109                         if (cache->disk_cache_state == BTRFS_DC_CLEAR)
3110                                 break;
3111                         cache = next_block_group(root, cache);
3112                 }
3113                 if (!cache) {
3114                         if (last == 0)
3115                                 break;
3116                         last = 0;
3117                         continue;
3118                 }
3119                 err = cache_save_setup(cache, trans, path);
3120                 last = cache->key.objectid + cache->key.offset;
3121                 btrfs_put_block_group(cache);
3122         }
3123
3124         while (1) {
3125                 if (last == 0) {
3126                         err = btrfs_run_delayed_refs(trans, root,
3127                                                      (unsigned long)-1);
3128                         if (err) /* File system offline */
3129                                 goto out;
3130                 }
3131
3132                 cache = btrfs_lookup_first_block_group(root->fs_info, last);
3133                 while (cache) {
3134                         if (cache->disk_cache_state == BTRFS_DC_CLEAR) {
3135                                 btrfs_put_block_group(cache);
3136                                 goto again;
3137                         }
3138
3139                         if (cache->dirty)
3140                                 break;
3141                         cache = next_block_group(root, cache);
3142                 }
3143                 if (!cache) {
3144                         if (last == 0)
3145                                 break;
3146                         last = 0;
3147                         continue;
3148                 }
3149
3150                 if (cache->disk_cache_state == BTRFS_DC_SETUP)
3151                         cache->disk_cache_state = BTRFS_DC_NEED_WRITE;
3152                 cache->dirty = 0;
3153                 last = cache->key.objectid + cache->key.offset;
3154
3155                 err = write_one_cache_group(trans, root, path, cache);
3156                 if (err) /* File system offline */
3157                         goto out;
3158
3159                 btrfs_put_block_group(cache);
3160         }
3161
3162         while (1) {
3163                 /*
3164                  * I don't think this is needed since we're just marking our
3165                  * preallocated extent as written, but just in case it can't
3166                  * hurt.
3167                  */
3168                 if (last == 0) {
3169                         err = btrfs_run_delayed_refs(trans, root,
3170                                                      (unsigned long)-1);
3171                         if (err) /* File system offline */
3172                                 goto out;
3173                 }
3174
3175                 cache = btrfs_lookup_first_block_group(root->fs_info, last);
3176                 while (cache) {
3177                         /*
3178                          * Really this shouldn't happen, but it could if we
3179                          * couldn't write the entire preallocated extent and
3180                          * splitting the extent resulted in a new block.
3181                          */
3182                         if (cache->dirty) {
3183                                 btrfs_put_block_group(cache);
3184                                 goto again;
3185                         }
3186                         if (cache->disk_cache_state == BTRFS_DC_NEED_WRITE)
3187                                 break;
3188                         cache = next_block_group(root, cache);
3189                 }
3190                 if (!cache) {
3191                         if (last == 0)
3192                                 break;
3193                         last = 0;
3194                         continue;
3195                 }
3196
3197                 err = btrfs_write_out_cache(root, trans, cache, path);
3198
3199                 /*
3200                  * If we didn't have an error then the cache state is still
3201                  * NEED_WRITE, so we can set it to WRITTEN.
3202                  */
3203                 if (!err && cache->disk_cache_state == BTRFS_DC_NEED_WRITE)
3204                         cache->disk_cache_state = BTRFS_DC_WRITTEN;
3205                 last = cache->key.objectid + cache->key.offset;
3206                 btrfs_put_block_group(cache);
3207         }
3208 out:
3209
3210         btrfs_free_path(path);
3211         return err;
3212 }
3213
3214 int btrfs_extent_readonly(struct btrfs_root *root, u64 bytenr)
3215 {
3216         struct btrfs_block_group_cache *block_group;
3217         int readonly = 0;
3218
3219         block_group = btrfs_lookup_block_group(root->fs_info, bytenr);
3220         if (!block_group || block_group->ro)
3221                 readonly = 1;
3222         if (block_group)
3223                 btrfs_put_block_group(block_group);
3224         return readonly;
3225 }
3226
3227 static int update_space_info(struct btrfs_fs_info *info, u64 flags,
3228                              u64 total_bytes, u64 bytes_used,
3229                              struct btrfs_space_info **space_info)
3230 {
3231         struct btrfs_space_info *found;
3232         int i;
3233         int factor;
3234
3235         if (flags & (BTRFS_BLOCK_GROUP_DUP | BTRFS_BLOCK_GROUP_RAID1 |
3236                      BTRFS_BLOCK_GROUP_RAID10))
3237                 factor = 2;
3238         else
3239                 factor = 1;
3240
3241         found = __find_space_info(info, flags);
3242         if (found) {
3243                 spin_lock(&found->lock);
3244                 found->total_bytes += total_bytes;
3245                 found->disk_total += total_bytes * factor;
3246                 found->bytes_used += bytes_used;
3247                 found->disk_used += bytes_used * factor;
3248                 found->full = 0;
3249                 spin_unlock(&found->lock);
3250                 *space_info = found;
3251                 return 0;
3252         }
3253         found = kzalloc(sizeof(*found), GFP_NOFS);
3254         if (!found)
3255                 return -ENOMEM;
3256
3257         for (i = 0; i < BTRFS_NR_RAID_TYPES; i++)
3258                 INIT_LIST_HEAD(&found->block_groups[i]);
3259         init_rwsem(&found->groups_sem);
3260         spin_lock_init(&found->lock);
3261         found->flags = flags & BTRFS_BLOCK_GROUP_TYPE_MASK;
3262         found->total_bytes = total_bytes;
3263         found->disk_total = total_bytes * factor;
3264         found->bytes_used = bytes_used;
3265         found->disk_used = bytes_used * factor;
3266         found->bytes_pinned = 0;
3267         found->bytes_reserved = 0;
3268         found->bytes_readonly = 0;
3269         found->bytes_may_use = 0;
3270         found->full = 0;
3271         found->force_alloc = CHUNK_ALLOC_NO_FORCE;
3272         found->chunk_alloc = 0;
3273         found->flush = 0;
3274         init_waitqueue_head(&found->wait);
3275         *space_info = found;
3276         list_add_rcu(&found->list, &info->space_info);
3277         if (flags & BTRFS_BLOCK_GROUP_DATA)
3278                 info->data_sinfo = found;
3279         return 0;
3280 }
3281
3282 static void set_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
3283 {
3284         u64 extra_flags = chunk_to_extended(flags) &
3285                                 BTRFS_EXTENDED_PROFILE_MASK;
3286
3287         write_seqlock(&fs_info->profiles_lock);
3288         if (flags & BTRFS_BLOCK_GROUP_DATA)
3289                 fs_info->avail_data_alloc_bits |= extra_flags;
3290         if (flags & BTRFS_BLOCK_GROUP_METADATA)
3291                 fs_info->avail_metadata_alloc_bits |= extra_flags;
3292         if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
3293                 fs_info->avail_system_alloc_bits |= extra_flags;
3294         write_sequnlock(&fs_info->profiles_lock);
3295 }
3296
3297 /*
3298  * returns target flags in extended format or 0 if restripe for this
3299  * chunk_type is not in progress
3300  *
3301  * should be called with either volume_mutex or balance_lock held
3302  */
3303 static u64 get_restripe_target(struct btrfs_fs_info *fs_info, u64 flags)
3304 {
3305         struct btrfs_balance_control *bctl = fs_info->balance_ctl;
3306         u64 target = 0;
3307
3308         if (!bctl)
3309                 return 0;
3310
3311         if (flags & BTRFS_BLOCK_GROUP_DATA &&
3312             bctl->data.flags & BTRFS_BALANCE_ARGS_CONVERT) {
3313                 target = BTRFS_BLOCK_GROUP_DATA | bctl->data.target;
3314         } else if (flags & BTRFS_BLOCK_GROUP_SYSTEM &&
3315                    bctl->sys.flags & BTRFS_BALANCE_ARGS_CONVERT) {
3316                 target = BTRFS_BLOCK_GROUP_SYSTEM | bctl->sys.target;
3317         } else if (flags & BTRFS_BLOCK_GROUP_METADATA &&
3318                    bctl->meta.flags & BTRFS_BALANCE_ARGS_CONVERT) {
3319                 target = BTRFS_BLOCK_GROUP_METADATA | bctl->meta.target;
3320         }
3321
3322         return target;
3323 }
3324
3325 /*
3326  * @flags: available profiles in extended format (see ctree.h)
3327  *
3328  * Returns reduced profile in chunk format.  If profile changing is in
3329  * progress (either running or paused) picks the target profile (if it's
3330  * already available), otherwise falls back to plain reducing.
3331  */
3332 u64 btrfs_reduce_alloc_profile(struct btrfs_root *root, u64 flags)
3333 {
3334         /*
3335          * we add in the count of missing devices because we want
3336          * to make sure that any RAID levels on a degraded FS
3337          * continue to be honored.
3338          */
3339         u64 num_devices = root->fs_info->fs_devices->rw_devices +
3340                 root->fs_info->fs_devices->missing_devices;
3341         u64 target;
3342         u64 tmp;
3343
3344         /*
3345          * see if restripe for this chunk_type is in progress, if so
3346          * try to reduce to the target profile
3347          */
3348         spin_lock(&root->fs_info->balance_lock);
3349         target = get_restripe_target(root->fs_info, flags);
3350         if (target) {
3351                 /* pick target profile only if it's already available */
3352                 if ((flags & target) & BTRFS_EXTENDED_PROFILE_MASK) {
3353                         spin_unlock(&root->fs_info->balance_lock);
3354                         return extended_to_chunk(target);
3355                 }
3356         }
3357         spin_unlock(&root->fs_info->balance_lock);
3358
3359         /* First, mask out the RAID levels which aren't possible */
3360         if (num_devices == 1)
3361                 flags &= ~(BTRFS_BLOCK_GROUP_RAID1 | BTRFS_BLOCK_GROUP_RAID0 |
3362                            BTRFS_BLOCK_GROUP_RAID5);
3363         if (num_devices < 3)
3364                 flags &= ~BTRFS_BLOCK_GROUP_RAID6;
3365         if (num_devices < 4)
3366                 flags &= ~BTRFS_BLOCK_GROUP_RAID10;
3367
3368         tmp = flags & (BTRFS_BLOCK_GROUP_DUP | BTRFS_BLOCK_GROUP_RAID0 |
3369                        BTRFS_BLOCK_GROUP_RAID1 | BTRFS_BLOCK_GROUP_RAID5 |
3370                        BTRFS_BLOCK_GROUP_RAID6 | BTRFS_BLOCK_GROUP_RAID10);
3371         flags &= ~tmp;
3372
3373         if (tmp & BTRFS_BLOCK_GROUP_RAID6)
3374                 tmp = BTRFS_BLOCK_GROUP_RAID6;
3375         else if (tmp & BTRFS_BLOCK_GROUP_RAID5)
3376                 tmp = BTRFS_BLOCK_GROUP_RAID5;
3377         else if (tmp & BTRFS_BLOCK_GROUP_RAID10)
3378                 tmp = BTRFS_BLOCK_GROUP_RAID10;
3379         else if (tmp & BTRFS_BLOCK_GROUP_RAID1)
3380                 tmp = BTRFS_BLOCK_GROUP_RAID1;
3381         else if (tmp & BTRFS_BLOCK_GROUP_RAID0)
3382                 tmp = BTRFS_BLOCK_GROUP_RAID0;
3383
3384         return extended_to_chunk(flags | tmp);
3385 }
3386
3387 static u64 get_alloc_profile(struct btrfs_root *root, u64 flags)
3388 {
3389         unsigned seq;
3390
3391         do {
3392                 seq = read_seqbegin(&root->fs_info->profiles_lock);
3393
3394                 if (flags & BTRFS_BLOCK_GROUP_DATA)
3395                         flags |= root->fs_info->avail_data_alloc_bits;
3396                 else if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
3397                         flags |= root->fs_info->avail_system_alloc_bits;
3398                 else if (flags & BTRFS_BLOCK_GROUP_METADATA)
3399                         flags |= root->fs_info->avail_metadata_alloc_bits;
3400         } while (read_seqretry(&root->fs_info->profiles_lock, seq));
3401
3402         return btrfs_reduce_alloc_profile(root, flags);
3403 }
3404
3405 u64 btrfs_get_alloc_profile(struct btrfs_root *root, int data)
3406 {
3407         u64 flags;
3408         u64 ret;
3409
3410         if (data)
3411                 flags = BTRFS_BLOCK_GROUP_DATA;
3412         else if (root == root->fs_info->chunk_root)
3413                 flags = BTRFS_BLOCK_GROUP_SYSTEM;
3414         else
3415                 flags = BTRFS_BLOCK_GROUP_METADATA;
3416
3417         ret = get_alloc_profile(root, flags);
3418         return ret;
3419 }
3420
3421 /*
3422  * This will check the space that the inode allocates from to make sure we have
3423  * enough space for bytes.
3424  */
3425 int btrfs_check_data_free_space(struct inode *inode, u64 bytes)
3426 {
3427         struct btrfs_space_info *data_sinfo;
3428         struct btrfs_root *root = BTRFS_I(inode)->root;
3429         struct btrfs_fs_info *fs_info = root->fs_info;
3430         u64 used;
3431         int ret = 0, committed = 0, alloc_chunk = 1;
3432
3433         /* make sure bytes are sectorsize aligned */
3434         bytes = ALIGN(bytes, root->sectorsize);
3435
3436         if (root == root->fs_info->tree_root ||
3437             BTRFS_I(inode)->location.objectid == BTRFS_FREE_INO_OBJECTID) {
3438                 alloc_chunk = 0;
3439                 committed = 1;
3440         }
3441
3442         data_sinfo = fs_info->data_sinfo;
3443         if (!data_sinfo)
3444                 goto alloc;
3445
3446 again:
3447         /* make sure we have enough space to handle the data first */
3448         spin_lock(&data_sinfo->lock);
3449         used = data_sinfo->bytes_used + data_sinfo->bytes_reserved +
3450                 data_sinfo->bytes_pinned + data_sinfo->bytes_readonly +
3451                 data_sinfo->bytes_may_use;
3452
3453         if (used + bytes > data_sinfo->total_bytes) {
3454                 struct btrfs_trans_handle *trans;
3455
3456                 /*
3457                  * if we don't have enough free bytes in this space then we need
3458                  * to alloc a new chunk.
3459                  */
3460                 if (!data_sinfo->full && alloc_chunk) {
3461                         u64 alloc_target;
3462
3463                         data_sinfo->force_alloc = CHUNK_ALLOC_FORCE;
3464                         spin_unlock(&data_sinfo->lock);
3465 alloc:
3466                         alloc_target = btrfs_get_alloc_profile(root, 1);
3467                         trans = btrfs_join_transaction(root);
3468                         if (IS_ERR(trans))
3469                                 return PTR_ERR(trans);
3470
3471                         ret = do_chunk_alloc(trans, root->fs_info->extent_root,
3472                                              alloc_target,
3473                                              CHUNK_ALLOC_NO_FORCE);
3474                         btrfs_end_transaction(trans, root);
3475                         if (ret < 0) {
3476                                 if (ret != -ENOSPC)
3477                                         return ret;
3478                                 else
3479                                         goto commit_trans;
3480                         }
3481
3482                         if (!data_sinfo)
3483                                 data_sinfo = fs_info->data_sinfo;
3484
3485                         goto again;
3486                 }
3487
3488                 /*
3489                  * If we have less pinned bytes than we want to allocate then
3490                  * don't bother committing the transaction, it won't help us.
3491                  */
3492                 if (data_sinfo->bytes_pinned < bytes)
3493                         committed = 1;
3494                 spin_unlock(&data_sinfo->lock);
3495
3496                 /* commit the current transaction and try again */
3497 commit_trans:
3498                 if (!committed &&
3499                     !atomic_read(&root->fs_info->open_ioctl_trans)) {
3500                         committed = 1;
3501                         trans = btrfs_join_transaction(root);
3502                         if (IS_ERR(trans))
3503                                 return PTR_ERR(trans);
3504                         ret = btrfs_commit_transaction(trans, root);
3505                         if (ret)
3506                                 return ret;
3507                         goto again;
3508                 }
3509
3510                 return -ENOSPC;
3511         }
3512         data_sinfo->bytes_may_use += bytes;
3513         trace_btrfs_space_reservation(root->fs_info, "space_info",
3514                                       data_sinfo->flags, bytes, 1);
3515         spin_unlock(&data_sinfo->lock);
3516
3517         return 0;
3518 }
3519
3520 /*
3521  * Called if we need to clear a data reservation for this inode.
3522  */
3523 void btrfs_free_reserved_data_space(struct inode *inode, u64 bytes)
3524 {
3525         struct btrfs_root *root = BTRFS_I(inode)->root;
3526         struct btrfs_space_info *data_sinfo;
3527
3528         /* make sure bytes are sectorsize aligned */
3529         bytes = ALIGN(bytes, root->sectorsize);
3530
3531         data_sinfo = root->fs_info->data_sinfo;
3532         spin_lock(&data_sinfo->lock);
3533         data_sinfo->bytes_may_use -= bytes;
3534         trace_btrfs_space_reservation(root->fs_info, "space_info",
3535                                       data_sinfo->flags, bytes, 0);
3536         spin_unlock(&data_sinfo->lock);
3537 }
3538
3539 static void force_metadata_allocation(struct btrfs_fs_info *info)
3540 {
3541         struct list_head *head = &info->space_info;
3542         struct btrfs_space_info *found;
3543
3544         rcu_read_lock();
3545         list_for_each_entry_rcu(found, head, list) {
3546                 if (found->flags & BTRFS_BLOCK_GROUP_METADATA)
3547                         found->force_alloc = CHUNK_ALLOC_FORCE;
3548         }
3549         rcu_read_unlock();
3550 }
3551
3552 static int should_alloc_chunk(struct btrfs_root *root,
3553                               struct btrfs_space_info *sinfo, int force)
3554 {
3555         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
3556         u64 num_bytes = sinfo->total_bytes - sinfo->bytes_readonly;
3557         u64 num_allocated = sinfo->bytes_used + sinfo->bytes_reserved;
3558         u64 thresh;
3559
3560         if (force == CHUNK_ALLOC_FORCE)
3561                 return 1;
3562
3563         /*
3564          * We need to take into account the global rsv because for all intents
3565          * and purposes it's used space.  Don't worry about locking the
3566          * global_rsv, it doesn't change except when the transaction commits.
3567          */
3568         if (sinfo->flags & BTRFS_BLOCK_GROUP_METADATA)
3569                 num_allocated += global_rsv->size;
3570
3571         /*
3572          * in limited mode, we want to have some free space up to
3573          * about 1% of the FS size.
3574          */
3575         if (force == CHUNK_ALLOC_LIMITED) {
3576                 thresh = btrfs_super_total_bytes(root->fs_info->super_copy);
3577                 thresh = max_t(u64, 64 * 1024 * 1024,
3578                                div_factor_fine(thresh, 1));
3579
3580                 if (num_bytes - num_allocated < thresh)
3581                         return 1;
3582         }
3583
3584         if (num_allocated + 2 * 1024 * 1024 < div_factor(num_bytes, 8))
3585                 return 0;
3586         return 1;
3587 }
3588
3589 static u64 get_system_chunk_thresh(struct btrfs_root *root, u64 type)
3590 {
3591         u64 num_dev;
3592
3593         if (type & (BTRFS_BLOCK_GROUP_RAID10 |
3594                     BTRFS_BLOCK_GROUP_RAID0 |
3595                     BTRFS_BLOCK_GROUP_RAID5 |
3596                     BTRFS_BLOCK_GROUP_RAID6))
3597                 num_dev = root->fs_info->fs_devices->rw_devices;
3598         else if (type & BTRFS_BLOCK_GROUP_RAID1)
3599                 num_dev = 2;
3600         else
3601                 num_dev = 1;    /* DUP or single */
3602
3603         /* metadata for updaing devices and chunk tree */
3604         return btrfs_calc_trans_metadata_size(root, num_dev + 1);
3605 }
3606
3607 static void check_system_chunk(struct btrfs_trans_handle *trans,
3608                                struct btrfs_root *root, u64 type)
3609 {
3610         struct btrfs_space_info *info;
3611         u64 left;
3612         u64 thresh;
3613
3614         info = __find_space_info(root->fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
3615         spin_lock(&info->lock);
3616         left = info->total_bytes - info->bytes_used - info->bytes_pinned -
3617                 info->bytes_reserved - info->bytes_readonly;
3618         spin_unlock(&info->lock);
3619
3620         thresh = get_system_chunk_thresh(root, type);
3621         if (left < thresh && btrfs_test_opt(root, ENOSPC_DEBUG)) {
3622                 printk(KERN_INFO "left=%llu, need=%llu, flags=%llu\n",
3623                        left, thresh, type);
3624                 dump_space_info(info, 0, 0);
3625         }
3626
3627         if (left < thresh) {
3628                 u64 flags;
3629
3630                 flags = btrfs_get_alloc_profile(root->fs_info->chunk_root, 0);
3631                 btrfs_alloc_chunk(trans, root, flags);
3632         }
3633 }
3634
3635 static int do_chunk_alloc(struct btrfs_trans_handle *trans,
3636                           struct btrfs_root *extent_root, u64 flags, int force)
3637 {
3638         struct btrfs_space_info *space_info;
3639         struct btrfs_fs_info *fs_info = extent_root->fs_info;
3640         int wait_for_alloc = 0;
3641         int ret = 0;
3642
3643         /* Don't re-enter if we're already allocating a chunk */
3644         if (trans->allocating_chunk)
3645                 return -ENOSPC;
3646
3647         space_info = __find_space_info(extent_root->fs_info, flags);
3648         if (!space_info) {
3649                 ret = update_space_info(extent_root->fs_info, flags,
3650                                         0, 0, &space_info);
3651                 BUG_ON(ret); /* -ENOMEM */
3652         }
3653         BUG_ON(!space_info); /* Logic error */
3654
3655 again:
3656         spin_lock(&space_info->lock);
3657         if (force < space_info->force_alloc)
3658                 force = space_info->force_alloc;
3659         if (space_info->full) {
3660                 spin_unlock(&space_info->lock);
3661                 return 0;
3662         }
3663
3664         if (!should_alloc_chunk(extent_root, space_info, force)) {
3665                 spin_unlock(&space_info->lock);
3666                 return 0;
3667         } else if (space_info->chunk_alloc) {
3668                 wait_for_alloc = 1;
3669         } else {
3670                 space_info->chunk_alloc = 1;
3671         }
3672
3673         spin_unlock(&space_info->lock);
3674
3675         mutex_lock(&fs_info->chunk_mutex);
3676
3677         /*
3678          * The chunk_mutex is held throughout the entirety of a chunk
3679          * allocation, so once we've acquired the chunk_mutex we know that the
3680          * other guy is done and we need to recheck and see if we should
3681          * allocate.
3682          */
3683         if (wait_for_alloc) {
3684                 mutex_unlock(&fs_info->chunk_mutex);
3685                 wait_for_alloc = 0;
3686                 goto again;
3687         }
3688
3689         trans->allocating_chunk = true;
3690
3691         /*
3692          * If we have mixed data/metadata chunks we want to make sure we keep
3693          * allocating mixed chunks instead of individual chunks.
3694          */
3695         if (btrfs_mixed_space_info(space_info))
3696                 flags |= (BTRFS_BLOCK_GROUP_DATA | BTRFS_BLOCK_GROUP_METADATA);
3697
3698         /*
3699          * if we're doing a data chunk, go ahead and make sure that
3700          * we keep a reasonable number of metadata chunks allocated in the
3701          * FS as well.
3702          */
3703         if (flags & BTRFS_BLOCK_GROUP_DATA && fs_info->metadata_ratio) {
3704                 fs_info->data_chunk_allocations++;
3705                 if (!(fs_info->data_chunk_allocations %
3706                       fs_info->metadata_ratio))
3707                         force_metadata_allocation(fs_info);
3708         }
3709
3710         /*
3711          * Check if we have enough space in SYSTEM chunk because we may need
3712          * to update devices.
3713          */
3714         check_system_chunk(trans, extent_root, flags);
3715
3716         ret = btrfs_alloc_chunk(trans, extent_root, flags);
3717         trans->allocating_chunk = false;
3718
3719         spin_lock(&space_info->lock);
3720         if (ret < 0 && ret != -ENOSPC)
3721                 goto out;
3722         if (ret)
3723                 space_info->full = 1;
3724         else
3725                 ret = 1;
3726
3727         space_info->force_alloc = CHUNK_ALLOC_NO_FORCE;
3728 out:
3729         space_info->chunk_alloc = 0;
3730         spin_unlock(&space_info->lock);
3731         mutex_unlock(&fs_info->chunk_mutex);
3732         return ret;
3733 }
3734
3735 static int can_overcommit(struct btrfs_root *root,
3736                           struct btrfs_space_info *space_info, u64 bytes,
3737                           enum btrfs_reserve_flush_enum flush)
3738 {
3739         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
3740         u64 profile = btrfs_get_alloc_profile(root, 0);
3741         u64 rsv_size = 0;
3742         u64 avail;
3743         u64 used;
3744         u64 to_add;
3745
3746         used = space_info->bytes_used + space_info->bytes_reserved +
3747                 space_info->bytes_pinned + space_info->bytes_readonly;
3748
3749         spin_lock(&global_rsv->lock);
3750         rsv_size = global_rsv->size;
3751         spin_unlock(&global_rsv->lock);
3752
3753         /*
3754          * We only want to allow over committing if we have lots of actual space
3755          * free, but if we don't have enough space to handle the global reserve
3756          * space then we could end up having a real enospc problem when trying
3757          * to allocate a chunk or some other such important allocation.
