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 <linux/percpu_counter.h>
  28 #include "compat.h"
  29 #include "hash.h"
  30 #include "ctree.h"
  31 #include "disk-io.h"
  32 #include "print-tree.h"
  33 #include "transaction.h"
  34 #include "volumes.h"
  35 #include "raid56.h"
  36 #include "locking.h"
  37 #include "free-space-cache.h"
  38 #include "math.h"
  39
  40 #undef SCRAMBLE_DELAYED_REFS
  41
  42 /*
  43  * control flags for do_chunk_alloc's force field
  44  * CHUNK_ALLOC_NO_FORCE means to only allocate a chunk
  45  * if we really need one.
  46  *
  47  * CHUNK_ALLOC_LIMITED means to only try and allocate one
  48  * if we have very few chunks already allocated.  This is
  49  * used as part of the clustering code to help make sure
  50  * we have a good pool of storage to cluster in, without
  51  * filling the FS with empty chunks
  52  *
  53  * CHUNK_ALLOC_FORCE means it must try to allocate one
  54  *
  55  */
  56 enum {
  57         CHUNK_ALLOC_NO_FORCE = 0,
  58         CHUNK_ALLOC_LIMITED = 1,
  59         CHUNK_ALLOC_FORCE = 2,
  60 };
  61
  62 /*
  63  * Control how reservations are dealt with.
  64  *
  65  * RESERVE_FREE - freeing a reservation.
  66  * RESERVE_ALLOC - allocating space and we need to update bytes_may_use for
  67  *   ENOSPC accounting
  68  * RESERVE_ALLOC_NO_ACCOUNT - allocating space and we should not update
  69  *   bytes_may_use as the ENOSPC accounting is done elsewhere
  70  */
  71 enum {
  72         RESERVE_FREE = 0,
  73         RESERVE_ALLOC = 1,
  74         RESERVE_ALLOC_NO_ACCOUNT = 2,
  75 };
  76
  77 static int update_block_group(struct btrfs_root *root,
  78                               u64 bytenr, u64 num_bytes, int alloc);
  79 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
  80                                 struct btrfs_root *root,
  81                                 u64 bytenr, u64 num_bytes, u64 parent,
  82                                 u64 root_objectid, u64 owner_objectid,
  83                                 u64 owner_offset, int refs_to_drop,
  84                                 struct btrfs_delayed_extent_op *extra_op);
  85 static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op,
  86                                     struct extent_buffer *leaf,
  87                                     struct btrfs_extent_item *ei);
  88 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
  89                                       struct btrfs_root *root,
  90                                       u64 parent, u64 root_objectid,
  91                                       u64 flags, u64 owner, u64 offset,
  92                                       struct btrfs_key *ins, int ref_mod);
  93 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
  94                                      struct btrfs_root *root,
  95                                      u64 parent, u64 root_objectid,
  96                                      u64 flags, struct btrfs_disk_key *key,
  97                                      int level, struct btrfs_key *ins);
  98 static int do_chunk_alloc(struct btrfs_trans_handle *trans,
  99                           struct btrfs_root *extent_root, u64 flags,
 100                           int force);
 101 static int find_next_key(struct btrfs_path *path, int level,
 102                          struct btrfs_key *key);
 103 static void dump_space_info(struct btrfs_space_info *info, u64 bytes,
 104                             int dump_block_groups);
 105 static int btrfs_update_reserved_bytes(struct btrfs_block_group_cache *cache,
 106                                        u64 num_bytes, int reserve);
 107 static int block_rsv_use_bytes(struct btrfs_block_rsv *block_rsv,
 108                                u64 num_bytes);
 109 int btrfs_pin_extent(struct btrfs_root *root,
 110                      u64 bytenr, u64 num_bytes, int reserved);
 111
 112 static noinline int
 113 block_group_cache_done(struct btrfs_block_group_cache *cache)
 114 {
 115         smp_mb();
 116         return cache->cached == BTRFS_CACHE_FINISHED;
 117 }
 118
 119 static int block_group_bits(struct btrfs_block_group_cache *cache, u64 bits)
 120 {
 121         return (cache->flags & bits) == bits;
 122 }
 123
 124 static void btrfs_get_block_group(struct btrfs_block_group_cache *cache)
 125 {
 126         atomic_inc(&cache->count);
 127 }
 128
 129 void btrfs_put_block_group(struct btrfs_block_group_cache *cache)
 130 {
 131         if (atomic_dec_and_test(&cache->count)) {
 132                 WARN_ON(cache->pinned > 0);
 133                 WARN_ON(cache->reserved > 0);
 134                 kfree(cache->free_space_ctl);
 135                 kfree(cache);
 136         }
 137 }
 138
 139 /*
 140  * this adds the block group to the fs_info rb tree for the block group
 141  * cache
 142  */
 143 static int btrfs_add_block_group_cache(struct btrfs_fs_info *info,
 144                                 struct btrfs_block_group_cache *block_group)
 145 {
 146         struct rb_node **p;
 147         struct rb_node *parent = NULL;
 148         struct btrfs_block_group_cache *cache;
 149
 150         spin_lock(&info->block_group_cache_lock);
 151         p = &info->block_group_cache_tree.rb_node;
 152
 153         while (*p) {
 154                 parent = *p;
 155                 cache = rb_entry(parent, struct btrfs_block_group_cache,
 156                                  cache_node);
 157                 if (block_group->key.objectid < cache->key.objectid) {
 158                         p = &(*p)->rb_left;
 159                 } else if (block_group->key.objectid > cache->key.objectid) {
 160                         p = &(*p)->rb_right;
 161                 } else {
 162                         spin_unlock(&info->block_group_cache_lock);
 163                         return -EEXIST;
 164                 }
 165         }
 166
 167         rb_link_node(&block_group->cache_node, parent, p);
 168         rb_insert_color(&block_group->cache_node,
 169                         &info->block_group_cache_tree);
 170
 171         if (info->first_logical_byte > block_group->key.objectid)
 172                 info->first_logical_byte = block_group->key.objectid;
 173
 174         spin_unlock(&info->block_group_cache_lock);
 175
 176         return 0;
 177 }
 178
 179 /*
 180  * This will return the block group at or after bytenr if contains is 0, else
 181  * it will return the block group that contains the bytenr
 182  */
 183 static struct btrfs_block_group_cache *
 184 block_group_cache_tree_search(struct btrfs_fs_info *info, u64 bytenr,
 185                               int contains)
 186 {
 187         struct btrfs_block_group_cache *cache, *ret = NULL;
 188         struct rb_node *n;
 189         u64 end, start;
 190
 191         spin_lock(&info->block_group_cache_lock);
 192         n = info->block_group_cache_tree.rb_node;
 193
 194         while (n) {
 195                 cache = rb_entry(n, struct btrfs_block_group_cache,
 196                                  cache_node);
 197                 end = cache->key.objectid + cache->key.offset - 1;
 198                 start = cache->key.objectid;
 199
 200                 if (bytenr < start) {
 201                         if (!contains && (!ret || start < ret->key.objectid))
 202                                 ret = cache;
 203                         n = n->rb_left;
 204                 } else if (bytenr > start) {
 205                         if (contains && bytenr <= end) {
 206                                 ret = cache;
 207                                 break;
 208                         }
 209                         n = n->rb_right;
 210                 } else {
 211                         ret = cache;
 212                         break;
 213                 }
 214         }
 215         if (ret) {
 216                 btrfs_get_block_group(ret);
 217                 if (bytenr == 0 && info->first_logical_byte > ret->key.objectid)
 218                         info->first_logical_byte = ret->key.objectid;
 219         }
 220         spin_unlock(&info->block_group_cache_lock);
 221
 222         return ret;
 223 }
 224
 225 static int add_excluded_extent(struct btrfs_root *root,
 226                                u64 start, u64 num_bytes)
 227 {
 228         u64 end = start + num_bytes - 1;
 229         set_extent_bits(&root->fs_info->freed_extents[0],
 230                         start, end, EXTENT_UPTODATE, GFP_NOFS);
 231         set_extent_bits(&root->fs_info->freed_extents[1],
 232                         start, end, EXTENT_UPTODATE, GFP_NOFS);
 233         return 0;
 234 }
 235
 236 static void free_excluded_extents(struct btrfs_root *root,
 237                                   struct btrfs_block_group_cache *cache)
 238 {
 239         u64 start, end;
 240
 241         start = cache->key.objectid;
 242         end = start + cache->key.offset - 1;
 243
 244         clear_extent_bits(&root->fs_info->freed_extents[0],
 245                           start, end, EXTENT_UPTODATE, GFP_NOFS);
 246         clear_extent_bits(&root->fs_info->freed_extents[1],
 247                           start, end, EXTENT_UPTODATE, GFP_NOFS);
 248 }
 249
 250 static int exclude_super_stripes(struct btrfs_root *root,
 251                                  struct btrfs_block_group_cache *cache)
 252 {
 253         u64 bytenr;
 254         u64 *logical;
 255         int stripe_len;
 256         int i, nr, ret;
 257
 258         if (cache->key.objectid < BTRFS_SUPER_INFO_OFFSET) {
 259                 stripe_len = BTRFS_SUPER_INFO_OFFSET - cache->key.objectid;
 260                 cache->bytes_super += stripe_len;
 261                 ret = add_excluded_extent(root, cache->key.objectid,
 262                                           stripe_len);
 263                 if (ret)
 264                         return ret;
 265         }
 266
 267         for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
 268                 bytenr = btrfs_sb_offset(i);
 269                 ret = btrfs_rmap_block(&root->fs_info->mapping_tree,
 270                                        cache->key.objectid, bytenr,
 271                                        0, &logical, &nr, &stripe_len);
 272                 if (ret)
 273                         return ret;
 274
 275                 while (nr--) {
 276                         u64 start, len;
 277
 278                         if (logical[nr] > cache->key.objectid +
 279                             cache->key.offset)
 280                                 continue;
 281
 282                         if (logical[nr] + stripe_len <= cache->key.objectid)
 283                                 continue;
 284
 285                         start = logical[nr];
 286                         if (start < cache->key.objectid) {
 287                                 start = cache->key.objectid;
 288                                 len = (logical[nr] + stripe_len) - start;
 289                         } else {
 290                                 len = min_t(u64, stripe_len,
 291                                             cache->key.objectid +
 292                                             cache->key.offset - start);
 293                         }
 294
 295                         cache->bytes_super += len;
 296                         ret = add_excluded_extent(root, start, len);
 297                         if (ret) {
 298                                 kfree(logical);
 299                                 return ret;
 300                         }
 301                 }
 302
 303                 kfree(logical);
 304         }
 305         return 0;
 306 }
 307
 308 static struct btrfs_caching_control *
 309 get_caching_control(struct btrfs_block_group_cache *cache)
 310 {
 311         struct btrfs_caching_control *ctl;
 312
 313         spin_lock(&cache->lock);
 314         if (cache->cached != BTRFS_CACHE_STARTED) {
 315                 spin_unlock(&cache->lock);
 316                 return NULL;
 317         }
 318
 319         /* We're loading it the fast way, so we don't have a caching_ctl. */
 320         if (!cache->caching_ctl) {
 321                 spin_unlock(&cache->lock);
 322                 return NULL;
 323         }
 324
 325         ctl = cache->caching_ctl;
 326         atomic_inc(&ctl->count);
 327         spin_unlock(&cache->lock);
 328         return ctl;
 329 }
 330
 331 static void put_caching_control(struct btrfs_caching_control *ctl)
 332 {
 333         if (atomic_dec_and_test(&ctl->count))
 334                 kfree(ctl);
 335 }
 336
 337 /*
 338  * this is only called by cache_block_group, since we could have freed extents
 339  * we need to check the pinned_extents for any extents that can't be used yet
 340  * since their free space will be released as soon as the transaction commits.
 341  */
 342 static u64 add_new_free_space(struct btrfs_block_group_cache *block_group,
 343                               struct btrfs_fs_info *info, u64 start, u64 end)
 344 {
 345         u64 extent_start, extent_end, size, total_added = 0;
 346         int ret;
 347
 348         while (start < end) {
 349                 ret = find_first_extent_bit(info->pinned_extents, start,
 350                                             &extent_start, &extent_end,
 351                                             EXTENT_DIRTY | EXTENT_UPTODATE,
 352                                             NULL);
 353                 if (ret)
 354                         break;
 355
 356                 if (extent_start <= start) {
 357                         start = extent_end + 1;
 358                 } else if (extent_start > start && extent_start < end) {
 359                         size = extent_start - start;
 360                         total_added += size;
 361                         ret = btrfs_add_free_space(block_group, start,
 362                                                    size);
 363                         BUG_ON(ret); /* -ENOMEM or logic error */
 364                         start = extent_end + 1;
 365                 } else {
 366                         break;
 367                 }
 368         }
 369
 370         if (start < end) {
 371                 size = end - start;
 372                 total_added += size;
 373                 ret = btrfs_add_free_space(block_group, start, size);
 374                 BUG_ON(ret); /* -ENOMEM or logic error */
 375         }
 376
 377         return total_added;
 378 }
 379
 380 static noinline void caching_thread(struct btrfs_work *work)
 381 {
 382         struct btrfs_block_group_cache *block_group;
 383         struct btrfs_fs_info *fs_info;
 384         struct btrfs_caching_control *caching_ctl;
 385         struct btrfs_root *extent_root;
 386         struct btrfs_path *path;
 387         struct extent_buffer *leaf;
 388         struct btrfs_key key;
 389         u64 total_found = 0;
 390         u64 last = 0;
 391         u32 nritems;
 392         int ret = 0;
 393
 394         caching_ctl = container_of(work, struct btrfs_caching_control, work);
 395         block_group = caching_ctl->block_group;
 396         fs_info = block_group->fs_info;
 397         extent_root = fs_info->extent_root;
 398
 399         path = btrfs_alloc_path();
 400         if (!path)
 401                 goto out;
 402
 403         last = max_t(u64, block_group->key.objectid, BTRFS_SUPER_INFO_OFFSET);
 404
 405         /*
 406          * We don't want to deadlock with somebody trying to allocate a new
 407          * extent for the extent root while also trying to search the extent
 408          * root to add free space.  So we skip locking and search the commit
 409          * root, since its read-only
 410          */
 411         path->skip_locking = 1;
 412         path->search_commit_root = 1;
 413         path->reada = 1;
 414
 415         key.objectid = last;
 416         key.offset = 0;
 417         key.type = BTRFS_EXTENT_ITEM_KEY;
 418 again:
 419         mutex_lock(&caching_ctl->mutex);
 420         /* need to make sure the commit_root doesn't disappear */
 421         down_read(&fs_info->extent_commit_sem);
 422
 423         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
 424         if (ret < 0)
 425                 goto err;
 426
 427         leaf = path->nodes[0];
 428         nritems = btrfs_header_nritems(leaf);
 429
 430         while (1) {
 431                 if (btrfs_fs_closing(fs_info) > 1) {
 432                         last = (u64)-1;
 433                         break;
 434                 }
 435
 436                 if (path->slots[0] < nritems) {
 437                         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
 438                 } else {
 439                         ret = find_next_key(path, 0, &key);
 440                         if (ret)
 441                                 break;
 442
 443                         if (need_resched()) {
 444                                 caching_ctl->progress = last;
 445                                 btrfs_release_path(path);
 446                                 up_read(&fs_info->extent_commit_sem);
 447                                 mutex_unlock(&caching_ctl->mutex);
 448                                 cond_resched();
 449                                 goto again;
 450                         }
 451
 452                         ret = btrfs_next_leaf(extent_root, path);
 453                         if (ret < 0)
 454                                 goto err;
 455                         if (ret)
 456                                 break;
 457                         leaf = path->nodes[0];
 458                         nritems = btrfs_header_nritems(leaf);
 459                         continue;
 460                 }
 461
 462                 if (key.objectid < block_group->key.objectid) {
 463                         path->slots[0]++;
 464                         continue;
 465                 }
 466
 467                 if (key.objectid >= block_group->key.objectid +
 468                     block_group->key.offset)
 469                         break;
 470
 471                 if (key.type == BTRFS_EXTENT_ITEM_KEY ||
 472                     key.type == BTRFS_METADATA_ITEM_KEY) {
 473                         total_found += add_new_free_space(block_group,
 474                                                           fs_info, last,
 475                                                           key.objectid);
 476                         if (key.type == BTRFS_METADATA_ITEM_KEY)
 477                                 last = key.objectid +
 478                                         fs_info->tree_root->leafsize;
 479                         else
 480                                 last = key.objectid + key.offset;
 481
 482                         if (total_found > (1024 * 1024 * 2)) {
 483                                 total_found = 0;
 484                                 wake_up(&caching_ctl->wait);
 485                         }
 486                 }
 487                 path->slots[0]++;
 488         }
 489         ret = 0;
 490
 491         total_found += add_new_free_space(block_group, fs_info, last,
 492                                           block_group->key.objectid +
 493                                           block_group->key.offset);
 494         caching_ctl->progress = (u64)-1;
 495
 496         spin_lock(&block_group->lock);
 497         block_group->caching_ctl = NULL;
 498         block_group->cached = BTRFS_CACHE_FINISHED;
 499         spin_unlock(&block_group->lock);
 500
 501 err:
 502         btrfs_free_path(path);
 503         up_read(&fs_info->extent_commit_sem);
 504
 505         free_excluded_extents(extent_root, block_group);
 506
 507         mutex_unlock(&caching_ctl->mutex);
 508 out:
 509         wake_up(&caching_ctl->wait);
 510
 511         put_caching_control(caching_ctl);
 512         btrfs_put_block_group(block_group);
 513 }
 514
 515 static int cache_block_group(struct btrfs_block_group_cache *cache,
 516                              int load_cache_only)
 517 {
 518         DEFINE_WAIT(wait);
 519         struct btrfs_fs_info *fs_info = cache->fs_info;
 520         struct btrfs_caching_control *caching_ctl;
 521         int ret = 0;
 522
 523         caching_ctl = kzalloc(sizeof(*caching_ctl), GFP_NOFS);
 524         if (!caching_ctl)
 525                 return -ENOMEM;
 526
 527         INIT_LIST_HEAD(&caching_ctl->list);
 528         mutex_init(&caching_ctl->mutex);
 529         init_waitqueue_head(&caching_ctl->wait);
 530         caching_ctl->block_group = cache;
 531         caching_ctl->progress = cache->key.objectid;
 532         atomic_set(&caching_ctl->count, 1);
 533         caching_ctl->work.func = caching_thread;
 534
 535         spin_lock(&cache->lock);
 536         /*
 537          * This should be a rare occasion, but this could happen I think in the
 538          * case where one thread starts to load the space cache info, and then
 539          * some other thread starts a transaction commit which tries to do an
 540          * allocation while the other thread is still loading the space cache
 541          * info.  The previous loop should have kept us from choosing this block
 542          * group, but if we've moved to the state where we will wait on caching
 543          * block groups we need to first check if we're doing a fast load here,
 544          * so we can wait for it to finish, otherwise we could end up allocating
 545          * from a block group who's cache gets evicted for one reason or
 546          * another.
 547          */
 548         while (cache->cached == BTRFS_CACHE_FAST) {
 549                 struct btrfs_caching_control *ctl;
 550
 551                 ctl = cache->caching_ctl;
 552                 atomic_inc(&ctl->count);
 553                 prepare_to_wait(&ctl->wait, &wait, TASK_UNINTERRUPTIBLE);
 554                 spin_unlock(&cache->lock);
 555
 556                 schedule();
 557
 558                 finish_wait(&ctl->wait, &wait);
 559                 put_caching_control(ctl);
 560                 spin_lock(&cache->lock);
 561         }
 562
 563         if (cache->cached != BTRFS_CACHE_NO) {
 564                 spin_unlock(&cache->lock);
 565                 kfree(caching_ctl);
 566                 return 0;
 567         }
 568         WARN_ON(cache->caching_ctl);
 569         cache->caching_ctl = caching_ctl;
 570         cache->cached = BTRFS_CACHE_FAST;
 571         spin_unlock(&cache->lock);
 572
 573         if (fs_info->mount_opt & BTRFS_MOUNT_SPACE_CACHE) {
 574                 ret = load_free_space_cache(fs_info, cache);
 575
 576                 spin_lock(&cache->lock);
 577                 if (ret == 1) {
 578                         cache->caching_ctl = NULL;
 579                         cache->cached = BTRFS_CACHE_FINISHED;
 580                         cache->last_byte_to_unpin = (u64)-1;
 581                 } else {
 582                         if (load_cache_only) {
 583                                 cache->caching_ctl = NULL;
 584                                 cache->cached = BTRFS_CACHE_NO;
 585                         } else {
 586                                 cache->cached = BTRFS_CACHE_STARTED;
 587                         }
 588                 }
 589                 spin_unlock(&cache->lock);
 590                 wake_up(&caching_ctl->wait);
 591                 if (ret == 1) {
 592                         put_caching_control(caching_ctl);
 593                         free_excluded_extents(fs_info->extent_root, cache);
 594                         return 0;
 595                 }
 596         } else {
 597                 /*
 598                  * We are not going to do the fast caching, set cached to the
 599                  * appropriate value and wakeup any waiters.
 600                  */
 601                 spin_lock(&cache->lock);
 602                 if (load_cache_only) {
 603                         cache->caching_ctl = NULL;
 604                         cache->cached = BTRFS_CACHE_NO;
 605                 } else {
 606                         cache->cached = BTRFS_CACHE_STARTED;
 607                 }
 608                 spin_unlock(&cache->lock);
 609                 wake_up(&caching_ctl->wait);
 610         }
 611
 612         if (load_cache_only) {
 613                 put_caching_control(caching_ctl);
 614                 return 0;
 615         }
 616
 617         down_write(&fs_info->extent_commit_sem);
 618         atomic_inc(&caching_ctl->count);
 619         list_add_tail(&caching_ctl->list, &fs_info->caching_block_groups);
 620         up_write(&fs_info->extent_commit_sem);
 621
 622         btrfs_get_block_group(cache);
 623
 624         btrfs_queue_worker(&fs_info->caching_workers, &caching_ctl->work);
 625
 626         return ret;
 627 }
 628
 629 /*
 630  * return the block group that starts at or after bytenr
 631  */
 632 static struct btrfs_block_group_cache *
 633 btrfs_lookup_first_block_group(struct btrfs_fs_info *info, u64 bytenr)
 634 {
 635         struct btrfs_block_group_cache *cache;
 636
 637         cache = block_group_cache_tree_search(info, bytenr, 0);
 638
 639         return cache;
 640 }
 641
 642 /*
 643  * return the block group that contains the given bytenr
 644  */
 645 struct btrfs_block_group_cache *btrfs_lookup_block_group(
 646                                                  struct btrfs_fs_info *info,
 647                                                  u64 bytenr)
 648 {
 649         struct btrfs_block_group_cache *cache;
 650
 651         cache = block_group_cache_tree_search(info, bytenr, 1);
 652
 653         return cache;
 654 }
 655
 656 static struct btrfs_space_info *__find_space_info(struct btrfs_fs_info *info,
 657                                                   u64 flags)
 658 {
 659         struct list_head *head = &info->space_info;
 660         struct btrfs_space_info *found;
 661
 662         flags &= BTRFS_BLOCK_GROUP_TYPE_MASK;
 663
 664         rcu_read_lock();
 665         list_for_each_entry_rcu(found, head, list) {
 666                 if (found->flags & flags) {
 667                         rcu_read_unlock();
 668                         return found;
 669                 }
 670         }
 671         rcu_read_unlock();
 672         return NULL;
 673 }
 674
 675 /*
 676  * after adding space to the filesystem, we need to clear the full flags
 677  * on all the space infos.
 678  */
 679 void btrfs_clear_space_info_full(struct btrfs_fs_info *info)
 680 {
 681         struct list_head *head = &info->space_info;
 682         struct btrfs_space_info *found;
 683
 684         rcu_read_lock();
 685         list_for_each_entry_rcu(found, head, list)
 686                 found->full = 0;
 687         rcu_read_unlock();
 688 }
 689
 690 /* simple helper to search for an existing extent at a given offset */
 691 int btrfs_lookup_extent(struct btrfs_root *root, u64 start, u64 len)
 692 {
 693         int ret;
 694         struct btrfs_key key;
 695         struct btrfs_path *path;
 696
 697         path = btrfs_alloc_path();
 698         if (!path)
 699                 return -ENOMEM;
 700
 701         key.objectid = start;
 702         key.offset = len;
 703         key.type = BTRFS_EXTENT_ITEM_KEY;
 704         ret = btrfs_search_slot(NULL, root->fs_info->extent_root, &key, path,
 705                                 0, 0);
 706         if (ret > 0) {
 707                 btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]);
 708                 if (key.objectid == start &&
 709                     key.type == BTRFS_METADATA_ITEM_KEY)
 710                         ret = 0;
 711         }
 712         btrfs_free_path(path);
 713         return ret;
 714 }
 715
 716 /*
 717  * helper function to lookup reference count and flags of a tree block.
 718  *
 719  * the head node for delayed ref is used to store the sum of all the
 720  * reference count modifications queued up in the rbtree. the head
 721  * node may also store the extent flags to set. This way you can check
 722  * to see what the reference count and extent flags would be if all of
 723  * the delayed refs are not processed.
 724  */
 725 int btrfs_lookup_extent_info(struct btrfs_trans_handle *trans,
 726                              struct btrfs_root *root, u64 bytenr,
 727                              u64 offset, int metadata, u64 *refs, u64 *flags)
 728 {
 729         struct btrfs_delayed_ref_head *head;
 730         struct btrfs_delayed_ref_root *delayed_refs;
 731         struct btrfs_path *path;
 732         struct btrfs_extent_item *ei;
 733         struct extent_buffer *leaf;
 734         struct btrfs_key key;
 735         u32 item_size;
 736         u64 num_refs;
 737         u64 extent_flags;
 738         int ret;
 739
 740         /*
 741          * If we don't have skinny metadata, don't bother doing anything
 742          * different
 743          */
 744         if (metadata && !btrfs_fs_incompat(root->fs_info, SKINNY_METADATA)) {
 745                 offset = root->leafsize;
 746                 metadata = 0;
 747         }
 748
 749         path = btrfs_alloc_path();
 750         if (!path)
 751                 return -ENOMEM;
 752
 753         if (metadata) {
 754                 key.objectid = bytenr;
 755                 key.type = BTRFS_METADATA_ITEM_KEY;
 756                 key.offset = offset;
 757         } else {
 758                 key.objectid = bytenr;
 759                 key.type = BTRFS_EXTENT_ITEM_KEY;
 760                 key.offset = offset;
 761         }
 762
 763         if (!trans) {
 764                 path->skip_locking = 1;
 765                 path->search_commit_root = 1;
 766         }
 767 again:
 768         ret = btrfs_search_slot(trans, root->fs_info->extent_root,
 769                                 &key, path, 0, 0);
 770         if (ret < 0)
 771                 goto out_free;
 772
 773         if (ret > 0 && metadata && key.type == BTRFS_METADATA_ITEM_KEY) {
 774                 key.type = BTRFS_EXTENT_ITEM_KEY;
 775                 key.offset = root->leafsize;
 776                 btrfs_release_path(path);
 777                 goto again;
 778         }
 779
 780         if (ret == 0) {
 781                 leaf = path->nodes[0];
 782                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
 783                 if (item_size >= sizeof(*ei)) {
 784                         ei = btrfs_item_ptr(leaf, path->slots[0],
 785                                             struct btrfs_extent_item);
 786                         num_refs = btrfs_extent_refs(leaf, ei);
 787                         extent_flags = btrfs_extent_flags(leaf, ei);
 788                 } else {
 789 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
 790                         struct btrfs_extent_item_v0 *ei0;
 791                         BUG_ON(item_size != sizeof(*ei0));
 792                         ei0 = btrfs_item_ptr(leaf, path->slots[0],
 793                                              struct btrfs_extent_item_v0);
 794                         num_refs = btrfs_extent_refs_v0(leaf, ei0);
 795                         /* FIXME: this isn't correct for data */
 796                         extent_flags = BTRFS_BLOCK_FLAG_FULL_BACKREF;
 797 #else
 798                         BUG();
 799 #endif
 800                 }
 801                 BUG_ON(num_refs == 0);
 802         } else {
 803                 num_refs = 0;
 804                 extent_flags = 0;
 805                 ret = 0;
 806         }
 807
 808         if (!trans)
 809                 goto out;
 810
 811         delayed_refs = &trans->transaction->delayed_refs;
 812         spin_lock(&delayed_refs->lock);
 813         head = btrfs_find_delayed_ref_head(trans, bytenr);
 814         if (head) {
 815                 if (!mutex_trylock(&head->mutex)) {
 816                         atomic_inc(&head->node.refs);
 817                         spin_unlock(&delayed_refs->lock);
 818
 819                         btrfs_release_path(path);
 820
 821                         /*
 822                          * Mutex was contended, block until it's released and try
 823                          * again
 824                          */
 825                         mutex_lock(&head->mutex);
 826                         mutex_unlock(&head->mutex);
 827                         btrfs_put_delayed_ref(&head->node);
 828                         goto again;
 829                 }
 830                 if (head->extent_op && head->extent_op->update_flags)
 831                         extent_flags |= head->extent_op->flags_to_set;
 832                 else
 833                         BUG_ON(num_refs == 0);
 834
 835                 num_refs += head->node.ref_mod;
 836                 mutex_unlock(&head->mutex);
 837         }
 838         spin_unlock(&delayed_refs->lock);
 839 out:
 840         WARN_ON(num_refs == 0);
 841         if (refs)
 842                 *refs = num_refs;
 843         if (flags)
 844                 *flags = extent_flags;
 845 out_free:
 846         btrfs_free_path(path);
 847         return ret;
 848 }
 849
 850 /*
 851  * Back reference rules.  Back refs have three main goals:
 852  *
 853  * 1) differentiate between all holders of references to an extent so that
 854  *    when a reference is dropped we can make sure it was a valid reference
 855  *    before freeing the extent.
 856  *
 857  * 2) Provide enough information to quickly find the holders of an extent
 858  *    if we notice a given block is corrupted or bad.
 859  *
 860  * 3) Make it easy to migrate blocks for FS shrinking or storage pool
 861  *    maintenance.  This is actually the same as #2, but with a slightly
 862  *    different use case.
 863  *
 864  * There are two kinds of back refs. The implicit back refs is optimized
 865  * for pointers in non-shared tree blocks. For a given pointer in a block,
 866  * back refs of this kind provide information about the block's owner tree
 867  * and the pointer's key. These information allow us to find the block by
 868  * b-tree searching. The full back refs is for pointers in tree blocks not
 869  * referenced by their owner trees. The location of tree block is recorded
 870  * in the back refs. Actually the full back refs is generic, and can be
 871  * used in all cases the implicit back refs is used. The major shortcoming
 872  * of the full back refs is its overhead. Every time a tree block gets
 873  * COWed, we have to update back refs entry for all pointers in it.
 874  *
 875  * For a newly allocated tree block, we use implicit back refs for
 876  * pointers in it. This means most tree related operations only involve
 877  * implicit back refs. For a tree block created in old transaction, the
 878  * only way to drop a reference to it is COW it. So we can detect the
 879  * event that tree block loses its owner tree's reference and do the
 880  * back refs conversion.
 881  *
 882  * When a tree block is COW'd through a tree, there are four cases:
 883  *
 884  * The reference count of the block is one and the tree is the block's
 885  * owner tree. Nothing to do in this case.
 886  *
 887  * The reference count of the block is one and the tree is not the
 888  * block's owner tree. In this case, full back refs is used for pointers
 889  * in the block. Remove these full back refs, add implicit back refs for
 890  * every pointers in the new block.
 891  *
 892  * The reference count of the block is greater than one and the tree is
 893  * the block's owner tree. In this case, implicit back refs is used for
 894  * pointers in the block. Add full back refs for every pointers in the
 895  * block, increase lower level extents' reference counts. The original
 896  * implicit back refs are entailed to the new block.
 897  *
 898  * The reference count of the block is greater than one and the tree is
 899  * not the block's owner tree. Add implicit back refs for every pointer in
 900  * the new block, increase lower level extents' reference count.
 901  *
 902  * Back Reference Key composing:
 903  *
 904  * The key objectid corresponds to the first byte in the extent,
 905  * The key type is used to differentiate between types of back refs.
 906  * There are different meanings of the key offset for different types
 907  * of back refs.
 908  *
 909  * File extents can be referenced by:
 910  *
 911  * - multiple snapshots, subvolumes, or different generations in one subvol
 912  * - different files inside a single subvolume
 913  * - different offsets inside a file (bookend extents in file.c)
 914  *
 915  * The extent ref structure for the implicit back refs has fields for:
 916  *
 917  * - Objectid of the subvolume root
 918  * - objectid of the file holding the reference
 919  * - original offset in the file
 920  * - how many bookend extents
 921  *
 922  * The key offset for the implicit back refs is hash of the first
 923  * three fields.
 924  *
 925  * The extent ref structure for the full back refs has field for:
 926  *
 927  * - number of pointers in the tree leaf
 928  *
 929  * The key offset for the implicit back refs is the first byte of
 930  * the tree leaf
 931  *
 932  * When a file extent is allocated, The implicit back refs is used.
 933  * the fields are filled in:
 934  *
 935  *     (root_key.objectid, inode objectid, offset in file, 1)
 936  *
 937  * When a file extent is removed file truncation, we find the
 938  * corresponding implicit back refs and check the following fields:
 939  *
 940  *     (btrfs_header_owner(leaf), inode objectid, offset in file)
 941  *
 942  * Btree extents can be referenced by:
 943  *
 944  * - Different subvolumes
 945  *
 946  * Both the implicit back refs and the full back refs for tree blocks
 947  * only consist of key. The key offset for the implicit back refs is
 948  * objectid of block's owner tree. The key offset for the full back refs
 949  * is the first byte of parent block.
 950  *
 951  * When implicit back refs is used, information about the lowest key and
 952  * level of the tree block are required. These information are stored in
 953  * tree block info structure.
