fs/btrfs/extent-tree.c

   1 /*
   2  * Copyright (C) 2007 Oracle.  All rights reserved.
   3  *
   4  * This program is free software; you can redistribute it and/or
   5  * modify it under the terms of the GNU General Public
   6  * License v2 as published by the Free Software Foundation.
   7  *
   8  * This program is distributed in the hope that it will be useful,
   9  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  10  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  11  * General Public License for more details.
  12  *
  13  * You should have received a copy of the GNU General Public
  14  * License along with this program; if not, write to the
  15  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
  16  * Boston, MA 021110-1307, USA.
  17  */
  18 #include <linux/sched.h>
  19 #include <linux/pagemap.h>
  20 #include <linux/writeback.h>
  21 #include <linux/blkdev.h>
  22 #include <linux/sort.h>
  23 #include <linux/rcupdate.h>
  24 #include <linux/kthread.h>
  25 #include <linux/slab.h>
  26 #include <linux/ratelimit.h>
  27 #include "compat.h"
  28 #include "hash.h"
  29 #include "ctree.h"
  30 #include "disk-io.h"
  31 #include "print-tree.h"
  32 #include "transaction.h"
  33 #include "volumes.h"
  34 #include "raid56.h"
  35 #include "locking.h"
  36 #include "free-space-cache.h"
  37 #include "math.h"
  38
  39 #undef SCRAMBLE_DELAYED_REFS
  40
  41 /*
  42  * control flags for do_chunk_alloc's force field
  43  * CHUNK_ALLOC_NO_FORCE means to only allocate a chunk
  44  * if we really need one.
  45  *
  46  * CHUNK_ALLOC_LIMITED means to only try and allocate one
  47  * if we have very few chunks already allocated.  This is
  48  * used as part of the clustering code to help make sure
  49  * we have a good pool of storage to cluster in, without
  50  * filling the FS with empty chunks
  51  *
  52  * CHUNK_ALLOC_FORCE means it must try to allocate one
  53  *
  54  */
  55 enum {
  56         CHUNK_ALLOC_NO_FORCE = 0,
  57         CHUNK_ALLOC_LIMITED = 1,
  58         CHUNK_ALLOC_FORCE = 2,
  59 };
  60
  61 /*
  62  * Control how reservations are dealt with.
  63  *
  64  * RESERVE_FREE - freeing a reservation.
  65  * RESERVE_ALLOC - allocating space and we need to update bytes_may_use for
  66  *   ENOSPC accounting
  67  * RESERVE_ALLOC_NO_ACCOUNT - allocating space and we should not update
  68  *   bytes_may_use as the ENOSPC accounting is done elsewhere
  69  */
  70 enum {
  71         RESERVE_FREE = 0,
  72         RESERVE_ALLOC = 1,
  73         RESERVE_ALLOC_NO_ACCOUNT = 2,
  74 };
  75
  76 static int update_block_group(struct btrfs_root *root,
  77                               u64 bytenr, u64 num_bytes, int alloc);
  78 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
  79                                 struct btrfs_root *root,
  80                                 u64 bytenr, u64 num_bytes, u64 parent,
  81                                 u64 root_objectid, u64 owner_objectid,
  82                                 u64 owner_offset, int refs_to_drop,
  83                                 struct btrfs_delayed_extent_op *extra_op);
  84 static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op,
  85                                     struct extent_buffer *leaf,
  86                                     struct btrfs_extent_item *ei);
  87 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
  88                                       struct btrfs_root *root,
  89                                       u64 parent, u64 root_objectid,
  90                                       u64 flags, u64 owner, u64 offset,
  91                                       struct btrfs_key *ins, int ref_mod);
  92 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
  93                                      struct btrfs_root *root,
  94                                      u64 parent, u64 root_objectid,
  95                                      u64 flags, struct btrfs_disk_key *key,
  96                                      int level, struct btrfs_key *ins);
  97 static int do_chunk_alloc(struct btrfs_trans_handle *trans,
  98                           struct btrfs_root *extent_root, u64 flags,
  99                           int force);
 100 static int find_next_key(struct btrfs_path *path, int level,
 101                          struct btrfs_key *key);
 102 static void dump_space_info(struct btrfs_space_info *info, u64 bytes,
 103                             int dump_block_groups);
 104 static int btrfs_update_reserved_bytes(struct btrfs_block_group_cache *cache,
 105                                        u64 num_bytes, int reserve);
 106 static int block_rsv_use_bytes(struct btrfs_block_rsv *block_rsv,
 107                                u64 num_bytes);
 108 int btrfs_pin_extent(struct btrfs_root *root,
 109                      u64 bytenr, u64 num_bytes, int reserved);
 110
 111 static noinline int
 112 block_group_cache_done(struct btrfs_block_group_cache *cache)
 113 {
 114         smp_mb();
 115         return cache->cached == BTRFS_CACHE_FINISHED;
 116 }
 117
 118 static int block_group_bits(struct btrfs_block_group_cache *cache, u64 bits)
 119 {
 120         return (cache->flags & bits) == bits;
 121 }
 122
 123 static void btrfs_get_block_group(struct btrfs_block_group_cache *cache)
 124 {
 125         atomic_inc(&cache->count);
 126 }
 127
 128 void btrfs_put_block_group(struct btrfs_block_group_cache *cache)
 129 {
 130         if (atomic_dec_and_test(&cache->count)) {
 131                 WARN_ON(cache->pinned > 0);
 132                 WARN_ON(cache->reserved > 0);
 133                 kfree(cache->free_space_ctl);
 134                 kfree(cache);
 135         }
 136 }
 137
 138 /*
 139  * this adds the block group to the fs_info rb tree for the block group
 140  * cache
 141  */
 142 static int btrfs_add_block_group_cache(struct btrfs_fs_info *info,
 143                                 struct btrfs_block_group_cache *block_group)
 144 {
 145         struct rb_node **p;
 146         struct rb_node *parent = NULL;
 147         struct btrfs_block_group_cache *cache;
 148
 149         spin_lock(&info->block_group_cache_lock);
 150         p = &info->block_group_cache_tree.rb_node;
 151
 152         while (*p) {
 153                 parent = *p;
 154                 cache = rb_entry(parent, struct btrfs_block_group_cache,
 155                                  cache_node);
 156                 if (block_group->key.objectid < cache->key.objectid) {
 157                         p = &(*p)->rb_left;
 158                 } else if (block_group->key.objectid > cache->key.objectid) {
 159                         p = &(*p)->rb_right;
 160                 } else {
 161                         spin_unlock(&info->block_group_cache_lock);
 162                         return -EEXIST;
 163                 }
 164         }
 165
 166         rb_link_node(&block_group->cache_node, parent, p);
 167         rb_insert_color(&block_group->cache_node,
 168                         &info->block_group_cache_tree);
 169
 170         if (info->first_logical_byte > block_group->key.objectid)
 171                 info->first_logical_byte = block_group->key.objectid;
 172
 173         spin_unlock(&info->block_group_cache_lock);
 174
 175         return 0;
 176 }
 177
 178 /*
 179  * This will return the block group at or after bytenr if contains is 0, else
 180  * it will return the block group that contains the bytenr
 181  */
 182 static struct btrfs_block_group_cache *
 183 block_group_cache_tree_search(struct btrfs_fs_info *info, u64 bytenr,
 184                               int contains)
 185 {
 186         struct btrfs_block_group_cache *cache, *ret = NULL;
 187         struct rb_node *n;
 188         u64 end, start;
 189
 190         spin_lock(&info->block_group_cache_lock);
 191         n = info->block_group_cache_tree.rb_node;
 192
 193         while (n) {
 194                 cache = rb_entry(n, struct btrfs_block_group_cache,
 195                                  cache_node);
 196                 end = cache->key.objectid + cache->key.offset - 1;
 197                 start = cache->key.objectid;
 198
 199                 if (bytenr < start) {
 200                         if (!contains && (!ret || start < ret->key.objectid))
 201                                 ret = cache;
 202                         n = n->rb_left;
 203                 } else if (bytenr > start) {
 204                         if (contains && bytenr <= end) {
 205                                 ret = cache;
 206                                 break;
 207                         }
 208                         n = n->rb_right;
 209                 } else {
 210                         ret = cache;
 211                         break;
 212                 }
 213         }
 214         if (ret) {
 215                 btrfs_get_block_group(ret);
 216                 if (bytenr == 0 && info->first_logical_byte > ret->key.objectid)
 217                         info->first_logical_byte = ret->key.objectid;
 218         }
 219         spin_unlock(&info->block_group_cache_lock);
 220
 221         return ret;
 222 }
 223
 224 static int add_excluded_extent(struct btrfs_root *root,
 225                                u64 start, u64 num_bytes)
 226 {
 227         u64 end = start + num_bytes - 1;
 228         set_extent_bits(&root->fs_info->freed_extents[0],
 229                         start, end, EXTENT_UPTODATE, GFP_NOFS);
 230         set_extent_bits(&root->fs_info->freed_extents[1],
 231                         start, end, EXTENT_UPTODATE, GFP_NOFS);
 232         return 0;
 233 }
 234
 235 static void free_excluded_extents(struct btrfs_root *root,
 236                                   struct btrfs_block_group_cache *cache)
 237 {
 238         u64 start, end;
 239
 240         start = cache->key.objectid;
 241         end = start + cache->key.offset - 1;
 242
 243         clear_extent_bits(&root->fs_info->freed_extents[0],
 244                           start, end, EXTENT_UPTODATE, GFP_NOFS);
 245         clear_extent_bits(&root->fs_info->freed_extents[1],
 246                           start, end, EXTENT_UPTODATE, GFP_NOFS);
 247 }
 248
 249 static int exclude_super_stripes(struct btrfs_root *root,
 250                                  struct btrfs_block_group_cache *cache)
 251 {
 252         u64 bytenr;
 253         u64 *logical;
 254         int stripe_len;
 255         int i, nr, ret;
 256
 257         if (cache->key.objectid < BTRFS_SUPER_INFO_OFFSET) {
 258                 stripe_len = BTRFS_SUPER_INFO_OFFSET - cache->key.objectid;
 259                 cache->bytes_super += stripe_len;
 260                 ret = add_excluded_extent(root, cache->key.objectid,
 261                                           stripe_len);
 262                 if (ret)
 263                         return ret;
 264         }
 265
 266         for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
 267                 bytenr = btrfs_sb_offset(i);
 268                 ret = btrfs_rmap_block(&root->fs_info->mapping_tree,
 269                                        cache->key.objectid, bytenr,
 270                                        0, &logical, &nr, &stripe_len);
 271                 if (ret)
 272                         return ret;
 273
 274                 while (nr--) {
 275                         u64 start, len;
 276
 277                         if (logical[nr] > cache->key.objectid +
 278                             cache->key.offset)
 279                                 continue;
 280
 281                         if (logical[nr] + stripe_len <= cache->key.objectid)
 282                                 continue;
 283
 284                         start = logical[nr];
 285                         if (start < cache->key.objectid) {
 286                                 start = cache->key.objectid;
 287                                 len = (logical[nr] + stripe_len) - start;
 288                         } else {
 289                                 len = min_t(u64, stripe_len,
 290                                             cache->key.objectid +
 291                                             cache->key.offset - start);
 292                         }
 293
 294                         cache->bytes_super += len;
 295                         ret = add_excluded_extent(root, start, len);
 296                         if (ret) {
 297                                 kfree(logical);
 298                                 return ret;
 299                         }
 300                 }
 301
 302                 kfree(logical);
 303         }
 304         return 0;
 305 }
 306
 307 static struct btrfs_caching_control *
 308 get_caching_control(struct btrfs_block_group_cache *cache)
 309 {
 310         struct btrfs_caching_control *ctl;
 311
 312         spin_lock(&cache->lock);
 313         if (cache->cached != BTRFS_CACHE_STARTED) {
 314                 spin_unlock(&cache->lock);
 315                 return NULL;
 316         }
 317
 318         /* We're loading it the fast way, so we don't have a caching_ctl. */
 319         if (!cache->caching_ctl) {
 320                 spin_unlock(&cache->lock);
 321                 return NULL;
 322         }
 323
 324         ctl = cache->caching_ctl;
 325         atomic_inc(&ctl->count);
 326         spin_unlock(&cache->lock);
 327         return ctl;
 328 }
 329
 330 static void put_caching_control(struct btrfs_caching_control *ctl)
 331 {
 332         if (atomic_dec_and_test(&ctl->count))
 333                 kfree(ctl);
 334 }
 335
 336 /*
 337  * this is only called by cache_block_group, since we could have freed extents
 338  * we need to check the pinned_extents for any extents that can't be used yet
 339  * since their free space will be released as soon as the transaction commits.
 340  */
 341 static u64 add_new_free_space(struct btrfs_block_group_cache *block_group,
 342                               struct btrfs_fs_info *info, u64 start, u64 end)
 343 {
 344         u64 extent_start, extent_end, size, total_added = 0;
 345         int ret;
 346
 347         while (start < end) {
 348                 ret = find_first_extent_bit(info->pinned_extents, start,
 349                                             &extent_start, &extent_end,
 350                                             EXTENT_DIRTY | EXTENT_UPTODATE,
 351                                             NULL);
 352                 if (ret)
 353                         break;
 354
 355                 if (extent_start <= start) {
 356                         start = extent_end + 1;
 357                 } else if (extent_start > start && extent_start < end) {
 358                         size = extent_start - start;
 359                         total_added += size;
 360                         ret = btrfs_add_free_space(block_group, start,
 361                                                    size);
 362                         BUG_ON(ret); /* -ENOMEM or logic error */
 363                         start = extent_end + 1;
 364                 } else {
 365                         break;
 366                 }
 367         }
 368
 369         if (start < end) {
 370                 size = end - start;
 371                 total_added += size;
 372                 ret = btrfs_add_free_space(block_group, start, size);
 373                 BUG_ON(ret); /* -ENOMEM or logic error */
 374         }
 375
 376         return total_added;
 377 }
 378
 379 static noinline void caching_thread(struct btrfs_work *work)
 380 {
 381         struct btrfs_block_group_cache *block_group;
 382         struct btrfs_fs_info *fs_info;
 383         struct btrfs_caching_control *caching_ctl;
 384         struct btrfs_root *extent_root;
 385         struct btrfs_path *path;
 386         struct extent_buffer *leaf;
 387         struct btrfs_key key;
 388         u64 total_found = 0;
 389         u64 last = 0;
 390         u32 nritems;
 391         int ret = 0;
 392
 393         caching_ctl = container_of(work, struct btrfs_caching_control, work);
 394         block_group = caching_ctl->block_group;
 395         fs_info = block_group->fs_info;
 396         extent_root = fs_info->extent_root;
 397
 398         path = btrfs_alloc_path();
 399         if (!path)
 400                 goto out;
 401
 402         last = max_t(u64, block_group->key.objectid, BTRFS_SUPER_INFO_OFFSET);
 403
 404         /*
 405          * We don't want to deadlock with somebody trying to allocate a new
 406          * extent for the extent root while also trying to search the extent
 407          * root to add free space.  So we skip locking and search the commit
 408          * root, since its read-only
 409          */
 410         path->skip_locking = 1;
 411         path->search_commit_root = 1;
 412         path->reada = 1;
 413
 414         key.objectid = last;
 415         key.offset = 0;
 416         key.type = BTRFS_EXTENT_ITEM_KEY;
 417 again:
 418         mutex_lock(&caching_ctl->mutex);
 419         /* need to make sure the commit_root doesn't disappear */
 420         down_read(&fs_info->extent_commit_sem);
 421
 422         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
 423         if (ret < 0)
 424                 goto err;
 425
 426         leaf = path->nodes[0];
 427         nritems = btrfs_header_nritems(leaf);
 428
 429         while (1) {
 430                 if (btrfs_fs_closing(fs_info) > 1) {
 431                         last = (u64)-1;
 432                         break;
 433                 }
 434
 435                 if (path->slots[0] < nritems) {
 436                         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
 437                 } else {
 438                         ret = find_next_key(path, 0, &key);
 439                         if (ret)
 440                                 break;
 441
 442                         if (need_resched()) {
 443                                 caching_ctl->progress = last;
 444                                 btrfs_release_path(path);
 445                                 up_read(&fs_info->extent_commit_sem);
 446                                 mutex_unlock(&caching_ctl->mutex);
 447                                 cond_resched();
 448                                 goto again;
 449                         }
 450
 451                         ret = btrfs_next_leaf(extent_root, path);
 452                         if (ret < 0)
 453                                 goto err;
 454                         if (ret)
 455                                 break;
 456                         leaf = path->nodes[0];
 457                         nritems = btrfs_header_nritems(leaf);
 458                         continue;
 459                 }
 460
 461                 if (key.objectid < block_group->key.objectid) {
 462                         path->slots[0]++;
 463                         continue;
 464                 }
 465
 466                 if (key.objectid >= block_group->key.objectid +
 467                     block_group->key.offset)
 468                         break;
 469
 470                 if (key.type == BTRFS_EXTENT_ITEM_KEY ||
 471                     key.type == BTRFS_METADATA_ITEM_KEY) {
 472                         total_found += add_new_free_space(block_group,
 473                                                           fs_info, last,
 474                                                           key.objectid);
 475                         if (key.type == BTRFS_METADATA_ITEM_KEY)
 476                                 last = key.objectid +
 477                                         fs_info->tree_root->leafsize;
 478                         else
 479                                 last = key.objectid + key.offset;
 480
 481                         if (total_found > (1024 * 1024 * 2)) {
 482                                 total_found = 0;
 483                                 wake_up(&caching_ctl->wait);
 484                         }
 485                 }
 486                 path->slots[0]++;
 487         }
 488         ret = 0;
 489
 490         total_found += add_new_free_space(block_group, fs_info, last,
 491                                           block_group->key.objectid +
 492                                           block_group->key.offset);
 493         caching_ctl->progress = (u64)-1;
 494
 495         spin_lock(&block_group->lock);
 496         block_group->caching_ctl = NULL;
 497         block_group->cached = BTRFS_CACHE_FINISHED;
 498         spin_unlock(&block_group->lock);
 499
 500 err:
 501         btrfs_free_path(path);
 502         up_read(&fs_info->extent_commit_sem);
 503
 504         free_excluded_extents(extent_root, block_group);
 505
 506         mutex_unlock(&caching_ctl->mutex);
 507 out:
 508         wake_up(&caching_ctl->wait);
 509
 510         put_caching_control(caching_ctl);
 511         btrfs_put_block_group(block_group);
 512 }
 513
 514 static int cache_block_group(struct btrfs_block_group_cache *cache,
 515                              int load_cache_only)
 516 {
 517         DEFINE_WAIT(wait);
 518         struct btrfs_fs_info *fs_info = cache->fs_info;
 519         struct btrfs_caching_control *caching_ctl;
 520         int ret = 0;
 521
 522         caching_ctl = kzalloc(sizeof(*caching_ctl), GFP_NOFS);
 523         if (!caching_ctl)
 524                 return -ENOMEM;
 525
 526         INIT_LIST_HEAD(&caching_ctl->list);
 527         mutex_init(&caching_ctl->mutex);
 528         init_waitqueue_head(&caching_ctl->wait);
 529         caching_ctl->block_group = cache;
 530         caching_ctl->progress = cache->key.objectid;
 531         atomic_set(&caching_ctl->count, 1);
 532         caching_ctl->work.func = caching_thread;
 533
 534         spin_lock(&cache->lock);
 535         /*
 536          * This should be a rare occasion, but this could happen I think in the
 537          * case where one thread starts to load the space cache info, and then
 538          * some other thread starts a transaction commit which tries to do an
 539          * allocation while the other thread is still loading the space cache
 540          * info.  The previous loop should have kept us from choosing this block
 541          * group, but if we've moved to the state where we will wait on caching
 542          * block groups we need to first check if we're doing a fast load here,
 543          * so we can wait for it to finish, otherwise we could end up allocating
 544          * from a block group who's cache gets evicted for one reason or
 545          * another.
 546          */
 547         while (cache->cached == BTRFS_CACHE_FAST) {
 548                 struct btrfs_caching_control *ctl;
 549
 550                 ctl = cache->caching_ctl;
 551                 atomic_inc(&ctl->count);
 552                 prepare_to_wait(&ctl->wait, &wait, TASK_UNINTERRUPTIBLE);
 553                 spin_unlock(&cache->lock);
 554
 555                 schedule();
 556
 557                 finish_wait(&ctl->wait, &wait);
 558                 put_caching_control(ctl);
 559                 spin_lock(&cache->lock);
 560         }
 561
 562         if (cache->cached != BTRFS_CACHE_NO) {
 563                 spin_unlock(&cache->lock);
 564                 kfree(caching_ctl);
 565                 return 0;
 566         }
 567         WARN_ON(cache->caching_ctl);
 568         cache->caching_ctl = caching_ctl;
 569         cache->cached = BTRFS_CACHE_FAST;
 570         spin_unlock(&cache->lock);
 571
 572         if (fs_info->mount_opt & BTRFS_MOUNT_SPACE_CACHE) {
 573                 ret = load_free_space_cache(fs_info, cache);
 574
 575                 spin_lock(&cache->lock);
 576                 if (ret == 1) {
 577                         cache->caching_ctl = NULL;
 578                         cache->cached = BTRFS_CACHE_FINISHED;
 579                         cache->last_byte_to_unpin = (u64)-1;
 580                 } else {
 581                         if (load_cache_only) {
 582                                 cache->caching_ctl = NULL;
 583                                 cache->cached = BTRFS_CACHE_NO;
 584                         } else {
 585                                 cache->cached = BTRFS_CACHE_STARTED;
 586                         }
 587                 }
 588                 spin_unlock(&cache->lock);
 589                 wake_up(&caching_ctl->wait);
 590                 if (ret == 1) {
 591                         put_caching_control(caching_ctl);
 592                         free_excluded_extents(fs_info->extent_root, cache);
 593                         return 0;
 594                 }
 595         } else {
 596                 /*
 597                  * We are not going to do the fast caching, set cached to the
 598                  * appropriate value and wakeup any waiters.
 599                  */
 600                 spin_lock(&cache->lock);
 601                 if (load_cache_only) {
 602                         cache->caching_ctl = NULL;
 603                         cache->cached = BTRFS_CACHE_NO;
 604                 } else {
 605                         cache->cached = BTRFS_CACHE_STARTED;
 606                 }
 607                 spin_unlock(&cache->lock);
 608                 wake_up(&caching_ctl->wait);
 609         }
 610
 611         if (load_cache_only) {
 612                 put_caching_control(caching_ctl);
 613                 return 0;
 614         }
 615
 616         down_write(&fs_info->extent_commit_sem);
 617         atomic_inc(&caching_ctl->count);
 618         list_add_tail(&caching_ctl->list, &fs_info->caching_block_groups);
 619         up_write(&fs_info->extent_commit_sem);
 620
 621         btrfs_get_block_group(cache);
 622
 623         btrfs_queue_worker(&fs_info->caching_workers, &caching_ctl->work);
 624
 625         return ret;
 626 }
 627
 628 /*
 629  * return the block group that starts at or after bytenr
 630  */
 631 static struct btrfs_block_group_cache *
 632 btrfs_lookup_first_block_group(struct btrfs_fs_info *info, u64 bytenr)
 633 {
 634         struct btrfs_block_group_cache *cache;
 635
 636         cache = block_group_cache_tree_search(info, bytenr, 0);
 637
 638         return cache;
 639 }
 640
 641 /*
 642  * return the block group that contains the given bytenr
 643  */
 644 struct btrfs_block_group_cache *btrfs_lookup_block_group(
 645                                                  struct btrfs_fs_info *info,
 646                                                  u64 bytenr)
 647 {
 648         struct btrfs_block_group_cache *cache;
 649
 650         cache = block_group_cache_tree_search(info, bytenr, 1);
 651
 652         return cache;
 653 }
 654
 655 static struct btrfs_space_info *__find_space_info(struct btrfs_fs_info *info,
 656                                                   u64 flags)
 657 {
 658         struct list_head *head = &info->space_info;
 659         struct btrfs_space_info *found;
 660
 661         flags &= BTRFS_BLOCK_GROUP_TYPE_MASK;
 662
 663         rcu_read_lock();
 664         list_for_each_entry_rcu(found, head, list) {
 665                 if (found->flags & flags) {
 666                         rcu_read_unlock();
 667                         return found;
 668                 }
 669         }
 670         rcu_read_unlock();
 671         return NULL;
 672 }
 673
 674 /*
 675  * after adding space to the filesystem, we need to clear the full flags
 676  * on all the space infos.
 677  */
 678 void btrfs_clear_space_info_full(struct btrfs_fs_info *info)
 679 {
 680         struct list_head *head = &info->space_info;
 681         struct btrfs_space_info *found;
 682
 683         rcu_read_lock();
 684         list_for_each_entry_rcu(found, head, list)
 685                 found->full = 0;
 686         rcu_read_unlock();
 687 }
 688
 689 /* simple helper to search for an existing extent at a given offset */
 690 int btrfs_lookup_extent(struct btrfs_root *root, u64 start, u64 len)
 691 {
 692         int ret;
 693         struct btrfs_key key;
 694         struct btrfs_path *path;
 695
 696         path = btrfs_alloc_path();
 697         if (!path)
 698                 return -ENOMEM;
 699
 700         key.objectid = start;
 701         key.offset = len;
 702         key.type = BTRFS_EXTENT_ITEM_KEY;
 703         ret = btrfs_search_slot(NULL, root->fs_info->extent_root, &key, path,
 704                                 0, 0);
 705         if (ret > 0) {
 706                 btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]);
 707                 if (key.objectid == start &&
 708                     key.type == BTRFS_METADATA_ITEM_KEY)
 709                         ret = 0;
 710         }
 711         btrfs_free_path(path);
 712         return ret;
 713 }
 714
 715 /*
 716  * helper function to lookup reference count and flags of a tree block.
 717  *
 718  * the head node for delayed ref is used to store the sum of all the
 719  * reference count modifications queued up in the rbtree. the head
 720  * node may also store the extent flags to set. This way you can check
 721  * to see what the reference count and extent flags would be if all of
 722  * the delayed refs are not processed.