3758          */
3759         rsv_size <<= 1;
3760         if (used + rsv_size >= space_info->total_bytes)
3761                 return 0;
3762
3763         used += space_info->bytes_may_use;
3764
3765         spin_lock(&root->fs_info->free_chunk_lock);
3766         avail = root->fs_info->free_chunk_space;
3767         spin_unlock(&root->fs_info->free_chunk_lock);
3768
3769         /*
3770          * If we have dup, raid1 or raid10 then only half of the free
3771          * space is actually useable.  For raid56, the space info used
3772          * doesn't include the parity drive, so we don't have to
3773          * change the math
3774          */
3775         if (profile & (BTRFS_BLOCK_GROUP_DUP |
3776                        BTRFS_BLOCK_GROUP_RAID1 |
3777                        BTRFS_BLOCK_GROUP_RAID10))
3778                 avail >>= 1;
3779
3780         to_add = space_info->total_bytes;
3781
3782         /*
3783          * If we aren't flushing all things, let us overcommit up to
3784          * 1/2th of the space. If we can flush, don't let us overcommit
3785          * too much, let it overcommit up to 1/8 of the space.
3786          */
3787         if (flush == BTRFS_RESERVE_FLUSH_ALL)
3788                 to_add >>= 3;
3789         else
3790                 to_add >>= 1;
3791
3792         /*
3793          * Limit the overcommit to the amount of free space we could possibly
3794          * allocate for chunks.
3795          */
3796         to_add = min(avail, to_add);
3797
3798         if (used + bytes < space_info->total_bytes + to_add)
3799                 return 1;
3800         return 0;
3801 }
3802
3803 static inline int writeback_inodes_sb_nr_if_idle_safe(struct super_block *sb,
3804                                                       unsigned long nr_pages,
3805                                                       enum wb_reason reason)
3806 {
3807         /* the flusher is dealing with the dirty inodes now. */
3808         if (writeback_in_progress(sb->s_bdi))
3809                 return 1;
3810
3811         if (down_read_trylock(&sb->s_umount)) {
3812                 writeback_inodes_sb_nr(sb, nr_pages, reason);
3813                 up_read(&sb->s_umount);
3814                 return 1;
3815         }
3816
3817         return 0;
3818 }
3819
3820 void btrfs_writeback_inodes_sb_nr(struct btrfs_root *root,
3821                                   unsigned long nr_pages)
3822 {
3823         struct super_block *sb = root->fs_info->sb;
3824         int started;
3825
3826         /* If we can not start writeback, just sync all the delalloc file. */
3827         started = writeback_inodes_sb_nr_if_idle_safe(sb, nr_pages,
3828                                                       WB_REASON_FS_FREE_SPACE);
3829         if (!started) {
3830                 /*
3831                  * We needn't worry the filesystem going from r/w to r/o though
3832                  * we don't acquire ->s_umount mutex, because the filesystem
3833                  * should guarantee the delalloc inodes list be empty after
3834                  * the filesystem is readonly(all dirty pages are written to
3835                  * the disk).
3836                  */
3837                 btrfs_start_delalloc_inodes(root, 0);
3838                 btrfs_wait_ordered_extents(root, 0);
3839         }
3840 }
3841
3842 /*
3843  * shrink metadata reservation for delalloc
3844  */
3845 static void shrink_delalloc(struct btrfs_root *root, u64 to_reclaim, u64 orig,
3846                             bool wait_ordered)
3847 {
3848         struct btrfs_block_rsv *block_rsv;
3849         struct btrfs_space_info *space_info;
3850         struct btrfs_trans_handle *trans;
3851         u64 delalloc_bytes;
3852         u64 max_reclaim;
3853         long time_left;
3854         unsigned long nr_pages = (2 * 1024 * 1024) >> PAGE_CACHE_SHIFT;
3855         int loops = 0;
3856         enum btrfs_reserve_flush_enum flush;
3857
3858         trans = (struct btrfs_trans_handle *)current->journal_info;
3859         block_rsv = &root->fs_info->delalloc_block_rsv;
3860         space_info = block_rsv->space_info;
3861
3862         smp_mb();
3863         delalloc_bytes = percpu_counter_sum_positive(
3864                                                 &root->fs_info->delalloc_bytes);
3865         if (delalloc_bytes == 0) {
3866                 if (trans)
3867                         return;
3868                 btrfs_wait_ordered_extents(root, 0);
3869                 return;
3870         }
3871
3872         while (delalloc_bytes && loops < 3) {
3873                 max_reclaim = min(delalloc_bytes, to_reclaim);
3874                 nr_pages = max_reclaim >> PAGE_CACHE_SHIFT;
3875                 btrfs_writeback_inodes_sb_nr(root, nr_pages);
3876                 /*
3877                  * We need to wait for the async pages to actually start before
3878                  * we do anything.
3879                  */
3880                 wait_event(root->fs_info->async_submit_wait,
3881                            !atomic_read(&root->fs_info->async_delalloc_pages));
3882
3883                 if (!trans)
3884                         flush = BTRFS_RESERVE_FLUSH_ALL;
3885                 else
3886                         flush = BTRFS_RESERVE_NO_FLUSH;
3887                 spin_lock(&space_info->lock);
3888                 if (can_overcommit(root, space_info, orig, flush)) {
3889                         spin_unlock(&space_info->lock);
3890                         break;
3891                 }
3892                 spin_unlock(&space_info->lock);
3893
3894                 loops++;
3895                 if (wait_ordered && !trans) {
3896                         btrfs_wait_ordered_extents(root, 0);
3897                 } else {
3898                         time_left = schedule_timeout_killable(1);
3899                         if (time_left)
3900                                 break;
3901                 }
3902                 smp_mb();
3903                 delalloc_bytes = percpu_counter_sum_positive(
3904                                                 &root->fs_info->delalloc_bytes);
3905         }
3906 }
3907
3908 /**
3909  * maybe_commit_transaction - possibly commit the transaction if its ok to
3910  * @root - the root we're allocating for
3911  * @bytes - the number of bytes we want to reserve
3912  * @force - force the commit
3913  *
3914  * This will check to make sure that committing the transaction will actually
3915  * get us somewhere and then commit the transaction if it does.  Otherwise it
3916  * will return -ENOSPC.
3917  */
3918 static int may_commit_transaction(struct btrfs_root *root,
3919                                   struct btrfs_space_info *space_info,
3920                                   u64 bytes, int force)
3921 {
3922         struct btrfs_block_rsv *delayed_rsv = &root->fs_info->delayed_block_rsv;
3923         struct btrfs_trans_handle *trans;
3924
3925         trans = (struct btrfs_trans_handle *)current->journal_info;
3926         if (trans)
3927                 return -EAGAIN;
3928
3929         if (force)
3930                 goto commit;
3931
3932         /* See if there is enough pinned space to make this reservation */
3933         spin_lock(&space_info->lock);
3934         if (space_info->bytes_pinned >= bytes) {
3935                 spin_unlock(&space_info->lock);
3936                 goto commit;
3937         }
3938         spin_unlock(&space_info->lock);
3939
3940         /*
3941          * See if there is some space in the delayed insertion reservation for
3942          * this reservation.
3943          */
3944         if (space_info != delayed_rsv->space_info)
3945                 return -ENOSPC;
3946
3947         spin_lock(&space_info->lock);
3948         spin_lock(&delayed_rsv->lock);
3949         if (space_info->bytes_pinned + delayed_rsv->size < bytes) {
3950                 spin_unlock(&delayed_rsv->lock);
3951                 spin_unlock(&space_info->lock);
3952                 return -ENOSPC;
3953         }
3954         spin_unlock(&delayed_rsv->lock);
3955         spin_unlock(&space_info->lock);
3956
3957 commit:
3958         trans = btrfs_join_transaction(root);
3959         if (IS_ERR(trans))
3960                 return -ENOSPC;
3961
3962         return btrfs_commit_transaction(trans, root);
3963 }
3964
3965 enum flush_state {
3966         FLUSH_DELAYED_ITEMS_NR  =       1,
3967         FLUSH_DELAYED_ITEMS     =       2,
3968         FLUSH_DELALLOC          =       3,
3969         FLUSH_DELALLOC_WAIT     =       4,
3970         ALLOC_CHUNK             =       5,
3971         COMMIT_TRANS            =       6,
3972 };
3973
3974 static int flush_space(struct btrfs_root *root,
3975                        struct btrfs_space_info *space_info, u64 num_bytes,
3976                        u64 orig_bytes, int state)
3977 {
3978         struct btrfs_trans_handle *trans;
3979         int nr;
3980         int ret = 0;
3981
3982         switch (state) {
3983         case FLUSH_DELAYED_ITEMS_NR:
3984         case FLUSH_DELAYED_ITEMS:
3985                 if (state == FLUSH_DELAYED_ITEMS_NR) {
3986                         u64 bytes = btrfs_calc_trans_metadata_size(root, 1);
3987
3988                         nr = (int)div64_u64(num_bytes, bytes);
3989                         if (!nr)
3990                                 nr = 1;
3991                         nr *= 2;
3992                 } else {
3993                         nr = -1;
3994                 }
3995                 trans = btrfs_join_transaction(root);
3996                 if (IS_ERR(trans)) {
3997                         ret = PTR_ERR(trans);
3998                         break;
3999                 }
4000                 ret = btrfs_run_delayed_items_nr(trans, root, nr);
4001                 btrfs_end_transaction(trans, root);
4002                 break;
4003         case FLUSH_DELALLOC:
4004         case FLUSH_DELALLOC_WAIT:
4005                 shrink_delalloc(root, num_bytes, orig_bytes,
4006                                 state == FLUSH_DELALLOC_WAIT);
4007                 break;
4008         case ALLOC_CHUNK:
4009                 trans = btrfs_join_transaction(root);
4010                 if (IS_ERR(trans)) {
4011                         ret = PTR_ERR(trans);
4012                         break;
4013                 }
4014                 ret = do_chunk_alloc(trans, root->fs_info->extent_root,
4015                                      btrfs_get_alloc_profile(root, 0),
4016                                      CHUNK_ALLOC_NO_FORCE);
4017                 btrfs_end_transaction(trans, root);
4018                 if (ret == -ENOSPC)
4019                         ret = 0;
4020                 break;
4021         case COMMIT_TRANS:
4022                 ret = may_commit_transaction(root, space_info, orig_bytes, 0);
4023                 break;
4024         default:
4025                 ret = -ENOSPC;
4026                 break;
4027         }
4028
4029         return ret;
4030 }
4031 /**
4032  * reserve_metadata_bytes - try to reserve bytes from the block_rsv's space
4033  * @root - the root we're allocating for
4034  * @block_rsv - the block_rsv we're allocating for
4035  * @orig_bytes - the number of bytes we want
4036  * @flush - whether or not we can flush to make our reservation
4037  *
4038  * This will reserve orgi_bytes number of bytes from the space info associated
4039  * with the block_rsv.  If there is not enough space it will make an attempt to
4040  * flush out space to make room.  It will do this by flushing delalloc if
4041  * possible or committing the transaction.  If flush is 0 then no attempts to
4042  * regain reservations will be made and this will fail if there is not enough
4043  * space already.
4044  */
4045 static int reserve_metadata_bytes(struct btrfs_root *root,
4046                                   struct btrfs_block_rsv *block_rsv,
4047                                   u64 orig_bytes,
4048                                   enum btrfs_reserve_flush_enum flush)
4049 {
4050         struct btrfs_space_info *space_info = block_rsv->space_info;
4051         u64 used;
4052         u64 num_bytes = orig_bytes;
4053         int flush_state = FLUSH_DELAYED_ITEMS_NR;
4054         int ret = 0;
4055         bool flushing = false;
4056
4057 again:
4058         ret = 0;
4059         spin_lock(&space_info->lock);
4060         /*
4061          * We only want to wait if somebody other than us is flushing and we
4062          * are actually allowed to flush all things.
4063          */
4064         while (flush == BTRFS_RESERVE_FLUSH_ALL && !flushing &&
4065                space_info->flush) {
4066                 spin_unlock(&space_info->lock);
4067                 /*
4068                  * If we have a trans handle we can't wait because the flusher
4069                  * may have to commit the transaction, which would mean we would
4070                  * deadlock since we are waiting for the flusher to finish, but
4071                  * hold the current transaction open.
4072                  */
4073                 if (current->journal_info)
4074                         return -EAGAIN;
4075                 ret = wait_event_killable(space_info->wait, !space_info->flush);
4076                 /* Must have been killed, return */
4077                 if (ret)
4078                         return -EINTR;
4079
4080                 spin_lock(&space_info->lock);
4081         }
4082
4083         ret = -ENOSPC;
4084         used = space_info->bytes_used + space_info->bytes_reserved +
4085                 space_info->bytes_pinned + space_info->bytes_readonly +
4086                 space_info->bytes_may_use;
4087
4088         /*
4089          * The idea here is that we've not already over-reserved the block group
4090          * then we can go ahead and save our reservation first and then start
4091          * flushing if we need to.  Otherwise if we've already overcommitted
4092          * lets start flushing stuff first and then come back and try to make
4093          * our reservation.
4094          */
4095         if (used <= space_info->total_bytes) {
4096                 if (used + orig_bytes <= space_info->total_bytes) {
4097                         space_info->bytes_may_use += orig_bytes;
4098                         trace_btrfs_space_reservation(root->fs_info,
4099                                 "space_info", space_info->flags, orig_bytes, 1);
4100                         ret = 0;
4101                 } else {
4102                         /*
4103                          * Ok set num_bytes to orig_bytes since we aren't
4104                          * overocmmitted, this way we only try and reclaim what
4105                          * we need.
4106                          */
4107                         num_bytes = orig_bytes;
4108                 }
4109         } else {
4110                 /*
4111                  * Ok we're over committed, set num_bytes to the overcommitted
4112                  * amount plus the amount of bytes that we need for this
4113                  * reservation.
4114                  */
4115                 num_bytes = used - space_info->total_bytes +
4116                         (orig_bytes * 2);
4117         }
4118
4119         if (ret && can_overcommit(root, space_info, orig_bytes, flush)) {
4120                 space_info->bytes_may_use += orig_bytes;
4121                 trace_btrfs_space_reservation(root->fs_info, "space_info",
4122                                               space_info->flags, orig_bytes,
4123                                               1);
4124                 ret = 0;
4125         }
4126
4127         /*
4128          * Couldn't make our reservation, save our place so while we're trying
4129          * to reclaim space we can actually use it instead of somebody else
4130          * stealing it from us.
4131          *
4132          * We make the other tasks wait for the flush only when we can flush
4133          * all things.
4134          */
4135         if (ret && flush != BTRFS_RESERVE_NO_FLUSH) {
4136                 flushing = true;
4137                 space_info->flush = 1;
4138         }
4139
4140         spin_unlock(&space_info->lock);
4141
4142         if (!ret || flush == BTRFS_RESERVE_NO_FLUSH)
4143                 goto out;
4144
4145         ret = flush_space(root, space_info, num_bytes, orig_bytes,
4146                           flush_state);
4147         flush_state++;
4148
4149         /*
4150          * If we are FLUSH_LIMIT, we can not flush delalloc, or the deadlock
4151          * would happen. So skip delalloc flush.
4152          */
4153         if (flush == BTRFS_RESERVE_FLUSH_LIMIT &&
4154             (flush_state == FLUSH_DELALLOC ||
4155              flush_state == FLUSH_DELALLOC_WAIT))
4156                 flush_state = ALLOC_CHUNK;
4157
4158         if (!ret)
4159                 goto again;
4160         else if (flush == BTRFS_RESERVE_FLUSH_LIMIT &&
4161                  flush_state < COMMIT_TRANS)
4162                 goto again;
4163         else if (flush == BTRFS_RESERVE_FLUSH_ALL &&
4164                  flush_state <= COMMIT_TRANS)
4165                 goto again;
4166
4167 out:
4168         if (ret == -ENOSPC &&
4169             unlikely(root->orphan_cleanup_state == ORPHAN_CLEANUP_STARTED)) {
4170                 struct btrfs_block_rsv *global_rsv =
4171                         &root->fs_info->global_block_rsv;
4172
4173                 if (block_rsv != global_rsv &&
4174                     !block_rsv_use_bytes(global_rsv, orig_bytes))
4175                         ret = 0;
4176         }
4177         if (flushing) {
4178                 spin_lock(&space_info->lock);
4179                 space_info->flush = 0;
4180                 wake_up_all(&space_info->wait);
4181                 spin_unlock(&space_info->lock);
4182         }
4183         return ret;
4184 }
4185
4186 static struct btrfs_block_rsv *get_block_rsv(
4187                                         const struct btrfs_trans_handle *trans,
4188                                         const struct btrfs_root *root)
4189 {
4190         struct btrfs_block_rsv *block_rsv = NULL;
4191
4192         if (root->ref_cows)
4193                 block_rsv = trans->block_rsv;
4194
4195         if (root == root->fs_info->csum_root && trans->adding_csums)
4196                 block_rsv = trans->block_rsv;
4197
4198         if (!block_rsv)
4199                 block_rsv = root->block_rsv;
4200
4201         if (!block_rsv)
4202                 block_rsv = &root->fs_info->empty_block_rsv;
4203
4204         return block_rsv;
4205 }
4206
4207 static int block_rsv_use_bytes(struct btrfs_block_rsv *block_rsv,
4208                                u64 num_bytes)
4209 {
4210         int ret = -ENOSPC;
4211         spin_lock(&block_rsv->lock);
4212         if (block_rsv->reserved >= num_bytes) {
4213                 block_rsv->reserved -= num_bytes;
4214                 if (block_rsv->reserved < block_rsv->size)
4215                         block_rsv->full = 0;
4216                 ret = 0;
4217         }
4218         spin_unlock(&block_rsv->lock);
4219         return ret;
4220 }
4221
4222 static void block_rsv_add_bytes(struct btrfs_block_rsv *block_rsv,
4223                                 u64 num_bytes, int update_size)
4224 {
4225         spin_lock(&block_rsv->lock);
4226         block_rsv->reserved += num_bytes;
4227         if (update_size)
4228                 block_rsv->size += num_bytes;
4229         else if (block_rsv->reserved >= block_rsv->size)
4230                 block_rsv->full = 1;
4231         spin_unlock(&block_rsv->lock);
4232 }
4233
4234 static void block_rsv_release_bytes(struct btrfs_fs_info *fs_info,
4235                                     struct btrfs_block_rsv *block_rsv,
4236                                     struct btrfs_block_rsv *dest, u64 num_bytes)
4237 {
4238         struct btrfs_space_info *space_info = block_rsv->space_info;
4239
4240         spin_lock(&block_rsv->lock);
4241         if (num_bytes == (u64)-1)
4242                 num_bytes = block_rsv->size;
4243         block_rsv->size -= num_bytes;
4244         if (block_rsv->reserved >= block_rsv->size) {
4245                 num_bytes = block_rsv->reserved - block_rsv->size;
4246                 block_rsv->reserved = block_rsv->size;
4247                 block_rsv->full = 1;
4248         } else {
4249                 num_bytes = 0;
4250         }
4251         spin_unlock(&block_rsv->lock);
4252
4253         if (num_bytes > 0) {
4254                 if (dest) {
4255                         spin_lock(&dest->lock);
4256                         if (!dest->full) {
4257                                 u64 bytes_to_add;
4258
4259                                 bytes_to_add = dest->size - dest->reserved;
4260                                 bytes_to_add = min(num_bytes, bytes_to_add);
4261                                 dest->reserved += bytes_to_add;
4262                                 if (dest->reserved >= dest->size)
4263                                         dest->full = 1;
4264                                 num_bytes -= bytes_to_add;
4265                         }
4266                         spin_unlock(&dest->lock);
4267                 }
4268                 if (num_bytes) {
4269                         spin_lock(&space_info->lock);
4270                         space_info->bytes_may_use -= num_bytes;
4271                         trace_btrfs_space_reservation(fs_info, "space_info",
4272                                         space_info->flags, num_bytes, 0);
4273                         space_info->reservation_progress++;
4274                         spin_unlock(&space_info->lock);
4275                 }
4276         }
4277 }
4278
4279 static int block_rsv_migrate_bytes(struct btrfs_block_rsv *src,
4280                                    struct btrfs_block_rsv *dst, u64 num_bytes)
4281 {
4282         int ret;
4283
4284         ret = block_rsv_use_bytes(src, num_bytes);
4285         if (ret)
4286                 return ret;
4287
4288         block_rsv_add_bytes(dst, num_bytes, 1);
4289         return 0;
4290 }
4291
4292 void btrfs_init_block_rsv(struct btrfs_block_rsv *rsv, unsigned short type)
4293 {
4294         memset(rsv, 0, sizeof(*rsv));
4295         spin_lock_init(&rsv->lock);
4296         rsv->type = type;
4297 }
4298
4299 struct btrfs_block_rsv *btrfs_alloc_block_rsv(struct btrfs_root *root,
4300                                               unsigned short type)
4301 {
4302         struct btrfs_block_rsv *block_rsv;
4303         struct btrfs_fs_info *fs_info = root->fs_info;
4304
4305         block_rsv = kmalloc(sizeof(*block_rsv), GFP_NOFS);
4306         if (!block_rsv)
4307                 return NULL;
4308
4309         btrfs_init_block_rsv(block_rsv, type);
4310         block_rsv->space_info = __find_space_info(fs_info,
4311                                                   BTRFS_BLOCK_GROUP_METADATA);
4312         return block_rsv;
4313 }
4314
4315 void btrfs_free_block_rsv(struct btrfs_root *root,
4316                           struct btrfs_block_rsv *rsv)
4317 {
4318         if (!rsv)
4319                 return;
4320         btrfs_block_rsv_release(root, rsv, (u64)-1);
4321         kfree(rsv);
4322 }
4323
4324 int btrfs_block_rsv_add(struct btrfs_root *root,
4325                         struct btrfs_block_rsv *block_rsv, u64 num_bytes,
4326                         enum btrfs_reserve_flush_enum flush)
4327 {
4328         int ret;
4329
4330         if (num_bytes == 0)
4331                 return 0;
4332
4333         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
4334         if (!ret) {
4335                 block_rsv_add_bytes(block_rsv, num_bytes, 1);
4336                 return 0;
4337         }
4338
4339         return ret;
4340 }
4341
4342 int btrfs_block_rsv_check(struct btrfs_root *root,
4343                           struct btrfs_block_rsv *block_rsv, int min_factor)
4344 {
4345         u64 num_bytes = 0;
4346         int ret = -ENOSPC;
4347
4348         if (!block_rsv)
4349                 return 0;
4350
4351         spin_lock(&block_rsv->lock);
4352         num_bytes = div_factor(block_rsv->size, min_factor);
4353         if (block_rsv->reserved >= num_bytes)
4354                 ret = 0;
4355         spin_unlock(&block_rsv->lock);
4356
4357         return ret;
4358 }
4359
4360 int btrfs_block_rsv_refill(struct btrfs_root *root,
4361                            struct btrfs_block_rsv *block_rsv, u64 min_reserved,
4362                            enum btrfs_reserve_flush_enum flush)
4363 {
4364         u64 num_bytes = 0;
4365         int ret = -ENOSPC;
4366
4367         if (!block_rsv)
4368                 return 0;
4369
4370         spin_lock(&block_rsv->lock);
4371         num_bytes = min_reserved;
4372         if (block_rsv->reserved >= num_bytes)
4373                 ret = 0;
4374         else
4375                 num_bytes -= block_rsv->reserved;
4376         spin_unlock(&block_rsv->lock);
4377
4378         if (!ret)
4379                 return 0;
4380
4381         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
4382         if (!ret) {
4383                 block_rsv_add_bytes(block_rsv, num_bytes, 0);
4384                 return 0;
4385         }
4386
4387         return ret;
4388 }
4389
4390 int btrfs_block_rsv_migrate(struct btrfs_block_rsv *src_rsv,
4391                             struct btrfs_block_rsv *dst_rsv,
4392                             u64 num_bytes)
4393 {
4394         return block_rsv_migrate_bytes(src_rsv, dst_rsv, num_bytes);
4395 }
4396
4397 void btrfs_block_rsv_release(struct btrfs_root *root,
4398                              struct btrfs_block_rsv *block_rsv,
4399                              u64 num_bytes)
4400 {
4401         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
4402         if (global_rsv->full || global_rsv == block_rsv ||
4403             block_rsv->space_info != global_rsv->space_info)
4404                 global_rsv = NULL;
4405         block_rsv_release_bytes(root->fs_info, block_rsv, global_rsv,
4406                                 num_bytes);
4407 }
4408
4409 /*
4410  * helper to calculate size of global block reservation.