 954  */
 955
 956 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
 957 static int convert_extent_item_v0(struct btrfs_trans_handle *trans,
 958                                   struct btrfs_root *root,
 959                                   struct btrfs_path *path,
 960                                   u64 owner, u32 extra_size)
 961 {
 962         struct btrfs_extent_item *item;
 963         struct btrfs_extent_item_v0 *ei0;
 964         struct btrfs_extent_ref_v0 *ref0;
 965         struct btrfs_tree_block_info *bi;
 966         struct extent_buffer *leaf;
 967         struct btrfs_key key;
 968         struct btrfs_key found_key;
 969         u32 new_size = sizeof(*item);
 970         u64 refs;
 971         int ret;
 972
 973         leaf = path->nodes[0];
 974         BUG_ON(btrfs_item_size_nr(leaf, path->slots[0]) != sizeof(*ei0));
 975
 976         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
 977         ei0 = btrfs_item_ptr(leaf, path->slots[0],
 978                              struct btrfs_extent_item_v0);
 979         refs = btrfs_extent_refs_v0(leaf, ei0);
 980
 981         if (owner == (u64)-1) {
 982                 while (1) {
 983                         if (path->slots[0] >= btrfs_header_nritems(leaf)) {
 984                                 ret = btrfs_next_leaf(root, path);
 985                                 if (ret < 0)
 986                                         return ret;
 987                                 BUG_ON(ret > 0); /* Corruption */
 988                                 leaf = path->nodes[0];
 989                         }
 990                         btrfs_item_key_to_cpu(leaf, &found_key,
 991                                               path->slots[0]);
 992                         BUG_ON(key.objectid != found_key.objectid);
 993                         if (found_key.type != BTRFS_EXTENT_REF_V0_KEY) {
 994                                 path->slots[0]++;
 995                                 continue;
 996                         }
 997                         ref0 = btrfs_item_ptr(leaf, path->slots[0],
 998                                               struct btrfs_extent_ref_v0);
 999                         owner = btrfs_ref_objectid_v0(leaf, ref0);
1000                         break;
1001                 }
1002         }
1003         btrfs_release_path(path);
1004
1005         if (owner < BTRFS_FIRST_FREE_OBJECTID)
1006                 new_size += sizeof(*bi);
1007
1008         new_size -= sizeof(*ei0);
1009         ret = btrfs_search_slot(trans, root, &key, path,
1010                                 new_size + extra_size, 1);
1011         if (ret < 0)
1012                 return ret;
1013         BUG_ON(ret); /* Corruption */
1014
1015         btrfs_extend_item(root, path, new_size);
1016
1017         leaf = path->nodes[0];
1018         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1019         btrfs_set_extent_refs(leaf, item, refs);
1020         /* FIXME: get real generation */
1021         btrfs_set_extent_generation(leaf, item, 0);
1022         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1023                 btrfs_set_extent_flags(leaf, item,
1024                                        BTRFS_EXTENT_FLAG_TREE_BLOCK |
1025                                        BTRFS_BLOCK_FLAG_FULL_BACKREF);
1026                 bi = (struct btrfs_tree_block_info *)(item + 1);
1027                 /* FIXME: get first key of the block */
1028                 memset_extent_buffer(leaf, 0, (unsigned long)bi, sizeof(*bi));
1029                 btrfs_set_tree_block_level(leaf, bi, (int)owner);
1030         } else {
1031                 btrfs_set_extent_flags(leaf, item, BTRFS_EXTENT_FLAG_DATA);
1032         }
1033         btrfs_mark_buffer_dirty(leaf);
1034         return 0;
1035 }
1036 #endif
1037
1038 static u64 hash_extent_data_ref(u64 root_objectid, u64 owner, u64 offset)
1039 {
1040         u32 high_crc = ~(u32)0;
1041         u32 low_crc = ~(u32)0;
1042         __le64 lenum;
1043
1044         lenum = cpu_to_le64(root_objectid);
1045         high_crc = crc32c(high_crc, &lenum, sizeof(lenum));
1046         lenum = cpu_to_le64(owner);
1047         low_crc = crc32c(low_crc, &lenum, sizeof(lenum));
1048         lenum = cpu_to_le64(offset);
1049         low_crc = crc32c(low_crc, &lenum, sizeof(lenum));
1050
1051         return ((u64)high_crc << 31) ^ (u64)low_crc;
1052 }
1053
1054 static u64 hash_extent_data_ref_item(struct extent_buffer *leaf,
1055                                      struct btrfs_extent_data_ref *ref)
1056 {
1057         return hash_extent_data_ref(btrfs_extent_data_ref_root(leaf, ref),
1058                                     btrfs_extent_data_ref_objectid(leaf, ref),
1059                                     btrfs_extent_data_ref_offset(leaf, ref));
1060 }
1061
1062 static int match_extent_data_ref(struct extent_buffer *leaf,
1063                                  struct btrfs_extent_data_ref *ref,
1064                                  u64 root_objectid, u64 owner, u64 offset)
1065 {
1066         if (btrfs_extent_data_ref_root(leaf, ref) != root_objectid ||
1067             btrfs_extent_data_ref_objectid(leaf, ref) != owner ||
1068             btrfs_extent_data_ref_offset(leaf, ref) != offset)
1069                 return 0;
1070         return 1;
1071 }
1072
1073 static noinline int lookup_extent_data_ref(struct btrfs_trans_handle *trans,
1074                                            struct btrfs_root *root,
1075                                            struct btrfs_path *path,
1076                                            u64 bytenr, u64 parent,
1077                                            u64 root_objectid,
1078                                            u64 owner, u64 offset)
1079 {
1080         struct btrfs_key key;
1081         struct btrfs_extent_data_ref *ref;
1082         struct extent_buffer *leaf;
1083         u32 nritems;
1084         int ret;
1085         int recow;
1086         int err = -ENOENT;
1087
1088         key.objectid = bytenr;
1089         if (parent) {
1090                 key.type = BTRFS_SHARED_DATA_REF_KEY;
1091                 key.offset = parent;
1092         } else {
1093                 key.type = BTRFS_EXTENT_DATA_REF_KEY;
1094                 key.offset = hash_extent_data_ref(root_objectid,
1095                                                   owner, offset);
1096         }
1097 again:
1098         recow = 0;
1099         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1100         if (ret < 0) {
1101                 err = ret;
1102                 goto fail;
1103         }
1104
1105         if (parent) {
1106                 if (!ret)
1107                         return 0;
1108 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1109                 key.type = BTRFS_EXTENT_REF_V0_KEY;
1110                 btrfs_release_path(path);
1111                 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1112                 if (ret < 0) {
1113                         err = ret;
1114                         goto fail;
1115                 }
1116                 if (!ret)
1117                         return 0;
1118 #endif
1119                 goto fail;
1120         }
1121
1122         leaf = path->nodes[0];
1123         nritems = btrfs_header_nritems(leaf);
1124         while (1) {
1125                 if (path->slots[0] >= nritems) {
1126                         ret = btrfs_next_leaf(root, path);
1127                         if (ret < 0)
1128                                 err = ret;
1129                         if (ret)
1130                                 goto fail;
1131
1132                         leaf = path->nodes[0];
1133                         nritems = btrfs_header_nritems(leaf);
1134                         recow = 1;
1135                 }
1136
1137                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1138                 if (key.objectid != bytenr ||
1139                     key.type != BTRFS_EXTENT_DATA_REF_KEY)
1140                         goto fail;
1141
1142                 ref = btrfs_item_ptr(leaf, path->slots[0],
1143                                      struct btrfs_extent_data_ref);
1144
1145                 if (match_extent_data_ref(leaf, ref, root_objectid,
1146                                           owner, offset)) {
1147                         if (recow) {
1148                                 btrfs_release_path(path);
1149                                 goto again;
1150                         }
1151                         err = 0;
1152                         break;
1153                 }
1154                 path->slots[0]++;
1155         }
1156 fail:
1157         return err;
1158 }
1159
1160 static noinline int insert_extent_data_ref(struct btrfs_trans_handle *trans,
1161                                            struct btrfs_root *root,
1162                                            struct btrfs_path *path,
1163                                            u64 bytenr, u64 parent,
1164                                            u64 root_objectid, u64 owner,
1165                                            u64 offset, int refs_to_add)
1166 {
1167         struct btrfs_key key;
1168         struct extent_buffer *leaf;
1169         u32 size;
1170         u32 num_refs;
1171         int ret;
1172
1173         key.objectid = bytenr;
1174         if (parent) {
1175                 key.type = BTRFS_SHARED_DATA_REF_KEY;
1176                 key.offset = parent;
1177                 size = sizeof(struct btrfs_shared_data_ref);
1178         } else {
1179                 key.type = BTRFS_EXTENT_DATA_REF_KEY;
1180                 key.offset = hash_extent_data_ref(root_objectid,
1181                                                   owner, offset);
1182                 size = sizeof(struct btrfs_extent_data_ref);
1183         }
1184
1185         ret = btrfs_insert_empty_item(trans, root, path, &key, size);
1186         if (ret && ret != -EEXIST)
1187                 goto fail;
1188
1189         leaf = path->nodes[0];
1190         if (parent) {
1191                 struct btrfs_shared_data_ref *ref;
1192                 ref = btrfs_item_ptr(leaf, path->slots[0],
1193                                      struct btrfs_shared_data_ref);
1194                 if (ret == 0) {
1195                         btrfs_set_shared_data_ref_count(leaf, ref, refs_to_add);
1196                 } else {
1197                         num_refs = btrfs_shared_data_ref_count(leaf, ref);
1198                         num_refs += refs_to_add;
1199                         btrfs_set_shared_data_ref_count(leaf, ref, num_refs);
1200                 }
1201         } else {
1202                 struct btrfs_extent_data_ref *ref;
1203                 while (ret == -EEXIST) {
1204                         ref = btrfs_item_ptr(leaf, path->slots[0],
1205                                              struct btrfs_extent_data_ref);
1206                         if (match_extent_data_ref(leaf, ref, root_objectid,
1207                                                   owner, offset))
1208                                 break;
1209                         btrfs_release_path(path);
1210                         key.offset++;
1211                         ret = btrfs_insert_empty_item(trans, root, path, &key,
1212                                                       size);
1213                         if (ret && ret != -EEXIST)
1214                                 goto fail;
1215
1216                         leaf = path->nodes[0];
1217                 }
1218                 ref = btrfs_item_ptr(leaf, path->slots[0],
1219                                      struct btrfs_extent_data_ref);
1220                 if (ret == 0) {
1221                         btrfs_set_extent_data_ref_root(leaf, ref,
1222                                                        root_objectid);
1223                         btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
1224                         btrfs_set_extent_data_ref_offset(leaf, ref, offset);
1225                         btrfs_set_extent_data_ref_count(leaf, ref, refs_to_add);
1226                 } else {
1227                         num_refs = btrfs_extent_data_ref_count(leaf, ref);
1228                         num_refs += refs_to_add;
1229                         btrfs_set_extent_data_ref_count(leaf, ref, num_refs);
1230                 }
1231         }
1232         btrfs_mark_buffer_dirty(leaf);
1233         ret = 0;
1234 fail:
1235         btrfs_release_path(path);
1236         return ret;
1237 }
1238
1239 static noinline int remove_extent_data_ref(struct btrfs_trans_handle *trans,
1240                                            struct btrfs_root *root,
1241                                            struct btrfs_path *path,
1242                                            int refs_to_drop)
1243 {
1244         struct btrfs_key key;
1245         struct btrfs_extent_data_ref *ref1 = NULL;
1246         struct btrfs_shared_data_ref *ref2 = NULL;
1247         struct extent_buffer *leaf;
1248         u32 num_refs = 0;
1249         int ret = 0;
1250
1251         leaf = path->nodes[0];
1252         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1253
1254         if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
1255                 ref1 = btrfs_item_ptr(leaf, path->slots[0],
1256                                       struct btrfs_extent_data_ref);
1257                 num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1258         } else if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
1259                 ref2 = btrfs_item_ptr(leaf, path->slots[0],
1260                                       struct btrfs_shared_data_ref);
1261                 num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1262 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1263         } else if (key.type == BTRFS_EXTENT_REF_V0_KEY) {
1264                 struct btrfs_extent_ref_v0 *ref0;
1265                 ref0 = btrfs_item_ptr(leaf, path->slots[0],
1266                                       struct btrfs_extent_ref_v0);
1267                 num_refs = btrfs_ref_count_v0(leaf, ref0);
1268 #endif
1269         } else {
1270                 BUG();
1271         }
1272
1273         BUG_ON(num_refs < refs_to_drop);
1274         num_refs -= refs_to_drop;
1275
1276         if (num_refs == 0) {
1277                 ret = btrfs_del_item(trans, root, path);
1278         } else {
1279                 if (key.type == BTRFS_EXTENT_DATA_REF_KEY)
1280                         btrfs_set_extent_data_ref_count(leaf, ref1, num_refs);
1281                 else if (key.type == BTRFS_SHARED_DATA_REF_KEY)
1282                         btrfs_set_shared_data_ref_count(leaf, ref2, num_refs);
1283 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1284                 else {
1285                         struct btrfs_extent_ref_v0 *ref0;
1286                         ref0 = btrfs_item_ptr(leaf, path->slots[0],
1287                                         struct btrfs_extent_ref_v0);
1288                         btrfs_set_ref_count_v0(leaf, ref0, num_refs);
1289                 }
1290 #endif
1291                 btrfs_mark_buffer_dirty(leaf);
1292         }
1293         return ret;
1294 }
1295
1296 static noinline u32 extent_data_ref_count(struct btrfs_root *root,
1297                                           struct btrfs_path *path,
1298                                           struct btrfs_extent_inline_ref *iref)
1299 {
1300         struct btrfs_key key;
1301         struct extent_buffer *leaf;
1302         struct btrfs_extent_data_ref *ref1;
1303         struct btrfs_shared_data_ref *ref2;
1304         u32 num_refs = 0;
1305
1306         leaf = path->nodes[0];
1307         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1308         if (iref) {
1309                 if (btrfs_extent_inline_ref_type(leaf, iref) ==
1310                     BTRFS_EXTENT_DATA_REF_KEY) {
1311                         ref1 = (struct btrfs_extent_data_ref *)(&iref->offset);
1312                         num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1313                 } else {
1314                         ref2 = (struct btrfs_shared_data_ref *)(iref + 1);
1315                         num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1316                 }
1317         } else if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
1318                 ref1 = btrfs_item_ptr(leaf, path->slots[0],
1319                                       struct btrfs_extent_data_ref);
1320                 num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1321         } else if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
1322                 ref2 = btrfs_item_ptr(leaf, path->slots[0],
1323                                       struct btrfs_shared_data_ref);
1324                 num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1325 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1326         } else if (key.type == BTRFS_EXTENT_REF_V0_KEY) {
1327                 struct btrfs_extent_ref_v0 *ref0;
1328                 ref0 = btrfs_item_ptr(leaf, path->slots[0],
1329                                       struct btrfs_extent_ref_v0);
1330                 num_refs = btrfs_ref_count_v0(leaf, ref0);
1331 #endif
1332         } else {
1333                 WARN_ON(1);
1334         }
1335         return num_refs;
1336 }
1337
1338 static noinline int lookup_tree_block_ref(struct btrfs_trans_handle *trans,
1339                                           struct btrfs_root *root,
1340                                           struct btrfs_path *path,
1341                                           u64 bytenr, u64 parent,
1342                                           u64 root_objectid)
1343 {
1344         struct btrfs_key key;
1345         int ret;
1346
1347         key.objectid = bytenr;
1348         if (parent) {
1349                 key.type = BTRFS_SHARED_BLOCK_REF_KEY;
1350                 key.offset = parent;
1351         } else {
1352                 key.type = BTRFS_TREE_BLOCK_REF_KEY;
1353                 key.offset = root_objectid;
1354         }
1355
1356         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1357         if (ret > 0)
1358                 ret = -ENOENT;
1359 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1360         if (ret == -ENOENT && parent) {
1361                 btrfs_release_path(path);
1362                 key.type = BTRFS_EXTENT_REF_V0_KEY;
1363                 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1364                 if (ret > 0)
1365                         ret = -ENOENT;
1366         }
1367 #endif
1368         return ret;
1369 }
1370
1371 static noinline int insert_tree_block_ref(struct btrfs_trans_handle *trans,
1372                                           struct btrfs_root *root,
1373                                           struct btrfs_path *path,
1374                                           u64 bytenr, u64 parent,
1375                                           u64 root_objectid)
1376 {
1377         struct btrfs_key key;
1378         int ret;
1379
1380         key.objectid = bytenr;
1381         if (parent) {
1382                 key.type = BTRFS_SHARED_BLOCK_REF_KEY;
1383                 key.offset = parent;
1384         } else {
1385                 key.type = BTRFS_TREE_BLOCK_REF_KEY;
1386                 key.offset = root_objectid;
1387         }
1388
1389         ret = btrfs_insert_empty_item(trans, root, path, &key, 0);
1390         btrfs_release_path(path);
1391         return ret;
1392 }
1393
1394 static inline int extent_ref_type(u64 parent, u64 owner)
1395 {
1396         int type;
1397         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1398                 if (parent > 0)
1399                         type = BTRFS_SHARED_BLOCK_REF_KEY;
1400                 else
1401                         type = BTRFS_TREE_BLOCK_REF_KEY;
1402         } else {
1403                 if (parent > 0)
1404                         type = BTRFS_SHARED_DATA_REF_KEY;
1405                 else
1406                         type = BTRFS_EXTENT_DATA_REF_KEY;
1407         }
1408         return type;
1409 }
1410
1411 static int find_next_key(struct btrfs_path *path, int level,
1412                          struct btrfs_key *key)
1413
1414 {
1415         for (; level < BTRFS_MAX_LEVEL; level++) {
1416                 if (!path->nodes[level])
1417                         break;
1418                 if (path->slots[level] + 1 >=
1419                     btrfs_header_nritems(path->nodes[level]))
1420                         continue;
1421                 if (level == 0)
1422                         btrfs_item_key_to_cpu(path->nodes[level], key,
1423                                               path->slots[level] + 1);
1424                 else
1425                         btrfs_node_key_to_cpu(path->nodes[level], key,
1426                                               path->slots[level] + 1);
1427                 return 0;
1428         }
1429         return 1;
1430 }
1431
1432 /*
1433  * look for inline back ref. if back ref is found, *ref_ret is set
1434  * to the address of inline back ref, and 0 is returned.
1435  *
1436  * if back ref isn't found, *ref_ret is set to the address where it
1437  * should be inserted, and -ENOENT is returned.
1438  *
1439  * if insert is true and there are too many inline back refs, the path
1440  * points to the extent item, and -EAGAIN is returned.
1441  *
1442  * NOTE: inline back refs are ordered in the same way that back ref
1443  *       items in the tree are ordered.
1444  */
1445 static noinline_for_stack
1446 int lookup_inline_extent_backref(struct btrfs_trans_handle *trans,
1447                                  struct btrfs_root *root,
1448                                  struct btrfs_path *path,
1449                                  struct btrfs_extent_inline_ref **ref_ret,
1450                                  u64 bytenr, u64 num_bytes,
1451                                  u64 parent, u64 root_objectid,
1452                                  u64 owner, u64 offset, int insert)
1453 {
1454         struct btrfs_key key;
1455         struct extent_buffer *leaf;
1456         struct btrfs_extent_item *ei;
1457         struct btrfs_extent_inline_ref *iref;
1458         u64 flags;
1459         u64 item_size;
1460         unsigned long ptr;
1461         unsigned long end;
1462         int extra_size;
1463         int type;
1464         int want;
1465         int ret;
1466         int err = 0;
1467         bool skinny_metadata = btrfs_fs_incompat(root->fs_info,
1468                                                  SKINNY_METADATA);
1469
1470         key.objectid = bytenr;
1471         key.type = BTRFS_EXTENT_ITEM_KEY;
1472         key.offset = num_bytes;
1473
1474         want = extent_ref_type(parent, owner);
1475         if (insert) {
1476                 extra_size = btrfs_extent_inline_ref_size(want);
1477                 path->keep_locks = 1;
1478         } else
1479                 extra_size = -1;
1480
1481         /*
1482          * Owner is our parent level, so we can just add one to get the level
1483          * for the block we are interested in.
1484          */
1485         if (skinny_metadata && owner < BTRFS_FIRST_FREE_OBJECTID) {
1486                 key.type = BTRFS_METADATA_ITEM_KEY;
1487                 key.offset = owner;
1488         }
1489
1490 again:
1491         ret = btrfs_search_slot(trans, root, &key, path, extra_size, 1);
1492         if (ret < 0) {
1493                 err = ret;
1494                 goto out;
1495         }
1496
1497         /*
1498          * We may be a newly converted file system which still has the old fat
1499          * extent entries for metadata, so try and see if we have one of those.
1500          */
1501         if (ret > 0 && skinny_metadata) {
1502                 skinny_metadata = false;
1503                 if (path->slots[0]) {
1504                         path->slots[0]--;
1505                         btrfs_item_key_to_cpu(path->nodes[0], &key,
1506                                               path->slots[0]);
1507                         if (key.objectid == bytenr &&
1508                             key.type == BTRFS_EXTENT_ITEM_KEY &&
1509                             key.offset == num_bytes)
1510                                 ret = 0;
1511                 }
1512                 if (ret) {
1513                         key.type = BTRFS_EXTENT_ITEM_KEY;
1514                         key.offset = num_bytes;
1515                         btrfs_release_path(path);
1516                         goto again;
1517                 }
1518         }
1519
1520         if (ret && !insert) {
1521                 err = -ENOENT;
1522                 goto out;
1523         } else if (ret) {
1524                 err = -EIO;
1525                 WARN_ON(1);
1526                 goto out;
1527         }
1528
1529         leaf = path->nodes[0];
1530         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1531 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1532         if (item_size < sizeof(*ei)) {
1533                 if (!insert) {
1534                         err = -ENOENT;
1535                         goto out;
1536                 }
1537                 ret = convert_extent_item_v0(trans, root, path, owner,
1538                                              extra_size);
1539                 if (ret < 0) {
1540                         err = ret;
1541                         goto out;
1542                 }
1543                 leaf = path->nodes[0];
1544                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1545         }
1546 #endif
1547         BUG_ON(item_size < sizeof(*ei));
1548
1549         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1550         flags = btrfs_extent_flags(leaf, ei);
1551
1552         ptr = (unsigned long)(ei + 1);
1553         end = (unsigned long)ei + item_size;
1554
1555         if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK && !skinny_metadata) {
1556                 ptr += sizeof(struct btrfs_tree_block_info);
1557                 BUG_ON(ptr > end);
1558         }
1559
1560         err = -ENOENT;
1561         while (1) {
1562                 if (ptr >= end) {
1563                         WARN_ON(ptr > end);
1564                         break;
1565                 }
1566                 iref = (struct btrfs_extent_inline_ref *)ptr;
1567                 type = btrfs_extent_inline_ref_type(leaf, iref);
1568                 if (want < type)
1569                         break;
1570                 if (want > type) {
1571                         ptr += btrfs_extent_inline_ref_size(type);
1572                         continue;
1573                 }
1574
1575                 if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1576                         struct btrfs_extent_data_ref *dref;
1577                         dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1578                         if (match_extent_data_ref(leaf, dref, root_objectid,
1579                                                   owner, offset)) {
1580                                 err = 0;
1581                                 break;
1582                         }
1583                         if (hash_extent_data_ref_item(leaf, dref) <
1584                             hash_extent_data_ref(root_objectid, owner, offset))
1585                                 break;
1586                 } else {
1587                         u64 ref_offset;
1588                         ref_offset = btrfs_extent_inline_ref_offset(leaf, iref);
1589                         if (parent > 0) {
1590                                 if (parent == ref_offset) {
1591                                         err = 0;
1592                                         break;
1593                                 }
1594                                 if (ref_offset < parent)
1595                                         break;
1596                         } else {
1597                                 if (root_objectid == ref_offset) {
1598                                         err = 0;
1599                                         break;
1600                                 }
1601                                 if (ref_offset < root_objectid)
1602                                         break;
1603                         }
1604                 }
1605                 ptr += btrfs_extent_inline_ref_size(type);
1606         }
1607         if (err == -ENOENT && insert) {
1608                 if (item_size + extra_size >=
1609                     BTRFS_MAX_EXTENT_ITEM_SIZE(root)) {
1610                         err = -EAGAIN;
1611                         goto out;
1612                 }
1613                 /*
1614                  * To add new inline back ref, we have to make sure
1615                  * there is no corresponding back ref item.
1616                  * For simplicity, we just do not add new inline back
1617                  * ref if there is any kind of item for this block
1618                  */
1619                 if (find_next_key(path, 0, &key) == 0 &&
1620                     key.objectid == bytenr &&
1621                     key.type < BTRFS_BLOCK_GROUP_ITEM_KEY) {
1622                         err = -EAGAIN;
1623                         goto out;
1624                 }
1625         }
1626         *ref_ret = (struct btrfs_extent_inline_ref *)ptr;
1627 out:
1628         if (insert) {
1629                 path->keep_locks = 0;
1630                 btrfs_unlock_up_safe(path, 1);
1631         }
1632         return err;
1633 }
1634
1635 /*
1636  * helper to add new inline back ref
1637  */
1638 static noinline_for_stack
1639 void setup_inline_extent_backref(struct btrfs_root *root,
1640                                  struct btrfs_path *path,
1641                                  struct btrfs_extent_inline_ref *iref,
1642                                  u64 parent, u64 root_objectid,
1643                                  u64 owner, u64 offset, int refs_to_add,
1644                                  struct btrfs_delayed_extent_op *extent_op)
1645 {
1646         struct extent_buffer *leaf;
1647         struct btrfs_extent_item *ei;
1648         unsigned long ptr;
1649         unsigned long end;
1650         unsigned long item_offset;
1651         u64 refs;
1652         int size;
1653         int type;
1654
1655         leaf = path->nodes[0];
1656         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1657         item_offset = (unsigned long)iref - (unsigned long)ei;
1658
1659         type = extent_ref_type(parent, owner);
1660         size = btrfs_extent_inline_ref_size(type);
1661
1662         btrfs_extend_item(root, path, size);
1663
1664         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1665         refs = btrfs_extent_refs(leaf, ei);
1666         refs += refs_to_add;
1667         btrfs_set_extent_refs(leaf, ei, refs);
1668         if (extent_op)
1669                 __run_delayed_extent_op(extent_op, leaf, ei);
1670
1671         ptr = (unsigned long)ei + item_offset;
1672         end = (unsigned long)ei + btrfs_item_size_nr(leaf, path->slots[0]);
1673         if (ptr < end - size)
1674                 memmove_extent_buffer(leaf, ptr + size, ptr,
1675                                       end - size - ptr);
1676
1677         iref = (struct btrfs_extent_inline_ref *)ptr;
1678         btrfs_set_extent_inline_ref_type(leaf, iref, type);
1679         if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1680                 struct btrfs_extent_data_ref *dref;
1681                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1682                 btrfs_set_extent_data_ref_root(leaf, dref, root_objectid);
1683                 btrfs_set_extent_data_ref_objectid(leaf, dref, owner);
1684                 btrfs_set_extent_data_ref_offset(leaf, dref, offset);
1685                 btrfs_set_extent_data_ref_count(leaf, dref, refs_to_add);
1686         } else if (type == BTRFS_SHARED_DATA_REF_KEY) {
1687                 struct btrfs_shared_data_ref *sref;
1688                 sref = (struct btrfs_shared_data_ref *)(iref + 1);
1689                 btrfs_set_shared_data_ref_count(leaf, sref, refs_to_add);
1690                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
1691         } else if (type == BTRFS_SHARED_BLOCK_REF_KEY) {
1692                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
1693         } else {
1694                 btrfs_set_extent_inline_ref_offset(leaf, iref, root_objectid);
1695         }
1696         btrfs_mark_buffer_dirty(leaf);
1697 }
1698
1699 static int lookup_extent_backref(struct btrfs_trans_handle *trans,
1700                                  struct btrfs_root *root,
1701                                  struct btrfs_path *path,
1702                                  struct btrfs_extent_inline_ref **ref_ret,
1703                                  u64 bytenr, u64 num_bytes, u64 parent,
1704                                  u64 root_objectid, u64 owner, u64 offset)
1705 {
1706         int ret;
1707
1708         ret = lookup_inline_extent_backref(trans, root, path, ref_ret,
1709                                            bytenr, num_bytes, parent,
1710                                            root_objectid, owner, offset, 0);
1711         if (ret != -ENOENT)
1712                 return ret;
1713
1714         btrfs_release_path(path);
1715         *ref_ret = NULL;
1716
1717         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1718                 ret = lookup_tree_block_ref(trans, root, path, bytenr, parent,
1719                                             root_objectid);
1720         } else {
1721                 ret = lookup_extent_data_ref(trans, root, path, bytenr, parent,
1722                                              root_objectid, owner, offset);
1723         }
1724         return ret;
1725 }
1726
1727 /*
1728  * helper to update/remove inline back ref
1729  */
1730 static noinline_for_stack
1731 void update_inline_extent_backref(struct btrfs_root *root,
1732                                   struct btrfs_path *path,
1733                                   struct btrfs_extent_inline_ref *iref,
1734                                   int refs_to_mod,
1735                                   struct btrfs_delayed_extent_op *extent_op)
1736 {
1737         struct extent_buffer *leaf;
1738         struct btrfs_extent_item *ei;
1739         struct btrfs_extent_data_ref *dref = NULL;
1740         struct btrfs_shared_data_ref *sref = NULL;
1741         unsigned long ptr;
1742         unsigned long end;
1743         u32 item_size;
1744         int size;
1745         int type;
1746         u64 refs;
1747
1748         leaf = path->nodes[0];
1749         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1750         refs = btrfs_extent_refs(leaf, ei);
1751         WARN_ON(refs_to_mod < 0 && refs + refs_to_mod <= 0);
1752         refs += refs_to_mod;
1753         btrfs_set_extent_refs(leaf, ei, refs);
1754         if (extent_op)
1755                 __run_delayed_extent_op(extent_op, leaf, ei);
1756
1757         type = btrfs_extent_inline_ref_type(leaf, iref);
1758
1759         if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1760                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1761                 refs = btrfs_extent_data_ref_count(leaf, dref);
1762         } else if (type == BTRFS_SHARED_DATA_REF_KEY) {
1763                 sref = (struct btrfs_shared_data_ref *)(iref + 1);
1764                 refs = btrfs_shared_data_ref_count(leaf, sref);
1765         } else {
1766                 refs = 1;
1767                 BUG_ON(refs_to_mod != -1);
1768         }
1769
1770         BUG_ON(refs_to_mod < 0 && refs < -refs_to_mod);
1771         refs += refs_to_mod;
1772
1773         if (refs > 0) {
1774                 if (type == BTRFS_EXTENT_DATA_REF_KEY)
1775                         btrfs_set_extent_data_ref_count(leaf, dref, refs);
1776                 else
1777                         btrfs_set_shared_data_ref_count(leaf, sref, refs);
1778         } else {
1779                 size =  btrfs_extent_inline_ref_size(type);
1780                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1781                 ptr = (unsigned long)iref;
1782                 end = (unsigned long)ei + item_size;
1783                 if (ptr + size < end)
1784                         memmove_extent_buffer(leaf, ptr, ptr + size,
1785                                               end - ptr - size);
1786                 item_size -= size;
1787                 btrfs_truncate_item(root, path, item_size, 1);
1788         }
1789         btrfs_mark_buffer_dirty(leaf);
1790 }
1791
1792 static noinline_for_stack
1793 int insert_inline_extent_backref(struct btrfs_trans_handle *trans,
1794                                  struct btrfs_root *root,
1795                                  struct btrfs_path *path,
1796                                  u64 bytenr, u64 num_bytes, u64 parent,
1797                                  u64 root_objectid, u64 owner,
1798                                  u64 offset, int refs_to_add,
1799                                  struct btrfs_delayed_extent_op *extent_op)
1800 {
1801         struct btrfs_extent_inline_ref *iref;
1802         int ret;
1803
1804         ret = lookup_inline_extent_backref(trans, root, path, &iref,
1805                                            bytenr, num_bytes, parent,
1806                                            root_objectid, owner, offset, 1);
1807         if (ret == 0) {
1808                 BUG_ON(owner < BTRFS_FIRST_FREE_OBJECTID);
1809                 update_inline_extent_backref(root, path, iref,
1810                                              refs_to_add, extent_op);
1811         } else if (ret == -ENOENT) {
1812                 setup_inline_extent_backref(root, path, iref, parent,
1813                                             root_objectid, owner, offset,
1814                                             refs_to_add, extent_op);
1815                 ret = 0;
1816         }
1817         return ret;
1818 }
1819
1820 static int insert_extent_backref(struct btrfs_trans_handle *trans,
1821                                  struct btrfs_root *root,
1822                                  struct btrfs_path *path,
1823                                  u64 bytenr, u64 parent, u64 root_objectid,
1824                                  u64 owner, u64 offset, int refs_to_add)
1825 {
1826         int ret;
1827         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1828                 BUG_ON(refs_to_add != 1);
1829                 ret = insert_tree_block_ref(trans, root, path, bytenr,
1830                                             parent, root_objectid);
1831         } else {
1832                 ret = insert_extent_data_ref(trans, root, path, bytenr,
1833                                              parent, root_objectid,
1834                                              owner, offset, refs_to_add);
1835         }
1836         return ret;
1837 }
1838
1839 static int remove_extent_backref(struct btrfs_trans_handle *trans,
1840                                  struct btrfs_root *root,
1841                                  struct btrfs_path *path,
1842                                  struct btrfs_extent_inline_ref *iref,
1843                                  int refs_to_drop, int is_data)
1844 {
1845         int ret = 0;
1846
1847         BUG_ON(!is_data && refs_to_drop != 1);
1848         if (iref) {
1849                 update_inline_extent_backref(root, path, iref,
1850                                              -refs_to_drop, NULL);
1851         } else if (is_data) {
1852                 ret = remove_extent_data_ref(trans, root, path, refs_to_drop);
1853         } else {
1854                 ret = btrfs_del_item(trans, root, path);
1855         }
1856         return ret;
1857 }
1858
1859 static int btrfs_issue_discard(struct block_device *bdev,
1860                                 u64 start, u64 len)
1861 {
1862         return blkdev_issue_discard(bdev, start >> 9, len >> 9, GFP_NOFS, 0);
1863 }
1864
1865 static int btrfs_discard_extent(struct btrfs_root *root, u64 bytenr,
1866                                 u64 num_bytes, u64 *actual_bytes)
1867 {
1868         int ret;
1869         u64 discarded_bytes = 0;
1870         struct btrfs_bio *bbio = NULL;
1871
1872
1873         /* Tell the block device(s) that the sectors can be discarded */
1874         ret = btrfs_map_block(root->fs_info, REQ_DISCARD,
1875                               bytenr, &num_bytes, &bbio, 0);
1876         /* Error condition is -ENOMEM */
1877         if (!ret) {
1878                 struct btrfs_bio_stripe *stripe = bbio->stripes;
1879                 int i;
1880
1881
1882                 for (i = 0; i < bbio->num_stripes; i++, stripe++) {
1883                         if (!stripe->dev->can_discard)
1884                                 continue;
1885
1886                         ret = btrfs_issue_discard(stripe->dev->bdev,
1887                                                   stripe->physical,
1888                                                   stripe->length);
1889                         if (!ret)
1890                                 discarded_bytes += stripe->length;
1891                         else if (ret != -EOPNOTSUPP)
1892                                 break; /* Logic errors or -ENOMEM, or -EIO but I don't know how that could happen JDM */
1893
1894                         /*
1895                          * Just in case we get back EOPNOTSUPP for some reason,
1896                          * just ignore the return value so we don't screw up
1897                          * people calling discard_extent.
1898                          */
1899                         ret = 0;
1900                 }
1901                 kfree(bbio);
1902         }
1903
1904         if (actual_bytes)
1905                 *actual_bytes = discarded_bytes;
1906
1907
1908         if (ret == -EOPNOTSUPP)
1909                 ret = 0;
1910         return ret;
1911 }
1912
1913 /* Can return -ENOMEM */
1914 int btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
1915                          struct btrfs_root *root,
1916                          u64 bytenr, u64 num_bytes, u64 parent,
1917                          u64 root_objectid, u64 owner, u64 offset, int for_cow)
1918 {
1919         int ret;
1920         struct btrfs_fs_info *fs_info = root->fs_info;
1921
1922         BUG_ON(owner < BTRFS_FIRST_FREE_OBJECTID &&
1923                root_objectid == BTRFS_TREE_LOG_OBJECTID);
1924
1925         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1926                 ret = btrfs_add_delayed_tree_ref(fs_info, trans, bytenr,
1927                                         num_bytes,
1928                                         parent, root_objectid, (int)owner,
1929                                         BTRFS_ADD_DELAYED_REF, NULL, for_cow);
1930         } else {
1931                 ret = btrfs_add_delayed_data_ref(fs_info, trans, bytenr,
1932                                         num_bytes,
1933                                         parent, root_objectid, owner, offset,
1934                                         BTRFS_ADD_DELAYED_REF, NULL, for_cow);
1935         }
1936         return ret;
1937 }
1938
1939 static int __btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
1940                                   struct btrfs_root *root,
1941                                   u64 bytenr, u64 num_bytes,
1942                                   u64 parent, u64 root_objectid,
1943                                   u64 owner, u64 offset, int refs_to_add,
1944                                   struct btrfs_delayed_extent_op *extent_op)
1945 {
1946         struct btrfs_path *path;
1947         struct extent_buffer *leaf;
1948         struct btrfs_extent_item *item;
1949         u64 refs;
1950         int ret;
1951         int err = 0;
1952
1953         path = btrfs_alloc_path();
1954         if (!path)
1955                 return -ENOMEM;
1956
1957         path->reada = 1;
1958         path->leave_spinning = 1;
1959         /* this will setup the path even if it fails to insert the back ref */
1960         ret = insert_inline_extent_backref(trans, root->fs_info->extent_root,
1961                                            path, bytenr, num_bytes, parent,
1962                                            root_objectid, owner, offset,
1963                                            refs_to_add, extent_op);
1964         if (ret == 0)
1965                 goto out;
1966
1967         if (ret != -EAGAIN) {
1968                 err = ret;
1969                 goto out;
1970         }
1971
1972         leaf = path->nodes[0];
1973         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1974         refs = btrfs_extent_refs(leaf, item);
1975         btrfs_set_extent_refs(leaf, item, refs + refs_to_add);
1976         if (extent_op)
1977                 __run_delayed_extent_op(extent_op, leaf, item);
1978
1979         btrfs_mark_buffer_dirty(leaf);
1980         btrfs_release_path(path);
1981
1982         path->reada = 1;
1983         path->leave_spinning = 1;
1984
1985         /* now insert the actual backref */
1986         ret = insert_extent_backref(trans, root->fs_info->extent_root,
1987                                     path, bytenr, parent, root_objectid,
1988                                     owner, offset, refs_to_add);
1989         if (ret)
1990                 btrfs_abort_transaction(trans, root, ret);
1991 out:
1992         btrfs_free_path(path);
1993         return err;
1994 }
1995
1996 static int run_delayed_data_ref(struct btrfs_trans_handle *trans,
1997                                 struct btrfs_root *root,
1998                                 struct btrfs_delayed_ref_node *node,
1999                                 struct btrfs_delayed_extent_op *extent_op,
2000                                 int insert_reserved)
2001 {
2002         int ret = 0;
2003         struct btrfs_delayed_data_ref *ref;
2004         struct btrfs_key ins;
2005         u64 parent = 0;
2006         u64 ref_root = 0;
2007         u64 flags = 0;
2008
2009         ins.objectid = node->bytenr;
2010         ins.offset = node->num_bytes;
2011         ins.type = BTRFS_EXTENT_ITEM_KEY;
2012
2013         ref = btrfs_delayed_node_to_data_ref(node);
2014         if (node->type == BTRFS_SHARED_DATA_REF_KEY)
2015                 parent = ref->parent;
2016         else
2017                 ref_root = ref->root;
2018
2019         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
2020                 if (extent_op)
2021                         flags |= extent_op->flags_to_set;
2022                 ret = alloc_reserved_file_extent(trans, root,
2023                                                  parent, ref_root, flags,
2024                                                  ref->objectid, ref->offset,
2025                                                  &ins, node->ref_mod);
2026         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
2027                 ret = __btrfs_inc_extent_ref(trans, root, node->bytenr,
2028                                              node->num_bytes, parent,
2029                                              ref_root, ref->objectid,
2030                                              ref->offset, node->ref_mod,
2031                                              extent_op);
2032         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
2033                 ret = __btrfs_free_extent(trans, root, node->bytenr,
2034                                           node->num_bytes, parent,
2035                                           ref_root, ref->objectid,
2036                                           ref->offset, node->ref_mod,
2037                                           extent_op);
2038         } else {
2039                 BUG();
2040         }
2041         return ret;
2042 }
2043
2044 static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op,
2045                                     struct extent_buffer *leaf,
2046                                     struct btrfs_extent_item *ei)
2047 {
2048         u64 flags = btrfs_extent_flags(leaf, ei);
2049         if (extent_op->update_flags) {
2050                 flags |= extent_op->flags_to_set;
2051                 btrfs_set_extent_flags(leaf, ei, flags);
2052         }
2053
2054         if (extent_op->update_key) {
2055                 struct btrfs_tree_block_info *bi;
2056                 BUG_ON(!(flags & BTRFS_EXTENT_FLAG_TREE_BLOCK));
2057                 bi = (struct btrfs_tree_block_info *)(ei + 1);
2058                 btrfs_set_tree_block_key(leaf, bi, &extent_op->key);
2059         }
2060 }
2061
2062 static int run_delayed_extent_op(struct btrfs_trans_handle *trans,
2063                                  struct btrfs_root *root,
2064                                  struct btrfs_delayed_ref_node *node,
2065                                  struct btrfs_delayed_extent_op *extent_op)
2066 {
2067         struct btrfs_key key;
2068         struct btrfs_path *path;
2069         struct btrfs_extent_item *ei;
2070         struct extent_buffer *leaf;
2071         u32 item_size;
2072         int ret;
2073         int err = 0;
2074         int metadata = !extent_op->is_data;
2075
2076         if (trans->aborted)
2077                 return 0;
2078
2079         if (metadata && !btrfs_fs_incompat(root->fs_info, SKINNY_METADATA))
2080                 metadata = 0;
2081
2082         path = btrfs_alloc_path();
2083         if (!path)
2084                 return -ENOMEM;
2085
2086         key.objectid = node->bytenr;
2087
2088         if (metadata) {
2089                 key.type = BTRFS_METADATA_ITEM_KEY;
2090                 key.offset = extent_op->level;
2091         } else {
2092                 key.type = BTRFS_EXTENT_ITEM_KEY;
2093                 key.offset = node->num_bytes;
2094         }
2095
2096 again:
2097         path->reada = 1;
2098         path->leave_spinning = 1;
2099         ret = btrfs_search_slot(trans, root->fs_info->extent_root, &key,
2100                                 path, 0, 1);
2101         if (ret < 0) {
2102                 err = ret;
2103                 goto out;
2104         }
2105         if (ret > 0) {
2106                 if (metadata) {
2107                         btrfs_release_path(path);
2108                         metadata = 0;
2109
2110                         key.offset = node->num_bytes;
2111                         key.type = BTRFS_EXTENT_ITEM_KEY;
2112                         goto again;
2113                 }
2114                 err = -EIO;
2115                 goto out;
2116         }
2117
2118         leaf = path->nodes[0];
2119         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2120 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
2121         if (item_size < sizeof(*ei)) {
2122                 ret = convert_extent_item_v0(trans, root->fs_info->extent_root,
2123                                              path, (u64)-1, 0);
2124                 if (ret < 0) {
2125                         err = ret;
2126                         goto out;
2127                 }
2128                 leaf = path->nodes[0];
2129                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2130         }
2131 #endif
2132         BUG_ON(item_size < sizeof(*ei));
2133         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2134         __run_delayed_extent_op(extent_op, leaf, ei);
2135
2136         btrfs_mark_buffer_dirty(leaf);
2137 out:
2138         btrfs_free_path(path);
2139         return err;
2140 }
2141
2142 static int run_delayed_tree_ref(struct btrfs_trans_handle *trans,
2143                                 struct btrfs_root *root,
2144                                 struct btrfs_delayed_ref_node *node,
2145                                 struct btrfs_delayed_extent_op *extent_op,
2146                                 int insert_reserved)
2147 {
2148         int ret = 0;
2149         struct btrfs_delayed_tree_ref *ref;
2150         struct btrfs_key ins;
2151         u64 parent = 0;
2152         u64 ref_root = 0;
2153         bool skinny_metadata = btrfs_fs_incompat(root->fs_info,
2154                                                  SKINNY_METADATA);
2155
2156         ref = btrfs_delayed_node_to_tree_ref(node);
2157         if (node->type == BTRFS_SHARED_BLOCK_REF_KEY)
2158                 parent = ref->parent;
2159         else
2160                 ref_root = ref->root;
2161
2162         ins.objectid = node->bytenr;
2163         if (skinny_metadata) {
2164                 ins.offset = ref->level;
2165                 ins.type = BTRFS_METADATA_ITEM_KEY;
2166         } else {
2167                 ins.offset = node->num_bytes;
2168                 ins.type = BTRFS_EXTENT_ITEM_KEY;
2169         }
2170
2171         BUG_ON(node->ref_mod != 1);
2172         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
2173                 BUG_ON(!extent_op || !extent_op->update_flags);
2174                 ret = alloc_reserved_tree_block(trans, root,
2175                                                 parent, ref_root,
2176                                                 extent_op->flags_to_set,
2177                                                 &extent_op->key,
2178                                                 ref->level, &ins);
2179         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
2180                 ret = __btrfs_inc_extent_ref(trans, root, node->bytenr,
2181                                              node->num_bytes, parent, ref_root,
2182                                              ref->level, 0, 1, extent_op);
2183         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
2184                 ret = __btrfs_free_extent(trans, root, node->bytenr,
2185                                           node->num_bytes, parent, ref_root,
2186                                           ref->level, 0, 1, extent_op);
2187         } else {
2188                 BUG();
2189         }
2190         return ret;
2191 }
2192
2193 /* helper function to actually process a single delayed ref entry */
2194 static int run_one_delayed_ref(struct btrfs_trans_handle *trans,
2195                                struct btrfs_root *root,
2196                                struct btrfs_delayed_ref_node *node,
2197                                struct btrfs_delayed_extent_op *extent_op,
2198                                int insert_reserved)
2199 {
2200         int ret = 0;
2201
2202         if (trans->aborted)
2203                 return 0;
2204
2205         if (btrfs_delayed_ref_is_head(node)) {
2206                 struct btrfs_delayed_ref_head *head;
2207                 /*
2208                  * we've hit the end of the chain and we were supposed
2209                  * to insert this extent into the tree.  But, it got
2210                  * deleted before we ever needed to insert it, so all
2211                  * we have to do is clean up the accounting
2212                  */
2213                 BUG_ON(extent_op);
2214                 head = btrfs_delayed_node_to_head(node);
2215                 if (insert_reserved) {
2216                         btrfs_pin_extent(root, node->bytenr,
2217                                          node->num_bytes, 1);
2218                         if (head->is_data) {
2219                                 ret = btrfs_del_csums(trans, root,
2220                                                       node->bytenr,
2221                                                       node->num_bytes);
2222                         }
2223                 }
2224                 return ret;
2225         }
2226
2227         if (node->type == BTRFS_TREE_BLOCK_REF_KEY ||
2228             node->type == BTRFS_SHARED_BLOCK_REF_KEY)
2229                 ret = run_delayed_tree_ref(trans, root, node, extent_op,
2230                                            insert_reserved);
2231         else if (node->type == BTRFS_EXTENT_DATA_REF_KEY ||
2232                  node->type == BTRFS_SHARED_DATA_REF_KEY)
2233                 ret = run_delayed_data_ref(trans, root, node, extent_op,
2234                                            insert_reserved);
2235         else
2236                 BUG();
2237         return ret;
2238 }
2239
2240 static noinline struct btrfs_delayed_ref_node *
2241 select_delayed_ref(struct btrfs_delayed_ref_head *head)
2242 {
2243         struct rb_node *node;
2244         struct btrfs_delayed_ref_node *ref;
2245         int action = BTRFS_ADD_DELAYED_REF;
2246 again:
2247         /*
2248          * select delayed ref of type BTRFS_ADD_DELAYED_REF first.
2249          * this prevents ref count from going down to zero when
2250          * there still are pending delayed ref.
2251          */
2252         node = rb_prev(&head->node.rb_node);
2253         while (1) {
2254                 if (!node)
2255                         break;
2256                 ref = rb_entry(node, struct btrfs_delayed_ref_node,
2257                                 rb_node);
2258                 if (ref->bytenr != head->node.bytenr)
2259                         break;
2260                 if (ref->action == action)
2261                         return ref;
2262                 node = rb_prev(node);
2263         }
2264         if (action == BTRFS_ADD_DELAYED_REF) {
2265                 action = BTRFS_DROP_DELAYED_REF;
2266                 goto again;
2267         }
2268         return NULL;
2269 }
2270
2271 /*
2272  * Returns 0 on success or if called with an already aborted transaction.