 723  */
 724 int btrfs_lookup_extent_info(struct btrfs_trans_handle *trans,
 725                              struct btrfs_root *root, u64 bytenr,
 726                              u64 offset, int metadata, u64 *refs, u64 *flags)
 727 {
 728         struct btrfs_delayed_ref_head *head;
 729         struct btrfs_delayed_ref_root *delayed_refs;
 730         struct btrfs_path *path;
 731         struct btrfs_extent_item *ei;
 732         struct extent_buffer *leaf;
 733         struct btrfs_key key;
 734         u32 item_size;
 735         u64 num_refs;
 736         u64 extent_flags;
 737         int ret;
 738
 739         /*
 740          * If we don't have skinny metadata, don't bother doing anything
 741          * different
 742          */
 743         if (metadata && !btrfs_fs_incompat(root->fs_info, SKINNY_METADATA)) {
 744                 offset = root->leafsize;
 745                 metadata = 0;
 746         }
 747
 748         path = btrfs_alloc_path();
 749         if (!path)
 750                 return -ENOMEM;
 751
 752         if (metadata) {
 753                 key.objectid = bytenr;
 754                 key.type = BTRFS_METADATA_ITEM_KEY;
 755                 key.offset = offset;
 756         } else {
 757                 key.objectid = bytenr;
 758                 key.type = BTRFS_EXTENT_ITEM_KEY;
 759                 key.offset = offset;
 760         }
 761
 762         if (!trans) {
 763                 path->skip_locking = 1;
 764                 path->search_commit_root = 1;
 765         }
 766 again:
 767         ret = btrfs_search_slot(trans, root->fs_info->extent_root,
 768                                 &key, path, 0, 0);
 769         if (ret < 0)
 770                 goto out_free;
 771
 772         if (ret > 0 && metadata && key.type == BTRFS_METADATA_ITEM_KEY) {
 773                 key.type = BTRFS_EXTENT_ITEM_KEY;
 774                 key.offset = root->leafsize;
 775                 btrfs_release_path(path);
 776                 goto again;
 777         }
 778
 779         if (ret == 0) {
 780                 leaf = path->nodes[0];
 781                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
 782                 if (item_size >= sizeof(*ei)) {
 783                         ei = btrfs_item_ptr(leaf, path->slots[0],
 784                                             struct btrfs_extent_item);
 785                         num_refs = btrfs_extent_refs(leaf, ei);
 786                         extent_flags = btrfs_extent_flags(leaf, ei);
 787                 } else {
 788 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
 789                         struct btrfs_extent_item_v0 *ei0;
 790                         BUG_ON(item_size != sizeof(*ei0));
 791                         ei0 = btrfs_item_ptr(leaf, path->slots[0],
 792                                              struct btrfs_extent_item_v0);
 793                         num_refs = btrfs_extent_refs_v0(leaf, ei0);
 794                         /* FIXME: this isn't correct for data */
 795                         extent_flags = BTRFS_BLOCK_FLAG_FULL_BACKREF;
 796 #else
 797                         BUG();
 798 #endif
 799                 }
 800                 BUG_ON(num_refs == 0);
 801         } else {
 802                 num_refs = 0;
 803                 extent_flags = 0;
 804                 ret = 0;
 805         }
 806
 807         if (!trans)
 808                 goto out;
 809
 810         delayed_refs = &trans->transaction->delayed_refs;
 811         spin_lock(&delayed_refs->lock);
 812         head = btrfs_find_delayed_ref_head(trans, bytenr);
 813         if (head) {
 814                 if (!mutex_trylock(&head->mutex)) {
 815                         atomic_inc(&head->node.refs);
 816                         spin_unlock(&delayed_refs->lock);
 817
 818                         btrfs_release_path(path);
 819
 820                         /*
 821                          * Mutex was contended, block until it's released and try
 822                          * again
 823                          */
 824                         mutex_lock(&head->mutex);
 825                         mutex_unlock(&head->mutex);
 826                         btrfs_put_delayed_ref(&head->node);
 827                         goto again;
 828                 }
 829                 if (head->extent_op && head->extent_op->update_flags)
 830                         extent_flags |= head->extent_op->flags_to_set;
 831                 else
 832                         BUG_ON(num_refs == 0);
 833
 834                 num_refs += head->node.ref_mod;
 835                 mutex_unlock(&head->mutex);
 836         }
 837         spin_unlock(&delayed_refs->lock);
 838 out:
 839         WARN_ON(num_refs == 0);
 840         if (refs)
 841                 *refs = num_refs;
 842         if (flags)
 843                 *flags = extent_flags;
 844 out_free:
 845         btrfs_free_path(path);
 846         return ret;
 847 }
 848
 849 /*
 850  * Back reference rules.  Back refs have three main goals:
 851  *
 852  * 1) differentiate between all holders of references to an extent so that
 853  *    when a reference is dropped we can make sure it was a valid reference
 854  *    before freeing the extent.
 855  *
 856  * 2) Provide enough information to quickly find the holders of an extent
 857  *    if we notice a given block is corrupted or bad.
 858  *
 859  * 3) Make it easy to migrate blocks for FS shrinking or storage pool
 860  *    maintenance.  This is actually the same as #2, but with a slightly
 861  *    different use case.
 862  *
 863  * There are two kinds of back refs. The implicit back refs is optimized
 864  * for pointers in non-shared tree blocks. For a given pointer in a block,
 865  * back refs of this kind provide information about the block's owner tree
 866  * and the pointer's key. These information allow us to find the block by
 867  * b-tree searching. The full back refs is for pointers in tree blocks not
 868  * referenced by their owner trees. The location of tree block is recorded
 869  * in the back refs. Actually the full back refs is generic, and can be
 870  * used in all cases the implicit back refs is used. The major shortcoming
 871  * of the full back refs is its overhead. Every time a tree block gets
 872  * COWed, we have to update back refs entry for all pointers in it.
 873  *
 874  * For a newly allocated tree block, we use implicit back refs for
 875  * pointers in it. This means most tree related operations only involve
 876  * implicit back refs. For a tree block created in old transaction, the
 877  * only way to drop a reference to it is COW it. So we can detect the
 878  * event that tree block loses its owner tree's reference and do the
 879  * back refs conversion.
 880  *
 881  * When a tree block is COW'd through a tree, there are four cases:
 882  *
 883  * The reference count of the block is one and the tree is the block's
 884  * owner tree. Nothing to do in this case.
 885  *
 886  * The reference count of the block is one and the tree is not the
 887  * block's owner tree. In this case, full back refs is used for pointers
 888  * in the block. Remove these full back refs, add implicit back refs for
 889  * every pointers in the new block.
 890  *
 891  * The reference count of the block is greater than one and the tree is
 892  * the block's owner tree. In this case, implicit back refs is used for
 893  * pointers in the block. Add full back refs for every pointers in the
 894  * block, increase lower level extents' reference counts. The original
 895  * implicit back refs are entailed to the new block.
 896  *
 897  * The reference count of the block is greater than one and the tree is
 898  * not the block's owner tree. Add implicit back refs for every pointer in
 899  * the new block, increase lower level extents' reference count.
 900  *
 901  * Back Reference Key composing:
 902  *
 903  * The key objectid corresponds to the first byte in the extent,
 904  * The key type is used to differentiate between types of back refs.
 905  * There are different meanings of the key offset for different types
 906  * of back refs.
 907  *
 908  * File extents can be referenced by:
 909  *
 910  * - multiple snapshots, subvolumes, or different generations in one subvol
 911  * - different files inside a single subvolume
 912  * - different offsets inside a file (bookend extents in file.c)
 913  *
 914  * The extent ref structure for the implicit back refs has fields for:
 915  *
 916  * - Objectid of the subvolume root
 917  * - objectid of the file holding the reference
 918  * - original offset in the file
 919  * - how many bookend extents
 920  *
 921  * The key offset for the implicit back refs is hash of the first
 922  * three fields.
 923  *
 924  * The extent ref structure for the full back refs has field for:
 925  *
 926  * - number of pointers in the tree leaf
 927  *
 928  * The key offset for the implicit back refs is the first byte of
 929  * the tree leaf
 930  *
 931  * When a file extent is allocated, The implicit back refs is used.
 932  * the fields are filled in:
 933  *
 934  *     (root_key.objectid, inode objectid, offset in file, 1)
 935  *
 936  * When a file extent is removed file truncation, we find the
 937  * corresponding implicit back refs and check the following fields:
 938  *
 939  *     (btrfs_header_owner(leaf), inode objectid, offset in file)
 940  *
 941  * Btree extents can be referenced by:
 942  *
 943  * - Different subvolumes
 944  *
 945  * Both the implicit back refs and the full back refs for tree blocks
 946  * only consist of key. The key offset for the implicit back refs is
 947  * objectid of block's owner tree. The key offset for the full back refs
 948  * is the first byte of parent block.
 949  *
 950  * When implicit back refs is used, information about the lowest key and
 951  * level of the tree block are required. These information are stored in
 952  * tree block info structure.
 953  */
 954
 955 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
 956 static int convert_extent_item_v0(struct btrfs_trans_handle *trans,
 957                                   struct btrfs_root *root,
 958                                   struct btrfs_path *path,
 959                                   u64 owner, u32 extra_size)
 960 {
 961         struct btrfs_extent_item *item;
 962         struct btrfs_extent_item_v0 *ei0;
 963         struct btrfs_extent_ref_v0 *ref0;
 964         struct btrfs_tree_block_info *bi;
 965         struct extent_buffer *leaf;
 966         struct btrfs_key key;
 967         struct btrfs_key found_key;
 968         u32 new_size = sizeof(*item);
 969         u64 refs;
 970         int ret;
 971
 972         leaf = path->nodes[0];
 973         BUG_ON(btrfs_item_size_nr(leaf, path->slots[0]) != sizeof(*ei0));
 974
 975         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
 976         ei0 = btrfs_item_ptr(leaf, path->slots[0],
 977                              struct btrfs_extent_item_v0);
 978         refs = btrfs_extent_refs_v0(leaf, ei0);
 979
 980         if (owner == (u64)-1) {
 981                 while (1) {
 982                         if (path->slots[0] >= btrfs_header_nritems(leaf)) {
 983                                 ret = btrfs_next_leaf(root, path);
 984                                 if (ret < 0)
 985                                         return ret;
 986                                 BUG_ON(ret > 0); /* Corruption */
 987                                 leaf = path->nodes[0];
 988                         }
 989                         btrfs_item_key_to_cpu(leaf, &found_key,
 990                                               path->slots[0]);
 991                         BUG_ON(key.objectid != found_key.objectid);
 992                         if (found_key.type != BTRFS_EXTENT_REF_V0_KEY) {
 993                                 path->slots[0]++;
 994                                 continue;
 995                         }
 996                         ref0 = btrfs_item_ptr(leaf, path->slots[0],
 997                                               struct btrfs_extent_ref_v0);
 998                         owner = btrfs_ref_objectid_v0(leaf, ref0);
 999                         break;
1000                 }
1001         }
1002         btrfs_release_path(path);
1003
1004         if (owner < BTRFS_FIRST_FREE_OBJECTID)
1005                 new_size += sizeof(*bi);
1006
1007         new_size -= sizeof(*ei0);
1008         ret = btrfs_search_slot(trans, root, &key, path,
1009                                 new_size + extra_size, 1);
1010         if (ret < 0)
1011                 return ret;
1012         BUG_ON(ret); /* Corruption */
1013
1014         btrfs_extend_item(root, path, new_size);
1015
1016         leaf = path->nodes[0];
1017         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1018         btrfs_set_extent_refs(leaf, item, refs);
1019         /* FIXME: get real generation */
1020         btrfs_set_extent_generation(leaf, item, 0);
1021         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1022                 btrfs_set_extent_flags(leaf, item,
1023                                        BTRFS_EXTENT_FLAG_TREE_BLOCK |
1024                                        BTRFS_BLOCK_FLAG_FULL_BACKREF);
1025                 bi = (struct btrfs_tree_block_info *)(item + 1);
1026                 /* FIXME: get first key of the block */
1027                 memset_extent_buffer(leaf, 0, (unsigned long)bi, sizeof(*bi));
1028                 btrfs_set_tree_block_level(leaf, bi, (int)owner);
1029         } else {
1030                 btrfs_set_extent_flags(leaf, item, BTRFS_EXTENT_FLAG_DATA);
1031         }
1032         btrfs_mark_buffer_dirty(leaf);
1033         return 0;
1034 }
1035 #endif
1036
1037 static u64 hash_extent_data_ref(u64 root_objectid, u64 owner, u64 offset)
1038 {
1039         u32 high_crc = ~(u32)0;
1040         u32 low_crc = ~(u32)0;
1041         __le64 lenum;
1042
1043         lenum = cpu_to_le64(root_objectid);
1044         high_crc = crc32c(high_crc, &lenum, sizeof(lenum));
1045         lenum = cpu_to_le64(owner);
1046         low_crc = crc32c(low_crc, &lenum, sizeof(lenum));
1047         lenum = cpu_to_le64(offset);
1048         low_crc = crc32c(low_crc, &lenum, sizeof(lenum));
1049
1050         return ((u64)high_crc << 31) ^ (u64)low_crc;
1051 }
1052
1053 static u64 hash_extent_data_ref_item(struct extent_buffer *leaf,
1054                                      struct btrfs_extent_data_ref *ref)
1055 {
1056         return hash_extent_data_ref(btrfs_extent_data_ref_root(leaf, ref),
1057                                     btrfs_extent_data_ref_objectid(leaf, ref),
1058                                     btrfs_extent_data_ref_offset(leaf, ref));
1059 }
1060
1061 static int match_extent_data_ref(struct extent_buffer *leaf,
1062                                  struct btrfs_extent_data_ref *ref,
1063                                  u64 root_objectid, u64 owner, u64 offset)
1064 {
1065         if (btrfs_extent_data_ref_root(leaf, ref) != root_objectid ||
1066             btrfs_extent_data_ref_objectid(leaf, ref) != owner ||
1067             btrfs_extent_data_ref_offset(leaf, ref) != offset)
1068                 return 0;
1069         return 1;
1070 }
1071
1072 static noinline int lookup_extent_data_ref(struct btrfs_trans_handle *trans,
1073                                            struct btrfs_root *root,
1074                                            struct btrfs_path *path,
1075                                            u64 bytenr, u64 parent,
1076                                            u64 root_objectid,
1077                                            u64 owner, u64 offset)
1078 {
1079         struct btrfs_key key;
1080         struct btrfs_extent_data_ref *ref;
1081         struct extent_buffer *leaf;
1082         u32 nritems;
1083         int ret;
1084         int recow;
1085         int err = -ENOENT;
1086
1087         key.objectid = bytenr;
1088         if (parent) {
1089                 key.type = BTRFS_SHARED_DATA_REF_KEY;
1090                 key.offset = parent;
1091         } else {
1092                 key.type = BTRFS_EXTENT_DATA_REF_KEY;
1093                 key.offset = hash_extent_data_ref(root_objectid,
1094                                                   owner, offset);
1095         }
1096 again:
1097         recow = 0;
1098         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1099         if (ret < 0) {
1100                 err = ret;
1101                 goto fail;
1102         }
1103
1104         if (parent) {
1105                 if (!ret)
1106                         return 0;
1107 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1108                 key.type = BTRFS_EXTENT_REF_V0_KEY;
1109                 btrfs_release_path(path);
1110                 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1111                 if (ret < 0) {
1112                         err = ret;
1113                         goto fail;
1114                 }
1115                 if (!ret)
1116                         return 0;
1117 #endif
1118                 goto fail;
1119         }
1120
1121         leaf = path->nodes[0];
1122         nritems = btrfs_header_nritems(leaf);
1123         while (1) {
1124                 if (path->slots[0] >= nritems) {
1125                         ret = btrfs_next_leaf(root, path);
1126                         if (ret < 0)
1127                                 err = ret;
1128                         if (ret)
1129                                 goto fail;
1130
1131                         leaf = path->nodes[0];
1132                         nritems = btrfs_header_nritems(leaf);
1133                         recow = 1;
1134                 }
1135
1136                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1137                 if (key.objectid != bytenr ||
1138                     key.type != BTRFS_EXTENT_DATA_REF_KEY)
1139                         goto fail;
1140
1141                 ref = btrfs_item_ptr(leaf, path->slots[0],
1142                                      struct btrfs_extent_data_ref);
1143
1144                 if (match_extent_data_ref(leaf, ref, root_objectid,
1145                                           owner, offset)) {
1146                         if (recow) {
1147                                 btrfs_release_path(path);
1148                                 goto again;
1149                         }
1150                         err = 0;
1151                         break;
1152                 }
1153                 path->slots[0]++;
1154         }
1155 fail:
1156         return err;
1157 }
1158
1159 static noinline int insert_extent_data_ref(struct btrfs_trans_handle *trans,
1160                                            struct btrfs_root *root,
1161                                            struct btrfs_path *path,
1162                                            u64 bytenr, u64 parent,
1163                                            u64 root_objectid, u64 owner,
1164                                            u64 offset, int refs_to_add)
1165 {
1166         struct btrfs_key key;
1167         struct extent_buffer *leaf;
1168         u32 size;
1169         u32 num_refs;
1170         int ret;
1171
1172         key.objectid = bytenr;
1173         if (parent) {
1174                 key.type = BTRFS_SHARED_DATA_REF_KEY;
1175                 key.offset = parent;
1176                 size = sizeof(struct btrfs_shared_data_ref);
1177         } else {
1178                 key.type = BTRFS_EXTENT_DATA_REF_KEY;
1179                 key.offset = hash_extent_data_ref(root_objectid,
1180                                                   owner, offset);
1181                 size = sizeof(struct btrfs_extent_data_ref);
1182         }
1183
1184         ret = btrfs_insert_empty_item(trans, root, path, &key, size);
1185         if (ret && ret != -EEXIST)
1186                 goto fail;
1187
1188         leaf = path->nodes[0];
1189         if (parent) {
1190                 struct btrfs_shared_data_ref *ref;
1191                 ref = btrfs_item_ptr(leaf, path->slots[0],
1192                                      struct btrfs_shared_data_ref);
1193                 if (ret == 0) {
1194                         btrfs_set_shared_data_ref_count(leaf, ref, refs_to_add);
1195                 } else {
1196                         num_refs = btrfs_shared_data_ref_count(leaf, ref);
1197                         num_refs += refs_to_add;
1198                         btrfs_set_shared_data_ref_count(leaf, ref, num_refs);
1199                 }
1200         } else {
1201                 struct btrfs_extent_data_ref *ref;
1202                 while (ret == -EEXIST) {
1203                         ref = btrfs_item_ptr(leaf, path->slots[0],
1204                                              struct btrfs_extent_data_ref);
1205                         if (match_extent_data_ref(leaf, ref, root_objectid,
1206                                                   owner, offset))
1207                                 break;
1208                         btrfs_release_path(path);
1209                         key.offset++;
1210                         ret = btrfs_insert_empty_item(trans, root, path, &key,
1211                                                       size);
1212                         if (ret && ret != -EEXIST)
1213                                 goto fail;
1214
1215                         leaf = path->nodes[0];
1216                 }
1217                 ref = btrfs_item_ptr(leaf, path->slots[0],
1218                                      struct btrfs_extent_data_ref);
1219                 if (ret == 0) {
1220                         btrfs_set_extent_data_ref_root(leaf, ref,
1221                                                        root_objectid);
1222                         btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
1223                         btrfs_set_extent_data_ref_offset(leaf, ref, offset);
1224                         btrfs_set_extent_data_ref_count(leaf, ref, refs_to_add);
1225                 } else {
1226                         num_refs = btrfs_extent_data_ref_count(leaf, ref);
1227                         num_refs += refs_to_add;
1228                         btrfs_set_extent_data_ref_count(leaf, ref, num_refs);
1229                 }
1230         }
1231         btrfs_mark_buffer_dirty(leaf);
1232         ret = 0;
1233 fail:
1234         btrfs_release_path(path);
1235         return ret;
1236 }
1237
1238 static noinline int remove_extent_data_ref(struct btrfs_trans_handle *trans,
1239                                            struct btrfs_root *root,
1240                                            struct btrfs_path *path,
1241                                            int refs_to_drop)
1242 {
1243         struct btrfs_key key;
1244         struct btrfs_extent_data_ref *ref1 = NULL;
1245         struct btrfs_shared_data_ref *ref2 = NULL;
1246         struct extent_buffer *leaf;
1247         u32 num_refs = 0;
1248         int ret = 0;
1249
1250         leaf = path->nodes[0];
1251         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1252
1253         if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
1254                 ref1 = btrfs_item_ptr(leaf, path->slots[0],
1255                                       struct btrfs_extent_data_ref);
1256                 num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1257         } else if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
1258                 ref2 = btrfs_item_ptr(leaf, path->slots[0],
1259                                       struct btrfs_shared_data_ref);
1260                 num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1261 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1262         } else if (key.type == BTRFS_EXTENT_REF_V0_KEY) {
1263                 struct btrfs_extent_ref_v0 *ref0;
1264                 ref0 = btrfs_item_ptr(leaf, path->slots[0],
1265                                       struct btrfs_extent_ref_v0);
1266                 num_refs = btrfs_ref_count_v0(leaf, ref0);
1267 #endif
1268         } else {
1269                 BUG();
1270         }
1271
1272         BUG_ON(num_refs < refs_to_drop);
1273         num_refs -= refs_to_drop;
1274
1275         if (num_refs == 0) {
1276                 ret = btrfs_del_item(trans, root, path);
1277         } else {
1278                 if (key.type == BTRFS_EXTENT_DATA_REF_KEY)
1279                         btrfs_set_extent_data_ref_count(leaf, ref1, num_refs);
1280                 else if (key.type == BTRFS_SHARED_DATA_REF_KEY)
1281                         btrfs_set_shared_data_ref_count(leaf, ref2, num_refs);
1282 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1283                 else {
1284                         struct btrfs_extent_ref_v0 *ref0;
1285                         ref0 = btrfs_item_ptr(leaf, path->slots[0],
1286                                         struct btrfs_extent_ref_v0);
1287                         btrfs_set_ref_count_v0(leaf, ref0, num_refs);
1288                 }
1289 #endif
1290                 btrfs_mark_buffer_dirty(leaf);
1291         }
1292         return ret;
1293 }
1294
1295 static noinline u32 extent_data_ref_count(struct btrfs_root *root,
1296                                           struct btrfs_path *path,
1297                                           struct btrfs_extent_inline_ref *iref)
1298 {
1299         struct btrfs_key key;
1300         struct extent_buffer *leaf;
1301         struct btrfs_extent_data_ref *ref1;
1302         struct btrfs_shared_data_ref *ref2;
1303         u32 num_refs = 0;
1304
1305         leaf = path->nodes[0];
1306         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1307         if (iref) {
1308                 if (btrfs_extent_inline_ref_type(leaf, iref) ==
1309                     BTRFS_EXTENT_DATA_REF_KEY) {
1310                         ref1 = (struct btrfs_extent_data_ref *)(&iref->offset);
1311                         num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1312                 } else {
1313                         ref2 = (struct btrfs_shared_data_ref *)(iref + 1);
1314                         num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1315                 }
1316         } else if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
1317                 ref1 = btrfs_item_ptr(leaf, path->slots[0],
1318                                       struct btrfs_extent_data_ref);
1319                 num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1320         } else if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
1321                 ref2 = btrfs_item_ptr(leaf, path->slots[0],
1322                                       struct btrfs_shared_data_ref);
1323                 num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1324 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1325         } else if (key.type == BTRFS_EXTENT_REF_V0_KEY) {
1326                 struct btrfs_extent_ref_v0 *ref0;
1327                 ref0 = btrfs_item_ptr(leaf, path->slots[0],
1328                                       struct btrfs_extent_ref_v0);
1329                 num_refs = btrfs_ref_count_v0(leaf, ref0);
1330 #endif
1331         } else {
1332                 WARN_ON(1);
1333         }
1334         return num_refs;
1335 }
1336
1337 static noinline int lookup_tree_block_ref(struct btrfs_trans_handle *trans,
1338                                           struct btrfs_root *root,
1339                                           struct btrfs_path *path,
1340                                           u64 bytenr, u64 parent,
1341                                           u64 root_objectid)
1342 {
1343         struct btrfs_key key;
1344         int ret;
1345
1346         key.objectid = bytenr;
1347         if (parent) {
1348                 key.type = BTRFS_SHARED_BLOCK_REF_KEY;
1349                 key.offset = parent;
1350         } else {
1351                 key.type = BTRFS_TREE_BLOCK_REF_KEY;
1352                 key.offset = root_objectid;
1353         }
1354
1355         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1356         if (ret > 0)
1357                 ret = -ENOENT;
1358 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1359         if (ret == -ENOENT && parent) {
1360                 btrfs_release_path(path);
1361                 key.type = BTRFS_EXTENT_REF_V0_KEY;
1362                 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1363                 if (ret > 0)
1364                         ret = -ENOENT;
1365         }
1366 #endif
1367         return ret;
1368 }
1369
1370 static noinline int insert_tree_block_ref(struct btrfs_trans_handle *trans,
1371                                           struct btrfs_root *root,
1372                                           struct btrfs_path *path,
1373                                           u64 bytenr, u64 parent,
1374                                           u64 root_objectid)
1375 {
1376         struct btrfs_key key;
1377         int ret;
1378
1379         key.objectid = bytenr;
1380         if (parent) {
1381                 key.type = BTRFS_SHARED_BLOCK_REF_KEY;
1382                 key.offset = parent;
1383         } else {
1384                 key.type = BTRFS_TREE_BLOCK_REF_KEY;
1385                 key.offset = root_objectid;
1386         }
1387
1388         ret = btrfs_insert_empty_item(trans, root, path, &key, 0);
1389         btrfs_release_path(path);
1390         return ret;
1391 }
1392
1393 static inline int extent_ref_type(u64 parent, u64 owner)
1394 {
1395         int type;
1396         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1397                 if (parent > 0)
1398                         type = BTRFS_SHARED_BLOCK_REF_KEY;
1399                 else
1400                         type = BTRFS_TREE_BLOCK_REF_KEY;
1401         } else {
1402                 if (parent > 0)
1403                         type = BTRFS_SHARED_DATA_REF_KEY;
1404                 else
1405                         type = BTRFS_EXTENT_DATA_REF_KEY;
1406         }
1407         return type;
1408 }
1409
1410 static int find_next_key(struct btrfs_path *path, int level,
1411                          struct btrfs_key *key)
1412
1413 {
1414         for (; level < BTRFS_MAX_LEVEL; level++) {
1415                 if (!path->nodes[level])
1416                         break;
1417                 if (path->slots[level] + 1 >=
1418                     btrfs_header_nritems(path->nodes[level]))
1419                         continue;
1420                 if (level == 0)
1421                         btrfs_item_key_to_cpu(path->nodes[level], key,
1422                                               path->slots[level] + 1);
1423                 else
1424                         btrfs_node_key_to_cpu(path->nodes[level], key,
1425                                               path->slots[level] + 1);
1426                 return 0;
1427         }
1428         return 1;
1429 }
1430
1431 /*
1432  * look for inline back ref. if back ref is found, *ref_ret is set
1433  * to the address of inline back ref, and 0 is returned.
1434  *
1435  * if back ref isn't found, *ref_ret is set to the address where it
1436  * should be inserted, and -ENOENT is returned.
1437  *
1438  * if insert is true and there are too many inline back refs, the path
1439  * points to the extent item, and -EAGAIN is returned.
1440  *
1441  * NOTE: inline back refs are ordered in the same way that back ref
1442  *       items in the tree are ordered.
1443  */
1444 static noinline_for_stack
1445 int lookup_inline_extent_backref(struct btrfs_trans_handle *trans,
1446                                  struct btrfs_root *root,
1447                                  struct btrfs_path *path,
1448                                  struct btrfs_extent_inline_ref **ref_ret,
1449                                  u64 bytenr, u64 num_bytes,
1450                                  u64 parent, u64 root_objectid,
1451                                  u64 owner, u64 offset, int insert)
1452 {
1453         struct btrfs_key key;
1454         struct extent_buffer *leaf;
1455         struct btrfs_extent_item *ei;
1456         struct btrfs_extent_inline_ref *iref;
1457         u64 flags;
1458         u64 item_size;
1459         unsigned long ptr;
1460         unsigned long end;
1461         int extra_size;
1462         int type;
1463         int want;
1464         int ret;
1465         int err = 0;
1466         bool skinny_metadata = btrfs_fs_incompat(root->fs_info,
1467                                                  SKINNY_METADATA);
1468
1469         key.objectid = bytenr;
1470         key.type = BTRFS_EXTENT_ITEM_KEY;
1471         key.offset = num_bytes;
1472
1473         want = extent_ref_type(parent, owner);
1474         if (insert) {
1475                 extra_size = btrfs_extent_inline_ref_size(want);
1476                 path->keep_locks = 1;
1477         } else
1478                 extra_size = -1;
1479
1480         /*
1481          * Owner is our parent level, so we can just add one to get the level
1482          * for the block we are interested in.
1483          */
1484         if (skinny_metadata && owner < BTRFS_FIRST_FREE_OBJECTID) {
1485                 key.type = BTRFS_METADATA_ITEM_KEY;
1486                 key.offset = owner;
1487         }
1488
1489 again:
1490         ret = btrfs_search_slot(trans, root, &key, path, extra_size, 1);
1491         if (ret < 0) {
1492                 err = ret;
1493                 goto out;
1494         }
1495
1496         /*
1497          * We may be a newly converted file system which still has the old fat
1498          * extent entries for metadata, so try and see if we have one of those.