4411  * the desired value is sum of space used by extent tree,
4412  * checksum tree and root tree
4413  */
4414 static u64 calc_global_metadata_size(struct btrfs_fs_info *fs_info)
4415 {
4416         struct btrfs_space_info *sinfo;
4417         u64 num_bytes;
4418         u64 meta_used;
4419         u64 data_used;
4420         int csum_size = btrfs_super_csum_size(fs_info->super_copy);
4421
4422         sinfo = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_DATA);
4423         spin_lock(&sinfo->lock);
4424         data_used = sinfo->bytes_used;
4425         spin_unlock(&sinfo->lock);
4426
4427         sinfo = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
4428         spin_lock(&sinfo->lock);
4429         if (sinfo->flags & BTRFS_BLOCK_GROUP_DATA)
4430                 data_used = 0;
4431         meta_used = sinfo->bytes_used;
4432         spin_unlock(&sinfo->lock);
4433
4434         num_bytes = (data_used >> fs_info->sb->s_blocksize_bits) *
4435                     csum_size * 2;
4436         num_bytes += div64_u64(data_used + meta_used, 50);
4437
4438         if (num_bytes * 3 > meta_used)
4439                 num_bytes = div64_u64(meta_used, 3);
4440
4441         return ALIGN(num_bytes, fs_info->extent_root->leafsize << 10);
4442 }
4443
4444 static void update_global_block_rsv(struct btrfs_fs_info *fs_info)
4445 {
4446         struct btrfs_block_rsv *block_rsv = &fs_info->global_block_rsv;
4447         struct btrfs_space_info *sinfo = block_rsv->space_info;
4448         u64 num_bytes;
4449
4450         num_bytes = calc_global_metadata_size(fs_info);
4451
4452         spin_lock(&sinfo->lock);
4453         spin_lock(&block_rsv->lock);
4454
4455         block_rsv->size = num_bytes;
4456
4457         num_bytes = sinfo->bytes_used + sinfo->bytes_pinned +
4458                     sinfo->bytes_reserved + sinfo->bytes_readonly +
4459                     sinfo->bytes_may_use;
4460
4461         if (sinfo->total_bytes > num_bytes) {
4462                 num_bytes = sinfo->total_bytes - num_bytes;
4463                 block_rsv->reserved += num_bytes;
4464                 sinfo->bytes_may_use += num_bytes;
4465                 trace_btrfs_space_reservation(fs_info, "space_info",
4466                                       sinfo->flags, num_bytes, 1);
4467         }
4468
4469         if (block_rsv->reserved >= block_rsv->size) {
4470                 num_bytes = block_rsv->reserved - block_rsv->size;
4471                 sinfo->bytes_may_use -= num_bytes;
4472                 trace_btrfs_space_reservation(fs_info, "space_info",
4473                                       sinfo->flags, num_bytes, 0);
4474                 sinfo->reservation_progress++;
4475                 block_rsv->reserved = block_rsv->size;
4476                 block_rsv->full = 1;
4477         }
4478
4479         spin_unlock(&block_rsv->lock);
4480         spin_unlock(&sinfo->lock);
4481 }
4482
4483 static void init_global_block_rsv(struct btrfs_fs_info *fs_info)
4484 {
4485         struct btrfs_space_info *space_info;
4486
4487         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
4488         fs_info->chunk_block_rsv.space_info = space_info;
4489
4490         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
4491         fs_info->global_block_rsv.space_info = space_info;
4492         fs_info->delalloc_block_rsv.space_info = space_info;
4493         fs_info->trans_block_rsv.space_info = space_info;
4494         fs_info->empty_block_rsv.space_info = space_info;
4495         fs_info->delayed_block_rsv.space_info = space_info;
4496
4497         fs_info->extent_root->block_rsv = &fs_info->global_block_rsv;
4498         fs_info->csum_root->block_rsv = &fs_info->global_block_rsv;
4499         fs_info->dev_root->block_rsv = &fs_info->global_block_rsv;
4500         fs_info->tree_root->block_rsv = &fs_info->global_block_rsv;
4501         fs_info->chunk_root->block_rsv = &fs_info->chunk_block_rsv;
4502
4503         update_global_block_rsv(fs_info);
4504 }
4505
4506 static void release_global_block_rsv(struct btrfs_fs_info *fs_info)
4507 {
4508         block_rsv_release_bytes(fs_info, &fs_info->global_block_rsv, NULL,
4509                                 (u64)-1);
4510         WARN_ON(fs_info->delalloc_block_rsv.size > 0);
4511         WARN_ON(fs_info->delalloc_block_rsv.reserved > 0);
4512         WARN_ON(fs_info->trans_block_rsv.size > 0);
4513         WARN_ON(fs_info->trans_block_rsv.reserved > 0);
4514         WARN_ON(fs_info->chunk_block_rsv.size > 0);
4515         WARN_ON(fs_info->chunk_block_rsv.reserved > 0);
4516         WARN_ON(fs_info->delayed_block_rsv.size > 0);
4517         WARN_ON(fs_info->delayed_block_rsv.reserved > 0);
4518 }
4519
4520 void btrfs_trans_release_metadata(struct btrfs_trans_handle *trans,
4521                                   struct btrfs_root *root)
4522 {
4523         if (!trans->block_rsv)
4524                 return;
4525
4526         if (!trans->bytes_reserved)
4527                 return;
4528
4529         trace_btrfs_space_reservation(root->fs_info, "transaction",
4530                                       trans->transid, trans->bytes_reserved, 0);
4531         btrfs_block_rsv_release(root, trans->block_rsv, trans->bytes_reserved);
4532         trans->bytes_reserved = 0;
4533 }
4534
4535 /* Can only return 0 or -ENOSPC */
4536 int btrfs_orphan_reserve_metadata(struct btrfs_trans_handle *trans,
4537                                   struct inode *inode)
4538 {
4539         struct btrfs_root *root = BTRFS_I(inode)->root;
4540         struct btrfs_block_rsv *src_rsv = get_block_rsv(trans, root);
4541         struct btrfs_block_rsv *dst_rsv = root->orphan_block_rsv;
4542
4543         /*
4544          * We need to hold space in order to delete our orphan item once we've
4545          * added it, so this takes the reservation so we can release it later
4546          * when we are truly done with the orphan item.
4547          */
4548         u64 num_bytes = btrfs_calc_trans_metadata_size(root, 1);
4549         trace_btrfs_space_reservation(root->fs_info, "orphan",
4550                                       btrfs_ino(inode), num_bytes, 1);
4551         return block_rsv_migrate_bytes(src_rsv, dst_rsv, num_bytes);
4552 }
4553
4554 void btrfs_orphan_release_metadata(struct inode *inode)
4555 {
4556         struct btrfs_root *root = BTRFS_I(inode)->root;
4557         u64 num_bytes = btrfs_calc_trans_metadata_size(root, 1);
4558         trace_btrfs_space_reservation(root->fs_info, "orphan",
4559                                       btrfs_ino(inode), num_bytes, 0);
4560         btrfs_block_rsv_release(root, root->orphan_block_rsv, num_bytes);
4561 }
4562
4563 /*
4564  * btrfs_subvolume_reserve_metadata() - reserve space for subvolume operation
4565  * root: the root of the parent directory
4566  * rsv: block reservation
4567  * items: the number of items that we need do reservation
4568  * qgroup_reserved: used to return the reserved size in qgroup
4569  *
4570  * This function is used to reserve the space for snapshot/subvolume
4571  * creation and deletion. Those operations are different with the
4572  * common file/directory operations, they change two fs/file trees
4573  * and root tree, the number of items that the qgroup reserves is
4574  * different with the free space reservation. So we can not use
4575  * the space reseravtion mechanism in start_transaction().
4576  */
4577 int btrfs_subvolume_reserve_metadata(struct btrfs_root *root,
4578                                      struct btrfs_block_rsv *rsv,
4579                                      int items,
4580                                      u64 *qgroup_reserved)
4581 {
4582         u64 num_bytes;
4583         int ret;
4584
4585         if (root->fs_info->quota_enabled) {
4586                 /* One for parent inode, two for dir entries */
4587                 num_bytes = 3 * root->leafsize;
4588                 ret = btrfs_qgroup_reserve(root, num_bytes);
4589                 if (ret)
4590                         return ret;
4591         } else {
4592                 num_bytes = 0;
4593         }
4594
4595         *qgroup_reserved = num_bytes;
4596
4597         num_bytes = btrfs_calc_trans_metadata_size(root, items);
4598         rsv->space_info = __find_space_info(root->fs_info,
4599                                             BTRFS_BLOCK_GROUP_METADATA);
4600         ret = btrfs_block_rsv_add(root, rsv, num_bytes,
4601                                   BTRFS_RESERVE_FLUSH_ALL);
4602         if (ret) {
4603                 if (*qgroup_reserved)
4604                         btrfs_qgroup_free(root, *qgroup_reserved);
4605         }
4606
4607         return ret;
4608 }
4609
4610 void btrfs_subvolume_release_metadata(struct btrfs_root *root,
4611                                       struct btrfs_block_rsv *rsv,
4612                                       u64 qgroup_reserved)
4613 {
4614         btrfs_block_rsv_release(root, rsv, (u64)-1);
4615         if (qgroup_reserved)
4616                 btrfs_qgroup_free(root, qgroup_reserved);
4617 }
4618
4619 /**
4620  * drop_outstanding_extent - drop an outstanding extent
4621  * @inode: the inode we're dropping the extent for
4622  *
4623  * This is called when we are freeing up an outstanding extent, either called
4624  * after an error or after an extent is written.  This will return the number of
4625  * reserved extents that need to be freed.  This must be called with
4626  * BTRFS_I(inode)->lock held.
4627  */
4628 static unsigned drop_outstanding_extent(struct inode *inode)
4629 {
4630         unsigned drop_inode_space = 0;
4631         unsigned dropped_extents = 0;
4632
4633         BUG_ON(!BTRFS_I(inode)->outstanding_extents);
4634         BTRFS_I(inode)->outstanding_extents--;
4635
4636         if (BTRFS_I(inode)->outstanding_extents == 0 &&
4637             test_and_clear_bit(BTRFS_INODE_DELALLOC_META_RESERVED,
4638                                &BTRFS_I(inode)->runtime_flags))
4639                 drop_inode_space = 1;
4640
4641         /*
4642          * If we have more or the same amount of outsanding extents than we have
4643          * reserved then we need to leave the reserved extents count alone.
4644          */
4645         if (BTRFS_I(inode)->outstanding_extents >=
4646             BTRFS_I(inode)->reserved_extents)
4647                 return drop_inode_space;
4648
4649         dropped_extents = BTRFS_I(inode)->reserved_extents -
4650                 BTRFS_I(inode)->outstanding_extents;
4651         BTRFS_I(inode)->reserved_extents -= dropped_extents;
4652         return dropped_extents + drop_inode_space;
4653 }
4654
4655 /**
4656  * calc_csum_metadata_size - return the amount of metada space that must be
4657  *      reserved/free'd for the given bytes.
4658  * @inode: the inode we're manipulating
4659  * @num_bytes: the number of bytes in question
4660  * @reserve: 1 if we are reserving space, 0 if we are freeing space
4661  *
4662  * This adjusts the number of csum_bytes in the inode and then returns the
4663  * correct amount of metadata that must either be reserved or freed.  We
4664  * calculate how many checksums we can fit into one leaf and then divide the
4665  * number of bytes that will need to be checksumed by this value to figure out
4666  * how many checksums will be required.  If we are adding bytes then the number
4667  * may go up and we will return the number of additional bytes that must be
4668  * reserved.  If it is going down we will return the number of bytes that must
4669  * be freed.
4670  *
4671  * This must be called with BTRFS_I(inode)->lock held.
4672  */
4673 static u64 calc_csum_metadata_size(struct inode *inode, u64 num_bytes,
4674                                    int reserve)
4675 {
4676         struct btrfs_root *root = BTRFS_I(inode)->root;
4677         u64 csum_size;
4678         int num_csums_per_leaf;
4679         int num_csums;
4680         int old_csums;
4681
4682         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM &&
4683             BTRFS_I(inode)->csum_bytes == 0)
4684                 return 0;
4685
4686         old_csums = (int)div64_u64(BTRFS_I(inode)->csum_bytes, root->sectorsize);
4687         if (reserve)
4688                 BTRFS_I(inode)->csum_bytes += num_bytes;
4689         else
4690                 BTRFS_I(inode)->csum_bytes -= num_bytes;
4691         csum_size = BTRFS_LEAF_DATA_SIZE(root) - sizeof(struct btrfs_item);
4692         num_csums_per_leaf = (int)div64_u64(csum_size,
4693                                             sizeof(struct btrfs_csum_item) +
4694                                             sizeof(struct btrfs_disk_key));
4695         num_csums = (int)div64_u64(BTRFS_I(inode)->csum_bytes, root->sectorsize);
4696         num_csums = num_csums + num_csums_per_leaf - 1;
4697         num_csums = num_csums / num_csums_per_leaf;
4698
4699         old_csums = old_csums + num_csums_per_leaf - 1;
4700         old_csums = old_csums / num_csums_per_leaf;
4701
4702         /* No change, no need to reserve more */
4703         if (old_csums == num_csums)
4704                 return 0;
4705
4706         if (reserve)
4707                 return btrfs_calc_trans_metadata_size(root,
4708                                                       num_csums - old_csums);
4709
4710         return btrfs_calc_trans_metadata_size(root, old_csums - num_csums);
4711 }
4712
4713 int btrfs_delalloc_reserve_metadata(struct inode *inode, u64 num_bytes)
4714 {
4715         struct btrfs_root *root = BTRFS_I(inode)->root;
4716         struct btrfs_block_rsv *block_rsv = &root->fs_info->delalloc_block_rsv;
4717         u64 to_reserve = 0;
4718         u64 csum_bytes;
4719         unsigned nr_extents = 0;
4720         int extra_reserve = 0;
4721         enum btrfs_reserve_flush_enum flush = BTRFS_RESERVE_FLUSH_ALL;
4722         int ret = 0;
4723         bool delalloc_lock = true;
4724
4725         /* If we are a free space inode we need to not flush since we will be in
4726          * the middle of a transaction commit.  We also don't need the delalloc
4727          * mutex since we won't race with anybody.  We need this mostly to make
4728          * lockdep shut its filthy mouth.
4729          */
4730         if (btrfs_is_free_space_inode(inode)) {
4731                 flush = BTRFS_RESERVE_NO_FLUSH;
4732                 delalloc_lock = false;
4733         }
4734
4735         if (flush != BTRFS_RESERVE_NO_FLUSH &&
4736             btrfs_transaction_in_commit(root->fs_info))
4737                 schedule_timeout(1);
4738
4739         if (delalloc_lock)
4740                 mutex_lock(&BTRFS_I(inode)->delalloc_mutex);
4741
4742         num_bytes = ALIGN(num_bytes, root->sectorsize);
4743
4744         spin_lock(&BTRFS_I(inode)->lock);
4745         BTRFS_I(inode)->outstanding_extents++;
4746
4747         if (BTRFS_I(inode)->outstanding_extents >
4748             BTRFS_I(inode)->reserved_extents)
4749                 nr_extents = BTRFS_I(inode)->outstanding_extents -
4750                         BTRFS_I(inode)->reserved_extents;
4751
4752         /*
4753          * Add an item to reserve for updating the inode when we complete the
4754          * delalloc io.
4755          */
4756         if (!test_bit(BTRFS_INODE_DELALLOC_META_RESERVED,
4757                       &BTRFS_I(inode)->runtime_flags)) {
4758                 nr_extents++;
4759                 extra_reserve = 1;
4760         }
4761
4762         to_reserve = btrfs_calc_trans_metadata_size(root, nr_extents);
4763         to_reserve += calc_csum_metadata_size(inode, num_bytes, 1);
4764         csum_bytes = BTRFS_I(inode)->csum_bytes;
4765         spin_unlock(&BTRFS_I(inode)->lock);
4766
4767         if (root->fs_info->quota_enabled)
4768                 ret = btrfs_qgroup_reserve(root, num_bytes +
4769                                            nr_extents * root->leafsize);
4770
4771         /*
4772          * ret != 0 here means the qgroup reservation failed, we go straight to
4773          * the shared error handling then.
4774          */
4775         if (ret == 0) {
4776                 ret = reserve_metadata_bytes(root, block_rsv,
4777                                              to_reserve, flush);
4778                 if (ret && root->fs_info->quota_enabled) {
4779                         btrfs_qgroup_free(root, num_bytes +
4780                                                 nr_extents * root->leafsize);
4781                 }
4782         }
4783
4784         if (ret) {
4785                 u64 to_free = 0;
4786                 unsigned dropped;
4787
4788                 spin_lock(&BTRFS_I(inode)->lock);
4789                 dropped = drop_outstanding_extent(inode);
4790                 /*
4791                  * If the inodes csum_bytes is the same as the original
4792                  * csum_bytes then we know we haven't raced with any free()ers
4793                  * so we can just reduce our inodes csum bytes and carry on.
4794                  * Otherwise we have to do the normal free thing to account for
4795                  * the case that the free side didn't free up its reserve
4796                  * because of this outstanding reservation.
4797                  */
4798                 if (BTRFS_I(inode)->csum_bytes == csum_bytes)
4799                         calc_csum_metadata_size(inode, num_bytes, 0);
4800                 else
4801                         to_free = calc_csum_metadata_size(inode, num_bytes, 0);
4802                 spin_unlock(&BTRFS_I(inode)->lock);
4803                 if (dropped)
4804                         to_free += btrfs_calc_trans_metadata_size(root, dropped);
4805
4806                 if (to_free) {
4807                         btrfs_block_rsv_release(root, block_rsv, to_free);
4808                         trace_btrfs_space_reservation(root->fs_info,
4809                                                       "delalloc",
4810                                                       btrfs_ino(inode),
4811                                                       to_free, 0);
4812                 }
4813                 if (delalloc_lock)
4814                         mutex_unlock(&BTRFS_I(inode)->delalloc_mutex);
4815                 return ret;
4816         }
4817
4818         spin_lock(&BTRFS_I(inode)->lock);
4819         if (extra_reserve) {
4820                 set_bit(BTRFS_INODE_DELALLOC_META_RESERVED,
4821                         &BTRFS_I(inode)->runtime_flags);
4822                 nr_extents--;
4823         }
4824         BTRFS_I(inode)->reserved_extents += nr_extents;
4825         spin_unlock(&BTRFS_I(inode)->lock);
4826
4827         if (delalloc_lock)
4828                 mutex_unlock(&BTRFS_I(inode)->delalloc_mutex);
4829
4830         if (to_reserve)
4831                 trace_btrfs_space_reservation(root->fs_info,"delalloc",
4832                                               btrfs_ino(inode), to_reserve, 1);
4833         block_rsv_add_bytes(block_rsv, to_reserve, 1);
4834
4835         return 0;
4836 }
4837
4838 /**
4839  * btrfs_delalloc_release_metadata - release a metadata reservation for an inode
4840  * @inode: the inode to release the reservation for
4841  * @num_bytes: the number of bytes we're releasing
4842  *
4843  * This will release the metadata reservation for an inode.  This can be called
4844  * once we complete IO for a given set of bytes to release their metadata
4845  * reservations.
4846  */
4847 void btrfs_delalloc_release_metadata(struct inode *inode, u64 num_bytes)
4848 {
4849         struct btrfs_root *root = BTRFS_I(inode)->root;
4850         u64 to_free = 0;
4851         unsigned dropped;
4852
4853         num_bytes = ALIGN(num_bytes, root->sectorsize);
4854         spin_lock(&BTRFS_I(inode)->lock);
4855         dropped = drop_outstanding_extent(inode);
4856
4857         if (num_bytes)
4858                 to_free = calc_csum_metadata_size(inode, num_bytes, 0);
4859         spin_unlock(&BTRFS_I(inode)->lock);
4860         if (dropped > 0)
4861                 to_free += btrfs_calc_trans_metadata_size(root, dropped);
4862
4863         trace_btrfs_space_reservation(root->fs_info, "delalloc",
4864                                       btrfs_ino(inode), to_free, 0);
4865         if (root->fs_info->quota_enabled) {
4866                 btrfs_qgroup_free(root, num_bytes +
4867                                         dropped * root->leafsize);
4868         }
4869
4870         btrfs_block_rsv_release(root, &root->fs_info->delalloc_block_rsv,
4871                                 to_free);
4872 }
4873
4874 /**
4875  * btrfs_delalloc_reserve_space - reserve data and metadata space for delalloc
4876  * @inode: inode we're writing to
4877  * @num_bytes: the number of bytes we want to allocate
4878  *
4879  * This will do the following things
4880  *
4881  * o reserve space in the data space info for num_bytes
4882  * o reserve space in the metadata space info based on number of outstanding
4883  *   extents and how much csums will be needed
4884  * o add to the inodes ->delalloc_bytes
4885  * o add it to the fs_info's delalloc inodes list.
4886  *
4887  * This will return 0 for success and -ENOSPC if there is no space left.
4888  */
4889 int btrfs_delalloc_reserve_space(struct inode *inode, u64 num_bytes)
4890 {
4891         int ret;
4892
4893         ret = btrfs_check_data_free_space(inode, num_bytes);
4894         if (ret)
4895                 return ret;
4896
4897         ret = btrfs_delalloc_reserve_metadata(inode, num_bytes);
4898         if (ret) {
4899                 btrfs_free_reserved_data_space(inode, num_bytes);
4900                 return ret;
4901         }
4902
4903         return 0;
4904 }
4905
4906 /**
4907  * btrfs_delalloc_release_space - release data and metadata space for delalloc
4908  * @inode: inode we're releasing space for
4909  * @num_bytes: the number of bytes we want to free up
4910  *
4911  * This must be matched with a call to btrfs_delalloc_reserve_space.  This is
4912  * called in the case that we don't need the metadata AND data reservations
4913  * anymore.  So if there is an error or we insert an inline extent.
4914  *
4915  * This function will release the metadata space that was not used and will
4916  * decrement ->delalloc_bytes and remove it from the fs_info delalloc_inodes
4917  * list if there are no delalloc bytes left.
4918  */
4919 void btrfs_delalloc_release_space(struct inode *inode, u64 num_bytes)
4920 {
4921         btrfs_delalloc_release_metadata(inode, num_bytes);
4922         btrfs_free_reserved_data_space(inode, num_bytes);
4923 }
4924
4925 static int update_block_group(struct btrfs_root *root,
4926                               u64 bytenr, u64 num_bytes, int alloc)
4927 {
4928         struct btrfs_block_group_cache *cache = NULL;
4929         struct btrfs_fs_info *info = root->fs_info;
4930         u64 total = num_bytes;
4931         u64 old_val;
4932         u64 byte_in_group;
4933         int factor;
4934
4935         /* block accounting for super block */
4936         spin_lock(&info->delalloc_lock);
4937         old_val = btrfs_super_bytes_used(info->super_copy);
4938         if (alloc)
4939                 old_val += num_bytes;
4940         else
4941                 old_val -= num_bytes;
4942         btrfs_set_super_bytes_used(info->super_copy, old_val);
4943         spin_unlock(&info->delalloc_lock);
4944
4945         while (total) {
4946                 cache = btrfs_lookup_block_group(info, bytenr);
4947                 if (!cache)
4948                         return -ENOENT;
4949                 if (cache->flags & (BTRFS_BLOCK_GROUP_DUP |
4950                                     BTRFS_BLOCK_GROUP_RAID1 |
4951                                     BTRFS_BLOCK_GROUP_RAID10))
4952                         factor = 2;
4953                 else
4954                         factor = 1;
4955                 /*
4956                  * If this block group has free space cache written out, we
4957                  * need to make sure to load it if we are removing space.  This
4958                  * is because we need the unpinning stage to actually add the
4959                  * space back to the block group, otherwise we will leak space.
4960                  */
4961                 if (!alloc && cache->cached == BTRFS_CACHE_NO)
4962                         cache_block_group(cache, 1);
4963
4964                 byte_in_group = bytenr - cache->key.objectid;
4965                 WARN_ON(byte_in_group > cache->key.offset);
4966
4967                 spin_lock(&cache->space_info->lock);
4968                 spin_lock(&cache->lock);
4969
4970                 if (btrfs_test_opt(root, SPACE_CACHE) &&
4971                     cache->disk_cache_state < BTRFS_DC_CLEAR)
4972                         cache->disk_cache_state = BTRFS_DC_CLEAR;
4973
4974                 cache->dirty = 1;
4975                 old_val = btrfs_block_group_used(&cache->item);
4976                 num_bytes = min(total, cache->key.offset - byte_in_group);
4977                 if (alloc) {
4978                         old_val += num_bytes;
4979                         btrfs_set_block_group_used(&cache->item, old_val);
4980                         cache->reserved -= num_bytes;
4981                         cache->space_info->bytes_reserved -= num_bytes;
4982                         cache->space_info->bytes_used += num_bytes;
4983                         cache->space_info->disk_used += num_bytes * factor;
4984                         spin_unlock(&cache->lock);
4985                         spin_unlock(&cache->space_info->lock);
4986                 } else {
4987                         old_val -= num_bytes;
4988                         btrfs_set_block_group_used(&cache->item, old_val);
4989                         cache->pinned += num_bytes;
4990                         cache->space_info->bytes_pinned += num_bytes;
4991                         cache->space_info->bytes_used -= num_bytes;
4992                         cache->space_info->disk_used -= num_bytes * factor;
4993                         spin_unlock(&cache->lock);
4994                         spin_unlock(&cache->space_info->lock);
4995
4996                         set_extent_dirty(info->pinned_extents,
4997                                          bytenr, bytenr + num_bytes - 1,
4998                                          GFP_NOFS | __GFP_NOFAIL);
4999                 }
5000                 btrfs_put_block_group(cache);
5001                 total -= num_bytes;
5002                 bytenr += num_bytes;
5003         }
5004         return 0;
5005 }
5006
5007 static u64 first_logical_byte(struct btrfs_root *root, u64 search_start)
5008 {
5009         struct btrfs_block_group_cache *cache;
5010         u64 bytenr;
5011
5012         spin_lock(&root->fs_info->block_group_cache_lock);
5013         bytenr = root->fs_info->first_logical_byte;
5014         spin_unlock(&root->fs_info->block_group_cache_lock);
5015
5016         if (bytenr < (u64)-1)
5017                 return bytenr;
5018
5019         cache = btrfs_lookup_first_block_group(root->fs_info, search_start);
5020         if (!cache)
5021                 return 0;
5022
5023         bytenr = cache->key.objectid;
5024         btrfs_put_block_group(cache);
5025
5026         return bytenr;
5027 }
5028
5029 static int pin_down_extent(struct btrfs_root *root,
5030                            struct btrfs_block_group_cache *cache,
5031                            u64 bytenr, u64 num_bytes, int reserved)
5032 {
5033         spin_lock(&cache->space_info->lock);
5034         spin_lock(&cache->lock);
5035         cache->pinned += num_bytes;
5036         cache->space_info->bytes_pinned += num_bytes;
5037         if (reserved) {
5038                 cache->reserved -= num_bytes;
5039                 cache->space_info->bytes_reserved -= num_bytes;
5040         }
5041         spin_unlock(&cache->lock);
5042         spin_unlock(&cache->space_info->lock);
5043
5044         set_extent_dirty(root->fs_info->pinned_extents, bytenr,
5045                          bytenr + num_bytes - 1, GFP_NOFS | __GFP_NOFAIL);
5046         return 0;
5047 }
5048
5049 /*
5050  * this function must be called within transaction
5051  */
5052 int btrfs_pin_extent(struct btrfs_root *root,
5053                      u64 bytenr, u64 num_bytes, int reserved)
5054 {
5055         struct btrfs_block_group_cache *cache;
5056
5057         cache = btrfs_lookup_block_group(root->fs_info, bytenr);
5058         BUG_ON(!cache); /* Logic error */
5059
5060         pin_down_extent(root, cache, bytenr, num_bytes, reserved);
5061
5062         btrfs_put_block_group(cache);
5063         return 0;
5064 }
5065
5066 /*
5067  * this function must be called within transaction
5068  */
5069 int btrfs_pin_extent_for_log_replay(struct btrfs_root *root,
5070                                     u64 bytenr, u64 num_bytes)
5071 {
5072         struct btrfs_block_group_cache *cache;
5073
5074         cache = btrfs_lookup_block_group(root->fs_info, bytenr);
5075         BUG_ON(!cache); /* Logic error */
5076
5077         /*
5078          * pull in the free space cache (if any) so that our pin
5079          * removes the free space from the cache.  We have load_only set
5080          * to one because the slow code to read in the free extents does check
5081          * the pinned extents.