2273  * Returns -ENOMEM or -EIO on failure and will abort the transaction.
2274  */
2275 static noinline int run_clustered_refs(struct btrfs_trans_handle *trans,
2276                                        struct btrfs_root *root,
2277                                        struct list_head *cluster)
2278 {
2279         struct btrfs_delayed_ref_root *delayed_refs;
2280         struct btrfs_delayed_ref_node *ref;
2281         struct btrfs_delayed_ref_head *locked_ref = NULL;
2282         struct btrfs_delayed_extent_op *extent_op;
2283         struct btrfs_fs_info *fs_info = root->fs_info;
2284         int ret;
2285         int count = 0;
2286         int must_insert_reserved = 0;
2287
2288         delayed_refs = &trans->transaction->delayed_refs;
2289         while (1) {
2290                 if (!locked_ref) {
2291                         /* pick a new head ref from the cluster list */
2292                         if (list_empty(cluster))
2293                                 break;
2294
2295                         locked_ref = list_entry(cluster->next,
2296                                      struct btrfs_delayed_ref_head, cluster);
2297
2298                         /* grab the lock that says we are going to process
2299                          * all the refs for this head */
2300                         ret = btrfs_delayed_ref_lock(trans, locked_ref);
2301
2302                         /*
2303                          * we may have dropped the spin lock to get the head
2304                          * mutex lock, and that might have given someone else
2305                          * time to free the head.  If that's true, it has been
2306                          * removed from our list and we can move on.
2307                          */
2308                         if (ret == -EAGAIN) {
2309                                 locked_ref = NULL;
2310                                 count++;
2311                                 continue;
2312                         }
2313                 }
2314
2315                 /*
2316                  * We need to try and merge add/drops of the same ref since we
2317                  * can run into issues with relocate dropping the implicit ref
2318                  * and then it being added back again before the drop can
2319                  * finish.  If we merged anything we need to re-loop so we can
2320                  * get a good ref.
2321                  */
2322                 btrfs_merge_delayed_refs(trans, fs_info, delayed_refs,
2323                                          locked_ref);
2324
2325                 /*
2326                  * locked_ref is the head node, so we have to go one
2327                  * node back for any delayed ref updates
2328                  */
2329                 ref = select_delayed_ref(locked_ref);
2330
2331                 if (ref && ref->seq &&
2332                     btrfs_check_delayed_seq(fs_info, delayed_refs, ref->seq)) {
2333                         /*
2334                          * there are still refs with lower seq numbers in the
2335                          * process of being added. Don't run this ref yet.
2336                          */
2337                         list_del_init(&locked_ref->cluster);
2338                         btrfs_delayed_ref_unlock(locked_ref);
2339                         locked_ref = NULL;
2340                         delayed_refs->num_heads_ready++;
2341                         spin_unlock(&delayed_refs->lock);
2342                         cond_resched();
2343                         spin_lock(&delayed_refs->lock);
2344                         continue;
2345                 }
2346
2347                 /*
2348                  * record the must insert reserved flag before we
2349                  * drop the spin lock.
2350                  */
2351                 must_insert_reserved = locked_ref->must_insert_reserved;
2352                 locked_ref->must_insert_reserved = 0;
2353
2354                 extent_op = locked_ref->extent_op;
2355                 locked_ref->extent_op = NULL;
2356
2357                 if (!ref) {
2358                         /* All delayed refs have been processed, Go ahead
2359                          * and send the head node to run_one_delayed_ref,
2360                          * so that any accounting fixes can happen
2361                          */
2362                         ref = &locked_ref->node;
2363
2364                         if (extent_op && must_insert_reserved) {
2365                                 btrfs_free_delayed_extent_op(extent_op);
2366                                 extent_op = NULL;
2367                         }
2368
2369                         if (extent_op) {
2370                                 spin_unlock(&delayed_refs->lock);
2371
2372                                 ret = run_delayed_extent_op(trans, root,
2373                                                             ref, extent_op);
2374                                 btrfs_free_delayed_extent_op(extent_op);
2375
2376                                 if (ret) {
2377                                         btrfs_debug(fs_info, "run_delayed_extent_op returned %d", ret);
2378                                         spin_lock(&delayed_refs->lock);
2379                                         btrfs_delayed_ref_unlock(locked_ref);
2380                                         return ret;
2381                                 }
2382
2383                                 goto next;
2384                         }
2385                 }
2386
2387                 ref->in_tree = 0;
2388                 rb_erase(&ref->rb_node, &delayed_refs->root);
2389                 delayed_refs->num_entries--;
2390                 if (!btrfs_delayed_ref_is_head(ref)) {
2391                         /*
2392                          * when we play the delayed ref, also correct the
2393                          * ref_mod on head
2394                          */
2395                         switch (ref->action) {
2396                         case BTRFS_ADD_DELAYED_REF:
2397                         case BTRFS_ADD_DELAYED_EXTENT:
2398                                 locked_ref->node.ref_mod -= ref->ref_mod;
2399                                 break;
2400                         case BTRFS_DROP_DELAYED_REF:
2401                                 locked_ref->node.ref_mod += ref->ref_mod;
2402                                 break;
2403                         default:
2404                                 WARN_ON(1);
2405                         }
2406                 }
2407                 spin_unlock(&delayed_refs->lock);
2408
2409                 ret = run_one_delayed_ref(trans, root, ref, extent_op,
2410                                           must_insert_reserved);
2411
2412                 btrfs_free_delayed_extent_op(extent_op);
2413                 if (ret) {
2414                         btrfs_delayed_ref_unlock(locked_ref);
2415                         btrfs_put_delayed_ref(ref);
2416                         btrfs_debug(fs_info, "run_one_delayed_ref returned %d", ret);
2417                         spin_lock(&delayed_refs->lock);
2418                         return ret;
2419                 }
2420
2421                 /*
2422                  * If this node is a head, that means all the refs in this head
2423                  * have been dealt with, and we will pick the next head to deal
2424                  * with, so we must unlock the head and drop it from the cluster
2425                  * list before we release it.
2426                  */
2427                 if (btrfs_delayed_ref_is_head(ref)) {
2428                         list_del_init(&locked_ref->cluster);
2429                         btrfs_delayed_ref_unlock(locked_ref);
2430                         locked_ref = NULL;
2431                 }
2432                 btrfs_put_delayed_ref(ref);
2433                 count++;
2434 next:
2435                 cond_resched();
2436                 spin_lock(&delayed_refs->lock);
2437         }
2438         return count;
2439 }
2440
2441 #ifdef SCRAMBLE_DELAYED_REFS
2442 /*
2443  * Normally delayed refs get processed in ascending bytenr order. This
2444  * correlates in most cases to the order added. To expose dependencies on this
2445  * order, we start to process the tree in the middle instead of the beginning
2446  */
2447 static u64 find_middle(struct rb_root *root)
2448 {
2449         struct rb_node *n = root->rb_node;
2450         struct btrfs_delayed_ref_node *entry;
2451         int alt = 1;
2452         u64 middle;
2453         u64 first = 0, last = 0;
2454
2455         n = rb_first(root);
2456         if (n) {
2457                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2458                 first = entry->bytenr;
2459         }
2460         n = rb_last(root);
2461         if (n) {
2462                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2463                 last = entry->bytenr;
2464         }
2465         n = root->rb_node;
2466
2467         while (n) {
2468                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2469                 WARN_ON(!entry->in_tree);
2470
2471                 middle = entry->bytenr;
2472
2473                 if (alt)
2474                         n = n->rb_left;
2475                 else
2476                         n = n->rb_right;
2477
2478                 alt = 1 - alt;
2479         }
2480         return middle;
2481 }
2482 #endif
2483
2484 int btrfs_delayed_refs_qgroup_accounting(struct btrfs_trans_handle *trans,
2485                                          struct btrfs_fs_info *fs_info)
2486 {
2487         struct qgroup_update *qgroup_update;
2488         int ret = 0;
2489
2490         if (list_empty(&trans->qgroup_ref_list) !=
2491             !trans->delayed_ref_elem.seq) {
2492                 /* list without seq or seq without list */
2493                 btrfs_err(fs_info,
2494                         "qgroup accounting update error, list is%s empty, seq is %#x.%x",
2495                         list_empty(&trans->qgroup_ref_list) ? "" : " not",
2496                         (u32)(trans->delayed_ref_elem.seq >> 32),
2497                         (u32)trans->delayed_ref_elem.seq);
2498                 BUG();
2499         }
2500
2501         if (!trans->delayed_ref_elem.seq)
2502                 return 0;
2503
2504         while (!list_empty(&trans->qgroup_ref_list)) {
2505                 qgroup_update = list_first_entry(&trans->qgroup_ref_list,
2506                                                  struct qgroup_update, list);
2507                 list_del(&qgroup_update->list);
2508                 if (!ret)
2509                         ret = btrfs_qgroup_account_ref(
2510                                         trans, fs_info, qgroup_update->node,
2511                                         qgroup_update->extent_op);
2512                 kfree(qgroup_update);
2513         }
2514
2515         btrfs_put_tree_mod_seq(fs_info, &trans->delayed_ref_elem);
2516
2517         return ret;
2518 }
2519
2520 static int refs_newer(struct btrfs_delayed_ref_root *delayed_refs, int seq,
2521                       int count)
2522 {
2523         int val = atomic_read(&delayed_refs->ref_seq);
2524
2525         if (val < seq || val >= seq + count)
2526                 return 1;
2527         return 0;
2528 }
2529
2530 static inline u64 heads_to_leaves(struct btrfs_root *root, u64 heads)
2531 {
2532         u64 num_bytes;
2533
2534         num_bytes = heads * (sizeof(struct btrfs_extent_item) +
2535                              sizeof(struct btrfs_extent_inline_ref));
2536         if (!btrfs_fs_incompat(root->fs_info, SKINNY_METADATA))
2537                 num_bytes += heads * sizeof(struct btrfs_tree_block_info);
2538
2539         /*
2540          * We don't ever fill up leaves all the way so multiply by 2 just to be
2541          * closer to what we're really going to want to ouse.
2542          */
2543         return div64_u64(num_bytes, BTRFS_LEAF_DATA_SIZE(root));
2544 }
2545
2546 int btrfs_should_throttle_delayed_refs(struct btrfs_trans_handle *trans,
2547                                        struct btrfs_root *root)
2548 {
2549         struct btrfs_block_rsv *global_rsv;
2550         u64 num_heads = trans->transaction->delayed_refs.num_heads_ready;
2551         u64 num_bytes;
2552         int ret = 0;
2553
2554         num_bytes = btrfs_calc_trans_metadata_size(root, 1);
2555         num_heads = heads_to_leaves(root, num_heads);
2556         if (num_heads > 1)
2557                 num_bytes += (num_heads - 1) * root->leafsize;
2558         num_bytes <<= 1;
2559         global_rsv = &root->fs_info->global_block_rsv;
2560
2561         /*
2562          * If we can't allocate any more chunks lets make sure we have _lots_ of
2563          * wiggle room since running delayed refs can create more delayed refs.
2564          */
2565         if (global_rsv->space_info->full)
2566                 num_bytes <<= 1;
2567
2568         spin_lock(&global_rsv->lock);
2569         if (global_rsv->reserved <= num_bytes)
2570                 ret = 1;
2571         spin_unlock(&global_rsv->lock);
2572         return ret;
2573 }
2574
2575 /*
2576  * this starts processing the delayed reference count updates and
2577  * extent insertions we have queued up so far.  count can be
2578  * 0, which means to process everything in the tree at the start
2579  * of the run (but not newly added entries), or it can be some target
2580  * number you'd like to process.
2581  *
2582  * Returns 0 on success or if called with an aborted transaction
2583  * Returns <0 on error and aborts the transaction
2584  */
2585 int btrfs_run_delayed_refs(struct btrfs_trans_handle *trans,
2586                            struct btrfs_root *root, unsigned long count)
2587 {
2588         struct rb_node *node;
2589         struct btrfs_delayed_ref_root *delayed_refs;
2590         struct btrfs_delayed_ref_node *ref;
2591         struct list_head cluster;
2592         int ret;
2593         u64 delayed_start;
2594         int run_all = count == (unsigned long)-1;
2595         int run_most = 0;
2596         int loops;
2597
2598         /* We'll clean this up in btrfs_cleanup_transaction */
2599         if (trans->aborted)
2600                 return 0;
2601
2602         if (root == root->fs_info->extent_root)
2603                 root = root->fs_info->tree_root;
2604
2605         btrfs_delayed_refs_qgroup_accounting(trans, root->fs_info);
2606
2607         delayed_refs = &trans->transaction->delayed_refs;
2608         INIT_LIST_HEAD(&cluster);
2609         if (count == 0) {
2610                 count = delayed_refs->num_entries * 2;
2611                 run_most = 1;
2612         }
2613
2614         if (!run_all && !run_most) {
2615                 int old;
2616                 int seq = atomic_read(&delayed_refs->ref_seq);
2617
2618 progress:
2619                 old = atomic_cmpxchg(&delayed_refs->procs_running_refs, 0, 1);
2620                 if (old) {
2621                         DEFINE_WAIT(__wait);
2622                         if (delayed_refs->flushing ||
2623                             !btrfs_should_throttle_delayed_refs(trans, root))
2624                                 return 0;
2625
2626                         prepare_to_wait(&delayed_refs->wait, &__wait,
2627                                         TASK_UNINTERRUPTIBLE);
2628
2629                         old = atomic_cmpxchg(&delayed_refs->procs_running_refs, 0, 1);
2630                         if (old) {
2631                                 schedule();
2632                                 finish_wait(&delayed_refs->wait, &__wait);
2633
2634                                 if (!refs_newer(delayed_refs, seq, 256))
2635                                         goto progress;
2636                                 else
2637                                         return 0;
2638                         } else {
2639                                 finish_wait(&delayed_refs->wait, &__wait);
2640                                 goto again;
2641                         }
2642                 }
2643
2644         } else {
2645                 atomic_inc(&delayed_refs->procs_running_refs);
2646         }
2647
2648 again:
2649         loops = 0;
2650         spin_lock(&delayed_refs->lock);
2651
2652 #ifdef SCRAMBLE_DELAYED_REFS
2653         delayed_refs->run_delayed_start = find_middle(&delayed_refs->root);
2654 #endif
2655
2656         while (1) {
2657                 if (!(run_all || run_most) &&
2658                     !btrfs_should_throttle_delayed_refs(trans, root))
2659                         break;
2660
2661                 /*
2662                  * go find something we can process in the rbtree.  We start at
2663                  * the beginning of the tree, and then build a cluster
2664                  * of refs to process starting at the first one we are able to
2665                  * lock
2666                  */
2667                 delayed_start = delayed_refs->run_delayed_start;
2668                 ret = btrfs_find_ref_cluster(trans, &cluster,
2669                                              delayed_refs->run_delayed_start);
2670                 if (ret)
2671                         break;
2672
2673                 ret = run_clustered_refs(trans, root, &cluster);
2674                 if (ret < 0) {
2675                         btrfs_release_ref_cluster(&cluster);
2676                         spin_unlock(&delayed_refs->lock);
2677                         btrfs_abort_transaction(trans, root, ret);
2678                         atomic_dec(&delayed_refs->procs_running_refs);
2679                         wake_up(&delayed_refs->wait);
2680                         return ret;
2681                 }
2682
2683                 atomic_add(ret, &delayed_refs->ref_seq);
2684
2685                 count -= min_t(unsigned long, ret, count);
2686
2687                 if (count == 0)
2688                         break;
2689
2690                 if (delayed_start >= delayed_refs->run_delayed_start) {
2691                         if (loops == 0) {
2692                                 /*
2693                                  * btrfs_find_ref_cluster looped. let's do one
2694                                  * more cycle. if we don't run any delayed ref
2695                                  * during that cycle (because we can't because
2696                                  * all of them are blocked), bail out.
2697                                  */
2698                                 loops = 1;
2699                         } else {
2700                                 /*
2701                                  * no runnable refs left, stop trying
2702                                  */
2703                                 BUG_ON(run_all);
2704                                 break;
2705                         }
2706                 }
2707                 if (ret) {
2708                         /* refs were run, let's reset staleness detection */
2709                         loops = 0;
2710                 }
2711         }
2712
2713         if (run_all) {
2714                 if (!list_empty(&trans->new_bgs)) {
2715                         spin_unlock(&delayed_refs->lock);
2716                         btrfs_create_pending_block_groups(trans, root);
2717                         spin_lock(&delayed_refs->lock);
2718                 }
2719
2720                 node = rb_first(&delayed_refs->root);
2721                 if (!node)
2722                         goto out;
2723                 count = (unsigned long)-1;
2724
2725                 while (node) {
2726                         ref = rb_entry(node, struct btrfs_delayed_ref_node,
2727                                        rb_node);
2728                         if (btrfs_delayed_ref_is_head(ref)) {
2729                                 struct btrfs_delayed_ref_head *head;
2730
2731                                 head = btrfs_delayed_node_to_head(ref);
2732                                 atomic_inc(&ref->refs);
2733
2734                                 spin_unlock(&delayed_refs->lock);
2735                                 /*
2736                                  * Mutex was contended, block until it's
2737                                  * released and try again
2738                                  */
2739                                 mutex_lock(&head->mutex);
2740                                 mutex_unlock(&head->mutex);
2741
2742                                 btrfs_put_delayed_ref(ref);
2743                                 cond_resched();
2744                                 goto again;
2745                         }
2746                         node = rb_next(node);
2747                 }
2748                 spin_unlock(&delayed_refs->lock);
2749                 schedule_timeout(1);
2750                 goto again;
2751         }
2752 out:
2753         atomic_dec(&delayed_refs->procs_running_refs);
2754         smp_mb();
2755         if (waitqueue_active(&delayed_refs->wait))
2756                 wake_up(&delayed_refs->wait);
2757
2758         spin_unlock(&delayed_refs->lock);
2759         assert_qgroups_uptodate(trans);
2760         return 0;
2761 }
2762
2763 int btrfs_set_disk_extent_flags(struct btrfs_trans_handle *trans,
2764                                 struct btrfs_root *root,
2765                                 u64 bytenr, u64 num_bytes, u64 flags,
2766                                 int level, int is_data)
2767 {
2768         struct btrfs_delayed_extent_op *extent_op;
2769         int ret;
2770
2771         extent_op = btrfs_alloc_delayed_extent_op();
2772         if (!extent_op)
2773                 return -ENOMEM;
2774
2775         extent_op->flags_to_set = flags;
2776         extent_op->update_flags = 1;
2777         extent_op->update_key = 0;
2778         extent_op->is_data = is_data ? 1 : 0;
2779         extent_op->level = level;
2780
2781         ret = btrfs_add_delayed_extent_op(root->fs_info, trans, bytenr,
2782                                           num_bytes, extent_op);
2783         if (ret)
2784                 btrfs_free_delayed_extent_op(extent_op);
2785         return ret;
2786 }
2787
2788 static noinline int check_delayed_ref(struct btrfs_trans_handle *trans,
2789                                       struct btrfs_root *root,
2790                                       struct btrfs_path *path,
2791                                       u64 objectid, u64 offset, u64 bytenr)
2792 {
2793         struct btrfs_delayed_ref_head *head;
2794         struct btrfs_delayed_ref_node *ref;
2795         struct btrfs_delayed_data_ref *data_ref;
2796         struct btrfs_delayed_ref_root *delayed_refs;
2797         struct rb_node *node;
2798         int ret = 0;
2799
2800         ret = -ENOENT;
2801         delayed_refs = &trans->transaction->delayed_refs;
2802         spin_lock(&delayed_refs->lock);
2803         head = btrfs_find_delayed_ref_head(trans, bytenr);
2804         if (!head)
2805                 goto out;
2806
2807         if (!mutex_trylock(&head->mutex)) {
2808                 atomic_inc(&head->node.refs);
2809                 spin_unlock(&delayed_refs->lock);
2810
2811                 btrfs_release_path(path);
2812
2813                 /*
2814                  * Mutex was contended, block until it's released and let
2815                  * caller try again
2816                  */
2817                 mutex_lock(&head->mutex);
2818                 mutex_unlock(&head->mutex);
2819                 btrfs_put_delayed_ref(&head->node);
2820                 return -EAGAIN;
2821         }
2822
2823         node = rb_prev(&head->node.rb_node);
2824         if (!node)
2825                 goto out_unlock;
2826
2827         ref = rb_entry(node, struct btrfs_delayed_ref_node, rb_node);
2828
2829         if (ref->bytenr != bytenr)
2830                 goto out_unlock;
2831
2832         ret = 1;
2833         if (ref->type != BTRFS_EXTENT_DATA_REF_KEY)
2834                 goto out_unlock;
2835
2836         data_ref = btrfs_delayed_node_to_data_ref(ref);
2837
2838         node = rb_prev(node);
2839         if (node) {
2840                 int seq = ref->seq;
2841
2842                 ref = rb_entry(node, struct btrfs_delayed_ref_node, rb_node);
2843                 if (ref->bytenr == bytenr && ref->seq == seq)
2844                         goto out_unlock;
2845         }
2846
2847         if (data_ref->root != root->root_key.objectid ||
2848             data_ref->objectid != objectid || data_ref->offset != offset)
2849                 goto out_unlock;
2850
2851         ret = 0;
2852 out_unlock:
2853         mutex_unlock(&head->mutex);
2854 out:
2855         spin_unlock(&delayed_refs->lock);
2856         return ret;
2857 }
2858
2859 static noinline int check_committed_ref(struct btrfs_trans_handle *trans,
2860                                         struct btrfs_root *root,
2861                                         struct btrfs_path *path,
2862                                         u64 objectid, u64 offset, u64 bytenr)
2863 {
2864         struct btrfs_root *extent_root = root->fs_info->extent_root;
2865         struct extent_buffer *leaf;
2866         struct btrfs_extent_data_ref *ref;
2867         struct btrfs_extent_inline_ref *iref;
2868         struct btrfs_extent_item *ei;
2869         struct btrfs_key key;
2870         u32 item_size;
2871         int ret;
2872
2873         key.objectid = bytenr;
2874         key.offset = (u64)-1;
2875         key.type = BTRFS_EXTENT_ITEM_KEY;
2876
2877         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
2878         if (ret < 0)
2879                 goto out;
2880         BUG_ON(ret == 0); /* Corruption */
2881
2882         ret = -ENOENT;
2883         if (path->slots[0] == 0)
2884                 goto out;
2885
2886         path->slots[0]--;
2887         leaf = path->nodes[0];
2888         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
2889
2890         if (key.objectid != bytenr || key.type != BTRFS_EXTENT_ITEM_KEY)
2891                 goto out;
2892
2893         ret = 1;
2894         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2895 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
2896         if (item_size < sizeof(*ei)) {
2897                 WARN_ON(item_size != sizeof(struct btrfs_extent_item_v0));
2898                 goto out;
2899         }
2900 #endif
2901         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2902
2903         if (item_size != sizeof(*ei) +
2904             btrfs_extent_inline_ref_size(BTRFS_EXTENT_DATA_REF_KEY))
2905                 goto out;
2906
2907         if (btrfs_extent_generation(leaf, ei) <=
2908             btrfs_root_last_snapshot(&root->root_item))
2909                 goto out;
2910
2911         iref = (struct btrfs_extent_inline_ref *)(ei + 1);
2912         if (btrfs_extent_inline_ref_type(leaf, iref) !=
2913             BTRFS_EXTENT_DATA_REF_KEY)
2914                 goto out;
2915
2916         ref = (struct btrfs_extent_data_ref *)(&iref->offset);
2917         if (btrfs_extent_refs(leaf, ei) !=
2918             btrfs_extent_data_ref_count(leaf, ref) ||
2919             btrfs_extent_data_ref_root(leaf, ref) !=
2920             root->root_key.objectid ||
2921             btrfs_extent_data_ref_objectid(leaf, ref) != objectid ||
2922             btrfs_extent_data_ref_offset(leaf, ref) != offset)
2923                 goto out;
2924
2925         ret = 0;
2926 out:
2927         return ret;
2928 }
2929
2930 int btrfs_cross_ref_exist(struct btrfs_trans_handle *trans,
2931                           struct btrfs_root *root,
2932                           u64 objectid, u64 offset, u64 bytenr)
2933 {
2934         struct btrfs_path *path;
2935         int ret;
2936         int ret2;
2937
2938         path = btrfs_alloc_path();
2939         if (!path)
2940                 return -ENOENT;
2941
2942         do {
2943                 ret = check_committed_ref(trans, root, path, objectid,
2944                                           offset, bytenr);
2945                 if (ret && ret != -ENOENT)
2946                         goto out;
2947
2948                 ret2 = check_delayed_ref(trans, root, path, objectid,
2949                                          offset, bytenr);
2950         } while (ret2 == -EAGAIN);
2951
2952         if (ret2 && ret2 != -ENOENT) {
2953                 ret = ret2;
2954                 goto out;
2955         }
2956
2957         if (ret != -ENOENT || ret2 != -ENOENT)
2958                 ret = 0;
2959 out:
2960         btrfs_free_path(path);
2961         if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID)
2962                 WARN_ON(ret > 0);
2963         return ret;
2964 }
2965
2966 static int __btrfs_mod_ref(struct btrfs_trans_handle *trans,
2967                            struct btrfs_root *root,
2968                            struct extent_buffer *buf,
2969                            int full_backref, int inc, int for_cow)
2970 {
2971         u64 bytenr;
2972         u64 num_bytes;
2973         u64 parent;
2974         u64 ref_root;
2975         u32 nritems;
2976         struct btrfs_key key;
2977         struct btrfs_file_extent_item *fi;
2978         int i;
2979         int level;
2980         int ret = 0;
2981         int (*process_func)(struct btrfs_trans_handle *, struct btrfs_root *,
2982                             u64, u64, u64, u64, u64, u64, int);
2983
2984         ref_root = btrfs_header_owner(buf);
2985         nritems = btrfs_header_nritems(buf);
2986         level = btrfs_header_level(buf);
2987
2988         if (!root->ref_cows && level == 0)
2989                 return 0;
2990
2991         if (inc)
2992                 process_func = btrfs_inc_extent_ref;
2993         else
2994                 process_func = btrfs_free_extent;
2995
2996         if (full_backref)
2997                 parent = buf->start;
2998         else
2999                 parent = 0;
3000
3001         for (i = 0; i < nritems; i++) {
3002                 if (level == 0) {
3003                         btrfs_item_key_to_cpu(buf, &key, i);
3004                         if (btrfs_key_type(&key) != BTRFS_EXTENT_DATA_KEY)
3005                                 continue;
3006                         fi = btrfs_item_ptr(buf, i,
3007                                             struct btrfs_file_extent_item);
3008                         if (btrfs_file_extent_type(buf, fi) ==
3009                             BTRFS_FILE_EXTENT_INLINE)
3010                                 continue;
3011                         bytenr = btrfs_file_extent_disk_bytenr(buf, fi);
3012                         if (bytenr == 0)
3013                                 continue;
3014
3015                         num_bytes = btrfs_file_extent_disk_num_bytes(buf, fi);
3016                         key.offset -= btrfs_file_extent_offset(buf, fi);
3017                         ret = process_func(trans, root, bytenr, num_bytes,
3018                                            parent, ref_root, key.objectid,
3019                                            key.offset, for_cow);
3020                         if (ret)
3021                                 goto fail;
3022                 } else {
3023                         bytenr = btrfs_node_blockptr(buf, i);
3024                         num_bytes = btrfs_level_size(root, level - 1);
3025                         ret = process_func(trans, root, bytenr, num_bytes,
3026                                            parent, ref_root, level - 1, 0,
3027                                            for_cow);
3028                         if (ret)
3029                                 goto fail;
3030                 }
3031         }
3032         return 0;
3033 fail:
3034         return ret;
3035 }
3036
3037 int btrfs_inc_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
3038                   struct extent_buffer *buf, int full_backref, int for_cow)
3039 {
3040         return __btrfs_mod_ref(trans, root, buf, full_backref, 1, for_cow);
3041 }
3042
3043 int btrfs_dec_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
3044                   struct extent_buffer *buf, int full_backref, int for_cow)
3045 {
3046         return __btrfs_mod_ref(trans, root, buf, full_backref, 0, for_cow);
3047 }
3048
3049 static int write_one_cache_group(struct btrfs_trans_handle *trans,
3050                                  struct btrfs_root *root,
3051                                  struct btrfs_path *path,
3052                                  struct btrfs_block_group_cache *cache)
3053 {
3054         int ret;
3055         struct btrfs_root *extent_root = root->fs_info->extent_root;
3056         unsigned long bi;
3057         struct extent_buffer *leaf;
3058
3059         ret = btrfs_search_slot(trans, extent_root, &cache->key, path, 0, 1);
3060         if (ret < 0)
3061                 goto fail;
3062         BUG_ON(ret); /* Corruption */
3063
3064         leaf = path->nodes[0];
3065         bi = btrfs_item_ptr_offset(leaf, path->slots[0]);
3066         write_extent_buffer(leaf, &cache->item, bi, sizeof(cache->item));
3067         btrfs_mark_buffer_dirty(leaf);
3068         btrfs_release_path(path);
3069 fail:
3070         if (ret) {
3071                 btrfs_abort_transaction(trans, root, ret);
3072                 return ret;
3073         }
3074         return 0;
3075
3076 }
3077
3078 static struct btrfs_block_group_cache *
3079 next_block_group(struct btrfs_root *root,
3080                  struct btrfs_block_group_cache *cache)
3081 {
3082         struct rb_node *node;
3083         spin_lock(&root->fs_info->block_group_cache_lock);
3084         node = rb_next(&cache->cache_node);
3085         btrfs_put_block_group(cache);
3086         if (node) {
3087                 cache = rb_entry(node, struct btrfs_block_group_cache,
3088                                  cache_node);
3089                 btrfs_get_block_group(cache);
3090         } else
3091                 cache = NULL;
3092         spin_unlock(&root->fs_info->block_group_cache_lock);
3093         return cache;
3094 }
3095
3096 static int cache_save_setup(struct btrfs_block_group_cache *block_group,
3097                             struct btrfs_trans_handle *trans,
3098                             struct btrfs_path *path)
3099 {
3100         struct btrfs_root *root = block_group->fs_info->tree_root;
3101         struct inode *inode = NULL;
3102         u64 alloc_hint = 0;
3103         int dcs = BTRFS_DC_ERROR;
3104         int num_pages = 0;
3105         int retries = 0;
3106         int ret = 0;
3107
3108         /*
3109          * If this block group is smaller than 100 megs don't bother caching the
3110          * block group.
3111          */
3112         if (block_group->key.offset < (100 * 1024 * 1024)) {
3113                 spin_lock(&block_group->lock);
3114                 block_group->disk_cache_state = BTRFS_DC_WRITTEN;
3115                 spin_unlock(&block_group->lock);
3116                 return 0;
3117         }
3118
3119 again:
3120         inode = lookup_free_space_inode(root, block_group, path);
3121         if (IS_ERR(inode) && PTR_ERR(inode) != -ENOENT) {
3122                 ret = PTR_ERR(inode);
3123                 btrfs_release_path(path);
3124                 goto out;
3125         }
3126
3127         if (IS_ERR(inode)) {
3128                 BUG_ON(retries);
3129                 retries++;
3130
3131                 if (block_group->ro)
3132                         goto out_free;
3133
3134                 ret = create_free_space_inode(root, trans, block_group, path);
3135                 if (ret)
3136                         goto out_free;
3137                 goto again;
3138         }
3139
3140         /* We've already setup this transaction, go ahead and exit */
3141         if (block_group->cache_generation == trans->transid &&
3142             i_size_read(inode)) {
3143                 dcs = BTRFS_DC_SETUP;
3144                 goto out_put;
3145         }
3146
3147         /*
3148          * We want to set the generation to 0, that way if anything goes wrong
3149          * from here on out we know not to trust this cache when we load up next
3150          * time.
3151          */
3152         BTRFS_I(inode)->generation = 0;
3153         ret = btrfs_update_inode(trans, root, inode);
3154         WARN_ON(ret);
3155
3156         if (i_size_read(inode) > 0) {
3157                 ret = btrfs_check_trunc_cache_free_space(root,
3158                                         &root->fs_info->global_block_rsv);
3159                 if (ret)
3160                         goto out_put;
3161
3162                 ret = btrfs_truncate_free_space_cache(root, trans, path,
3163                                                       inode);
3164                 if (ret)
3165                         goto out_put;
3166         }
3167
3168         spin_lock(&block_group->lock);
3169         if (block_group->cached != BTRFS_CACHE_FINISHED ||
3170             !btrfs_test_opt(root, SPACE_CACHE)) {
3171                 /*
3172                  * don't bother trying to write stuff out _if_
3173                  * a) we're not cached,
3174                  * b) we're with nospace_cache mount option.
3175                  */
3176                 dcs = BTRFS_DC_WRITTEN;
3177                 spin_unlock(&block_group->lock);
3178                 goto out_put;
3179         }
3180         spin_unlock(&block_group->lock);
3181
3182         /*
3183          * Try to preallocate enough space based on how big the block group is.
3184          * Keep in mind this has to include any pinned space which could end up
3185          * taking up quite a bit since it's not folded into the other space
3186          * cache.
3187          */
3188         num_pages = (int)div64_u64(block_group->key.offset, 256 * 1024 * 1024);
3189         if (!num_pages)
3190                 num_pages = 1;
3191
3192         num_pages *= 16;
3193         num_pages *= PAGE_CACHE_SIZE;
3194
3195         ret = btrfs_check_data_free_space(inode, num_pages);
3196         if (ret)
3197                 goto out_put;
3198
3199         ret = btrfs_prealloc_file_range_trans(inode, trans, 0, 0, num_pages,
3200                                               num_pages, num_pages,
3201                                               &alloc_hint);
3202         if (!ret)
3203                 dcs = BTRFS_DC_SETUP;
3204         btrfs_free_reserved_data_space(inode, num_pages);
3205
3206 out_put:
3207         iput(inode);
3208 out_free:
3209         btrfs_release_path(path);
3210 out:
3211         spin_lock(&block_group->lock);
3212         if (!ret && dcs == BTRFS_DC_SETUP)
3213                 block_group->cache_generation = trans->transid;
3214         block_group->disk_cache_state = dcs;
3215         spin_unlock(&block_group->lock);
3216
3217         return ret;
3218 }
3219
3220 int btrfs_write_dirty_block_groups(struct btrfs_trans_handle *trans,
3221                                    struct btrfs_root *root)
3222 {
3223         struct btrfs_block_group_cache *cache;
3224         int err = 0;
3225         struct btrfs_path *path;
3226         u64 last = 0;
3227
3228         path = btrfs_alloc_path();
3229         if (!path)
3230                 return -ENOMEM;
3231
3232 again:
3233         while (1) {
3234                 cache = btrfs_lookup_first_block_group(root->fs_info, last);
3235                 while (cache) {
3236                         if (cache->disk_cache_state == BTRFS_DC_CLEAR)
3237                                 break;
3238                         cache = next_block_group(root, cache);
3239                 }
3240                 if (!cache) {
3241                         if (last == 0)
3242                                 break;
3243                         last = 0;
3244                         continue;
3245                 }
3246                 err = cache_save_setup(cache, trans, path);
3247                 last = cache->key.objectid + cache->key.offset;
3248                 btrfs_put_block_group(cache);
3249         }
3250
3251         while (1) {
3252                 if (last == 0) {
3253                         err = btrfs_run_delayed_refs(trans, root,
3254                                                      (unsigned long)-1);
3255                         if (err) /* File system offline */
3256                                 goto out;
3257                 }
3258
3259                 cache = btrfs_lookup_first_block_group(root->fs_info, last);
3260                 while (cache) {
3261                         if (cache->disk_cache_state == BTRFS_DC_CLEAR) {
3262                                 btrfs_put_block_group(cache);
3263                                 goto again;
3264                         }
3265
3266                         if (cache->dirty)
3267                                 break;
3268                         cache = next_block_group(root, cache);
3269                 }
3270                 if (!cache) {
3271                         if (last == 0)
3272                                 break;
3273                         last = 0;
3274                         continue;
3275                 }
3276
3277                 if (cache->disk_cache_state == BTRFS_DC_SETUP)
3278                         cache->disk_cache_state = BTRFS_DC_NEED_WRITE;
3279                 cache->dirty = 0;
3280                 last = cache->key.objectid + cache->key.offset;
3281
3282                 err = write_one_cache_group(trans, root, path, cache);
3283                 if (err) /* File system offline */
3284                         goto out;
3285
3286                 btrfs_put_block_group(cache);
3287         }
3288
3289         while (1) {
3290                 /*
3291                  * I don't think this is needed since we're just marking our
3292                  * preallocated extent as written, but just in case it can't
3293                  * hurt.
3294                  */
3295                 if (last == 0) {
3296                         err = btrfs_run_delayed_refs(trans, root,
3297                                                      (unsigned long)-1);
3298                         if (err) /* File system offline */
3299                                 goto out;
3300                 }
3301
3302                 cache = btrfs_lookup_first_block_group(root->fs_info, last);
3303                 while (cache) {
3304                         /*
3305                          * Really this shouldn't happen, but it could if we
3306                          * couldn't write the entire preallocated extent and
3307                          * splitting the extent resulted in a new block.
3308                          */
3309                         if (cache->dirty) {
3310                                 btrfs_put_block_group(cache);
3311                                 goto again;
3312                         }
3313                         if (cache->disk_cache_state == BTRFS_DC_NEED_WRITE)
3314                                 break;
3315                         cache = next_block_group(root, cache);
3316                 }
3317                 if (!cache) {
3318                         if (last == 0)
3319                                 break;
3320                         last = 0;
3321                         continue;
3322                 }
3323
3324                 err = btrfs_write_out_cache(root, trans, cache, path);
3325
3326                 /*
3327                  * If we didn't have an error then the cache state is still
3328                  * NEED_WRITE, so we can set it to WRITTEN.
3329                  */
3330                 if (!err && cache->disk_cache_state == BTRFS_DC_NEED_WRITE)
3331                         cache->disk_cache_state = BTRFS_DC_WRITTEN;
3332                 last = cache->key.objectid + cache->key.offset;
3333                 btrfs_put_block_group(cache);
3334         }
3335 out:
3336
3337         btrfs_free_path(path);
3338         return err;
3339 }
3340
3341 int btrfs_extent_readonly(struct btrfs_root *root, u64 bytenr)
3342 {
3343         struct btrfs_block_group_cache *block_group;
3344         int readonly = 0;
3345
3346         block_group = btrfs_lookup_block_group(root->fs_info, bytenr);
3347         if (!block_group || block_group->ro)
3348                 readonly = 1;
3349         if (block_group)
3350                 btrfs_put_block_group(block_group);
3351         return readonly;
3352 }
3353
3354 static int update_space_info(struct btrfs_fs_info *info, u64 flags,
3355                              u64 total_bytes, u64 bytes_used,
3356                              struct btrfs_space_info **space_info)
3357 {
3358         struct btrfs_space_info *found;
3359         int i;
3360         int factor;
3361         int ret;
3362
3363         if (flags & (BTRFS_BLOCK_GROUP_DUP | BTRFS_BLOCK_GROUP_RAID1 |
3364                      BTRFS_BLOCK_GROUP_RAID10))
3365                 factor = 2;
3366         else
3367                 factor = 1;
3368
3369         found = __find_space_info(info, flags);
3370         if (found) {
3371                 spin_lock(&found->lock);
3372                 found->total_bytes += total_bytes;
3373                 found->disk_total += total_bytes * factor;
3374                 found->bytes_used += bytes_used;
3375                 found->disk_used += bytes_used * factor;
3376                 found->full = 0;
3377                 spin_unlock(&found->lock);
3378                 *space_info = found;
3379                 return 0;
3380         }
3381         found = kzalloc(sizeof(*found), GFP_NOFS);
3382         if (!found)
3383                 return -ENOMEM;
3384
3385         ret = percpu_counter_init(&found->total_bytes_pinned, 0);
3386         if (ret) {
3387                 kfree(found);
3388                 return ret;
3389         }
3390
3391         for (i = 0; i < BTRFS_NR_RAID_TYPES; i++)
3392                 INIT_LIST_HEAD(&found->block_groups[i]);
3393         init_rwsem(&found->groups_sem);
3394         spin_lock_init(&found->lock);
3395         found->flags = flags & BTRFS_BLOCK_GROUP_TYPE_MASK;
3396         found->total_bytes = total_bytes;
3397         found->disk_total = total_bytes * factor;
3398         found->bytes_used = bytes_used;
3399         found->disk_used = bytes_used * factor;
3400         found->bytes_pinned = 0;
3401         found->bytes_reserved = 0;
3402         found->bytes_readonly = 0;
3403         found->bytes_may_use = 0;
3404         found->full = 0;
3405         found->force_alloc = CHUNK_ALLOC_NO_FORCE;
3406         found->chunk_alloc = 0;
3407         found->flush = 0;
3408         init_waitqueue_head(&found->wait);
3409         *space_info = found;
3410         list_add_rcu(&found->list, &info->space_info);
3411         if (flags & BTRFS_BLOCK_GROUP_DATA)
3412                 info->data_sinfo = found;
3413         return 0;
3414 }
3415
3416 static void set_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
3417 {
3418         u64 extra_flags = chunk_to_extended(flags) &
3419                                 BTRFS_EXTENDED_PROFILE_MASK;
3420
3421         write_seqlock(&fs_info->profiles_lock);
3422         if (flags & BTRFS_BLOCK_GROUP_DATA)
3423                 fs_info->avail_data_alloc_bits |= extra_flags;
3424         if (flags & BTRFS_BLOCK_GROUP_METADATA)
3425                 fs_info->avail_metadata_alloc_bits |= extra_flags;
3426         if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
3427                 fs_info->avail_system_alloc_bits |= extra_flags;
3428         write_sequnlock(&fs_info->profiles_lock);
3429 }
3430
3431 /*
3432  * returns target flags in extended format or 0 if restripe for this
3433  * chunk_type is not in progress
3434  *
3435  * should be called with either volume_mutex or balance_lock held
3436  */
3437 static u64 get_restripe_target(struct btrfs_fs_info *fs_info, u64 flags)
3438 {
3439         struct btrfs_balance_control *bctl = fs_info->balance_ctl;
3440         u64 target = 0;
3441
3442         if (!bctl)
3443                 return 0;
3444
3445         if (flags & BTRFS_BLOCK_GROUP_DATA &&
3446             bctl->data.flags & BTRFS_BALANCE_ARGS_CONVERT) {
3447                 target = BTRFS_BLOCK_GROUP_DATA | bctl->data.target;
3448         } else if (flags & BTRFS_BLOCK_GROUP_SYSTEM &&
3449                    bctl->sys.flags & BTRFS_BALANCE_ARGS_CONVERT) {
3450                 target = BTRFS_BLOCK_GROUP_SYSTEM | bctl->sys.target;
3451         } else if (flags & BTRFS_BLOCK_GROUP_METADATA &&
3452                    bctl->meta.flags & BTRFS_BALANCE_ARGS_CONVERT) {
3453                 target = BTRFS_BLOCK_GROUP_METADATA | bctl->meta.target;
3454         }
3455
3456         return target;
3457 }
3458
3459 /*
3460  * @flags: available profiles in extended format (see ctree.h)
3461  *
3462  * Returns reduced profile in chunk format.  If profile changing is in
3463  * progress (either running or paused) picks the target profile (if it's
3464  * already available), otherwise falls back to plain reducing.