1499          */
1500         if (ret > 0 && skinny_metadata) {
1501                 skinny_metadata = false;
1502                 if (path->slots[0]) {
1503                         path->slots[0]--;
1504                         btrfs_item_key_to_cpu(path->nodes[0], &key,
1505                                               path->slots[0]);
1506                         if (key.objectid == bytenr &&
1507                             key.type == BTRFS_EXTENT_ITEM_KEY &&
1508                             key.offset == num_bytes)
1509                                 ret = 0;
1510                 }
1511                 if (ret) {
1512                         key.type = BTRFS_EXTENT_ITEM_KEY;
1513                         key.offset = num_bytes;
1514                         btrfs_release_path(path);
1515                         goto again;
1516                 }
1517         }
1518
1519         if (ret && !insert) {
1520                 err = -ENOENT;
1521                 goto out;
1522         } else if (ret) {
1523                 err = -EIO;
1524                 WARN_ON(1);
1525                 goto out;
1526         }
1527
1528         leaf = path->nodes[0];
1529         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1530 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1531         if (item_size < sizeof(*ei)) {
1532                 if (!insert) {
1533                         err = -ENOENT;
1534                         goto out;
1535                 }
1536                 ret = convert_extent_item_v0(trans, root, path, owner,
1537                                              extra_size);
1538                 if (ret < 0) {
1539                         err = ret;
1540                         goto out;
1541                 }
1542                 leaf = path->nodes[0];
1543                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1544         }
1545 #endif
1546         BUG_ON(item_size < sizeof(*ei));
1547
1548         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1549         flags = btrfs_extent_flags(leaf, ei);
1550
1551         ptr = (unsigned long)(ei + 1);
1552         end = (unsigned long)ei + item_size;
1553
1554         if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK && !skinny_metadata) {
1555                 ptr += sizeof(struct btrfs_tree_block_info);
1556                 BUG_ON(ptr > end);
1557         }
1558
1559         err = -ENOENT;
1560         while (1) {
1561                 if (ptr >= end) {
1562                         WARN_ON(ptr > end);
1563                         break;
1564                 }
1565                 iref = (struct btrfs_extent_inline_ref *)ptr;
1566                 type = btrfs_extent_inline_ref_type(leaf, iref);
1567                 if (want < type)
1568                         break;
1569                 if (want > type) {
1570                         ptr += btrfs_extent_inline_ref_size(type);
1571                         continue;
1572                 }
1573
1574                 if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1575                         struct btrfs_extent_data_ref *dref;
1576                         dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1577                         if (match_extent_data_ref(leaf, dref, root_objectid,
1578                                                   owner, offset)) {
1579                                 err = 0;
1580                                 break;
1581                         }
1582                         if (hash_extent_data_ref_item(leaf, dref) <
1583                             hash_extent_data_ref(root_objectid, owner, offset))
1584                                 break;
1585                 } else {
1586                         u64 ref_offset;
1587                         ref_offset = btrfs_extent_inline_ref_offset(leaf, iref);
1588                         if (parent > 0) {
1589                                 if (parent == ref_offset) {
1590                                         err = 0;
1591                                         break;
1592                                 }
1593                                 if (ref_offset < parent)
1594                                         break;
1595                         } else {
1596                                 if (root_objectid == ref_offset) {
1597                                         err = 0;
1598                                         break;
1599                                 }
1600                                 if (ref_offset < root_objectid)
1601                                         break;
1602                         }
1603                 }
1604                 ptr += btrfs_extent_inline_ref_size(type);
1605         }
1606         if (err == -ENOENT && insert) {
1607                 if (item_size + extra_size >=
1608                     BTRFS_MAX_EXTENT_ITEM_SIZE(root)) {
1609                         err = -EAGAIN;
1610                         goto out;
1611                 }
1612                 /*
1613                  * To add new inline back ref, we have to make sure
1614                  * there is no corresponding back ref item.
1615                  * For simplicity, we just do not add new inline back
1616                  * ref if there is any kind of item for this block
1617                  */
1618                 if (find_next_key(path, 0, &key) == 0 &&
1619                     key.objectid == bytenr &&
1620                     key.type < BTRFS_BLOCK_GROUP_ITEM_KEY) {
1621                         err = -EAGAIN;
1622                         goto out;
1623                 }
1624         }
1625         *ref_ret = (struct btrfs_extent_inline_ref *)ptr;
1626 out:
1627         if (insert) {
1628                 path->keep_locks = 0;
1629                 btrfs_unlock_up_safe(path, 1);
1630         }
1631         return err;
1632 }
1633
1634 /*
1635  * helper to add new inline back ref
1636  */
1637 static noinline_for_stack
1638 void setup_inline_extent_backref(struct btrfs_root *root,
1639                                  struct btrfs_path *path,
1640                                  struct btrfs_extent_inline_ref *iref,
1641                                  u64 parent, u64 root_objectid,
1642                                  u64 owner, u64 offset, int refs_to_add,
1643                                  struct btrfs_delayed_extent_op *extent_op)
1644 {
1645         struct extent_buffer *leaf;
1646         struct btrfs_extent_item *ei;
1647         unsigned long ptr;
1648         unsigned long end;
1649         unsigned long item_offset;
1650         u64 refs;
1651         int size;
1652         int type;
1653
1654         leaf = path->nodes[0];
1655         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1656         item_offset = (unsigned long)iref - (unsigned long)ei;
1657
1658         type = extent_ref_type(parent, owner);
1659         size = btrfs_extent_inline_ref_size(type);
1660
1661         btrfs_extend_item(root, path, size);
1662
1663         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1664         refs = btrfs_extent_refs(leaf, ei);
1665         refs += refs_to_add;
1666         btrfs_set_extent_refs(leaf, ei, refs);
1667         if (extent_op)
1668                 __run_delayed_extent_op(extent_op, leaf, ei);
1669
1670         ptr = (unsigned long)ei + item_offset;
1671         end = (unsigned long)ei + btrfs_item_size_nr(leaf, path->slots[0]);
1672         if (ptr < end - size)
1673                 memmove_extent_buffer(leaf, ptr + size, ptr,
1674                                       end - size - ptr);
1675
1676         iref = (struct btrfs_extent_inline_ref *)ptr;
1677         btrfs_set_extent_inline_ref_type(leaf, iref, type);
1678         if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1679                 struct btrfs_extent_data_ref *dref;
1680                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1681                 btrfs_set_extent_data_ref_root(leaf, dref, root_objectid);
1682                 btrfs_set_extent_data_ref_objectid(leaf, dref, owner);
1683                 btrfs_set_extent_data_ref_offset(leaf, dref, offset);
1684                 btrfs_set_extent_data_ref_count(leaf, dref, refs_to_add);
1685         } else if (type == BTRFS_SHARED_DATA_REF_KEY) {
1686                 struct btrfs_shared_data_ref *sref;
1687                 sref = (struct btrfs_shared_data_ref *)(iref + 1);
1688                 btrfs_set_shared_data_ref_count(leaf, sref, refs_to_add);
1689                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
1690         } else if (type == BTRFS_SHARED_BLOCK_REF_KEY) {
1691                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
1692         } else {
1693                 btrfs_set_extent_inline_ref_offset(leaf, iref, root_objectid);
1694         }
1695         btrfs_mark_buffer_dirty(leaf);
1696 }
1697
1698 static int lookup_extent_backref(struct btrfs_trans_handle *trans,
1699                                  struct btrfs_root *root,
1700                                  struct btrfs_path *path,
1701                                  struct btrfs_extent_inline_ref **ref_ret,
1702                                  u64 bytenr, u64 num_bytes, u64 parent,
1703                                  u64 root_objectid, u64 owner, u64 offset)
1704 {
1705         int ret;
1706
1707         ret = lookup_inline_extent_backref(trans, root, path, ref_ret,
1708                                            bytenr, num_bytes, parent,
1709                                            root_objectid, owner, offset, 0);
1710         if (ret != -ENOENT)
1711                 return ret;
1712
1713         btrfs_release_path(path);
1714         *ref_ret = NULL;
1715
1716         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1717                 ret = lookup_tree_block_ref(trans, root, path, bytenr, parent,
1718                                             root_objectid);
1719         } else {
1720                 ret = lookup_extent_data_ref(trans, root, path, bytenr, parent,
1721                                              root_objectid, owner, offset);
1722         }
1723         return ret;
1724 }
1725
1726 /*
1727  * helper to update/remove inline back ref
1728  */
1729 static noinline_for_stack
1730 void update_inline_extent_backref(struct btrfs_root *root,
1731                                   struct btrfs_path *path,
1732                                   struct btrfs_extent_inline_ref *iref,
1733                                   int refs_to_mod,
1734                                   struct btrfs_delayed_extent_op *extent_op)
1735 {
1736         struct extent_buffer *leaf;
1737         struct btrfs_extent_item *ei;
1738         struct btrfs_extent_data_ref *dref = NULL;
1739         struct btrfs_shared_data_ref *sref = NULL;
1740         unsigned long ptr;
1741         unsigned long end;
1742         u32 item_size;
1743         int size;
1744         int type;
1745         u64 refs;
1746
1747         leaf = path->nodes[0];
1748         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1749         refs = btrfs_extent_refs(leaf, ei);
1750         WARN_ON(refs_to_mod < 0 && refs + refs_to_mod <= 0);
1751         refs += refs_to_mod;
1752         btrfs_set_extent_refs(leaf, ei, refs);
1753         if (extent_op)
1754                 __run_delayed_extent_op(extent_op, leaf, ei);
1755
1756         type = btrfs_extent_inline_ref_type(leaf, iref);
1757
1758         if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1759                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1760                 refs = btrfs_extent_data_ref_count(leaf, dref);
1761         } else if (type == BTRFS_SHARED_DATA_REF_KEY) {
1762                 sref = (struct btrfs_shared_data_ref *)(iref + 1);
1763                 refs = btrfs_shared_data_ref_count(leaf, sref);
1764         } else {
1765                 refs = 1;
1766                 BUG_ON(refs_to_mod != -1);
1767         }
1768
1769         BUG_ON(refs_to_mod < 0 && refs < -refs_to_mod);
1770         refs += refs_to_mod;
1771
1772         if (refs > 0) {
1773                 if (type == BTRFS_EXTENT_DATA_REF_KEY)
1774                         btrfs_set_extent_data_ref_count(leaf, dref, refs);
1775                 else
1776                         btrfs_set_shared_data_ref_count(leaf, sref, refs);
1777         } else {
1778                 size =  btrfs_extent_inline_ref_size(type);
1779                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1780                 ptr = (unsigned long)iref;
1781                 end = (unsigned long)ei + item_size;
1782                 if (ptr + size < end)
1783                         memmove_extent_buffer(leaf, ptr, ptr + size,
1784                                               end - ptr - size);
1785                 item_size -= size;
1786                 btrfs_truncate_item(root, path, item_size, 1);
1787         }
1788         btrfs_mark_buffer_dirty(leaf);
1789 }
1790
1791 static noinline_for_stack
1792 int insert_inline_extent_backref(struct btrfs_trans_handle *trans,
1793                                  struct btrfs_root *root,
1794                                  struct btrfs_path *path,
1795                                  u64 bytenr, u64 num_bytes, u64 parent,
1796                                  u64 root_objectid, u64 owner,
1797                                  u64 offset, int refs_to_add,
1798                                  struct btrfs_delayed_extent_op *extent_op)
1799 {
1800         struct btrfs_extent_inline_ref *iref;
1801         int ret;
1802
1803         ret = lookup_inline_extent_backref(trans, root, path, &iref,
1804                                            bytenr, num_bytes, parent,
1805                                            root_objectid, owner, offset, 1);
1806         if (ret == 0) {
1807                 BUG_ON(owner < BTRFS_FIRST_FREE_OBJECTID);
1808                 update_inline_extent_backref(root, path, iref,
1809                                              refs_to_add, extent_op);
1810         } else if (ret == -ENOENT) {
1811                 setup_inline_extent_backref(root, path, iref, parent,
1812                                             root_objectid, owner, offset,
1813                                             refs_to_add, extent_op);
1814                 ret = 0;
1815         }
1816         return ret;
1817 }
1818
1819 static int insert_extent_backref(struct btrfs_trans_handle *trans,
1820                                  struct btrfs_root *root,
1821                                  struct btrfs_path *path,
1822                                  u64 bytenr, u64 parent, u64 root_objectid,
1823                                  u64 owner, u64 offset, int refs_to_add)
1824 {
1825         int ret;
1826         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1827                 BUG_ON(refs_to_add != 1);
1828                 ret = insert_tree_block_ref(trans, root, path, bytenr,
1829                                             parent, root_objectid);
1830         } else {
1831                 ret = insert_extent_data_ref(trans, root, path, bytenr,
1832                                              parent, root_objectid,
1833                                              owner, offset, refs_to_add);
1834         }
1835         return ret;
1836 }
1837
1838 static int remove_extent_backref(struct btrfs_trans_handle *trans,
1839                                  struct btrfs_root *root,
1840                                  struct btrfs_path *path,
1841                                  struct btrfs_extent_inline_ref *iref,
1842                                  int refs_to_drop, int is_data)
1843 {
1844         int ret = 0;
1845
1846         BUG_ON(!is_data && refs_to_drop != 1);
1847         if (iref) {
1848                 update_inline_extent_backref(root, path, iref,
1849                                              -refs_to_drop, NULL);
1850         } else if (is_data) {
1851                 ret = remove_extent_data_ref(trans, root, path, refs_to_drop);
1852         } else {
1853                 ret = btrfs_del_item(trans, root, path);
1854         }
1855         return ret;
1856 }
1857
1858 static int btrfs_issue_discard(struct block_device *bdev,
1859                                 u64 start, u64 len)
1860 {
1861         return blkdev_issue_discard(bdev, start >> 9, len >> 9, GFP_NOFS, 0);
1862 }
1863
1864 static int btrfs_discard_extent(struct btrfs_root *root, u64 bytenr,
1865                                 u64 num_bytes, u64 *actual_bytes)
1866 {
1867         int ret;
1868         u64 discarded_bytes = 0;
1869         struct btrfs_bio *bbio = NULL;
1870
1871
1872         /* Tell the block device(s) that the sectors can be discarded */
1873         ret = btrfs_map_block(root->fs_info, REQ_DISCARD,
1874                               bytenr, &num_bytes, &bbio, 0);
1875         /* Error condition is -ENOMEM */
1876         if (!ret) {
1877                 struct btrfs_bio_stripe *stripe = bbio->stripes;
1878                 int i;
1879
1880
1881                 for (i = 0; i < bbio->num_stripes; i++, stripe++) {
1882                         if (!stripe->dev->can_discard)
1883                                 continue;
1884
1885                         ret = btrfs_issue_discard(stripe->dev->bdev,
1886                                                   stripe->physical,
1887                                                   stripe->length);
1888                         if (!ret)
1889                                 discarded_bytes += stripe->length;
1890                         else if (ret != -EOPNOTSUPP)
1891                                 break; /* Logic errors or -ENOMEM, or -EIO but I don't know how that could happen JDM */
1892
1893                         /*
1894                          * Just in case we get back EOPNOTSUPP for some reason,
1895                          * just ignore the return value so we don't screw up
1896                          * people calling discard_extent.
1897                          */
1898                         ret = 0;
1899                 }
1900                 kfree(bbio);
1901         }
1902
1903         if (actual_bytes)
1904                 *actual_bytes = discarded_bytes;
1905
1906
1907         if (ret == -EOPNOTSUPP)
1908                 ret = 0;
1909         return ret;
1910 }
1911
1912 /* Can return -ENOMEM */
1913 int btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
1914                          struct btrfs_root *root,
1915                          u64 bytenr, u64 num_bytes, u64 parent,
1916                          u64 root_objectid, u64 owner, u64 offset, int for_cow)
1917 {
1918         int ret;
1919         struct btrfs_fs_info *fs_info = root->fs_info;
1920
1921         BUG_ON(owner < BTRFS_FIRST_FREE_OBJECTID &&
1922                root_objectid == BTRFS_TREE_LOG_OBJECTID);
1923
1924         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1925                 ret = btrfs_add_delayed_tree_ref(fs_info, trans, bytenr,
1926                                         num_bytes,
1927                                         parent, root_objectid, (int)owner,
1928                                         BTRFS_ADD_DELAYED_REF, NULL, for_cow);
1929         } else {
1930                 ret = btrfs_add_delayed_data_ref(fs_info, trans, bytenr,
1931                                         num_bytes,
1932                                         parent, root_objectid, owner, offset,
1933                                         BTRFS_ADD_DELAYED_REF, NULL, for_cow);
1934         }
1935         return ret;
1936 }
1937
1938 static int __btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
1939                                   struct btrfs_root *root,
1940                                   u64 bytenr, u64 num_bytes,
1941                                   u64 parent, u64 root_objectid,
1942                                   u64 owner, u64 offset, int refs_to_add,
1943                                   struct btrfs_delayed_extent_op *extent_op)
1944 {
1945         struct btrfs_path *path;
1946         struct extent_buffer *leaf;
1947         struct btrfs_extent_item *item;
1948         u64 refs;
1949         int ret;
1950         int err = 0;
1951
1952         path = btrfs_alloc_path();
1953         if (!path)
1954                 return -ENOMEM;
1955
1956         path->reada = 1;
1957         path->leave_spinning = 1;
1958         /* this will setup the path even if it fails to insert the back ref */
1959         ret = insert_inline_extent_backref(trans, root->fs_info->extent_root,
1960                                            path, bytenr, num_bytes, parent,
1961                                            root_objectid, owner, offset,
1962                                            refs_to_add, extent_op);
1963         if (ret == 0)
1964                 goto out;
1965
1966         if (ret != -EAGAIN) {
1967                 err = ret;
1968                 goto out;
1969         }
1970
1971         leaf = path->nodes[0];
1972         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1973         refs = btrfs_extent_refs(leaf, item);
1974         btrfs_set_extent_refs(leaf, item, refs + refs_to_add);
1975         if (extent_op)
1976                 __run_delayed_extent_op(extent_op, leaf, item);
1977
1978         btrfs_mark_buffer_dirty(leaf);
1979         btrfs_release_path(path);
1980
1981         path->reada = 1;
1982         path->leave_spinning = 1;
1983
1984         /* now insert the actual backref */
1985         ret = insert_extent_backref(trans, root->fs_info->extent_root,
1986                                     path, bytenr, parent, root_objectid,
1987                                     owner, offset, refs_to_add);
1988         if (ret)
1989                 btrfs_abort_transaction(trans, root, ret);
1990 out:
1991         btrfs_free_path(path);
1992         return err;
1993 }
1994
1995 static int run_delayed_data_ref(struct btrfs_trans_handle *trans,
1996                                 struct btrfs_root *root,
1997                                 struct btrfs_delayed_ref_node *node,
1998                                 struct btrfs_delayed_extent_op *extent_op,
1999                                 int insert_reserved)
2000 {
2001         int ret = 0;
2002         struct btrfs_delayed_data_ref *ref;
2003         struct btrfs_key ins;
2004         u64 parent = 0;
2005         u64 ref_root = 0;
2006         u64 flags = 0;
2007
2008         ins.objectid = node->bytenr;
2009         ins.offset = node->num_bytes;
2010         ins.type = BTRFS_EXTENT_ITEM_KEY;
2011
2012         ref = btrfs_delayed_node_to_data_ref(node);
2013         if (node->type == BTRFS_SHARED_DATA_REF_KEY)
2014                 parent = ref->parent;
2015         else
2016                 ref_root = ref->root;
2017
2018         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
2019                 if (extent_op)
2020                         flags |= extent_op->flags_to_set;
2021                 ret = alloc_reserved_file_extent(trans, root,
2022                                                  parent, ref_root, flags,
2023                                                  ref->objectid, ref->offset,
2024                                                  &ins, node->ref_mod);
2025         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
2026                 ret = __btrfs_inc_extent_ref(trans, root, node->bytenr,
2027                                              node->num_bytes, parent,
2028                                              ref_root, ref->objectid,
2029                                              ref->offset, node->ref_mod,
2030                                              extent_op);
2031         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
2032                 ret = __btrfs_free_extent(trans, root, node->bytenr,
2033                                           node->num_bytes, parent,
2034                                           ref_root, ref->objectid,
2035                                           ref->offset, node->ref_mod,
2036                                           extent_op);
2037         } else {
2038                 BUG();
2039         }
2040         return ret;
2041 }
2042
2043 static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op,
2044                                     struct extent_buffer *leaf,
2045                                     struct btrfs_extent_item *ei)
2046 {
2047         u64 flags = btrfs_extent_flags(leaf, ei);
2048         if (extent_op->update_flags) {
2049                 flags |= extent_op->flags_to_set;
2050                 btrfs_set_extent_flags(leaf, ei, flags);
2051         }
2052
2053         if (extent_op->update_key) {
2054                 struct btrfs_tree_block_info *bi;
2055                 BUG_ON(!(flags & BTRFS_EXTENT_FLAG_TREE_BLOCK));
2056                 bi = (struct btrfs_tree_block_info *)(ei + 1);
2057                 btrfs_set_tree_block_key(leaf, bi, &extent_op->key);
2058         }
2059 }
2060
2061 static int run_delayed_extent_op(struct btrfs_trans_handle *trans,
2062                                  struct btrfs_root *root,
2063                                  struct btrfs_delayed_ref_node *node,
2064                                  struct btrfs_delayed_extent_op *extent_op)
2065 {
2066         struct btrfs_key key;
2067         struct btrfs_path *path;
2068         struct btrfs_extent_item *ei;
2069         struct extent_buffer *leaf;
2070         u32 item_size;
2071         int ret;
2072         int err = 0;
2073         int metadata = !extent_op->is_data;
2074
2075         if (trans->aborted)
2076                 return 0;
2077
2078         if (metadata && !btrfs_fs_incompat(root->fs_info, SKINNY_METADATA))
2079                 metadata = 0;
2080
2081         path = btrfs_alloc_path();
2082         if (!path)
2083                 return -ENOMEM;
2084
2085         key.objectid = node->bytenr;
2086
2087         if (metadata) {
2088                 key.type = BTRFS_METADATA_ITEM_KEY;
2089                 key.offset = extent_op->level;
2090         } else {
2091                 key.type = BTRFS_EXTENT_ITEM_KEY;
2092                 key.offset = node->num_bytes;
2093         }
2094
2095 again:
2096         path->reada = 1;
2097         path->leave_spinning = 1;
2098         ret = btrfs_search_slot(trans, root->fs_info->extent_root, &key,
2099                                 path, 0, 1);
2100         if (ret < 0) {
2101                 err = ret;
2102                 goto out;
2103         }
2104         if (ret > 0) {
2105                 if (metadata) {
2106                         btrfs_release_path(path);
2107                         metadata = 0;
2108
2109                         key.offset = node->num_bytes;
2110                         key.type = BTRFS_EXTENT_ITEM_KEY;
2111                         goto again;
2112                 }
2113                 err = -EIO;
2114                 goto out;
2115         }
2116
2117         leaf = path->nodes[0];
2118         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2119 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
2120         if (item_size < sizeof(*ei)) {
2121                 ret = convert_extent_item_v0(trans, root->fs_info->extent_root,
2122                                              path, (u64)-1, 0);
2123                 if (ret < 0) {
2124                         err = ret;
2125                         goto out;
2126                 }
2127                 leaf = path->nodes[0];
2128                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2129         }
2130 #endif
2131         BUG_ON(item_size < sizeof(*ei));
2132         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2133         __run_delayed_extent_op(extent_op, leaf, ei);
2134
2135         btrfs_mark_buffer_dirty(leaf);
2136 out:
2137         btrfs_free_path(path);
2138         return err;
2139 }
2140
2141 static int run_delayed_tree_ref(struct btrfs_trans_handle *trans,
2142                                 struct btrfs_root *root,
2143                                 struct btrfs_delayed_ref_node *node,
2144                                 struct btrfs_delayed_extent_op *extent_op,
2145                                 int insert_reserved)
2146 {
2147         int ret = 0;
2148         struct btrfs_delayed_tree_ref *ref;
2149         struct btrfs_key ins;
2150         u64 parent = 0;
2151         u64 ref_root = 0;
2152         bool skinny_metadata = btrfs_fs_incompat(root->fs_info,
2153                                                  SKINNY_METADATA);
2154
2155         ref = btrfs_delayed_node_to_tree_ref(node);
2156         if (node->type == BTRFS_SHARED_BLOCK_REF_KEY)
2157                 parent = ref->parent;
2158         else
2159                 ref_root = ref->root;
2160
2161         ins.objectid = node->bytenr;
2162         if (skinny_metadata) {
2163                 ins.offset = ref->level;
2164                 ins.type = BTRFS_METADATA_ITEM_KEY;
2165         } else {
2166                 ins.offset = node->num_bytes;
2167                 ins.type = BTRFS_EXTENT_ITEM_KEY;
2168         }
2169
2170         BUG_ON(node->ref_mod != 1);
2171         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
2172                 BUG_ON(!extent_op || !extent_op->update_flags);
2173                 ret = alloc_reserved_tree_block(trans, root,
2174                                                 parent, ref_root,
2175                                                 extent_op->flags_to_set,
2176                                                 &extent_op->key,
2177                                                 ref->level, &ins);
2178         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
2179                 ret = __btrfs_inc_extent_ref(trans, root, node->bytenr,
2180                                              node->num_bytes, parent, ref_root,
2181                                              ref->level, 0, 1, extent_op);
2182         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
2183                 ret = __btrfs_free_extent(trans, root, node->bytenr,
2184                                           node->num_bytes, parent, ref_root,
2185                                           ref->level, 0, 1, extent_op);
2186         } else {
2187                 BUG();
2188         }
2189         return ret;
2190 }
2191
2192 /* helper function to actually process a single delayed ref entry */
2193 static int run_one_delayed_ref(struct btrfs_trans_handle *trans,
2194                                struct btrfs_root *root,
2195                                struct btrfs_delayed_ref_node *node,
2196                                struct btrfs_delayed_extent_op *extent_op,
2197                                int insert_reserved)
2198 {
2199         int ret = 0;
2200
2201         if (trans->aborted)
2202                 return 0;
2203
2204         if (btrfs_delayed_ref_is_head(node)) {
2205                 struct btrfs_delayed_ref_head *head;
2206                 /*
2207                  * we've hit the end of the chain and we were supposed
2208                  * to insert this extent into the tree.  But, it got
2209                  * deleted before we ever needed to insert it, so all
2210                  * we have to do is clean up the accounting
2211                  */
2212                 BUG_ON(extent_op);
2213                 head = btrfs_delayed_node_to_head(node);
2214                 if (insert_reserved) {
2215                         btrfs_pin_extent(root, node->bytenr,
2216                                          node->num_bytes, 1);
2217                         if (head->is_data) {
2218                                 ret = btrfs_del_csums(trans, root,
2219                                                       node->bytenr,
2220                                                       node->num_bytes);
2221                         }
2222                 }
2223                 return ret;
2224         }
2225
2226         if (node->type == BTRFS_TREE_BLOCK_REF_KEY ||
2227             node->type == BTRFS_SHARED_BLOCK_REF_KEY)
2228                 ret = run_delayed_tree_ref(trans, root, node, extent_op,
2229                                            insert_reserved);
2230         else if (node->type == BTRFS_EXTENT_DATA_REF_KEY ||
2231                  node->type == BTRFS_SHARED_DATA_REF_KEY)
2232                 ret = run_delayed_data_ref(trans, root, node, extent_op,
2233                                            insert_reserved);
2234         else
2235                 BUG();
2236         return ret;
2237 }
2238
2239 static noinline struct btrfs_delayed_ref_node *
2240 select_delayed_ref(struct btrfs_delayed_ref_head *head)
2241 {
2242         struct rb_node *node;
2243         struct btrfs_delayed_ref_node *ref;
2244         int action = BTRFS_ADD_DELAYED_REF;
2245 again:
2246         /*
2247          * select delayed ref of type BTRFS_ADD_DELAYED_REF first.
2248          * this prevents ref count from going down to zero when
2249          * there still are pending delayed ref.