5082          */
5083         cache_block_group(cache, 1);
5084
5085         pin_down_extent(root, cache, bytenr, num_bytes, 0);
5086
5087         /* remove us from the free space cache (if we're there at all) */
5088         btrfs_remove_free_space(cache, bytenr, num_bytes);
5089         btrfs_put_block_group(cache);
5090         return 0;
5091 }
5092
5093 /**
5094  * btrfs_update_reserved_bytes - update the block_group and space info counters
5095  * @cache:      The cache we are manipulating
5096  * @num_bytes:  The number of bytes in question
5097  * @reserve:    One of the reservation enums
5098  *
5099  * This is called by the allocator when it reserves space, or by somebody who is
5100  * freeing space that was never actually used on disk.  For example if you
5101  * reserve some space for a new leaf in transaction A and before transaction A
5102  * commits you free that leaf, you call this with reserve set to 0 in order to
5103  * clear the reservation.
5104  *
5105  * Metadata reservations should be called with RESERVE_ALLOC so we do the proper
5106  * ENOSPC accounting.  For data we handle the reservation through clearing the
5107  * delalloc bits in the io_tree.  We have to do this since we could end up
5108  * allocating less disk space for the amount of data we have reserved in the
5109  * case of compression.
5110  *
5111  * If this is a reservation and the block group has become read only we cannot
5112  * make the reservation and return -EAGAIN, otherwise this function always
5113  * succeeds.
5114  */
5115 static int btrfs_update_reserved_bytes(struct btrfs_block_group_cache *cache,
5116                                        u64 num_bytes, int reserve)
5117 {
5118         struct btrfs_space_info *space_info = cache->space_info;
5119         int ret = 0;
5120
5121         spin_lock(&space_info->lock);
5122         spin_lock(&cache->lock);
5123         if (reserve != RESERVE_FREE) {
5124                 if (cache->ro) {
5125                         ret = -EAGAIN;
5126                 } else {
5127                         cache->reserved += num_bytes;
5128                         space_info->bytes_reserved += num_bytes;
5129                         if (reserve == RESERVE_ALLOC) {
5130                                 trace_btrfs_space_reservation(cache->fs_info,
5131                                                 "space_info", space_info->flags,
5132                                                 num_bytes, 0);
5133                                 space_info->bytes_may_use -= num_bytes;
5134                         }
5135                 }
5136         } else {
5137                 if (cache->ro)
5138                         space_info->bytes_readonly += num_bytes;
5139                 cache->reserved -= num_bytes;
5140                 space_info->bytes_reserved -= num_bytes;
5141                 space_info->reservation_progress++;
5142         }
5143         spin_unlock(&cache->lock);
5144         spin_unlock(&space_info->lock);
5145         return ret;
5146 }
5147
5148 void btrfs_prepare_extent_commit(struct btrfs_trans_handle *trans,
5149                                 struct btrfs_root *root)
5150 {
5151         struct btrfs_fs_info *fs_info = root->fs_info;
5152         struct btrfs_caching_control *next;
5153         struct btrfs_caching_control *caching_ctl;
5154         struct btrfs_block_group_cache *cache;
5155
5156         down_write(&fs_info->extent_commit_sem);
5157
5158         list_for_each_entry_safe(caching_ctl, next,
5159                                  &fs_info->caching_block_groups, list) {
5160                 cache = caching_ctl->block_group;
5161                 if (block_group_cache_done(cache)) {
5162                         cache->last_byte_to_unpin = (u64)-1;
5163                         list_del_init(&caching_ctl->list);
5164                         put_caching_control(caching_ctl);
5165                 } else {
5166                         cache->last_byte_to_unpin = caching_ctl->progress;
5167                 }
5168         }
5169
5170         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
5171                 fs_info->pinned_extents = &fs_info->freed_extents[1];
5172         else
5173                 fs_info->pinned_extents = &fs_info->freed_extents[0];
5174
5175         up_write(&fs_info->extent_commit_sem);
5176
5177         update_global_block_rsv(fs_info);
5178 }
5179
5180 static int unpin_extent_range(struct btrfs_root *root, u64 start, u64 end)
5181 {
5182         struct btrfs_fs_info *fs_info = root->fs_info;
5183         struct btrfs_block_group_cache *cache = NULL;
5184         struct btrfs_space_info *space_info;
5185         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
5186         u64 len;
5187         bool readonly;
5188
5189         while (start <= end) {
5190                 readonly = false;
5191                 if (!cache ||
5192                     start >= cache->key.objectid + cache->key.offset) {
5193                         if (cache)
5194                                 btrfs_put_block_group(cache);
5195                         cache = btrfs_lookup_block_group(fs_info, start);
5196                         BUG_ON(!cache); /* Logic error */
5197                 }
5198
5199                 len = cache->key.objectid + cache->key.offset - start;
5200                 len = min(len, end + 1 - start);
5201
5202                 if (start < cache->last_byte_to_unpin) {
5203                         len = min(len, cache->last_byte_to_unpin - start);
5204                         btrfs_add_free_space(cache, start, len);
5205                 }
5206
5207                 start += len;
5208                 space_info = cache->space_info;
5209
5210                 spin_lock(&space_info->lock);
5211                 spin_lock(&cache->lock);
5212                 cache->pinned -= len;
5213                 space_info->bytes_pinned -= len;
5214                 if (cache->ro) {
5215                         space_info->bytes_readonly += len;
5216                         readonly = true;
5217                 }
5218                 spin_unlock(&cache->lock);
5219                 if (!readonly && global_rsv->space_info == space_info) {
5220                         spin_lock(&global_rsv->lock);
5221                         if (!global_rsv->full) {
5222                                 len = min(len, global_rsv->size -
5223                                           global_rsv->reserved);
5224                                 global_rsv->reserved += len;
5225                                 space_info->bytes_may_use += len;
5226                                 if (global_rsv->reserved >= global_rsv->size)
5227                                         global_rsv->full = 1;
5228                         }
5229                         spin_unlock(&global_rsv->lock);
5230                 }
5231                 spin_unlock(&space_info->lock);
5232         }
5233
5234         if (cache)
5235                 btrfs_put_block_group(cache);
5236         return 0;
5237 }
5238
5239 int btrfs_finish_extent_commit(struct btrfs_trans_handle *trans,
5240                                struct btrfs_root *root)
5241 {
5242         struct btrfs_fs_info *fs_info = root->fs_info;
5243         struct extent_io_tree *unpin;
5244         u64 start;
5245         u64 end;
5246         int ret;
5247
5248         if (trans->aborted)
5249                 return 0;
5250
5251         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
5252                 unpin = &fs_info->freed_extents[1];
5253         else
5254                 unpin = &fs_info->freed_extents[0];
5255
5256         while (1) {
5257                 ret = find_first_extent_bit(unpin, 0, &start, &end,
5258                                             EXTENT_DIRTY, NULL);
5259                 if (ret)
5260                         break;
5261
5262                 if (btrfs_test_opt(root, DISCARD))
5263                         ret = btrfs_discard_extent(root, start,
5264                                                    end + 1 - start, NULL);
5265
5266                 clear_extent_dirty(unpin, start, end, GFP_NOFS);
5267                 unpin_extent_range(root, start, end);
5268                 cond_resched();
5269         }
5270
5271         return 0;
5272 }
5273
5274 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
5275                                 struct btrfs_root *root,
5276                                 u64 bytenr, u64 num_bytes, u64 parent,
5277                                 u64 root_objectid, u64 owner_objectid,
5278                                 u64 owner_offset, int refs_to_drop,
5279                                 struct btrfs_delayed_extent_op *extent_op)
5280 {
5281         struct btrfs_key key;
5282         struct btrfs_path *path;
5283         struct btrfs_fs_info *info = root->fs_info;
5284         struct btrfs_root *extent_root = info->extent_root;
5285         struct extent_buffer *leaf;
5286         struct btrfs_extent_item *ei;
5287         struct btrfs_extent_inline_ref *iref;
5288         int ret;
5289         int is_data;
5290         int extent_slot = 0;
5291         int found_extent = 0;
5292         int num_to_del = 1;
5293         u32 item_size;
5294         u64 refs;
5295
5296         path = btrfs_alloc_path();
5297         if (!path)
5298                 return -ENOMEM;
5299
5300         path->reada = 1;
5301         path->leave_spinning = 1;
5302
5303         is_data = owner_objectid >= BTRFS_FIRST_FREE_OBJECTID;
5304         BUG_ON(!is_data && refs_to_drop != 1);
5305
5306         ret = lookup_extent_backref(trans, extent_root, path, &iref,
5307                                     bytenr, num_bytes, parent,
5308                                     root_objectid, owner_objectid,
5309                                     owner_offset);
5310         if (ret == 0) {
5311                 extent_slot = path->slots[0];
5312                 while (extent_slot >= 0) {
5313                         btrfs_item_key_to_cpu(path->nodes[0], &key,
5314                                               extent_slot);
5315                         if (key.objectid != bytenr)
5316                                 break;
5317                         if (key.type == BTRFS_EXTENT_ITEM_KEY &&
5318                             key.offset == num_bytes) {
5319                                 found_extent = 1;
5320                                 break;
5321                         }
5322                         if (path->slots[0] - extent_slot > 5)
5323                                 break;
5324                         extent_slot--;
5325                 }
5326 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
5327                 item_size = btrfs_item_size_nr(path->nodes[0], extent_slot);
5328                 if (found_extent && item_size < sizeof(*ei))
5329                         found_extent = 0;
5330 #endif
5331                 if (!found_extent) {
5332                         BUG_ON(iref);
5333                         ret = remove_extent_backref(trans, extent_root, path,
5334                                                     NULL, refs_to_drop,
5335                                                     is_data);
5336                         if (ret) {
5337                                 btrfs_abort_transaction(trans, extent_root, ret);
5338                                 goto out;
5339                         }
5340                         btrfs_release_path(path);
5341                         path->leave_spinning = 1;
5342
5343                         key.objectid = bytenr;
5344                         key.type = BTRFS_EXTENT_ITEM_KEY;
5345                         key.offset = num_bytes;
5346
5347                         ret = btrfs_search_slot(trans, extent_root,
5348                                                 &key, path, -1, 1);
5349                         if (ret) {
5350                                 printk(KERN_ERR "umm, got %d back from search"
5351                                        ", was looking for %llu\n", ret,
5352                                        (unsigned long long)bytenr);
5353                                 if (ret > 0)
5354                                         btrfs_print_leaf(extent_root,
5355                                                          path->nodes[0]);
5356                         }
5357                         if (ret < 0) {
5358                                 btrfs_abort_transaction(trans, extent_root, ret);
5359                                 goto out;
5360                         }
5361                         extent_slot = path->slots[0];
5362                 }
5363         } else if (ret == -ENOENT) {
5364                 btrfs_print_leaf(extent_root, path->nodes[0]);
5365                 WARN_ON(1);
5366                 printk(KERN_ERR "btrfs unable to find ref byte nr %llu "
5367                        "parent %llu root %llu  owner %llu offset %llu\n",
5368                        (unsigned long long)bytenr,
5369                        (unsigned long long)parent,
5370                        (unsigned long long)root_objectid,
5371                        (unsigned long long)owner_objectid,
5372                        (unsigned long long)owner_offset);
5373         } else {
5374                 btrfs_abort_transaction(trans, extent_root, ret);
5375                 goto out;
5376         }
5377
5378         leaf = path->nodes[0];
5379         item_size = btrfs_item_size_nr(leaf, extent_slot);
5380 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
5381         if (item_size < sizeof(*ei)) {
5382                 BUG_ON(found_extent || extent_slot != path->slots[0]);
5383                 ret = convert_extent_item_v0(trans, extent_root, path,
5384                                              owner_objectid, 0);
5385                 if (ret < 0) {
5386                         btrfs_abort_transaction(trans, extent_root, ret);
5387                         goto out;
5388                 }
5389
5390                 btrfs_release_path(path);
5391                 path->leave_spinning = 1;
5392
5393                 key.objectid = bytenr;
5394                 key.type = BTRFS_EXTENT_ITEM_KEY;
5395                 key.offset = num_bytes;
5396
5397                 ret = btrfs_search_slot(trans, extent_root, &key, path,
5398                                         -1, 1);
5399                 if (ret) {
5400                         printk(KERN_ERR "umm, got %d back from search"
5401                                ", was looking for %llu\n", ret,
5402                                (unsigned long long)bytenr);
5403                         btrfs_print_leaf(extent_root, path->nodes[0]);
5404                 }
5405                 if (ret < 0) {
5406                         btrfs_abort_transaction(trans, extent_root, ret);
5407                         goto out;
5408                 }
5409
5410                 extent_slot = path->slots[0];
5411                 leaf = path->nodes[0];
5412                 item_size = btrfs_item_size_nr(leaf, extent_slot);
5413         }
5414 #endif
5415         BUG_ON(item_size < sizeof(*ei));
5416         ei = btrfs_item_ptr(leaf, extent_slot,
5417                             struct btrfs_extent_item);
5418         if (owner_objectid < BTRFS_FIRST_FREE_OBJECTID) {
5419                 struct btrfs_tree_block_info *bi;
5420                 BUG_ON(item_size < sizeof(*ei) + sizeof(*bi));
5421                 bi = (struct btrfs_tree_block_info *)(ei + 1);
5422                 WARN_ON(owner_objectid != btrfs_tree_block_level(leaf, bi));
5423         }
5424
5425         refs = btrfs_extent_refs(leaf, ei);
5426         BUG_ON(refs < refs_to_drop);
5427         refs -= refs_to_drop;
5428
5429         if (refs > 0) {
5430                 if (extent_op)
5431                         __run_delayed_extent_op(extent_op, leaf, ei);
5432                 /*
5433                  * In the case of inline back ref, reference count will
5434                  * be updated by remove_extent_backref
5435                  */
5436                 if (iref) {
5437                         BUG_ON(!found_extent);
5438                 } else {
5439                         btrfs_set_extent_refs(leaf, ei, refs);
5440                         btrfs_mark_buffer_dirty(leaf);
5441                 }
5442                 if (found_extent) {
5443                         ret = remove_extent_backref(trans, extent_root, path,
5444                                                     iref, refs_to_drop,
5445                                                     is_data);
5446                         if (ret) {
5447                                 btrfs_abort_transaction(trans, extent_root, ret);
5448                                 goto out;
5449                         }
5450                 }
5451         } else {
5452                 if (found_extent) {
5453                         BUG_ON(is_data && refs_to_drop !=
5454                                extent_data_ref_count(root, path, iref));
5455                         if (iref) {
5456                                 BUG_ON(path->slots[0] != extent_slot);
5457                         } else {
5458                                 BUG_ON(path->slots[0] != extent_slot + 1);
5459                                 path->slots[0] = extent_slot;
5460                                 num_to_del = 2;
5461                         }
5462                 }
5463
5464                 ret = btrfs_del_items(trans, extent_root, path, path->slots[0],
5465                                       num_to_del);
5466                 if (ret) {
5467                         btrfs_abort_transaction(trans, extent_root, ret);
5468                         goto out;
5469                 }
5470                 btrfs_release_path(path);
5471
5472                 if (is_data) {
5473                         ret = btrfs_del_csums(trans, root, bytenr, num_bytes);
5474                         if (ret) {
5475                                 btrfs_abort_transaction(trans, extent_root, ret);
5476                                 goto out;
5477                         }
5478                 }
5479
5480                 ret = update_block_group(root, bytenr, num_bytes, 0);
5481                 if (ret) {
5482                         btrfs_abort_transaction(trans, extent_root, ret);
5483                         goto out;
5484                 }
5485         }
5486 out:
5487         btrfs_free_path(path);
5488         return ret;
5489 }
5490
5491 /*
5492  * when we free an block, it is possible (and likely) that we free the last
5493  * delayed ref for that extent as well.  This searches the delayed ref tree for
5494  * a given extent, and if there are no other delayed refs to be processed, it
5495  * removes it from the tree.
5496  */
5497 static noinline int check_ref_cleanup(struct btrfs_trans_handle *trans,
5498                                       struct btrfs_root *root, u64 bytenr)
5499 {
5500         struct btrfs_delayed_ref_head *head;
5501         struct btrfs_delayed_ref_root *delayed_refs;
5502         struct btrfs_delayed_ref_node *ref;
5503         struct rb_node *node;
5504         int ret = 0;
5505
5506         delayed_refs = &trans->transaction->delayed_refs;
5507         spin_lock(&delayed_refs->lock);
5508         head = btrfs_find_delayed_ref_head(trans, bytenr);
5509         if (!head)
5510                 goto out;
5511
5512         node = rb_prev(&head->node.rb_node);
5513         if (!node)
5514                 goto out;
5515
5516         ref = rb_entry(node, struct btrfs_delayed_ref_node, rb_node);
5517
5518         /* there are still entries for this ref, we can't drop it */
5519         if (ref->bytenr == bytenr)
5520                 goto out;
5521
5522         if (head->extent_op) {
5523                 if (!head->must_insert_reserved)
5524                         goto out;
5525                 btrfs_free_delayed_extent_op(head->extent_op);
5526                 head->extent_op = NULL;
5527         }
5528
5529         /*
5530          * waiting for the lock here would deadlock.  If someone else has it
5531          * locked they are already in the process of dropping it anyway
5532          */
5533         if (!mutex_trylock(&head->mutex))
5534                 goto out;
5535
5536         /*
5537          * at this point we have a head with no other entries.  Go
5538          * ahead and process it.
5539          */
5540         head->node.in_tree = 0;
5541         rb_erase(&head->node.rb_node, &delayed_refs->root);
5542
5543         delayed_refs->num_entries--;
5544
5545         /*
5546          * we don't take a ref on the node because we're removing it from the
5547          * tree, so we just steal the ref the tree was holding.
5548          */
5549         delayed_refs->num_heads--;
5550         if (list_empty(&head->cluster))
5551                 delayed_refs->num_heads_ready--;
5552
5553         list_del_init(&head->cluster);
5554         spin_unlock(&delayed_refs->lock);
5555
5556         BUG_ON(head->extent_op);
5557         if (head->must_insert_reserved)
5558                 ret = 1;
5559
5560         mutex_unlock(&head->mutex);
5561         btrfs_put_delayed_ref(&head->node);
5562         return ret;
5563 out:
5564         spin_unlock(&delayed_refs->lock);
5565         return 0;
5566 }
5567
5568 void btrfs_free_tree_block(struct btrfs_trans_handle *trans,
5569                            struct btrfs_root *root,
5570                            struct extent_buffer *buf,
5571                            u64 parent, int last_ref)
5572 {
5573         struct btrfs_block_group_cache *cache = NULL;
5574         int ret;
5575
5576         if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
5577                 ret = btrfs_add_delayed_tree_ref(root->fs_info, trans,
5578                                         buf->start, buf->len,
5579                                         parent, root->root_key.objectid,
5580                                         btrfs_header_level(buf),
5581                                         BTRFS_DROP_DELAYED_REF, NULL, 0);
5582                 BUG_ON(ret); /* -ENOMEM */
5583         }
5584
5585         if (!last_ref)
5586                 return;
5587
5588         cache = btrfs_lookup_block_group(root->fs_info, buf->start);
5589
5590         if (btrfs_header_generation(buf) == trans->transid) {
5591                 if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
5592                         ret = check_ref_cleanup(trans, root, buf->start);
5593                         if (!ret)
5594                                 goto out;
5595                 }
5596
5597                 if (btrfs_header_flag(buf, BTRFS_HEADER_FLAG_WRITTEN)) {
5598                         pin_down_extent(root, cache, buf->start, buf->len, 1);
5599                         goto out;
5600                 }
5601
5602                 WARN_ON(test_bit(EXTENT_BUFFER_DIRTY, &buf->bflags));
5603
5604                 btrfs_add_free_space(cache, buf->start, buf->len);
5605                 btrfs_update_reserved_bytes(cache, buf->len, RESERVE_FREE);
5606         }
5607 out:
5608         /*
5609          * Deleting the buffer, clear the corrupt flag since it doesn't matter
5610          * anymore.
5611          */
5612         clear_bit(EXTENT_BUFFER_CORRUPT, &buf->bflags);
5613         btrfs_put_block_group(cache);
5614 }
5615
5616 /* Can return -ENOMEM */
5617 int btrfs_free_extent(struct btrfs_trans_handle *trans, struct btrfs_root *root,
5618                       u64 bytenr, u64 num_bytes, u64 parent, u64 root_objectid,
5619                       u64 owner, u64 offset, int for_cow)
5620 {
5621         int ret;
5622         struct btrfs_fs_info *fs_info = root->fs_info;
5623
5624         /*
5625          * tree log blocks never actually go into the extent allocation
5626          * tree, just update pinning info and exit early.
5627          */
5628         if (root_objectid == BTRFS_TREE_LOG_OBJECTID) {
5629                 WARN_ON(owner >= BTRFS_FIRST_FREE_OBJECTID);
5630                 /* unlocks the pinned mutex */
5631                 btrfs_pin_extent(root, bytenr, num_bytes, 1);
5632                 ret = 0;
5633         } else if (owner < BTRFS_FIRST_FREE_OBJECTID) {
5634                 ret = btrfs_add_delayed_tree_ref(fs_info, trans, bytenr,
5635                                         num_bytes,
5636                                         parent, root_objectid, (int)owner,
5637                                         BTRFS_DROP_DELAYED_REF, NULL, for_cow);
5638         } else {
5639                 ret = btrfs_add_delayed_data_ref(fs_info, trans, bytenr,
5640                                                 num_bytes,
5641                                                 parent, root_objectid, owner,
5642                                                 offset, BTRFS_DROP_DELAYED_REF,
5643                                                 NULL, for_cow);
5644         }
5645         return ret;
5646 }
5647
5648 static u64 stripe_align(struct btrfs_root *root,
5649                         struct btrfs_block_group_cache *cache,
5650                         u64 val, u64 num_bytes)
5651 {
5652         u64 ret = ALIGN(val, root->stripesize);
5653         return ret;
5654 }
5655
5656 /*
5657  * when we wait for progress in the block group caching, its because
5658  * our allocation attempt failed at least once.  So, we must sleep
5659  * and let some progress happen before we try again.
5660  *
5661  * This function will sleep at least once waiting for new free space to
5662  * show up, and then it will check the block group free space numbers
5663  * for our min num_bytes.  Another option is to have it go ahead
5664  * and look in the rbtree for a free extent of a given size, but this
5665  * is a good start.