3465  */
3466 static u64 btrfs_reduce_alloc_profile(struct btrfs_root *root, u64 flags)
3467 {
3468         /*
3469          * we add in the count of missing devices because we want
3470          * to make sure that any RAID levels on a degraded FS
3471          * continue to be honored.
3472          */
3473         u64 num_devices = root->fs_info->fs_devices->rw_devices +
3474                 root->fs_info->fs_devices->missing_devices;
3475         u64 target;
3476         u64 tmp;
3477
3478         /*
3479          * see if restripe for this chunk_type is in progress, if so
3480          * try to reduce to the target profile
3481          */
3482         spin_lock(&root->fs_info->balance_lock);
3483         target = get_restripe_target(root->fs_info, flags);
3484         if (target) {
3485                 /* pick target profile only if it's already available */
3486                 if ((flags & target) & BTRFS_EXTENDED_PROFILE_MASK) {
3487                         spin_unlock(&root->fs_info->balance_lock);
3488                         return extended_to_chunk(target);
3489                 }
3490         }
3491         spin_unlock(&root->fs_info->balance_lock);
3492
3493         /* First, mask out the RAID levels which aren't possible */
3494         if (num_devices == 1)
3495                 flags &= ~(BTRFS_BLOCK_GROUP_RAID1 | BTRFS_BLOCK_GROUP_RAID0 |
3496                            BTRFS_BLOCK_GROUP_RAID5);
3497         if (num_devices < 3)
3498                 flags &= ~BTRFS_BLOCK_GROUP_RAID6;
3499         if (num_devices < 4)
3500                 flags &= ~BTRFS_BLOCK_GROUP_RAID10;
3501
3502         tmp = flags & (BTRFS_BLOCK_GROUP_DUP | BTRFS_BLOCK_GROUP_RAID0 |
3503                        BTRFS_BLOCK_GROUP_RAID1 | BTRFS_BLOCK_GROUP_RAID5 |
3504                        BTRFS_BLOCK_GROUP_RAID6 | BTRFS_BLOCK_GROUP_RAID10);
3505         flags &= ~tmp;
3506
3507         if (tmp & BTRFS_BLOCK_GROUP_RAID6)
3508                 tmp = BTRFS_BLOCK_GROUP_RAID6;
3509         else if (tmp & BTRFS_BLOCK_GROUP_RAID5)
3510                 tmp = BTRFS_BLOCK_GROUP_RAID5;
3511         else if (tmp & BTRFS_BLOCK_GROUP_RAID10)
3512                 tmp = BTRFS_BLOCK_GROUP_RAID10;
3513         else if (tmp & BTRFS_BLOCK_GROUP_RAID1)
3514                 tmp = BTRFS_BLOCK_GROUP_RAID1;
3515         else if (tmp & BTRFS_BLOCK_GROUP_RAID0)
3516                 tmp = BTRFS_BLOCK_GROUP_RAID0;
3517
3518         return extended_to_chunk(flags | tmp);
3519 }
3520
3521 static u64 get_alloc_profile(struct btrfs_root *root, u64 flags)
3522 {
3523         unsigned seq;
3524
3525         do {
3526                 seq = read_seqbegin(&root->fs_info->profiles_lock);
3527
3528                 if (flags & BTRFS_BLOCK_GROUP_DATA)
3529                         flags |= root->fs_info->avail_data_alloc_bits;
3530                 else if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
3531                         flags |= root->fs_info->avail_system_alloc_bits;
3532                 else if (flags & BTRFS_BLOCK_GROUP_METADATA)
3533                         flags |= root->fs_info->avail_metadata_alloc_bits;
3534         } while (read_seqretry(&root->fs_info->profiles_lock, seq));
3535
3536         return btrfs_reduce_alloc_profile(root, flags);
3537 }
3538
3539 u64 btrfs_get_alloc_profile(struct btrfs_root *root, int data)
3540 {
3541         u64 flags;
3542         u64 ret;
3543
3544         if (data)
3545                 flags = BTRFS_BLOCK_GROUP_DATA;
3546         else if (root == root->fs_info->chunk_root)
3547                 flags = BTRFS_BLOCK_GROUP_SYSTEM;
3548         else
3549                 flags = BTRFS_BLOCK_GROUP_METADATA;
3550
3551         ret = get_alloc_profile(root, flags);
3552         return ret;
3553 }
3554
3555 /*
3556  * This will check the space that the inode allocates from to make sure we have
3557  * enough space for bytes.
3558  */
3559 int btrfs_check_data_free_space(struct inode *inode, u64 bytes)
3560 {
3561         struct btrfs_space_info *data_sinfo;
3562         struct btrfs_root *root = BTRFS_I(inode)->root;
3563         struct btrfs_fs_info *fs_info = root->fs_info;
3564         u64 used;
3565         int ret = 0, committed = 0, alloc_chunk = 1;
3566
3567         /* make sure bytes are sectorsize aligned */
3568         bytes = ALIGN(bytes, root->sectorsize);
3569
3570         if (root == root->fs_info->tree_root ||
3571             BTRFS_I(inode)->location.objectid == BTRFS_FREE_INO_OBJECTID) {
3572                 alloc_chunk = 0;
3573                 committed = 1;
3574         }
3575
3576         data_sinfo = fs_info->data_sinfo;
3577         if (!data_sinfo)
3578                 goto alloc;
3579
3580 again:
3581         /* make sure we have enough space to handle the data first */
3582         spin_lock(&data_sinfo->lock);
3583         used = data_sinfo->bytes_used + data_sinfo->bytes_reserved +
3584                 data_sinfo->bytes_pinned + data_sinfo->bytes_readonly +
3585                 data_sinfo->bytes_may_use;
3586
3587         if (used + bytes > data_sinfo->total_bytes) {
3588                 struct btrfs_trans_handle *trans;
3589
3590                 /*
3591                  * if we don't have enough free bytes in this space then we need
3592                  * to alloc a new chunk.
3593                  */
3594                 if (!data_sinfo->full && alloc_chunk) {
3595                         u64 alloc_target;
3596
3597                         data_sinfo->force_alloc = CHUNK_ALLOC_FORCE;
3598                         spin_unlock(&data_sinfo->lock);
3599 alloc:
3600                         alloc_target = btrfs_get_alloc_profile(root, 1);
3601                         trans = btrfs_join_transaction(root);
3602                         if (IS_ERR(trans))
3603                                 return PTR_ERR(trans);
3604
3605                         ret = do_chunk_alloc(trans, root->fs_info->extent_root,
3606                                              alloc_target,
3607                                              CHUNK_ALLOC_NO_FORCE);
3608                         btrfs_end_transaction(trans, root);
3609                         if (ret < 0) {
3610                                 if (ret != -ENOSPC)
3611                                         return ret;
3612                                 else
3613                                         goto commit_trans;
3614                         }
3615
3616                         if (!data_sinfo)
3617                                 data_sinfo = fs_info->data_sinfo;
3618
3619                         goto again;
3620                 }
3621
3622                 /*
3623                  * If we don't have enough pinned space to deal with this
3624                  * allocation don't bother committing the transaction.
3625                  */
3626                 if (percpu_counter_compare(&data_sinfo->total_bytes_pinned,
3627                                            bytes) < 0)
3628                         committed = 1;
3629                 spin_unlock(&data_sinfo->lock);
3630
3631                 /* commit the current transaction and try again */
3632 commit_trans:
3633                 if (!committed &&
3634                     !atomic_read(&root->fs_info->open_ioctl_trans)) {
3635                         committed = 1;
3636
3637                         trans = btrfs_join_transaction(root);
3638                         if (IS_ERR(trans))
3639                                 return PTR_ERR(trans);
3640                         ret = btrfs_commit_transaction(trans, root);
3641                         if (ret)
3642                                 return ret;
3643                         goto again;
3644                 }
3645
3646                 return -ENOSPC;
3647         }
3648         data_sinfo->bytes_may_use += bytes;
3649         trace_btrfs_space_reservation(root->fs_info, "space_info",
3650                                       data_sinfo->flags, bytes, 1);
3651         spin_unlock(&data_sinfo->lock);
3652
3653         return 0;
3654 }
3655
3656 /*
3657  * Called if we need to clear a data reservation for this inode.
3658  */
3659 void btrfs_free_reserved_data_space(struct inode *inode, u64 bytes)
3660 {
3661         struct btrfs_root *root = BTRFS_I(inode)->root;
3662         struct btrfs_space_info *data_sinfo;
3663
3664         /* make sure bytes are sectorsize aligned */
3665         bytes = ALIGN(bytes, root->sectorsize);
3666
3667         data_sinfo = root->fs_info->data_sinfo;
3668         spin_lock(&data_sinfo->lock);
3669         WARN_ON(data_sinfo->bytes_may_use < bytes);
3670         data_sinfo->bytes_may_use -= bytes;
3671         trace_btrfs_space_reservation(root->fs_info, "space_info",
3672                                       data_sinfo->flags, bytes, 0);
3673         spin_unlock(&data_sinfo->lock);
3674 }
3675
3676 static void force_metadata_allocation(struct btrfs_fs_info *info)
3677 {
3678         struct list_head *head = &info->space_info;
3679         struct btrfs_space_info *found;
3680
3681         rcu_read_lock();
3682         list_for_each_entry_rcu(found, head, list) {
3683                 if (found->flags & BTRFS_BLOCK_GROUP_METADATA)
3684                         found->force_alloc = CHUNK_ALLOC_FORCE;
3685         }
3686         rcu_read_unlock();
3687 }
3688
3689 static inline u64 calc_global_rsv_need_space(struct btrfs_block_rsv *global)
3690 {
3691         return (global->size << 1);
3692 }
3693
3694 static int should_alloc_chunk(struct btrfs_root *root,
3695                               struct btrfs_space_info *sinfo, int force)
3696 {
3697         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
3698         u64 num_bytes = sinfo->total_bytes - sinfo->bytes_readonly;
3699         u64 num_allocated = sinfo->bytes_used + sinfo->bytes_reserved;
3700         u64 thresh;
3701
3702         if (force == CHUNK_ALLOC_FORCE)
3703                 return 1;
3704
3705         /*
3706          * We need to take into account the global rsv because for all intents
3707          * and purposes it's used space.  Don't worry about locking the
3708          * global_rsv, it doesn't change except when the transaction commits.
3709          */
3710         if (sinfo->flags & BTRFS_BLOCK_GROUP_METADATA)
3711                 num_allocated += calc_global_rsv_need_space(global_rsv);
3712
3713         /*
3714          * in limited mode, we want to have some free space up to
3715          * about 1% of the FS size.
3716          */
3717         if (force == CHUNK_ALLOC_LIMITED) {
3718                 thresh = btrfs_super_total_bytes(root->fs_info->super_copy);
3719                 thresh = max_t(u64, 64 * 1024 * 1024,
3720                                div_factor_fine(thresh, 1));
3721
3722                 if (num_bytes - num_allocated < thresh)
3723                         return 1;
3724         }
3725
3726         if (num_allocated + 2 * 1024 * 1024 < div_factor(num_bytes, 8))
3727                 return 0;
3728         return 1;
3729 }
3730
3731 static u64 get_system_chunk_thresh(struct btrfs_root *root, u64 type)
3732 {
3733         u64 num_dev;
3734
3735         if (type & (BTRFS_BLOCK_GROUP_RAID10 |
3736                     BTRFS_BLOCK_GROUP_RAID0 |
3737                     BTRFS_BLOCK_GROUP_RAID5 |
3738                     BTRFS_BLOCK_GROUP_RAID6))
3739                 num_dev = root->fs_info->fs_devices->rw_devices;
3740         else if (type & BTRFS_BLOCK_GROUP_RAID1)
3741                 num_dev = 2;
3742         else
3743                 num_dev = 1;    /* DUP or single */
3744
3745         /* metadata for updaing devices and chunk tree */
3746         return btrfs_calc_trans_metadata_size(root, num_dev + 1);
3747 }
3748
3749 static void check_system_chunk(struct btrfs_trans_handle *trans,
3750                                struct btrfs_root *root, u64 type)
3751 {
3752         struct btrfs_space_info *info;
3753         u64 left;
3754         u64 thresh;
3755
3756         info = __find_space_info(root->fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
3757         spin_lock(&info->lock);
3758         left = info->total_bytes - info->bytes_used - info->bytes_pinned -
3759                 info->bytes_reserved - info->bytes_readonly;
3760         spin_unlock(&info->lock);
3761
3762         thresh = get_system_chunk_thresh(root, type);
3763         if (left < thresh && btrfs_test_opt(root, ENOSPC_DEBUG)) {
3764                 btrfs_info(root->fs_info, "left=%llu, need=%llu, flags=%llu",
3765                         left, thresh, type);
3766                 dump_space_info(info, 0, 0);
3767         }
3768
3769         if (left < thresh) {
3770                 u64 flags;
3771
3772                 flags = btrfs_get_alloc_profile(root->fs_info->chunk_root, 0);
3773                 btrfs_alloc_chunk(trans, root, flags);
3774         }
3775 }
3776
3777 static int do_chunk_alloc(struct btrfs_trans_handle *trans,
3778                           struct btrfs_root *extent_root, u64 flags, int force)
3779 {
3780         struct btrfs_space_info *space_info;
3781         struct btrfs_fs_info *fs_info = extent_root->fs_info;
3782         int wait_for_alloc = 0;
3783         int ret = 0;
3784
3785         /* Don't re-enter if we're already allocating a chunk */
3786         if (trans->allocating_chunk)
3787                 return -ENOSPC;
3788
3789         space_info = __find_space_info(extent_root->fs_info, flags);
3790         if (!space_info) {
3791                 ret = update_space_info(extent_root->fs_info, flags,
3792                                         0, 0, &space_info);
3793                 BUG_ON(ret); /* -ENOMEM */
3794         }
3795         BUG_ON(!space_info); /* Logic error */
3796
3797 again:
3798         spin_lock(&space_info->lock);
3799         if (force < space_info->force_alloc)
3800                 force = space_info->force_alloc;
3801         if (space_info->full) {
3802                 spin_unlock(&space_info->lock);
3803                 return 0;
3804         }
3805
3806         if (!should_alloc_chunk(extent_root, space_info, force)) {
3807                 spin_unlock(&space_info->lock);
3808                 return 0;
3809         } else if (space_info->chunk_alloc) {
3810                 wait_for_alloc = 1;
3811         } else {
3812                 space_info->chunk_alloc = 1;
3813         }
3814
3815         spin_unlock(&space_info->lock);
3816
3817         mutex_lock(&fs_info->chunk_mutex);
3818
3819         /*
3820          * The chunk_mutex is held throughout the entirety of a chunk
3821          * allocation, so once we've acquired the chunk_mutex we know that the
3822          * other guy is done and we need to recheck and see if we should
3823          * allocate.
3824          */
3825         if (wait_for_alloc) {
3826                 mutex_unlock(&fs_info->chunk_mutex);
3827                 wait_for_alloc = 0;
3828                 goto again;
3829         }
3830
3831         trans->allocating_chunk = true;
3832
3833         /*
3834          * If we have mixed data/metadata chunks we want to make sure we keep
3835          * allocating mixed chunks instead of individual chunks.
3836          */
3837         if (btrfs_mixed_space_info(space_info))
3838                 flags |= (BTRFS_BLOCK_GROUP_DATA | BTRFS_BLOCK_GROUP_METADATA);
3839
3840         /*
3841          * if we're doing a data chunk, go ahead and make sure that
3842          * we keep a reasonable number of metadata chunks allocated in the
3843          * FS as well.
3844          */
3845         if (flags & BTRFS_BLOCK_GROUP_DATA && fs_info->metadata_ratio) {
3846                 fs_info->data_chunk_allocations++;
3847                 if (!(fs_info->data_chunk_allocations %
3848                       fs_info->metadata_ratio))
3849                         force_metadata_allocation(fs_info);
3850         }
3851
3852         /*
3853          * Check if we have enough space in SYSTEM chunk because we may need
3854          * to update devices.
3855          */
3856         check_system_chunk(trans, extent_root, flags);
3857
3858         ret = btrfs_alloc_chunk(trans, extent_root, flags);
3859         trans->allocating_chunk = false;
3860
3861         spin_lock(&space_info->lock);
3862         if (ret < 0 && ret != -ENOSPC)
3863                 goto out;
3864         if (ret)
3865                 space_info->full = 1;
3866         else
3867                 ret = 1;
3868
3869         space_info->force_alloc = CHUNK_ALLOC_NO_FORCE;
3870 out:
3871         space_info->chunk_alloc = 0;
3872         spin_unlock(&space_info->lock);
3873         mutex_unlock(&fs_info->chunk_mutex);
3874         return ret;
3875 }
3876
3877 static int can_overcommit(struct btrfs_root *root,
3878                           struct btrfs_space_info *space_info, u64 bytes,
3879                           enum btrfs_reserve_flush_enum flush)
3880 {
3881         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
3882         u64 profile = btrfs_get_alloc_profile(root, 0);
3883         u64 space_size;
3884         u64 avail;
3885         u64 used;
3886         u64 to_add;
3887
3888         used = space_info->bytes_used + space_info->bytes_reserved +
3889                 space_info->bytes_pinned + space_info->bytes_readonly;
3890
3891         /*
3892          * We only want to allow over committing if we have lots of actual space
3893          * free, but if we don't have enough space to handle the global reserve
3894          * space then we could end up having a real enospc problem when trying
3895          * to allocate a chunk or some other such important allocation.
3896          */
3897         spin_lock(&global_rsv->lock);
3898         space_size = calc_global_rsv_need_space(global_rsv);
3899         spin_unlock(&global_rsv->lock);
3900         if (used + space_size >= space_info->total_bytes)
3901                 return 0;
3902
3903         used += space_info->bytes_may_use;
3904
3905         spin_lock(&root->fs_info->free_chunk_lock);
3906         avail = root->fs_info->free_chunk_space;
3907         spin_unlock(&root->fs_info->free_chunk_lock);
3908
3909         /*
3910          * If we have dup, raid1 or raid10 then only half of the free
3911          * space is actually useable.  For raid56, the space info used
3912          * doesn't include the parity drive, so we don't have to
3913          * change the math
3914          */
3915         if (profile & (BTRFS_BLOCK_GROUP_DUP |
3916                        BTRFS_BLOCK_GROUP_RAID1 |
3917                        BTRFS_BLOCK_GROUP_RAID10))
3918                 avail >>= 1;
3919
3920         to_add = space_info->total_bytes;
3921
3922         /*
3923          * If we aren't flushing all things, let us overcommit up to
3924          * 1/2th of the space. If we can flush, don't let us overcommit
3925          * too much, let it overcommit up to 1/8 of the space.
3926          */
3927         if (flush == BTRFS_RESERVE_FLUSH_ALL)
3928                 to_add >>= 3;
3929         else
3930                 to_add >>= 1;
3931
3932         /*
3933          * Limit the overcommit to the amount of free space we could possibly
3934          * allocate for chunks.
3935          */
3936         to_add = min(avail, to_add);
3937
3938         if (used + bytes < space_info->total_bytes + to_add)
3939                 return 1;
3940         return 0;
3941 }
3942
3943 static void btrfs_writeback_inodes_sb_nr(struct btrfs_root *root,
3944                                          unsigned long nr_pages)
3945 {
3946         struct super_block *sb = root->fs_info->sb;
3947
3948         if (down_read_trylock(&sb->s_umount)) {
3949                 writeback_inodes_sb_nr(sb, nr_pages, WB_REASON_FS_FREE_SPACE);
3950                 up_read(&sb->s_umount);
3951         } else {
3952                 /*
3953                  * We needn't worry the filesystem going from r/w to r/o though
3954                  * we don't acquire ->s_umount mutex, because the filesystem
3955                  * should guarantee the delalloc inodes list be empty after
3956                  * the filesystem is readonly(all dirty pages are written to
3957                  * the disk).
3958                  */
3959                 btrfs_start_all_delalloc_inodes(root->fs_info, 0);
3960                 if (!current->journal_info)
3961                         btrfs_wait_all_ordered_extents(root->fs_info, 0);
3962         }
3963 }
3964
3965 /*
3966  * shrink metadata reservation for delalloc
3967  */
3968 static void shrink_delalloc(struct btrfs_root *root, u64 to_reclaim, u64 orig,
3969                             bool wait_ordered)
3970 {
3971         struct btrfs_block_rsv *block_rsv;
3972         struct btrfs_space_info *space_info;
3973         struct btrfs_trans_handle *trans;
3974         u64 delalloc_bytes;
3975         u64 max_reclaim;
3976         long time_left;
3977         unsigned long nr_pages = (2 * 1024 * 1024) >> PAGE_CACHE_SHIFT;
3978         int loops = 0;
3979         enum btrfs_reserve_flush_enum flush;
3980
3981         trans = (struct btrfs_trans_handle *)current->journal_info;
3982         block_rsv = &root->fs_info->delalloc_block_rsv;
3983         space_info = block_rsv->space_info;
3984
3985         smp_mb();
3986         delalloc_bytes = percpu_counter_sum_positive(
3987                                                 &root->fs_info->delalloc_bytes);
3988         if (delalloc_bytes == 0) {
3989                 if (trans)
3990                         return;
3991                 btrfs_wait_all_ordered_extents(root->fs_info, 0);
3992                 return;
3993         }
3994
3995         while (delalloc_bytes && loops < 3) {
3996                 max_reclaim = min(delalloc_bytes, to_reclaim);
3997                 nr_pages = max_reclaim >> PAGE_CACHE_SHIFT;
3998                 btrfs_writeback_inodes_sb_nr(root, nr_pages);
3999                 /*
4000                  * We need to wait for the async pages to actually start before
4001                  * we do anything.
4002                  */
4003                 wait_event(root->fs_info->async_submit_wait,
4004                            !atomic_read(&root->fs_info->async_delalloc_pages));
4005
4006                 if (!trans)
4007                         flush = BTRFS_RESERVE_FLUSH_ALL;
4008                 else
4009                         flush = BTRFS_RESERVE_NO_FLUSH;
4010                 spin_lock(&space_info->lock);
4011                 if (can_overcommit(root, space_info, orig, flush)) {
4012                         spin_unlock(&space_info->lock);
4013                         break;
4014                 }
4015                 spin_unlock(&space_info->lock);
4016
4017                 loops++;
4018                 if (wait_ordered && !trans) {
4019                         btrfs_wait_all_ordered_extents(root->fs_info, 0);
4020                 } else {
4021                         time_left = schedule_timeout_killable(1);
4022                         if (time_left)
4023                                 break;
4024                 }
4025                 smp_mb();
4026                 delalloc_bytes = percpu_counter_sum_positive(
4027                                                 &root->fs_info->delalloc_bytes);
4028         }
4029 }
4030
4031 /**
4032  * maybe_commit_transaction - possibly commit the transaction if its ok to
4033  * @root - the root we're allocating for
4034  * @bytes - the number of bytes we want to reserve
4035  * @force - force the commit
4036  *
4037  * This will check to make sure that committing the transaction will actually
4038  * get us somewhere and then commit the transaction if it does.  Otherwise it
4039  * will return -ENOSPC.
4040  */
4041 static int may_commit_transaction(struct btrfs_root *root,
4042                                   struct btrfs_space_info *space_info,
4043                                   u64 bytes, int force)
4044 {
4045         struct btrfs_block_rsv *delayed_rsv = &root->fs_info->delayed_block_rsv;
4046         struct btrfs_trans_handle *trans;
4047
4048         trans = (struct btrfs_trans_handle *)current->journal_info;
4049         if (trans)
4050                 return -EAGAIN;
4051
4052         if (force)
4053                 goto commit;
4054
4055         /* See if there is enough pinned space to make this reservation */
4056         spin_lock(&space_info->lock);
4057         if (percpu_counter_compare(&space_info->total_bytes_pinned,
4058                                    bytes) >= 0) {
4059                 spin_unlock(&space_info->lock);
4060                 goto commit;
4061         }
4062         spin_unlock(&space_info->lock);
4063
4064         /*
4065          * See if there is some space in the delayed insertion reservation for
4066          * this reservation.
4067          */
4068         if (space_info != delayed_rsv->space_info)
4069                 return -ENOSPC;
4070
4071         spin_lock(&space_info->lock);
4072         spin_lock(&delayed_rsv->lock);
4073         if (percpu_counter_compare(&space_info->total_bytes_pinned,
4074                                    bytes - delayed_rsv->size) >= 0) {
4075                 spin_unlock(&delayed_rsv->lock);
4076                 spin_unlock(&space_info->lock);
4077                 return -ENOSPC;
4078         }
4079         spin_unlock(&delayed_rsv->lock);
4080         spin_unlock(&space_info->lock);
4081
4082 commit:
4083         trans = btrfs_join_transaction(root);
4084         if (IS_ERR(trans))
4085                 return -ENOSPC;
4086
4087         return btrfs_commit_transaction(trans, root);
4088 }
4089
4090 enum flush_state {
4091         FLUSH_DELAYED_ITEMS_NR  =       1,
4092         FLUSH_DELAYED_ITEMS     =       2,
4093         FLUSH_DELALLOC          =       3,
4094         FLUSH_DELALLOC_WAIT     =       4,
4095         ALLOC_CHUNK             =       5,
4096         COMMIT_TRANS            =       6,
4097 };
4098
4099 static int flush_space(struct btrfs_root *root,
4100                        struct btrfs_space_info *space_info, u64 num_bytes,
4101                        u64 orig_bytes, int state)
4102 {
4103         struct btrfs_trans_handle *trans;
4104         int nr;
4105         int ret = 0;
4106
4107         switch (state) {
4108         case FLUSH_DELAYED_ITEMS_NR:
4109         case FLUSH_DELAYED_ITEMS:
4110                 if (state == FLUSH_DELAYED_ITEMS_NR) {
4111                         u64 bytes = btrfs_calc_trans_metadata_size(root, 1);
4112
4113                         nr = (int)div64_u64(num_bytes, bytes);
4114                         if (!nr)
4115                                 nr = 1;
4116                         nr *= 2;
4117                 } else {
4118                         nr = -1;
4119                 }
4120                 trans = btrfs_join_transaction(root);
4121                 if (IS_ERR(trans)) {
4122                         ret = PTR_ERR(trans);
4123                         break;
4124                 }
4125                 ret = btrfs_run_delayed_items_nr(trans, root, nr);
4126                 btrfs_end_transaction(trans, root);
4127                 break;
4128         case FLUSH_DELALLOC:
4129         case FLUSH_DELALLOC_WAIT:
4130                 shrink_delalloc(root, num_bytes, orig_bytes,
4131                                 state == FLUSH_DELALLOC_WAIT);
4132                 break;
4133         case ALLOC_CHUNK:
4134                 trans = btrfs_join_transaction(root);
4135                 if (IS_ERR(trans)) {
4136                         ret = PTR_ERR(trans);
4137                         break;
4138                 }
4139                 ret = do_chunk_alloc(trans, root->fs_info->extent_root,
4140                                      btrfs_get_alloc_profile(root, 0),
4141                                      CHUNK_ALLOC_NO_FORCE);
4142                 btrfs_end_transaction(trans, root);
4143                 if (ret == -ENOSPC)
4144                         ret = 0;
4145                 break;
4146         case COMMIT_TRANS:
4147                 ret = may_commit_transaction(root, space_info, orig_bytes, 0);
4148                 break;
4149         default:
4150                 ret = -ENOSPC;
4151                 break;
4152         }
4153
4154         return ret;
4155 }
4156 /**
4157  * reserve_metadata_bytes - try to reserve bytes from the block_rsv's space
4158  * @root - the root we're allocating for
4159  * @block_rsv - the block_rsv we're allocating for
4160  * @orig_bytes - the number of bytes we want
4161  * @flush - whether or not we can flush to make our reservation
4162  *
4163  * This will reserve orgi_bytes number of bytes from the space info associated
4164  * with the block_rsv.  If there is not enough space it will make an attempt to
4165  * flush out space to make room.  It will do this by flushing delalloc if
4166  * possible or committing the transaction.  If flush is 0 then no attempts to
4167  * regain reservations will be made and this will fail if there is not enough
4168  * space already.
4169  */
4170 static int reserve_metadata_bytes(struct btrfs_root *root,
4171                                   struct btrfs_block_rsv *block_rsv,
4172                                   u64 orig_bytes,
4173                                   enum btrfs_reserve_flush_enum flush)
4174 {
4175         struct btrfs_space_info *space_info = block_rsv->space_info;
4176         u64 used;
4177         u64 num_bytes = orig_bytes;
4178         int flush_state = FLUSH_DELAYED_ITEMS_NR;
4179         int ret = 0;
4180         bool flushing = false;
4181
4182 again:
4183         ret = 0;
4184         spin_lock(&space_info->lock);
4185         /*
4186          * We only want to wait if somebody other than us is flushing and we
4187          * are actually allowed to flush all things.
4188          */
4189         while (flush == BTRFS_RESERVE_FLUSH_ALL && !flushing &&
4190                space_info->flush) {
4191                 spin_unlock(&space_info->lock);
4192                 /*
4193                  * If we have a trans handle we can't wait because the flusher
4194                  * may have to commit the transaction, which would mean we would
4195                  * deadlock since we are waiting for the flusher to finish, but
4196                  * hold the current transaction open.
4197                  */
4198                 if (current->journal_info)
4199                         return -EAGAIN;
4200                 ret = wait_event_killable(space_info->wait, !space_info->flush);
4201                 /* Must have been killed, return */
4202                 if (ret)
4203                         return -EINTR;
4204
4205                 spin_lock(&space_info->lock);
4206         }
4207
4208         ret = -ENOSPC;
4209         used = space_info->bytes_used + space_info->bytes_reserved +
4210                 space_info->bytes_pinned + space_info->bytes_readonly +
4211                 space_info->bytes_may_use;
4212
4213         /*
4214          * The idea here is that we've not already over-reserved the block group
4215          * then we can go ahead and save our reservation first and then start
4216          * flushing if we need to.  Otherwise if we've already overcommitted
4217          * lets start flushing stuff first and then come back and try to make
4218          * our reservation.
4219          */
4220         if (used <= space_info->total_bytes) {
4221                 if (used + orig_bytes <= space_info->total_bytes) {
4222                         space_info->bytes_may_use += orig_bytes;
4223                         trace_btrfs_space_reservation(root->fs_info,
4224                                 "space_info", space_info->flags, orig_bytes, 1);
4225                         ret = 0;
4226                 } else {
4227                         /*
4228                          * Ok set num_bytes to orig_bytes since we aren't
4229                          * overocmmitted, this way we only try and reclaim what
4230                          * we need.
4231                          */
4232                         num_bytes = orig_bytes;
4233                 }
4234         } else {
4235                 /*
4236                  * Ok we're over committed, set num_bytes to the overcommitted
4237                  * amount plus the amount of bytes that we need for this
4238                  * reservation.
4239                  */
4240                 num_bytes = used - space_info->total_bytes +
4241                         (orig_bytes * 2);
4242         }
4243
4244         if (ret && can_overcommit(root, space_info, orig_bytes, flush)) {
4245                 space_info->bytes_may_use += orig_bytes;
4246                 trace_btrfs_space_reservation(root->fs_info, "space_info",
4247                                               space_info->flags, orig_bytes,
4248                                               1);
4249                 ret = 0;
4250         }
4251
4252         /*
4253          * Couldn't make our reservation, save our place so while we're trying
4254          * to reclaim space we can actually use it instead of somebody else
4255          * stealing it from us.
4256          *
4257          * We make the other tasks wait for the flush only when we can flush
4258          * all things.
4259          */
4260         if (ret && flush != BTRFS_RESERVE_NO_FLUSH) {
4261                 flushing = true;
4262                 space_info->flush = 1;
4263         }
4264
4265         spin_unlock(&space_info->lock);
4266
4267         if (!ret || flush == BTRFS_RESERVE_NO_FLUSH)
4268                 goto out;
4269
4270         ret = flush_space(root, space_info, num_bytes, orig_bytes,
4271                           flush_state);
4272         flush_state++;
4273
4274         /*
4275          * If we are FLUSH_LIMIT, we can not flush delalloc, or the deadlock
4276          * would happen. So skip delalloc flush.
4277          */
4278         if (flush == BTRFS_RESERVE_FLUSH_LIMIT &&
4279             (flush_state == FLUSH_DELALLOC ||
4280              flush_state == FLUSH_DELALLOC_WAIT))
4281                 flush_state = ALLOC_CHUNK;
4282
4283         if (!ret)
4284                 goto again;
4285         else if (flush == BTRFS_RESERVE_FLUSH_LIMIT &&
4286                  flush_state < COMMIT_TRANS)
4287                 goto again;
4288         else if (flush == BTRFS_RESERVE_FLUSH_ALL &&
4289                  flush_state <= COMMIT_TRANS)
4290                 goto again;
4291
4292 out:
4293         if (ret == -ENOSPC &&
4294             unlikely(root->orphan_cleanup_state == ORPHAN_CLEANUP_STARTED)) {
4295                 struct btrfs_block_rsv *global_rsv =
4296                         &root->fs_info->global_block_rsv;
4297
4298                 if (block_rsv != global_rsv &&
4299                     !block_rsv_use_bytes(global_rsv, orig_bytes))
4300                         ret = 0;
4301         }
4302         if (flushing) {
4303                 spin_lock(&space_info->lock);
4304                 space_info->flush = 0;
4305                 wake_up_all(&space_info->wait);
4306                 spin_unlock(&space_info->lock);
4307         }
4308         return ret;
4309 }
4310
4311 static struct btrfs_block_rsv *get_block_rsv(
4312                                         const struct btrfs_trans_handle *trans,
4313                                         const struct btrfs_root *root)
4314 {
4315         struct btrfs_block_rsv *block_rsv = NULL;
4316
4317         if (root->ref_cows)
4318                 block_rsv = trans->block_rsv;
4319
4320         if (root == root->fs_info->csum_root && trans->adding_csums)
4321                 block_rsv = trans->block_rsv;
4322
4323         if (!block_rsv)
4324                 block_rsv = root->block_rsv;
4325
4326         if (!block_rsv)
4327                 block_rsv = &root->fs_info->empty_block_rsv;
4328
4329         return block_rsv;
4330 }
4331
4332 static int block_rsv_use_bytes(struct btrfs_block_rsv *block_rsv,
4333                                u64 num_bytes)
4334 {
4335         int ret = -ENOSPC;
4336         spin_lock(&block_rsv->lock);
4337         if (block_rsv->reserved >= num_bytes) {
4338                 block_rsv->reserved -= num_bytes;
4339                 if (block_rsv->reserved < block_rsv->size)
4340                         block_rsv->full = 0;
4341                 ret = 0;
4342         }
4343         spin_unlock(&block_rsv->lock);
4344         return ret;
4345 }
4346
4347 static void block_rsv_add_bytes(struct btrfs_block_rsv *block_rsv,
4348                                 u64 num_bytes, int update_size)
4349 {
4350         spin_lock(&block_rsv->lock);
4351         block_rsv->reserved += num_bytes;
4352         if (update_size)
4353                 block_rsv->size += num_bytes;
4354         else if (block_rsv->reserved >= block_rsv->size)
4355                 block_rsv->full = 1;
4356         spin_unlock(&block_rsv->lock);
4357 }
4358
4359 int btrfs_cond_migrate_bytes(struct btrfs_fs_info *fs_info,
4360                              struct btrfs_block_rsv *dest, u64 num_bytes,
4361                              int min_factor)
4362 {
4363         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
4364         u64 min_bytes;
4365
4366         if (global_rsv->space_info != dest->space_info)
4367                 return -ENOSPC;
4368
4369         spin_lock(&global_rsv->lock);
4370         min_bytes = div_factor(global_rsv->size, min_factor);
4371         if (global_rsv->reserved < min_bytes + num_bytes) {
4372                 spin_unlock(&global_rsv->lock);
4373                 return -ENOSPC;
4374         }
4375         global_rsv->reserved -= num_bytes;
4376         if (global_rsv->reserved < global_rsv->size)
4377                 global_rsv->full = 0;
4378         spin_unlock(&global_rsv->lock);
4379
4380         block_rsv_add_bytes(dest, num_bytes, 1);
4381         return 0;
4382 }
4383
4384 static void block_rsv_release_bytes(struct btrfs_fs_info *fs_info,
4385                                     struct btrfs_block_rsv *block_rsv,
4386                                     struct btrfs_block_rsv *dest, u64 num_bytes)
4387 {
4388         struct btrfs_space_info *space_info = block_rsv->space_info;
4389
4390         spin_lock(&block_rsv->lock);
4391         if (num_bytes == (u64)-1)
4392                 num_bytes = block_rsv->size;
4393         block_rsv->size -= num_bytes;
4394         if (block_rsv->reserved >= block_rsv->size) {
4395                 num_bytes = block_rsv->reserved - block_rsv->size;
4396                 block_rsv->reserved = block_rsv->size;
4397                 block_rsv->full = 1;
4398         } else {
4399                 num_bytes = 0;
4400         }
4401         spin_unlock(&block_rsv->lock);
4402
4403         if (num_bytes > 0) {
4404                 if (dest) {
4405                         spin_lock(&dest->lock);
4406                         if (!dest->full) {
4407                                 u64 bytes_to_add;
4408
4409                                 bytes_to_add = dest->size - dest->reserved;
4410                                 bytes_to_add = min(num_bytes, bytes_to_add);
4411                                 dest->reserved += bytes_to_add;
4412                                 if (dest->reserved >= dest->size)
4413                                         dest->full = 1;
4414                                 num_bytes -= bytes_to_add;
4415                         }
4416                         spin_unlock(&dest->lock);
4417                 }
4418                 if (num_bytes) {
4419                         spin_lock(&space_info->lock);
4420                         space_info->bytes_may_use -= num_bytes;
4421                         trace_btrfs_space_reservation(fs_info, "space_info",
4422                                         space_info->flags, num_bytes, 0);
4423                         space_info->reservation_progress++;
4424                         spin_unlock(&space_info->lock);
4425                 }
4426         }
4427 }
4428
4429 static int block_rsv_migrate_bytes(struct btrfs_block_rsv *src,
4430                                    struct btrfs_block_rsv *dst, u64 num_bytes)
4431 {
4432         int ret;
4433
4434         ret = block_rsv_use_bytes(src, num_bytes);
4435         if (ret)
4436                 return ret;
4437
4438         block_rsv_add_bytes(dst, num_bytes, 1);
4439         return 0;
4440 }
4441
4442 void btrfs_init_block_rsv(struct btrfs_block_rsv *rsv, unsigned short type)
4443 {
4444         memset(rsv, 0, sizeof(*rsv));
4445         spin_lock_init(&rsv->lock);
4446         rsv->type = type;
4447 }
4448
4449 struct btrfs_block_rsv *btrfs_alloc_block_rsv(struct btrfs_root *root,
4450                                               unsigned short type)
4451 {
4452         struct btrfs_block_rsv *block_rsv;
4453         struct btrfs_fs_info *fs_info = root->fs_info;
4454
4455         block_rsv = kmalloc(sizeof(*block_rsv), GFP_NOFS);
4456         if (!block_rsv)
4457                 return NULL;
4458
4459         btrfs_init_block_rsv(block_rsv, type);
4460         block_rsv->space_info = __find_space_info(fs_info,
4461                                                   BTRFS_BLOCK_GROUP_METADATA);
4462         return block_rsv;
4463 }
4464
4465 void btrfs_free_block_rsv(struct btrfs_root *root,
4466                           struct btrfs_block_rsv *rsv)
4467 {
4468         if (!rsv)
4469                 return;
4470         btrfs_block_rsv_release(root, rsv, (u64)-1);
4471         kfree(rsv);
4472 }
4473
4474 int btrfs_block_rsv_add(struct btrfs_root *root,
4475                         struct btrfs_block_rsv *block_rsv, u64 num_bytes,
4476                         enum btrfs_reserve_flush_enum flush)
4477 {
4478         int ret;
4479
4480         if (num_bytes == 0)
4481                 return 0;
4482
4483         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
4484         if (!ret) {
4485                 block_rsv_add_bytes(block_rsv, num_bytes, 1);
4486                 return 0;
4487         }
4488
4489         return ret;
4490 }
4491
4492 int btrfs_block_rsv_check(struct btrfs_root *root,
4493                           struct btrfs_block_rsv *block_rsv, int min_factor)
4494 {
4495         u64 num_bytes = 0;
4496         int ret = -ENOSPC;
4497
4498         if (!block_rsv)
4499                 return 0;
4500
4501         spin_lock(&block_rsv->lock);
4502         num_bytes = div_factor(block_rsv->size, min_factor);
4503         if (block_rsv->reserved >= num_bytes)
4504                 ret = 0;
4505         spin_unlock(&block_rsv->lock);
4506
4507         return ret;
4508 }
4509
4510 int btrfs_block_rsv_refill(struct btrfs_root *root,
4511                            struct btrfs_block_rsv *block_rsv, u64 min_reserved,
4512                            enum btrfs_reserve_flush_enum flush)
4513 {
4514         u64 num_bytes = 0;
4515         int ret = -ENOSPC;
4516
4517         if (!block_rsv)
4518                 return 0;
4519
4520         spin_lock(&block_rsv->lock);
4521         num_bytes = min_reserved;
4522         if (block_rsv->reserved >= num_bytes)
4523                 ret = 0;
4524         else
4525                 num_bytes -= block_rsv->reserved;
4526         spin_unlock(&block_rsv->lock);
4527
4528         if (!ret)
4529                 return 0;
4530
4531         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
4532         if (!ret) {
4533                 block_rsv_add_bytes(block_rsv, num_bytes, 0);
4534                 return 0;
4535         }
4536
4537         return ret;
4538 }
4539
4540 int btrfs_block_rsv_migrate(struct btrfs_block_rsv *src_rsv,
4541                             struct btrfs_block_rsv *dst_rsv,
4542                             u64 num_bytes)
4543 {
4544         return block_rsv_migrate_bytes(src_rsv, dst_rsv, num_bytes);
4545 }
4546
4547 void btrfs_block_rsv_release(struct btrfs_root *root,
4548                              struct btrfs_block_rsv *block_rsv,
4549                              u64 num_bytes)
4550 {
4551         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
4552         if (global_rsv->full || global_rsv == block_rsv ||
4553             block_rsv->space_info != global_rsv->space_info)
4554                 global_rsv = NULL;
4555         block_rsv_release_bytes(root->fs_info, block_rsv, global_rsv,
4556                                 num_bytes);
4557 }
4558
4559 /*
4560  * helper to calculate size of global block reservation.