2250          */
2251         node = rb_prev(&head->node.rb_node);
2252         while (1) {
2253                 if (!node)
2254                         break;
2255                 ref = rb_entry(node, struct btrfs_delayed_ref_node,
2256                                 rb_node);
2257                 if (ref->bytenr != head->node.bytenr)
2258                         break;
2259                 if (ref->action == action)
2260                         return ref;
2261                 node = rb_prev(node);
2262         }
2263         if (action == BTRFS_ADD_DELAYED_REF) {
2264                 action = BTRFS_DROP_DELAYED_REF;
2265                 goto again;
2266         }
2267         return NULL;
2268 }
2269
2270 /*
2271  * Returns 0 on success or if called with an already aborted transaction.
2272  * Returns -ENOMEM or -EIO on failure and will abort the transaction.
2273  */
2274 static noinline int run_clustered_refs(struct btrfs_trans_handle *trans,
2275                                        struct btrfs_root *root,
2276                                        struct list_head *cluster)
2277 {
2278         struct btrfs_delayed_ref_root *delayed_refs;
2279         struct btrfs_delayed_ref_node *ref;
2280         struct btrfs_delayed_ref_head *locked_ref = NULL;
2281         struct btrfs_delayed_extent_op *extent_op;
2282         struct btrfs_fs_info *fs_info = root->fs_info;
2283         int ret;
2284         int count = 0;
2285         int must_insert_reserved = 0;
2286
2287         delayed_refs = &trans->transaction->delayed_refs;
2288         while (1) {
2289                 if (!locked_ref) {
2290                         /* pick a new head ref from the cluster list */
2291                         if (list_empty(cluster))
2292                                 break;
2293
2294                         locked_ref = list_entry(cluster->next,
2295                                      struct btrfs_delayed_ref_head, cluster);
2296
2297                         /* grab the lock that says we are going to process
2298                          * all the refs for this head */
2299                         ret = btrfs_delayed_ref_lock(trans, locked_ref);
2300
2301                         /*
2302                          * we may have dropped the spin lock to get the head
2303                          * mutex lock, and that might have given someone else
2304                          * time to free the head.  If that's true, it has been
2305                          * removed from our list and we can move on.
2306                          */
2307                         if (ret == -EAGAIN) {
2308                                 locked_ref = NULL;
2309                                 count++;
2310                                 continue;
2311                         }
2312                 }
2313
2314                 /*
2315                  * We need to try and merge add/drops of the same ref since we
2316                  * can run into issues with relocate dropping the implicit ref
2317                  * and then it being added back again before the drop can
2318                  * finish.  If we merged anything we need to re-loop so we can
2319                  * get a good ref.
2320                  */
2321                 btrfs_merge_delayed_refs(trans, fs_info, delayed_refs,
2322                                          locked_ref);
2323
2324                 /*
2325                  * locked_ref is the head node, so we have to go one
2326                  * node back for any delayed ref updates
2327                  */
2328                 ref = select_delayed_ref(locked_ref);
2329
2330                 if (ref && ref->seq &&
2331                     btrfs_check_delayed_seq(fs_info, delayed_refs, ref->seq)) {
2332                         /*
2333                          * there are still refs with lower seq numbers in the
2334                          * process of being added. Don't run this ref yet.
2335                          */
2336                         list_del_init(&locked_ref->cluster);
2337                         btrfs_delayed_ref_unlock(locked_ref);
2338                         locked_ref = NULL;
2339                         delayed_refs->num_heads_ready++;
2340                         spin_unlock(&delayed_refs->lock);
2341                         cond_resched();
2342                         spin_lock(&delayed_refs->lock);
2343                         continue;
2344                 }
2345
2346                 /*
2347                  * record the must insert reserved flag before we
2348                  * drop the spin lock.
2349                  */
2350                 must_insert_reserved = locked_ref->must_insert_reserved;
2351                 locked_ref->must_insert_reserved = 0;
2352
2353                 extent_op = locked_ref->extent_op;
2354                 locked_ref->extent_op = NULL;
2355
2356                 if (!ref) {
2357                         /* All delayed refs have been processed, Go ahead
2358                          * and send the head node to run_one_delayed_ref,
2359                          * so that any accounting fixes can happen
2360                          */
2361                         ref = &locked_ref->node;
2362
2363                         if (extent_op && must_insert_reserved) {
2364                                 btrfs_free_delayed_extent_op(extent_op);
2365                                 extent_op = NULL;
2366                         }
2367
2368                         if (extent_op) {
2369                                 spin_unlock(&delayed_refs->lock);
2370
2371                                 ret = run_delayed_extent_op(trans, root,
2372                                                             ref, extent_op);
2373                                 btrfs_free_delayed_extent_op(extent_op);
2374
2375                                 if (ret) {
2376                                         btrfs_debug(fs_info, "run_delayed_extent_op returned %d", ret);
2377                                         spin_lock(&delayed_refs->lock);
2378                                         btrfs_delayed_ref_unlock(locked_ref);
2379                                         return ret;
2380                                 }
2381
2382                                 goto next;
2383                         }
2384                 }
2385
2386                 ref->in_tree = 0;
2387                 rb_erase(&ref->rb_node, &delayed_refs->root);
2388                 delayed_refs->num_entries--;
2389                 if (!btrfs_delayed_ref_is_head(ref)) {
2390                         /*
2391                          * when we play the delayed ref, also correct the
2392                          * ref_mod on head
2393                          */
2394                         switch (ref->action) {
2395                         case BTRFS_ADD_DELAYED_REF:
2396                         case BTRFS_ADD_DELAYED_EXTENT:
2397                                 locked_ref->node.ref_mod -= ref->ref_mod;
2398                                 break;
2399                         case BTRFS_DROP_DELAYED_REF:
2400                                 locked_ref->node.ref_mod += ref->ref_mod;
2401                                 break;
2402                         default:
2403                                 WARN_ON(1);
2404                         }
2405                 }
2406                 spin_unlock(&delayed_refs->lock);
2407
2408                 ret = run_one_delayed_ref(trans, root, ref, extent_op,
2409                                           must_insert_reserved);
2410
2411                 btrfs_free_delayed_extent_op(extent_op);
2412                 if (ret) {
2413                         btrfs_delayed_ref_unlock(locked_ref);
2414                         btrfs_put_delayed_ref(ref);
2415                         btrfs_debug(fs_info, "run_one_delayed_ref returned %d", ret);
2416                         spin_lock(&delayed_refs->lock);
2417                         return ret;
2418                 }
2419
2420                 /*
2421                  * If this node is a head, that means all the refs in this head
2422                  * have been dealt with, and we will pick the next head to deal
2423                  * with, so we must unlock the head and drop it from the cluster
2424                  * list before we release it.
2425                  */
2426                 if (btrfs_delayed_ref_is_head(ref)) {
2427                         list_del_init(&locked_ref->cluster);
2428                         btrfs_delayed_ref_unlock(locked_ref);
2429                         locked_ref = NULL;
2430                 }
2431                 btrfs_put_delayed_ref(ref);
2432                 count++;
2433 next:
2434                 cond_resched();
2435                 spin_lock(&delayed_refs->lock);
2436         }
2437         return count;
2438 }
2439
2440 #ifdef SCRAMBLE_DELAYED_REFS
2441 /*
2442  * Normally delayed refs get processed in ascending bytenr order. This
2443  * correlates in most cases to the order added. To expose dependencies on this
2444  * order, we start to process the tree in the middle instead of the beginning
2445  */
2446 static u64 find_middle(struct rb_root *root)
2447 {
2448         struct rb_node *n = root->rb_node;
2449         struct btrfs_delayed_ref_node *entry;
2450         int alt = 1;
2451         u64 middle;
2452         u64 first = 0, last = 0;
2453
2454         n = rb_first(root);
2455         if (n) {
2456                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2457                 first = entry->bytenr;
2458         }
2459         n = rb_last(root);
2460         if (n) {
2461                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2462                 last = entry->bytenr;
2463         }
2464         n = root->rb_node;
2465
2466         while (n) {
2467                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2468                 WARN_ON(!entry->in_tree);
2469
2470                 middle = entry->bytenr;
2471
2472                 if (alt)
2473                         n = n->rb_left;
2474                 else
2475                         n = n->rb_right;
2476
2477                 alt = 1 - alt;
2478         }
2479         return middle;
2480 }
2481 #endif
2482
2483 int btrfs_delayed_refs_qgroup_accounting(struct btrfs_trans_handle *trans,
2484                                          struct btrfs_fs_info *fs_info)
2485 {
2486         struct qgroup_update *qgroup_update;
2487         int ret = 0;
2488
2489         if (list_empty(&trans->qgroup_ref_list) !=
2490             !trans->delayed_ref_elem.seq) {
2491                 /* list without seq or seq without list */
2492                 btrfs_err(fs_info,
2493                         "qgroup accounting update error, list is%s empty, seq is %#x.%x",
2494                         list_empty(&trans->qgroup_ref_list) ? "" : " not",
2495                         (u32)(trans->delayed_ref_elem.seq >> 32),
2496                         (u32)trans->delayed_ref_elem.seq);
2497                 BUG();
2498         }
2499
2500         if (!trans->delayed_ref_elem.seq)
2501                 return 0;
2502
2503         while (!list_empty(&trans->qgroup_ref_list)) {
2504                 qgroup_update = list_first_entry(&trans->qgroup_ref_list,
2505                                                  struct qgroup_update, list);
2506                 list_del(&qgroup_update->list);
2507                 if (!ret)
2508                         ret = btrfs_qgroup_account_ref(
2509                                         trans, fs_info, qgroup_update->node,
2510                                         qgroup_update->extent_op);
2511                 kfree(qgroup_update);
2512         }
2513
2514         btrfs_put_tree_mod_seq(fs_info, &trans->delayed_ref_elem);
2515
2516         return ret;
2517 }
2518
2519 static int refs_newer(struct btrfs_delayed_ref_root *delayed_refs, int seq,
2520                       int count)
2521 {
2522         int val = atomic_read(&delayed_refs->ref_seq);
2523
2524         if (val < seq || val >= seq + count)
2525                 return 1;
2526         return 0;
2527 }
2528
2529 static inline u64 heads_to_leaves(struct btrfs_root *root, u64 heads)
2530 {
2531         u64 num_bytes;
2532
2533         num_bytes = heads * (sizeof(struct btrfs_extent_item) +
2534                              sizeof(struct btrfs_extent_inline_ref));
2535         if (!btrfs_fs_incompat(root->fs_info, SKINNY_METADATA))
2536                 num_bytes += heads * sizeof(struct btrfs_tree_block_info);
2537
2538         /*
2539          * We don't ever fill up leaves all the way so multiply by 2 just to be
2540          * closer to what we're really going to want to ouse.
2541          */
2542         return div64_u64(num_bytes, BTRFS_LEAF_DATA_SIZE(root));
2543 }
2544
2545 int btrfs_should_throttle_delayed_refs(struct btrfs_trans_handle *trans,
2546                                        struct btrfs_root *root)
2547 {
2548         struct btrfs_block_rsv *global_rsv;
2549         u64 num_heads = trans->transaction->delayed_refs.num_heads_ready;
2550         u64 num_bytes;
2551         int ret = 0;
2552
2553         num_bytes = btrfs_calc_trans_metadata_size(root, 1);
2554         num_heads = heads_to_leaves(root, num_heads);
2555         if (num_heads > 1)
2556                 num_bytes += (num_heads - 1) * root->leafsize;
2557         num_bytes <<= 1;
2558         global_rsv = &root->fs_info->global_block_rsv;
2559
2560         /*
2561          * If we can't allocate any more chunks lets make sure we have _lots_ of
2562          * wiggle room since running delayed refs can create more delayed refs.
2563          */
2564         if (global_rsv->space_info->full)
2565                 num_bytes <<= 1;
2566
2567         spin_lock(&global_rsv->lock);
2568         if (global_rsv->reserved <= num_bytes)
2569                 ret = 1;
2570         spin_unlock(&global_rsv->lock);
2571         return ret;
2572 }
2573
2574 /*
2575  * this starts processing the delayed reference count updates and
2576  * extent insertions we have queued up so far.  count can be
2577  * 0, which means to process everything in the tree at the start
2578  * of the run (but not newly added entries), or it can be some target
2579  * number you'd like to process.
2580  *
2581  * Returns 0 on success or if called with an aborted transaction
2582  * Returns <0 on error and aborts the transaction
2583  */
2584 int btrfs_run_delayed_refs(struct btrfs_trans_handle *trans,
2585                            struct btrfs_root *root, unsigned long count)
2586 {
2587         struct rb_node *node;
2588         struct btrfs_delayed_ref_root *delayed_refs;
2589         struct btrfs_delayed_ref_node *ref;
2590         struct list_head cluster;
2591         int ret;
2592         u64 delayed_start;
2593         int run_all = count == (unsigned long)-1;
2594         int run_most = 0;
2595         int loops;
2596
2597         /* We'll clean this up in btrfs_cleanup_transaction */
2598         if (trans->aborted)
2599                 return 0;
2600
2601         if (root == root->fs_info->extent_root)
2602                 root = root->fs_info->tree_root;
2603
2604         btrfs_delayed_refs_qgroup_accounting(trans, root->fs_info);
2605
2606         delayed_refs = &trans->transaction->delayed_refs;
2607         INIT_LIST_HEAD(&cluster);
2608         if (count == 0) {
2609                 count = delayed_refs->num_entries * 2;
2610                 run_most = 1;
2611         }
2612
2613         if (!run_all && !run_most) {
2614                 int old;
2615                 int seq = atomic_read(&delayed_refs->ref_seq);
2616
2617 progress:
2618                 old = atomic_cmpxchg(&delayed_refs->procs_running_refs, 0, 1);
2619                 if (old) {
2620                         DEFINE_WAIT(__wait);
2621                         if (delayed_refs->flushing ||
2622                             !btrfs_should_throttle_delayed_refs(trans, root))
2623                                 return 0;
2624
2625                         prepare_to_wait(&delayed_refs->wait, &__wait,
2626                                         TASK_UNINTERRUPTIBLE);
2627
2628                         old = atomic_cmpxchg(&delayed_refs->procs_running_refs, 0, 1);
2629                         if (old) {
2630                                 schedule();
2631                                 finish_wait(&delayed_refs->wait, &__wait);
2632
2633                                 if (!refs_newer(delayed_refs, seq, 256))
2634                                         goto progress;
2635                                 else
2636                                         return 0;
2637                         } else {
2638                                 finish_wait(&delayed_refs->wait, &__wait);
2639                                 goto again;
2640                         }
2641                 }
2642
2643         } else {
2644                 atomic_inc(&delayed_refs->procs_running_refs);
2645         }
2646
2647 again:
2648         loops = 0;
2649         spin_lock(&delayed_refs->lock);
2650
2651 #ifdef SCRAMBLE_DELAYED_REFS
2652         delayed_refs->run_delayed_start = find_middle(&delayed_refs->root);
2653 #endif
2654
2655         while (1) {
2656                 if (!(run_all || run_most) &&
2657                     !btrfs_should_throttle_delayed_refs(trans, root))
2658                         break;
2659
2660                 /*
2661                  * go find something we can process in the rbtree.  We start at
2662                  * the beginning of the tree, and then build a cluster
2663                  * of refs to process starting at the first one we are able to
2664                  * lock
2665                  */
2666                 delayed_start = delayed_refs->run_delayed_start;
2667                 ret = btrfs_find_ref_cluster(trans, &cluster,
2668                                              delayed_refs->run_delayed_start);
2669                 if (ret)
2670                         break;
2671
2672                 ret = run_clustered_refs(trans, root, &cluster);
2673                 if (ret < 0) {
2674                         btrfs_release_ref_cluster(&cluster);
2675                         spin_unlock(&delayed_refs->lock);
2676                         btrfs_abort_transaction(trans, root, ret);
2677                         atomic_dec(&delayed_refs->procs_running_refs);
2678                         wake_up(&delayed_refs->wait);
2679                         return ret;
2680                 }
2681
2682                 atomic_add(ret, &delayed_refs->ref_seq);
2683
2684                 count -= min_t(unsigned long, ret, count);
2685
2686                 if (count == 0)
2687                         break;
2688
2689                 if (delayed_start >= delayed_refs->run_delayed_start) {
2690                         if (loops == 0) {
2691                                 /*
2692                                  * btrfs_find_ref_cluster looped. let's do one
2693                                  * more cycle. if we don't run any delayed ref
2694                                  * during that cycle (because we can't because
2695                                  * all of them are blocked), bail out.
2696                                  */
2697                                 loops = 1;
2698                         } else {
2699                                 /*
2700                                  * no runnable refs left, stop trying
2701                                  */
2702                                 BUG_ON(run_all);
2703                                 break;
2704                         }
2705                 }
2706                 if (ret) {
2707                         /* refs were run, let's reset staleness detection */
2708                         loops = 0;
2709                 }
2710         }
2711
2712         if (run_all) {
2713                 if (!list_empty(&trans->new_bgs)) {
2714                         spin_unlock(&delayed_refs->lock);
2715                         btrfs_create_pending_block_groups(trans, root);
2716                         spin_lock(&delayed_refs->lock);
2717                 }
2718
2719                 node = rb_first(&delayed_refs->root);
2720                 if (!node)
2721                         goto out;
2722                 count = (unsigned long)-1;
2723
2724                 while (node) {
2725                         ref = rb_entry(node, struct btrfs_delayed_ref_node,
2726                                        rb_node);
2727                         if (btrfs_delayed_ref_is_head(ref)) {
2728                                 struct btrfs_delayed_ref_head *head;
2729
2730                                 head = btrfs_delayed_node_to_head(ref);
2731                                 atomic_inc(&ref->refs);
2732
2733                                 spin_unlock(&delayed_refs->lock);
2734                                 /*
2735                                  * Mutex was contended, block until it's
2736                                  * released and try again
2737                                  */
2738                                 mutex_lock(&head->mutex);
2739                                 mutex_unlock(&head->mutex);
2740
2741                                 btrfs_put_delayed_ref(ref);
2742                                 cond_resched();
2743                                 goto again;
2744                         }
2745                         node = rb_next(node);
2746                 }
2747                 spin_unlock(&delayed_refs->lock);
2748                 schedule_timeout(1);
2749                 goto again;
2750         }
2751 out:
2752         atomic_dec(&delayed_refs->procs_running_refs);
2753         smp_mb();
2754         if (waitqueue_active(&delayed_refs->wait))
2755                 wake_up(&delayed_refs->wait);
2756
2757         spin_unlock(&delayed_refs->lock);
2758         assert_qgroups_uptodate(trans);
2759         return 0;
2760 }
2761
2762 int btrfs_set_disk_extent_flags(struct btrfs_trans_handle *trans,
2763                                 struct btrfs_root *root,
2764                                 u64 bytenr, u64 num_bytes, u64 flags,
2765                                 int level, int is_data)
2766 {
2767         struct btrfs_delayed_extent_op *extent_op;
2768         int ret;
2769
2770         extent_op = btrfs_alloc_delayed_extent_op();
2771         if (!extent_op)
2772                 return -ENOMEM;
2773
2774         extent_op->flags_to_set = flags;
2775         extent_op->update_flags = 1;
2776         extent_op->update_key = 0;
2777         extent_op->is_data = is_data ? 1 : 0;
2778         extent_op->level = level;
2779
2780         ret = btrfs_add_delayed_extent_op(root->fs_info, trans, bytenr,
2781                                           num_bytes, extent_op);
2782         if (ret)
2783                 btrfs_free_delayed_extent_op(extent_op);
2784         return ret;
2785 }
2786
2787 static noinline int check_delayed_ref(struct btrfs_trans_handle *trans,
2788                                       struct btrfs_root *root,
2789                                       struct btrfs_path *path,
2790                                       u64 objectid, u64 offset, u64 bytenr)
2791 {
2792         struct btrfs_delayed_ref_head *head;
2793         struct btrfs_delayed_ref_node *ref;
2794         struct btrfs_delayed_data_ref *data_ref;
2795         struct btrfs_delayed_ref_root *delayed_refs;
2796         struct rb_node *node;
2797         int ret = 0;
2798
2799         ret = -ENOENT;
2800         delayed_refs = &trans->transaction->delayed_refs;
2801         spin_lock(&delayed_refs->lock);
2802         head = btrfs_find_delayed_ref_head(trans, bytenr);
2803         if (!head)
2804                 goto out;
2805
2806         if (!mutex_trylock(&head->mutex)) {
2807                 atomic_inc(&head->node.refs);
2808                 spin_unlock(&delayed_refs->lock);
2809
2810                 btrfs_release_path(path);
2811
2812                 /*
2813                  * Mutex was contended, block until it's released and let
2814                  * caller try again
2815                  */
2816                 mutex_lock(&head->mutex);
2817                 mutex_unlock(&head->mutex);
2818                 btrfs_put_delayed_ref(&head->node);
2819                 return -EAGAIN;
2820         }
2821
2822         node = rb_prev(&head->node.rb_node);
2823         if (!node)
2824                 goto out_unlock;
2825
2826         ref = rb_entry(node, struct btrfs_delayed_ref_node, rb_node);
2827
2828         if (ref->bytenr != bytenr)
2829                 goto out_unlock;
2830
2831         ret = 1;
2832         if (ref->type != BTRFS_EXTENT_DATA_REF_KEY)
2833                 goto out_unlock;
2834
2835         data_ref = btrfs_delayed_node_to_data_ref(ref);
2836
2837         node = rb_prev(node);
2838         if (node) {
2839                 int seq = ref->seq;
2840
2841                 ref = rb_entry(node, struct btrfs_delayed_ref_node, rb_node);
2842                 if (ref->bytenr == bytenr && ref->seq == seq)
2843                         goto out_unlock;
2844         }
2845
2846         if (data_ref->root != root->root_key.objectid ||
2847             data_ref->objectid != objectid || data_ref->offset != offset)
2848                 goto out_unlock;
2849
2850         ret = 0;
2851 out_unlock:
2852         mutex_unlock(&head->mutex);
2853 out:
2854         spin_unlock(&delayed_refs->lock);
2855         return ret;
2856 }
2857
2858 static noinline int check_committed_ref(struct btrfs_trans_handle *trans,
2859                                         struct btrfs_root *root,
2860                                         struct btrfs_path *path,
2861                                         u64 objectid, u64 offset, u64 bytenr)
2862 {
2863         struct btrfs_root *extent_root = root->fs_info->extent_root;
2864         struct extent_buffer *leaf;
2865         struct btrfs_extent_data_ref *ref;
2866         struct btrfs_extent_inline_ref *iref;
2867         struct btrfs_extent_item *ei;
2868         struct btrfs_key key;
2869         u32 item_size;
2870         int ret;
2871
2872         key.objectid = bytenr;
2873         key.offset = (u64)-1;
2874         key.type = BTRFS_EXTENT_ITEM_KEY;
2875
2876         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
2877         if (ret < 0)
2878                 goto out;
2879         BUG_ON(ret == 0); /* Corruption */
2880
2881         ret = -ENOENT;
2882         if (path->slots[0] == 0)
2883                 goto out;
2884
2885         path->slots[0]--;
2886         leaf = path->nodes[0];
2887         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
2888
2889         if (key.objectid != bytenr || key.type != BTRFS_EXTENT_ITEM_KEY)
2890                 goto out;
2891
2892         ret = 1;
2893         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2894 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
2895         if (item_size < sizeof(*ei)) {
2896                 WARN_ON(item_size != sizeof(struct btrfs_extent_item_v0));
2897                 goto out;
2898         }
2899 #endif
2900         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2901
2902         if (item_size != sizeof(*ei) +
2903             btrfs_extent_inline_ref_size(BTRFS_EXTENT_DATA_REF_KEY))
2904                 goto out;
2905
2906         if (btrfs_extent_generation(leaf, ei) <=
2907             btrfs_root_last_snapshot(&root->root_item))
2908                 goto out;
2909
2910         iref = (struct btrfs_extent_inline_ref *)(ei + 1);
2911         if (btrfs_extent_inline_ref_type(leaf, iref) !=
2912             BTRFS_EXTENT_DATA_REF_KEY)
2913                 goto out;
2914
2915         ref = (struct btrfs_extent_data_ref *)(&iref->offset);
2916         if (btrfs_extent_refs(leaf, ei) !=
2917             btrfs_extent_data_ref_count(leaf, ref) ||
2918             btrfs_extent_data_ref_root(leaf, ref) !=
2919             root->root_key.objectid ||
2920             btrfs_extent_data_ref_objectid(leaf, ref) != objectid ||
2921             btrfs_extent_data_ref_offset(leaf, ref) != offset)
2922                 goto out;
2923
2924         ret = 0;
2925 out:
2926         return ret;
2927 }
2928
2929 int btrfs_cross_ref_exist(struct btrfs_trans_handle *trans,
2930                           struct btrfs_root *root,
2931                           u64 objectid, u64 offset, u64 bytenr)
2932 {
2933         struct btrfs_path *path;
2934         int ret;
2935         int ret2;
2936
2937         path = btrfs_alloc_path();
2938         if (!path)
2939                 return -ENOENT;
2940
2941         do {
2942                 ret = check_committed_ref(trans, root, path, objectid,
2943                                           offset, bytenr);
2944                 if (ret && ret != -ENOENT)
2945                         goto out;
2946
2947                 ret2 = check_delayed_ref(trans, root, path, objectid,
2948                                          offset, bytenr);
2949         } while (ret2 == -EAGAIN);
2950
2951         if (ret2 && ret2 != -ENOENT) {
2952                 ret = ret2;
2953                 goto out;
2954         }
2955
2956         if (ret != -ENOENT || ret2 != -ENOENT)
2957                 ret = 0;
2958 out:
2959         btrfs_free_path(path);
2960         if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID)
2961                 WARN_ON(ret > 0);
2962         return ret;
2963 }
2964
2965 static int __btrfs_mod_ref(struct btrfs_trans_handle *trans,
2966                            struct btrfs_root *root,
2967                            struct extent_buffer *buf,
2968                            int full_backref, int inc, int for_cow)
2969 {
2970         u64 bytenr;
2971         u64 num_bytes;
2972         u64 parent;
2973         u64 ref_root;
2974         u32 nritems;
2975         struct btrfs_key key;
2976         struct btrfs_file_extent_item *fi;
2977         int i;
2978         int level;
2979         int ret = 0;
2980         int (*process_func)(struct btrfs_trans_handle *, struct btrfs_root *,
2981                             u64, u64, u64, u64, u64, u64, int);
2982
2983         ref_root = btrfs_header_owner(buf);
2984         nritems = btrfs_header_nritems(buf);
2985         level = btrfs_header_level(buf);
2986
2987         if (!root->ref_cows && level == 0)
2988                 return 0;
2989
2990         if (inc)
2991                 process_func = btrfs_inc_extent_ref;
2992         else
2993                 process_func = btrfs_free_extent;
2994
2995         if (full_backref)
2996                 parent = buf->start;
2997         else
2998                 parent = 0;
2999
3000         for (i = 0; i < nritems; i++) {
3001                 if (level == 0) {
3002                         btrfs_item_key_to_cpu(buf, &key, i);
3003                         if (btrfs_key_type(&key) != BTRFS_EXTENT_DATA_KEY)
3004                                 continue;
3005                         fi = btrfs_item_ptr(buf, i,
3006                                             struct btrfs_file_extent_item);
3007                         if (btrfs_file_extent_type(buf, fi) ==
3008                             BTRFS_FILE_EXTENT_INLINE)
3009                                 continue;
3010                         bytenr = btrfs_file_extent_disk_bytenr(buf, fi);
3011                         if (bytenr == 0)
3012                                 continue;
3013
3014                         num_bytes = btrfs_file_extent_disk_num_bytes(buf, fi);
3015                         key.offset -= btrfs_file_extent_offset(buf, fi);
3016                         ret = process_func(trans, root, bytenr, num_bytes,
3017                                            parent, ref_root, key.objectid,
3018                                            key.offset, for_cow);
3019                         if (ret)
3020                                 goto fail;
3021                 } else {
3022                         bytenr = btrfs_node_blockptr(buf, i);
3023                         num_bytes = btrfs_level_size(root, level - 1);
3024                         ret = process_func(trans, root, bytenr, num_bytes,
3025                                            parent, ref_root, level - 1, 0,
3026                                            for_cow);
3027                         if (ret)
3028                                 goto fail;
3029                 }
3030         }
3031         return 0;
3032 fail:
3033         return ret;
3034 }
3035
3036 int btrfs_inc_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
3037                   struct extent_buffer *buf, int full_backref, int for_cow)
3038 {
3039         return __btrfs_mod_ref(trans, root, buf, full_backref, 1, for_cow);
3040 }
3041
3042 int btrfs_dec_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
3043                   struct extent_buffer *buf, int full_backref, int for_cow)
3044 {
3045         return __btrfs_mod_ref(trans, root, buf, full_backref, 0, for_cow);
3046 }
3047
3048 static int write_one_cache_group(struct btrfs_trans_handle *trans,
3049                                  struct btrfs_root *root,
3050                                  struct btrfs_path *path,
3051                                  struct btrfs_block_group_cache *cache)
3052 {
3053         int ret;
3054         struct btrfs_root *extent_root = root->fs_info->extent_root;
3055         unsigned long bi;
3056         struct extent_buffer *leaf;
3057
3058         ret = btrfs_search_slot(trans, extent_root, &cache->key, path, 0, 1);
3059         if (ret < 0)
3060                 goto fail;
3061         BUG_ON(ret); /* Corruption */
3062
3063         leaf = path->nodes[0];
3064         bi = btrfs_item_ptr_offset(leaf, path->slots[0]);
3065         write_extent_buffer(leaf, &cache->item, bi, sizeof(cache->item));
3066         btrfs_mark_buffer_dirty(leaf);
3067         btrfs_release_path(path);
3068 fail:
3069         if (ret) {
3070                 btrfs_abort_transaction(trans, root, ret);
3071                 return ret;
3072         }
3073         return 0;
3074
3075 }
3076
3077 static struct btrfs_block_group_cache *
3078 next_block_group(struct btrfs_root *root,
3079                  struct btrfs_block_group_cache *cache)
3080 {
3081         struct rb_node *node;
3082         spin_lock(&root->fs_info->block_group_cache_lock);
3083         node = rb_next(&cache->cache_node);
3084         btrfs_put_block_group(cache);
3085         if (node) {
3086                 cache = rb_entry(node, struct btrfs_block_group_cache,
3087                                  cache_node);
3088                 btrfs_get_block_group(cache);
3089         } else
3090                 cache = NULL;
3091         spin_unlock(&root->fs_info->block_group_cache_lock);
3092         return cache;
3093 }
3094
3095 static int cache_save_setup(struct btrfs_block_group_cache *block_group,
3096                             struct btrfs_trans_handle *trans,
3097                             struct btrfs_path *path)
3098 {
3099         struct btrfs_root *root = block_group->fs_info->tree_root;
3100         struct inode *inode = NULL;
3101         u64 alloc_hint = 0;
3102         int dcs = BTRFS_DC_ERROR;
3103         int num_pages = 0;
3104         int retries = 0;
3105         int ret = 0;
3106
3107         /*
3108          * If this block group is smaller than 100 megs don't bother caching the
3109          * block group.