5666  */
5667 static noinline int
5668 wait_block_group_cache_progress(struct btrfs_block_group_cache *cache,
5669                                 u64 num_bytes)
5670 {
5671         struct btrfs_caching_control *caching_ctl;
5672
5673         caching_ctl = get_caching_control(cache);
5674         if (!caching_ctl)
5675                 return 0;
5676
5677         wait_event(caching_ctl->wait, block_group_cache_done(cache) ||
5678                    (cache->free_space_ctl->free_space >= num_bytes));
5679
5680         put_caching_control(caching_ctl);
5681         return 0;
5682 }
5683
5684 static noinline int
5685 wait_block_group_cache_done(struct btrfs_block_group_cache *cache)
5686 {
5687         struct btrfs_caching_control *caching_ctl;
5688
5689         caching_ctl = get_caching_control(cache);
5690         if (!caching_ctl)
5691                 return 0;
5692
5693         wait_event(caching_ctl->wait, block_group_cache_done(cache));
5694
5695         put_caching_control(caching_ctl);
5696         return 0;
5697 }
5698
5699 int __get_raid_index(u64 flags)
5700 {
5701         if (flags & BTRFS_BLOCK_GROUP_RAID10)
5702                 return BTRFS_RAID_RAID10;
5703         else if (flags & BTRFS_BLOCK_GROUP_RAID1)
5704                 return BTRFS_RAID_RAID1;
5705         else if (flags & BTRFS_BLOCK_GROUP_DUP)
5706                 return BTRFS_RAID_DUP;
5707         else if (flags & BTRFS_BLOCK_GROUP_RAID0)
5708                 return BTRFS_RAID_RAID0;
5709         else if (flags & BTRFS_BLOCK_GROUP_RAID5)
5710                 return BTRFS_RAID_RAID5;
5711         else if (flags & BTRFS_BLOCK_GROUP_RAID6)
5712                 return BTRFS_RAID_RAID6;
5713
5714         return BTRFS_RAID_SINGLE; /* BTRFS_BLOCK_GROUP_SINGLE */
5715 }
5716
5717 static int get_block_group_index(struct btrfs_block_group_cache *cache)
5718 {
5719         return __get_raid_index(cache->flags);
5720 }
5721
5722 enum btrfs_loop_type {
5723         LOOP_CACHING_NOWAIT = 0,
5724         LOOP_CACHING_WAIT = 1,
5725         LOOP_ALLOC_CHUNK = 2,
5726         LOOP_NO_EMPTY_SIZE = 3,
5727 };
5728
5729 /*
5730  * walks the btree of allocated extents and find a hole of a given size.
5731  * The key ins is changed to record the hole:
5732  * ins->objectid == block start
5733  * ins->flags = BTRFS_EXTENT_ITEM_KEY
5734  * ins->offset == number of blocks
5735  * Any available blocks before search_start are skipped.
5736  */
5737 static noinline int find_free_extent(struct btrfs_trans_handle *trans,
5738                                      struct btrfs_root *orig_root,
5739                                      u64 num_bytes, u64 empty_size,
5740                                      u64 hint_byte, struct btrfs_key *ins,
5741                                      u64 data)
5742 {
5743         int ret = 0;
5744         struct btrfs_root *root = orig_root->fs_info->extent_root;
5745         struct btrfs_free_cluster *last_ptr = NULL;
5746         struct btrfs_block_group_cache *block_group = NULL;
5747         struct btrfs_block_group_cache *used_block_group;
5748         u64 search_start = 0;
5749         int empty_cluster = 2 * 1024 * 1024;
5750         struct btrfs_space_info *space_info;
5751         int loop = 0;
5752         int index = __get_raid_index(data);
5753         int alloc_type = (data & BTRFS_BLOCK_GROUP_DATA) ?
5754                 RESERVE_ALLOC_NO_ACCOUNT : RESERVE_ALLOC;
5755         bool found_uncached_bg = false;
5756         bool failed_cluster_refill = false;
5757         bool failed_alloc = false;
5758         bool use_cluster = true;
5759         bool have_caching_bg = false;
5760
5761         WARN_ON(num_bytes < root->sectorsize);
5762         btrfs_set_key_type(ins, BTRFS_EXTENT_ITEM_KEY);
5763         ins->objectid = 0;
5764         ins->offset = 0;
5765
5766         trace_find_free_extent(orig_root, num_bytes, empty_size, data);
5767
5768         space_info = __find_space_info(root->fs_info, data);
5769         if (!space_info) {
5770                 printk(KERN_ERR "No space info for %llu\n", data);
5771                 return -ENOSPC;
5772         }
5773
5774         /*
5775          * If the space info is for both data and metadata it means we have a
5776          * small filesystem and we can't use the clustering stuff.
5777          */
5778         if (btrfs_mixed_space_info(space_info))
5779                 use_cluster = false;
5780
5781         if (data & BTRFS_BLOCK_GROUP_METADATA && use_cluster) {
5782                 last_ptr = &root->fs_info->meta_alloc_cluster;
5783                 if (!btrfs_test_opt(root, SSD))
5784                         empty_cluster = 64 * 1024;
5785         }
5786
5787         if ((data & BTRFS_BLOCK_GROUP_DATA) && use_cluster &&
5788             btrfs_test_opt(root, SSD)) {
5789                 last_ptr = &root->fs_info->data_alloc_cluster;
5790         }
5791
5792         if (last_ptr) {
5793                 spin_lock(&last_ptr->lock);
5794                 if (last_ptr->block_group)
5795                         hint_byte = last_ptr->window_start;
5796                 spin_unlock(&last_ptr->lock);
5797         }
5798
5799         search_start = max(search_start, first_logical_byte(root, 0));
5800         search_start = max(search_start, hint_byte);
5801
5802         if (!last_ptr)
5803                 empty_cluster = 0;
5804
5805         if (search_start == hint_byte) {
5806                 block_group = btrfs_lookup_block_group(root->fs_info,
5807                                                        search_start);
5808                 used_block_group = block_group;
5809                 /*
5810                  * we don't want to use the block group if it doesn't match our
5811                  * allocation bits, or if its not cached.
5812                  *
5813                  * However if we are re-searching with an ideal block group
5814                  * picked out then we don't care that the block group is cached.
5815                  */
5816                 if (block_group && block_group_bits(block_group, data) &&
5817                     block_group->cached != BTRFS_CACHE_NO) {
5818                         down_read(&space_info->groups_sem);
5819                         if (list_empty(&block_group->list) ||
5820                             block_group->ro) {
5821                                 /*
5822                                  * someone is removing this block group,
5823                                  * we can't jump into the have_block_group
5824                                  * target because our list pointers are not
5825                                  * valid
5826                                  */
5827                                 btrfs_put_block_group(block_group);
5828                                 up_read(&space_info->groups_sem);
5829                         } else {
5830                                 index = get_block_group_index(block_group);
5831                                 goto have_block_group;
5832                         }
5833                 } else if (block_group) {
5834                         btrfs_put_block_group(block_group);
5835                 }
5836         }
5837 search:
5838         have_caching_bg = false;
5839         down_read(&space_info->groups_sem);
5840         list_for_each_entry(block_group, &space_info->block_groups[index],
5841                             list) {
5842                 u64 offset;
5843                 int cached;
5844
5845                 used_block_group = block_group;
5846                 btrfs_get_block_group(block_group);
5847                 search_start = block_group->key.objectid;
5848
5849                 /*
5850                  * this can happen if we end up cycling through all the
5851                  * raid types, but we want to make sure we only allocate
5852                  * for the proper type.
5853                  */
5854                 if (!block_group_bits(block_group, data)) {
5855                     u64 extra = BTRFS_BLOCK_GROUP_DUP |
5856                                 BTRFS_BLOCK_GROUP_RAID1 |
5857                                 BTRFS_BLOCK_GROUP_RAID5 |
5858                                 BTRFS_BLOCK_GROUP_RAID6 |
5859                                 BTRFS_BLOCK_GROUP_RAID10;
5860
5861                         /*
5862                          * if they asked for extra copies and this block group
5863                          * doesn't provide them, bail.  This does allow us to
5864                          * fill raid0 from raid1.
5865                          */
5866                         if ((data & extra) && !(block_group->flags & extra))
5867                                 goto loop;
5868                 }
5869
5870 have_block_group:
5871                 cached = block_group_cache_done(block_group);
5872                 if (unlikely(!cached)) {
5873                         found_uncached_bg = true;
5874                         ret = cache_block_group(block_group, 0);
5875                         BUG_ON(ret < 0);
5876                         ret = 0;
5877                 }
5878
5879                 if (unlikely(block_group->ro))
5880                         goto loop;
5881
5882                 /*
5883                  * Ok we want to try and use the cluster allocator, so
5884                  * lets look there
5885                  */
5886                 if (last_ptr) {
5887                         unsigned long aligned_cluster;
5888                         /*
5889                          * the refill lock keeps out other
5890                          * people trying to start a new cluster
5891                          */
5892                         spin_lock(&last_ptr->refill_lock);
5893                         used_block_group = last_ptr->block_group;
5894                         if (used_block_group != block_group &&
5895                             (!used_block_group ||
5896                              used_block_group->ro ||
5897                              !block_group_bits(used_block_group, data))) {
5898                                 used_block_group = block_group;
5899                                 goto refill_cluster;
5900                         }
5901
5902                         if (used_block_group != block_group)
5903                                 btrfs_get_block_group(used_block_group);
5904
5905                         offset = btrfs_alloc_from_cluster(used_block_group,
5906                           last_ptr, num_bytes, used_block_group->key.objectid);
5907                         if (offset) {
5908                                 /* we have a block, we're done */
5909                                 spin_unlock(&last_ptr->refill_lock);
5910                                 trace_btrfs_reserve_extent_cluster(root,
5911                                         block_group, search_start, num_bytes);
5912                                 goto checks;
5913                         }
5914
5915                         WARN_ON(last_ptr->block_group != used_block_group);
5916                         if (used_block_group != block_group) {
5917                                 btrfs_put_block_group(used_block_group);
5918                                 used_block_group = block_group;
5919                         }
5920 refill_cluster:
5921                         BUG_ON(used_block_group != block_group);
5922                         /* If we are on LOOP_NO_EMPTY_SIZE, we can't
5923                          * set up a new clusters, so lets just skip it
5924                          * and let the allocator find whatever block
5925                          * it can find.  If we reach this point, we
5926                          * will have tried the cluster allocator
5927                          * plenty of times and not have found
5928                          * anything, so we are likely way too
5929                          * fragmented for the clustering stuff to find
5930                          * anything.
5931                          *
5932                          * However, if the cluster is taken from the
5933                          * current block group, release the cluster
5934                          * first, so that we stand a better chance of
5935                          * succeeding in the unclustered
5936                          * allocation.  */
5937                         if (loop >= LOOP_NO_EMPTY_SIZE &&
5938                             last_ptr->block_group != block_group) {
5939                                 spin_unlock(&last_ptr->refill_lock);
5940                                 goto unclustered_alloc;
5941                         }
5942
5943                         /*
5944                          * this cluster didn't work out, free it and
5945                          * start over
5946                          */
5947                         btrfs_return_cluster_to_free_space(NULL, last_ptr);
5948
5949                         if (loop >= LOOP_NO_EMPTY_SIZE) {
5950                                 spin_unlock(&last_ptr->refill_lock);
5951                                 goto unclustered_alloc;
5952                         }
5953
5954                         aligned_cluster = max_t(unsigned long,
5955                                                 empty_cluster + empty_size,
5956                                               block_group->full_stripe_len);
5957
5958                         /* allocate a cluster in this block group */
5959                         ret = btrfs_find_space_cluster(trans, root,
5960                                                block_group, last_ptr,
5961                                                search_start, num_bytes,
5962                                                aligned_cluster);
5963                         if (ret == 0) {
5964                                 /*
5965                                  * now pull our allocation out of this
5966                                  * cluster
5967                                  */
5968                                 offset = btrfs_alloc_from_cluster(block_group,
5969                                                   last_ptr, num_bytes,
5970                                                   search_start);
5971                                 if (offset) {
5972                                         /* we found one, proceed */
5973                                         spin_unlock(&last_ptr->refill_lock);
5974                                         trace_btrfs_reserve_extent_cluster(root,
5975                                                 block_group, search_start,
5976                                                 num_bytes);
5977                                         goto checks;
5978                                 }
5979                         } else if (!cached && loop > LOOP_CACHING_NOWAIT
5980                                    && !failed_cluster_refill) {
5981                                 spin_unlock(&last_ptr->refill_lock);
5982
5983                                 failed_cluster_refill = true;
5984                                 wait_block_group_cache_progress(block_group,
5985                                        num_bytes + empty_cluster + empty_size);
5986                                 goto have_block_group;
5987                         }
5988
5989                         /*
5990                          * at this point we either didn't find a cluster
5991                          * or we weren't able to allocate a block from our
5992                          * cluster.  Free the cluster we've been trying
5993                          * to use, and go to the next block group
5994                          */
5995                         btrfs_return_cluster_to_free_space(NULL, last_ptr);
5996                         spin_unlock(&last_ptr->refill_lock);
5997                         goto loop;
5998                 }
5999
6000 unclustered_alloc:
6001                 spin_lock(&block_group->free_space_ctl->tree_lock);
6002                 if (cached &&
6003                     block_group->free_space_ctl->free_space <
6004                     num_bytes + empty_cluster + empty_size) {
6005                         spin_unlock(&block_group->free_space_ctl->tree_lock);
6006                         goto loop;
6007                 }
6008                 spin_unlock(&block_group->free_space_ctl->tree_lock);
6009
6010                 offset = btrfs_find_space_for_alloc(block_group, search_start,
6011                                                     num_bytes, empty_size);
6012                 /*
6013                  * If we didn't find a chunk, and we haven't failed on this
6014                  * block group before, and this block group is in the middle of
6015                  * caching and we are ok with waiting, then go ahead and wait
6016                  * for progress to be made, and set failed_alloc to true.
6017                  *
6018                  * If failed_alloc is true then we've already waited on this
6019                  * block group once and should move on to the next block group.
6020                  */
6021                 if (!offset && !failed_alloc && !cached &&
6022                     loop > LOOP_CACHING_NOWAIT) {
6023                         wait_block_group_cache_progress(block_group,
6024                                                 num_bytes + empty_size);
6025                         failed_alloc = true;
6026                         goto have_block_group;
6027                 } else if (!offset) {
6028                         if (!cached)
6029                                 have_caching_bg = true;
6030                         goto loop;
6031                 }
6032 checks:
6033                 search_start = stripe_align(root, used_block_group,
6034                                             offset, num_bytes);
6035
6036                 /* move on to the next group */
6037                 if (search_start + num_bytes >
6038                     used_block_group->key.objectid + used_block_group->key.offset) {
6039                         btrfs_add_free_space(used_block_group, offset, num_bytes);
6040                         goto loop;
6041                 }
6042
6043                 if (offset < search_start)
6044                         btrfs_add_free_space(used_block_group, offset,
6045                                              search_start - offset);
6046                 BUG_ON(offset > search_start);
6047
6048                 ret = btrfs_update_reserved_bytes(used_block_group, num_bytes,
6049                                                   alloc_type);
6050                 if (ret == -EAGAIN) {
6051                         btrfs_add_free_space(used_block_group, offset, num_bytes);
6052                         goto loop;
6053                 }
6054
6055                 /* we are all good, lets return */
6056                 ins->objectid = search_start;
6057                 ins->offset = num_bytes;
6058
6059                 trace_btrfs_reserve_extent(orig_root, block_group,
6060                                            search_start, num_bytes);
6061                 if (used_block_group != block_group)
6062                         btrfs_put_block_group(used_block_group);
6063                 btrfs_put_block_group(block_group);
6064                 break;
6065 loop:
6066                 failed_cluster_refill = false;
6067                 failed_alloc = false;
6068                 BUG_ON(index != get_block_group_index(block_group));
6069                 if (used_block_group != block_group)
6070                         btrfs_put_block_group(used_block_group);
6071                 btrfs_put_block_group(block_group);
6072         }
6073         up_read(&space_info->groups_sem);
6074
6075         if (!ins->objectid && loop >= LOOP_CACHING_WAIT && have_caching_bg)
6076                 goto search;
6077
6078         if (!ins->objectid && ++index < BTRFS_NR_RAID_TYPES)
6079                 goto search;
6080
6081         /*
6082          * LOOP_CACHING_NOWAIT, search partially cached block groups, kicking
6083          *                      caching kthreads as we move along
6084          * LOOP_CACHING_WAIT, search everything, and wait if our bg is caching
6085          * LOOP_ALLOC_CHUNK, force a chunk allocation and try again
6086          * LOOP_NO_EMPTY_SIZE, set empty_size and empty_cluster to 0 and try
6087          *                      again
6088          */
6089         if (!ins->objectid && loop < LOOP_NO_EMPTY_SIZE) {
6090                 index = 0;
6091                 loop++;
6092                 if (loop == LOOP_ALLOC_CHUNK) {
6093                         ret = do_chunk_alloc(trans, root, data,
6094                                              CHUNK_ALLOC_FORCE);
6095                         /*
6096                          * Do not bail out on ENOSPC since we
6097                          * can do more things.
6098                          */
6099                         if (ret < 0 && ret != -ENOSPC) {
6100                                 btrfs_abort_transaction(trans,
6101                                                         root, ret);
6102                                 goto out;
6103                         }
6104                 }
6105
6106                 if (loop == LOOP_NO_EMPTY_SIZE) {
6107                         empty_size = 0;
6108                         empty_cluster = 0;
6109                 }
6110
6111                 goto search;
6112         } else if (!ins->objectid) {
6113                 ret = -ENOSPC;
6114         } else if (ins->objectid) {
6115                 ret = 0;
6116         }
6117 out:
6118
6119         return ret;
6120 }
6121
6122 static void dump_space_info(struct btrfs_space_info *info, u64 bytes,
6123                             int dump_block_groups)
6124 {
6125         struct btrfs_block_group_cache *cache;
6126         int index = 0;
6127
6128         spin_lock(&info->lock);
6129         printk(KERN_INFO "space_info %llu has %llu free, is %sfull\n",
6130                (unsigned long long)info->flags,
6131                (unsigned long long)(info->total_bytes - info->bytes_used -
6132                                     info->bytes_pinned - info->bytes_reserved -
6133                                     info->bytes_readonly),
6134                (info->full) ? "" : "not ");
6135         printk(KERN_INFO "space_info total=%llu, used=%llu, pinned=%llu, "
6136                "reserved=%llu, may_use=%llu, readonly=%llu\n",
6137                (unsigned long long)info->total_bytes,
6138                (unsigned long long)info->bytes_used,
6139                (unsigned long long)info->bytes_pinned,
6140                (unsigned long long)info->bytes_reserved,
6141                (unsigned long long)info->bytes_may_use,
6142                (unsigned long long)info->bytes_readonly);
6143         spin_unlock(&info->lock);
6144
6145         if (!dump_block_groups)
6146                 return;
6147
6148         down_read(&info->groups_sem);
6149 again:
6150         list_for_each_entry(cache, &info->block_groups[index], list) {
6151                 spin_lock(&cache->lock);
6152                 printk(KERN_INFO "block group %llu has %llu bytes, %llu used %llu pinned %llu reserved %s\n",
6153                        (unsigned long long)cache->key.objectid,
6154                        (unsigned long long)cache->key.offset,
6155                        (unsigned long long)btrfs_block_group_used(&cache->item),
6156                        (unsigned long long)cache->pinned,
6157                        (unsigned long long)cache->reserved,
6158                        cache->ro ? "[readonly]" : "");
6159                 btrfs_dump_free_space(cache, bytes);
6160                 spin_unlock(&cache->lock);
6161         }
6162         if (++index < BTRFS_NR_RAID_TYPES)
6163                 goto again;
6164         up_read(&info->groups_sem);
6165 }
6166
6167 int btrfs_reserve_extent(struct btrfs_trans_handle *trans,
6168                          struct btrfs_root *root,
6169                          u64 num_bytes, u64 min_alloc_size,
6170                          u64 empty_size, u64 hint_byte,
6171                          struct btrfs_key *ins, u64 data)
6172 {
6173         bool final_tried = false;
6174         int ret;
6175
6176         data = btrfs_get_alloc_profile(root, data);
6177 again:
6178         WARN_ON(num_bytes < root->sectorsize);
6179         ret = find_free_extent(trans, root, num_bytes, empty_size,
6180                                hint_byte, ins, data);
6181
6182         if (ret == -ENOSPC) {
6183                 if (!final_tried) {
6184                         num_bytes = num_bytes >> 1;
6185                         num_bytes = round_down(num_bytes, root->sectorsize);
6186                         num_bytes = max(num_bytes, min_alloc_size);
6187                         if (num_bytes == min_alloc_size)
6188                                 final_tried = true;
6189                         goto again;
6190                 } else if (btrfs_test_opt(root, ENOSPC_DEBUG)) {
6191                         struct btrfs_space_info *sinfo;
6192
6193                         sinfo = __find_space_info(root->fs_info, data);
6194                         printk(KERN_ERR "btrfs allocation failed flags %llu, "
6195                                "wanted %llu\n", (unsigned long long)data,
6196                                (unsigned long long)num_bytes);
6197                         if (sinfo)
6198                                 dump_space_info(sinfo, num_bytes, 1);
6199                 }
6200         }
6201
6202         trace_btrfs_reserved_extent_alloc(root, ins->objectid, ins->offset);
6203
6204         return ret;
6205 }
6206
6207 static int __btrfs_free_reserved_extent(struct btrfs_root *root,
6208                                         u64 start, u64 len, int pin)
6209 {
6210         struct btrfs_block_group_cache *cache;
6211         int ret = 0;
6212
6213         cache = btrfs_lookup_block_group(root->fs_info, start);
6214         if (!cache) {
6215                 printk(KERN_ERR "Unable to find block group for %llu\n",
6216                        (unsigned long long)start);
6217                 return -ENOSPC;
6218         }
6219
6220         if (btrfs_test_opt(root, DISCARD))
6221                 ret = btrfs_discard_extent(root, start, len, NULL);
6222
6223         if (pin)
6224                 pin_down_extent(root, cache, start, len, 1);
6225         else {
6226                 btrfs_add_free_space(cache, start, len);
6227                 btrfs_update_reserved_bytes(cache, len, RESERVE_FREE);
6228         }
6229         btrfs_put_block_group(cache);
6230
6231         trace_btrfs_reserved_extent_free(root, start, len);
6232
6233         return ret;
6234 }
6235
6236 int btrfs_free_reserved_extent(struct btrfs_root *root,
6237                                         u64 start, u64 len)
6238 {
6239         return __btrfs_free_reserved_extent(root, start, len, 0);
6240 }
6241
6242 int btrfs_free_and_pin_reserved_extent(struct btrfs_root *root,
6243                                        u64 start, u64 len)
6244 {
6245         return __btrfs_free_reserved_extent(root, start, len, 1);
6246 }
6247
6248 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
6249                                       struct btrfs_root *root,
6250                                       u64 parent, u64 root_objectid,
6251                                       u64 flags, u64 owner, u64 offset,
6252                                       struct btrfs_key *ins, int ref_mod)
6253 {
6254         int ret;
6255         struct btrfs_fs_info *fs_info = root->fs_info;
6256         struct btrfs_extent_item *extent_item;
6257         struct btrfs_extent_inline_ref *iref;
6258         struct btrfs_path *path;
6259         struct extent_buffer *leaf;
6260         int type;
6261         u32 size;
6262
6263         if (parent > 0)
6264                 type = BTRFS_SHARED_DATA_REF_KEY;
6265         else
6266                 type = BTRFS_EXTENT_DATA_REF_KEY;
6267
6268         size = sizeof(*extent_item) + btrfs_extent_inline_ref_size(type);
6269
6270         path = btrfs_alloc_path();
6271         if (!path)
6272                 return -ENOMEM;
6273
6274         path->leave_spinning = 1;
6275         ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path,
6276                                       ins, size);
6277         if (ret) {
6278                 btrfs_free_path(path);
6279                 return ret;
6280         }
6281
6282         leaf = path->nodes[0];
6283         extent_item = btrfs_item_ptr(leaf, path->slots[0],
6284                                      struct btrfs_extent_item);
6285         btrfs_set_extent_refs(leaf, extent_item, ref_mod);
6286         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
6287         btrfs_set_extent_flags(leaf, extent_item,
6288                                flags | BTRFS_EXTENT_FLAG_DATA);
6289
6290         iref = (struct btrfs_extent_inline_ref *)(extent_item + 1);
6291         btrfs_set_extent_inline_ref_type(leaf, iref, type);
6292         if (parent > 0) {
6293                 struct btrfs_shared_data_ref *ref;
6294                 ref = (struct btrfs_shared_data_ref *)(iref + 1);
6295                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
6296                 btrfs_set_shared_data_ref_count(leaf, ref, ref_mod);
6297         } else {
6298                 struct btrfs_extent_data_ref *ref;
6299                 ref = (struct btrfs_extent_data_ref *)(&iref->offset);
6300                 btrfs_set_extent_data_ref_root(leaf, ref, root_objectid);
6301                 btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
6302                 btrfs_set_extent_data_ref_offset(leaf, ref, offset);
6303                 btrfs_set_extent_data_ref_count(leaf, ref, ref_mod);
6304         }
6305
6306         btrfs_mark_buffer_dirty(path->nodes[0]);
6307         btrfs_free_path(path);
6308
6309         ret = update_block_group(root, ins->objectid, ins->offset, 1);
6310         if (ret) { /* -ENOENT, logic error */
6311                 printk(KERN_ERR "btrfs update block group failed for %llu "
6312                        "%llu\n", (unsigned long long)ins->objectid,
6313                        (unsigned long long)ins->offset);
6314                 BUG();
6315         }
6316         return ret;
6317 }
6318
6319 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
6320                                      struct btrfs_root *root,
6321                                      u64 parent, u64 root_objectid,
6322                                      u64 flags, struct btrfs_disk_key *key,
6323                                      int level, struct btrfs_key *ins)
6324 {
6325         int ret;
6326         struct btrfs_fs_info *fs_info = root->fs_info;
6327         struct btrfs_extent_item *extent_item;
6328         struct btrfs_tree_block_info *block_info;
6329         struct btrfs_extent_inline_ref *iref;
6330         struct btrfs_path *path;
6331         struct extent_buffer *leaf;
6332         u32 size = sizeof(*extent_item) + sizeof(*block_info) + sizeof(*iref);
6333
6334         path = btrfs_alloc_path();
6335         if (!path)
6336                 return -ENOMEM;
6337
6338         path->leave_spinning = 1;
6339         ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path,
6340                                       ins, size);
6341         if (ret) {
6342                 btrfs_free_path(path);
6343                 return ret;
6344         }
6345
6346         leaf = path->nodes[0];
6347         extent_item = btrfs_item_ptr(leaf, path->slots[0],
6348                                      struct btrfs_extent_item);
6349         btrfs_set_extent_refs(leaf, extent_item, 1);
6350         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
6351         btrfs_set_extent_flags(leaf, extent_item,
6352                                flags | BTRFS_EXTENT_FLAG_TREE_BLOCK);
6353         block_info = (struct btrfs_tree_block_info *)(extent_item + 1);
6354
6355         btrfs_set_tree_block_key(leaf, block_info, key);
6356         btrfs_set_tree_block_level(leaf, block_info, level);
6357
6358         iref = (struct btrfs_extent_inline_ref *)(block_info + 1);
6359         if (parent > 0) {
6360                 BUG_ON(!(flags & BTRFS_BLOCK_FLAG_FULL_BACKREF));
6361                 btrfs_set_extent_inline_ref_type(leaf, iref,
6362                                                  BTRFS_SHARED_BLOCK_REF_KEY);
6363                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
6364         } else {
6365                 btrfs_set_extent_inline_ref_type(leaf, iref,
6366                                                  BTRFS_TREE_BLOCK_REF_KEY);
6367                 btrfs_set_extent_inline_ref_offset(leaf, iref, root_objectid);
6368         }
6369
6370         btrfs_mark_buffer_dirty(leaf);
6371         btrfs_free_path(path);
6372
6373         ret = update_block_group(root, ins->objectid, ins->offset, 1);
6374         if (ret) { /* -ENOENT, logic error */
6375                 printk(KERN_ERR "btrfs update block group failed for %llu "
6376                        "%llu\n", (unsigned long long)ins->objectid,
6377                        (unsigned long long)ins->offset);
6378                 BUG();
6379         }
6380         return ret;
6381 }
6382
6383 int btrfs_alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
6384                                      struct btrfs_root *root,
6385                                      u64 root_objectid, u64 owner,
6386                                      u64 offset, struct btrfs_key *ins)
6387 {
6388         int ret;
6389
6390         BUG_ON(root_objectid == BTRFS_TREE_LOG_OBJECTID);
6391
6392         ret = btrfs_add_delayed_data_ref(root->fs_info, trans, ins->objectid,
6393                                          ins->offset, 0,
6394                                          root_objectid, owner, offset,
6395                                          BTRFS_ADD_DELAYED_EXTENT, NULL, 0);
6396         return ret;
6397 }
6398
6399 /*
6400  * this is used by the tree logging recovery code.  It records that
6401  * an extent has been allocated and makes sure to clear the free
6402  * space cache bits as well
6403  */
6404 int btrfs_alloc_logged_file_extent(struct btrfs_trans_handle *trans,
6405                                    struct btrfs_root *root,
6406                                    u64 root_objectid, u64 owner, u64 offset,
6407                                    struct btrfs_key *ins)
6408 {
6409         int ret;
6410         struct btrfs_block_group_cache *block_group;
6411         struct btrfs_caching_control *caching_ctl;
6412         u64 start = ins->objectid;
6413         u64 num_bytes = ins->offset;
6414
6415         block_group = btrfs_lookup_block_group(root->fs_info, ins->objectid);
6416         cache_block_group(block_group, 0);
6417         caching_ctl = get_caching_control(block_group);
6418
6419         if (!caching_ctl) {
6420                 BUG_ON(!block_group_cache_done(block_group));
6421                 ret = btrfs_remove_free_space(block_group, start, num_bytes);
6422                 BUG_ON(ret); /* -ENOMEM */
6423         } else {
6424                 mutex_lock(&caching_ctl->mutex);
6425
6426                 if (start >= caching_ctl->progress) {
6427                         ret = add_excluded_extent(root, start, num_bytes);
6428                         BUG_ON(ret); /* -ENOMEM */
6429                 } else if (start + num_bytes <= caching_ctl->progress) {
6430                         ret = btrfs_remove_free_space(block_group,
6431                                                       start, num_bytes);
6432                         BUG_ON(ret); /* -ENOMEM */
6433                 } else {
6434                         num_bytes = caching_ctl->progress - start;
6435                         ret = btrfs_remove_free_space(block_group,
6436                                                       start, num_bytes);
6437                         BUG_ON(ret); /* -ENOMEM */
6438
6439                         start = caching_ctl->progress;
6440                         num_bytes = ins->objectid + ins->offset -
6441                                     caching_ctl->progress;
6442                         ret = add_excluded_extent(root, start, num_bytes);
6443                         BUG_ON(ret); /* -ENOMEM */
6444                 }
6445
6446                 mutex_unlock(&caching_ctl->mutex);
6447                 put_caching_control(caching_ctl);
6448         }
6449
6450         ret = btrfs_update_reserved_bytes(block_group, ins->offset,
6451                                           RESERVE_ALLOC_NO_ACCOUNT);
6452         BUG_ON(ret); /* logic error */
6453         btrfs_put_block_group(block_group);
6454         ret = alloc_reserved_file_extent(trans, root, 0, root_objectid,
6455                                          0, owner, offset, ins, 1);
6456         return ret;
6457 }
6458
6459 struct extent_buffer *btrfs_init_new_buffer(struct btrfs_trans_handle *trans,
6460                                             struct btrfs_root *root,
6461                                             u64 bytenr, u32 blocksize,
6462                                             int level)
6463 {
6464         struct extent_buffer *buf;
6465
6466         buf = btrfs_find_create_tree_block(root, bytenr, blocksize);
6467         if (!buf)
6468                 return ERR_PTR(-ENOMEM);
6469         btrfs_set_header_generation(buf, trans->transid);
6470         btrfs_set_buffer_lockdep_class(root->root_key.objectid, buf, level);
6471         btrfs_tree_lock(buf);
6472         clean_tree_block(trans, root, buf);
6473         clear_bit(EXTENT_BUFFER_STALE, &buf->bflags);
6474
6475         btrfs_set_lock_blocking(buf);
6476         btrfs_set_buffer_uptodate(buf);
6477
6478         if (root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID) {
6479                 /*
6480                  * we allow two log transactions at a time, use different
6481                  * EXENT bit to differentiate dirty pages.