4561  * the desired value is sum of space used by extent tree,
4562  * checksum tree and root tree
4563  */
4564 static u64 calc_global_metadata_size(struct btrfs_fs_info *fs_info)
4565 {
4566         struct btrfs_space_info *sinfo;
4567         u64 num_bytes;
4568         u64 meta_used;
4569         u64 data_used;
4570         int csum_size = btrfs_super_csum_size(fs_info->super_copy);
4571
4572         sinfo = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_DATA);
4573         spin_lock(&sinfo->lock);
4574         data_used = sinfo->bytes_used;
4575         spin_unlock(&sinfo->lock);
4576
4577         sinfo = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
4578         spin_lock(&sinfo->lock);
4579         if (sinfo->flags & BTRFS_BLOCK_GROUP_DATA)
4580                 data_used = 0;
4581         meta_used = sinfo->bytes_used;
4582         spin_unlock(&sinfo->lock);
4583
4584         num_bytes = (data_used >> fs_info->sb->s_blocksize_bits) *
4585                     csum_size * 2;
4586         num_bytes += div64_u64(data_used + meta_used, 50);
4587
4588         if (num_bytes * 3 > meta_used)
4589                 num_bytes = div64_u64(meta_used, 3);
4590
4591         return ALIGN(num_bytes, fs_info->extent_root->leafsize << 10);
4592 }
4593
4594 static void update_global_block_rsv(struct btrfs_fs_info *fs_info)
4595 {
4596         struct btrfs_block_rsv *block_rsv = &fs_info->global_block_rsv;
4597         struct btrfs_space_info *sinfo = block_rsv->space_info;
4598         u64 num_bytes;
4599
4600         num_bytes = calc_global_metadata_size(fs_info);
4601
4602         spin_lock(&sinfo->lock);
4603         spin_lock(&block_rsv->lock);
4604
4605         block_rsv->size = min_t(u64, num_bytes, 512 * 1024 * 1024);
4606
4607         num_bytes = sinfo->bytes_used + sinfo->bytes_pinned +
4608                     sinfo->bytes_reserved + sinfo->bytes_readonly +
4609                     sinfo->bytes_may_use;
4610
4611         if (sinfo->total_bytes > num_bytes) {
4612                 num_bytes = sinfo->total_bytes - num_bytes;
4613                 block_rsv->reserved += num_bytes;
4614                 sinfo->bytes_may_use += num_bytes;
4615                 trace_btrfs_space_reservation(fs_info, "space_info",
4616                                       sinfo->flags, num_bytes, 1);
4617         }
4618
4619         if (block_rsv->reserved >= block_rsv->size) {
4620                 num_bytes = block_rsv->reserved - block_rsv->size;
4621                 sinfo->bytes_may_use -= num_bytes;
4622                 trace_btrfs_space_reservation(fs_info, "space_info",
4623                                       sinfo->flags, num_bytes, 0);
4624                 sinfo->reservation_progress++;
4625                 block_rsv->reserved = block_rsv->size;
4626                 block_rsv->full = 1;
4627         }
4628
4629         spin_unlock(&block_rsv->lock);
4630         spin_unlock(&sinfo->lock);
4631 }
4632
4633 static void init_global_block_rsv(struct btrfs_fs_info *fs_info)
4634 {
4635         struct btrfs_space_info *space_info;
4636
4637         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
4638         fs_info->chunk_block_rsv.space_info = space_info;
4639
4640         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
4641         fs_info->global_block_rsv.space_info = space_info;
4642         fs_info->delalloc_block_rsv.space_info = space_info;
4643         fs_info->trans_block_rsv.space_info = space_info;
4644         fs_info->empty_block_rsv.space_info = space_info;
4645         fs_info->delayed_block_rsv.space_info = space_info;
4646
4647         fs_info->extent_root->block_rsv = &fs_info->global_block_rsv;
4648         fs_info->csum_root->block_rsv = &fs_info->global_block_rsv;
4649         fs_info->dev_root->block_rsv = &fs_info->global_block_rsv;
4650         fs_info->tree_root->block_rsv = &fs_info->global_block_rsv;
4651         if (fs_info->quota_root)
4652                 fs_info->quota_root->block_rsv = &fs_info->global_block_rsv;
4653         fs_info->chunk_root->block_rsv = &fs_info->chunk_block_rsv;
4654
4655         update_global_block_rsv(fs_info);
4656 }
4657
4658 static void release_global_block_rsv(struct btrfs_fs_info *fs_info)
4659 {
4660         block_rsv_release_bytes(fs_info, &fs_info->global_block_rsv, NULL,
4661                                 (u64)-1);
4662         WARN_ON(fs_info->delalloc_block_rsv.size > 0);
4663         WARN_ON(fs_info->delalloc_block_rsv.reserved > 0);
4664         WARN_ON(fs_info->trans_block_rsv.size > 0);
4665         WARN_ON(fs_info->trans_block_rsv.reserved > 0);
4666         WARN_ON(fs_info->chunk_block_rsv.size > 0);
4667         WARN_ON(fs_info->chunk_block_rsv.reserved > 0);
4668         WARN_ON(fs_info->delayed_block_rsv.size > 0);
4669         WARN_ON(fs_info->delayed_block_rsv.reserved > 0);
4670 }
4671
4672 void btrfs_trans_release_metadata(struct btrfs_trans_handle *trans,
4673                                   struct btrfs_root *root)
4674 {
4675         if (!trans->block_rsv)
4676                 return;
4677
4678         if (!trans->bytes_reserved)
4679                 return;
4680
4681         trace_btrfs_space_reservation(root->fs_info, "transaction",
4682                                       trans->transid, trans->bytes_reserved, 0);
4683         btrfs_block_rsv_release(root, trans->block_rsv, trans->bytes_reserved);
4684         trans->bytes_reserved = 0;
4685 }
4686
4687 /* Can only return 0 or -ENOSPC */
4688 int btrfs_orphan_reserve_metadata(struct btrfs_trans_handle *trans,
4689                                   struct inode *inode)
4690 {
4691         struct btrfs_root *root = BTRFS_I(inode)->root;
4692         struct btrfs_block_rsv *src_rsv = get_block_rsv(trans, root);
4693         struct btrfs_block_rsv *dst_rsv = root->orphan_block_rsv;
4694
4695         /*
4696          * We need to hold space in order to delete our orphan item once we've
4697          * added it, so this takes the reservation so we can release it later
4698          * when we are truly done with the orphan item.
4699          */
4700         u64 num_bytes = btrfs_calc_trans_metadata_size(root, 1);
4701         trace_btrfs_space_reservation(root->fs_info, "orphan",
4702                                       btrfs_ino(inode), num_bytes, 1);
4703         return block_rsv_migrate_bytes(src_rsv, dst_rsv, num_bytes);
4704 }
4705
4706 void btrfs_orphan_release_metadata(struct inode *inode)
4707 {
4708         struct btrfs_root *root = BTRFS_I(inode)->root;
4709         u64 num_bytes = btrfs_calc_trans_metadata_size(root, 1);
4710         trace_btrfs_space_reservation(root->fs_info, "orphan",
4711                                       btrfs_ino(inode), num_bytes, 0);
4712         btrfs_block_rsv_release(root, root->orphan_block_rsv, num_bytes);
4713 }
4714
4715 /*
4716  * btrfs_subvolume_reserve_metadata() - reserve space for subvolume operation
4717  * root: the root of the parent directory
4718  * rsv: block reservation
4719  * items: the number of items that we need do reservation
4720  * qgroup_reserved: used to return the reserved size in qgroup
4721  *
4722  * This function is used to reserve the space for snapshot/subvolume
4723  * creation and deletion. Those operations are different with the
4724  * common file/directory operations, they change two fs/file trees
4725  * and root tree, the number of items that the qgroup reserves is
4726  * different with the free space reservation. So we can not use
4727  * the space reseravtion mechanism in start_transaction().
4728  */
4729 int btrfs_subvolume_reserve_metadata(struct btrfs_root *root,
4730                                      struct btrfs_block_rsv *rsv,
4731                                      int items,
4732                                      u64 *qgroup_reserved)
4733 {
4734         u64 num_bytes;
4735         int ret;
4736
4737         if (root->fs_info->quota_enabled) {
4738                 /* One for parent inode, two for dir entries */
4739                 num_bytes = 3 * root->leafsize;
4740                 ret = btrfs_qgroup_reserve(root, num_bytes);
4741                 if (ret)
4742                         return ret;
4743         } else {
4744                 num_bytes = 0;
4745         }
4746
4747         *qgroup_reserved = num_bytes;
4748
4749         num_bytes = btrfs_calc_trans_metadata_size(root, items);
4750         rsv->space_info = __find_space_info(root->fs_info,
4751                                             BTRFS_BLOCK_GROUP_METADATA);
4752         ret = btrfs_block_rsv_add(root, rsv, num_bytes,
4753                                   BTRFS_RESERVE_FLUSH_ALL);
4754         if (ret) {
4755                 if (*qgroup_reserved)
4756                         btrfs_qgroup_free(root, *qgroup_reserved);
4757         }
4758
4759         return ret;
4760 }
4761
4762 void btrfs_subvolume_release_metadata(struct btrfs_root *root,
4763                                       struct btrfs_block_rsv *rsv,
4764                                       u64 qgroup_reserved)
4765 {
4766         btrfs_block_rsv_release(root, rsv, (u64)-1);
4767         if (qgroup_reserved)
4768                 btrfs_qgroup_free(root, qgroup_reserved);
4769 }
4770
4771 /**
4772  * drop_outstanding_extent - drop an outstanding extent
4773  * @inode: the inode we're dropping the extent for
4774  *
4775  * This is called when we are freeing up an outstanding extent, either called
4776  * after an error or after an extent is written.  This will return the number of
4777  * reserved extents that need to be freed.  This must be called with
4778  * BTRFS_I(inode)->lock held.
4779  */
4780 static unsigned drop_outstanding_extent(struct inode *inode)
4781 {
4782         unsigned drop_inode_space = 0;
4783         unsigned dropped_extents = 0;
4784
4785         BUG_ON(!BTRFS_I(inode)->outstanding_extents);
4786         BTRFS_I(inode)->outstanding_extents--;
4787
4788         if (BTRFS_I(inode)->outstanding_extents == 0 &&
4789             test_and_clear_bit(BTRFS_INODE_DELALLOC_META_RESERVED,
4790                                &BTRFS_I(inode)->runtime_flags))
4791                 drop_inode_space = 1;
4792
4793         /*
4794          * If we have more or the same amount of outsanding extents than we have
4795          * reserved then we need to leave the reserved extents count alone.
4796          */
4797         if (BTRFS_I(inode)->outstanding_extents >=
4798             BTRFS_I(inode)->reserved_extents)
4799                 return drop_inode_space;
4800
4801         dropped_extents = BTRFS_I(inode)->reserved_extents -
4802                 BTRFS_I(inode)->outstanding_extents;
4803         BTRFS_I(inode)->reserved_extents -= dropped_extents;
4804         return dropped_extents + drop_inode_space;
4805 }
4806
4807 /**
4808  * calc_csum_metadata_size - return the amount of metada space that must be
4809  *      reserved/free'd for the given bytes.
4810  * @inode: the inode we're manipulating
4811  * @num_bytes: the number of bytes in question
4812  * @reserve: 1 if we are reserving space, 0 if we are freeing space
4813  *
4814  * This adjusts the number of csum_bytes in the inode and then returns the
4815  * correct amount of metadata that must either be reserved or freed.  We
4816  * calculate how many checksums we can fit into one leaf and then divide the
4817  * number of bytes that will need to be checksumed by this value to figure out
4818  * how many checksums will be required.  If we are adding bytes then the number
4819  * may go up and we will return the number of additional bytes that must be
4820  * reserved.  If it is going down we will return the number of bytes that must
4821  * be freed.
4822  *
4823  * This must be called with BTRFS_I(inode)->lock held.
4824  */
4825 static u64 calc_csum_metadata_size(struct inode *inode, u64 num_bytes,
4826                                    int reserve)
4827 {
4828         struct btrfs_root *root = BTRFS_I(inode)->root;
4829         u64 csum_size;
4830         int num_csums_per_leaf;
4831         int num_csums;
4832         int old_csums;
4833
4834         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM &&
4835             BTRFS_I(inode)->csum_bytes == 0)
4836                 return 0;
4837
4838         old_csums = (int)div64_u64(BTRFS_I(inode)->csum_bytes, root->sectorsize);
4839         if (reserve)
4840                 BTRFS_I(inode)->csum_bytes += num_bytes;
4841         else
4842                 BTRFS_I(inode)->csum_bytes -= num_bytes;
4843         csum_size = BTRFS_LEAF_DATA_SIZE(root) - sizeof(struct btrfs_item);
4844         num_csums_per_leaf = (int)div64_u64(csum_size,
4845                                             sizeof(struct btrfs_csum_item) +
4846                                             sizeof(struct btrfs_disk_key));
4847         num_csums = (int)div64_u64(BTRFS_I(inode)->csum_bytes, root->sectorsize);
4848         num_csums = num_csums + num_csums_per_leaf - 1;
4849         num_csums = num_csums / num_csums_per_leaf;
4850
4851         old_csums = old_csums + num_csums_per_leaf - 1;
4852         old_csums = old_csums / num_csums_per_leaf;
4853
4854         /* No change, no need to reserve more */
4855         if (old_csums == num_csums)
4856                 return 0;
4857
4858         if (reserve)
4859                 return btrfs_calc_trans_metadata_size(root,
4860                                                       num_csums - old_csums);
4861
4862         return btrfs_calc_trans_metadata_size(root, old_csums - num_csums);
4863 }
4864
4865 int btrfs_delalloc_reserve_metadata(struct inode *inode, u64 num_bytes)
4866 {
4867         struct btrfs_root *root = BTRFS_I(inode)->root;
4868         struct btrfs_block_rsv *block_rsv = &root->fs_info->delalloc_block_rsv;
4869         u64 to_reserve = 0;
4870         u64 csum_bytes;
4871         unsigned nr_extents = 0;
4872         int extra_reserve = 0;
4873         enum btrfs_reserve_flush_enum flush = BTRFS_RESERVE_FLUSH_ALL;
4874         int ret = 0;
4875         bool delalloc_lock = true;
4876         u64 to_free = 0;
4877         unsigned dropped;
4878
4879         /* If we are a free space inode we need to not flush since we will be in
4880          * the middle of a transaction commit.  We also don't need the delalloc
4881          * mutex since we won't race with anybody.  We need this mostly to make
4882          * lockdep shut its filthy mouth.
4883          */
4884         if (btrfs_is_free_space_inode(inode)) {
4885                 flush = BTRFS_RESERVE_NO_FLUSH;
4886                 delalloc_lock = false;
4887         }
4888
4889         if (flush != BTRFS_RESERVE_NO_FLUSH &&
4890             btrfs_transaction_in_commit(root->fs_info))
4891                 schedule_timeout(1);
4892
4893         if (delalloc_lock)
4894                 mutex_lock(&BTRFS_I(inode)->delalloc_mutex);
4895
4896         num_bytes = ALIGN(num_bytes, root->sectorsize);
4897
4898         spin_lock(&BTRFS_I(inode)->lock);
4899         BTRFS_I(inode)->outstanding_extents++;
4900
4901         if (BTRFS_I(inode)->outstanding_extents >
4902             BTRFS_I(inode)->reserved_extents)
4903                 nr_extents = BTRFS_I(inode)->outstanding_extents -
4904                         BTRFS_I(inode)->reserved_extents;
4905
4906         /*
4907          * Add an item to reserve for updating the inode when we complete the
4908          * delalloc io.
4909          */
4910         if (!test_bit(BTRFS_INODE_DELALLOC_META_RESERVED,
4911                       &BTRFS_I(inode)->runtime_flags)) {
4912                 nr_extents++;
4913                 extra_reserve = 1;
4914         }
4915
4916         to_reserve = btrfs_calc_trans_metadata_size(root, nr_extents);
4917         to_reserve += calc_csum_metadata_size(inode, num_bytes, 1);
4918         csum_bytes = BTRFS_I(inode)->csum_bytes;
4919         spin_unlock(&BTRFS_I(inode)->lock);
4920
4921         if (root->fs_info->quota_enabled) {
4922                 ret = btrfs_qgroup_reserve(root, num_bytes +
4923                                            nr_extents * root->leafsize);
4924                 if (ret)
4925                         goto out_fail;
4926         }
4927
4928         ret = reserve_metadata_bytes(root, block_rsv, to_reserve, flush);
4929         if (unlikely(ret)) {
4930                 if (root->fs_info->quota_enabled)
4931                         btrfs_qgroup_free(root, num_bytes +
4932                                                 nr_extents * root->leafsize);
4933                 goto out_fail;
4934         }
4935
4936         spin_lock(&BTRFS_I(inode)->lock);
4937         if (extra_reserve) {
4938                 set_bit(BTRFS_INODE_DELALLOC_META_RESERVED,
4939                         &BTRFS_I(inode)->runtime_flags);
4940                 nr_extents--;
4941         }
4942         BTRFS_I(inode)->reserved_extents += nr_extents;
4943         spin_unlock(&BTRFS_I(inode)->lock);
4944
4945         if (delalloc_lock)
4946                 mutex_unlock(&BTRFS_I(inode)->delalloc_mutex);
4947
4948         if (to_reserve)
4949                 trace_btrfs_space_reservation(root->fs_info,"delalloc",
4950                                               btrfs_ino(inode), to_reserve, 1);
4951         block_rsv_add_bytes(block_rsv, to_reserve, 1);
4952
4953         return 0;
4954
4955 out_fail:
4956         spin_lock(&BTRFS_I(inode)->lock);
4957         dropped = drop_outstanding_extent(inode);
4958         /*
4959          * If the inodes csum_bytes is the same as the original
4960          * csum_bytes then we know we haven't raced with any free()ers
4961          * so we can just reduce our inodes csum bytes and carry on.
4962          */
4963         if (BTRFS_I(inode)->csum_bytes == csum_bytes) {
4964                 calc_csum_metadata_size(inode, num_bytes, 0);
4965         } else {
4966                 u64 orig_csum_bytes = BTRFS_I(inode)->csum_bytes;
4967                 u64 bytes;
4968
4969                 /*
4970                  * This is tricky, but first we need to figure out how much we
4971                  * free'd from any free-ers that occured during this
4972                  * reservation, so we reset ->csum_bytes to the csum_bytes
4973                  * before we dropped our lock, and then call the free for the
4974                  * number of bytes that were freed while we were trying our
4975                  * reservation.
4976                  */
4977                 bytes = csum_bytes - BTRFS_I(inode)->csum_bytes;
4978                 BTRFS_I(inode)->csum_bytes = csum_bytes;
4979                 to_free = calc_csum_metadata_size(inode, bytes, 0);
4980
4981
4982                 /*
4983                  * Now we need to see how much we would have freed had we not
4984                  * been making this reservation and our ->csum_bytes were not
4985                  * artificially inflated.
4986                  */
4987                 BTRFS_I(inode)->csum_bytes = csum_bytes - num_bytes;
4988                 bytes = csum_bytes - orig_csum_bytes;
4989                 bytes = calc_csum_metadata_size(inode, bytes, 0);
4990
4991                 /*
4992                  * Now reset ->csum_bytes to what it should be.  If bytes is
4993                  * more than to_free then we would have free'd more space had we
4994                  * not had an artificially high ->csum_bytes, so we need to free
4995                  * the remainder.  If bytes is the same or less then we don't
4996                  * need to do anything, the other free-ers did the correct
4997                  * thing.
4998                  */
4999                 BTRFS_I(inode)->csum_bytes = orig_csum_bytes - num_bytes;
5000                 if (bytes > to_free)
5001                         to_free = bytes - to_free;
5002                 else
5003                         to_free = 0;
5004         }
5005         spin_unlock(&BTRFS_I(inode)->lock);
5006         if (dropped)
5007                 to_free += btrfs_calc_trans_metadata_size(root, dropped);
5008
5009         if (to_free) {
5010                 btrfs_block_rsv_release(root, block_rsv, to_free);
5011                 trace_btrfs_space_reservation(root->fs_info, "delalloc",
5012                                               btrfs_ino(inode), to_free, 0);
5013         }
5014         if (delalloc_lock)
5015                 mutex_unlock(&BTRFS_I(inode)->delalloc_mutex);
5016         return ret;
5017 }
5018
5019 /**
5020  * btrfs_delalloc_release_metadata - release a metadata reservation for an inode
5021  * @inode: the inode to release the reservation for
5022  * @num_bytes: the number of bytes we're releasing
5023  *
5024  * This will release the metadata reservation for an inode.  This can be called
5025  * once we complete IO for a given set of bytes to release their metadata
5026  * reservations.
5027  */
5028 void btrfs_delalloc_release_metadata(struct inode *inode, u64 num_bytes)
5029 {
5030         struct btrfs_root *root = BTRFS_I(inode)->root;
5031         u64 to_free = 0;
5032         unsigned dropped;
5033
5034         num_bytes = ALIGN(num_bytes, root->sectorsize);
5035         spin_lock(&BTRFS_I(inode)->lock);
5036         dropped = drop_outstanding_extent(inode);
5037
5038         if (num_bytes)
5039                 to_free = calc_csum_metadata_size(inode, num_bytes, 0);
5040         spin_unlock(&BTRFS_I(inode)->lock);
5041         if (dropped > 0)
5042                 to_free += btrfs_calc_trans_metadata_size(root, dropped);
5043
5044         trace_btrfs_space_reservation(root->fs_info, "delalloc",
5045                                       btrfs_ino(inode), to_free, 0);
5046         if (root->fs_info->quota_enabled) {
5047                 btrfs_qgroup_free(root, num_bytes +
5048                                         dropped * root->leafsize);
5049         }
5050
5051         btrfs_block_rsv_release(root, &root->fs_info->delalloc_block_rsv,
5052                                 to_free);
5053 }
5054
5055 /**
5056  * btrfs_delalloc_reserve_space - reserve data and metadata space for delalloc
5057  * @inode: inode we're writing to
5058  * @num_bytes: the number of bytes we want to allocate
5059  *
5060  * This will do the following things
5061  *
5062  * o reserve space in the data space info for num_bytes
5063  * o reserve space in the metadata space info based on number of outstanding
5064  *   extents and how much csums will be needed
5065  * o add to the inodes ->delalloc_bytes
5066  * o add it to the fs_info's delalloc inodes list.
5067  *
5068  * This will return 0 for success and -ENOSPC if there is no space left.
5069  */
5070 int btrfs_delalloc_reserve_space(struct inode *inode, u64 num_bytes)
5071 {
5072         int ret;
5073
5074         ret = btrfs_check_data_free_space(inode, num_bytes);
5075         if (ret)
5076                 return ret;
5077
5078         ret = btrfs_delalloc_reserve_metadata(inode, num_bytes);
5079         if (ret) {
5080                 btrfs_free_reserved_data_space(inode, num_bytes);
5081                 return ret;
5082         }
5083
5084         return 0;
5085 }
5086
5087 /**
5088  * btrfs_delalloc_release_space - release data and metadata space for delalloc
5089  * @inode: inode we're releasing space for
5090  * @num_bytes: the number of bytes we want to free up
5091  *
5092  * This must be matched with a call to btrfs_delalloc_reserve_space.  This is
5093  * called in the case that we don't need the metadata AND data reservations
5094  * anymore.  So if there is an error or we insert an inline extent.
5095  *
5096  * This function will release the metadata space that was not used and will
5097  * decrement ->delalloc_bytes and remove it from the fs_info delalloc_inodes
5098  * list if there are no delalloc bytes left.
5099  */
5100 void btrfs_delalloc_release_space(struct inode *inode, u64 num_bytes)
5101 {
5102         btrfs_delalloc_release_metadata(inode, num_bytes);
5103         btrfs_free_reserved_data_space(inode, num_bytes);
5104 }
5105
5106 static int update_block_group(struct btrfs_root *root,
5107                               u64 bytenr, u64 num_bytes, int alloc)
5108 {
5109         struct btrfs_block_group_cache *cache = NULL;
5110         struct btrfs_fs_info *info = root->fs_info;
5111         u64 total = num_bytes;
5112         u64 old_val;
5113         u64 byte_in_group;
5114         int factor;
5115
5116         /* block accounting for super block */
5117         spin_lock(&info->delalloc_root_lock);
5118         old_val = btrfs_super_bytes_used(info->super_copy);
5119         if (alloc)
5120                 old_val += num_bytes;
5121         else
5122                 old_val -= num_bytes;
5123         btrfs_set_super_bytes_used(info->super_copy, old_val);
5124         spin_unlock(&info->delalloc_root_lock);
5125
5126         while (total) {
5127                 cache = btrfs_lookup_block_group(info, bytenr);
5128                 if (!cache)
5129                         return -ENOENT;
5130                 if (cache->flags & (BTRFS_BLOCK_GROUP_DUP |
5131                                     BTRFS_BLOCK_GROUP_RAID1 |
5132                                     BTRFS_BLOCK_GROUP_RAID10))
5133                         factor = 2;
5134                 else
5135                         factor = 1;
5136                 /*
5137                  * If this block group has free space cache written out, we
5138                  * need to make sure to load it if we are removing space.  This
5139                  * is because we need the unpinning stage to actually add the
5140                  * space back to the block group, otherwise we will leak space.
5141                  */
5142                 if (!alloc && cache->cached == BTRFS_CACHE_NO)
5143                         cache_block_group(cache, 1);
5144
5145                 byte_in_group = bytenr - cache->key.objectid;
5146                 WARN_ON(byte_in_group > cache->key.offset);
5147
5148                 spin_lock(&cache->space_info->lock);
5149                 spin_lock(&cache->lock);
5150
5151                 if (btrfs_test_opt(root, SPACE_CACHE) &&
5152                     cache->disk_cache_state < BTRFS_DC_CLEAR)
5153                         cache->disk_cache_state = BTRFS_DC_CLEAR;
5154
5155                 cache->dirty = 1;
5156                 old_val = btrfs_block_group_used(&cache->item);
5157                 num_bytes = min(total, cache->key.offset - byte_in_group);
5158                 if (alloc) {
5159                         old_val += num_bytes;
5160                         btrfs_set_block_group_used(&cache->item, old_val);
5161                         cache->reserved -= num_bytes;
5162                         cache->space_info->bytes_reserved -= num_bytes;
5163                         cache->space_info->bytes_used += num_bytes;
5164                         cache->space_info->disk_used += num_bytes * factor;
5165                         spin_unlock(&cache->lock);
5166                         spin_unlock(&cache->space_info->lock);
5167                 } else {
5168                         old_val -= num_bytes;
5169                         btrfs_set_block_group_used(&cache->item, old_val);
5170                         cache->pinned += num_bytes;
5171                         cache->space_info->bytes_pinned += num_bytes;
5172                         cache->space_info->bytes_used -= num_bytes;
5173                         cache->space_info->disk_used -= num_bytes * factor;
5174                         spin_unlock(&cache->lock);
5175                         spin_unlock(&cache->space_info->lock);
5176
5177                         set_extent_dirty(info->pinned_extents,
5178                                          bytenr, bytenr + num_bytes - 1,
5179                                          GFP_NOFS | __GFP_NOFAIL);
5180                 }
5181                 btrfs_put_block_group(cache);
5182                 total -= num_bytes;
5183                 bytenr += num_bytes;
5184         }
5185         return 0;
5186 }
5187
5188 static u64 first_logical_byte(struct btrfs_root *root, u64 search_start)
5189 {
5190         struct btrfs_block_group_cache *cache;
5191         u64 bytenr;
5192
5193         spin_lock(&root->fs_info->block_group_cache_lock);
5194         bytenr = root->fs_info->first_logical_byte;
5195         spin_unlock(&root->fs_info->block_group_cache_lock);
5196
5197         if (bytenr < (u64)-1)
5198                 return bytenr;
5199
5200         cache = btrfs_lookup_first_block_group(root->fs_info, search_start);
5201         if (!cache)
5202                 return 0;
5203
5204         bytenr = cache->key.objectid;
5205         btrfs_put_block_group(cache);
5206
5207         return bytenr;
5208 }
5209
5210 static int pin_down_extent(struct btrfs_root *root,
5211                            struct btrfs_block_group_cache *cache,
5212                            u64 bytenr, u64 num_bytes, int reserved)
5213 {
5214         spin_lock(&cache->space_info->lock);
5215         spin_lock(&cache->lock);
5216         cache->pinned += num_bytes;
5217         cache->space_info->bytes_pinned += num_bytes;
5218         if (reserved) {
5219                 cache->reserved -= num_bytes;
5220                 cache->space_info->bytes_reserved -= num_bytes;
5221         }
5222         spin_unlock(&cache->lock);
5223         spin_unlock(&cache->space_info->lock);
5224
5225         set_extent_dirty(root->fs_info->pinned_extents, bytenr,
5226                          bytenr + num_bytes - 1, GFP_NOFS | __GFP_NOFAIL);
5227         return 0;
5228 }
5229
5230 /*
5231  * this function must be called within transaction
5232  */
5233 int btrfs_pin_extent(struct btrfs_root *root,
5234                      u64 bytenr, u64 num_bytes, int reserved)
5235 {
5236         struct btrfs_block_group_cache *cache;
5237
5238         cache = btrfs_lookup_block_group(root->fs_info, bytenr);
5239         BUG_ON(!cache); /* Logic error */
5240
5241         pin_down_extent(root, cache, bytenr, num_bytes, reserved);
5242
5243         btrfs_put_block_group(cache);
5244         return 0;
5245 }
5246
5247 /*
5248  * this function must be called within transaction
5249  */
5250 int btrfs_pin_extent_for_log_replay(struct btrfs_root *root,
5251                                     u64 bytenr, u64 num_bytes)
5252 {
5253         struct btrfs_block_group_cache *cache;
5254         int ret;
5255
5256         cache = btrfs_lookup_block_group(root->fs_info, bytenr);
5257         if (!cache)
5258                 return -EINVAL;
5259
5260         /*
5261          * pull in the free space cache (if any) so that our pin
5262          * removes the free space from the cache.  We have load_only set
5263          * to one because the slow code to read in the free extents does check
5264          * the pinned extents.
5265          */
5266         cache_block_group(cache, 1);
5267
5268         pin_down_extent(root, cache, bytenr, num_bytes, 0);
5269
5270         /* remove us from the free space cache (if we're there at all) */
5271         ret = btrfs_remove_free_space(cache, bytenr, num_bytes);
5272         btrfs_put_block_group(cache);
5273         return ret;
5274 }
5275
5276 static int __exclude_logged_extent(struct btrfs_root *root, u64 start, u64 num_bytes)
5277 {
5278         int ret;
5279         struct btrfs_block_group_cache *block_group;
5280         struct btrfs_caching_control *caching_ctl;
5281
5282         block_group = btrfs_lookup_block_group(root->fs_info, start);
5283         if (!block_group)
5284                 return -EINVAL;
5285
5286         cache_block_group(block_group, 0);
5287         caching_ctl = get_caching_control(block_group);
5288
5289         if (!caching_ctl) {
5290                 /* Logic error */
5291                 BUG_ON(!block_group_cache_done(block_group));
5292                 ret = btrfs_remove_free_space(block_group, start, num_bytes);
5293         } else {
5294                 mutex_lock(&caching_ctl->mutex);
5295
5296                 if (start >= caching_ctl->progress) {
5297                         ret = add_excluded_extent(root, start, num_bytes);
5298                 } else if (start + num_bytes <= caching_ctl->progress) {
5299                         ret = btrfs_remove_free_space(block_group,
5300                                                       start, num_bytes);
5301                 } else {
5302                         num_bytes = caching_ctl->progress - start;
5303                         ret = btrfs_remove_free_space(block_group,
5304                                                       start, num_bytes);
5305                         if (ret)
5306                                 goto out_lock;
5307
5308                         num_bytes = (start + num_bytes) -
5309                                 caching_ctl->progress;
5310                         start = caching_ctl->progress;
5311                         ret = add_excluded_extent(root, start, num_bytes);
5312                 }
5313 out_lock:
5314                 mutex_unlock(&caching_ctl->mutex);
5315                 put_caching_control(caching_ctl);
5316         }
5317         btrfs_put_block_group(block_group);
5318         return ret;
5319 }
5320
5321 int btrfs_exclude_logged_extents(struct btrfs_root *log,
5322                                  struct extent_buffer *eb)
5323 {
5324         struct btrfs_file_extent_item *item;
5325         struct btrfs_key key;
5326         int found_type;
5327         int i;
5328
5329         if (!btrfs_fs_incompat(log->fs_info, MIXED_GROUPS))
5330                 return 0;
5331
5332         for (i = 0; i < btrfs_header_nritems(eb); i++) {
5333                 btrfs_item_key_to_cpu(eb, &key, i);
5334                 if (key.type != BTRFS_EXTENT_DATA_KEY)
5335                         continue;
5336                 item = btrfs_item_ptr(eb, i, struct btrfs_file_extent_item);
5337                 found_type = btrfs_file_extent_type(eb, item);
5338                 if (found_type == BTRFS_FILE_EXTENT_INLINE)
5339                         continue;
5340                 if (btrfs_file_extent_disk_bytenr(eb, item) == 0)
5341                         continue;
5342                 key.objectid = btrfs_file_extent_disk_bytenr(eb, item);
5343                 key.offset = btrfs_file_extent_disk_num_bytes(eb, item);
5344                 __exclude_logged_extent(log, key.objectid, key.offset);
5345         }
5346
5347         return 0;
5348 }
5349
5350 /**
5351  * btrfs_update_reserved_bytes - update the block_group and space info counters
5352  * @cache:      The cache we are manipulating
5353  * @num_bytes:  The number of bytes in question
5354  * @reserve:    One of the reservation enums
5355  *
5356  * This is called by the allocator when it reserves space, or by somebody who is
5357  * freeing space that was never actually used on disk.  For example if you
5358  * reserve some space for a new leaf in transaction A and before transaction A
5359  * commits you free that leaf, you call this with reserve set to 0 in order to
5360  * clear the reservation.
5361  *
5362  * Metadata reservations should be called with RESERVE_ALLOC so we do the proper
5363  * ENOSPC accounting.  For data we handle the reservation through clearing the
5364  * delalloc bits in the io_tree.  We have to do this since we could end up
5365  * allocating less disk space for the amount of data we have reserved in the
5366  * case of compression.
5367  *
5368  * If this is a reservation and the block group has become read only we cannot
5369  * make the reservation and return -EAGAIN, otherwise this function always
5370  * succeeds.