3110          */
3111         if (block_group->key.offset < (100 * 1024 * 1024)) {
3112                 spin_lock(&block_group->lock);
3113                 block_group->disk_cache_state = BTRFS_DC_WRITTEN;
3114                 spin_unlock(&block_group->lock);
3115                 return 0;
3116         }
3117
3118 again:
3119         inode = lookup_free_space_inode(root, block_group, path);
3120         if (IS_ERR(inode) && PTR_ERR(inode) != -ENOENT) {
3121                 ret = PTR_ERR(inode);
3122                 btrfs_release_path(path);
3123                 goto out;
3124         }
3125
3126         if (IS_ERR(inode)) {
3127                 BUG_ON(retries);
3128                 retries++;
3129
3130                 if (block_group->ro)
3131                         goto out_free;
3132
3133                 ret = create_free_space_inode(root, trans, block_group, path);
3134                 if (ret)
3135                         goto out_free;
3136                 goto again;
3137         }
3138
3139         /* We've already setup this transaction, go ahead and exit */
3140         if (block_group->cache_generation == trans->transid &&
3141             i_size_read(inode)) {
3142                 dcs = BTRFS_DC_SETUP;
3143                 goto out_put;
3144         }
3145
3146         /*
3147          * We want to set the generation to 0, that way if anything goes wrong
3148          * from here on out we know not to trust this cache when we load up next
3149          * time.
3150          */
3151         BTRFS_I(inode)->generation = 0;
3152         ret = btrfs_update_inode(trans, root, inode);
3153         WARN_ON(ret);
3154
3155         if (i_size_read(inode) > 0) {
3156                 ret = btrfs_check_trunc_cache_free_space(root,
3157                                         &root->fs_info->global_block_rsv);
3158                 if (ret)
3159                         goto out_put;
3160
3161                 ret = btrfs_truncate_free_space_cache(root, trans, path,
3162                                                       inode);
3163                 if (ret)
3164                         goto out_put;
3165         }
3166
3167         spin_lock(&block_group->lock);
3168         if (block_group->cached != BTRFS_CACHE_FINISHED ||
3169             !btrfs_test_opt(root, SPACE_CACHE)) {
3170                 /*
3171                  * don't bother trying to write stuff out _if_
3172                  * a) we're not cached,
3173                  * b) we're with nospace_cache mount option.
3174                  */
3175                 dcs = BTRFS_DC_WRITTEN;
3176                 spin_unlock(&block_group->lock);
3177                 goto out_put;
3178         }
3179         spin_unlock(&block_group->lock);
3180
3181         /*
3182          * Try to preallocate enough space based on how big the block group is.
3183          * Keep in mind this has to include any pinned space which could end up
3184          * taking up quite a bit since it's not folded into the other space
3185          * cache.
3186          */
3187         num_pages = (int)div64_u64(block_group->key.offset, 256 * 1024 * 1024);
3188         if (!num_pages)
3189                 num_pages = 1;
3190
3191         num_pages *= 16;
3192         num_pages *= PAGE_CACHE_SIZE;
3193
3194         ret = btrfs_check_data_free_space(inode, num_pages);
3195         if (ret)
3196                 goto out_put;
3197
3198         ret = btrfs_prealloc_file_range_trans(inode, trans, 0, 0, num_pages,
3199                                               num_pages, num_pages,
3200                                               &alloc_hint);
3201         if (!ret)
3202                 dcs = BTRFS_DC_SETUP;
3203         btrfs_free_reserved_data_space(inode, num_pages);
3204
3205 out_put:
3206         iput(inode);
3207 out_free:
3208         btrfs_release_path(path);
3209 out:
3210         spin_lock(&block_group->lock);
3211         if (!ret && dcs == BTRFS_DC_SETUP)
3212                 block_group->cache_generation = trans->transid;
3213         block_group->disk_cache_state = dcs;
3214         spin_unlock(&block_group->lock);
3215
3216         return ret;
3217 }
3218
3219 int btrfs_write_dirty_block_groups(struct btrfs_trans_handle *trans,
3220                                    struct btrfs_root *root)
3221 {
3222         struct btrfs_block_group_cache *cache;
3223         int err = 0;
3224         struct btrfs_path *path;
3225         u64 last = 0;
3226
3227         path = btrfs_alloc_path();
3228         if (!path)
3229                 return -ENOMEM;
3230
3231 again:
3232         while (1) {
3233                 cache = btrfs_lookup_first_block_group(root->fs_info, last);
3234                 while (cache) {
3235                         if (cache->disk_cache_state == BTRFS_DC_CLEAR)
3236                                 break;
3237                         cache = next_block_group(root, cache);
3238                 }
3239                 if (!cache) {
3240                         if (last == 0)
3241                                 break;
3242                         last = 0;
3243                         continue;
3244                 }
3245                 err = cache_save_setup(cache, trans, path);
3246                 last = cache->key.objectid + cache->key.offset;
3247                 btrfs_put_block_group(cache);
3248         }
3249
3250         while (1) {
3251                 if (last == 0) {
3252                         err = btrfs_run_delayed_refs(trans, root,
3253                                                      (unsigned long)-1);
3254                         if (err) /* File system offline */
3255                                 goto out;
3256                 }
3257
3258                 cache = btrfs_lookup_first_block_group(root->fs_info, last);
3259                 while (cache) {
3260                         if (cache->disk_cache_state == BTRFS_DC_CLEAR) {
3261                                 btrfs_put_block_group(cache);
3262                                 goto again;
3263                         }
3264
3265                         if (cache->dirty)
3266                                 break;
3267                         cache = next_block_group(root, cache);
3268                 }
3269                 if (!cache) {
3270                         if (last == 0)
3271                                 break;
3272                         last = 0;
3273                         continue;
3274                 }
3275
3276                 if (cache->disk_cache_state == BTRFS_DC_SETUP)
3277                         cache->disk_cache_state = BTRFS_DC_NEED_WRITE;
3278                 cache->dirty = 0;
3279                 last = cache->key.objectid + cache->key.offset;
3280
3281                 err = write_one_cache_group(trans, root, path, cache);
3282                 if (err) /* File system offline */
3283                         goto out;
3284
3285                 btrfs_put_block_group(cache);
3286         }
3287
3288         while (1) {
3289                 /*
3290                  * I don't think this is needed since we're just marking our
3291                  * preallocated extent as written, but just in case it can't
3292                  * hurt.
3293                  */
3294                 if (last == 0) {
3295                         err = btrfs_run_delayed_refs(trans, root,
3296                                                      (unsigned long)-1);
3297                         if (err) /* File system offline */
3298                                 goto out;
3299                 }
3300
3301                 cache = btrfs_lookup_first_block_group(root->fs_info, last);
3302                 while (cache) {
3303                         /*
3304                          * Really this shouldn't happen, but it could if we
3305                          * couldn't write the entire preallocated extent and
3306                          * splitting the extent resulted in a new block.
3307                          */
3308                         if (cache->dirty) {
3309                                 btrfs_put_block_group(cache);
3310                                 goto again;
3311                         }
3312                         if (cache->disk_cache_state == BTRFS_DC_NEED_WRITE)
3313                                 break;
3314                         cache = next_block_group(root, cache);
3315                 }
3316                 if (!cache) {
3317                         if (last == 0)
3318                                 break;
3319                         last = 0;
3320                         continue;
3321                 }
3322
3323                 err = btrfs_write_out_cache(root, trans, cache, path);
3324
3325                 /*
3326                  * If we didn't have an error then the cache state is still
3327                  * NEED_WRITE, so we can set it to WRITTEN.
3328                  */
3329                 if (!err && cache->disk_cache_state == BTRFS_DC_NEED_WRITE)
3330                         cache->disk_cache_state = BTRFS_DC_WRITTEN;
3331                 last = cache->key.objectid + cache->key.offset;
3332                 btrfs_put_block_group(cache);
3333         }
3334 out:
3335
3336         btrfs_free_path(path);
3337         return err;
3338 }
3339
3340 int btrfs_extent_readonly(struct btrfs_root *root, u64 bytenr)
3341 {
3342         struct btrfs_block_group_cache *block_group;
3343         int readonly = 0;
3344
3345         block_group = btrfs_lookup_block_group(root->fs_info, bytenr);
3346         if (!block_group || block_group->ro)
3347                 readonly = 1;
3348         if (block_group)
3349                 btrfs_put_block_group(block_group);
3350         return readonly;
3351 }
3352
3353 static int update_space_info(struct btrfs_fs_info *info, u64 flags,
3354                              u64 total_bytes, u64 bytes_used,
3355                              struct btrfs_space_info **space_info)
3356 {
3357         struct btrfs_space_info *found;
3358         int i;
3359         int factor;
3360
3361         if (flags & (BTRFS_BLOCK_GROUP_DUP | BTRFS_BLOCK_GROUP_RAID1 |
3362                      BTRFS_BLOCK_GROUP_RAID10))
3363                 factor = 2;
3364         else
3365                 factor = 1;
3366
3367         found = __find_space_info(info, flags);
3368         if (found) {
3369                 spin_lock(&found->lock);
3370                 found->total_bytes += total_bytes;
3371                 found->disk_total += total_bytes * factor;
3372                 found->bytes_used += bytes_used;
3373                 found->disk_used += bytes_used * factor;
3374                 found->full = 0;
3375                 spin_unlock(&found->lock);
3376                 *space_info = found;
3377                 return 0;
3378         }
3379         found = kzalloc(sizeof(*found), GFP_NOFS);
3380         if (!found)
3381                 return -ENOMEM;
3382
3383         for (i = 0; i < BTRFS_NR_RAID_TYPES; i++)
3384                 INIT_LIST_HEAD(&found->block_groups[i]);
3385         init_rwsem(&found->groups_sem);
3386         spin_lock_init(&found->lock);
3387         found->flags = flags & BTRFS_BLOCK_GROUP_TYPE_MASK;
3388         found->total_bytes = total_bytes;
3389         found->disk_total = total_bytes * factor;
3390         found->bytes_used = bytes_used;
3391         found->disk_used = bytes_used * factor;
3392         found->bytes_pinned = 0;
3393         found->bytes_reserved = 0;
3394         found->bytes_readonly = 0;
3395         found->bytes_may_use = 0;
3396         found->full = 0;
3397         found->force_alloc = CHUNK_ALLOC_NO_FORCE;
3398         found->chunk_alloc = 0;
3399         found->flush = 0;
3400         init_waitqueue_head(&found->wait);
3401         *space_info = found;
3402         list_add_rcu(&found->list, &info->space_info);
3403         if (flags & BTRFS_BLOCK_GROUP_DATA)
3404                 info->data_sinfo = found;
3405         return 0;
3406 }
3407
3408 static void set_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
3409 {
3410         u64 extra_flags = chunk_to_extended(flags) &
3411                                 BTRFS_EXTENDED_PROFILE_MASK;
3412
3413         write_seqlock(&fs_info->profiles_lock);
3414         if (flags & BTRFS_BLOCK_GROUP_DATA)
3415                 fs_info->avail_data_alloc_bits |= extra_flags;
3416         if (flags & BTRFS_BLOCK_GROUP_METADATA)
3417                 fs_info->avail_metadata_alloc_bits |= extra_flags;
3418         if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
3419                 fs_info->avail_system_alloc_bits |= extra_flags;
3420         write_sequnlock(&fs_info->profiles_lock);
3421 }
3422
3423 /*
3424  * returns target flags in extended format or 0 if restripe for this
3425  * chunk_type is not in progress
3426  *
3427  * should be called with either volume_mutex or balance_lock held
3428  */
3429 static u64 get_restripe_target(struct btrfs_fs_info *fs_info, u64 flags)
3430 {
3431         struct btrfs_balance_control *bctl = fs_info->balance_ctl;
3432         u64 target = 0;
3433
3434         if (!bctl)
3435                 return 0;
3436
3437         if (flags & BTRFS_BLOCK_GROUP_DATA &&
3438             bctl->data.flags & BTRFS_BALANCE_ARGS_CONVERT) {
3439                 target = BTRFS_BLOCK_GROUP_DATA | bctl->data.target;
3440         } else if (flags & BTRFS_BLOCK_GROUP_SYSTEM &&
3441                    bctl->sys.flags & BTRFS_BALANCE_ARGS_CONVERT) {
3442                 target = BTRFS_BLOCK_GROUP_SYSTEM | bctl->sys.target;
3443         } else if (flags & BTRFS_BLOCK_GROUP_METADATA &&
3444                    bctl->meta.flags & BTRFS_BALANCE_ARGS_CONVERT) {
3445                 target = BTRFS_BLOCK_GROUP_METADATA | bctl->meta.target;
3446         }
3447
3448         return target;
3449 }
3450
3451 /*
3452  * @flags: available profiles in extended format (see ctree.h)
3453  *
3454  * Returns reduced profile in chunk format.  If profile changing is in
3455  * progress (either running or paused) picks the target profile (if it's
3456  * already available), otherwise falls back to plain reducing.
3457  */
3458 static u64 btrfs_reduce_alloc_profile(struct btrfs_root *root, u64 flags)
3459 {
3460         /*
3461          * we add in the count of missing devices because we want
3462          * to make sure that any RAID levels on a degraded FS
3463          * continue to be honored.
3464          */
3465         u64 num_devices = root->fs_info->fs_devices->rw_devices +
3466                 root->fs_info->fs_devices->missing_devices;
3467         u64 target;
3468         u64 tmp;
3469
3470         /*
3471          * see if restripe for this chunk_type is in progress, if so
3472          * try to reduce to the target profile
3473          */
3474         spin_lock(&root->fs_info->balance_lock);
3475         target = get_restripe_target(root->fs_info, flags);
3476         if (target) {
3477                 /* pick target profile only if it's already available */
3478                 if ((flags & target) & BTRFS_EXTENDED_PROFILE_MASK) {
3479                         spin_unlock(&root->fs_info->balance_lock);
3480                         return extended_to_chunk(target);
3481                 }
3482         }
3483         spin_unlock(&root->fs_info->balance_lock);
3484
3485         /* First, mask out the RAID levels which aren't possible */
3486         if (num_devices == 1)
3487                 flags &= ~(BTRFS_BLOCK_GROUP_RAID1 | BTRFS_BLOCK_GROUP_RAID0 |
3488                            BTRFS_BLOCK_GROUP_RAID5);
3489         if (num_devices < 3)
3490                 flags &= ~BTRFS_BLOCK_GROUP_RAID6;
3491         if (num_devices < 4)
3492                 flags &= ~BTRFS_BLOCK_GROUP_RAID10;
3493
3494         tmp = flags & (BTRFS_BLOCK_GROUP_DUP | BTRFS_BLOCK_GROUP_RAID0 |
3495                        BTRFS_BLOCK_GROUP_RAID1 | BTRFS_BLOCK_GROUP_RAID5 |
3496                        BTRFS_BLOCK_GROUP_RAID6 | BTRFS_BLOCK_GROUP_RAID10);
3497         flags &= ~tmp;
3498
3499         if (tmp & BTRFS_BLOCK_GROUP_RAID6)
3500                 tmp = BTRFS_BLOCK_GROUP_RAID6;
3501         else if (tmp & BTRFS_BLOCK_GROUP_RAID5)
3502                 tmp = BTRFS_BLOCK_GROUP_RAID5;
3503         else if (tmp & BTRFS_BLOCK_GROUP_RAID10)
3504                 tmp = BTRFS_BLOCK_GROUP_RAID10;
3505         else if (tmp & BTRFS_BLOCK_GROUP_RAID1)
3506                 tmp = BTRFS_BLOCK_GROUP_RAID1;
3507         else if (tmp & BTRFS_BLOCK_GROUP_RAID0)
3508                 tmp = BTRFS_BLOCK_GROUP_RAID0;
3509
3510         return extended_to_chunk(flags | tmp);
3511 }
3512
3513 static u64 get_alloc_profile(struct btrfs_root *root, u64 flags)
3514 {
3515         unsigned seq;
3516
3517         do {
3518                 seq = read_seqbegin(&root->fs_info->profiles_lock);
3519
3520                 if (flags & BTRFS_BLOCK_GROUP_DATA)
3521                         flags |= root->fs_info->avail_data_alloc_bits;
3522                 else if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
3523                         flags |= root->fs_info->avail_system_alloc_bits;
3524                 else if (flags & BTRFS_BLOCK_GROUP_METADATA)
3525                         flags |= root->fs_info->avail_metadata_alloc_bits;
3526         } while (read_seqretry(&root->fs_info->profiles_lock, seq));
3527
3528         return btrfs_reduce_alloc_profile(root, flags);
3529 }
3530
3531 u64 btrfs_get_alloc_profile(struct btrfs_root *root, int data)
3532 {
3533         u64 flags;
3534         u64 ret;
3535
3536         if (data)
3537                 flags = BTRFS_BLOCK_GROUP_DATA;
3538         else if (root == root->fs_info->chunk_root)
3539                 flags = BTRFS_BLOCK_GROUP_SYSTEM;
3540         else
3541                 flags = BTRFS_BLOCK_GROUP_METADATA;
3542
3543         ret = get_alloc_profile(root, flags);
3544         return ret;
3545 }
3546
3547 /*
3548  * This will check the space that the inode allocates from to make sure we have
3549  * enough space for bytes.
3550  */
3551 int btrfs_check_data_free_space(struct inode *inode, u64 bytes)
3552 {
3553         struct btrfs_space_info *data_sinfo;
3554         struct btrfs_root *root = BTRFS_I(inode)->root;
3555         struct btrfs_fs_info *fs_info = root->fs_info;
3556         u64 used;
3557         int ret = 0, committed = 0, alloc_chunk = 1;
3558
3559         /* make sure bytes are sectorsize aligned */
3560         bytes = ALIGN(bytes, root->sectorsize);
3561
3562         if (root == root->fs_info->tree_root ||
3563             BTRFS_I(inode)->location.objectid == BTRFS_FREE_INO_OBJECTID) {
3564                 alloc_chunk = 0;
3565                 committed = 1;
3566         }
3567
3568         data_sinfo = fs_info->data_sinfo;
3569         if (!data_sinfo)
3570                 goto alloc;
3571
3572 again:
3573         /* make sure we have enough space to handle the data first */
3574         spin_lock(&data_sinfo->lock);
3575         used = data_sinfo->bytes_used + data_sinfo->bytes_reserved +
3576                 data_sinfo->bytes_pinned + data_sinfo->bytes_readonly +
3577                 data_sinfo->bytes_may_use;
3578
3579         if (used + bytes > data_sinfo->total_bytes) {
3580                 struct btrfs_trans_handle *trans;
3581
3582                 /*
3583                  * if we don't have enough free bytes in this space then we need
3584                  * to alloc a new chunk.
3585                  */
3586                 if (!data_sinfo->full && alloc_chunk) {
3587                         u64 alloc_target;
3588
3589                         data_sinfo->force_alloc = CHUNK_ALLOC_FORCE;
3590                         spin_unlock(&data_sinfo->lock);
3591 alloc:
3592                         alloc_target = btrfs_get_alloc_profile(root, 1);
3593                         trans = btrfs_join_transaction(root);
3594                         if (IS_ERR(trans))
3595                                 return PTR_ERR(trans);
3596
3597                         ret = do_chunk_alloc(trans, root->fs_info->extent_root,
3598                                              alloc_target,
3599                                              CHUNK_ALLOC_NO_FORCE);
3600                         btrfs_end_transaction(trans, root);
3601                         if (ret < 0) {
3602                                 if (ret != -ENOSPC)
3603                                         return ret;
3604                                 else
3605                                         goto commit_trans;
3606                         }
3607
3608                         if (!data_sinfo)
3609                                 data_sinfo = fs_info->data_sinfo;
3610
3611                         goto again;
3612                 }
3613
3614                 /*
3615                  * If we have less pinned bytes than we want to allocate then
3616                  * don't bother committing the transaction, it won't help us.
3617                  */
3618                 if (data_sinfo->bytes_pinned < bytes)
3619                         committed = 1;
3620                 spin_unlock(&data_sinfo->lock);
3621
3622                 /* commit the current transaction and try again */
3623 commit_trans:
3624                 if (!committed &&
3625                     !atomic_read(&root->fs_info->open_ioctl_trans)) {
3626                         committed = 1;
3627                         trans = btrfs_join_transaction(root);
3628                         if (IS_ERR(trans))
3629                                 return PTR_ERR(trans);
3630                         ret = btrfs_commit_transaction(trans, root);
3631                         if (ret)
3632                                 return ret;
3633                         goto again;
3634                 }
3635
3636                 return -ENOSPC;
3637         }
3638         data_sinfo->bytes_may_use += bytes;
3639         trace_btrfs_space_reservation(root->fs_info, "space_info",
3640                                       data_sinfo->flags, bytes, 1);
3641         spin_unlock(&data_sinfo->lock);
3642
3643         return 0;
3644 }
3645
3646 /*
3647  * Called if we need to clear a data reservation for this inode.
3648  */
3649 void btrfs_free_reserved_data_space(struct inode *inode, u64 bytes)
3650 {
3651         struct btrfs_root *root = BTRFS_I(inode)->root;
3652         struct btrfs_space_info *data_sinfo;
3653
3654         /* make sure bytes are sectorsize aligned */
3655         bytes = ALIGN(bytes, root->sectorsize);
3656
3657         data_sinfo = root->fs_info->data_sinfo;
3658         spin_lock(&data_sinfo->lock);
3659         data_sinfo->bytes_may_use -= bytes;
3660         trace_btrfs_space_reservation(root->fs_info, "space_info",
3661                                       data_sinfo->flags, bytes, 0);
3662         spin_unlock(&data_sinfo->lock);
3663 }
3664
3665 static void force_metadata_allocation(struct btrfs_fs_info *info)
3666 {
3667         struct list_head *head = &info->space_info;
3668         struct btrfs_space_info *found;
3669
3670         rcu_read_lock();
3671         list_for_each_entry_rcu(found, head, list) {
3672                 if (found->flags & BTRFS_BLOCK_GROUP_METADATA)
3673                         found->force_alloc = CHUNK_ALLOC_FORCE;
3674         }
3675         rcu_read_unlock();
3676 }
3677
3678 static inline u64 calc_global_rsv_need_space(struct btrfs_block_rsv *global)
3679 {
3680         return (global->size << 1);
3681 }
3682
3683 static int should_alloc_chunk(struct btrfs_root *root,
3684                               struct btrfs_space_info *sinfo, int force)
3685 {
3686         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
3687         u64 num_bytes = sinfo->total_bytes - sinfo->bytes_readonly;
3688         u64 num_allocated = sinfo->bytes_used + sinfo->bytes_reserved;
3689         u64 thresh;
3690
3691         if (force == CHUNK_ALLOC_FORCE)
3692                 return 1;
3693
3694         /*
3695          * We need to take into account the global rsv because for all intents
3696          * and purposes it's used space.  Don't worry about locking the
3697          * global_rsv, it doesn't change except when the transaction commits.
3698          */
3699         if (sinfo->flags & BTRFS_BLOCK_GROUP_METADATA)
3700                 num_allocated += calc_global_rsv_need_space(global_rsv);
3701
3702         /*
3703          * in limited mode, we want to have some free space up to
3704          * about 1% of the FS size.