6482                  */
6483                 if (root->log_transid % 2 == 0)
6484                         set_extent_dirty(&root->dirty_log_pages, buf->start,
6485                                         buf->start + buf->len - 1, GFP_NOFS);
6486                 else
6487                         set_extent_new(&root->dirty_log_pages, buf->start,
6488                                         buf->start + buf->len - 1, GFP_NOFS);
6489         } else {
6490                 set_extent_dirty(&trans->transaction->dirty_pages, buf->start,
6491                          buf->start + buf->len - 1, GFP_NOFS);
6492         }
6493         trans->blocks_used++;
6494         /* this returns a buffer locked for blocking */
6495         return buf;
6496 }
6497
6498 static struct btrfs_block_rsv *
6499 use_block_rsv(struct btrfs_trans_handle *trans,
6500               struct btrfs_root *root, u32 blocksize)
6501 {
6502         struct btrfs_block_rsv *block_rsv;
6503         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
6504         int ret;
6505
6506         block_rsv = get_block_rsv(trans, root);
6507
6508         if (block_rsv->size == 0) {
6509                 ret = reserve_metadata_bytes(root, block_rsv, blocksize,
6510                                              BTRFS_RESERVE_NO_FLUSH);
6511                 /*
6512                  * If we couldn't reserve metadata bytes try and use some from
6513                  * the global reserve.
6514                  */
6515                 if (ret && block_rsv != global_rsv) {
6516                         ret = block_rsv_use_bytes(global_rsv, blocksize);
6517                         if (!ret)
6518                                 return global_rsv;
6519                         return ERR_PTR(ret);
6520                 } else if (ret) {
6521                         return ERR_PTR(ret);
6522                 }
6523                 return block_rsv;
6524         }
6525
6526         ret = block_rsv_use_bytes(block_rsv, blocksize);
6527         if (!ret)
6528                 return block_rsv;
6529         if (ret && !block_rsv->failfast) {
6530                 if (btrfs_test_opt(root, ENOSPC_DEBUG)) {
6531                         static DEFINE_RATELIMIT_STATE(_rs,
6532                                         DEFAULT_RATELIMIT_INTERVAL * 10,
6533                                         /*DEFAULT_RATELIMIT_BURST*/ 1);
6534                         if (__ratelimit(&_rs))
6535                                 WARN(1, KERN_DEBUG
6536                                         "btrfs: block rsv returned %d\n", ret);
6537                 }
6538                 ret = reserve_metadata_bytes(root, block_rsv, blocksize,
6539                                              BTRFS_RESERVE_NO_FLUSH);
6540                 if (!ret) {
6541                         return block_rsv;
6542                 } else if (ret && block_rsv != global_rsv) {
6543                         ret = block_rsv_use_bytes(global_rsv, blocksize);
6544                         if (!ret)
6545                                 return global_rsv;
6546                 }
6547         }
6548
6549         return ERR_PTR(-ENOSPC);
6550 }
6551
6552 static void unuse_block_rsv(struct btrfs_fs_info *fs_info,
6553                             struct btrfs_block_rsv *block_rsv, u32 blocksize)
6554 {
6555         block_rsv_add_bytes(block_rsv, blocksize, 0);
6556         block_rsv_release_bytes(fs_info, block_rsv, NULL, 0);
6557 }
6558
6559 /*
6560  * finds a free extent and does all the dirty work required for allocation
6561  * returns the key for the extent through ins, and a tree buffer for
6562  * the first block of the extent through buf.
6563  *
6564  * returns the tree buffer or NULL.
6565  */
6566 struct extent_buffer *btrfs_alloc_free_block(struct btrfs_trans_handle *trans,
6567                                         struct btrfs_root *root, u32 blocksize,
6568                                         u64 parent, u64 root_objectid,
6569                                         struct btrfs_disk_key *key, int level,
6570                                         u64 hint, u64 empty_size)
6571 {
6572         struct btrfs_key ins;
6573         struct btrfs_block_rsv *block_rsv;
6574         struct extent_buffer *buf;
6575         u64 flags = 0;
6576         int ret;
6577
6578
6579         block_rsv = use_block_rsv(trans, root, blocksize);
6580         if (IS_ERR(block_rsv))
6581                 return ERR_CAST(block_rsv);
6582
6583         ret = btrfs_reserve_extent(trans, root, blocksize, blocksize,
6584                                    empty_size, hint, &ins, 0);
6585         if (ret) {
6586                 unuse_block_rsv(root->fs_info, block_rsv, blocksize);
6587                 return ERR_PTR(ret);
6588         }
6589
6590         buf = btrfs_init_new_buffer(trans, root, ins.objectid,
6591                                     blocksize, level);
6592         BUG_ON(IS_ERR(buf)); /* -ENOMEM */
6593
6594         if (root_objectid == BTRFS_TREE_RELOC_OBJECTID) {
6595                 if (parent == 0)
6596                         parent = ins.objectid;
6597                 flags |= BTRFS_BLOCK_FLAG_FULL_BACKREF;
6598         } else
6599                 BUG_ON(parent > 0);
6600
6601         if (root_objectid != BTRFS_TREE_LOG_OBJECTID) {
6602                 struct btrfs_delayed_extent_op *extent_op;
6603                 extent_op = btrfs_alloc_delayed_extent_op();
6604                 BUG_ON(!extent_op); /* -ENOMEM */
6605                 if (key)
6606                         memcpy(&extent_op->key, key, sizeof(extent_op->key));
6607                 else
6608                         memset(&extent_op->key, 0, sizeof(extent_op->key));
6609                 extent_op->flags_to_set = flags;
6610                 extent_op->update_key = 1;
6611                 extent_op->update_flags = 1;
6612                 extent_op->is_data = 0;
6613
6614                 ret = btrfs_add_delayed_tree_ref(root->fs_info, trans,
6615                                         ins.objectid,
6616                                         ins.offset, parent, root_objectid,
6617                                         level, BTRFS_ADD_DELAYED_EXTENT,
6618                                         extent_op, 0);
6619                 BUG_ON(ret); /* -ENOMEM */
6620         }
6621         return buf;
6622 }
6623
6624 struct walk_control {
6625         u64 refs[BTRFS_MAX_LEVEL];
6626         u64 flags[BTRFS_MAX_LEVEL];
6627         struct btrfs_key update_progress;
6628         int stage;
6629         int level;
6630         int shared_level;
6631         int update_ref;
6632         int keep_locks;
6633         int reada_slot;
6634         int reada_count;
6635         int for_reloc;
6636 };
6637
6638 #define DROP_REFERENCE  1
6639 #define UPDATE_BACKREF  2
6640
6641 static noinline void reada_walk_down(struct btrfs_trans_handle *trans,
6642                                      struct btrfs_root *root,
6643                                      struct walk_control *wc,
6644                                      struct btrfs_path *path)
6645 {
6646         u64 bytenr;
6647         u64 generation;
6648         u64 refs;
6649         u64 flags;
6650         u32 nritems;
6651         u32 blocksize;
6652         struct btrfs_key key;
6653         struct extent_buffer *eb;
6654         int ret;
6655         int slot;
6656         int nread = 0;
6657
6658         if (path->slots[wc->level] < wc->reada_slot) {
6659                 wc->reada_count = wc->reada_count * 2 / 3;
6660                 wc->reada_count = max(wc->reada_count, 2);
6661         } else {
6662                 wc->reada_count = wc->reada_count * 3 / 2;
6663                 wc->reada_count = min_t(int, wc->reada_count,
6664                                         BTRFS_NODEPTRS_PER_BLOCK(root));
6665         }
6666
6667         eb = path->nodes[wc->level];
6668         nritems = btrfs_header_nritems(eb);
6669         blocksize = btrfs_level_size(root, wc->level - 1);
6670
6671         for (slot = path->slots[wc->level]; slot < nritems; slot++) {
6672                 if (nread >= wc->reada_count)
6673                         break;
6674
6675                 cond_resched();
6676                 bytenr = btrfs_node_blockptr(eb, slot);
6677                 generation = btrfs_node_ptr_generation(eb, slot);
6678
6679                 if (slot == path->slots[wc->level])
6680                         goto reada;
6681
6682                 if (wc->stage == UPDATE_BACKREF &&
6683                     generation <= root->root_key.offset)
6684                         continue;
6685
6686                 /* We don't lock the tree block, it's OK to be racy here */
6687                 ret = btrfs_lookup_extent_info(trans, root, bytenr, blocksize,
6688                                                &refs, &flags);
6689                 /* We don't care about errors in readahead. */
6690                 if (ret < 0)
6691                         continue;
6692                 BUG_ON(refs == 0);
6693
6694                 if (wc->stage == DROP_REFERENCE) {
6695                         if (refs == 1)
6696                                 goto reada;
6697
6698                         if (wc->level == 1 &&
6699                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
6700                                 continue;
6701                         if (!wc->update_ref ||
6702                             generation <= root->root_key.offset)
6703                                 continue;
6704                         btrfs_node_key_to_cpu(eb, &key, slot);
6705                         ret = btrfs_comp_cpu_keys(&key,
6706                                                   &wc->update_progress);
6707                         if (ret < 0)
6708                                 continue;
6709                 } else {
6710                         if (wc->level == 1 &&
6711                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
6712                                 continue;
6713                 }
6714 reada:
6715                 ret = readahead_tree_block(root, bytenr, blocksize,
6716                                            generation);
6717                 if (ret)
6718                         break;
6719                 nread++;
6720         }
6721         wc->reada_slot = slot;
6722 }
6723
6724 /*
6725  * hepler to process tree block while walking down the tree.
6726  *
6727  * when wc->stage == UPDATE_BACKREF, this function updates
6728  * back refs for pointers in the block.
6729  *
6730  * NOTE: return value 1 means we should stop walking down.
6731  */
6732 static noinline int walk_down_proc(struct btrfs_trans_handle *trans,
6733                                    struct btrfs_root *root,
6734                                    struct btrfs_path *path,
6735                                    struct walk_control *wc, int lookup_info)
6736 {
6737         int level = wc->level;
6738         struct extent_buffer *eb = path->nodes[level];
6739         u64 flag = BTRFS_BLOCK_FLAG_FULL_BACKREF;
6740         int ret;
6741
6742         if (wc->stage == UPDATE_BACKREF &&
6743             btrfs_header_owner(eb) != root->root_key.objectid)
6744                 return 1;
6745
6746         /*
6747          * when reference count of tree block is 1, it won't increase
6748          * again. once full backref flag is set, we never clear it.
6749          */
6750         if (lookup_info &&
6751             ((wc->stage == DROP_REFERENCE && wc->refs[level] != 1) ||
6752              (wc->stage == UPDATE_BACKREF && !(wc->flags[level] & flag)))) {
6753                 BUG_ON(!path->locks[level]);
6754                 ret = btrfs_lookup_extent_info(trans, root,
6755                                                eb->start, eb->len,
6756                                                &wc->refs[level],
6757                                                &wc->flags[level]);
6758                 BUG_ON(ret == -ENOMEM);
6759                 if (ret)
6760                         return ret;
6761                 BUG_ON(wc->refs[level] == 0);
6762         }
6763
6764         if (wc->stage == DROP_REFERENCE) {
6765                 if (wc->refs[level] > 1)
6766                         return 1;
6767
6768                 if (path->locks[level] && !wc->keep_locks) {
6769                         btrfs_tree_unlock_rw(eb, path->locks[level]);
6770                         path->locks[level] = 0;
6771                 }
6772                 return 0;
6773         }
6774
6775         /* wc->stage == UPDATE_BACKREF */
6776         if (!(wc->flags[level] & flag)) {
6777                 BUG_ON(!path->locks[level]);
6778                 ret = btrfs_inc_ref(trans, root, eb, 1, wc->for_reloc);
6779                 BUG_ON(ret); /* -ENOMEM */
6780                 ret = btrfs_dec_ref(trans, root, eb, 0, wc->for_reloc);
6781                 BUG_ON(ret); /* -ENOMEM */
6782                 ret = btrfs_set_disk_extent_flags(trans, root, eb->start,
6783                                                   eb->len, flag, 0);
6784                 BUG_ON(ret); /* -ENOMEM */
6785                 wc->flags[level] |= flag;
6786         }
6787
6788         /*
6789          * the block is shared by multiple trees, so it's not good to
6790          * keep the tree lock
6791          */
6792         if (path->locks[level] && level > 0) {
6793                 btrfs_tree_unlock_rw(eb, path->locks[level]);
6794                 path->locks[level] = 0;
6795         }
6796         return 0;
6797 }
6798
6799 /*
6800  * hepler to process tree block pointer.
6801  *
6802  * when wc->stage == DROP_REFERENCE, this function checks
6803  * reference count of the block pointed to. if the block
6804  * is shared and we need update back refs for the subtree
6805  * rooted at the block, this function changes wc->stage to
6806  * UPDATE_BACKREF. if the block is shared and there is no
6807  * need to update back, this function drops the reference
6808  * to the block.
6809  *
6810  * NOTE: return value 1 means we should stop walking down.
6811  */
6812 static noinline int do_walk_down(struct btrfs_trans_handle *trans,
6813                                  struct btrfs_root *root,
6814                                  struct btrfs_path *path,
6815                                  struct walk_control *wc, int *lookup_info)
6816 {
6817         u64 bytenr;
6818         u64 generation;
6819         u64 parent;
6820         u32 blocksize;
6821         struct btrfs_key key;
6822         struct extent_buffer *next;
6823         int level = wc->level;
6824         int reada = 0;
6825         int ret = 0;
6826
6827         generation = btrfs_node_ptr_generation(path->nodes[level],
6828                                                path->slots[level]);
6829         /*
6830          * if the lower level block was created before the snapshot
6831          * was created, we know there is no need to update back refs
6832          * for the subtree
6833          */
6834         if (wc->stage == UPDATE_BACKREF &&
6835             generation <= root->root_key.offset) {
6836                 *lookup_info = 1;
6837                 return 1;
6838         }
6839
6840         bytenr = btrfs_node_blockptr(path->nodes[level], path->slots[level]);
6841         blocksize = btrfs_level_size(root, level - 1);
6842
6843         next = btrfs_find_tree_block(root, bytenr, blocksize);
6844         if (!next) {
6845                 next = btrfs_find_create_tree_block(root, bytenr, blocksize);
6846                 if (!next)
6847                         return -ENOMEM;
6848                 reada = 1;
6849         }
6850         btrfs_tree_lock(next);
6851         btrfs_set_lock_blocking(next);
6852
6853         ret = btrfs_lookup_extent_info(trans, root, bytenr, blocksize,
6854                                        &wc->refs[level - 1],
6855                                        &wc->flags[level - 1]);
6856         if (ret < 0) {
6857                 btrfs_tree_unlock(next);
6858                 return ret;
6859         }
6860
6861         BUG_ON(wc->refs[level - 1] == 0);
6862         *lookup_info = 0;
6863
6864         if (wc->stage == DROP_REFERENCE) {
6865                 if (wc->refs[level - 1] > 1) {
6866                         if (level == 1 &&
6867                             (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
6868                                 goto skip;
6869
6870                         if (!wc->update_ref ||
6871                             generation <= root->root_key.offset)
6872                                 goto skip;
6873
6874                         btrfs_node_key_to_cpu(path->nodes[level], &key,
6875                                               path->slots[level]);
6876                         ret = btrfs_comp_cpu_keys(&key, &wc->update_progress);
6877                         if (ret < 0)
6878                                 goto skip;
6879
6880                         wc->stage = UPDATE_BACKREF;
6881                         wc->shared_level = level - 1;
6882                 }
6883         } else {
6884                 if (level == 1 &&
6885                     (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
6886                         goto skip;
6887         }
6888
6889         if (!btrfs_buffer_uptodate(next, generation, 0)) {
6890                 btrfs_tree_unlock(next);
6891                 free_extent_buffer(next);
6892                 next = NULL;
6893                 *lookup_info = 1;
6894         }
6895
6896         if (!next) {
6897                 if (reada && level == 1)
6898                         reada_walk_down(trans, root, wc, path);
6899                 next = read_tree_block(root, bytenr, blocksize, generation);
6900                 if (!next)
6901                         return -EIO;
6902                 btrfs_tree_lock(next);
6903                 btrfs_set_lock_blocking(next);
6904         }
6905
6906         level--;
6907         BUG_ON(level != btrfs_header_level(next));
6908         path->nodes[level] = next;
6909         path->slots[level] = 0;
6910         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
6911         wc->level = level;
6912         if (wc->level == 1)
6913                 wc->reada_slot = 0;
6914         return 0;
6915 skip:
6916         wc->refs[level - 1] = 0;
6917         wc->flags[level - 1] = 0;
6918         if (wc->stage == DROP_REFERENCE) {
6919                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF) {
6920                         parent = path->nodes[level]->start;
6921                 } else {
6922                         BUG_ON(root->root_key.objectid !=
6923                                btrfs_header_owner(path->nodes[level]));
6924                         parent = 0;
6925                 }
6926
6927                 ret = btrfs_free_extent(trans, root, bytenr, blocksize, parent,
6928                                 root->root_key.objectid, level - 1, 0, 0);
6929                 BUG_ON(ret); /* -ENOMEM */
6930         }
6931         btrfs_tree_unlock(next);
6932         free_extent_buffer(next);
6933         *lookup_info = 1;
6934         return 1;
6935 }
6936
6937 /*
6938  * hepler to process tree block while walking up the tree.
6939  *
6940  * when wc->stage == DROP_REFERENCE, this function drops
6941  * reference count on the block.
6942  *
6943  * when wc->stage == UPDATE_BACKREF, this function changes
6944  * wc->stage back to DROP_REFERENCE if we changed wc->stage
6945  * to UPDATE_BACKREF previously while processing the block.
6946  *
6947  * NOTE: return value 1 means we should stop walking up.
6948  */
6949 static noinline int walk_up_proc(struct btrfs_trans_handle *trans,
6950                                  struct btrfs_root *root,
6951                                  struct btrfs_path *path,
6952                                  struct walk_control *wc)
6953 {
6954         int ret;
6955         int level = wc->level;
6956         struct extent_buffer *eb = path->nodes[level];
6957         u64 parent = 0;
6958
6959         if (wc->stage == UPDATE_BACKREF) {
6960                 BUG_ON(wc->shared_level < level);
6961                 if (level < wc->shared_level)
6962                         goto out;
6963
6964                 ret = find_next_key(path, level + 1, &wc->update_progress);
6965                 if (ret > 0)
6966                         wc->update_ref = 0;
6967
6968                 wc->stage = DROP_REFERENCE;
6969                 wc->shared_level = -1;
6970                 path->slots[level] = 0;
6971
6972                 /*
6973                  * check reference count again if the block isn't locked.
6974                  * we should start walking down the tree again if reference
6975                  * count is one.