5371  */
5372 static int btrfs_update_reserved_bytes(struct btrfs_block_group_cache *cache,
5373                                        u64 num_bytes, int reserve)
5374 {
5375         struct btrfs_space_info *space_info = cache->space_info;
5376         int ret = 0;
5377
5378         spin_lock(&space_info->lock);
5379         spin_lock(&cache->lock);
5380         if (reserve != RESERVE_FREE) {
5381                 if (cache->ro) {
5382                         ret = -EAGAIN;
5383                 } else {
5384                         cache->reserved += num_bytes;
5385                         space_info->bytes_reserved += num_bytes;
5386                         if (reserve == RESERVE_ALLOC) {
5387                                 trace_btrfs_space_reservation(cache->fs_info,
5388                                                 "space_info", space_info->flags,
5389                                                 num_bytes, 0);
5390                                 space_info->bytes_may_use -= num_bytes;
5391                         }
5392                 }
5393         } else {
5394                 if (cache->ro)
5395                         space_info->bytes_readonly += num_bytes;
5396                 cache->reserved -= num_bytes;
5397                 space_info->bytes_reserved -= num_bytes;
5398                 space_info->reservation_progress++;
5399         }
5400         spin_unlock(&cache->lock);
5401         spin_unlock(&space_info->lock);
5402         return ret;
5403 }
5404
5405 void btrfs_prepare_extent_commit(struct btrfs_trans_handle *trans,
5406                                 struct btrfs_root *root)
5407 {
5408         struct btrfs_fs_info *fs_info = root->fs_info;
5409         struct btrfs_caching_control *next;
5410         struct btrfs_caching_control *caching_ctl;
5411         struct btrfs_block_group_cache *cache;
5412         struct btrfs_space_info *space_info;
5413
5414         down_write(&fs_info->extent_commit_sem);
5415
5416         list_for_each_entry_safe(caching_ctl, next,
5417                                  &fs_info->caching_block_groups, list) {
5418                 cache = caching_ctl->block_group;
5419                 if (block_group_cache_done(cache)) {
5420                         cache->last_byte_to_unpin = (u64)-1;
5421                         list_del_init(&caching_ctl->list);
5422                         put_caching_control(caching_ctl);
5423                 } else {
5424                         cache->last_byte_to_unpin = caching_ctl->progress;
5425                 }
5426         }
5427
5428         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
5429                 fs_info->pinned_extents = &fs_info->freed_extents[1];
5430         else
5431                 fs_info->pinned_extents = &fs_info->freed_extents[0];
5432
5433         up_write(&fs_info->extent_commit_sem);
5434
5435         list_for_each_entry_rcu(space_info, &fs_info->space_info, list)
5436                 percpu_counter_set(&space_info->total_bytes_pinned, 0);
5437
5438         update_global_block_rsv(fs_info);
5439 }
5440
5441 static int unpin_extent_range(struct btrfs_root *root, u64 start, u64 end)
5442 {
5443         struct btrfs_fs_info *fs_info = root->fs_info;
5444         struct btrfs_block_group_cache *cache = NULL;
5445         struct btrfs_space_info *space_info;
5446         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
5447         u64 len;
5448         bool readonly;
5449
5450         while (start <= end) {
5451                 readonly = false;
5452                 if (!cache ||
5453                     start >= cache->key.objectid + cache->key.offset) {
5454                         if (cache)
5455                                 btrfs_put_block_group(cache);
5456                         cache = btrfs_lookup_block_group(fs_info, start);
5457                         BUG_ON(!cache); /* Logic error */
5458                 }
5459
5460                 len = cache->key.objectid + cache->key.offset - start;
5461                 len = min(len, end + 1 - start);
5462
5463                 if (start < cache->last_byte_to_unpin) {
5464                         len = min(len, cache->last_byte_to_unpin - start);
5465                         btrfs_add_free_space(cache, start, len);
5466                 }
5467
5468                 start += len;
5469                 space_info = cache->space_info;
5470
5471                 spin_lock(&space_info->lock);
5472                 spin_lock(&cache->lock);
5473                 cache->pinned -= len;
5474                 space_info->bytes_pinned -= len;
5475                 if (cache->ro) {
5476                         space_info->bytes_readonly += len;
5477                         readonly = true;
5478                 }
5479                 spin_unlock(&cache->lock);
5480                 if (!readonly && global_rsv->space_info == space_info) {
5481                         spin_lock(&global_rsv->lock);
5482                         if (!global_rsv->full) {
5483                                 len = min(len, global_rsv->size -
5484                                           global_rsv->reserved);
5485                                 global_rsv->reserved += len;
5486                                 space_info->bytes_may_use += len;
5487                                 if (global_rsv->reserved >= global_rsv->size)
5488                                         global_rsv->full = 1;
5489                         }
5490                         spin_unlock(&global_rsv->lock);
5491                 }
5492                 spin_unlock(&space_info->lock);
5493         }
5494
5495         if (cache)
5496                 btrfs_put_block_group(cache);
5497         return 0;
5498 }
5499
5500 int btrfs_finish_extent_commit(struct btrfs_trans_handle *trans,
5501                                struct btrfs_root *root)
5502 {
5503         struct btrfs_fs_info *fs_info = root->fs_info;
5504         struct extent_io_tree *unpin;
5505         u64 start;
5506         u64 end;
5507         int ret;
5508
5509         if (trans->aborted)
5510                 return 0;
5511
5512         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
5513                 unpin = &fs_info->freed_extents[1];
5514         else
5515                 unpin = &fs_info->freed_extents[0];
5516
5517         while (1) {
5518                 ret = find_first_extent_bit(unpin, 0, &start, &end,
5519                                             EXTENT_DIRTY, NULL);
5520                 if (ret)
5521                         break;
5522
5523                 if (btrfs_test_opt(root, DISCARD))
5524                         ret = btrfs_discard_extent(root, start,
5525                                                    end + 1 - start, NULL);
5526
5527                 clear_extent_dirty(unpin, start, end, GFP_NOFS);
5528                 unpin_extent_range(root, start, end);
5529                 cond_resched();
5530         }
5531
5532         return 0;
5533 }
5534
5535 static void add_pinned_bytes(struct btrfs_fs_info *fs_info, u64 num_bytes,
5536                              u64 owner, u64 root_objectid)
5537 {
5538         struct btrfs_space_info *space_info;
5539         u64 flags;
5540
5541         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
5542                 if (root_objectid == BTRFS_CHUNK_TREE_OBJECTID)
5543                         flags = BTRFS_BLOCK_GROUP_SYSTEM;
5544                 else
5545                         flags = BTRFS_BLOCK_GROUP_METADATA;
5546         } else {
5547                 flags = BTRFS_BLOCK_GROUP_DATA;
5548         }
5549
5550         space_info = __find_space_info(fs_info, flags);
5551         BUG_ON(!space_info); /* Logic bug */
5552         percpu_counter_add(&space_info->total_bytes_pinned, num_bytes);
5553 }
5554
5555
5556 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
5557                                 struct btrfs_root *root,
5558                                 u64 bytenr, u64 num_bytes, u64 parent,
5559                                 u64 root_objectid, u64 owner_objectid,
5560                                 u64 owner_offset, int refs_to_drop,
5561                                 struct btrfs_delayed_extent_op *extent_op)
5562 {
5563         struct btrfs_key key;
5564         struct btrfs_path *path;
5565         struct btrfs_fs_info *info = root->fs_info;
5566         struct btrfs_root *extent_root = info->extent_root;
5567         struct extent_buffer *leaf;
5568         struct btrfs_extent_item *ei;
5569         struct btrfs_extent_inline_ref *iref;
5570         int ret;
5571         int is_data;
5572         int extent_slot = 0;
5573         int found_extent = 0;
5574         int num_to_del = 1;
5575         u32 item_size;
5576         u64 refs;
5577         bool skinny_metadata = btrfs_fs_incompat(root->fs_info,
5578                                                  SKINNY_METADATA);
5579
5580         path = btrfs_alloc_path();
5581         if (!path)
5582                 return -ENOMEM;
5583
5584         path->reada = 1;
5585         path->leave_spinning = 1;
5586
5587         is_data = owner_objectid >= BTRFS_FIRST_FREE_OBJECTID;
5588         BUG_ON(!is_data && refs_to_drop != 1);
5589
5590         if (is_data)
5591                 skinny_metadata = 0;
5592
5593         ret = lookup_extent_backref(trans, extent_root, path, &iref,
5594                                     bytenr, num_bytes, parent,
5595                                     root_objectid, owner_objectid,
5596                                     owner_offset);
5597         if (ret == 0) {
5598                 extent_slot = path->slots[0];
5599                 while (extent_slot >= 0) {
5600                         btrfs_item_key_to_cpu(path->nodes[0], &key,
5601                                               extent_slot);
5602                         if (key.objectid != bytenr)
5603                                 break;
5604                         if (key.type == BTRFS_EXTENT_ITEM_KEY &&
5605                             key.offset == num_bytes) {
5606                                 found_extent = 1;
5607                                 break;
5608                         }
5609                         if (key.type == BTRFS_METADATA_ITEM_KEY &&
5610                             key.offset == owner_objectid) {
5611                                 found_extent = 1;
5612                                 break;
5613                         }
5614                         if (path->slots[0] - extent_slot > 5)
5615                                 break;
5616                         extent_slot--;
5617                 }
5618 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
5619                 item_size = btrfs_item_size_nr(path->nodes[0], extent_slot);
5620                 if (found_extent && item_size < sizeof(*ei))
5621                         found_extent = 0;
5622 #endif
5623                 if (!found_extent) {
5624                         BUG_ON(iref);
5625                         ret = remove_extent_backref(trans, extent_root, path,
5626                                                     NULL, refs_to_drop,
5627                                                     is_data);
5628                         if (ret) {
5629                                 btrfs_abort_transaction(trans, extent_root, ret);
5630                                 goto out;
5631                         }
5632                         btrfs_release_path(path);
5633                         path->leave_spinning = 1;
5634
5635                         key.objectid = bytenr;
5636                         key.type = BTRFS_EXTENT_ITEM_KEY;
5637                         key.offset = num_bytes;
5638
5639                         if (!is_data && skinny_metadata) {
5640                                 key.type = BTRFS_METADATA_ITEM_KEY;
5641                                 key.offset = owner_objectid;
5642                         }
5643
5644                         ret = btrfs_search_slot(trans, extent_root,
5645                                                 &key, path, -1, 1);
5646                         if (ret > 0 && skinny_metadata && path->slots[0]) {
5647                                 /*
5648                                  * Couldn't find our skinny metadata item,
5649                                  * see if we have ye olde extent item.
5650                                  */
5651                                 path->slots[0]--;
5652                                 btrfs_item_key_to_cpu(path->nodes[0], &key,
5653                                                       path->slots[0]);
5654                                 if (key.objectid == bytenr &&
5655                                     key.type == BTRFS_EXTENT_ITEM_KEY &&
5656                                     key.offset == num_bytes)
5657                                         ret = 0;
5658                         }
5659
5660                         if (ret > 0 && skinny_metadata) {
5661                                 skinny_metadata = false;
5662                                 key.type = BTRFS_EXTENT_ITEM_KEY;
5663                                 key.offset = num_bytes;
5664                                 btrfs_release_path(path);
5665                                 ret = btrfs_search_slot(trans, extent_root,
5666                                                         &key, path, -1, 1);
5667                         }
5668
5669                         if (ret) {
5670                                 btrfs_err(info, "umm, got %d back from search, was looking for %llu",
5671                                         ret, (unsigned long long)bytenr);
5672                                 if (ret > 0)
5673                                         btrfs_print_leaf(extent_root,
5674                                                          path->nodes[0]);
5675                         }
5676                         if (ret < 0) {
5677                                 btrfs_abort_transaction(trans, extent_root, ret);
5678                                 goto out;
5679                         }
5680                         extent_slot = path->slots[0];
5681                 }
5682         } else if (ret == -ENOENT) {
5683                 btrfs_print_leaf(extent_root, path->nodes[0]);
5684                 WARN_ON(1);
5685                 btrfs_err(info,
5686                         "unable to find ref byte nr %llu parent %llu root %llu  owner %llu offset %llu",
5687                         (unsigned long long)bytenr,
5688                         (unsigned long long)parent,
5689                         (unsigned long long)root_objectid,
5690                         (unsigned long long)owner_objectid,
5691                         (unsigned long long)owner_offset);
5692         } else {
5693                 btrfs_abort_transaction(trans, extent_root, ret);
5694                 goto out;
5695         }
5696
5697         leaf = path->nodes[0];
5698         item_size = btrfs_item_size_nr(leaf, extent_slot);
5699 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
5700         if (item_size < sizeof(*ei)) {
5701                 BUG_ON(found_extent || extent_slot != path->slots[0]);
5702                 ret = convert_extent_item_v0(trans, extent_root, path,
5703                                              owner_objectid, 0);
5704                 if (ret < 0) {
5705                         btrfs_abort_transaction(trans, extent_root, ret);
5706                         goto out;
5707                 }
5708
5709                 btrfs_release_path(path);
5710                 path->leave_spinning = 1;
5711
5712                 key.objectid = bytenr;
5713                 key.type = BTRFS_EXTENT_ITEM_KEY;
5714                 key.offset = num_bytes;
5715
5716                 ret = btrfs_search_slot(trans, extent_root, &key, path,
5717                                         -1, 1);
5718                 if (ret) {
5719                         btrfs_err(info, "umm, got %d back from search, was looking for %llu",
5720                                 ret, (unsigned long long)bytenr);
5721                         btrfs_print_leaf(extent_root, path->nodes[0]);
5722                 }
5723                 if (ret < 0) {
5724                         btrfs_abort_transaction(trans, extent_root, ret);
5725                         goto out;
5726                 }
5727
5728                 extent_slot = path->slots[0];
5729                 leaf = path->nodes[0];
5730                 item_size = btrfs_item_size_nr(leaf, extent_slot);
5731         }
5732 #endif
5733         BUG_ON(item_size < sizeof(*ei));
5734         ei = btrfs_item_ptr(leaf, extent_slot,
5735                             struct btrfs_extent_item);
5736         if (owner_objectid < BTRFS_FIRST_FREE_OBJECTID &&
5737             key.type == BTRFS_EXTENT_ITEM_KEY) {
5738                 struct btrfs_tree_block_info *bi;
5739                 BUG_ON(item_size < sizeof(*ei) + sizeof(*bi));
5740                 bi = (struct btrfs_tree_block_info *)(ei + 1);
5741                 WARN_ON(owner_objectid != btrfs_tree_block_level(leaf, bi));
5742         }
5743
5744         refs = btrfs_extent_refs(leaf, ei);
5745         if (refs < refs_to_drop) {
5746                 btrfs_err(info, "trying to drop %d refs but we only have %Lu "
5747                           "for bytenr %Lu\n", refs_to_drop, refs, bytenr);
5748                 ret = -EINVAL;
5749                 btrfs_abort_transaction(trans, extent_root, ret);
5750                 goto out;
5751         }
5752         refs -= refs_to_drop;
5753
5754         if (refs > 0) {
5755                 if (extent_op)
5756                         __run_delayed_extent_op(extent_op, leaf, ei);
5757                 /*
5758                  * In the case of inline back ref, reference count will
5759                  * be updated by remove_extent_backref
5760                  */
5761                 if (iref) {
5762                         BUG_ON(!found_extent);
5763                 } else {
5764                         btrfs_set_extent_refs(leaf, ei, refs);
5765                         btrfs_mark_buffer_dirty(leaf);
5766                 }
5767                 if (found_extent) {
5768                         ret = remove_extent_backref(trans, extent_root, path,
5769                                                     iref, refs_to_drop,
5770                                                     is_data);
5771                         if (ret) {
5772                                 btrfs_abort_transaction(trans, extent_root, ret);
5773                                 goto out;
5774                         }
5775                 }
5776                 add_pinned_bytes(root->fs_info, -num_bytes, owner_objectid,
5777                                  root_objectid);
5778         } else {
5779                 if (found_extent) {
5780                         BUG_ON(is_data && refs_to_drop !=
5781                                extent_data_ref_count(root, path, iref));
5782                         if (iref) {
5783                                 BUG_ON(path->slots[0] != extent_slot);
5784                         } else {
5785                                 BUG_ON(path->slots[0] != extent_slot + 1);
5786                                 path->slots[0] = extent_slot;
5787                                 num_to_del = 2;
5788                         }
5789                 }
5790
5791                 ret = btrfs_del_items(trans, extent_root, path, path->slots[0],
5792                                       num_to_del);
5793                 if (ret) {
5794                         btrfs_abort_transaction(trans, extent_root, ret);
5795                         goto out;
5796                 }
5797                 btrfs_release_path(path);
5798
5799                 if (is_data) {
5800                         ret = btrfs_del_csums(trans, root, bytenr, num_bytes);
5801                         if (ret) {
5802                                 btrfs_abort_transaction(trans, extent_root, ret);
5803                                 goto out;
5804                         }
5805                 }
5806
5807                 ret = update_block_group(root, bytenr, num_bytes, 0);
5808                 if (ret) {
5809                         btrfs_abort_transaction(trans, extent_root, ret);
5810                         goto out;
5811                 }
5812         }
5813 out:
5814         btrfs_free_path(path);
5815         return ret;
5816 }
5817
5818 /*
5819  * when we free an block, it is possible (and likely) that we free the last
5820  * delayed ref for that extent as well.  This searches the delayed ref tree for
5821  * a given extent, and if there are no other delayed refs to be processed, it
5822  * removes it from the tree.
5823  */
5824 static noinline int check_ref_cleanup(struct btrfs_trans_handle *trans,
5825                                       struct btrfs_root *root, u64 bytenr)
5826 {
5827         struct btrfs_delayed_ref_head *head;
5828         struct btrfs_delayed_ref_root *delayed_refs;
5829         struct btrfs_delayed_ref_node *ref;
5830         struct rb_node *node;
5831         int ret = 0;
5832
5833         delayed_refs = &trans->transaction->delayed_refs;
5834         spin_lock(&delayed_refs->lock);
5835         head = btrfs_find_delayed_ref_head(trans, bytenr);
5836         if (!head)
5837                 goto out;
5838
5839         node = rb_prev(&head->node.rb_node);
5840         if (!node)
5841                 goto out;
5842
5843         ref = rb_entry(node, struct btrfs_delayed_ref_node, rb_node);
5844
5845         /* there are still entries for this ref, we can't drop it */
5846         if (ref->bytenr == bytenr)
5847                 goto out;
5848
5849         if (head->extent_op) {
5850                 if (!head->must_insert_reserved)
5851                         goto out;
5852                 btrfs_free_delayed_extent_op(head->extent_op);
5853                 head->extent_op = NULL;
5854         }
5855
5856         /*
5857          * waiting for the lock here would deadlock.  If someone else has it
5858          * locked they are already in the process of dropping it anyway
5859          */
5860         if (!mutex_trylock(&head->mutex))
5861                 goto out;
5862
5863         /*
5864          * at this point we have a head with no other entries.  Go
5865          * ahead and process it.
5866          */
5867         head->node.in_tree = 0;
5868         rb_erase(&head->node.rb_node, &delayed_refs->root);
5869
5870         delayed_refs->num_entries--;
5871
5872         /*
5873          * we don't take a ref on the node because we're removing it from the
5874          * tree, so we just steal the ref the tree was holding.
5875          */
5876         delayed_refs->num_heads--;
5877         if (list_empty(&head->cluster))
5878                 delayed_refs->num_heads_ready--;
5879
5880         list_del_init(&head->cluster);
5881         spin_unlock(&delayed_refs->lock);
5882
5883         BUG_ON(head->extent_op);
5884         if (head->must_insert_reserved)
5885                 ret = 1;
5886
5887         mutex_unlock(&head->mutex);
5888         btrfs_put_delayed_ref(&head->node);
5889         return ret;
5890 out:
5891         spin_unlock(&delayed_refs->lock);
5892         return 0;
5893 }
5894
5895 void btrfs_free_tree_block(struct btrfs_trans_handle *trans,
5896                            struct btrfs_root *root,
5897                            struct extent_buffer *buf,
5898                            u64 parent, int last_ref)
5899 {
5900         struct btrfs_block_group_cache *cache = NULL;
5901         int pin = 1;
5902         int ret;
5903
5904         if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
5905                 ret = btrfs_add_delayed_tree_ref(root->fs_info, trans,
5906                                         buf->start, buf->len,
5907                                         parent, root->root_key.objectid,
5908                                         btrfs_header_level(buf),
5909                                         BTRFS_DROP_DELAYED_REF, NULL, 0);
5910                 BUG_ON(ret); /* -ENOMEM */
5911         }
5912
5913         if (!last_ref)
5914                 return;
5915
5916         cache = btrfs_lookup_block_group(root->fs_info, buf->start);
5917
5918         if (btrfs_header_generation(buf) == trans->transid) {
5919                 if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
5920                         ret = check_ref_cleanup(trans, root, buf->start);
5921                         if (!ret)
5922                                 goto out;
5923                 }
5924
5925                 if (btrfs_header_flag(buf, BTRFS_HEADER_FLAG_WRITTEN)) {
5926                         pin_down_extent(root, cache, buf->start, buf->len, 1);
5927                         goto out;
5928                 }
5929
5930                 WARN_ON(test_bit(EXTENT_BUFFER_DIRTY, &buf->bflags));
5931
5932                 btrfs_add_free_space(cache, buf->start, buf->len);
5933                 btrfs_update_reserved_bytes(cache, buf->len, RESERVE_FREE);
5934                 pin = 0;
5935         }
5936 out:
5937         if (pin)
5938                 add_pinned_bytes(root->fs_info, buf->len,
5939                                  btrfs_header_level(buf),
5940                                  root->root_key.objectid);
5941
5942         /*
5943          * Deleting the buffer, clear the corrupt flag since it doesn't matter
5944          * anymore.
5945          */
5946         clear_bit(EXTENT_BUFFER_CORRUPT, &buf->bflags);
5947         btrfs_put_block_group(cache);
5948 }
5949
5950 /* Can return -ENOMEM */
5951 int btrfs_free_extent(struct btrfs_trans_handle *trans, struct btrfs_root *root,
5952                       u64 bytenr, u64 num_bytes, u64 parent, u64 root_objectid,
5953                       u64 owner, u64 offset, int for_cow)
5954 {
5955         int ret;
5956         struct btrfs_fs_info *fs_info = root->fs_info;
5957
5958         add_pinned_bytes(root->fs_info, num_bytes, owner, root_objectid);
5959
5960         /*
5961          * tree log blocks never actually go into the extent allocation
5962          * tree, just update pinning info and exit early.
5963          */
5964         if (root_objectid == BTRFS_TREE_LOG_OBJECTID) {
5965                 WARN_ON(owner >= BTRFS_FIRST_FREE_OBJECTID);
5966                 /* unlocks the pinned mutex */
5967                 btrfs_pin_extent(root, bytenr, num_bytes, 1);
5968                 ret = 0;
5969         } else if (owner < BTRFS_FIRST_FREE_OBJECTID) {
5970                 ret = btrfs_add_delayed_tree_ref(fs_info, trans, bytenr,
5971                                         num_bytes,
5972                                         parent, root_objectid, (int)owner,
5973                                         BTRFS_DROP_DELAYED_REF, NULL, for_cow);
5974         } else {
5975                 ret = btrfs_add_delayed_data_ref(fs_info, trans, bytenr,
5976                                                 num_bytes,
5977                                                 parent, root_objectid, owner,
5978                                                 offset, BTRFS_DROP_DELAYED_REF,
5979                                                 NULL, for_cow);
5980         }
5981         return ret;
5982 }
5983
5984 static u64 stripe_align(struct btrfs_root *root,
5985                         struct btrfs_block_group_cache *cache,
5986                         u64 val, u64 num_bytes)
5987 {
5988         u64 ret = ALIGN(val, root->stripesize);
5989         return ret;
5990 }
5991
5992 /*
5993  * when we wait for progress in the block group caching, its because
5994  * our allocation attempt failed at least once.  So, we must sleep
5995  * and let some progress happen before we try again.
5996  *
5997  * This function will sleep at least once waiting for new free space to
5998  * show up, and then it will check the block group free space numbers
5999  * for our min num_bytes.  Another option is to have it go ahead
6000  * and look in the rbtree for a free extent of a given size, but this
6001  * is a good start.
6002  */
6003 static noinline int
6004 wait_block_group_cache_progress(struct btrfs_block_group_cache *cache,
6005                                 u64 num_bytes)
6006 {
6007         struct btrfs_caching_control *caching_ctl;
6008
6009         caching_ctl = get_caching_control(cache);
6010         if (!caching_ctl)
6011                 return 0;
6012
6013         wait_event(caching_ctl->wait, block_group_cache_done(cache) ||
6014                    (cache->free_space_ctl->free_space >= num_bytes));
6015
6016         put_caching_control(caching_ctl);
6017         return 0;
6018 }
6019
6020 static noinline int
6021 wait_block_group_cache_done(struct btrfs_block_group_cache *cache)
6022 {
6023         struct btrfs_caching_control *caching_ctl;
6024
6025         caching_ctl = get_caching_control(cache);
6026         if (!caching_ctl)
6027                 return 0;
6028
6029         wait_event(caching_ctl->wait, block_group_cache_done(cache));
6030
6031         put_caching_control(caching_ctl);
6032         return 0;
6033 }
6034
6035 int __get_raid_index(u64 flags)
6036 {
6037         if (flags & BTRFS_BLOCK_GROUP_RAID10)
6038                 return BTRFS_RAID_RAID10;
6039         else if (flags & BTRFS_BLOCK_GROUP_RAID1)
6040                 return BTRFS_RAID_RAID1;
6041         else if (flags & BTRFS_BLOCK_GROUP_DUP)
6042                 return BTRFS_RAID_DUP;
6043         else if (flags & BTRFS_BLOCK_GROUP_RAID0)
6044                 return BTRFS_RAID_RAID0;
6045         else if (flags & BTRFS_BLOCK_GROUP_RAID5)
6046                 return BTRFS_RAID_RAID5;
6047         else if (flags & BTRFS_BLOCK_GROUP_RAID6)
6048                 return BTRFS_RAID_RAID6;
6049
6050         return BTRFS_RAID_SINGLE; /* BTRFS_BLOCK_GROUP_SINGLE */
6051 }
6052
6053 static int get_block_group_index(struct btrfs_block_group_cache *cache)
6054 {
6055         return __get_raid_index(cache->flags);
6056 }
6057
6058 enum btrfs_loop_type {
6059         LOOP_CACHING_NOWAIT = 0,
6060         LOOP_CACHING_WAIT = 1,
6061         LOOP_ALLOC_CHUNK = 2,
6062         LOOP_NO_EMPTY_SIZE = 3,
6063 };
6064
6065 /*
6066  * walks the btree of allocated extents and find a hole of a given size.
6067  * The key ins is changed to record the hole:
6068  * ins->objectid == block start
6069  * ins->flags = BTRFS_EXTENT_ITEM_KEY
6070  * ins->offset == number of blocks
6071  * Any available blocks before search_start are skipped.
6072  */
6073 static noinline int find_free_extent(struct btrfs_trans_handle *trans,
6074                                      struct btrfs_root *orig_root,
6075                                      u64 num_bytes, u64 empty_size,
6076                                      u64 hint_byte, struct btrfs_key *ins,
6077                                      u64 flags)
6078 {
6079         int ret = 0;
6080         struct btrfs_root *root = orig_root->fs_info->extent_root;
6081         struct btrfs_free_cluster *last_ptr = NULL;
6082         struct btrfs_block_group_cache *block_group = NULL;
6083         struct btrfs_block_group_cache *used_block_group;
6084         u64 search_start = 0;
6085         int empty_cluster = 2 * 1024 * 1024;
6086         struct btrfs_space_info *space_info;
6087         int loop = 0;
6088         int index = __get_raid_index(flags);
6089         int alloc_type = (flags & BTRFS_BLOCK_GROUP_DATA) ?
6090                 RESERVE_ALLOC_NO_ACCOUNT : RESERVE_ALLOC;
6091         bool found_uncached_bg = false;
6092         bool failed_cluster_refill = false;
6093         bool failed_alloc = false;
6094         bool use_cluster = true;
6095         bool have_caching_bg = false;
6096
6097         WARN_ON(num_bytes < root->sectorsize);
6098         btrfs_set_key_type(ins, BTRFS_EXTENT_ITEM_KEY);
6099         ins->objectid = 0;
6100         ins->offset = 0;
6101
6102         trace_find_free_extent(orig_root, num_bytes, empty_size, flags);
6103
6104         space_info = __find_space_info(root->fs_info, flags);
6105         if (!space_info) {
6106                 btrfs_err(root->fs_info, "No space info for %llu", flags);
6107                 return -ENOSPC;
6108         }
6109
6110         /*
6111          * If the space info is for both data and metadata it means we have a
6112          * small filesystem and we can't use the clustering stuff.
6113          */
6114         if (btrfs_mixed_space_info(space_info))
6115                 use_cluster = false;
6116
6117         if (flags & BTRFS_BLOCK_GROUP_METADATA && use_cluster) {
6118                 last_ptr = &root->fs_info->meta_alloc_cluster;
6119                 if (!btrfs_test_opt(root, SSD))
6120                         empty_cluster = 64 * 1024;
6121         }
6122
6123         if ((flags & BTRFS_BLOCK_GROUP_DATA) && use_cluster &&
6124             btrfs_test_opt(root, SSD)) {
6125                 last_ptr = &root->fs_info->data_alloc_cluster;
6126         }
6127
6128         if (last_ptr) {
6129                 spin_lock(&last_ptr->lock);
6130                 if (last_ptr->block_group)
6131                         hint_byte = last_ptr->window_start;
6132                 spin_unlock(&last_ptr->lock);
6133         }
6134
6135         search_start = max(search_start, first_logical_byte(root, 0));
6136         search_start = max(search_start, hint_byte);
6137
6138         if (!last_ptr)
6139                 empty_cluster = 0;
6140
6141         if (search_start == hint_byte) {
6142                 block_group = btrfs_lookup_block_group(root->fs_info,
6143                                                        search_start);
6144                 used_block_group = block_group;
6145                 /*
6146                  * we don't want to use the block group if it doesn't match our
6147                  * allocation bits, or if its not cached.
6148                  *
6149                  * However if we are re-searching with an ideal block group
6150                  * picked out then we don't care that the block group is cached.
6151                  */
6152                 if (block_group && block_group_bits(block_group, flags) &&
6153                     block_group->cached != BTRFS_CACHE_NO) {
6154                         down_read(&space_info->groups_sem);
6155                         if (list_empty(&block_group->list) ||
6156                             block_group->ro) {
6157                                 /*
6158                                  * someone is removing this block group,
6159                                  * we can't jump into the have_block_group
6160                                  * target because our list pointers are not
6161                                  * valid
6162                                  */
6163                                 btrfs_put_block_group(block_group);
6164                                 up_read(&space_info->groups_sem);
6165                         } else {
6166                                 index = get_block_group_index(block_group);
6167                                 goto have_block_group;
6168                         }
6169                 } else if (block_group) {
6170                         btrfs_put_block_group(block_group);
6171                 }
6172         }
6173 search:
6174         have_caching_bg = false;
6175         down_read(&space_info->groups_sem);
6176         list_for_each_entry(block_group, &space_info->block_groups[index],
6177                             list) {
6178                 u64 offset;
6179                 int cached;
6180
6181                 used_block_group = block_group;
6182                 btrfs_get_block_group(block_group);
6183                 search_start = block_group->key.objectid;
6184
6185                 /*
6186                  * this can happen if we end up cycling through all the
6187                  * raid types, but we want to make sure we only allocate
6188                  * for the proper type.
6189                  */
6190                 if (!block_group_bits(block_group, flags)) {
6191                     u64 extra = BTRFS_BLOCK_GROUP_DUP |
6192                                 BTRFS_BLOCK_GROUP_RAID1 |
6193                                 BTRFS_BLOCK_GROUP_RAID5 |
6194                                 BTRFS_BLOCK_GROUP_RAID6 |
6195                                 BTRFS_BLOCK_GROUP_RAID10;
6196
6197                         /*
6198                          * if they asked for extra copies and this block group
6199                          * doesn't provide them, bail.  This does allow us to
6200                          * fill raid0 from raid1.
6201                          */
6202                         if ((flags & extra) && !(block_group->flags & extra))
6203                                 goto loop;
6204                 }
6205
6206 have_block_group:
6207                 cached = block_group_cache_done(block_group);
6208                 if (unlikely(!cached)) {
6209                         found_uncached_bg = true;
6210                         ret = cache_block_group(block_group, 0);
6211                         BUG_ON(ret < 0);
6212                         ret = 0;
6213                 }
6214
6215                 if (unlikely(block_group->ro))
6216                         goto loop;
6217
6218                 /*
6219                  * Ok we want to try and use the cluster allocator, so
6220                  * lets look there
6221                  */
6222                 if (last_ptr) {
6223                         unsigned long aligned_cluster;
6224                         /*
6225                          * the refill lock keeps out other
6226                          * people trying to start a new cluster
6227                          */
6228                         spin_lock(&last_ptr->refill_lock);
6229                         used_block_group = last_ptr->block_group;
6230                         if (used_block_group != block_group &&
6231                             (!used_block_group ||
6232                              used_block_group->ro ||
6233                              !block_group_bits(used_block_group, flags))) {
6234                                 used_block_group = block_group;
6235                                 goto refill_cluster;
6236                         }
6237
6238                         if (used_block_group != block_group)
6239                                 btrfs_get_block_group(used_block_group);
6240
6241                         offset = btrfs_alloc_from_cluster(used_block_group,
6242                           last_ptr, num_bytes, used_block_group->key.objectid);
6243                         if (offset) {
6244                                 /* we have a block, we're done */
6245                                 spin_unlock(&last_ptr->refill_lock);
6246                                 trace_btrfs_reserve_extent_cluster(root,
6247                                         block_group, search_start, num_bytes);
6248                                 goto checks;
6249                         }
6250
6251                         WARN_ON(last_ptr->block_group != used_block_group);
6252                         if (used_block_group != block_group) {
6253                                 btrfs_put_block_group(used_block_group);
6254                                 used_block_group = block_group;
6255                         }
6256 refill_cluster:
6257                         BUG_ON(used_block_group != block_group);
6258                         /* If we are on LOOP_NO_EMPTY_SIZE, we can't
6259                          * set up a new clusters, so lets just skip it
6260                          * and let the allocator find whatever block
6261                          * it can find.  If we reach this point, we
6262                          * will have tried the cluster allocator
6263                          * plenty of times and not have found
6264                          * anything, so we are likely way too
6265                          * fragmented for the clustering stuff to find
6266                          * anything.
6267                          *
6268                          * However, if the cluster is taken from the
6269                          * current block group, release the cluster
6270                          * first, so that we stand a better chance of
6271                          * succeeding in the unclustered
6272                          * allocation.  */
6273                         if (loop >= LOOP_NO_EMPTY_SIZE &&
6274                             last_ptr->block_group != block_group) {
6275                                 spin_unlock(&last_ptr->refill_lock);
6276                                 goto unclustered_alloc;
6277                         }
6278
6279                         /*
6280                          * this cluster didn't work out, free it and
6281                          * start over
6282                          */
6283                         btrfs_return_cluster_to_free_space(NULL, last_ptr);
6284
6285                         if (loop >= LOOP_NO_EMPTY_SIZE) {
6286                                 spin_unlock(&last_ptr->refill_lock);
6287                                 goto unclustered_alloc;
6288                         }
6289
6290                         aligned_cluster = max_t(unsigned long,
6291                                                 empty_cluster + empty_size,
6292                                               block_group->full_stripe_len);
6293
6294                         /* allocate a cluster in this block group */
6295                         ret = btrfs_find_space_cluster(trans, root,
6296                                                block_group, last_ptr,
6297                                                search_start, num_bytes,
6298                                                aligned_cluster);
6299                         if (ret == 0) {
6300                                 /*
6301                                  * now pull our allocation out of this
6302                                  * cluster
6303                                  */
6304                                 offset = btrfs_alloc_from_cluster(block_group,
6305                                                   last_ptr, num_bytes,
6306                                                   search_start);
6307                                 if (offset) {
6308                                         /* we found one, proceed */
6309                                         spin_unlock(&last_ptr->refill_lock);
6310                                         trace_btrfs_reserve_extent_cluster(root,
6311                                                 block_group, search_start,
6312                                                 num_bytes);
6313                                         goto checks;
6314                                 }
6315                         } else if (!cached && loop > LOOP_CACHING_NOWAIT
6316                                    && !failed_cluster_refill) {
6317                                 spin_unlock(&last_ptr->refill_lock);
6318
6319                                 failed_cluster_refill = true;
6320                                 wait_block_group_cache_progress(block_group,
6321                                        num_bytes + empty_cluster + empty_size);
6322                                 goto have_block_group;
6323                         }
6324
6325                         /*
6326                          * at this point we either didn't find a cluster
6327                          * or we weren't able to allocate a block from our
6328                          * cluster.  Free the cluster we've been trying
6329                          * to use, and go to the next block group
6330                          */
6331                         btrfs_return_cluster_to_free_space(NULL, last_ptr);
6332                         spin_unlock(&last_ptr->refill_lock);
6333                         goto loop;
6334                 }
6335
6336 unclustered_alloc:
6337                 spin_lock(&block_group->free_space_ctl->tree_lock);
6338                 if (cached &&
6339                     block_group->free_space_ctl->free_space <
6340                     num_bytes + empty_cluster + empty_size) {
6341                         spin_unlock(&block_group->free_space_ctl->tree_lock);
6342                         goto loop;
6343                 }
6344                 spin_unlock(&block_group->free_space_ctl->tree_lock);
6345
6346                 offset = btrfs_find_space_for_alloc(block_group, search_start,
6347                                                     num_bytes, empty_size);
6348                 /*
6349                  * If we didn't find a chunk, and we haven't failed on this
6350                  * block group before, and this block group is in the middle of
6351                  * caching and we are ok with waiting, then go ahead and wait
6352                  * for progress to be made, and set failed_alloc to true.
6353                  *
6354                  * If failed_alloc is true then we've already waited on this
6355                  * block group once and should move on to the next block group.
6356                  */
6357                 if (!offset && !failed_alloc && !cached &&
6358                     loop > LOOP_CACHING_NOWAIT) {
6359                         wait_block_group_cache_progress(block_group,
6360                                                 num_bytes + empty_size);
6361                         failed_alloc = true;
6362                         goto have_block_group;
6363                 } else if (!offset) {
6364                         if (!cached)
6365                                 have_caching_bg = true;
6366                         goto loop;
6367                 }
6368 checks:
6369                 search_start = stripe_align(root, used_block_group,
6370                                             offset, num_bytes);
6371
6372                 /* move on to the next group */
6373                 if (search_start + num_bytes >
6374                     used_block_group->key.objectid + used_block_group->key.offset) {
6375                         btrfs_add_free_space(used_block_group, offset, num_bytes);
6376                         goto loop;
6377                 }
6378
6379                 if (offset < search_start)
6380                         btrfs_add_free_space(used_block_group, offset,
6381                                              search_start - offset);
6382                 BUG_ON(offset > search_start);
6383
6384                 ret = btrfs_update_reserved_bytes(used_block_group, num_bytes,
6385                                                   alloc_type);
6386                 if (ret == -EAGAIN) {
6387                         btrfs_add_free_space(used_block_group, offset, num_bytes);
6388                         goto loop;
6389                 }
6390
6391                 /* we are all good, lets return */
6392                 ins->objectid = search_start;
6393                 ins->offset = num_bytes;
6394
6395                 trace_btrfs_reserve_extent(orig_root, block_group,
6396                                            search_start, num_bytes);
6397                 if (used_block_group != block_group)
6398                         btrfs_put_block_group(used_block_group);
6399                 btrfs_put_block_group(block_group);
6400                 break;
6401 loop:
6402                 failed_cluster_refill = false;
6403                 failed_alloc = false;
6404                 BUG_ON(index != get_block_group_index(block_group));
6405                 if (used_block_group != block_group)
6406                         btrfs_put_block_group(used_block_group);
6407                 btrfs_put_block_group(block_group);
6408         }
6409         up_read(&space_info->groups_sem);
6410
6411         if (!ins->objectid && loop >= LOOP_CACHING_WAIT && have_caching_bg)
6412                 goto search;
6413
6414         if (!ins->objectid && ++index < BTRFS_NR_RAID_TYPES)
6415                 goto search;
6416
6417         /*
6418          * LOOP_CACHING_NOWAIT, search partially cached block groups, kicking
6419          *                      caching kthreads as we move along
6420          * LOOP_CACHING_WAIT, search everything, and wait if our bg is caching
6421          * LOOP_ALLOC_CHUNK, force a chunk allocation and try again
6422          * LOOP_NO_EMPTY_SIZE, set empty_size and empty_cluster to 0 and try
6423          *                      again
6424          */
6425         if (!ins->objectid && loop < LOOP_NO_EMPTY_SIZE) {
6426                 index = 0;
6427                 loop++;
6428                 if (loop == LOOP_ALLOC_CHUNK) {
6429                         ret = do_chunk_alloc(trans, root, flags,
6430                                              CHUNK_ALLOC_FORCE);
6431                         /*
6432                          * Do not bail out on ENOSPC since we
6433                          * can do more things.