3705          */
3706         if (force == CHUNK_ALLOC_LIMITED) {
3707                 thresh = btrfs_super_total_bytes(root->fs_info->super_copy);
3708                 thresh = max_t(u64, 64 * 1024 * 1024,
3709                                div_factor_fine(thresh, 1));
3710
3711                 if (num_bytes - num_allocated < thresh)
3712                         return 1;
3713         }
3714
3715         if (num_allocated + 2 * 1024 * 1024 < div_factor(num_bytes, 8))
3716                 return 0;
3717         return 1;
3718 }
3719
3720 static u64 get_system_chunk_thresh(struct btrfs_root *root, u64 type)
3721 {
3722         u64 num_dev;
3723
3724         if (type & (BTRFS_BLOCK_GROUP_RAID10 |
3725                     BTRFS_BLOCK_GROUP_RAID0 |
3726                     BTRFS_BLOCK_GROUP_RAID5 |
3727                     BTRFS_BLOCK_GROUP_RAID6))
3728                 num_dev = root->fs_info->fs_devices->rw_devices;
3729         else if (type & BTRFS_BLOCK_GROUP_RAID1)
3730                 num_dev = 2;
3731         else
3732                 num_dev = 1;    /* DUP or single */
3733
3734         /* metadata for updaing devices and chunk tree */
3735         return btrfs_calc_trans_metadata_size(root, num_dev + 1);
3736 }
3737
3738 static void check_system_chunk(struct btrfs_trans_handle *trans,
3739                                struct btrfs_root *root, u64 type)
3740 {
3741         struct btrfs_space_info *info;
3742         u64 left;
3743         u64 thresh;
3744
3745         info = __find_space_info(root->fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
3746         spin_lock(&info->lock);
3747         left = info->total_bytes - info->bytes_used - info->bytes_pinned -
3748                 info->bytes_reserved - info->bytes_readonly;
3749         spin_unlock(&info->lock);
3750
3751         thresh = get_system_chunk_thresh(root, type);
3752         if (left < thresh && btrfs_test_opt(root, ENOSPC_DEBUG)) {
3753                 btrfs_info(root->fs_info, "left=%llu, need=%llu, flags=%llu",
3754                         left, thresh, type);
3755                 dump_space_info(info, 0, 0);
3756         }
3757
3758         if (left < thresh) {
3759                 u64 flags;
3760
3761                 flags = btrfs_get_alloc_profile(root->fs_info->chunk_root, 0);
3762                 btrfs_alloc_chunk(trans, root, flags);
3763         }
3764 }
3765
3766 static int do_chunk_alloc(struct btrfs_trans_handle *trans,
3767                           struct btrfs_root *extent_root, u64 flags, int force)
3768 {
3769         struct btrfs_space_info *space_info;
3770         struct btrfs_fs_info *fs_info = extent_root->fs_info;
3771         int wait_for_alloc = 0;
3772         int ret = 0;
3773
3774         /* Don't re-enter if we're already allocating a chunk */
3775         if (trans->allocating_chunk)
3776                 return -ENOSPC;
3777
3778         space_info = __find_space_info(extent_root->fs_info, flags);
3779         if (!space_info) {
3780                 ret = update_space_info(extent_root->fs_info, flags,
3781                                         0, 0, &space_info);
3782                 BUG_ON(ret); /* -ENOMEM */
3783         }
3784         BUG_ON(!space_info); /* Logic error */
3785
3786 again:
3787         spin_lock(&space_info->lock);
3788         if (force < space_info->force_alloc)
3789                 force = space_info->force_alloc;
3790         if (space_info->full) {
3791                 spin_unlock(&space_info->lock);
3792                 return 0;
3793         }
3794
3795         if (!should_alloc_chunk(extent_root, space_info, force)) {
3796                 spin_unlock(&space_info->lock);
3797                 return 0;
3798         } else if (space_info->chunk_alloc) {
3799                 wait_for_alloc = 1;
3800         } else {
3801                 space_info->chunk_alloc = 1;
3802         }
3803
3804         spin_unlock(&space_info->lock);
3805
3806         mutex_lock(&fs_info->chunk_mutex);
3807
3808         /*
3809          * The chunk_mutex is held throughout the entirety of a chunk
3810          * allocation, so once we've acquired the chunk_mutex we know that the
3811          * other guy is done and we need to recheck and see if we should
3812          * allocate.
3813          */
3814         if (wait_for_alloc) {
3815                 mutex_unlock(&fs_info->chunk_mutex);
3816                 wait_for_alloc = 0;
3817                 goto again;
3818         }
3819
3820         trans->allocating_chunk = true;
3821
3822         /*
3823          * If we have mixed data/metadata chunks we want to make sure we keep
3824          * allocating mixed chunks instead of individual chunks.
3825          */
3826         if (btrfs_mixed_space_info(space_info))
3827                 flags |= (BTRFS_BLOCK_GROUP_DATA | BTRFS_BLOCK_GROUP_METADATA);
3828
3829         /*
3830          * if we're doing a data chunk, go ahead and make sure that
3831          * we keep a reasonable number of metadata chunks allocated in the
3832          * FS as well.
3833          */
3834         if (flags & BTRFS_BLOCK_GROUP_DATA && fs_info->metadata_ratio) {
3835                 fs_info->data_chunk_allocations++;
3836                 if (!(fs_info->data_chunk_allocations %
3837                       fs_info->metadata_ratio))
3838                         force_metadata_allocation(fs_info);
3839         }
3840
3841         /*
3842          * Check if we have enough space in SYSTEM chunk because we may need
3843          * to update devices.
3844          */
3845         check_system_chunk(trans, extent_root, flags);
3846
3847         ret = btrfs_alloc_chunk(trans, extent_root, flags);
3848         trans->allocating_chunk = false;
3849
3850         spin_lock(&space_info->lock);
3851         if (ret < 0 && ret != -ENOSPC)
3852                 goto out;
3853         if (ret)
3854                 space_info->full = 1;
3855         else
3856                 ret = 1;
3857
3858         space_info->force_alloc = CHUNK_ALLOC_NO_FORCE;
3859 out:
3860         space_info->chunk_alloc = 0;
3861         spin_unlock(&space_info->lock);
3862         mutex_unlock(&fs_info->chunk_mutex);
3863         return ret;
3864 }
3865
3866 static int can_overcommit(struct btrfs_root *root,
3867                           struct btrfs_space_info *space_info, u64 bytes,
3868                           enum btrfs_reserve_flush_enum flush)
3869 {
3870         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
3871         u64 profile = btrfs_get_alloc_profile(root, 0);
3872         u64 space_size;
3873         u64 avail;
3874         u64 used;
3875         u64 to_add;
3876
3877         used = space_info->bytes_used + space_info->bytes_reserved +
3878                 space_info->bytes_pinned + space_info->bytes_readonly;
3879
3880         /*
3881          * We only want to allow over committing if we have lots of actual space
3882          * free, but if we don't have enough space to handle the global reserve
3883          * space then we could end up having a real enospc problem when trying
3884          * to allocate a chunk or some other such important allocation.
3885          */
3886         spin_lock(&global_rsv->lock);
3887         space_size = calc_global_rsv_need_space(global_rsv);
3888         spin_unlock(&global_rsv->lock);
3889         if (used + space_size >= space_info->total_bytes)
3890                 return 0;
3891
3892         used += space_info->bytes_may_use;
3893
3894         spin_lock(&root->fs_info->free_chunk_lock);
3895         avail = root->fs_info->free_chunk_space;
3896         spin_unlock(&root->fs_info->free_chunk_lock);
3897
3898         /*
3899          * If we have dup, raid1 or raid10 then only half of the free
3900          * space is actually useable.  For raid56, the space info used
3901          * doesn't include the parity drive, so we don't have to
3902          * change the math
3903          */
3904         if (profile & (BTRFS_BLOCK_GROUP_DUP |
3905                        BTRFS_BLOCK_GROUP_RAID1 |
3906                        BTRFS_BLOCK_GROUP_RAID10))
3907                 avail >>= 1;
3908
3909         to_add = space_info->total_bytes;
3910
3911         /*
3912          * If we aren't flushing all things, let us overcommit up to
3913          * 1/2th of the space. If we can flush, don't let us overcommit
3914          * too much, let it overcommit up to 1/8 of the space.
3915          */
3916         if (flush == BTRFS_RESERVE_FLUSH_ALL)
3917                 to_add >>= 3;
3918         else
3919                 to_add >>= 1;
3920
3921         /*
3922          * Limit the overcommit to the amount of free space we could possibly
3923          * allocate for chunks.
3924          */
3925         to_add = min(avail, to_add);
3926
3927         if (used + bytes < space_info->total_bytes + to_add)
3928                 return 1;
3929         return 0;
3930 }
3931
3932 static void btrfs_writeback_inodes_sb_nr(struct btrfs_root *root,
3933                                          unsigned long nr_pages)
3934 {
3935         struct super_block *sb = root->fs_info->sb;
3936         int started;
3937
3938         /* If we can not start writeback, just sync all the delalloc file. */
3939         started = try_to_writeback_inodes_sb_nr(sb, nr_pages,
3940                                                       WB_REASON_FS_FREE_SPACE);
3941         if (!started) {
3942                 /*
3943                  * We needn't worry the filesystem going from r/w to r/o though
3944                  * we don't acquire ->s_umount mutex, because the filesystem
3945                  * should guarantee the delalloc inodes list be empty after
3946                  * the filesystem is readonly(all dirty pages are written to
3947                  * the disk).
3948                  */
3949                 btrfs_start_all_delalloc_inodes(root->fs_info, 0);
3950                 if (!current->journal_info)
3951                         btrfs_wait_all_ordered_extents(root->fs_info, 0);
3952         }
3953 }
3954
3955 /*
3956  * shrink metadata reservation for delalloc
3957  */
3958 static void shrink_delalloc(struct btrfs_root *root, u64 to_reclaim, u64 orig,
3959                             bool wait_ordered)
3960 {
3961         struct btrfs_block_rsv *block_rsv;
3962         struct btrfs_space_info *space_info;
3963         struct btrfs_trans_handle *trans;
3964         u64 delalloc_bytes;
3965         u64 max_reclaim;
3966         long time_left;
3967         unsigned long nr_pages = (2 * 1024 * 1024) >> PAGE_CACHE_SHIFT;
3968         int loops = 0;
3969         enum btrfs_reserve_flush_enum flush;
3970
3971         trans = (struct btrfs_trans_handle *)current->journal_info;
3972         block_rsv = &root->fs_info->delalloc_block_rsv;
3973         space_info = block_rsv->space_info;
3974
3975         smp_mb();
3976         delalloc_bytes = percpu_counter_sum_positive(
3977                                                 &root->fs_info->delalloc_bytes);
3978         if (delalloc_bytes == 0) {
3979                 if (trans)
3980                         return;
3981                 btrfs_wait_all_ordered_extents(root->fs_info, 0);
3982                 return;
3983         }
3984
3985         while (delalloc_bytes && loops < 3) {
3986                 max_reclaim = min(delalloc_bytes, to_reclaim);
3987                 nr_pages = max_reclaim >> PAGE_CACHE_SHIFT;
3988                 btrfs_writeback_inodes_sb_nr(root, nr_pages);
3989                 /*
3990                  * We need to wait for the async pages to actually start before
3991                  * we do anything.
3992                  */
3993                 wait_event(root->fs_info->async_submit_wait,
3994                            !atomic_read(&root->fs_info->async_delalloc_pages));
3995
3996                 if (!trans)
3997                         flush = BTRFS_RESERVE_FLUSH_ALL;
3998                 else
3999                         flush = BTRFS_RESERVE_NO_FLUSH;
4000                 spin_lock(&space_info->lock);
4001                 if (can_overcommit(root, space_info, orig, flush)) {
4002                         spin_unlock(&space_info->lock);
4003                         break;
4004                 }
4005                 spin_unlock(&space_info->lock);
4006
4007                 loops++;
4008                 if (wait_ordered && !trans) {
4009                         btrfs_wait_all_ordered_extents(root->fs_info, 0);
4010                 } else {
4011                         time_left = schedule_timeout_killable(1);
4012                         if (time_left)
4013                                 break;
4014                 }
4015                 smp_mb();
4016                 delalloc_bytes = percpu_counter_sum_positive(
4017                                                 &root->fs_info->delalloc_bytes);
4018         }
4019 }
4020
4021 /**
4022  * maybe_commit_transaction - possibly commit the transaction if its ok to
4023  * @root - the root we're allocating for
4024  * @bytes - the number of bytes we want to reserve
4025  * @force - force the commit
4026  *
4027  * This will check to make sure that committing the transaction will actually
4028  * get us somewhere and then commit the transaction if it does.  Otherwise it
4029  * will return -ENOSPC.
4030  */
4031 static int may_commit_transaction(struct btrfs_root *root,
4032                                   struct btrfs_space_info *space_info,
4033                                   u64 bytes, int force)
4034 {
4035         struct btrfs_block_rsv *delayed_rsv = &root->fs_info->delayed_block_rsv;
4036         struct btrfs_trans_handle *trans;
4037
4038         trans = (struct btrfs_trans_handle *)current->journal_info;
4039         if (trans)
4040                 return -EAGAIN;
4041
4042         if (force)
4043                 goto commit;
4044
4045         /* See if there is enough pinned space to make this reservation */
4046         spin_lock(&space_info->lock);
4047         if (space_info->bytes_pinned >= bytes) {
4048                 spin_unlock(&space_info->lock);
4049                 goto commit;
4050         }
4051         spin_unlock(&space_info->lock);
4052
4053         /*
4054          * See if there is some space in the delayed insertion reservation for
4055          * this reservation.
4056          */
4057         if (space_info != delayed_rsv->space_info)
4058                 return -ENOSPC;
4059
4060         spin_lock(&space_info->lock);
4061         spin_lock(&delayed_rsv->lock);
4062         if (space_info->bytes_pinned + delayed_rsv->size < bytes) {
4063                 spin_unlock(&delayed_rsv->lock);
4064                 spin_unlock(&space_info->lock);
4065                 return -ENOSPC;
4066         }
4067         spin_unlock(&delayed_rsv->lock);
4068         spin_unlock(&space_info->lock);
4069
4070 commit:
4071         trans = btrfs_join_transaction(root);
4072         if (IS_ERR(trans))
4073                 return -ENOSPC;
4074
4075         return btrfs_commit_transaction(trans, root);
4076 }
4077
4078 enum flush_state {
4079         FLUSH_DELAYED_ITEMS_NR  =       1,
4080         FLUSH_DELAYED_ITEMS     =       2,
4081         FLUSH_DELALLOC          =       3,
4082         FLUSH_DELALLOC_WAIT     =       4,
4083         ALLOC_CHUNK             =       5,
4084         COMMIT_TRANS            =       6,
4085 };
4086
4087 static int flush_space(struct btrfs_root *root,
4088                        struct btrfs_space_info *space_info, u64 num_bytes,
4089                        u64 orig_bytes, int state)
4090 {
4091         struct btrfs_trans_handle *trans;
4092         int nr;
4093         int ret = 0;
4094
4095         switch (state) {
4096         case FLUSH_DELAYED_ITEMS_NR:
4097         case FLUSH_DELAYED_ITEMS:
4098                 if (state == FLUSH_DELAYED_ITEMS_NR) {
4099                         u64 bytes = btrfs_calc_trans_metadata_size(root, 1);
4100
4101                         nr = (int)div64_u64(num_bytes, bytes);
4102                         if (!nr)
4103                                 nr = 1;
4104                         nr *= 2;
4105                 } else {
4106                         nr = -1;
4107                 }
4108                 trans = btrfs_join_transaction(root);
4109                 if (IS_ERR(trans)) {
4110                         ret = PTR_ERR(trans);
4111                         break;
4112                 }
4113                 ret = btrfs_run_delayed_items_nr(trans, root, nr);
4114                 btrfs_end_transaction(trans, root);
4115                 break;
4116         case FLUSH_DELALLOC:
4117         case FLUSH_DELALLOC_WAIT:
4118                 shrink_delalloc(root, num_bytes, orig_bytes,
4119                                 state == FLUSH_DELALLOC_WAIT);
4120                 break;
4121         case ALLOC_CHUNK:
4122                 trans = btrfs_join_transaction(root);
4123                 if (IS_ERR(trans)) {
4124                         ret = PTR_ERR(trans);
4125                         break;
4126                 }
4127                 ret = do_chunk_alloc(trans, root->fs_info->extent_root,
4128                                      btrfs_get_alloc_profile(root, 0),
4129                                      CHUNK_ALLOC_NO_FORCE);
4130                 btrfs_end_transaction(trans, root);
4131                 if (ret == -ENOSPC)
4132                         ret = 0;
4133                 break;
4134         case COMMIT_TRANS:
4135                 ret = may_commit_transaction(root, space_info, orig_bytes, 0);
4136                 break;
4137         default:
4138                 ret = -ENOSPC;
4139                 break;
4140         }
4141
4142         return ret;
4143 }
4144 /**
4145  * reserve_metadata_bytes - try to reserve bytes from the block_rsv's space
4146  * @root - the root we're allocating for
4147  * @block_rsv - the block_rsv we're allocating for
4148  * @orig_bytes - the number of bytes we want
4149  * @flush - whether or not we can flush to make our reservation
4150  *
4151  * This will reserve orgi_bytes number of bytes from the space info associated
4152  * with the block_rsv.  If there is not enough space it will make an attempt to
4153  * flush out space to make room.  It will do this by flushing delalloc if
4154  * possible or committing the transaction.  If flush is 0 then no attempts to
4155  * regain reservations will be made and this will fail if there is not enough
4156  * space already.
4157  */
4158 static int reserve_metadata_bytes(struct btrfs_root *root,
4159                                   struct btrfs_block_rsv *block_rsv,
4160                                   u64 orig_bytes,
4161                                   enum btrfs_reserve_flush_enum flush)
4162 {
4163         struct btrfs_space_info *space_info = block_rsv->space_info;
4164         u64 used;
4165         u64 num_bytes = orig_bytes;
4166         int flush_state = FLUSH_DELAYED_ITEMS_NR;
4167         int ret = 0;
4168         bool flushing = false;
4169
4170 again:
4171         ret = 0;
4172         spin_lock(&space_info->lock);
4173         /*
4174          * We only want to wait if somebody other than us is flushing and we
4175          * are actually allowed to flush all things.
4176          */
4177         while (flush == BTRFS_RESERVE_FLUSH_ALL && !flushing &&
4178                space_info->flush) {
4179                 spin_unlock(&space_info->lock);
4180                 /*
4181                  * If we have a trans handle we can't wait because the flusher
4182                  * may have to commit the transaction, which would mean we would
4183                  * deadlock since we are waiting for the flusher to finish, but
4184                  * hold the current transaction open.
4185                  */
4186                 if (current->journal_info)
4187                         return -EAGAIN;
4188                 ret = wait_event_killable(space_info->wait, !space_info->flush);
4189                 /* Must have been killed, return */
4190                 if (ret)
4191                         return -EINTR;
4192
4193                 spin_lock(&space_info->lock);
4194         }
4195
4196         ret = -ENOSPC;
4197         used = space_info->bytes_used + space_info->bytes_reserved +
4198                 space_info->bytes_pinned + space_info->bytes_readonly +
4199                 space_info->bytes_may_use;
4200
4201         /*
4202          * The idea here is that we've not already over-reserved the block group
4203          * then we can go ahead and save our reservation first and then start
4204          * flushing if we need to.  Otherwise if we've already overcommitted
4205          * lets start flushing stuff first and then come back and try to make
4206          * our reservation.
4207          */
4208         if (used <= space_info->total_bytes) {
4209                 if (used + orig_bytes <= space_info->total_bytes) {
4210                         space_info->bytes_may_use += orig_bytes;
4211                         trace_btrfs_space_reservation(root->fs_info,
4212                                 "space_info", space_info->flags, orig_bytes, 1);
4213                         ret = 0;
4214                 } else {
4215                         /*
4216                          * Ok set num_bytes to orig_bytes since we aren't
4217                          * overocmmitted, this way we only try and reclaim what
4218                          * we need.
4219                          */
4220                         num_bytes = orig_bytes;
4221                 }
4222         } else {
4223                 /*
4224                  * Ok we're over committed, set num_bytes to the overcommitted
4225                  * amount plus the amount of bytes that we need for this
4226                  * reservation.
4227                  */
4228                 num_bytes = used - space_info->total_bytes +
4229                         (orig_bytes * 2);
4230         }
4231
4232         if (ret && can_overcommit(root, space_info, orig_bytes, flush)) {
4233                 space_info->bytes_may_use += orig_bytes;
4234                 trace_btrfs_space_reservation(root->fs_info, "space_info",
4235                                               space_info->flags, orig_bytes,
4236                                               1);
4237                 ret = 0;
4238         }
4239
4240         /*
4241          * Couldn't make our reservation, save our place so while we're trying
4242          * to reclaim space we can actually use it instead of somebody else
4243          * stealing it from us.
4244          *
4245          * We make the other tasks wait for the flush only when we can flush
4246          * all things.
4247          */
4248         if (ret && flush != BTRFS_RESERVE_NO_FLUSH) {
4249                 flushing = true;
4250                 space_info->flush = 1;
4251         }
4252
4253         spin_unlock(&space_info->lock);
4254
4255         if (!ret || flush == BTRFS_RESERVE_NO_FLUSH)
4256                 goto out;
4257
4258         ret = flush_space(root, space_info, num_bytes, orig_bytes,
4259                           flush_state);
4260         flush_state++;
4261
4262         /*
4263          * If we are FLUSH_LIMIT, we can not flush delalloc, or the deadlock
4264          * would happen. So skip delalloc flush.
4265          */
4266         if (flush == BTRFS_RESERVE_FLUSH_LIMIT &&
4267             (flush_state == FLUSH_DELALLOC ||
4268              flush_state == FLUSH_DELALLOC_WAIT))
4269                 flush_state = ALLOC_CHUNK;
4270
4271         if (!ret)
4272                 goto again;
4273         else if (flush == BTRFS_RESERVE_FLUSH_LIMIT &&
4274                  flush_state < COMMIT_TRANS)
4275                 goto again;
4276         else if (flush == BTRFS_RESERVE_FLUSH_ALL &&
4277                  flush_state <= COMMIT_TRANS)
4278                 goto again;
4279
4280 out:
4281         if (ret == -ENOSPC &&
4282             unlikely(root->orphan_cleanup_state == ORPHAN_CLEANUP_STARTED)) {
4283                 struct btrfs_block_rsv *global_rsv =
4284                         &root->fs_info->global_block_rsv;
4285
4286                 if (block_rsv != global_rsv &&
4287                     !block_rsv_use_bytes(global_rsv, orig_bytes))
4288                         ret = 0;
4289         }
4290         if (flushing) {
4291                 spin_lock(&space_info->lock);
4292                 space_info->flush = 0;
4293                 wake_up_all(&space_info->wait);
4294                 spin_unlock(&space_info->lock);
4295         }
4296         return ret;
4297 }
4298
4299 static struct btrfs_block_rsv *get_block_rsv(
4300                                         const struct btrfs_trans_handle *trans,
4301                                         const struct btrfs_root *root)
4302 {
4303         struct btrfs_block_rsv *block_rsv = NULL;
4304
4305         if (root->ref_cows)
4306                 block_rsv = trans->block_rsv;
4307
4308         if (root == root->fs_info->csum_root && trans->adding_csums)
4309                 block_rsv = trans->block_rsv;
4310
4311         if (!block_rsv)
4312                 block_rsv = root->block_rsv;
4313
4314         if (!block_rsv)
4315                 block_rsv = &root->fs_info->empty_block_rsv;
4316
4317         return block_rsv;
4318 }
4319
4320 static int block_rsv_use_bytes(struct btrfs_block_rsv *block_rsv,
4321                                u64 num_bytes)
4322 {
4323         int ret = -ENOSPC;
4324         spin_lock(&block_rsv->lock);
4325         if (block_rsv->reserved >= num_bytes) {
4326                 block_rsv->reserved -= num_bytes;
4327                 if (block_rsv->reserved < block_rsv->size)
4328                         block_rsv->full = 0;
4329                 ret = 0;
4330         }
4331         spin_unlock(&block_rsv->lock);
4332         return ret;
4333 }
4334
4335 static void block_rsv_add_bytes(struct btrfs_block_rsv *block_rsv,
4336                                 u64 num_bytes, int update_size)
4337 {
4338         spin_lock(&block_rsv->lock);
4339         block_rsv->reserved += num_bytes;
4340         if (update_size)
4341                 block_rsv->size += num_bytes;
4342         else if (block_rsv->reserved >= block_rsv->size)
4343                 block_rsv->full = 1;
4344         spin_unlock(&block_rsv->lock);
4345 }
4346
4347 int btrfs_cond_migrate_bytes(struct btrfs_fs_info *fs_info,
4348                              struct btrfs_block_rsv *dest, u64 num_bytes,
4349                              int min_factor)
4350 {
4351         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
4352         u64 min_bytes;
4353
4354         if (global_rsv->space_info != dest->space_info)
4355                 return -ENOSPC;
4356
4357         spin_lock(&global_rsv->lock);
4358         min_bytes = div_factor(global_rsv->size, min_factor);
4359         if (global_rsv->reserved < min_bytes + num_bytes) {
4360                 spin_unlock(&global_rsv->lock);
4361                 return -ENOSPC;
4362         }
4363         global_rsv->reserved -= num_bytes;
4364         if (global_rsv->reserved < global_rsv->size)
4365                 global_rsv->full = 0;
4366         spin_unlock(&global_rsv->lock);
4367
4368         block_rsv_add_bytes(dest, num_bytes, 1);
4369         return 0;
4370 }
4371
4372 static void block_rsv_release_bytes(struct btrfs_fs_info *fs_info,
4373                                     struct btrfs_block_rsv *block_rsv,
4374                                     struct btrfs_block_rsv *dest, u64 num_bytes)
4375 {
4376         struct btrfs_space_info *space_info = block_rsv->space_info;
4377
4378         spin_lock(&block_rsv->lock);
4379         if (num_bytes == (u64)-1)
4380                 num_bytes = block_rsv->size;
4381         block_rsv->size -= num_bytes;
4382         if (block_rsv->reserved >= block_rsv->size) {
4383                 num_bytes = block_rsv->reserved - block_rsv->size;
4384                 block_rsv->reserved = block_rsv->size;
4385                 block_rsv->full = 1;
4386         } else {
4387                 num_bytes = 0;
4388         }
4389         spin_unlock(&block_rsv->lock);
4390
4391         if (num_bytes > 0) {
4392                 if (dest) {
4393                         spin_lock(&dest->lock);
4394                         if (!dest->full) {
4395                                 u64 bytes_to_add;
4396
4397                                 bytes_to_add = dest->size - dest->reserved;
4398                                 bytes_to_add = min(num_bytes, bytes_to_add);
4399                                 dest->reserved += bytes_to_add;
4400                                 if (dest->reserved >= dest->size)
4401                                         dest->full = 1;
4402                                 num_bytes -= bytes_to_add;
4403                         }
4404                         spin_unlock(&dest->lock);
4405                 }
4406                 if (num_bytes) {
4407                         spin_lock(&space_info->lock);
4408                         space_info->bytes_may_use -= num_bytes;
4409                         trace_btrfs_space_reservation(fs_info, "space_info",
4410                                         space_info->flags, num_bytes, 0);
4411                         space_info->reservation_progress++;
4412                         spin_unlock(&space_info->lock);
4413                 }
4414         }
4415 }
4416
4417 static int block_rsv_migrate_bytes(struct btrfs_block_rsv *src,
4418                                    struct btrfs_block_rsv *dst, u64 num_bytes)
4419 {
4420         int ret;
4421
4422         ret = block_rsv_use_bytes(src, num_bytes);
4423         if (ret)
4424                 return ret;
4425
4426         block_rsv_add_bytes(dst, num_bytes, 1);
4427         return 0;
4428 }
4429
4430 void btrfs_init_block_rsv(struct btrfs_block_rsv *rsv, unsigned short type)
4431 {
4432         memset(rsv, 0, sizeof(*rsv));
4433         spin_lock_init(&rsv->lock);
4434         rsv->type = type;
4435 }
4436
4437 struct btrfs_block_rsv *btrfs_alloc_block_rsv(struct btrfs_root *root,
4438                                               unsigned short type)
4439 {
4440         struct btrfs_block_rsv *block_rsv;
4441         struct btrfs_fs_info *fs_info = root->fs_info;
4442
4443         block_rsv = kmalloc(sizeof(*block_rsv), GFP_NOFS);
4444         if (!block_rsv)
4445                 return NULL;
4446
4447         btrfs_init_block_rsv(block_rsv, type);
4448         block_rsv->space_info = __find_space_info(fs_info,
4449                                                   BTRFS_BLOCK_GROUP_METADATA);
4450         return block_rsv;
4451 }
4452
4453 void btrfs_free_block_rsv(struct btrfs_root *root,
4454                           struct btrfs_block_rsv *rsv)
4455 {
4456         if (!rsv)
4457                 return;
4458         btrfs_block_rsv_release(root, rsv, (u64)-1);
4459         kfree(rsv);
4460 }
4461
4462 int btrfs_block_rsv_add(struct btrfs_root *root,
4463                         struct btrfs_block_rsv *block_rsv, u64 num_bytes,
4464                         enum btrfs_reserve_flush_enum flush)
4465 {
4466         int ret;
4467
4468         if (num_bytes == 0)
4469                 return 0;
4470
4471         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
4472         if (!ret) {
4473                 block_rsv_add_bytes(block_rsv, num_bytes, 1);
4474                 return 0;
4475         }
4476
4477         return ret;
4478 }
4479
4480 int btrfs_block_rsv_check(struct btrfs_root *root,
4481                           struct btrfs_block_rsv *block_rsv, int min_factor)
4482 {
4483         u64 num_bytes = 0;
4484         int ret = -ENOSPC;
4485
4486         if (!block_rsv)
4487                 return 0;
4488
4489         spin_lock(&block_rsv->lock);
4490         num_bytes = div_factor(block_rsv->size, min_factor);
4491         if (block_rsv->reserved >= num_bytes)
4492                 ret = 0;
4493         spin_unlock(&block_rsv->lock);
4494
4495         return ret;
4496 }
4497
4498 int btrfs_block_rsv_refill(struct btrfs_root *root,
4499                            struct btrfs_block_rsv *block_rsv, u64 min_reserved,
4500                            enum btrfs_reserve_flush_enum flush)
4501 {
4502         u64 num_bytes = 0;
4503         int ret = -ENOSPC;
4504
4505         if (!block_rsv)
4506                 return 0;
4507
4508         spin_lock(&block_rsv->lock);
4509         num_bytes = min_reserved;
4510         if (block_rsv->reserved >= num_bytes)
4511                 ret = 0;
4512         else
4513                 num_bytes -= block_rsv->reserved;
4514         spin_unlock(&block_rsv->lock);
4515
4516         if (!ret)
4517                 return 0;
4518
4519         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
4520         if (!ret) {
4521                 block_rsv_add_bytes(block_rsv, num_bytes, 0);
4522                 return 0;
4523         }
4524
4525         return ret;
4526 }
4527
4528 int btrfs_block_rsv_migrate(struct btrfs_block_rsv *src_rsv,
4529                             struct btrfs_block_rsv *dst_rsv,
4530                             u64 num_bytes)
4531 {
4532         return block_rsv_migrate_bytes(src_rsv, dst_rsv, num_bytes);
4533 }
4534
4535 void btrfs_block_rsv_release(struct btrfs_root *root,
4536                              struct btrfs_block_rsv *block_rsv,
4537                              u64 num_bytes)
4538 {
4539         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
4540         if (global_rsv->full || global_rsv == block_rsv ||
4541             block_rsv->space_info != global_rsv->space_info)
4542                 global_rsv = NULL;
4543         block_rsv_release_bytes(root->fs_info, block_rsv, global_rsv,
4544                                 num_bytes);
4545 }
4546
4547 /*
4548  * helper to calculate size of global block reservation.
