fs/btrfs/extent-tree.c

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