6434                          */
6435                         if (ret < 0 && ret != -ENOSPC) {
6436                                 btrfs_abort_transaction(trans,
6437                                                         root, ret);
6438                                 goto out;
6439                         }
6440                 }
6441
6442                 if (loop == LOOP_NO_EMPTY_SIZE) {
6443                         empty_size = 0;
6444                         empty_cluster = 0;
6445                 }
6446
6447                 goto search;
6448         } else if (!ins->objectid) {
6449                 ret = -ENOSPC;
6450         } else if (ins->objectid) {
6451                 ret = 0;
6452         }
6453 out:
6454
6455         return ret;
6456 }
6457
6458 static void dump_space_info(struct btrfs_space_info *info, u64 bytes,
6459                             int dump_block_groups)
6460 {
6461         struct btrfs_block_group_cache *cache;
6462         int index = 0;
6463
6464         spin_lock(&info->lock);
6465         printk(KERN_INFO "space_info %llu has %llu free, is %sfull\n",
6466                (unsigned long long)info->flags,
6467                (unsigned long long)(info->total_bytes - info->bytes_used -
6468                                     info->bytes_pinned - info->bytes_reserved -
6469                                     info->bytes_readonly),
6470                (info->full) ? "" : "not ");
6471         printk(KERN_INFO "space_info total=%llu, used=%llu, pinned=%llu, "
6472                "reserved=%llu, may_use=%llu, readonly=%llu\n",
6473                (unsigned long long)info->total_bytes,
6474                (unsigned long long)info->bytes_used,
6475                (unsigned long long)info->bytes_pinned,
6476                (unsigned long long)info->bytes_reserved,
6477                (unsigned long long)info->bytes_may_use,
6478                (unsigned long long)info->bytes_readonly);
6479         spin_unlock(&info->lock);
6480
6481         if (!dump_block_groups)
6482                 return;
6483
6484         down_read(&info->groups_sem);
6485 again:
6486         list_for_each_entry(cache, &info->block_groups[index], list) {
6487                 spin_lock(&cache->lock);
6488                 printk(KERN_INFO "block group %llu has %llu bytes, %llu used %llu pinned %llu reserved %s\n",
6489                        (unsigned long long)cache->key.objectid,
6490                        (unsigned long long)cache->key.offset,
6491                        (unsigned long long)btrfs_block_group_used(&cache->item),
6492                        (unsigned long long)cache->pinned,
6493                        (unsigned long long)cache->reserved,
6494                        cache->ro ? "[readonly]" : "");
6495                 btrfs_dump_free_space(cache, bytes);
6496                 spin_unlock(&cache->lock);
6497         }
6498         if (++index < BTRFS_NR_RAID_TYPES)
6499                 goto again;
6500         up_read(&info->groups_sem);
6501 }
6502
6503 int btrfs_reserve_extent(struct btrfs_trans_handle *trans,
6504                          struct btrfs_root *root,
6505                          u64 num_bytes, u64 min_alloc_size,
6506                          u64 empty_size, u64 hint_byte,
6507                          struct btrfs_key *ins, int is_data)
6508 {
6509         bool final_tried = false;
6510         u64 flags;
6511         int ret;
6512
6513         flags = btrfs_get_alloc_profile(root, is_data);
6514 again:
6515         WARN_ON(num_bytes < root->sectorsize);
6516         ret = find_free_extent(trans, root, num_bytes, empty_size,
6517                                hint_byte, ins, flags);
6518
6519         if (ret == -ENOSPC) {
6520                 if (!final_tried) {
6521                         num_bytes = num_bytes >> 1;
6522                         num_bytes = round_down(num_bytes, root->sectorsize);
6523                         num_bytes = max(num_bytes, min_alloc_size);
6524                         if (num_bytes == min_alloc_size)
6525                                 final_tried = true;
6526                         goto again;
6527                 } else if (btrfs_test_opt(root, ENOSPC_DEBUG)) {
6528                         struct btrfs_space_info *sinfo;
6529
6530                         sinfo = __find_space_info(root->fs_info, flags);
6531                         btrfs_err(root->fs_info, "allocation failed flags %llu, wanted %llu",
6532                                 (unsigned long long)flags,
6533                                 (unsigned long long)num_bytes);
6534                         if (sinfo)
6535                                 dump_space_info(sinfo, num_bytes, 1);
6536                 }
6537         }
6538
6539         trace_btrfs_reserved_extent_alloc(root, ins->objectid, ins->offset);
6540
6541         return ret;
6542 }
6543
6544 static int __btrfs_free_reserved_extent(struct btrfs_root *root,
6545                                         u64 start, u64 len, int pin)
6546 {
6547         struct btrfs_block_group_cache *cache;
6548         int ret = 0;
6549
6550         cache = btrfs_lookup_block_group(root->fs_info, start);
6551         if (!cache) {
6552                 btrfs_err(root->fs_info, "Unable to find block group for %llu",
6553                         (unsigned long long)start);
6554                 return -ENOSPC;
6555         }
6556
6557         if (btrfs_test_opt(root, DISCARD))
6558                 ret = btrfs_discard_extent(root, start, len, NULL);
6559
6560         if (pin)
6561                 pin_down_extent(root, cache, start, len, 1);
6562         else {
6563                 btrfs_add_free_space(cache, start, len);
6564                 btrfs_update_reserved_bytes(cache, len, RESERVE_FREE);
6565         }
6566         btrfs_put_block_group(cache);
6567
6568         trace_btrfs_reserved_extent_free(root, start, len);
6569
6570         return ret;
6571 }
6572
6573 int btrfs_free_reserved_extent(struct btrfs_root *root,
6574                                         u64 start, u64 len)
6575 {
6576         return __btrfs_free_reserved_extent(root, start, len, 0);
6577 }
6578
6579 int btrfs_free_and_pin_reserved_extent(struct btrfs_root *root,
6580                                        u64 start, u64 len)
6581 {
6582         return __btrfs_free_reserved_extent(root, start, len, 1);
6583 }
6584
6585 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
6586                                       struct btrfs_root *root,
6587                                       u64 parent, u64 root_objectid,
6588                                       u64 flags, u64 owner, u64 offset,
6589                                       struct btrfs_key *ins, int ref_mod)
6590 {
6591         int ret;
6592         struct btrfs_fs_info *fs_info = root->fs_info;
6593         struct btrfs_extent_item *extent_item;
6594         struct btrfs_extent_inline_ref *iref;
6595         struct btrfs_path *path;
6596         struct extent_buffer *leaf;
6597         int type;
6598         u32 size;
6599
6600         if (parent > 0)
6601                 type = BTRFS_SHARED_DATA_REF_KEY;
6602         else
6603                 type = BTRFS_EXTENT_DATA_REF_KEY;
6604
6605         size = sizeof(*extent_item) + btrfs_extent_inline_ref_size(type);
6606
6607         path = btrfs_alloc_path();
6608         if (!path)
6609                 return -ENOMEM;
6610
6611         path->leave_spinning = 1;
6612         ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path,
6613                                       ins, size);
6614         if (ret) {
6615                 btrfs_free_path(path);
6616                 return ret;
6617         }
6618
6619         leaf = path->nodes[0];
6620         extent_item = btrfs_item_ptr(leaf, path->slots[0],
6621                                      struct btrfs_extent_item);
6622         btrfs_set_extent_refs(leaf, extent_item, ref_mod);
6623         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
6624         btrfs_set_extent_flags(leaf, extent_item,
6625                                flags | BTRFS_EXTENT_FLAG_DATA);
6626
6627         iref = (struct btrfs_extent_inline_ref *)(extent_item + 1);
6628         btrfs_set_extent_inline_ref_type(leaf, iref, type);
6629         if (parent > 0) {
6630                 struct btrfs_shared_data_ref *ref;
6631                 ref = (struct btrfs_shared_data_ref *)(iref + 1);
6632                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
6633                 btrfs_set_shared_data_ref_count(leaf, ref, ref_mod);
6634         } else {
6635                 struct btrfs_extent_data_ref *ref;
6636                 ref = (struct btrfs_extent_data_ref *)(&iref->offset);
6637                 btrfs_set_extent_data_ref_root(leaf, ref, root_objectid);
6638                 btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
6639                 btrfs_set_extent_data_ref_offset(leaf, ref, offset);
6640                 btrfs_set_extent_data_ref_count(leaf, ref, ref_mod);
6641         }
6642
6643         btrfs_mark_buffer_dirty(path->nodes[0]);
6644         btrfs_free_path(path);
6645
6646         ret = update_block_group(root, ins->objectid, ins->offset, 1);
6647         if (ret) { /* -ENOENT, logic error */
6648                 btrfs_err(fs_info, "update block group failed for %llu %llu",
6649                         (unsigned long long)ins->objectid,
6650                         (unsigned long long)ins->offset);
6651                 BUG();
6652         }
6653         return ret;
6654 }
6655
6656 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
6657                                      struct btrfs_root *root,
6658                                      u64 parent, u64 root_objectid,
6659                                      u64 flags, struct btrfs_disk_key *key,
6660                                      int level, struct btrfs_key *ins)
6661 {
6662         int ret;
6663         struct btrfs_fs_info *fs_info = root->fs_info;
6664         struct btrfs_extent_item *extent_item;
6665         struct btrfs_tree_block_info *block_info;
6666         struct btrfs_extent_inline_ref *iref;
6667         struct btrfs_path *path;
6668         struct extent_buffer *leaf;
6669         u32 size = sizeof(*extent_item) + sizeof(*iref);
6670         bool skinny_metadata = btrfs_fs_incompat(root->fs_info,
6671                                                  SKINNY_METADATA);
6672
6673         if (!skinny_metadata)
6674                 size += sizeof(*block_info);
6675
6676         path = btrfs_alloc_path();
6677         if (!path)
6678                 return -ENOMEM;
6679
6680         path->leave_spinning = 1;
6681         ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path,
6682                                       ins, size);
6683         if (ret) {
6684                 btrfs_free_path(path);
6685                 return ret;
6686         }
6687
6688         leaf = path->nodes[0];
6689         extent_item = btrfs_item_ptr(leaf, path->slots[0],
6690                                      struct btrfs_extent_item);
6691         btrfs_set_extent_refs(leaf, extent_item, 1);
6692         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
6693         btrfs_set_extent_flags(leaf, extent_item,
6694                                flags | BTRFS_EXTENT_FLAG_TREE_BLOCK);
6695
6696         if (skinny_metadata) {
6697                 iref = (struct btrfs_extent_inline_ref *)(extent_item + 1);
6698         } else {
6699                 block_info = (struct btrfs_tree_block_info *)(extent_item + 1);
6700                 btrfs_set_tree_block_key(leaf, block_info, key);
6701                 btrfs_set_tree_block_level(leaf, block_info, level);
6702                 iref = (struct btrfs_extent_inline_ref *)(block_info + 1);
6703         }
6704
6705         if (parent > 0) {
6706                 BUG_ON(!(flags & BTRFS_BLOCK_FLAG_FULL_BACKREF));
6707                 btrfs_set_extent_inline_ref_type(leaf, iref,
6708                                                  BTRFS_SHARED_BLOCK_REF_KEY);
6709                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
6710         } else {
6711                 btrfs_set_extent_inline_ref_type(leaf, iref,
6712                                                  BTRFS_TREE_BLOCK_REF_KEY);
6713                 btrfs_set_extent_inline_ref_offset(leaf, iref, root_objectid);
6714         }
6715
6716         btrfs_mark_buffer_dirty(leaf);
6717         btrfs_free_path(path);
6718
6719         ret = update_block_group(root, ins->objectid, root->leafsize, 1);
6720         if (ret) { /* -ENOENT, logic error */
6721                 btrfs_err(fs_info, "update block group failed for %llu %llu",
6722                         (unsigned long long)ins->objectid,
6723                         (unsigned long long)ins->offset);
6724                 BUG();
6725         }
6726         return ret;
6727 }
6728
6729 int btrfs_alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
6730                                      struct btrfs_root *root,
6731                                      u64 root_objectid, u64 owner,
6732                                      u64 offset, struct btrfs_key *ins)
6733 {
6734         int ret;
6735
6736         BUG_ON(root_objectid == BTRFS_TREE_LOG_OBJECTID);
6737
6738         ret = btrfs_add_delayed_data_ref(root->fs_info, trans, ins->objectid,
6739                                          ins->offset, 0,
6740                                          root_objectid, owner, offset,
6741                                          BTRFS_ADD_DELAYED_EXTENT, NULL, 0);
6742         return ret;
6743 }
6744
6745 /*
6746  * this is used by the tree logging recovery code.  It records that
6747  * an extent has been allocated and makes sure to clear the free
6748  * space cache bits as well
6749  */
6750 int btrfs_alloc_logged_file_extent(struct btrfs_trans_handle *trans,
6751                                    struct btrfs_root *root,
6752                                    u64 root_objectid, u64 owner, u64 offset,
6753                                    struct btrfs_key *ins)
6754 {
6755         int ret;
6756         struct btrfs_block_group_cache *block_group;
6757
6758         /*
6759          * Mixed block groups will exclude before processing the log so we only
6760          * need to do the exlude dance if this fs isn't mixed.
6761          */
6762         if (!btrfs_fs_incompat(root->fs_info, MIXED_GROUPS)) {
6763                 ret = __exclude_logged_extent(root, ins->objectid, ins->offset);
6764                 if (ret)
6765                         return ret;
6766         }
6767
6768         block_group = btrfs_lookup_block_group(root->fs_info, ins->objectid);
6769         if (!block_group)
6770                 return -EINVAL;
6771
6772         ret = btrfs_update_reserved_bytes(block_group, ins->offset,
6773                                           RESERVE_ALLOC_NO_ACCOUNT);
6774         BUG_ON(ret); /* logic error */
6775         ret = alloc_reserved_file_extent(trans, root, 0, root_objectid,
6776                                          0, owner, offset, ins, 1);
6777         btrfs_put_block_group(block_group);
6778         return ret;
6779 }
6780
6781 static struct extent_buffer *
6782 btrfs_init_new_buffer(struct btrfs_trans_handle *trans, struct btrfs_root *root,
6783                       u64 bytenr, u32 blocksize, int level)
6784 {
6785         struct extent_buffer *buf;
6786
6787         buf = btrfs_find_create_tree_block(root, bytenr, blocksize);
6788         if (!buf)
6789                 return ERR_PTR(-ENOMEM);
6790         btrfs_set_header_generation(buf, trans->transid);
6791         btrfs_set_buffer_lockdep_class(root->root_key.objectid, buf, level);
6792         btrfs_tree_lock(buf);
6793         clean_tree_block(trans, root, buf);
6794         clear_bit(EXTENT_BUFFER_STALE, &buf->bflags);
6795
6796         btrfs_set_lock_blocking(buf);
6797         btrfs_set_buffer_uptodate(buf);
6798
6799         if (root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID) {
6800                 /*
6801                  * we allow two log transactions at a time, use different
6802                  * EXENT bit to differentiate dirty pages.
6803                  */
6804                 if (root->log_transid % 2 == 0)
6805                         set_extent_dirty(&root->dirty_log_pages, buf->start,
6806                                         buf->start + buf->len - 1, GFP_NOFS);
6807                 else
6808                         set_extent_new(&root->dirty_log_pages, buf->start,
6809                                         buf->start + buf->len - 1, GFP_NOFS);
6810         } else {
6811                 set_extent_dirty(&trans->transaction->dirty_pages, buf->start,
6812                          buf->start + buf->len - 1, GFP_NOFS);
6813         }
6814         trans->blocks_used++;
6815         /* this returns a buffer locked for blocking */
6816         return buf;
6817 }
6818
6819 static struct btrfs_block_rsv *
6820 use_block_rsv(struct btrfs_trans_handle *trans,
6821               struct btrfs_root *root, u32 blocksize)
6822 {
6823         struct btrfs_block_rsv *block_rsv;
6824         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
6825         int ret;
6826         bool global_updated = false;
6827
6828         block_rsv = get_block_rsv(trans, root);
6829
6830         if (unlikely(block_rsv->size == 0))
6831                 goto try_reserve;
6832 again:
6833         ret = block_rsv_use_bytes(block_rsv, blocksize);
6834         if (!ret)
6835                 return block_rsv;
6836
6837         if (block_rsv->failfast)
6838                 return ERR_PTR(ret);
6839
6840         if (block_rsv->type == BTRFS_BLOCK_RSV_GLOBAL && !global_updated) {
6841                 global_updated = true;
6842                 update_global_block_rsv(root->fs_info);
6843                 goto again;
6844         }
6845
6846         if (btrfs_test_opt(root, ENOSPC_DEBUG)) {
6847                 static DEFINE_RATELIMIT_STATE(_rs,
6848                                 DEFAULT_RATELIMIT_INTERVAL * 10,
6849                                 /*DEFAULT_RATELIMIT_BURST*/ 1);
6850                 if (__ratelimit(&_rs))
6851                         WARN(1, KERN_DEBUG
6852                                 "btrfs: block rsv returned %d\n", ret);
6853         }
6854 try_reserve:
6855         ret = reserve_metadata_bytes(root, block_rsv, blocksize,
6856                                      BTRFS_RESERVE_NO_FLUSH);
6857         if (!ret)
6858                 return block_rsv;
6859         /*
6860          * If we couldn't reserve metadata bytes try and use some from
6861          * the global reserve if its space type is the same as the global
6862          * reservation.
6863          */
6864         if (block_rsv->type != BTRFS_BLOCK_RSV_GLOBAL &&
6865             block_rsv->space_info == global_rsv->space_info) {
6866                 ret = block_rsv_use_bytes(global_rsv, blocksize);
6867                 if (!ret)
6868                         return global_rsv;
6869         }
6870         return ERR_PTR(ret);
6871 }
6872
6873 static void unuse_block_rsv(struct btrfs_fs_info *fs_info,
6874                             struct btrfs_block_rsv *block_rsv, u32 blocksize)
6875 {
6876         block_rsv_add_bytes(block_rsv, blocksize, 0);
6877         block_rsv_release_bytes(fs_info, block_rsv, NULL, 0);
6878 }
6879
6880 /*
6881  * finds a free extent and does all the dirty work required for allocation
6882  * returns the key for the extent through ins, and a tree buffer for
6883  * the first block of the extent through buf.
6884  *
6885  * returns the tree buffer or NULL.
6886  */
6887 struct extent_buffer *btrfs_alloc_free_block(struct btrfs_trans_handle *trans,
6888                                         struct btrfs_root *root, u32 blocksize,
6889                                         u64 parent, u64 root_objectid,
6890                                         struct btrfs_disk_key *key, int level,
6891                                         u64 hint, u64 empty_size)
6892 {
6893         struct btrfs_key ins;
6894         struct btrfs_block_rsv *block_rsv;
6895         struct extent_buffer *buf;
6896         u64 flags = 0;
6897         int ret;
6898         bool skinny_metadata = btrfs_fs_incompat(root->fs_info,
6899                                                  SKINNY_METADATA);
6900
6901         block_rsv = use_block_rsv(trans, root, blocksize);
6902         if (IS_ERR(block_rsv))
6903                 return ERR_CAST(block_rsv);
6904
6905         ret = btrfs_reserve_extent(trans, root, blocksize, blocksize,
6906                                    empty_size, hint, &ins, 0);
6907         if (ret) {
6908                 unuse_block_rsv(root->fs_info, block_rsv, blocksize);
6909                 return ERR_PTR(ret);
6910         }
6911
6912         buf = btrfs_init_new_buffer(trans, root, ins.objectid,
6913                                     blocksize, level);
6914         BUG_ON(IS_ERR(buf)); /* -ENOMEM */
6915
6916         if (root_objectid == BTRFS_TREE_RELOC_OBJECTID) {
6917                 if (parent == 0)
6918                         parent = ins.objectid;
6919                 flags |= BTRFS_BLOCK_FLAG_FULL_BACKREF;
6920         } else
6921                 BUG_ON(parent > 0);
6922
6923         if (root_objectid != BTRFS_TREE_LOG_OBJECTID) {
6924                 struct btrfs_delayed_extent_op *extent_op;
6925                 extent_op = btrfs_alloc_delayed_extent_op();
6926                 BUG_ON(!extent_op); /* -ENOMEM */
6927                 if (key)
6928                         memcpy(&extent_op->key, key, sizeof(extent_op->key));
6929                 else
6930                         memset(&extent_op->key, 0, sizeof(extent_op->key));
6931                 extent_op->flags_to_set = flags;
6932                 if (skinny_metadata)
6933                         extent_op->update_key = 0;
6934                 else
6935                         extent_op->update_key = 1;
6936                 extent_op->update_flags = 1;
6937                 extent_op->is_data = 0;
6938                 extent_op->level = level;
6939
6940                 ret = btrfs_add_delayed_tree_ref(root->fs_info, trans,
6941                                         ins.objectid,
6942                                         ins.offset, parent, root_objectid,
6943                                         level, BTRFS_ADD_DELAYED_EXTENT,
6944                                         extent_op, 0);
6945                 BUG_ON(ret); /* -ENOMEM */
6946         }
6947         return buf;
6948 }
6949
6950 struct walk_control {
6951         u64 refs[BTRFS_MAX_LEVEL];
6952         u64 flags[BTRFS_MAX_LEVEL];
6953         struct btrfs_key update_progress;
6954         int stage;
6955         int level;
6956         int shared_level;
6957         int update_ref;
6958         int keep_locks;
6959         int reada_slot;
6960         int reada_count;
6961         int for_reloc;
6962 };
6963
6964 #define DROP_REFERENCE  1
6965 #define UPDATE_BACKREF  2
6966
6967 static noinline void reada_walk_down(struct btrfs_trans_handle *trans,
6968                                      struct btrfs_root *root,
6969                                      struct walk_control *wc,
6970                                      struct btrfs_path *path)
6971 {
6972         u64 bytenr;
6973         u64 generation;
6974         u64 refs;
6975         u64 flags;
6976         u32 nritems;
6977         u32 blocksize;
6978         struct btrfs_key key;
6979         struct extent_buffer *eb;
6980         int ret;
6981         int slot;
6982         int nread = 0;
6983
6984         if (path->slots[wc->level] < wc->reada_slot) {
6985                 wc->reada_count = wc->reada_count * 2 / 3;
6986                 wc->reada_count = max(wc->reada_count, 2);
6987         } else {
6988                 wc->reada_count = wc->reada_count * 3 / 2;
6989                 wc->reada_count = min_t(int, wc->reada_count,
6990                                         BTRFS_NODEPTRS_PER_BLOCK(root));
6991         }
6992
6993         eb = path->nodes[wc->level];
6994         nritems = btrfs_header_nritems(eb);
6995         blocksize = btrfs_level_size(root, wc->level - 1);
6996
6997         for (slot = path->slots[wc->level]; slot < nritems; slot++) {
6998                 if (nread >= wc->reada_count)
6999                         break;
7000
7001                 cond_resched();
7002                 bytenr = btrfs_node_blockptr(eb, slot);
7003                 generation = btrfs_node_ptr_generation(eb, slot);
7004
7005                 if (slot == path->slots[wc->level])
7006                         goto reada;
7007
7008                 if (wc->stage == UPDATE_BACKREF &&
7009                     generation <= root->root_key.offset)
7010                         continue;
7011
7012                 /* We don't lock the tree block, it's OK to be racy here */
7013                 ret = btrfs_lookup_extent_info(trans, root, bytenr,
7014                                                wc->level - 1, 1, &refs,
7015                                                &flags);
7016                 /* We don't care about errors in readahead. */
7017                 if (ret < 0)
7018                         continue;
7019                 BUG_ON(refs == 0);
7020
7021                 if (wc->stage == DROP_REFERENCE) {
7022                         if (refs == 1)
7023                                 goto reada;
7024
7025                         if (wc->level == 1 &&
7026                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
7027                                 continue;
7028                         if (!wc->update_ref ||
7029                             generation <= root->root_key.offset)
7030                                 continue;
7031                         btrfs_node_key_to_cpu(eb, &key, slot);
7032                         ret = btrfs_comp_cpu_keys(&key,
7033                                                   &wc->update_progress);
7034                         if (ret < 0)
7035                                 continue;
7036                 } else {
7037                         if (wc->level == 1 &&
7038                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
7039                                 continue;
7040                 }
7041 reada:
7042                 ret = readahead_tree_block(root, bytenr, blocksize,
7043                                            generation);
7044                 if (ret)
7045                         break;
7046                 nread++;
7047         }
7048         wc->reada_slot = slot;
7049 }
7050
7051 /*
7052  * helper to process tree block while walking down the tree.
7053  *
7054  * when wc->stage == UPDATE_BACKREF, this function updates
7055  * back refs for pointers in the block.
7056  *
7057  * NOTE: return value 1 means we should stop walking down.
7058  */
7059 static noinline int walk_down_proc(struct btrfs_trans_handle *trans,
7060                                    struct btrfs_root *root,
7061                                    struct btrfs_path *path,
7062                                    struct walk_control *wc, int lookup_info)
7063 {
7064         int level = wc->level;
7065         struct extent_buffer *eb = path->nodes[level];
7066         u64 flag = BTRFS_BLOCK_FLAG_FULL_BACKREF;
7067         int ret;
7068
7069         if (wc->stage == UPDATE_BACKREF &&
7070             btrfs_header_owner(eb) != root->root_key.objectid)
7071                 return 1;
7072
7073         /*
7074          * when reference count of tree block is 1, it won't increase
7075          * again. once full backref flag is set, we never clear it.
7076          */
7077         if (lookup_info &&
7078             ((wc->stage == DROP_REFERENCE && wc->refs[level] != 1) ||
7079              (wc->stage == UPDATE_BACKREF && !(wc->flags[level] & flag)))) {
7080                 BUG_ON(!path->locks[level]);
7081                 ret = btrfs_lookup_extent_info(trans, root,
7082                                                eb->start, level, 1,
7083                                                &wc->refs[level],
7084                                                &wc->flags[level]);
7085                 BUG_ON(ret == -ENOMEM);
7086                 if (ret)
7087                         return ret;
7088                 BUG_ON(wc->refs[level] == 0);
7089         }
7090
7091         if (wc->stage == DROP_REFERENCE) {
7092                 if (wc->refs[level] > 1)
7093                         return 1;
7094
7095                 if (path->locks[level] && !wc->keep_locks) {
7096                         btrfs_tree_unlock_rw(eb, path->locks[level]);
7097                         path->locks[level] = 0;
7098                 }
7099                 return 0;
7100         }
7101
7102         /* wc->stage == UPDATE_BACKREF */
7103         if (!(wc->flags[level] & flag)) {
7104                 BUG_ON(!path->locks[level]);
7105                 ret = btrfs_inc_ref(trans, root, eb, 1, wc->for_reloc);
7106                 BUG_ON(ret); /* -ENOMEM */
7107                 ret = btrfs_dec_ref(trans, root, eb, 0, wc->for_reloc);
7108                 BUG_ON(ret); /* -ENOMEM */
7109                 ret = btrfs_set_disk_extent_flags(trans, root, eb->start,
7110                                                   eb->len, flag,
7111                                                   btrfs_header_level(eb), 0);
7112                 BUG_ON(ret); /* -ENOMEM */
7113                 wc->flags[level] |= flag;
7114         }
7115
7116         /*
7117          * the block is shared by multiple trees, so it's not good to
7118          * keep the tree lock
7119          */
7120         if (path->locks[level] && level > 0) {
7121                 btrfs_tree_unlock_rw(eb, path->locks[level]);
7122                 path->locks[level] = 0;
7123         }
7124         return 0;
7125 }
7126
7127 /*
7128  * helper to process tree block pointer.
7129  *
7130  * when wc->stage == DROP_REFERENCE, this function checks
7131  * reference count of the block pointed to. if the block
7132  * is shared and we need update back refs for the subtree
7133  * rooted at the block, this function changes wc->stage to
7134  * UPDATE_BACKREF. if the block is shared and there is no
7135  * need to update back, this function drops the reference
7136  * to the block.
7137  *
7138  * NOTE: return value 1 means we should stop walking down.
7139  */
7140 static noinline int do_walk_down(struct btrfs_trans_handle *trans,
7141                                  struct btrfs_root *root,
7142                                  struct btrfs_path *path,
7143                                  struct walk_control *wc, int *lookup_info)
7144 {
7145         u64 bytenr;
7146         u64 generation;
7147         u64 parent;
7148         u32 blocksize;
7149         struct btrfs_key key;
7150         struct extent_buffer *next;
7151         int level = wc->level;
7152         int reada = 0;
7153         int ret = 0;
7154
7155         generation = btrfs_node_ptr_generation(path->nodes[level],
7156                                                path->slots[level]);
7157         /*
7158          * if the lower level block was created before the snapshot
7159          * was created, we know there is no need to update back refs
7160          * for the subtree
7161          */
7162         if (wc->stage == UPDATE_BACKREF &&
7163             generation <= root->root_key.offset) {
7164                 *lookup_info = 1;
7165                 return 1;
7166         }
7167
7168         bytenr = btrfs_node_blockptr(path->nodes[level], path->slots[level]);
7169         blocksize = btrfs_level_size(root, level - 1);
7170
7171         next = btrfs_find_tree_block(root, bytenr, blocksize);
7172         if (!next) {
7173                 next = btrfs_find_create_tree_block(root, bytenr, blocksize);
7174                 if (!next)
7175                         return -ENOMEM;
7176                 reada = 1;
7177         }
7178         btrfs_tree_lock(next);
7179         btrfs_set_lock_blocking(next);
7180
7181         ret = btrfs_lookup_extent_info(trans, root, bytenr, level - 1, 1,
7182                                        &wc->refs[level - 1],
7183                                        &wc->flags[level - 1]);
7184         if (ret < 0) {
7185                 btrfs_tree_unlock(next);
7186                 return ret;
7187         }
7188
7189         if (unlikely(wc->refs[level - 1] == 0)) {
7190                 btrfs_err(root->fs_info, "Missing references.");
7191                 BUG();
7192         }
7193         *lookup_info = 0;
7194
7195         if (wc->stage == DROP_REFERENCE) {
7196                 if (wc->refs[level - 1] > 1) {
7197                         if (level == 1 &&
7198                             (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
7199                                 goto skip;
7200
7201                         if (!wc->update_ref ||
7202                             generation <= root->root_key.offset)
7203                                 goto skip;
7204
7205                         btrfs_node_key_to_cpu(path->nodes[level], &key,
7206                                               path->slots[level]);
7207                         ret = btrfs_comp_cpu_keys(&key, &wc->update_progress);
7208                         if (ret < 0)
7209                                 goto skip;
7210
7211                         wc->stage = UPDATE_BACKREF;
7212                         wc->shared_level = level - 1;
7213                 }
7214         } else {
7215                 if (level == 1 &&
7216                     (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
7217                         goto skip;
7218         }
7219
7220         if (!btrfs_buffer_uptodate(next, generation, 0)) {
7221                 btrfs_tree_unlock(next);
7222                 free_extent_buffer(next);
7223                 next = NULL;
7224                 *lookup_info = 1;
7225         }
7226
7227         if (!next) {
7228                 if (reada && level == 1)
7229                         reada_walk_down(trans, root, wc, path);
7230                 next = read_tree_block(root, bytenr, blocksize, generation);
7231                 if (!next || !extent_buffer_uptodate(next)) {
7232                         free_extent_buffer(next);
7233                         return -EIO;
7234                 }
7235                 btrfs_tree_lock(next);
7236                 btrfs_set_lock_blocking(next);
7237         }
7238
7239         level--;
7240         BUG_ON(level != btrfs_header_level(next));
7241         path->nodes[level] = next;
7242         path->slots[level] = 0;
7243         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
7244         wc->level = level;
7245         if (wc->level == 1)
7246                 wc->reada_slot = 0;
7247         return 0;
7248 skip:
7249         wc->refs[level - 1] = 0;
7250         wc->flags[level - 1] = 0;
7251         if (wc->stage == DROP_REFERENCE) {
7252                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF) {
7253                         parent = path->nodes[level]->start;
7254                 } else {
7255                         BUG_ON(root->root_key.objectid !=
7256                                btrfs_header_owner(path->nodes[level]));
7257                         parent = 0;
7258                 }
7259
7260                 ret = btrfs_free_extent(trans, root, bytenr, blocksize, parent,
7261                                 root->root_key.objectid, level - 1, 0, 0);
7262                 BUG_ON(ret); /* -ENOMEM */
7263         }
7264         btrfs_tree_unlock(next);
7265         free_extent_buffer(next);
7266         *lookup_info = 1;
7267         return 1;
7268 }
7269
7270 /*
7271  * helper to process tree block while walking up the tree.
7272  *
7273  * when wc->stage == DROP_REFERENCE, this function drops
7274  * reference count on the block.
7275  *
7276  * when wc->stage == UPDATE_BACKREF, this function changes
7277  * wc->stage back to DROP_REFERENCE if we changed wc->stage
7278  * to UPDATE_BACKREF previously while processing the block.
7279  *
7280  * NOTE: return value 1 means we should stop walking up.
7281  */
7282 static noinline int walk_up_proc(struct btrfs_trans_handle *trans,
7283                                  struct btrfs_root *root,
7284                                  struct btrfs_path *path,
7285                                  struct walk_control *wc)
7286 {
7287         int ret;
7288         int level = wc->level;
7289         struct extent_buffer *eb = path->nodes[level];
7290         u64 parent = 0;
7291
7292         if (wc->stage == UPDATE_BACKREF) {
7293                 BUG_ON(wc->shared_level < level);
7294                 if (level < wc->shared_level)
7295                         goto out;
7296
7297                 ret = find_next_key(path, level + 1, &wc->update_progress);
7298                 if (ret > 0)
7299                         wc->update_ref = 0;
7300
7301                 wc->stage = DROP_REFERENCE;
7302                 wc->shared_level = -1;
7303                 path->slots[level] = 0;
7304
7305                 /*
7306                  * check reference count again if the block isn't locked.
7307                  * we should start walking down the tree again if reference
7308                  * count is one.
7309                  */
7310                 if (!path->locks[level]) {
7311                         BUG_ON(level == 0);
7312                         btrfs_tree_lock(eb);
7313                         btrfs_set_lock_blocking(eb);
7314                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
7315
7316                         ret = btrfs_lookup_extent_info(trans, root,
7317                                                        eb->start, level, 1,
7318                                                        &wc->refs[level],
7319                                                        &wc->flags[level]);
7320                         if (ret < 0) {
7321                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
7322                                 path->locks[level] = 0;
7323                                 return ret;
7324                         }
7325                         BUG_ON(wc->refs[level] == 0);
7326                         if (wc->refs[level] == 1) {
7327                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
7328                                 path->locks[level] = 0;
7329                                 return 1;
7330                         }
7331                 }
7332         }
7333
7334         /* wc->stage == DROP_REFERENCE */
7335         BUG_ON(wc->refs[level] > 1 && !path->locks[level]);
7336
7337         if (wc->refs[level] == 1) {
7338                 if (level == 0) {
7339                         if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
7340                                 ret = btrfs_dec_ref(trans, root, eb, 1,
7341                                                     wc->for_reloc);
7342                         else
7343                                 ret = btrfs_dec_ref(trans, root, eb, 0,
7344                                                     wc->for_reloc);
7345                         BUG_ON(ret); /* -ENOMEM */
7346                 }
7347                 /* make block locked assertion in clean_tree_block happy */
7348                 if (!path->locks[level] &&
7349                     btrfs_header_generation(eb) == trans->transid) {
7350                         btrfs_tree_lock(eb);
7351                         btrfs_set_lock_blocking(eb);
7352                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
7353                 }
7354                 clean_tree_block(trans, root, eb);
7355         }
7356
7357         if (eb == root->node) {
7358                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
7359                         parent = eb->start;
7360                 else
7361                         BUG_ON(root->root_key.objectid !=
7362                                btrfs_header_owner(eb));
7363         } else {
7364                 if (wc->flags[level + 1] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
7365                         parent = path->nodes[level + 1]->start;
7366                 else
7367                         BUG_ON(root->root_key.objectid !=
7368                                btrfs_header_owner(path->nodes[level + 1]));
7369         }
7370
7371         btrfs_free_tree_block(trans, root, eb, parent, wc->refs[level] == 1);
7372 out:
7373         wc->refs[level] = 0;
7374         wc->flags[level] = 0;
7375         return 0;
7376 }
7377
7378 static noinline int walk_down_tree(struct btrfs_trans_handle *trans,
7379                                    struct btrfs_root *root,
7380                                    struct btrfs_path *path,
7381                                    struct walk_control *wc)
7382 {
7383         int level = wc->level;
7384         int lookup_info = 1;
7385         int ret;
7386
7387         while (level >= 0) {
7388                 ret = walk_down_proc(trans, root, path, wc, lookup_info);
7389                 if (ret > 0)
7390                         break;
7391
7392                 if (level == 0)
7393                         break;
7394
7395                 if (path->slots[level] >=
7396                     btrfs_header_nritems(path->nodes[level]))
7397                         break;
7398
7399                 ret = do_walk_down(trans, root, path, wc, &lookup_info);
7400                 if (ret > 0) {
7401                         path->slots[level]++;
7402                         continue;
7403                 } else if (ret < 0)
7404                         return ret;
7405                 level = wc->level;
7406         }
7407         return 0;
7408 }
7409
7410 static noinline int walk_up_tree(struct btrfs_trans_handle *trans,
7411                                  struct btrfs_root *root,
7412                                  struct btrfs_path *path,
7413                                  struct walk_control *wc, int max_level)
7414 {
7415         int level = wc->level;
7416         int ret;
7417
7418         path->slots[level] = btrfs_header_nritems(path->nodes[level]);
7419         while (level < max_level && path->nodes[level]) {
7420                 wc->level = level;
7421                 if (path->slots[level] + 1 <
7422                     btrfs_header_nritems(path->nodes[level])) {
7423                         path->slots[level]++;
7424                         return 0;
7425                 } else {
7426                         ret = walk_up_proc(trans, root, path, wc);
7427                         if (ret > 0)
7428                                 return 0;
7429
7430                         if (path->locks[level]) {
7431                                 btrfs_tree_unlock_rw(path->nodes[level],
7432                                                      path->locks[level]);
7433                                 path->locks[level] = 0;
7434                         }
7435                         free_extent_buffer(path->nodes[level]);
7436                         path->nodes[level] = NULL;
7437                         level++;
7438                 }
7439         }
7440         return 1;
7441 }
7442
7443 /*
7444  * drop a subvolume tree.
7445  *
7446  * this function traverses the tree freeing any blocks that only
7447  * referenced by the tree.
7448  *
7449  * when a shared tree block is found. this function decreases its
7450  * reference count by one. if update_ref is true, this function
7451  * also make sure backrefs for the shared block and all lower level
7452  * blocks are properly updated.
7453  *
7454  * If called with for_reloc == 0, may exit early with -EAGAIN
7455  */
7456 int btrfs_drop_snapshot(struct btrfs_root *root,
7457                          struct btrfs_block_rsv *block_rsv, int update_ref,
7458                          int for_reloc)
7459 {
7460         struct btrfs_path *path;
7461         struct btrfs_trans_handle *trans;
7462         struct btrfs_root *tree_root = root->fs_info->tree_root;
7463         struct btrfs_root_item *root_item = &root->root_item;
7464         struct walk_control *wc;
7465         struct btrfs_key key;
7466         int err = 0;
7467         int ret;
7468         int level;
7469
7470         path = btrfs_alloc_path();
7471         if (!path) {
7472                 err = -ENOMEM;
7473                 goto out;
7474         }
7475
7476         wc = kzalloc(sizeof(*wc), GFP_NOFS);
7477         if (!wc) {
7478                 btrfs_free_path(path);
7479                 err = -ENOMEM;
7480                 goto out;
7481         }
7482
7483         trans = btrfs_start_transaction(tree_root, 0);
7484         if (IS_ERR(trans)) {
7485                 err = PTR_ERR(trans);
7486                 goto out_free;
7487         }
7488
7489         if (block_rsv)
7490                 trans->block_rsv = block_rsv;
7491
7492         if (btrfs_disk_key_objectid(&root_item->drop_progress) == 0) {
7493                 level = btrfs_header_level(root->node);
7494                 path->nodes[level] = btrfs_lock_root_node(root);
7495                 btrfs_set_lock_blocking(path->nodes[level]);
7496                 path->slots[level] = 0;
7497                 path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
7498                 memset(&wc->update_progress, 0,
7499                        sizeof(wc->update_progress));
7500         } else {
7501                 btrfs_disk_key_to_cpu(&key, &root_item->drop_progress);
7502                 memcpy(&wc->update_progress, &key,
7503                        sizeof(wc->update_progress));
7504
7505                 level = root_item->drop_level;
7506                 BUG_ON(level == 0);
7507                 path->lowest_level = level;
7508                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
7509                 path->lowest_level = 0;
7510                 if (ret < 0) {
7511                         err = ret;
7512                         goto out_end_trans;
7513                 }
7514                 WARN_ON(ret > 0);
7515
7516                 /*
7517                  * unlock our path, this is safe because only this
7518                  * function is allowed to delete this snapshot
7519                  */
7520                 btrfs_unlock_up_safe(path, 0);
7521
7522                 level = btrfs_header_level(root->node);
7523                 while (1) {
7524                         btrfs_tree_lock(path->nodes[level]);
7525                         btrfs_set_lock_blocking(path->nodes[level]);
7526
7527                         ret = btrfs_lookup_extent_info(trans, root,
7528                                                 path->nodes[level]->start,
7529                                                 level, 1, &wc->refs[level],
7530                                                 &wc->flags[level]);
7531                         if (ret < 0) {
7532                                 err = ret;
7533                                 goto out_end_trans;
7534                         }
7535                         BUG_ON(wc->refs[level] == 0);
7536
7537                         if (level == root_item->drop_level)
7538                                 break;
7539
7540                         btrfs_tree_unlock(path->nodes[level]);
7541                         WARN_ON(wc->refs[level] != 1);
7542                         level--;
7543                 }
7544         }
7545
7546         wc->level = level;
7547         wc->shared_level = -1;
7548         wc->stage = DROP_REFERENCE;
7549         wc->update_ref = update_ref;
7550         wc->keep_locks = 0;
7551         wc->for_reloc = for_reloc;
7552         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(root);
7553
7554         while (1) {
7555                 if (!for_reloc && btrfs_need_cleaner_sleep(root)) {
7556                         pr_debug("btrfs: drop snapshot early exit\n");
7557                         err = -EAGAIN;
7558                         goto out_end_trans;
7559                 }
7560
7561                 ret = walk_down_tree(trans, root, path, wc);
7562                 if (ret < 0) {
7563                         err = ret;
7564                         break;
7565                 }
7566
7567                 ret = walk_up_tree(trans, root, path, wc, BTRFS_MAX_LEVEL);
7568                 if (ret < 0) {
7569                         err = ret;
7570                         break;
7571                 }
7572
7573                 if (ret > 0) {
7574                         BUG_ON(wc->stage != DROP_REFERENCE);
7575                         break;
7576                 }
7577
7578                 if (wc->stage == DROP_REFERENCE) {
7579                         level = wc->level;
7580                         btrfs_node_key(path->nodes[level],
7581                                        &root_item->drop_progress,
7582                                        path->slots[level]);
7583                         root_item->drop_level = level;
7584                 }
7585
7586                 BUG_ON(wc->level == 0);
7587                 if (btrfs_should_end_transaction(trans, tree_root)) {
7588                         ret = btrfs_update_root(trans, tree_root,
7589                                                 &root->root_key,
7590                                                 root_item);
7591                         if (ret) {
7592                                 btrfs_abort_transaction(trans, tree_root, ret);
7593                                 err = ret;
7594                                 goto out_end_trans;
7595                         }
7596
7597                         btrfs_end_transaction_throttle(trans, tree_root);
7598                         trans = btrfs_start_transaction(tree_root, 0);
7599                         if (IS_ERR(trans)) {
7600                                 err = PTR_ERR(trans);
7601                                 goto out_free;
7602                         }
7603                         if (block_rsv)
7604                                 trans->block_rsv = block_rsv;
7605                 }
7606         }
7607         btrfs_release_path(path);
7608         if (err)
7609                 goto out_end_trans;
7610
7611         ret = btrfs_del_root(trans, tree_root, &root->root_key);
7612         if (ret) {
7613                 btrfs_abort_transaction(trans, tree_root, ret);
7614                 goto out_end_trans;
7615         }
7616
7617         if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) {
7618                 ret = btrfs_find_root(tree_root, &root->root_key, path,
7619                                       NULL, NULL);
7620                 if (ret < 0) {
7621                         btrfs_abort_transaction(trans, tree_root, ret);
7622                         err = ret;
7623                         goto out_end_trans;
7624                 } else if (ret > 0) {
7625                         /* if we fail to delete the orphan item this time
7626                          * around, it'll get picked up the next time.
7627                          *
7628                          * The most common failure here is just -ENOENT.
7629                          */
7630                         btrfs_del_orphan_item(trans, tree_root,
7631                                               root->root_key.objectid);
7632                 }
7633         }
7634
7635         if (root->in_radix) {
7636                 btrfs_drop_and_free_fs_root(tree_root->fs_info, root);
7637         } else {
7638                 free_extent_buffer(root->node);
7639                 free_extent_buffer(root->commit_root);
7640                 btrfs_put_fs_root(root);
7641         }
7642 out_end_trans:
7643         btrfs_end_transaction_throttle(trans, tree_root);
7644 out_free:
7645         kfree(wc);
7646         btrfs_free_path(path);
7647 out:
7648         if (err)
7649                 btrfs_std_error(root->fs_info, err);
7650         return err;
7651 }
7652
7653 /*
7654  * drop subtree rooted at tree block 'node'.