4549  * the desired value is sum of space used by extent tree,
4550  * checksum tree and root tree
4551  */
4552 static u64 calc_global_metadata_size(struct btrfs_fs_info *fs_info)
4553 {
4554         struct btrfs_space_info *sinfo;
4555         u64 num_bytes;
4556         u64 meta_used;
4557         u64 data_used;
4558         int csum_size = btrfs_super_csum_size(fs_info->super_copy);
4559
4560         sinfo = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_DATA);
4561         spin_lock(&sinfo->lock);
4562         data_used = sinfo->bytes_used;
4563         spin_unlock(&sinfo->lock);
4564
4565         sinfo = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
4566         spin_lock(&sinfo->lock);
4567         if (sinfo->flags & BTRFS_BLOCK_GROUP_DATA)
4568                 data_used = 0;
4569         meta_used = sinfo->bytes_used;
4570         spin_unlock(&sinfo->lock);
4571
4572         num_bytes = (data_used >> fs_info->sb->s_blocksize_bits) *
4573                     csum_size * 2;
4574         num_bytes += div64_u64(data_used + meta_used, 50);
4575
4576         if (num_bytes * 3 > meta_used)
4577                 num_bytes = div64_u64(meta_used, 3);
4578
4579         return ALIGN(num_bytes, fs_info->extent_root->leafsize << 10);
4580 }
4581
4582 static void update_global_block_rsv(struct btrfs_fs_info *fs_info)
4583 {
4584         struct btrfs_block_rsv *block_rsv = &fs_info->global_block_rsv;
4585         struct btrfs_space_info *sinfo = block_rsv->space_info;
4586         u64 num_bytes;
4587
4588         num_bytes = calc_global_metadata_size(fs_info);
4589
4590         spin_lock(&sinfo->lock);
4591         spin_lock(&block_rsv->lock);
4592
4593         block_rsv->size = min_t(u64, num_bytes, 512 * 1024 * 1024);
4594
4595         num_bytes = sinfo->bytes_used + sinfo->bytes_pinned +
4596                     sinfo->bytes_reserved + sinfo->bytes_readonly +
4597                     sinfo->bytes_may_use;
4598
4599         if (sinfo->total_bytes > num_bytes) {
4600                 num_bytes = sinfo->total_bytes - num_bytes;
4601                 block_rsv->reserved += num_bytes;
4602                 sinfo->bytes_may_use += num_bytes;
4603                 trace_btrfs_space_reservation(fs_info, "space_info",
4604                                       sinfo->flags, num_bytes, 1);
4605         }
4606
4607         if (block_rsv->reserved >= block_rsv->size) {
4608                 num_bytes = block_rsv->reserved - block_rsv->size;
4609                 sinfo->bytes_may_use -= num_bytes;
4610                 trace_btrfs_space_reservation(fs_info, "space_info",
4611                                       sinfo->flags, num_bytes, 0);
4612                 sinfo->reservation_progress++;
4613                 block_rsv->reserved = block_rsv->size;
4614                 block_rsv->full = 1;
4615         }
4616
4617         spin_unlock(&block_rsv->lock);
4618         spin_unlock(&sinfo->lock);
4619 }
4620
4621 static void init_global_block_rsv(struct btrfs_fs_info *fs_info)
4622 {
4623         struct btrfs_space_info *space_info;
4624
4625         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
4626         fs_info->chunk_block_rsv.space_info = space_info;
4627
4628         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
4629         fs_info->global_block_rsv.space_info = space_info;
4630         fs_info->delalloc_block_rsv.space_info = space_info;
4631         fs_info->trans_block_rsv.space_info = space_info;
4632         fs_info->empty_block_rsv.space_info = space_info;
4633         fs_info->delayed_block_rsv.space_info = space_info;
4634
4635         fs_info->extent_root->block_rsv = &fs_info->global_block_rsv;
4636         fs_info->csum_root->block_rsv = &fs_info->global_block_rsv;
4637         fs_info->dev_root->block_rsv = &fs_info->global_block_rsv;
4638         fs_info->tree_root->block_rsv = &fs_info->global_block_rsv;
4639         if (fs_info->quota_root)
4640                 fs_info->quota_root->block_rsv = &fs_info->global_block_rsv;
4641         fs_info->chunk_root->block_rsv = &fs_info->chunk_block_rsv;
4642
4643         update_global_block_rsv(fs_info);
4644 }
4645
4646 static void release_global_block_rsv(struct btrfs_fs_info *fs_info)
4647 {
4648         block_rsv_release_bytes(fs_info, &fs_info->global_block_rsv, NULL,
4649                                 (u64)-1);
4650         WARN_ON(fs_info->delalloc_block_rsv.size > 0);
4651         WARN_ON(fs_info->delalloc_block_rsv.reserved > 0);
4652         WARN_ON(fs_info->trans_block_rsv.size > 0);
4653         WARN_ON(fs_info->trans_block_rsv.reserved > 0);
4654         WARN_ON(fs_info->chunk_block_rsv.size > 0);
4655         WARN_ON(fs_info->chunk_block_rsv.reserved > 0);
4656         WARN_ON(fs_info->delayed_block_rsv.size > 0);
4657         WARN_ON(fs_info->delayed_block_rsv.reserved > 0);
4658 }
4659
4660 void btrfs_trans_release_metadata(struct btrfs_trans_handle *trans,
4661                                   struct btrfs_root *root)
4662 {
4663         if (!trans->block_rsv)
4664                 return;
4665
4666         if (!trans->bytes_reserved)
4667                 return;
4668
4669         trace_btrfs_space_reservation(root->fs_info, "transaction",
4670                                       trans->transid, trans->bytes_reserved, 0);
4671         btrfs_block_rsv_release(root, trans->block_rsv, trans->bytes_reserved);
4672         trans->bytes_reserved = 0;
4673 }
4674
4675 /* Can only return 0 or -ENOSPC */
4676 int btrfs_orphan_reserve_metadata(struct btrfs_trans_handle *trans,
4677                                   struct inode *inode)
4678 {
4679         struct btrfs_root *root = BTRFS_I(inode)->root;
4680         struct btrfs_block_rsv *src_rsv = get_block_rsv(trans, root);
4681         struct btrfs_block_rsv *dst_rsv = root->orphan_block_rsv;
4682
4683         /*
4684          * We need to hold space in order to delete our orphan item once we've
4685          * added it, so this takes the reservation so we can release it later
4686          * when we are truly done with the orphan item.
4687          */
4688         u64 num_bytes = btrfs_calc_trans_metadata_size(root, 1);
4689         trace_btrfs_space_reservation(root->fs_info, "orphan",
4690                                       btrfs_ino(inode), num_bytes, 1);
4691         return block_rsv_migrate_bytes(src_rsv, dst_rsv, num_bytes);
4692 }
4693
4694 void btrfs_orphan_release_metadata(struct inode *inode)
4695 {
4696         struct btrfs_root *root = BTRFS_I(inode)->root;
4697         u64 num_bytes = btrfs_calc_trans_metadata_size(root, 1);
4698         trace_btrfs_space_reservation(root->fs_info, "orphan",
4699                                       btrfs_ino(inode), num_bytes, 0);
4700         btrfs_block_rsv_release(root, root->orphan_block_rsv, num_bytes);
4701 }
4702
4703 /*
4704  * btrfs_subvolume_reserve_metadata() - reserve space for subvolume operation
4705  * root: the root of the parent directory
4706  * rsv: block reservation
4707  * items: the number of items that we need do reservation
4708  * qgroup_reserved: used to return the reserved size in qgroup
4709  *
4710  * This function is used to reserve the space for snapshot/subvolume
4711  * creation and deletion. Those operations are different with the
4712  * common file/directory operations, they change two fs/file trees
4713  * and root tree, the number of items that the qgroup reserves is
4714  * different with the free space reservation. So we can not use
4715  * the space reseravtion mechanism in start_transaction().
4716  */
4717 int btrfs_subvolume_reserve_metadata(struct btrfs_root *root,
4718                                      struct btrfs_block_rsv *rsv,
4719                                      int items,
4720                                      u64 *qgroup_reserved)
4721 {
4722         u64 num_bytes;
4723         int ret;
4724
4725         if (root->fs_info->quota_enabled) {
4726                 /* One for parent inode, two for dir entries */
4727                 num_bytes = 3 * root->leafsize;
4728                 ret = btrfs_qgroup_reserve(root, num_bytes);
4729                 if (ret)
4730                         return ret;
4731         } else {
4732                 num_bytes = 0;
4733         }
4734
4735         *qgroup_reserved = num_bytes;
4736
4737         num_bytes = btrfs_calc_trans_metadata_size(root, items);
4738         rsv->space_info = __find_space_info(root->fs_info,
4739                                             BTRFS_BLOCK_GROUP_METADATA);
4740         ret = btrfs_block_rsv_add(root, rsv, num_bytes,
4741                                   BTRFS_RESERVE_FLUSH_ALL);
4742         if (ret) {
4743                 if (*qgroup_reserved)
4744                         btrfs_qgroup_free(root, *qgroup_reserved);
4745         }
4746
4747         return ret;
4748 }
4749
4750 void btrfs_subvolume_release_metadata(struct btrfs_root *root,
4751                                       struct btrfs_block_rsv *rsv,
4752                                       u64 qgroup_reserved)
4753 {
4754         btrfs_block_rsv_release(root, rsv, (u64)-1);
4755         if (qgroup_reserved)
4756                 btrfs_qgroup_free(root, qgroup_reserved);
4757 }
4758
4759 /**
4760  * drop_outstanding_extent - drop an outstanding extent
4761  * @inode: the inode we're dropping the extent for
4762  *
4763  * This is called when we are freeing up an outstanding extent, either called
4764  * after an error or after an extent is written.  This will return the number of
4765  * reserved extents that need to be freed.  This must be called with
4766  * BTRFS_I(inode)->lock held.
4767  */
4768 static unsigned drop_outstanding_extent(struct inode *inode)
4769 {
4770         unsigned drop_inode_space = 0;
4771         unsigned dropped_extents = 0;
4772
4773         BUG_ON(!BTRFS_I(inode)->outstanding_extents);
4774         BTRFS_I(inode)->outstanding_extents--;
4775
4776         if (BTRFS_I(inode)->outstanding_extents == 0 &&
4777             test_and_clear_bit(BTRFS_INODE_DELALLOC_META_RESERVED,
4778                                &BTRFS_I(inode)->runtime_flags))
4779                 drop_inode_space = 1;
4780
4781         /*
4782          * If we have more or the same amount of outsanding extents than we have
4783          * reserved then we need to leave the reserved extents count alone.
4784          */
4785         if (BTRFS_I(inode)->outstanding_extents >=
4786             BTRFS_I(inode)->reserved_extents)
4787                 return drop_inode_space;
4788
4789         dropped_extents = BTRFS_I(inode)->reserved_extents -
4790                 BTRFS_I(inode)->outstanding_extents;
4791         BTRFS_I(inode)->reserved_extents -= dropped_extents;
4792         return dropped_extents + drop_inode_space;
4793 }
4794
4795 /**
4796  * calc_csum_metadata_size - return the amount of metada space that must be
4797  *      reserved/free'd for the given bytes.
4798  * @inode: the inode we're manipulating
4799  * @num_bytes: the number of bytes in question
4800  * @reserve: 1 if we are reserving space, 0 if we are freeing space
4801  *
4802  * This adjusts the number of csum_bytes in the inode and then returns the
4803  * correct amount of metadata that must either be reserved or freed.  We
4804  * calculate how many checksums we can fit into one leaf and then divide the
4805  * number of bytes that will need to be checksumed by this value to figure out
4806  * how many checksums will be required.  If we are adding bytes then the number
4807  * may go up and we will return the number of additional bytes that must be
4808  * reserved.  If it is going down we will return the number of bytes that must
4809  * be freed.
4810  *
4811  * This must be called with BTRFS_I(inode)->lock held.
4812  */
4813 static u64 calc_csum_metadata_size(struct inode *inode, u64 num_bytes,
4814                                    int reserve)
4815 {
4816         struct btrfs_root *root = BTRFS_I(inode)->root;
4817         u64 csum_size;
4818         int num_csums_per_leaf;
4819         int num_csums;
4820         int old_csums;
4821
4822         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM &&
4823             BTRFS_I(inode)->csum_bytes == 0)
4824                 return 0;
4825
4826         old_csums = (int)div64_u64(BTRFS_I(inode)->csum_bytes, root->sectorsize);
4827         if (reserve)
4828                 BTRFS_I(inode)->csum_bytes += num_bytes;
4829         else
4830                 BTRFS_I(inode)->csum_bytes -= num_bytes;
4831         csum_size = BTRFS_LEAF_DATA_SIZE(root) - sizeof(struct btrfs_item);
4832         num_csums_per_leaf = (int)div64_u64(csum_size,
4833                                             sizeof(struct btrfs_csum_item) +
4834                                             sizeof(struct btrfs_disk_key));
4835         num_csums = (int)div64_u64(BTRFS_I(inode)->csum_bytes, root->sectorsize);
4836         num_csums = num_csums + num_csums_per_leaf - 1;
4837         num_csums = num_csums / num_csums_per_leaf;
4838
4839         old_csums = old_csums + num_csums_per_leaf - 1;
4840         old_csums = old_csums / num_csums_per_leaf;
4841
4842         /* No change, no need to reserve more */
4843         if (old_csums == num_csums)
4844                 return 0;
4845
4846         if (reserve)
4847                 return btrfs_calc_trans_metadata_size(root,
4848                                                       num_csums - old_csums);
4849
4850         return btrfs_calc_trans_metadata_size(root, old_csums - num_csums);
4851 }
4852
4853 int btrfs_delalloc_reserve_metadata(struct inode *inode, u64 num_bytes)
4854 {
4855         struct btrfs_root *root = BTRFS_I(inode)->root;
4856         struct btrfs_block_rsv *block_rsv = &root->fs_info->delalloc_block_rsv;
4857         u64 to_reserve = 0;
4858         u64 csum_bytes;
4859         unsigned nr_extents = 0;
4860         int extra_reserve = 0;
4861         enum btrfs_reserve_flush_enum flush = BTRFS_RESERVE_FLUSH_ALL;
4862         int ret = 0;
4863         bool delalloc_lock = true;
4864         u64 to_free = 0;
4865         unsigned dropped;
4866
4867         /* If we are a free space inode we need to not flush since we will be in
4868          * the middle of a transaction commit.  We also don't need the delalloc
4869          * mutex since we won't race with anybody.  We need this mostly to make
4870          * lockdep shut its filthy mouth.
4871          */
4872         if (btrfs_is_free_space_inode(inode)) {
4873                 flush = BTRFS_RESERVE_NO_FLUSH;
4874                 delalloc_lock = false;
4875         }
4876
4877         if (flush != BTRFS_RESERVE_NO_FLUSH &&
4878             btrfs_transaction_in_commit(root->fs_info))
4879                 schedule_timeout(1);
4880
4881         if (delalloc_lock)
4882                 mutex_lock(&BTRFS_I(inode)->delalloc_mutex);
4883
4884         num_bytes = ALIGN(num_bytes, root->sectorsize);
4885
4886         spin_lock(&BTRFS_I(inode)->lock);
4887         BTRFS_I(inode)->outstanding_extents++;
4888
4889         if (BTRFS_I(inode)->outstanding_extents >
4890             BTRFS_I(inode)->reserved_extents)
4891                 nr_extents = BTRFS_I(inode)->outstanding_extents -
4892                         BTRFS_I(inode)->reserved_extents;
4893
4894         /*
4895          * Add an item to reserve for updating the inode when we complete the
4896          * delalloc io.
4897          */
4898         if (!test_bit(BTRFS_INODE_DELALLOC_META_RESERVED,
4899                       &BTRFS_I(inode)->runtime_flags)) {
4900                 nr_extents++;
4901                 extra_reserve = 1;
4902         }
4903
4904         to_reserve = btrfs_calc_trans_metadata_size(root, nr_extents);
4905         to_reserve += calc_csum_metadata_size(inode, num_bytes, 1);
4906         csum_bytes = BTRFS_I(inode)->csum_bytes;
4907         spin_unlock(&BTRFS_I(inode)->lock);
4908
4909         if (root->fs_info->quota_enabled) {
4910                 ret = btrfs_qgroup_reserve(root, num_bytes +
4911                                            nr_extents * root->leafsize);
4912                 if (ret)
4913                         goto out_fail;
4914         }
4915
4916         ret = reserve_metadata_bytes(root, block_rsv, to_reserve, flush);
4917         if (unlikely(ret)) {
4918                 if (root->fs_info->quota_enabled)
4919                         btrfs_qgroup_free(root, num_bytes +
4920                                                 nr_extents * root->leafsize);
4921                 goto out_fail;
4922         }
4923
4924         spin_lock(&BTRFS_I(inode)->lock);
4925         if (extra_reserve) {
4926                 set_bit(BTRFS_INODE_DELALLOC_META_RESERVED,
4927                         &BTRFS_I(inode)->runtime_flags);
4928                 nr_extents--;
4929         }
4930         BTRFS_I(inode)->reserved_extents += nr_extents;
4931         spin_unlock(&BTRFS_I(inode)->lock);
4932
4933         if (delalloc_lock)
4934                 mutex_unlock(&BTRFS_I(inode)->delalloc_mutex);
4935
4936         if (to_reserve)
4937                 trace_btrfs_space_reservation(root->fs_info,"delalloc",
4938                                               btrfs_ino(inode), to_reserve, 1);
4939         block_rsv_add_bytes(block_rsv, to_reserve, 1);
4940
4941         return 0;
4942
4943 out_fail:
4944         spin_lock(&BTRFS_I(inode)->lock);
4945         dropped = drop_outstanding_extent(inode);
4946         /*
4947          * If the inodes csum_bytes is the same as the original
4948          * csum_bytes then we know we haven't raced with any free()ers
4949          * so we can just reduce our inodes csum bytes and carry on.
4950          */
4951         if (BTRFS_I(inode)->csum_bytes == csum_bytes) {
4952                 calc_csum_metadata_size(inode, num_bytes, 0);
4953         } else {
4954                 u64 orig_csum_bytes = BTRFS_I(inode)->csum_bytes;
4955                 u64 bytes;
4956
4957                 /*
4958                  * This is tricky, but first we need to figure out how much we
4959                  * free'd from any free-ers that occured during this
4960                  * reservation, so we reset ->csum_bytes to the csum_bytes
4961                  * before we dropped our lock, and then call the free for the
4962                  * number of bytes that were freed while we were trying our
4963                  * reservation.
4964                  */
4965                 bytes = csum_bytes - BTRFS_I(inode)->csum_bytes;
4966                 BTRFS_I(inode)->csum_bytes = csum_bytes;
4967                 to_free = calc_csum_metadata_size(inode, bytes, 0);
4968
4969
4970                 /*
4971                  * Now we need to see how much we would have freed had we not
4972                  * been making this reservation and our ->csum_bytes were not
4973                  * artificially inflated.
4974                  */
4975                 BTRFS_I(inode)->csum_bytes = csum_bytes - num_bytes;
4976                 bytes = csum_bytes - orig_csum_bytes;
4977                 bytes = calc_csum_metadata_size(inode, bytes, 0);
4978
4979                 /*
4980                  * Now reset ->csum_bytes to what it should be.  If bytes is
4981                  * more than to_free then we would have free'd more space had we
4982                  * not had an artificially high ->csum_bytes, so we need to free
4983                  * the remainder.  If bytes is the same or less then we don't
4984                  * need to do anything, the other free-ers did the correct
4985                  * thing.
4986                  */
4987                 BTRFS_I(inode)->csum_bytes = orig_csum_bytes - num_bytes;
4988                 if (bytes > to_free)
4989                         to_free = bytes - to_free;
4990                 else
4991                         to_free = 0;
4992         }
4993         spin_unlock(&BTRFS_I(inode)->lock);
4994         if (dropped)
4995                 to_free += btrfs_calc_trans_metadata_size(root, dropped);
4996
4997         if (to_free) {
4998                 btrfs_block_rsv_release(root, block_rsv, to_free);
4999                 trace_btrfs_space_reservation(root->fs_info, "delalloc",
5000                                               btrfs_ino(inode), to_free, 0);
5001         }
5002         if (delalloc_lock)
5003                 mutex_unlock(&BTRFS_I(inode)->delalloc_mutex);
5004         return ret;
5005 }
5006
5007 /**
5008  * btrfs_delalloc_release_metadata - release a metadata reservation for an inode
5009  * @inode: the inode to release the reservation for
5010  * @num_bytes: the number of bytes we're releasing
5011  *
5012  * This will release the metadata reservation for an inode.  This can be called
5013  * once we complete IO for a given set of bytes to release their metadata
5014  * reservations.
5015  */
5016 void btrfs_delalloc_release_metadata(struct inode *inode, u64 num_bytes)
5017 {
5018         struct btrfs_root *root = BTRFS_I(inode)->root;
5019         u64 to_free = 0;
5020         unsigned dropped;
5021
5022         num_bytes = ALIGN(num_bytes, root->sectorsize);
5023         spin_lock(&BTRFS_I(inode)->lock);
5024         dropped = drop_outstanding_extent(inode);
5025
5026         if (num_bytes)
5027                 to_free = calc_csum_metadata_size(inode, num_bytes, 0);
5028         spin_unlock(&BTRFS_I(inode)->lock);
5029         if (dropped > 0)
5030                 to_free += btrfs_calc_trans_metadata_size(root, dropped);
5031
5032         trace_btrfs_space_reservation(root->fs_info, "delalloc",
5033                                       btrfs_ino(inode), to_free, 0);
5034         if (root->fs_info->quota_enabled) {
5035                 btrfs_qgroup_free(root, num_bytes +
5036                                         dropped * root->leafsize);
5037         }
5038
5039         btrfs_block_rsv_release(root, &root->fs_info->delalloc_block_rsv,
5040                                 to_free);
5041 }
5042
5043 /**
5044  * btrfs_delalloc_reserve_space - reserve data and metadata space for delalloc
5045  * @inode: inode we're writing to
5046  * @num_bytes: the number of bytes we want to allocate
5047  *
5048  * This will do the following things
5049  *
5050  * o reserve space in the data space info for num_bytes
5051  * o reserve space in the metadata space info based on number of outstanding
5052  *   extents and how much csums will be needed
5053  * o add to the inodes ->delalloc_bytes
5054  * o add it to the fs_info's delalloc inodes list.
5055  *
5056  * This will return 0 for success and -ENOSPC if there is no space left.
5057  */
5058 int btrfs_delalloc_reserve_space(struct inode *inode, u64 num_bytes)
5059 {
5060         int ret;
5061
5062         ret = btrfs_check_data_free_space(inode, num_bytes);
5063         if (ret)
5064                 return ret;
5065
5066         ret = btrfs_delalloc_reserve_metadata(inode, num_bytes);
5067         if (ret) {
5068                 btrfs_free_reserved_data_space(inode, num_bytes);
5069                 return ret;
5070         }
5071
5072         return 0;
5073 }
5074
5075 /**
5076  * btrfs_delalloc_release_space - release data and metadata space for delalloc
5077  * @inode: inode we're releasing space for
5078  * @num_bytes: the number of bytes we want to free up
5079  *
5080  * This must be matched with a call to btrfs_delalloc_reserve_space.  This is
5081  * called in the case that we don't need the metadata AND data reservations
5082  * anymore.  So if there is an error or we insert an inline extent.
5083  *
5084  * This function will release the metadata space that was not used and will
5085  * decrement ->delalloc_bytes and remove it from the fs_info delalloc_inodes
5086  * list if there are no delalloc bytes left.
5087  */
5088 void btrfs_delalloc_release_space(struct inode *inode, u64 num_bytes)
5089 {
5090         btrfs_delalloc_release_metadata(inode, num_bytes);
5091         btrfs_free_reserved_data_space(inode, num_bytes);
5092 }
5093
5094 static int update_block_group(struct btrfs_root *root,
5095                               u64 bytenr, u64 num_bytes, int alloc)
5096 {
5097         struct btrfs_block_group_cache *cache = NULL;
5098         struct btrfs_fs_info *info = root->fs_info;
5099         u64 total = num_bytes;
5100         u64 old_val;
5101         u64 byte_in_group;
5102         int factor;
5103
5104         /* block accounting for super block */
5105         spin_lock(&info->delalloc_root_lock);
5106         old_val = btrfs_super_bytes_used(info->super_copy);
5107         if (alloc)
5108                 old_val += num_bytes;
5109         else
5110                 old_val -= num_bytes;
5111         btrfs_set_super_bytes_used(info->super_copy, old_val);
5112         spin_unlock(&info->delalloc_root_lock);
5113
5114         while (total) {
5115                 cache = btrfs_lookup_block_group(info, bytenr);
5116                 if (!cache)
5117                         return -ENOENT;
5118                 if (cache->flags & (BTRFS_BLOCK_GROUP_DUP |
5119                                     BTRFS_BLOCK_GROUP_RAID1 |
5120                                     BTRFS_BLOCK_GROUP_RAID10))
5121                         factor = 2;
5122                 else
5123                         factor = 1;
5124                 /*
5125                  * If this block group has free space cache written out, we
5126                  * need to make sure to load it if we are removing space.  This
5127                  * is because we need the unpinning stage to actually add the
5128                  * space back to the block group, otherwise we will leak space.