6976                  */
6977                 if (!path->locks[level]) {
6978                         BUG_ON(level == 0);
6979                         btrfs_tree_lock(eb);
6980                         btrfs_set_lock_blocking(eb);
6981                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
6982
6983                         ret = btrfs_lookup_extent_info(trans, root,
6984                                                        eb->start, eb->len,
6985                                                        &wc->refs[level],
6986                                                        &wc->flags[level]);
6987                         if (ret < 0) {
6988                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
6989                                 path->locks[level] = 0;
6990                                 return ret;
6991                         }
6992                         BUG_ON(wc->refs[level] == 0);
6993                         if (wc->refs[level] == 1) {
6994                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
6995                                 path->locks[level] = 0;
6996                                 return 1;
6997                         }
6998                 }
6999         }
7000
7001         /* wc->stage == DROP_REFERENCE */
7002         BUG_ON(wc->refs[level] > 1 && !path->locks[level]);
7003
7004         if (wc->refs[level] == 1) {
7005                 if (level == 0) {
7006                         if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
7007                                 ret = btrfs_dec_ref(trans, root, eb, 1,
7008                                                     wc->for_reloc);
7009                         else
7010                                 ret = btrfs_dec_ref(trans, root, eb, 0,
7011                                                     wc->for_reloc);
7012                         BUG_ON(ret); /* -ENOMEM */
7013                 }
7014                 /* make block locked assertion in clean_tree_block happy */
7015                 if (!path->locks[level] &&
7016                     btrfs_header_generation(eb) == trans->transid) {
7017                         btrfs_tree_lock(eb);
7018                         btrfs_set_lock_blocking(eb);
7019                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
7020                 }
7021                 clean_tree_block(trans, root, eb);
7022         }
7023
7024         if (eb == root->node) {
7025                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
7026                         parent = eb->start;
7027                 else
7028                         BUG_ON(root->root_key.objectid !=
7029                                btrfs_header_owner(eb));
7030         } else {
7031                 if (wc->flags[level + 1] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
7032                         parent = path->nodes[level + 1]->start;
7033                 else
7034                         BUG_ON(root->root_key.objectid !=
7035                                btrfs_header_owner(path->nodes[level + 1]));
7036         }
7037
7038         btrfs_free_tree_block(trans, root, eb, parent, wc->refs[level] == 1);
7039 out:
7040         wc->refs[level] = 0;
7041         wc->flags[level] = 0;
7042         return 0;
7043 }
7044
7045 static noinline int walk_down_tree(struct btrfs_trans_handle *trans,
7046                                    struct btrfs_root *root,
7047                                    struct btrfs_path *path,
7048                                    struct walk_control *wc)
7049 {
7050         int level = wc->level;
7051         int lookup_info = 1;
7052         int ret;
7053
7054         while (level >= 0) {
7055                 ret = walk_down_proc(trans, root, path, wc, lookup_info);
7056                 if (ret > 0)
7057                         break;
7058
7059                 if (level == 0)
7060                         break;
7061
7062                 if (path->slots[level] >=
7063                     btrfs_header_nritems(path->nodes[level]))
7064                         break;
7065
7066                 ret = do_walk_down(trans, root, path, wc, &lookup_info);
7067                 if (ret > 0) {
7068                         path->slots[level]++;
7069                         continue;
7070                 } else if (ret < 0)
7071                         return ret;
7072                 level = wc->level;
7073         }
7074         return 0;
7075 }
7076
7077 static noinline int walk_up_tree(struct btrfs_trans_handle *trans,
7078                                  struct btrfs_root *root,
7079                                  struct btrfs_path *path,
7080                                  struct walk_control *wc, int max_level)
7081 {
7082         int level = wc->level;
7083         int ret;
7084
7085         path->slots[level] = btrfs_header_nritems(path->nodes[level]);
7086         while (level < max_level && path->nodes[level]) {
7087                 wc->level = level;
7088                 if (path->slots[level] + 1 <
7089                     btrfs_header_nritems(path->nodes[level])) {
7090                         path->slots[level]++;
7091                         return 0;
7092                 } else {
7093                         ret = walk_up_proc(trans, root, path, wc);
7094                         if (ret > 0)
7095                                 return 0;
7096
7097                         if (path->locks[level]) {
7098                                 btrfs_tree_unlock_rw(path->nodes[level],
7099                                                      path->locks[level]);
7100                                 path->locks[level] = 0;
7101                         }
7102                         free_extent_buffer(path->nodes[level]);
7103                         path->nodes[level] = NULL;
7104                         level++;
7105                 }
7106         }
7107         return 1;
7108 }
7109
7110 /*
7111  * drop a subvolume tree.
7112  *
7113  * this function traverses the tree freeing any blocks that only
7114  * referenced by the tree.
7115  *
7116  * when a shared tree block is found. this function decreases its
7117  * reference count by one. if update_ref is true, this function
7118  * also make sure backrefs for the shared block and all lower level
7119  * blocks are properly updated.
7120  */
7121 int btrfs_drop_snapshot(struct btrfs_root *root,
7122                          struct btrfs_block_rsv *block_rsv, int update_ref,
7123                          int for_reloc)
7124 {
7125         struct btrfs_path *path;
7126         struct btrfs_trans_handle *trans;
7127         struct btrfs_root *tree_root = root->fs_info->tree_root;
7128         struct btrfs_root_item *root_item = &root->root_item;
7129         struct walk_control *wc;
7130         struct btrfs_key key;
7131         int err = 0;
7132         int ret;
7133         int level;
7134
7135         path = btrfs_alloc_path();
7136         if (!path) {
7137                 err = -ENOMEM;
7138                 goto out;
7139         }
7140
7141         wc = kzalloc(sizeof(*wc), GFP_NOFS);
7142         if (!wc) {
7143                 btrfs_free_path(path);
7144                 err = -ENOMEM;
7145                 goto out;
7146         }
7147
7148         trans = btrfs_start_transaction(tree_root, 0);
7149         if (IS_ERR(trans)) {
7150                 err = PTR_ERR(trans);
7151                 goto out_free;
7152         }
7153
7154         if (block_rsv)
7155                 trans->block_rsv = block_rsv;
7156
7157         if (btrfs_disk_key_objectid(&root_item->drop_progress) == 0) {
7158                 level = btrfs_header_level(root->node);
7159                 path->nodes[level] = btrfs_lock_root_node(root);
7160                 btrfs_set_lock_blocking(path->nodes[level]);
7161                 path->slots[level] = 0;
7162                 path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
7163                 memset(&wc->update_progress, 0,
7164                        sizeof(wc->update_progress));
7165         } else {
7166                 btrfs_disk_key_to_cpu(&key, &root_item->drop_progress);
7167                 memcpy(&wc->update_progress, &key,
7168                        sizeof(wc->update_progress));
7169
7170                 level = root_item->drop_level;
7171                 BUG_ON(level == 0);
7172                 path->lowest_level = level;
7173                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
7174                 path->lowest_level = 0;
7175                 if (ret < 0) {
7176                         err = ret;
7177                         goto out_end_trans;
7178                 }
7179                 WARN_ON(ret > 0);
7180
7181                 /*
7182                  * unlock our path, this is safe because only this
7183                  * function is allowed to delete this snapshot
7184                  */
7185                 btrfs_unlock_up_safe(path, 0);
7186
7187                 level = btrfs_header_level(root->node);
7188                 while (1) {
7189                         btrfs_tree_lock(path->nodes[level]);
7190                         btrfs_set_lock_blocking(path->nodes[level]);
7191
7192                         ret = btrfs_lookup_extent_info(trans, root,
7193                                                 path->nodes[level]->start,
7194                                                 path->nodes[level]->len,
7195                                                 &wc->refs[level],
7196                                                 &wc->flags[level]);
7197                         if (ret < 0) {
7198                                 err = ret;
7199                                 goto out_end_trans;
7200                         }
7201                         BUG_ON(wc->refs[level] == 0);
7202
7203                         if (level == root_item->drop_level)
7204                                 break;
7205
7206                         btrfs_tree_unlock(path->nodes[level]);
7207                         WARN_ON(wc->refs[level] != 1);
7208                         level--;
7209                 }
7210         }
7211
7212         wc->level = level;
7213         wc->shared_level = -1;
7214         wc->stage = DROP_REFERENCE;
7215         wc->update_ref = update_ref;
7216         wc->keep_locks = 0;
7217         wc->for_reloc = for_reloc;
7218         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(root);
7219
7220         while (1) {
7221                 ret = walk_down_tree(trans, root, path, wc);
7222                 if (ret < 0) {
7223                         err = ret;
7224                         break;
7225                 }
7226
7227                 ret = walk_up_tree(trans, root, path, wc, BTRFS_MAX_LEVEL);
7228                 if (ret < 0) {
7229                         err = ret;
7230                         break;
7231                 }
7232
7233                 if (ret > 0) {
7234                         BUG_ON(wc->stage != DROP_REFERENCE);
7235                         break;
7236                 }
7237
7238                 if (wc->stage == DROP_REFERENCE) {
7239                         level = wc->level;
7240                         btrfs_node_key(path->nodes[level],
7241                                        &root_item->drop_progress,
7242                                        path->slots[level]);
7243                         root_item->drop_level = level;
7244                 }
7245
7246                 BUG_ON(wc->level == 0);
7247                 if (btrfs_should_end_transaction(trans, tree_root)) {
7248                         ret = btrfs_update_root(trans, tree_root,
7249                                                 &root->root_key,
7250                                                 root_item);
7251                         if (ret) {
7252                                 btrfs_abort_transaction(trans, tree_root, ret);
7253                                 err = ret;
7254                                 goto out_end_trans;
7255                         }
7256
7257                         btrfs_end_transaction_throttle(trans, tree_root);
7258                         trans = btrfs_start_transaction(tree_root, 0);
7259                         if (IS_ERR(trans)) {
7260                                 err = PTR_ERR(trans);
7261                                 goto out_free;
7262                         }
7263                         if (block_rsv)
7264                                 trans->block_rsv = block_rsv;
7265                 }
7266         }
7267         btrfs_release_path(path);
7268         if (err)
7269                 goto out_end_trans;
7270
7271         ret = btrfs_del_root(trans, tree_root, &root->root_key);
7272         if (ret) {
7273                 btrfs_abort_transaction(trans, tree_root, ret);
7274                 goto out_end_trans;
7275         }
7276
7277         if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) {
7278                 ret = btrfs_find_last_root(tree_root, root->root_key.objectid,
7279                                            NULL, NULL);
7280                 if (ret < 0) {
7281                         btrfs_abort_transaction(trans, tree_root, ret);
7282                         err = ret;
7283                         goto out_end_trans;
7284                 } else if (ret > 0) {
7285                         /* if we fail to delete the orphan item this time
7286                          * around, it'll get picked up the next time.
7287                          *
7288                          * The most common failure here is just -ENOENT.
7289                          */
7290                         btrfs_del_orphan_item(trans, tree_root,
7291                                               root->root_key.objectid);
7292                 }
7293         }
7294
7295         if (root->in_radix) {
7296                 btrfs_free_fs_root(tree_root->fs_info, root);
7297         } else {
7298                 free_extent_buffer(root->node);
7299                 free_extent_buffer(root->commit_root);
7300                 kfree(root);
7301         }
7302 out_end_trans:
7303         btrfs_end_transaction_throttle(trans, tree_root);
7304 out_free:
7305         kfree(wc);
7306         btrfs_free_path(path);
7307 out:
7308         if (err)
7309                 btrfs_std_error(root->fs_info, err);
7310         return err;
7311 }
7312
7313 /*
7314  * drop subtree rooted at tree block 'node'.
7315  *
7316  * NOTE: this function will unlock and release tree block 'node'
7317  * only used by relocation code
7318  */
7319 int btrfs_drop_subtree(struct btrfs_trans_handle *trans,
7320                         struct btrfs_root *root,
7321                         struct extent_buffer *node,
7322                         struct extent_buffer *parent)
7323 {
7324         struct btrfs_path *path;
7325         struct walk_control *wc;
7326         int level;
7327         int parent_level;
7328         int ret = 0;
7329         int wret;
7330
7331         BUG_ON(root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID);
7332
7333         path = btrfs_alloc_path();
7334         if (!path)
7335                 return -ENOMEM;
7336
7337         wc = kzalloc(sizeof(*wc), GFP_NOFS);
7338         if (!wc) {
7339                 btrfs_free_path(path);
7340                 return -ENOMEM;
7341         }
7342
7343         btrfs_assert_tree_locked(parent);
7344         parent_level = btrfs_header_level(parent);
7345         extent_buffer_get(parent);
7346         path->nodes[parent_level] = parent;
7347         path->slots[parent_level] = btrfs_header_nritems(parent);
7348
7349         btrfs_assert_tree_locked(node);
7350         level = btrfs_header_level(node);
7351         path->nodes[level] = node;
7352         path->slots[level] = 0;
7353         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
7354
7355         wc->refs[parent_level] = 1;
7356         wc->flags[parent_level] = BTRFS_BLOCK_FLAG_FULL_BACKREF;
7357         wc->level = level;
7358         wc->shared_level = -1;
7359         wc->stage = DROP_REFERENCE;
7360         wc->update_ref = 0;
7361         wc->keep_locks = 1;
7362         wc->for_reloc = 1;
7363         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(root);
7364
7365         while (1) {
7366                 wret = walk_down_tree(trans, root, path, wc);
7367                 if (wret < 0) {
7368                         ret = wret;
7369                         break;
7370                 }
7371
7372                 wret = walk_up_tree(trans, root, path, wc, parent_level);
7373                 if (wret < 0)
7374                         ret = wret;
7375                 if (wret != 0)
7376                         break;
7377         }
7378
7379         kfree(wc);
7380         btrfs_free_path(path);
7381         return ret;
7382 }
7383
7384 static u64 update_block_group_flags(struct btrfs_root *root, u64 flags)
7385 {
7386         u64 num_devices;
7387         u64 stripped;
7388
7389         /*
7390          * if restripe for this chunk_type is on pick target profile and
7391          * return, otherwise do the usual balance
7392          */
7393         stripped = get_restripe_target(root->fs_info, flags);
7394         if (stripped)
7395                 return extended_to_chunk(stripped);
7396
7397         /*
7398          * we add in the count of missing devices because we want
7399          * to make sure that any RAID levels on a degraded FS
7400          * continue to be honored.
7401          */
7402         num_devices = root->fs_info->fs_devices->rw_devices +
7403                 root->fs_info->fs_devices->missing_devices;
7404
7405         stripped = BTRFS_BLOCK_GROUP_RAID0 |
7406                 BTRFS_BLOCK_GROUP_RAID5 | BTRFS_BLOCK_GROUP_RAID6 |
7407                 BTRFS_BLOCK_GROUP_RAID1 | BTRFS_BLOCK_GROUP_RAID10;
7408
7409         if (num_devices == 1) {
7410                 stripped |= BTRFS_BLOCK_GROUP_DUP;
7411                 stripped = flags & ~stripped;
7412
7413                 /* turn raid0 into single device chunks */
7414                 if (flags & BTRFS_BLOCK_GROUP_RAID0)
7415                         return stripped;
7416
7417                 /* turn mirroring into duplication */
7418                 if (flags & (BTRFS_BLOCK_GROUP_RAID1 |
7419                              BTRFS_BLOCK_GROUP_RAID10))
7420                         return stripped | BTRFS_BLOCK_GROUP_DUP;
7421         } else {
7422                 /* they already had raid on here, just return */
7423                 if (flags & stripped)
7424                         return flags;
7425
7426                 stripped |= BTRFS_BLOCK_GROUP_DUP;
7427                 stripped = flags & ~stripped;
7428
7429                 /* switch duplicated blocks with raid1 */
7430                 if (flags & BTRFS_BLOCK_GROUP_DUP)
7431                         return stripped | BTRFS_BLOCK_GROUP_RAID1;
7432
7433                 /* this is drive concat, leave it alone */
7434         }
7435
7436         return flags;
7437 }
7438
7439 static int set_block_group_ro(struct btrfs_block_group_cache *cache, int force)
7440 {
7441         struct btrfs_space_info *sinfo = cache->space_info;
7442         u64 num_bytes;
7443         u64 min_allocable_bytes;
7444         int ret = -ENOSPC;
7445
7446
7447         /*
7448          * We need some metadata space and system metadata space for
7449          * allocating chunks in some corner cases until we force to set
7450          * it to be readonly.
7451          */
7452         if ((sinfo->flags &
7453              (BTRFS_BLOCK_GROUP_SYSTEM | BTRFS_BLOCK_GROUP_METADATA)) &&
7454             !force)
7455                 min_allocable_bytes = 1 * 1024 * 1024;
7456         else
7457                 min_allocable_bytes = 0;
7458
7459         spin_lock(&sinfo->lock);
7460         spin_lock(&cache->lock);
7461
7462         if (cache->ro) {
7463                 ret = 0;
7464                 goto out;
7465         }
7466
7467         num_bytes = cache->key.offset - cache->reserved - cache->pinned -
7468                     cache->bytes_super - btrfs_block_group_used(&cache->item);
7469
7470         if (sinfo->bytes_used + sinfo->bytes_reserved + sinfo->bytes_pinned +
7471             sinfo->bytes_may_use + sinfo->bytes_readonly + num_bytes +
7472             min_allocable_bytes <= sinfo->total_bytes) {
7473                 sinfo->bytes_readonly += num_bytes;
7474                 cache->ro = 1;
7475                 ret = 0;
7476         }
7477 out:
7478         spin_unlock(&cache->lock);
7479         spin_unlock(&sinfo->lock);
7480         return ret;
7481 }
7482
7483 int btrfs_set_block_group_ro(struct btrfs_root *root,
7484                              struct btrfs_block_group_cache *cache)
7485
7486 {
7487         struct btrfs_trans_handle *trans;
7488         u64 alloc_flags;
7489         int ret;
7490
7491         BUG_ON(cache->ro);
7492
7493         trans = btrfs_join_transaction(root);
7494         if (IS_ERR(trans))
7495                 return PTR_ERR(trans);
7496
7497         alloc_flags = update_block_group_flags(root, cache->flags);
7498         if (alloc_flags != cache->flags) {
7499                 ret = do_chunk_alloc(trans, root, alloc_flags,
7500                                      CHUNK_ALLOC_FORCE);
7501                 if (ret < 0)
7502                         goto out;
7503         }
7504
7505         ret = set_block_group_ro(cache, 0);
7506         if (!ret)
7507                 goto out;
7508         alloc_flags = get_alloc_profile(root, cache->space_info->flags);
7509         ret = do_chunk_alloc(trans, root, alloc_flags,
7510                              CHUNK_ALLOC_FORCE);
7511         if (ret < 0)
7512                 goto out;
7513         ret = set_block_group_ro(cache, 0);
7514 out:
7515         btrfs_end_transaction(trans, root);
7516         return ret;
7517 }
7518
7519 int btrfs_force_chunk_alloc(struct btrfs_trans_handle *trans,
7520                             struct btrfs_root *root, u64 type)
7521 {
7522         u64 alloc_flags = get_alloc_profile(root, type);
7523         return do_chunk_alloc(trans, root, alloc_flags,
7524                               CHUNK_ALLOC_FORCE);
7525 }
7526
7527 /*
7528  * helper to account the unused space of all the readonly block group in the
7529  * list. takes mirrors into account.
7530  */
7531 static u64 __btrfs_get_ro_block_group_free_space(struct list_head *groups_list)
7532 {
7533         struct btrfs_block_group_cache *block_group;
7534         u64 free_bytes = 0;
7535         int factor;
7536
7537         list_for_each_entry(block_group, groups_list, list) {
7538                 spin_lock(&block_group->lock);
7539
7540                 if (!block_group->ro) {
7541                         spin_unlock(&block_group->lock);
7542                         continue;
7543                 }
7544
7545                 if (block_group->flags & (BTRFS_BLOCK_GROUP_RAID1 |
7546                                           BTRFS_BLOCK_GROUP_RAID10 |
7547                                           BTRFS_BLOCK_GROUP_DUP))
7548                         factor = 2;
7549                 else
7550                         factor = 1;
7551
7552                 free_bytes += (block_group->key.offset -
7553                                btrfs_block_group_used(&block_group->item)) *
7554                                factor;
7555
7556                 spin_unlock(&block_group->lock);
7557         }
7558
7559         return free_bytes;
7560 }
7561
7562 /*
7563  * helper to account the unused space of all the readonly block group in the
7564  * space_info. takes mirrors into account.
7565  */
7566 u64 btrfs_account_ro_block_groups_free_space(struct btrfs_space_info *sinfo)
7567 {
7568         int i;
7569         u64 free_bytes = 0;
7570
7571         spin_lock(&sinfo->lock);
7572
7573         for(i = 0; i < BTRFS_NR_RAID_TYPES; i++)
7574                 if (!list_empty(&sinfo->block_groups[i]))
7575                         free_bytes += __btrfs_get_ro_block_group_free_space(
7576                                                 &sinfo->block_groups[i]);
7577
7578         spin_unlock(&sinfo->lock);
7579
7580         return free_bytes;
7581 }
7582
7583 void btrfs_set_block_group_rw(struct btrfs_root *root,
7584                               struct btrfs_block_group_cache *cache)
7585 {
7586         struct btrfs_space_info *sinfo = cache->space_info;
7587         u64 num_bytes;
7588
7589         BUG_ON(!cache->ro);
7590
7591         spin_lock(&sinfo->lock);
7592         spin_lock(&cache->lock);
7593         num_bytes = cache->key.offset - cache->reserved - cache->pinned -
7594                     cache->bytes_super - btrfs_block_group_used(&cache->item);
7595         sinfo->bytes_readonly -= num_bytes;
7596         cache->ro = 0;
7597         spin_unlock(&cache->lock);
7598         spin_unlock(&sinfo->lock);
7599 }
7600
7601 /*
7602  * checks to see if its even possible to relocate this block group.
7603  *
7604  * @return - -1 if it's not a good idea to relocate this block group, 0 if its
7605  * ok to go ahead and try.
7606  */
7607 int btrfs_can_relocate(struct btrfs_root *root, u64 bytenr)
7608 {
7609         struct btrfs_block_group_cache *block_group;
7610         struct btrfs_space_info *space_info;
7611         struct btrfs_fs_devices *fs_devices = root->fs_info->fs_devices;
7612         struct btrfs_device *device;
7613         u64 min_free;
7614         u64 dev_min = 1;
7615         u64 dev_nr = 0;
7616         u64 target;
7617         int index;
7618         int full = 0;
7619         int ret = 0;
7620
7621         block_group = btrfs_lookup_block_group(root->fs_info, bytenr);
7622
7623         /* odd, couldn't find the block group, leave it alone */
7624         if (!block_group)
7625                 return -1;
7626
7627         min_free = btrfs_block_group_used(&block_group->item);
7628
7629         /* no bytes used, we're good */
7630         if (!min_free)
7631                 goto out;
7632
7633         space_info = block_group->space_info;
7634         spin_lock(&space_info->lock);
7635
7636         full = space_info->full;
7637
7638         /*
7639          * if this is the last block group we have in this space, we can't
7640          * relocate it unless we're able to allocate a new chunk below.
7641          *
7642          * Otherwise, we need to make sure we have room in the space to handle
7643          * all of the extents from this block group.  If we can, we're good
7644          */
7645         if ((space_info->total_bytes != block_group->key.offset) &&
7646             (space_info->bytes_used + space_info->bytes_reserved +
7647              space_info->bytes_pinned + space_info->bytes_readonly +
7648              min_free < space_info->total_bytes)) {
7649                 spin_unlock(&space_info->lock);
7650                 goto out;
7651         }
7652         spin_unlock(&space_info->lock);
7653
7654         /*
7655          * ok we don't have enough space, but maybe we have free space on our
7656          * devices to allocate new chunks for relocation, so loop through our
7657          * alloc devices and guess if we have enough space.  if this block
7658          * group is going to be restriped, run checks against the target
7659          * profile instead of the current one.
7660          */
7661         ret = -1;
7662
7663         /*
7664          * index:
7665          *      0: raid10
7666          *      1: raid1
7667          *      2: dup
7668          *      3: raid0
7669          *      4: single
7670          */
7671         target = get_restripe_target(root->fs_info, block_group->flags);
7672         if (target) {
7673                 index = __get_raid_index(extended_to_chunk(target));
7674         } else {
7675                 /*
7676                  * this is just a balance, so if we were marked as full
7677                  * we know there is no space for a new chunk
7678                  */
7679                 if (full)
7680                         goto out;
7681
7682                 index = get_block_group_index(block_group);
7683         }
7684
7685         if (index == BTRFS_RAID_RAID10) {
7686                 dev_min = 4;
7687                 /* Divide by 2 */
7688                 min_free >>= 1;
7689         } else if (index == BTRFS_RAID_RAID1) {
7690                 dev_min = 2;
7691         } else if (index == BTRFS_RAID_DUP) {
7692                 /* Multiply by 2 */
7693                 min_free <<= 1;
7694         } else if (index == BTRFS_RAID_RAID0) {
7695                 dev_min = fs_devices->rw_devices;
7696                 do_div(min_free, dev_min);
7697         }
7698
7699         mutex_lock(&root->fs_info->chunk_mutex);
7700         list_for_each_entry(device, &fs_devices->alloc_list, dev_alloc_list) {
7701                 u64 dev_offset;
7702
7703                 /*
7704                  * check to make sure we can actually find a chunk with enough
7705                  * space to fit our block group in.