7655  *
7656  * NOTE: this function will unlock and release tree block 'node'
7657  * only used by relocation code
7658  */
7659 int btrfs_drop_subtree(struct btrfs_trans_handle *trans,
7660                         struct btrfs_root *root,
7661                         struct extent_buffer *node,
7662                         struct extent_buffer *parent)
7663 {
7664         struct btrfs_path *path;
7665         struct walk_control *wc;
7666         int level;
7667         int parent_level;
7668         int ret = 0;
7669         int wret;
7670
7671         BUG_ON(root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID);
7672
7673         path = btrfs_alloc_path();
7674         if (!path)
7675                 return -ENOMEM;
7676
7677         wc = kzalloc(sizeof(*wc), GFP_NOFS);
7678         if (!wc) {
7679                 btrfs_free_path(path);
7680                 return -ENOMEM;
7681         }
7682
7683         btrfs_assert_tree_locked(parent);
7684         parent_level = btrfs_header_level(parent);
7685         extent_buffer_get(parent);
7686         path->nodes[parent_level] = parent;
7687         path->slots[parent_level] = btrfs_header_nritems(parent);
7688
7689         btrfs_assert_tree_locked(node);
7690         level = btrfs_header_level(node);
7691         path->nodes[level] = node;
7692         path->slots[level] = 0;
7693         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
7694
7695         wc->refs[parent_level] = 1;
7696         wc->flags[parent_level] = BTRFS_BLOCK_FLAG_FULL_BACKREF;
7697         wc->level = level;
7698         wc->shared_level = -1;
7699         wc->stage = DROP_REFERENCE;
7700         wc->update_ref = 0;
7701         wc->keep_locks = 1;
7702         wc->for_reloc = 1;
7703         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(root);
7704
7705         while (1) {
7706                 wret = walk_down_tree(trans, root, path, wc);
7707                 if (wret < 0) {
7708                         ret = wret;
7709                         break;
7710                 }
7711
7712                 wret = walk_up_tree(trans, root, path, wc, parent_level);
7713                 if (wret < 0)
7714                         ret = wret;
7715                 if (wret != 0)
7716                         break;
7717         }
7718
7719         kfree(wc);
7720         btrfs_free_path(path);
7721         return ret;
7722 }
7723
7724 static u64 update_block_group_flags(struct btrfs_root *root, u64 flags)
7725 {
7726         u64 num_devices;
7727         u64 stripped;
7728
7729         /*
7730          * if restripe for this chunk_type is on pick target profile and
7731          * return, otherwise do the usual balance
7732          */
7733         stripped = get_restripe_target(root->fs_info, flags);
7734         if (stripped)
7735                 return extended_to_chunk(stripped);
7736
7737         /*
7738          * we add in the count of missing devices because we want
7739          * to make sure that any RAID levels on a degraded FS
7740          * continue to be honored.
7741          */
7742         num_devices = root->fs_info->fs_devices->rw_devices +
7743                 root->fs_info->fs_devices->missing_devices;
7744
7745         stripped = BTRFS_BLOCK_GROUP_RAID0 |
7746                 BTRFS_BLOCK_GROUP_RAID5 | BTRFS_BLOCK_GROUP_RAID6 |
7747                 BTRFS_BLOCK_GROUP_RAID1 | BTRFS_BLOCK_GROUP_RAID10;
7748
7749         if (num_devices == 1) {
7750                 stripped |= BTRFS_BLOCK_GROUP_DUP;
7751                 stripped = flags & ~stripped;
7752
7753                 /* turn raid0 into single device chunks */
7754                 if (flags & BTRFS_BLOCK_GROUP_RAID0)
7755                         return stripped;
7756
7757                 /* turn mirroring into duplication */
7758                 if (flags & (BTRFS_BLOCK_GROUP_RAID1 |
7759                              BTRFS_BLOCK_GROUP_RAID10))
7760                         return stripped | BTRFS_BLOCK_GROUP_DUP;
7761         } else {
7762                 /* they already had raid on here, just return */
7763                 if (flags & stripped)
7764                         return flags;
7765
7766                 stripped |= BTRFS_BLOCK_GROUP_DUP;
7767                 stripped = flags & ~stripped;
7768
7769                 /* switch duplicated blocks with raid1 */
7770                 if (flags & BTRFS_BLOCK_GROUP_DUP)
7771                         return stripped | BTRFS_BLOCK_GROUP_RAID1;
7772
7773                 /* this is drive concat, leave it alone */
7774         }
7775
7776         return flags;
7777 }
7778
7779 static int set_block_group_ro(struct btrfs_block_group_cache *cache, int force)
7780 {
7781         struct btrfs_space_info *sinfo = cache->space_info;
7782         u64 num_bytes;
7783         u64 min_allocable_bytes;
7784         int ret = -ENOSPC;
7785
7786
7787         /*
7788          * We need some metadata space and system metadata space for
7789          * allocating chunks in some corner cases until we force to set
7790          * it to be readonly.
7791          */
7792         if ((sinfo->flags &
7793              (BTRFS_BLOCK_GROUP_SYSTEM | BTRFS_BLOCK_GROUP_METADATA)) &&
7794             !force)
7795                 min_allocable_bytes = 1 * 1024 * 1024;
7796         else
7797                 min_allocable_bytes = 0;
7798
7799         spin_lock(&sinfo->lock);
7800         spin_lock(&cache->lock);
7801
7802         if (cache->ro) {
7803                 ret = 0;
7804                 goto out;
7805         }
7806
7807         num_bytes = cache->key.offset - cache->reserved - cache->pinned -
7808                     cache->bytes_super - btrfs_block_group_used(&cache->item);
7809
7810         if (sinfo->bytes_used + sinfo->bytes_reserved + sinfo->bytes_pinned +
7811             sinfo->bytes_may_use + sinfo->bytes_readonly + num_bytes +
7812             min_allocable_bytes <= sinfo->total_bytes) {
7813                 sinfo->bytes_readonly += num_bytes;
7814                 cache->ro = 1;
7815                 ret = 0;
7816         }
7817 out:
7818         spin_unlock(&cache->lock);
7819         spin_unlock(&sinfo->lock);
7820         return ret;
7821 }
7822
7823 int btrfs_set_block_group_ro(struct btrfs_root *root,
7824                              struct btrfs_block_group_cache *cache)
7825
7826 {
7827         struct btrfs_trans_handle *trans;
7828         u64 alloc_flags;
7829         int ret;
7830
7831         BUG_ON(cache->ro);
7832
7833         trans = btrfs_join_transaction(root);
7834         if (IS_ERR(trans))
7835                 return PTR_ERR(trans);
7836
7837         alloc_flags = update_block_group_flags(root, cache->flags);
7838         if (alloc_flags != cache->flags) {
7839                 ret = do_chunk_alloc(trans, root, alloc_flags,
7840                                      CHUNK_ALLOC_FORCE);
7841                 if (ret < 0)
7842                         goto out;
7843         }
7844
7845         ret = set_block_group_ro(cache, 0);
7846         if (!ret)
7847                 goto out;
7848         alloc_flags = get_alloc_profile(root, cache->space_info->flags);
7849         ret = do_chunk_alloc(trans, root, alloc_flags,
7850                              CHUNK_ALLOC_FORCE);
7851         if (ret < 0)
7852                 goto out;
7853         ret = set_block_group_ro(cache, 0);
7854 out:
7855         btrfs_end_transaction(trans, root);
7856         return ret;
7857 }
7858
7859 int btrfs_force_chunk_alloc(struct btrfs_trans_handle *trans,
7860                             struct btrfs_root *root, u64 type)
7861 {
7862         u64 alloc_flags = get_alloc_profile(root, type);
7863         return do_chunk_alloc(trans, root, alloc_flags,
7864                               CHUNK_ALLOC_FORCE);
7865 }
7866
7867 /*
7868  * helper to account the unused space of all the readonly block group in the
7869  * list. takes mirrors into account.
7870  */
7871 static u64 __btrfs_get_ro_block_group_free_space(struct list_head *groups_list)
7872 {
7873         struct btrfs_block_group_cache *block_group;
7874         u64 free_bytes = 0;
7875         int factor;
7876
7877         list_for_each_entry(block_group, groups_list, list) {
7878                 spin_lock(&block_group->lock);
7879
7880                 if (!block_group->ro) {
7881                         spin_unlock(&block_group->lock);
7882                         continue;
7883                 }
7884
7885                 if (block_group->flags & (BTRFS_BLOCK_GROUP_RAID1 |
7886                                           BTRFS_BLOCK_GROUP_RAID10 |
7887                                           BTRFS_BLOCK_GROUP_DUP))
7888                         factor = 2;
7889                 else
7890                         factor = 1;
7891
7892                 free_bytes += (block_group->key.offset -
7893                                btrfs_block_group_used(&block_group->item)) *
7894                                factor;
7895
7896                 spin_unlock(&block_group->lock);
7897         }
7898
7899         return free_bytes;
7900 }
7901
7902 /*
7903  * helper to account the unused space of all the readonly block group in the
7904  * space_info. takes mirrors into account.
7905  */
7906 u64 btrfs_account_ro_block_groups_free_space(struct btrfs_space_info *sinfo)
7907 {
7908         int i;
7909         u64 free_bytes = 0;
7910
7911         spin_lock(&sinfo->lock);
7912
7913         for(i = 0; i < BTRFS_NR_RAID_TYPES; i++)
7914                 if (!list_empty(&sinfo->block_groups[i]))
7915                         free_bytes += __btrfs_get_ro_block_group_free_space(
7916                                                 &sinfo->block_groups[i]);
7917
7918         spin_unlock(&sinfo->lock);
7919
7920         return free_bytes;
7921 }
7922
7923 void btrfs_set_block_group_rw(struct btrfs_root *root,
7924                               struct btrfs_block_group_cache *cache)
7925 {
7926         struct btrfs_space_info *sinfo = cache->space_info;
7927         u64 num_bytes;
7928
7929         BUG_ON(!cache->ro);
7930
7931         spin_lock(&sinfo->lock);
7932         spin_lock(&cache->lock);
7933         num_bytes = cache->key.offset - cache->reserved - cache->pinned -
7934                     cache->bytes_super - btrfs_block_group_used(&cache->item);
7935         sinfo->bytes_readonly -= num_bytes;
7936         cache->ro = 0;
7937         spin_unlock(&cache->lock);
7938         spin_unlock(&sinfo->lock);
7939 }
7940
7941 /*
7942  * checks to see if its even possible to relocate this block group.
7943  *
7944  * @return - -1 if it's not a good idea to relocate this block group, 0 if its
7945  * ok to go ahead and try.
7946  */
7947 int btrfs_can_relocate(struct btrfs_root *root, u64 bytenr)
7948 {
7949         struct btrfs_block_group_cache *block_group;
7950         struct btrfs_space_info *space_info;
7951         struct btrfs_fs_devices *fs_devices = root->fs_info->fs_devices;
7952         struct btrfs_device *device;
7953         u64 min_free;
7954         u64 dev_min = 1;
7955         u64 dev_nr = 0;
7956         u64 target;
7957         int index;
7958         int full = 0;
7959         int ret = 0;
7960
7961         block_group = btrfs_lookup_block_group(root->fs_info, bytenr);
7962
7963         /* odd, couldn't find the block group, leave it alone */
7964         if (!block_group)
7965                 return -1;
7966
7967         min_free = btrfs_block_group_used(&block_group->item);
7968
7969         /* no bytes used, we're good */
7970         if (!min_free)
7971                 goto out;
7972
7973         space_info = block_group->space_info;
7974         spin_lock(&space_info->lock);
7975
7976         full = space_info->full;
7977
7978         /*
7979          * if this is the last block group we have in this space, we can't
7980          * relocate it unless we're able to allocate a new chunk below.
7981          *
7982          * Otherwise, we need to make sure we have room in the space to handle
7983          * all of the extents from this block group.  If we can, we're good
7984          */
7985         if ((space_info->total_bytes != block_group->key.offset) &&
7986             (space_info->bytes_used + space_info->bytes_reserved +
7987              space_info->bytes_pinned + space_info->bytes_readonly +
7988              min_free < space_info->total_bytes)) {
7989                 spin_unlock(&space_info->lock);
7990                 goto out;
7991         }
7992         spin_unlock(&space_info->lock);
7993
7994         /*
7995          * ok we don't have enough space, but maybe we have free space on our
7996          * devices to allocate new chunks for relocation, so loop through our
7997          * alloc devices and guess if we have enough space.  if this block
7998          * group is going to be restriped, run checks against the target
7999          * profile instead of the current one.
8000          */
8001         ret = -1;
8002
8003         /*
8004          * index:
8005          *      0: raid10
8006          *      1: raid1
8007          *      2: dup
8008          *      3: raid0
8009          *      4: single
8010          */
8011         target = get_restripe_target(root->fs_info, block_group->flags);
8012         if (target) {
8013                 index = __get_raid_index(extended_to_chunk(target));
8014         } else {
8015                 /*
8016                  * this is just a balance, so if we were marked as full
8017                  * we know there is no space for a new chunk
8018                  */
8019                 if (full)
8020                         goto out;
8021
8022                 index = get_block_group_index(block_group);
8023         }
8024
8025         if (index == BTRFS_RAID_RAID10) {
8026                 dev_min = 4;
8027                 /* Divide by 2 */
8028                 min_free >>= 1;
8029         } else if (index == BTRFS_RAID_RAID1) {
8030                 dev_min = 2;
8031         } else if (index == BTRFS_RAID_DUP) {
8032                 /* Multiply by 2 */
8033                 min_free <<= 1;
8034         } else if (index == BTRFS_RAID_RAID0) {
8035                 dev_min = fs_devices->rw_devices;
8036                 do_div(min_free, dev_min);
8037         }
8038
8039         mutex_lock(&root->fs_info->chunk_mutex);
8040         list_for_each_entry(device, &fs_devices->alloc_list, dev_alloc_list) {
8041                 u64 dev_offset;
8042
8043                 /*
8044                  * check to make sure we can actually find a chunk with enough
8045                  * space to fit our block group in.
8046                  */
8047                 if (device->total_bytes > device->bytes_used + min_free &&
8048                     !device->is_tgtdev_for_dev_replace) {
8049                         ret = find_free_dev_extent(device, min_free,
8050                                                    &dev_offset, NULL);
8051                         if (!ret)
8052                                 dev_nr++;
8053
8054                         if (dev_nr >= dev_min)
8055                                 break;
8056
8057                         ret = -1;
8058                 }
8059         }
8060         mutex_unlock(&root->fs_info->chunk_mutex);
8061 out:
8062         btrfs_put_block_group(block_group);
8063         return ret;
8064 }
8065
8066 static int find_first_block_group(struct btrfs_root *root,
8067                 struct btrfs_path *path, struct btrfs_key *key)
8068 {
8069         int ret = 0;
8070         struct btrfs_key found_key;
8071         struct extent_buffer *leaf;
8072         int slot;
8073
8074         ret = btrfs_search_slot(NULL, root, key, path, 0, 0);
8075         if (ret < 0)
8076                 goto out;
8077
8078         while (1) {
8079                 slot = path->slots[0];
8080                 leaf = path->nodes[0];
8081                 if (slot >= btrfs_header_nritems(leaf)) {
8082                         ret = btrfs_next_leaf(root, path);
8083                         if (ret == 0)
8084                                 continue;
8085                         if (ret < 0)
8086                                 goto out;
8087                         break;
8088                 }
8089                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
8090
8091                 if (found_key.objectid >= key->objectid &&
8092                     found_key.type == BTRFS_BLOCK_GROUP_ITEM_KEY) {
8093                         ret = 0;
8094                         goto out;
8095                 }
8096                 path->slots[0]++;
8097         }
8098 out:
8099         return ret;
8100 }
8101
8102 void btrfs_put_block_group_cache(struct btrfs_fs_info *info)
8103 {
8104         struct btrfs_block_group_cache *block_group;
8105         u64 last = 0;
8106
8107         while (1) {
8108                 struct inode *inode;
8109
8110                 block_group = btrfs_lookup_first_block_group(info, last);
8111                 while (block_group) {
8112                         spin_lock(&block_group->lock);
8113                         if (block_group->iref)
8114                                 break;
8115                         spin_unlock(&block_group->lock);
8116                         block_group = next_block_group(info->tree_root,
8117                                                        block_group);
8118                 }
8119                 if (!block_group) {
8120                         if (last == 0)
8121                                 break;
8122                         last = 0;
8123                         continue;
8124                 }
8125
8126                 inode = block_group->inode;
8127                 block_group->iref = 0;
8128                 block_group->inode = NULL;
8129                 spin_unlock(&block_group->lock);
8130                 iput(inode);
8131                 last = block_group->key.objectid + block_group->key.offset;
8132                 btrfs_put_block_group(block_group);
8133         }
8134 }
8135
8136 int btrfs_free_block_groups(struct btrfs_fs_info *info)
8137 {
8138         struct btrfs_block_group_cache *block_group;
8139         struct btrfs_space_info *space_info;
8140         struct btrfs_caching_control *caching_ctl;
8141         struct rb_node *n;
8142
8143         down_write(&info->extent_commit_sem);
8144         while (!list_empty(&info->caching_block_groups)) {
8145                 caching_ctl = list_entry(info->caching_block_groups.next,
8146                                          struct btrfs_caching_control, list);
8147                 list_del(&caching_ctl->list);
8148                 put_caching_control(caching_ctl);
8149         }
8150         up_write(&info->extent_commit_sem);
8151
8152         spin_lock(&info->block_group_cache_lock);
8153         while ((n = rb_last(&info->block_group_cache_tree)) != NULL) {
8154                 block_group = rb_entry(n, struct btrfs_block_group_cache,
8155                                        cache_node);
8156                 rb_erase(&block_group->cache_node,
8157                          &info->block_group_cache_tree);
8158                 spin_unlock(&info->block_group_cache_lock);
8159
8160                 down_write(&block_group->space_info->groups_sem);
8161                 list_del(&block_group->list);
8162                 up_write(&block_group->space_info->groups_sem);
8163
8164                 if (block_group->cached == BTRFS_CACHE_STARTED)
8165                         wait_block_group_cache_done(block_group);
8166
8167                 /*
8168                  * We haven't cached this block group, which means we could
8169                  * possibly have excluded extents on this block group.
8170                  */
8171                 if (block_group->cached == BTRFS_CACHE_NO)
8172                         free_excluded_extents(info->extent_root, block_group);
8173
8174                 btrfs_remove_free_space_cache(block_group);
8175                 btrfs_put_block_group(block_group);
8176
8177                 spin_lock(&info->block_group_cache_lock);
8178         }
8179         spin_unlock(&info->block_group_cache_lock);
8180
8181         /* now that all the block groups are freed, go through and
8182          * free all the space_info structs.  This is only called during
8183          * the final stages of unmount, and so we know nobody is
8184          * using them.  We call synchronize_rcu() once before we start,
8185          * just to be on the safe side.
8186          */
8187         synchronize_rcu();
8188
8189         release_global_block_rsv(info);
8190
8191         while(!list_empty(&info->space_info)) {
8192                 space_info = list_entry(info->space_info.next,
8193                                         struct btrfs_space_info,
8194                                         list);
8195                 if (btrfs_test_opt(info->tree_root, ENOSPC_DEBUG)) {
8196                         if (space_info->bytes_pinned > 0 ||
8197                             space_info->bytes_reserved > 0 ||
8198                             space_info->bytes_may_use > 0) {
8199                                 WARN_ON(1);
8200                                 dump_space_info(space_info, 0, 0);
8201                         }
8202                 }
8203                 percpu_counter_destroy(&space_info->total_bytes_pinned);
8204                 list_del(&space_info->list);
8205                 kfree(space_info);
8206         }
8207         return 0;
8208 }
8209
8210 static void __link_block_group(struct btrfs_space_info *space_info,
8211                                struct btrfs_block_group_cache *cache)
8212 {
8213         int index = get_block_group_index(cache);
8214
8215         down_write(&space_info->groups_sem);
8216         list_add_tail(&cache->list, &space_info->block_groups[index]);
8217         up_write(&space_info->groups_sem);
8218 }
8219
8220 int btrfs_read_block_groups(struct btrfs_root *root)
8221 {
8222         struct btrfs_path *path;
8223         int ret;
8224         struct btrfs_block_group_cache *cache;
8225         struct btrfs_fs_info *info = root->fs_info;
8226         struct btrfs_space_info *space_info;
8227         struct btrfs_key key;
8228         struct btrfs_key found_key;
8229         struct extent_buffer *leaf;
8230         int need_clear = 0;
8231         u64 cache_gen;
8232
8233         root = info->extent_root;
8234         key.objectid = 0;
8235         key.offset = 0;
8236         btrfs_set_key_type(&key, BTRFS_BLOCK_GROUP_ITEM_KEY);
8237         path = btrfs_alloc_path();
8238         if (!path)
8239                 return -ENOMEM;
8240         path->reada = 1;
8241
8242         cache_gen = btrfs_super_cache_generation(root->fs_info->super_copy);
8243         if (btrfs_test_opt(root, SPACE_CACHE) &&
8244             btrfs_super_generation(root->fs_info->super_copy) != cache_gen)
8245                 need_clear = 1;
8246         if (btrfs_test_opt(root, CLEAR_CACHE))
8247                 need_clear = 1;
8248
8249         while (1) {
8250                 ret = find_first_block_group(root, path, &key);
8251                 if (ret > 0)
8252                         break;
8253                 if (ret != 0)
8254                         goto error;
8255                 leaf = path->nodes[0];
8256                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
8257                 cache = kzalloc(sizeof(*cache), GFP_NOFS);
8258                 if (!cache) {
8259                         ret = -ENOMEM;
8260                         goto error;
8261                 }
8262                 cache->free_space_ctl = kzalloc(sizeof(*cache->free_space_ctl),
8263                                                 GFP_NOFS);
8264                 if (!cache->free_space_ctl) {
8265                         kfree(cache);
8266                         ret = -ENOMEM;
8267                         goto error;
8268                 }
8269
8270                 atomic_set(&cache->count, 1);
8271                 spin_lock_init(&cache->lock);
8272                 cache->fs_info = info;
8273                 INIT_LIST_HEAD(&cache->list);
8274                 INIT_LIST_HEAD(&cache->cluster_list);
8275
8276                 if (need_clear) {
8277                         /*
8278                          * When we mount with old space cache, we need to
8279                          * set BTRFS_DC_CLEAR and set dirty flag.
8280                          *
8281                          * a) Setting 'BTRFS_DC_CLEAR' makes sure that we
8282                          *    truncate the old free space cache inode and
8283                          *    setup a new one.
8284                          * b) Setting 'dirty flag' makes sure that we flush
8285                          *    the new space cache info onto disk.
8286                          */
8287                         cache->disk_cache_state = BTRFS_DC_CLEAR;
8288                         if (btrfs_test_opt(root, SPACE_CACHE))
8289                                 cache->dirty = 1;
8290                 }
8291
8292                 read_extent_buffer(leaf, &cache->item,
8293                                    btrfs_item_ptr_offset(leaf, path->slots[0]),
8294                                    sizeof(cache->item));
8295                 memcpy(&cache->key, &found_key, sizeof(found_key));
8296
8297                 key.objectid = found_key.objectid + found_key.offset;
8298                 btrfs_release_path(path);
8299                 cache->flags = btrfs_block_group_flags(&cache->item);
8300                 cache->sectorsize = root->sectorsize;
8301                 cache->full_stripe_len = btrfs_full_stripe_len(root,
8302                                                &root->fs_info->mapping_tree,
8303                                                found_key.objectid);
8304                 btrfs_init_free_space_ctl(cache);
8305
8306                 /*
8307                  * We need to exclude the super stripes now so that the space
8308                  * info has super bytes accounted for, otherwise we'll think
8309                  * we have more space than we actually do.
8310                  */
8311                 ret = exclude_super_stripes(root, cache);
8312                 if (ret) {
8313                         /*
8314                          * We may have excluded something, so call this just in
8315                          * case.
8316                          */
8317                         free_excluded_extents(root, cache);
8318                         kfree(cache->free_space_ctl);
8319                         kfree(cache);
8320                         goto error;
8321                 }
8322
8323                 /*
8324                  * check for two cases, either we are full, and therefore
8325                  * don't need to bother with the caching work since we won't
8326                  * find any space, or we are empty, and we can just add all
8327                  * the space in and be done with it.  This saves us _alot_ of
8328                  * time, particularly in the full case.
8329                  */
8330                 if (found_key.offset == btrfs_block_group_used(&cache->item)) {
8331                         cache->last_byte_to_unpin = (u64)-1;
8332                         cache->cached = BTRFS_CACHE_FINISHED;
8333                         free_excluded_extents(root, cache);
8334                 } else if (btrfs_block_group_used(&cache->item) == 0) {
8335                         cache->last_byte_to_unpin = (u64)-1;
8336                         cache->cached = BTRFS_CACHE_FINISHED;
8337                         add_new_free_space(cache, root->fs_info,
8338                                            found_key.objectid,
8339                                            found_key.objectid +
8340                                            found_key.offset);
8341                         free_excluded_extents(root, cache);
8342                 }
8343
8344                 ret = btrfs_add_block_group_cache(root->fs_info, cache);
8345                 if (ret) {
8346                         btrfs_remove_free_space_cache(cache);
8347                         btrfs_put_block_group(cache);
8348                         goto error;
8349                 }
8350
8351                 ret = update_space_info(info, cache->flags, found_key.offset,
8352                                         btrfs_block_group_used(&cache->item),
8353                                         &space_info);
8354                 if (ret) {
8355                         btrfs_remove_free_space_cache(cache);
8356                         spin_lock(&info->block_group_cache_lock);
8357                         rb_erase(&cache->cache_node,
8358                                  &info->block_group_cache_tree);
8359                         spin_unlock(&info->block_group_cache_lock);
8360                         btrfs_put_block_group(cache);
8361                         goto error;
8362                 }
8363
8364                 cache->space_info = space_info;
8365                 spin_lock(&cache->space_info->lock);
8366                 cache->space_info->bytes_readonly += cache->bytes_super;
8367                 spin_unlock(&cache->space_info->lock);
8368
8369                 __link_block_group(space_info, cache);
8370
8371                 set_avail_alloc_bits(root->fs_info, cache->flags);
8372                 if (btrfs_chunk_readonly(root, cache->key.objectid))
8373                         set_block_group_ro(cache, 1);
8374         }
8375
8376         list_for_each_entry_rcu(space_info, &root->fs_info->space_info, list) {
8377                 if (!(get_alloc_profile(root, space_info->flags) &
8378                       (BTRFS_BLOCK_GROUP_RAID10 |
8379                        BTRFS_BLOCK_GROUP_RAID1 |
8380                        BTRFS_BLOCK_GROUP_RAID5 |
8381                        BTRFS_BLOCK_GROUP_RAID6 |
8382                        BTRFS_BLOCK_GROUP_DUP)))
8383                         continue;
8384                 /*
8385                  * avoid allocating from un-mirrored block group if there are
8386                  * mirrored block groups.
8387                  */
8388                 list_for_each_entry(cache, &space_info->block_groups[3], list)
8389                         set_block_group_ro(cache, 1);
8390                 list_for_each_entry(cache, &space_info->block_groups[4], list)
8391                         set_block_group_ro(cache, 1);
8392         }
8393
8394         init_global_block_rsv(info);
8395         ret = 0;
8396 error:
8397         btrfs_free_path(path);
8398         return ret;
8399 }
8400
8401 void btrfs_create_pending_block_groups(struct btrfs_trans_handle *trans,
8402                                        struct btrfs_root *root)
8403 {
8404         struct btrfs_block_group_cache *block_group, *tmp;
8405         struct btrfs_root *extent_root = root->fs_info->extent_root;
8406         struct btrfs_block_group_item item;
8407         struct btrfs_key key;
8408         int ret = 0;
8409
8410         list_for_each_entry_safe(block_group, tmp, &trans->new_bgs,
8411                                  new_bg_list) {
8412                 list_del_init(&block_group->new_bg_list);
8413
8414                 if (ret)
8415                         continue;
8416
8417                 spin_lock(&block_group->lock);
8418                 memcpy(&item, &block_group->item, sizeof(item));
8419                 memcpy(&key, &block_group->key, sizeof(key));
8420                 spin_unlock(&block_group->lock);
8421
8422                 ret = btrfs_insert_item(trans, extent_root, &key, &item,
8423                                         sizeof(item));
8424                 if (ret)
8425                         btrfs_abort_transaction(trans, extent_root, ret);
8426         }
8427 }
8428
8429 int btrfs_make_block_group(struct btrfs_trans_handle *trans,
8430                            struct btrfs_root *root, u64 bytes_used,
8431                            u64 type, u64 chunk_objectid, u64 chunk_offset,
8432                            u64 size)
8433 {
8434         int ret;
8435         struct btrfs_root *extent_root;
8436         struct btrfs_block_group_cache *cache;
8437
8438         extent_root = root->fs_info->extent_root;
8439
8440         root->fs_info->last_trans_log_full_commit = trans->transid;
8441
8442         cache = kzalloc(sizeof(*cache), GFP_NOFS);
8443         if (!cache)
8444                 return -ENOMEM;
8445         cache->free_space_ctl = kzalloc(sizeof(*cache->free_space_ctl),
8446                                         GFP_NOFS);
8447         if (!cache->free_space_ctl) {
8448                 kfree(cache);
8449                 return -ENOMEM;
8450         }
8451
8452         cache->key.objectid = chunk_offset;
8453         cache->key.offset = size;
8454         cache->key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
8455         cache->sectorsize = root->sectorsize;
8456         cache->fs_info = root->fs_info;
8457         cache->full_stripe_len = btrfs_full_stripe_len(root,
8458                                                &root->fs_info->mapping_tree,
8459                                                chunk_offset);
8460
8461         atomic_set(&cache->count, 1);
8462         spin_lock_init(&cache->lock);
8463         INIT_LIST_HEAD(&cache->list);
8464         INIT_LIST_HEAD(&cache->cluster_list);
8465         INIT_LIST_HEAD(&cache->new_bg_list);
8466
8467         btrfs_init_free_space_ctl(cache);
8468
8469         btrfs_set_block_group_used(&cache->item, bytes_used);
8470         btrfs_set_block_group_chunk_objectid(&cache->item, chunk_objectid);
8471         cache->flags = type;
8472         btrfs_set_block_group_flags(&cache->item, type);
8473
8474         cache->last_byte_to_unpin = (u64)-1;
8475         cache->cached = BTRFS_CACHE_FINISHED;
8476         ret = exclude_super_stripes(root, cache);
8477         if (ret) {
8478                 /*
8479                  * We may have excluded something, so call this just in
8480                  * case.
8481                  */
8482                 free_excluded_extents(root, cache);
8483                 kfree(cache->free_space_ctl);
8484                 kfree(cache);
8485                 return ret;
8486         }
8487
8488         add_new_free_space(cache, root->fs_info, chunk_offset,
8489                            chunk_offset + size);
8490
8491         free_excluded_extents(root, cache);
8492
8493         ret = btrfs_add_block_group_cache(root->fs_info, cache);
8494         if (ret) {
8495                 btrfs_remove_free_space_cache(cache);
8496                 btrfs_put_block_group(cache);
8497                 return ret;
8498         }
8499
8500         ret = update_space_info(root->fs_info, cache->flags, size, bytes_used,
8501                                 &cache->space_info);
8502         if (ret) {
8503                 btrfs_remove_free_space_cache(cache);
8504                 spin_lock(&root->fs_info->block_group_cache_lock);
8505                 rb_erase(&cache->cache_node,
8506                          &root->fs_info->block_group_cache_tree);
8507                 spin_unlock(&root->fs_info->block_group_cache_lock);
8508                 btrfs_put_block_group(cache);
8509                 return ret;
8510         }
8511         update_global_block_rsv(root->fs_info);
8512
8513         spin_lock(&cache->space_info->lock);
8514         cache->space_info->bytes_readonly += cache->bytes_super;
8515         spin_unlock(&cache->space_info->lock);
8516
8517         __link_block_group(cache->space_info, cache);
8518
8519         list_add_tail(&cache->new_bg_list, &trans->new_bgs);
8520
8521         set_avail_alloc_bits(extent_root->fs_info, type);
8522
8523         return 0;
8524 }
8525
8526 static void clear_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
8527 {
8528         u64 extra_flags = chunk_to_extended(flags) &
8529                                 BTRFS_EXTENDED_PROFILE_MASK;
8530
8531         write_seqlock(&fs_info->profiles_lock);
8532         if (flags & BTRFS_BLOCK_GROUP_DATA)
8533                 fs_info->avail_data_alloc_bits &= ~extra_flags;
8534         if (flags & BTRFS_BLOCK_GROUP_METADATA)
8535                 fs_info->avail_metadata_alloc_bits &= ~extra_flags;
8536         if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
8537                 fs_info->avail_system_alloc_bits &= ~extra_flags;
8538         write_sequnlock(&fs_info->profiles_lock);
8539 }
8540
8541 int btrfs_remove_block_group(struct btrfs_trans_handle *trans,
8542                              struct btrfs_root *root, u64 group_start)
8543 {
8544         struct btrfs_path *path;
8545         struct btrfs_block_group_cache *block_group;
8546         struct btrfs_free_cluster *cluster;
8547         struct btrfs_root *tree_root = root->fs_info->tree_root;
8548         struct btrfs_key key;
8549         struct inode *inode;
8550         int ret;
8551         int index;
8552         int factor;
8553
8554         root = root->fs_info->extent_root;
8555
8556         block_group = btrfs_lookup_block_group(root->fs_info, group_start);
8557         BUG_ON(!block_group);
8558         BUG_ON(!block_group->ro);
8559
8560         /*
8561          * Free the reserved super bytes from this block group before
8562          * remove it.
8563          */
8564         free_excluded_extents(root, block_group);
8565
8566         memcpy(&key, &block_group->key, sizeof(key));
8567         index = get_block_group_index(block_group);
8568         if (block_group->flags & (BTRFS_BLOCK_GROUP_DUP |
8569                                   BTRFS_BLOCK_GROUP_RAID1 |
8570                                   BTRFS_BLOCK_GROUP_RAID10))
8571                 factor = 2;
8572         else
8573                 factor = 1;
8574
8575         /* make sure this block group isn't part of an allocation cluster */
8576         cluster = &root->fs_info->data_alloc_cluster;
8577         spin_lock(&cluster->refill_lock);
8578         btrfs_return_cluster_to_free_space(block_group, cluster);
8579         spin_unlock(&cluster->refill_lock);
8580
8581         /*
8582          * make sure this block group isn't part of a metadata
8583          * allocation cluster
8584          */
8585         cluster = &root->fs_info->meta_alloc_cluster;
8586         spin_lock(&cluster->refill_lock);
8587         btrfs_return_cluster_to_free_space(block_group, cluster);
8588         spin_unlock(&cluster->refill_lock);
8589
8590         path = btrfs_alloc_path();
8591         if (!path) {
8592                 ret = -ENOMEM;
8593                 goto out;
8594         }
8595
8596         inode = lookup_free_space_inode(tree_root, block_group, path);
8597         if (!IS_ERR(inode)) {
8598                 ret = btrfs_orphan_add(trans, inode);
8599                 if (ret) {
8600                         btrfs_add_delayed_iput(inode);
8601                         goto out;
8602                 }
8603                 clear_nlink(inode);
8604                 /* One for the block groups ref */
8605                 spin_lock(&block_group->lock);
8606                 if (block_group->iref) {
8607                         block_group->iref = 0;
8608                         block_group->inode = NULL;
8609                         spin_unlock(&block_group->lock);
8610                         iput(inode);
8611                 } else {
8612                         spin_unlock(&block_group->lock);
8613                 }
8614                 /* One for our lookup ref */
8615                 btrfs_add_delayed_iput(inode);
8616         }
8617
8618         key.objectid = BTRFS_FREE_SPACE_OBJECTID;
8619         key.offset = block_group->key.objectid;
8620         key.type = 0;
8621
8622         ret = btrfs_search_slot(trans, tree_root, &key, path, -1, 1);
8623         if (ret < 0)
8624                 goto out;
8625         if (ret > 0)
8626                 btrfs_release_path(path);
8627         if (ret == 0) {
8628                 ret = btrfs_del_item(trans, tree_root, path);
8629                 if (ret)
8630                         goto out;
8631                 btrfs_release_path(path);
8632         }
8633
8634         spin_lock(&root->fs_info->block_group_cache_lock);
8635         rb_erase(&block_group->cache_node,
8636                  &root->fs_info->block_group_cache_tree);
8637
8638         if (root->fs_info->first_logical_byte == block_group->key.objectid)
8639                 root->fs_info->first_logical_byte = (u64)-1;
8640         spin_unlock(&root->fs_info->block_group_cache_lock);
8641
8642         down_write(&block_group->space_info->groups_sem);
8643         /*
8644          * we must use list_del_init so people can check to see if they
8645          * are still on the list after taking the semaphore
8646          */
8647         list_del_init(&block_group->list);
8648         if (list_empty(&block_group->space_info->block_groups[index]))
8649                 clear_avail_alloc_bits(root->fs_info, block_group->flags);
8650         up_write(&block_group->space_info->groups_sem);
8651
8652         if (block_group->cached == BTRFS_CACHE_STARTED)
8653                 wait_block_group_cache_done(block_group);
8654
8655         btrfs_remove_free_space_cache(block_group);
8656
8657         spin_lock(&block_group->space_info->lock);
8658         block_group->space_info->total_bytes -= block_group->key.offset;
8659         block_group->space_info->bytes_readonly -= block_group->key.offset;
8660         block_group->space_info->disk_total -= block_group->key.offset * factor;
8661         spin_unlock(&block_group->space_info->lock);
8662
8663         memcpy(&key, &block_group->key, sizeof(key));
8664
8665         btrfs_clear_space_info_full(root->fs_info);
8666
8667         btrfs_put_block_group(block_group);
8668         btrfs_put_block_group(block_group);
8669
8670         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
8671         if (ret > 0)
8672                 ret = -EIO;
8673         if (ret < 0)
8674                 goto out;
8675
8676         ret = btrfs_del_item(trans, root, path);
8677 out:
8678         btrfs_free_path(path);
8679         return ret;
8680 }
8681
8682 int btrfs_init_space_info(struct btrfs_fs_info *fs_info)
8683 {
8684         struct btrfs_space_info *space_info;
8685         struct btrfs_super_block *disk_super;
8686         u64 features;
8687         u64 flags;
8688         int mixed = 0;
8689         int ret;
8690
8691         disk_super = fs_info->super_copy;
8692         if (!btrfs_super_root(disk_super))
8693                 return 1;
8694
8695         features = btrfs_super_incompat_flags(disk_super);
8696         if (features & BTRFS_FEATURE_INCOMPAT_MIXED_GROUPS)
8697                 mixed = 1;
8698
8699         flags = BTRFS_BLOCK_GROUP_SYSTEM;
8700         ret = update_space_info(fs_info, flags, 0, 0, &space_info);
8701         if (ret)
8702                 goto out;
8703
8704         if (mixed) {
8705                 flags = BTRFS_BLOCK_GROUP_METADATA | BTRFS_BLOCK_GROUP_DATA;
8706                 ret = update_space_info(fs_info, flags, 0, 0, &space_info);
8707         } else {
8708                 flags = BTRFS_BLOCK_GROUP_METADATA;
8709                 ret = update_space_info(fs_info, flags, 0, 0, &space_info);
8710                 if (ret)
8711                         goto out;
8712
8713                 flags = BTRFS_BLOCK_GROUP_DATA;
8714                 ret = update_space_info(fs_info, flags, 0, 0, &space_info);
8715         }
8716 out:
8717         return ret;
8718 }
8719
8720 int btrfs_error_unpin_extent_range(struct btrfs_root *root, u64 start, u64 end)
8721 {
8722         return unpin_extent_range(root, start, end);
8723 }
8724
8725 int btrfs_error_discard_extent(struct btrfs_root *root, u64 bytenr,
8726                                u64 num_bytes, u64 *actual_bytes)
8727 {
8728         return btrfs_discard_extent(root, bytenr, num_bytes, actual_bytes);
8729 }
8730
8731 int btrfs_trim_fs(struct btrfs_root *root, struct fstrim_range *range)
8732 {
8733         struct btrfs_fs_info *fs_info = root->fs_info;
8734         struct btrfs_block_group_cache *cache = NULL;
8735         u64 group_trimmed;
8736         u64 start;
8737         u64 end;
8738         u64 trimmed = 0;
8739         u64 total_bytes = btrfs_super_total_bytes(fs_info->super_copy);
8740         int ret = 0;
8741
8742         /*
8743          * try to trim all FS space, our block group may start from non-zero.
8744          */
8745         if (range->len == total_bytes)
8746                 cache = btrfs_lookup_first_block_group(fs_info, range->start);
8747         else
8748                 cache = btrfs_lookup_block_group(fs_info, range->start);
8749
8750         while (cache) {
8751                 if (cache->key.objectid >= (range->start + range->len)) {
8752                         btrfs_put_block_group(cache);
8753                         break;
8754                 }
8755
8756                 start = max(range->start, cache->key.objectid);
8757                 end = min(range->start + range->len,
8758                                 cache->key.objectid + cache->key.offset);
8759
8760                 if (end - start >= range->minlen) {
8761                         if (!block_group_cache_done(cache)) {
8762                                 ret = cache_block_group(cache, 0);
8763                                 if (ret) {
8764                                         btrfs_put_block_group(cache);
8765                                         break;
8766                                 }
8767                                 ret = wait_block_group_cache_done(cache);
8768                                 if (ret) {
8769                                         btrfs_put_block_group(cache);
8770                                         break;
8771                                 }
8772                         }
8773                         ret = btrfs_trim_block_group(cache,
8774                                                      &group_trimmed,
8775                                                      start,
8776                                                      end,
8777                                                      range->minlen);
8778
8779                         trimmed += group_trimmed;
8780                         if (ret) {
8781                                 btrfs_put_block_group(cache);
8782                                 break;
8783                         }
8784                 }
8785
8786                 cache = next_block_group(fs_info->tree_root, cache);
8787         }
8788
8789         range->len = trimmed;
8790         return ret;
8791 }