7706                  */
7707                 if (device->total_bytes > device->bytes_used + min_free &&
7708                     !device->is_tgtdev_for_dev_replace) {
7709                         ret = find_free_dev_extent(device, min_free,
7710                                                    &dev_offset, NULL);
7711                         if (!ret)
7712                                 dev_nr++;
7713
7714                         if (dev_nr >= dev_min)
7715                                 break;
7716
7717                         ret = -1;
7718                 }
7719         }
7720         mutex_unlock(&root->fs_info->chunk_mutex);
7721 out:
7722         btrfs_put_block_group(block_group);
7723         return ret;
7724 }
7725
7726 static int find_first_block_group(struct btrfs_root *root,
7727                 struct btrfs_path *path, struct btrfs_key *key)
7728 {
7729         int ret = 0;
7730         struct btrfs_key found_key;
7731         struct extent_buffer *leaf;
7732         int slot;
7733
7734         ret = btrfs_search_slot(NULL, root, key, path, 0, 0);
7735         if (ret < 0)
7736                 goto out;
7737
7738         while (1) {
7739                 slot = path->slots[0];
7740                 leaf = path->nodes[0];
7741                 if (slot >= btrfs_header_nritems(leaf)) {
7742                         ret = btrfs_next_leaf(root, path);
7743                         if (ret == 0)
7744                                 continue;
7745                         if (ret < 0)
7746                                 goto out;
7747                         break;
7748                 }
7749                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
7750
7751                 if (found_key.objectid >= key->objectid &&
7752                     found_key.type == BTRFS_BLOCK_GROUP_ITEM_KEY) {
7753                         ret = 0;
7754                         goto out;
7755                 }
7756                 path->slots[0]++;
7757         }
7758 out:
7759         return ret;
7760 }
7761
7762 void btrfs_put_block_group_cache(struct btrfs_fs_info *info)
7763 {
7764         struct btrfs_block_group_cache *block_group;
7765         u64 last = 0;
7766
7767         while (1) {
7768                 struct inode *inode;
7769
7770                 block_group = btrfs_lookup_first_block_group(info, last);
7771                 while (block_group) {
7772                         spin_lock(&block_group->lock);
7773                         if (block_group->iref)
7774                                 break;
7775                         spin_unlock(&block_group->lock);
7776                         block_group = next_block_group(info->tree_root,
7777                                                        block_group);
7778                 }
7779                 if (!block_group) {
7780                         if (last == 0)
7781                                 break;
7782                         last = 0;
7783                         continue;
7784                 }
7785
7786                 inode = block_group->inode;
7787                 block_group->iref = 0;
7788                 block_group->inode = NULL;
7789                 spin_unlock(&block_group->lock);
7790                 iput(inode);
7791                 last = block_group->key.objectid + block_group->key.offset;
7792                 btrfs_put_block_group(block_group);
7793         }
7794 }
7795
7796 int btrfs_free_block_groups(struct btrfs_fs_info *info)
7797 {
7798         struct btrfs_block_group_cache *block_group;
7799         struct btrfs_space_info *space_info;
7800         struct btrfs_caching_control *caching_ctl;
7801         struct rb_node *n;
7802
7803         down_write(&info->extent_commit_sem);
7804         while (!list_empty(&info->caching_block_groups)) {
7805                 caching_ctl = list_entry(info->caching_block_groups.next,
7806                                          struct btrfs_caching_control, list);
7807                 list_del(&caching_ctl->list);
7808                 put_caching_control(caching_ctl);
7809         }
7810         up_write(&info->extent_commit_sem);
7811
7812         spin_lock(&info->block_group_cache_lock);
7813         while ((n = rb_last(&info->block_group_cache_tree)) != NULL) {
7814                 block_group = rb_entry(n, struct btrfs_block_group_cache,
7815                                        cache_node);
7816                 rb_erase(&block_group->cache_node,
7817                          &info->block_group_cache_tree);
7818                 spin_unlock(&info->block_group_cache_lock);
7819
7820                 down_write(&block_group->space_info->groups_sem);
7821                 list_del(&block_group->list);
7822                 up_write(&block_group->space_info->groups_sem);
7823
7824                 if (block_group->cached == BTRFS_CACHE_STARTED)
7825                         wait_block_group_cache_done(block_group);
7826
7827                 /*
7828                  * We haven't cached this block group, which means we could
7829                  * possibly have excluded extents on this block group.
7830                  */
7831                 if (block_group->cached == BTRFS_CACHE_NO)
7832                         free_excluded_extents(info->extent_root, block_group);
7833
7834                 btrfs_remove_free_space_cache(block_group);
7835                 btrfs_put_block_group(block_group);
7836
7837                 spin_lock(&info->block_group_cache_lock);
7838         }
7839         spin_unlock(&info->block_group_cache_lock);
7840
7841         /* now that all the block groups are freed, go through and
7842          * free all the space_info structs.  This is only called during
7843          * the final stages of unmount, and so we know nobody is
7844          * using them.  We call synchronize_rcu() once before we start,
7845          * just to be on the safe side.
7846          */
7847         synchronize_rcu();
7848
7849         release_global_block_rsv(info);
7850
7851         while(!list_empty(&info->space_info)) {
7852                 space_info = list_entry(info->space_info.next,
7853                                         struct btrfs_space_info,
7854                                         list);
7855                 if (btrfs_test_opt(info->tree_root, ENOSPC_DEBUG)) {
7856                         if (space_info->bytes_pinned > 0 ||
7857                             space_info->bytes_reserved > 0 ||
7858                             space_info->bytes_may_use > 0) {
7859                                 WARN_ON(1);
7860                                 dump_space_info(space_info, 0, 0);
7861                         }
7862                 }
7863                 list_del(&space_info->list);
7864                 kfree(space_info);
7865         }
7866         return 0;
7867 }
7868
7869 static void __link_block_group(struct btrfs_space_info *space_info,
7870                                struct btrfs_block_group_cache *cache)
7871 {
7872         int index = get_block_group_index(cache);
7873
7874         down_write(&space_info->groups_sem);
7875         list_add_tail(&cache->list, &space_info->block_groups[index]);
7876         up_write(&space_info->groups_sem);
7877 }
7878
7879 int btrfs_read_block_groups(struct btrfs_root *root)
7880 {
7881         struct btrfs_path *path;
7882         int ret;
7883         struct btrfs_block_group_cache *cache;
7884         struct btrfs_fs_info *info = root->fs_info;
7885         struct btrfs_space_info *space_info;
7886         struct btrfs_key key;
7887         struct btrfs_key found_key;
7888         struct extent_buffer *leaf;
7889         int need_clear = 0;
7890         u64 cache_gen;
7891
7892         root = info->extent_root;
7893         key.objectid = 0;
7894         key.offset = 0;
7895         btrfs_set_key_type(&key, BTRFS_BLOCK_GROUP_ITEM_KEY);
7896         path = btrfs_alloc_path();
7897         if (!path)
7898                 return -ENOMEM;
7899         path->reada = 1;
7900
7901         cache_gen = btrfs_super_cache_generation(root->fs_info->super_copy);
7902         if (btrfs_test_opt(root, SPACE_CACHE) &&
7903             btrfs_super_generation(root->fs_info->super_copy) != cache_gen)
7904                 need_clear = 1;
7905         if (btrfs_test_opt(root, CLEAR_CACHE))
7906                 need_clear = 1;
7907
7908         while (1) {
7909                 ret = find_first_block_group(root, path, &key);
7910                 if (ret > 0)
7911                         break;
7912                 if (ret != 0)
7913                         goto error;
7914                 leaf = path->nodes[0];
7915                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
7916                 cache = kzalloc(sizeof(*cache), GFP_NOFS);
7917                 if (!cache) {
7918                         ret = -ENOMEM;
7919                         goto error;
7920                 }
7921                 cache->free_space_ctl = kzalloc(sizeof(*cache->free_space_ctl),
7922                                                 GFP_NOFS);
7923                 if (!cache->free_space_ctl) {
7924                         kfree(cache);
7925                         ret = -ENOMEM;
7926                         goto error;
7927                 }
7928
7929                 atomic_set(&cache->count, 1);
7930                 spin_lock_init(&cache->lock);
7931                 cache->fs_info = info;
7932                 INIT_LIST_HEAD(&cache->list);
7933                 INIT_LIST_HEAD(&cache->cluster_list);
7934
7935                 if (need_clear) {
7936                         /*
7937                          * When we mount with old space cache, we need to
7938                          * set BTRFS_DC_CLEAR and set dirty flag.
7939                          *
7940                          * a) Setting 'BTRFS_DC_CLEAR' makes sure that we
7941                          *    truncate the old free space cache inode and
7942                          *    setup a new one.
7943                          * b) Setting 'dirty flag' makes sure that we flush
7944                          *    the new space cache info onto disk.
7945                          */
7946                         cache->disk_cache_state = BTRFS_DC_CLEAR;
7947                         if (btrfs_test_opt(root, SPACE_CACHE))
7948                                 cache->dirty = 1;
7949                 }
7950
7951                 read_extent_buffer(leaf, &cache->item,
7952                                    btrfs_item_ptr_offset(leaf, path->slots[0]),
7953                                    sizeof(cache->item));
7954                 memcpy(&cache->key, &found_key, sizeof(found_key));
7955
7956                 key.objectid = found_key.objectid + found_key.offset;
7957                 btrfs_release_path(path);
7958                 cache->flags = btrfs_block_group_flags(&cache->item);
7959                 cache->sectorsize = root->sectorsize;
7960                 cache->full_stripe_len = btrfs_full_stripe_len(root,
7961                                                &root->fs_info->mapping_tree,
7962                                                found_key.objectid);
7963                 btrfs_init_free_space_ctl(cache);
7964
7965                 /*
7966                  * We need to exclude the super stripes now so that the space
7967                  * info has super bytes accounted for, otherwise we'll think
7968                  * we have more space than we actually do.
7969                  */
7970                 exclude_super_stripes(root, cache);
7971
7972                 /*
7973                  * check for two cases, either we are full, and therefore
7974                  * don't need to bother with the caching work since we won't
7975                  * find any space, or we are empty, and we can just add all
7976                  * the space in and be done with it.  This saves us _alot_ of
7977                  * time, particularly in the full case.
7978                  */
7979                 if (found_key.offset == btrfs_block_group_used(&cache->item)) {
7980                         cache->last_byte_to_unpin = (u64)-1;
7981                         cache->cached = BTRFS_CACHE_FINISHED;
7982                         free_excluded_extents(root, cache);
7983                 } else if (btrfs_block_group_used(&cache->item) == 0) {
7984                         cache->last_byte_to_unpin = (u64)-1;
7985                         cache->cached = BTRFS_CACHE_FINISHED;
7986                         add_new_free_space(cache, root->fs_info,
7987                                            found_key.objectid,
7988                                            found_key.objectid +
7989                                            found_key.offset);
7990                         free_excluded_extents(root, cache);
7991                 }
7992
7993                 ret = update_space_info(info, cache->flags, found_key.offset,
7994                                         btrfs_block_group_used(&cache->item),
7995                                         &space_info);
7996                 BUG_ON(ret); /* -ENOMEM */
7997                 cache->space_info = space_info;
7998                 spin_lock(&cache->space_info->lock);
7999                 cache->space_info->bytes_readonly += cache->bytes_super;
8000                 spin_unlock(&cache->space_info->lock);
8001
8002                 __link_block_group(space_info, cache);
8003
8004                 ret = btrfs_add_block_group_cache(root->fs_info, cache);
8005                 BUG_ON(ret); /* Logic error */
8006
8007                 set_avail_alloc_bits(root->fs_info, cache->flags);
8008                 if (btrfs_chunk_readonly(root, cache->key.objectid))
8009                         set_block_group_ro(cache, 1);
8010         }
8011
8012         list_for_each_entry_rcu(space_info, &root->fs_info->space_info, list) {
8013                 if (!(get_alloc_profile(root, space_info->flags) &
8014                       (BTRFS_BLOCK_GROUP_RAID10 |
8015                        BTRFS_BLOCK_GROUP_RAID1 |
8016                        BTRFS_BLOCK_GROUP_RAID5 |
8017                        BTRFS_BLOCK_GROUP_RAID6 |
8018                        BTRFS_BLOCK_GROUP_DUP)))
8019                         continue;
8020                 /*
8021                  * avoid allocating from un-mirrored block group if there are
8022                  * mirrored block groups.
8023                  */
8024                 list_for_each_entry(cache, &space_info->block_groups[3], list)
8025                         set_block_group_ro(cache, 1);
8026                 list_for_each_entry(cache, &space_info->block_groups[4], list)
8027                         set_block_group_ro(cache, 1);
8028         }
8029
8030         init_global_block_rsv(info);
8031         ret = 0;
8032 error:
8033         btrfs_free_path(path);
8034         return ret;
8035 }
8036
8037 void btrfs_create_pending_block_groups(struct btrfs_trans_handle *trans,
8038                                        struct btrfs_root *root)
8039 {
8040         struct btrfs_block_group_cache *block_group, *tmp;
8041         struct btrfs_root *extent_root = root->fs_info->extent_root;
8042         struct btrfs_block_group_item item;
8043         struct btrfs_key key;
8044         int ret = 0;
8045
8046         list_for_each_entry_safe(block_group, tmp, &trans->new_bgs,
8047                                  new_bg_list) {
8048                 list_del_init(&block_group->new_bg_list);
8049
8050                 if (ret)
8051                         continue;
8052
8053                 spin_lock(&block_group->lock);
8054                 memcpy(&item, &block_group->item, sizeof(item));
8055                 memcpy(&key, &block_group->key, sizeof(key));
8056                 spin_unlock(&block_group->lock);
8057
8058                 ret = btrfs_insert_item(trans, extent_root, &key, &item,
8059                                         sizeof(item));
8060                 if (ret)
8061                         btrfs_abort_transaction(trans, extent_root, ret);
8062         }
8063 }
8064
8065 int btrfs_make_block_group(struct btrfs_trans_handle *trans,
8066                            struct btrfs_root *root, u64 bytes_used,
8067                            u64 type, u64 chunk_objectid, u64 chunk_offset,
8068                            u64 size)
8069 {
8070         int ret;
8071         struct btrfs_root *extent_root;
8072         struct btrfs_block_group_cache *cache;
8073
8074         extent_root = root->fs_info->extent_root;
8075
8076         root->fs_info->last_trans_log_full_commit = trans->transid;
8077
8078         cache = kzalloc(sizeof(*cache), GFP_NOFS);
8079         if (!cache)
8080                 return -ENOMEM;
8081         cache->free_space_ctl = kzalloc(sizeof(*cache->free_space_ctl),
8082                                         GFP_NOFS);
8083         if (!cache->free_space_ctl) {
8084                 kfree(cache);
8085                 return -ENOMEM;
8086         }
8087
8088         cache->key.objectid = chunk_offset;
8089         cache->key.offset = size;
8090         cache->key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
8091         cache->sectorsize = root->sectorsize;
8092         cache->fs_info = root->fs_info;
8093         cache->full_stripe_len = btrfs_full_stripe_len(root,
8094                                                &root->fs_info->mapping_tree,
8095                                                chunk_offset);
8096
8097         atomic_set(&cache->count, 1);
8098         spin_lock_init(&cache->lock);
8099         INIT_LIST_HEAD(&cache->list);
8100         INIT_LIST_HEAD(&cache->cluster_list);
8101         INIT_LIST_HEAD(&cache->new_bg_list);
8102
8103         btrfs_init_free_space_ctl(cache);
8104
8105         btrfs_set_block_group_used(&cache->item, bytes_used);
8106         btrfs_set_block_group_chunk_objectid(&cache->item, chunk_objectid);
8107         cache->flags = type;
8108         btrfs_set_block_group_flags(&cache->item, type);
8109
8110         cache->last_byte_to_unpin = (u64)-1;
8111         cache->cached = BTRFS_CACHE_FINISHED;
8112         exclude_super_stripes(root, cache);
8113
8114         add_new_free_space(cache, root->fs_info, chunk_offset,
8115                            chunk_offset + size);
8116
8117         free_excluded_extents(root, cache);
8118
8119         ret = update_space_info(root->fs_info, cache->flags, size, bytes_used,
8120                                 &cache->space_info);
8121         BUG_ON(ret); /* -ENOMEM */
8122         update_global_block_rsv(root->fs_info);
8123
8124         spin_lock(&cache->space_info->lock);
8125         cache->space_info->bytes_readonly += cache->bytes_super;
8126         spin_unlock(&cache->space_info->lock);
8127
8128         __link_block_group(cache->space_info, cache);
8129
8130         ret = btrfs_add_block_group_cache(root->fs_info, cache);
8131         BUG_ON(ret); /* Logic error */
8132
8133         list_add_tail(&cache->new_bg_list, &trans->new_bgs);
8134
8135         set_avail_alloc_bits(extent_root->fs_info, type);
8136
8137         return 0;
8138 }
8139
8140 static void clear_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
8141 {
8142         u64 extra_flags = chunk_to_extended(flags) &
8143                                 BTRFS_EXTENDED_PROFILE_MASK;
8144
8145         write_seqlock(&fs_info->profiles_lock);
8146         if (flags & BTRFS_BLOCK_GROUP_DATA)
8147                 fs_info->avail_data_alloc_bits &= ~extra_flags;
8148         if (flags & BTRFS_BLOCK_GROUP_METADATA)
8149                 fs_info->avail_metadata_alloc_bits &= ~extra_flags;
8150         if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
8151                 fs_info->avail_system_alloc_bits &= ~extra_flags;
8152         write_sequnlock(&fs_info->profiles_lock);
8153 }
8154
8155 int btrfs_remove_block_group(struct btrfs_trans_handle *trans,
8156                              struct btrfs_root *root, u64 group_start)
8157 {
8158         struct btrfs_path *path;
8159         struct btrfs_block_group_cache *block_group;
8160         struct btrfs_free_cluster *cluster;
8161         struct btrfs_root *tree_root = root->fs_info->tree_root;
8162         struct btrfs_key key;
8163         struct inode *inode;
8164         int ret;
8165         int index;
8166         int factor;
8167
8168         root = root->fs_info->extent_root;
8169
8170         block_group = btrfs_lookup_block_group(root->fs_info, group_start);
8171         BUG_ON(!block_group);
8172         BUG_ON(!block_group->ro);
8173
8174         /*
8175          * Free the reserved super bytes from this block group before
8176          * remove it.
8177          */
8178         free_excluded_extents(root, block_group);
8179
8180         memcpy(&key, &block_group->key, sizeof(key));
8181         index = get_block_group_index(block_group);
8182         if (block_group->flags & (BTRFS_BLOCK_GROUP_DUP |
8183                                   BTRFS_BLOCK_GROUP_RAID1 |
8184                                   BTRFS_BLOCK_GROUP_RAID10))
8185                 factor = 2;
8186         else
8187                 factor = 1;
8188
8189         /* make sure this block group isn't part of an allocation cluster */
8190         cluster = &root->fs_info->data_alloc_cluster;
8191         spin_lock(&cluster->refill_lock);
8192         btrfs_return_cluster_to_free_space(block_group, cluster);
8193         spin_unlock(&cluster->refill_lock);
8194
8195         /*
8196          * make sure this block group isn't part of a metadata
8197          * allocation cluster
8198          */
8199         cluster = &root->fs_info->meta_alloc_cluster;
8200         spin_lock(&cluster->refill_lock);
8201         btrfs_return_cluster_to_free_space(block_group, cluster);
8202         spin_unlock(&cluster->refill_lock);
8203
8204         path = btrfs_alloc_path();
8205         if (!path) {
8206                 ret = -ENOMEM;
8207                 goto out;
8208         }
8209
8210         inode = lookup_free_space_inode(tree_root, block_group, path);
8211         if (!IS_ERR(inode)) {
8212                 ret = btrfs_orphan_add(trans, inode);
8213                 if (ret) {
8214                         btrfs_add_delayed_iput(inode);
8215                         goto out;
8216                 }
8217                 clear_nlink(inode);
8218                 /* One for the block groups ref */
8219                 spin_lock(&block_group->lock);
8220                 if (block_group->iref) {
8221                         block_group->iref = 0;
8222                         block_group->inode = NULL;
8223                         spin_unlock(&block_group->lock);
8224                         iput(inode);
8225                 } else {
8226                         spin_unlock(&block_group->lock);
8227                 }
8228                 /* One for our lookup ref */
8229                 btrfs_add_delayed_iput(inode);
8230         }
8231
8232         key.objectid = BTRFS_FREE_SPACE_OBJECTID;
8233         key.offset = block_group->key.objectid;
8234         key.type = 0;
8235
8236         ret = btrfs_search_slot(trans, tree_root, &key, path, -1, 1);
8237         if (ret < 0)
8238                 goto out;
8239         if (ret > 0)
8240                 btrfs_release_path(path);
8241         if (ret == 0) {
8242                 ret = btrfs_del_item(trans, tree_root, path);
8243                 if (ret)
8244                         goto out;
8245                 btrfs_release_path(path);
8246         }
8247
8248         spin_lock(&root->fs_info->block_group_cache_lock);
8249         rb_erase(&block_group->cache_node,
8250                  &root->fs_info->block_group_cache_tree);
8251
8252         if (root->fs_info->first_logical_byte == block_group->key.objectid)
8253                 root->fs_info->first_logical_byte = (u64)-1;
8254         spin_unlock(&root->fs_info->block_group_cache_lock);
8255
8256         down_write(&block_group->space_info->groups_sem);
8257         /*
8258          * we must use list_del_init so people can check to see if they
8259          * are still on the list after taking the semaphore
8260          */
8261         list_del_init(&block_group->list);
8262         if (list_empty(&block_group->space_info->block_groups[index]))
8263                 clear_avail_alloc_bits(root->fs_info, block_group->flags);
8264         up_write(&block_group->space_info->groups_sem);
8265
8266         if (block_group->cached == BTRFS_CACHE_STARTED)
8267                 wait_block_group_cache_done(block_group);
8268
8269         btrfs_remove_free_space_cache(block_group);
8270
8271         spin_lock(&block_group->space_info->lock);
8272         block_group->space_info->total_bytes -= block_group->key.offset;
8273         block_group->space_info->bytes_readonly -= block_group->key.offset;
8274         block_group->space_info->disk_total -= block_group->key.offset * factor;
8275         spin_unlock(&block_group->space_info->lock);
8276
8277         memcpy(&key, &block_group->key, sizeof(key));
8278
8279         btrfs_clear_space_info_full(root->fs_info);
8280
8281         btrfs_put_block_group(block_group);
8282         btrfs_put_block_group(block_group);
8283
8284         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
8285         if (ret > 0)
8286                 ret = -EIO;
8287         if (ret < 0)
8288                 goto out;
8289
8290         ret = btrfs_del_item(trans, root, path);
8291 out:
8292         btrfs_free_path(path);
8293         return ret;
8294 }
8295
8296 int btrfs_init_space_info(struct btrfs_fs_info *fs_info)
8297 {
8298         struct btrfs_space_info *space_info;
8299         struct btrfs_super_block *disk_super;
8300         u64 features;
8301         u64 flags;
8302         int mixed = 0;
8303         int ret;
8304
8305         disk_super = fs_info->super_copy;
8306         if (!btrfs_super_root(disk_super))
8307                 return 1;
8308
8309         features = btrfs_super_incompat_flags(disk_super);
8310         if (features & BTRFS_FEATURE_INCOMPAT_MIXED_GROUPS)
8311                 mixed = 1;
8312
8313         flags = BTRFS_BLOCK_GROUP_SYSTEM;
8314         ret = update_space_info(fs_info, flags, 0, 0, &space_info);
8315         if (ret)
8316                 goto out;
8317
8318         if (mixed) {
8319                 flags = BTRFS_BLOCK_GROUP_METADATA | BTRFS_BLOCK_GROUP_DATA;
8320                 ret = update_space_info(fs_info, flags, 0, 0, &space_info);
8321         } else {
8322                 flags = BTRFS_BLOCK_GROUP_METADATA;
8323                 ret = update_space_info(fs_info, flags, 0, 0, &space_info);
8324                 if (ret)
8325                         goto out;
8326
8327                 flags = BTRFS_BLOCK_GROUP_DATA;
8328                 ret = update_space_info(fs_info, flags, 0, 0, &space_info);
8329         }
8330 out:
8331         return ret;
8332 }
8333
8334 int btrfs_error_unpin_extent_range(struct btrfs_root *root, u64 start, u64 end)
8335 {
8336         return unpin_extent_range(root, start, end);
8337 }
8338
8339 int btrfs_error_discard_extent(struct btrfs_root *root, u64 bytenr,
8340                                u64 num_bytes, u64 *actual_bytes)
8341 {
8342         return btrfs_discard_extent(root, bytenr, num_bytes, actual_bytes);
8343 }
8344
8345 int btrfs_trim_fs(struct btrfs_root *root, struct fstrim_range *range)
8346 {
8347         struct btrfs_fs_info *fs_info = root->fs_info;
8348         struct btrfs_block_group_cache *cache = NULL;
8349         u64 group_trimmed;
8350         u64 start;
8351         u64 end;
8352         u64 trimmed = 0;
8353         u64 total_bytes = btrfs_super_total_bytes(fs_info->super_copy);
8354         int ret = 0;
8355
8356         /*
8357          * try to trim all FS space, our block group may start from non-zero.
8358          */
8359         if (range->len == total_bytes)
8360                 cache = btrfs_lookup_first_block_group(fs_info, range->start);
8361         else
8362                 cache = btrfs_lookup_block_group(fs_info, range->start);
8363
8364         while (cache) {
8365                 if (cache->key.objectid >= (range->start + range->len)) {
8366                         btrfs_put_block_group(cache);
8367                         break;
8368                 }
8369
8370                 start = max(range->start, cache->key.objectid);
8371                 end = min(range->start + range->len,
8372                                 cache->key.objectid + cache->key.offset);
8373
8374                 if (end - start >= range->minlen) {
8375                         if (!block_group_cache_done(cache)) {
8376                                 ret = cache_block_group(cache, 0);
8377                                 if (!ret)
8378                                         wait_block_group_cache_done(cache);
8379                         }
8380                         ret = btrfs_trim_block_group(cache,
8381                                                      &group_trimmed,
8382                                                      start,
8383                                                      end,
8384                                                      range->minlen);
8385
8386                         trimmed += group_trimmed;
8387                         if (ret) {
8388                                 btrfs_put_block_group(cache);
8389                                 break;
8390                         }
8391                 }
8392
8393                 cache = next_block_group(fs_info->tree_root, cache);
8394         }
8395
8396         range->len = trimmed;
8397         return ret;
8398 }