5129                  */
5130                 if (!alloc && cache->cached == BTRFS_CACHE_NO)
5131                         cache_block_group(cache, 1);
5132
5133                 byte_in_group = bytenr - cache->key.objectid;
5134                 WARN_ON(byte_in_group > cache->key.offset);
5135
5136                 spin_lock(&cache->space_info->lock);
5137                 spin_lock(&cache->lock);
5138
5139                 if (btrfs_test_opt(root, SPACE_CACHE) &&
5140                     cache->disk_cache_state < BTRFS_DC_CLEAR)
5141                         cache->disk_cache_state = BTRFS_DC_CLEAR;
5142
5143                 cache->dirty = 1;
5144                 old_val = btrfs_block_group_used(&cache->item);
5145                 num_bytes = min(total, cache->key.offset - byte_in_group);
5146                 if (alloc) {
5147                         old_val += num_bytes;
5148                         btrfs_set_block_group_used(&cache->item, old_val);
5149                         cache->reserved -= num_bytes;
5150                         cache->space_info->bytes_reserved -= num_bytes;
5151                         cache->space_info->bytes_used += num_bytes;
5152                         cache->space_info->disk_used += num_bytes * factor;
5153                         spin_unlock(&cache->lock);
5154                         spin_unlock(&cache->space_info->lock);
5155                 } else {
5156                         old_val -= num_bytes;
5157                         btrfs_set_block_group_used(&cache->item, old_val);
5158                         cache->pinned += num_bytes;
5159                         cache->space_info->bytes_pinned += num_bytes;
5160                         cache->space_info->bytes_used -= num_bytes;
5161                         cache->space_info->disk_used -= num_bytes * factor;
5162                         spin_unlock(&cache->lock);
5163                         spin_unlock(&cache->space_info->lock);
5164
5165                         set_extent_dirty(info->pinned_extents,
5166                                          bytenr, bytenr + num_bytes - 1,
5167                                          GFP_NOFS | __GFP_NOFAIL);
5168                 }
5169                 btrfs_put_block_group(cache);
5170                 total -= num_bytes;
5171                 bytenr += num_bytes;
5172         }
5173         return 0;
5174 }
5175
5176 static u64 first_logical_byte(struct btrfs_root *root, u64 search_start)
5177 {
5178         struct btrfs_block_group_cache *cache;
5179         u64 bytenr;
5180
5181         spin_lock(&root->fs_info->block_group_cache_lock);
5182         bytenr = root->fs_info->first_logical_byte;
5183         spin_unlock(&root->fs_info->block_group_cache_lock);
5184
5185         if (bytenr < (u64)-1)
5186                 return bytenr;
5187
5188         cache = btrfs_lookup_first_block_group(root->fs_info, search_start);
5189         if (!cache)
5190                 return 0;
5191
5192         bytenr = cache->key.objectid;
5193         btrfs_put_block_group(cache);
5194
5195         return bytenr;
5196 }
5197
5198 static int pin_down_extent(struct btrfs_root *root,
5199                            struct btrfs_block_group_cache *cache,
5200                            u64 bytenr, u64 num_bytes, int reserved)
5201 {
5202         spin_lock(&cache->space_info->lock);
5203         spin_lock(&cache->lock);
5204         cache->pinned += num_bytes;
5205         cache->space_info->bytes_pinned += num_bytes;
5206         if (reserved) {
5207                 cache->reserved -= num_bytes;
5208                 cache->space_info->bytes_reserved -= num_bytes;
5209         }
5210         spin_unlock(&cache->lock);
5211         spin_unlock(&cache->space_info->lock);
5212
5213         set_extent_dirty(root->fs_info->pinned_extents, bytenr,
5214                          bytenr + num_bytes - 1, GFP_NOFS | __GFP_NOFAIL);
5215         return 0;
5216 }
5217
5218 /*
5219  * this function must be called within transaction
5220  */
5221 int btrfs_pin_extent(struct btrfs_root *root,
5222                      u64 bytenr, u64 num_bytes, int reserved)
5223 {
5224         struct btrfs_block_group_cache *cache;
5225
5226         cache = btrfs_lookup_block_group(root->fs_info, bytenr);
5227         BUG_ON(!cache); /* Logic error */
5228
5229         pin_down_extent(root, cache, bytenr, num_bytes, reserved);
5230
5231         btrfs_put_block_group(cache);
5232         return 0;
5233 }
5234
5235 /*
5236  * this function must be called within transaction
5237  */
5238 int btrfs_pin_extent_for_log_replay(struct btrfs_root *root,
5239                                     u64 bytenr, u64 num_bytes)
5240 {
5241         struct btrfs_block_group_cache *cache;
5242         int ret;
5243
5244         cache = btrfs_lookup_block_group(root->fs_info, bytenr);
5245         if (!cache)
5246                 return -EINVAL;
5247
5248         /*
5249          * pull in the free space cache (if any) so that our pin
5250          * removes the free space from the cache.  We have load_only set
5251          * to one because the slow code to read in the free extents does check
5252          * the pinned extents.
5253          */
5254         cache_block_group(cache, 1);
5255
5256         pin_down_extent(root, cache, bytenr, num_bytes, 0);
5257
5258         /* remove us from the free space cache (if we're there at all) */
5259         ret = btrfs_remove_free_space(cache, bytenr, num_bytes);
5260         btrfs_put_block_group(cache);
5261         return ret;
5262 }
5263
5264 static int __exclude_logged_extent(struct btrfs_root *root, u64 start, u64 num_bytes)
5265 {
5266         int ret;
5267         struct btrfs_block_group_cache *block_group;
5268         struct btrfs_caching_control *caching_ctl;
5269
5270         block_group = btrfs_lookup_block_group(root->fs_info, start);
5271         if (!block_group)
5272                 return -EINVAL;
5273
5274         cache_block_group(block_group, 0);
5275         caching_ctl = get_caching_control(block_group);
5276
5277         if (!caching_ctl) {
5278                 /* Logic error */
5279                 BUG_ON(!block_group_cache_done(block_group));
5280                 ret = btrfs_remove_free_space(block_group, start, num_bytes);
5281         } else {
5282                 mutex_lock(&caching_ctl->mutex);
5283
5284                 if (start >= caching_ctl->progress) {
5285                         ret = add_excluded_extent(root, start, num_bytes);
5286                 } else if (start + num_bytes <= caching_ctl->progress) {
5287                         ret = btrfs_remove_free_space(block_group,
5288                                                       start, num_bytes);
5289                 } else {
5290                         num_bytes = caching_ctl->progress - start;
5291                         ret = btrfs_remove_free_space(block_group,
5292                                                       start, num_bytes);
5293                         if (ret)
5294                                 goto out_lock;
5295
5296                         num_bytes = (start + num_bytes) -
5297                                 caching_ctl->progress;
5298                         start = caching_ctl->progress;
5299                         ret = add_excluded_extent(root, start, num_bytes);
5300                 }
5301 out_lock:
5302                 mutex_unlock(&caching_ctl->mutex);
5303                 put_caching_control(caching_ctl);
5304         }
5305         btrfs_put_block_group(block_group);
5306         return ret;
5307 }
5308
5309 int btrfs_exclude_logged_extents(struct btrfs_root *log,
5310                                  struct extent_buffer *eb)
5311 {
5312         struct btrfs_file_extent_item *item;
5313         struct btrfs_key key;
5314         int found_type;
5315         int i;
5316
5317         if (!btrfs_fs_incompat(log->fs_info, MIXED_GROUPS))
5318                 return 0;
5319
5320         for (i = 0; i < btrfs_header_nritems(eb); i++) {
5321                 btrfs_item_key_to_cpu(eb, &key, i);
5322                 if (key.type != BTRFS_EXTENT_DATA_KEY)
5323                         continue;
5324                 item = btrfs_item_ptr(eb, i, struct btrfs_file_extent_item);
5325                 found_type = btrfs_file_extent_type(eb, item);
5326                 if (found_type == BTRFS_FILE_EXTENT_INLINE)
5327                         continue;
5328                 if (btrfs_file_extent_disk_bytenr(eb, item) == 0)
5329                         continue;
5330                 key.objectid = btrfs_file_extent_disk_bytenr(eb, item);
5331                 key.offset = btrfs_file_extent_disk_num_bytes(eb, item);
5332                 __exclude_logged_extent(log, key.objectid, key.offset);
5333         }
5334
5335         return 0;
5336 }
5337
5338 /**
5339  * btrfs_update_reserved_bytes - update the block_group and space info counters
5340  * @cache:      The cache we are manipulating
5341  * @num_bytes:  The number of bytes in question
5342  * @reserve:    One of the reservation enums
5343  *
5344  * This is called by the allocator when it reserves space, or by somebody who is
5345  * freeing space that was never actually used on disk.  For example if you
5346  * reserve some space for a new leaf in transaction A and before transaction A
5347  * commits you free that leaf, you call this with reserve set to 0 in order to
5348  * clear the reservation.
5349  *
5350  * Metadata reservations should be called with RESERVE_ALLOC so we do the proper
5351  * ENOSPC accounting.  For data we handle the reservation through clearing the
5352  * delalloc bits in the io_tree.  We have to do this since we could end up
5353  * allocating less disk space for the amount of data we have reserved in the
5354  * case of compression.
5355  *
5356  * If this is a reservation and the block group has become read only we cannot
5357  * make the reservation and return -EAGAIN, otherwise this function always
5358  * succeeds.
5359  */
5360 static int btrfs_update_reserved_bytes(struct btrfs_block_group_cache *cache,
5361                                        u64 num_bytes, int reserve)
5362 {
5363         struct btrfs_space_info *space_info = cache->space_info;
5364         int ret = 0;
5365
5366         spin_lock(&space_info->lock);
5367         spin_lock(&cache->lock);
5368         if (reserve != RESERVE_FREE) {
5369                 if (cache->ro) {
5370                         ret = -EAGAIN;
5371                 } else {
5372                         cache->reserved += num_bytes;
5373                         space_info->bytes_reserved += num_bytes;
5374                         if (reserve == RESERVE_ALLOC) {
5375                                 trace_btrfs_space_reservation(cache->fs_info,
5376                                                 "space_info", space_info->flags,
5377                                                 num_bytes, 0);
5378                                 space_info->bytes_may_use -= num_bytes;
5379                         }
5380                 }
5381         } else {
5382                 if (cache->ro)
5383                         space_info->bytes_readonly += num_bytes;
5384                 cache->reserved -= num_bytes;
5385                 space_info->bytes_reserved -= num_bytes;
5386                 space_info->reservation_progress++;
5387         }
5388         spin_unlock(&cache->lock);
5389         spin_unlock(&space_info->lock);
5390         return ret;
5391 }
5392
5393 void btrfs_prepare_extent_commit(struct btrfs_trans_handle *trans,
5394                                 struct btrfs_root *root)
5395 {
5396         struct btrfs_fs_info *fs_info = root->fs_info;
5397         struct btrfs_caching_control *next;
5398         struct btrfs_caching_control *caching_ctl;
5399         struct btrfs_block_group_cache *cache;
5400
5401         down_write(&fs_info->extent_commit_sem);
5402
5403         list_for_each_entry_safe(caching_ctl, next,
5404                                  &fs_info->caching_block_groups, list) {
5405                 cache = caching_ctl->block_group;
5406                 if (block_group_cache_done(cache)) {
5407                         cache->last_byte_to_unpin = (u64)-1;
5408                         list_del_init(&caching_ctl->list);
5409                         put_caching_control(caching_ctl);
5410                 } else {
5411                         cache->last_byte_to_unpin = caching_ctl->progress;
5412                 }
5413         }
5414
5415         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
5416                 fs_info->pinned_extents = &fs_info->freed_extents[1];
5417         else
5418                 fs_info->pinned_extents = &fs_info->freed_extents[0];
5419
5420         up_write(&fs_info->extent_commit_sem);
5421
5422         update_global_block_rsv(fs_info);
5423 }
5424
5425 static int unpin_extent_range(struct btrfs_root *root, u64 start, u64 end)
5426 {
5427         struct btrfs_fs_info *fs_info = root->fs_info;
5428         struct btrfs_block_group_cache *cache = NULL;
5429         struct btrfs_space_info *space_info;
5430         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
5431         u64 len;
5432         bool readonly;
5433
5434         while (start <= end) {
5435                 readonly = false;
5436                 if (!cache ||
5437                     start >= cache->key.objectid + cache->key.offset) {
5438                         if (cache)
5439                                 btrfs_put_block_group(cache);
5440                         cache = btrfs_lookup_block_group(fs_info, start);
5441                         BUG_ON(!cache); /* Logic error */
5442                 }
5443
5444                 len = cache->key.objectid + cache->key.offset - start;
5445                 len = min(len, end + 1 - start);
5446
5447                 if (start < cache->last_byte_to_unpin) {
5448                         len = min(len, cache->last_byte_to_unpin - start);
5449                         btrfs_add_free_space(cache, start, len);
5450                 }
5451
5452                 start += len;
5453                 space_info = cache->space_info;
5454
5455                 spin_lock(&space_info->lock);
5456                 spin_lock(&cache->lock);
5457                 cache->pinned -= len;
5458                 space_info->bytes_pinned -= len;
5459                 if (cache->ro) {
5460                         space_info->bytes_readonly += len;
5461                         readonly = true;
5462                 }
5463                 spin_unlock(&cache->lock);
5464                 if (!readonly && global_rsv->space_info == space_info) {
5465                         spin_lock(&global_rsv->lock);
5466                         if (!global_rsv->full) {
5467                                 len = min(len, global_rsv->size -
5468                                           global_rsv->reserved);
5469                                 global_rsv->reserved += len;
5470                                 space_info->bytes_may_use += len;
5471                                 if (global_rsv->reserved >= global_rsv->size)
5472                                         global_rsv->full = 1;
5473                         }
5474                         spin_unlock(&global_rsv->lock);
5475                 }
5476                 spin_unlock(&space_info->lock);
5477         }
5478
5479         if (cache)
5480                 btrfs_put_block_group(cache);
5481         return 0;
5482 }
5483
5484 int btrfs_finish_extent_commit(struct btrfs_trans_handle *trans,
5485                                struct btrfs_root *root)
5486 {
5487         struct btrfs_fs_info *fs_info = root->fs_info;
5488         struct extent_io_tree *unpin;
5489         u64 start;
5490         u64 end;
5491         int ret;
5492
5493         if (trans->aborted)
5494                 return 0;
5495
5496         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
5497                 unpin = &fs_info->freed_extents[1];
5498         else
5499                 unpin = &fs_info->freed_extents[0];
5500
5501         while (1) {
5502                 ret = find_first_extent_bit(unpin, 0, &start, &end,
5503                                             EXTENT_DIRTY, NULL);
5504                 if (ret)
5505                         break;
5506
5507                 if (btrfs_test_opt(root, DISCARD))
5508                         ret = btrfs_discard_extent(root, start,
5509                                                    end + 1 - start, NULL);
5510
5511                 clear_extent_dirty(unpin, start, end, GFP_NOFS);
5512                 unpin_extent_range(root, start, end);
5513                 cond_resched();
5514         }
5515
5516         return 0;
5517 }
5518
5519 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
5520                                 struct btrfs_root *root,
5521                                 u64 bytenr, u64 num_bytes, u64 parent,
5522                                 u64 root_objectid, u64 owner_objectid,
5523                                 u64 owner_offset, int refs_to_drop,
5524                                 struct btrfs_delayed_extent_op *extent_op)
5525 {
5526         struct btrfs_key key;
5527         struct btrfs_path *path;
5528         struct btrfs_fs_info *info = root->fs_info;
5529         struct btrfs_root *extent_root = info->extent_root;
5530         struct extent_buffer *leaf;
5531         struct btrfs_extent_item *ei;
5532         struct btrfs_extent_inline_ref *iref;
5533         int ret;
5534         int is_data;
5535         int extent_slot = 0;
5536         int found_extent = 0;
5537         int num_to_del = 1;
5538         u32 item_size;
5539         u64 refs;
5540         bool skinny_metadata = btrfs_fs_incompat(root->fs_info,
5541                                                  SKINNY_METADATA);
5542
5543         path = btrfs_alloc_path();
5544         if (!path)
5545                 return -ENOMEM;
5546
5547         path->reada = 1;
5548         path->leave_spinning = 1;
5549
5550         is_data = owner_objectid >= BTRFS_FIRST_FREE_OBJECTID;
5551         BUG_ON(!is_data && refs_to_drop != 1);
5552
5553         if (is_data)
5554                 skinny_metadata = 0;
5555
5556         ret = lookup_extent_backref(trans, extent_root, path, &iref,
5557                                     bytenr, num_bytes, parent,
5558                                     root_objectid, owner_objectid,
5559                                     owner_offset);
5560         if (ret == 0) {
5561                 extent_slot = path->slots[0];
5562                 while (extent_slot >= 0) {
5563                         btrfs_item_key_to_cpu(path->nodes[0], &key,
5564                                               extent_slot);
5565                         if (key.objectid != bytenr)
5566                                 break;
5567                         if (key.type == BTRFS_EXTENT_ITEM_KEY &&
5568                             key.offset == num_bytes) {
5569                                 found_extent = 1;
5570                                 break;
5571                         }
5572                         if (key.type == BTRFS_METADATA_ITEM_KEY &&
5573                             key.offset == owner_objectid) {
5574                                 found_extent = 1;
5575                                 break;
5576                         }
5577                         if (path->slots[0] - extent_slot > 5)
5578                                 break;
5579                         extent_slot--;
5580                 }
5581 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
5582                 item_size = btrfs_item_size_nr(path->nodes[0], extent_slot);
5583                 if (found_extent && item_size < sizeof(*ei))
5584                         found_extent = 0;
5585 #endif
5586                 if (!found_extent) {
5587                         BUG_ON(iref);
5588                         ret = remove_extent_backref(trans, extent_root, path,
5589                                                     NULL, refs_to_drop,
5590                                                     is_data);
5591                         if (ret) {
5592                                 btrfs_abort_transaction(trans, extent_root, ret);
5593                                 goto out;
5594                         }
5595                         btrfs_release_path(path);
5596                         path->leave_spinning = 1;
5597
5598                         key.objectid = bytenr;
5599                         key.type = BTRFS_EXTENT_ITEM_KEY;
5600                         key.offset = num_bytes;
5601
5602                         if (!is_data && skinny_metadata) {
5603                                 key.type = BTRFS_METADATA_ITEM_KEY;
5604                                 key.offset = owner_objectid;
5605                         }
5606
5607                         ret = btrfs_search_slot(trans, extent_root,
5608                                                 &key, path, -1, 1);
5609                         if (ret > 0 && skinny_metadata && path->slots[0]) {
5610                                 /*
5611                                  * Couldn't find our skinny metadata item,
5612                                  * see if we have ye olde extent item.
5613                                  */
5614                                 path->slots[0]--;
5615                                 btrfs_item_key_to_cpu(path->nodes[0], &key,
5616                                                       path->slots[0]);
5617                                 if (key.objectid == bytenr &&
5618                                     key.type == BTRFS_EXTENT_ITEM_KEY &&
5619                                     key.offset == num_bytes)
5620                                         ret = 0;
5621                         }
5622
5623                         if (ret > 0 && skinny_metadata) {
5624                                 skinny_metadata = false;
5625                                 key.type = BTRFS_EXTENT_ITEM_KEY;
5626                                 key.offset = num_bytes;
5627                                 btrfs_release_path(path);
5628                                 ret = btrfs_search_slot(trans, extent_root,
5629                                                         &key, path, -1, 1);
5630                         }
5631
5632                         if (ret) {
5633                                 btrfs_err(info, "umm, got %d back from search, was looking for %llu",
5634                                         ret, (unsigned long long)bytenr);
5635                                 if (ret > 0)
5636                                         btrfs_print_leaf(extent_root,
5637                                                          path->nodes[0]);
5638                         }
5639                         if (ret < 0) {
5640                                 btrfs_abort_transaction(trans, extent_root, ret);
5641                                 goto out;
5642                         }
5643                         extent_slot = path->slots[0];
5644                 }
5645         } else if (ret == -ENOENT) {
5646                 btrfs_print_leaf(extent_root, path->nodes[0]);
5647                 WARN_ON(1);
5648                 btrfs_err(info,
5649                         "unable to find ref byte nr %llu parent %llu root %llu  owner %llu offset %llu",
5650                         (unsigned long long)bytenr,
5651                         (unsigned long long)parent,
5652                         (unsigned long long)root_objectid,
5653                         (unsigned long long)owner_objectid,
5654                         (unsigned long long)owner_offset);
5655         } else {
5656                 btrfs_abort_transaction(trans, extent_root, ret);
5657                 goto out;
5658         }
5659
5660         leaf = path->nodes[0];
5661         item_size = btrfs_item_size_nr(leaf, extent_slot);
5662 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
5663         if (item_size < sizeof(*ei)) {
5664                 BUG_ON(found_extent || extent_slot != path->slots[0]);
5665                 ret = convert_extent_item_v0(trans, extent_root, path,
5666                                              owner_objectid, 0);
5667                 if (ret < 0) {
5668                         btrfs_abort_transaction(trans, extent_root, ret);
5669                         goto out;
5670                 }
5671
5672                 btrfs_release_path(path);
5673                 path->leave_spinning = 1;
5674
5675                 key.objectid = bytenr;
5676                 key.type = BTRFS_EXTENT_ITEM_KEY;
5677                 key.offset = num_bytes;
5678
5679                 ret = btrfs_search_slot(trans, extent_root, &key, path,
5680                                         -1, 1);
5681                 if (ret) {
5682                         btrfs_err(info, "umm, got %d back from search, was looking for %llu",
5683                                 ret, (unsigned long long)bytenr);
5684                         btrfs_print_leaf(extent_root, path->nodes[0]);
5685                 }
5686                 if (ret < 0) {
5687                         btrfs_abort_transaction(trans, extent_root, ret);
5688                         goto out;
5689                 }
5690
5691                 extent_slot = path->slots[0];
5692                 leaf = path->nodes[0];
5693                 item_size = btrfs_item_size_nr(leaf, extent_slot);
5694         }
5695 #endif
5696         BUG_ON(item_size < sizeof(*ei));
5697         ei = btrfs_item_ptr(leaf, extent_slot,
5698                             struct btrfs_extent_item);
5699         if (owner_objectid < BTRFS_FIRST_FREE_OBJECTID &&
5700             key.type == BTRFS_EXTENT_ITEM_KEY) {
5701                 struct btrfs_tree_block_info *bi;
5702                 BUG_ON(item_size < sizeof(*ei) + sizeof(*bi));
5703                 bi = (struct btrfs_tree_block_info *)(ei + 1);
5704                 WARN_ON(owner_objectid != btrfs_tree_block_level(leaf, bi));
5705         }
5706
5707         refs = btrfs_extent_refs(leaf, ei);
5708         if (refs < refs_to_drop) {
5709                 btrfs_err(info, "trying to drop %d refs but we only have %Lu "
5710                           "for bytenr %Lu\n", refs_to_drop, refs, bytenr);
5711                 ret = -EINVAL;
5712                 btrfs_abort_transaction(trans, extent_root, ret);
5713                 goto out;
5714         }
5715         refs -= refs_to_drop;
5716
5717         if (refs > 0) {
5718                 if (extent_op)
5719                         __run_delayed_extent_op(extent_op, leaf, ei);
5720                 /*
5721                  * In the case of inline back ref, reference count will
5722                  * be updated by remove_extent_backref
5723                  */
5724                 if (iref) {
5725                         BUG_ON(!found_extent);
5726                 } else {
5727                         btrfs_set_extent_refs(leaf, ei, refs);
5728                         btrfs_mark_buffer_dirty(leaf);
5729                 }
5730                 if (found_extent) {
5731                         ret = remove_extent_backref(trans, extent_root, path,
5732                                                     iref, refs_to_drop,
5733                                                     is_data);
5734                         if (ret) {
5735                                 btrfs_abort_transaction(trans, extent_root, ret);
5736                                 goto out;
5737                         }
5738                 }
5739         } else {
5740                 if (found_extent) {
5741                         BUG_ON(is_data && refs_to_drop !=
5742                                extent_data_ref_count(root, path, iref));
5743                         if (iref) {
5744                                 BUG_ON(path->slots[0] != extent_slot);
5745                         } else {
5746                                 BUG_ON(path->slots[0] != extent_slot + 1);
5747                                 path->slots[0] = extent_slot;
5748                                 num_to_del = 2;
5749                         }
5750                 }
5751
5752                 ret = btrfs_del_items(trans, extent_root, path, path->slots[0],
5753                                       num_to_del);
5754                 if (ret) {
5755                         btrfs_abort_transaction(trans, extent_root, ret);
5756                         goto out;
5757                 }
5758                 btrfs_release_path(path);
5759
5760                 if (is_data) {
5761                         ret = btrfs_del_csums(trans, root, bytenr, num_bytes);
5762                         if (ret) {
5763                                 btrfs_abort_transaction(trans, extent_root, ret);
5764                                 goto out;
5765                         }
5766                 }
5767
5768                 ret = update_block_group(root, bytenr, num_bytes, 0);
5769                 if (ret) {
5770                         btrfs_abort_transaction(trans, extent_root, ret);
5771                         goto out;
5772                 }
5773         }
5774 out:
5775         btrfs_free_path(path);
5776         return ret;
5777 }
5778
5779 /*
5780  * when we free an block, it is possible (and likely) that we free the last
5781  * delayed ref for that extent as well.  This searches the delayed ref tree for
5782  * a given extent, and if there are no other delayed refs to be processed, it
5783  * removes it from the tree.
5784  */
5785 static noinline int check_ref_cleanup(struct btrfs_trans_handle *trans,
5786                                       struct btrfs_root *root, u64 bytenr)
5787 {
5788         struct btrfs_delayed_ref_head *head;
5789         struct btrfs_delayed_ref_root *delayed_refs;
5790         struct btrfs_delayed_ref_node *ref;
5791         struct rb_node *node;
5792         int ret = 0;
5793
5794         delayed_refs = &trans->transaction->delayed_refs;
5795         spin_lock(&delayed_refs->lock);
5796         head = btrfs_find_delayed_ref_head(trans, bytenr);
5797         if (!head)
5798                 goto out;
5799
5800         node = rb_prev(&head->node.rb_node);
5801         if (!node)
5802                 goto out;
5803
5804         ref = rb_entry(node, struct btrfs_delayed_ref_node, rb_node);
5805
5806         /* there are still entries for this ref, we can't drop it */
5807         if (ref->bytenr == bytenr)
5808                 goto out;
5809
5810         if (head->extent_op) {
5811                 if (!head->must_insert_reserved)
5812                         goto out;
5813                 btrfs_free_delayed_extent_op(head->extent_op);
5814                 head->extent_op = NULL;
5815         }
5816
5817         /*
5818          * waiting for the lock here would deadlock.  If someone else has it
5819          * locked they are already in the process of dropping it anyway
5820          */
5821         if (!mutex_trylock(&head->mutex))
5822                 goto out;
5823
5824         /*
5825          * at this point we have a head with no other entries.  Go
5826          * ahead and process it.
5827          */
5828         head->node.in_tree = 0;
5829         rb_erase(&head->node.rb_node, &delayed_refs->root);
5830
5831         delayed_refs->num_entries--;
5832
5833         /*
5834          * we don't take a ref on the node because we're removing it from the
5835          * tree, so we just steal the ref the tree was holding.
5836          */
5837         delayed_refs->num_heads--;
5838         if (list_empty(&head->cluster))
5839                 delayed_refs->num_heads_ready--;
5840
5841         list_del_init(&head->cluster);
5842         spin_unlock(&delayed_refs->lock);
5843
5844         BUG_ON(head->extent_op);
5845         if (head->must_insert_reserved)
5846                 ret = 1;
5847
5848         mutex_unlock(&head->mutex);
5849         btrfs_put_delayed_ref(&head->node);
5850         return ret;
5851 out:
5852         spin_unlock(&delayed_refs->lock);
5853         return 0;
5854 }
5855
5856 void btrfs_free_tree_block(struct btrfs_trans_handle *trans,
5857                            struct btrfs_root *root,
5858                            struct extent_buffer *buf,
5859                            u64 parent, int last_ref)
5860 {
5861         struct btrfs_block_group_cache *cache = NULL;
5862         int ret;
5863
5864         if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
5865                 ret = btrfs_add_delayed_tree_ref(root->fs_info, trans,
5866                                         buf->start, buf->len,
5867                                         parent, root->root_key.objectid,
5868                                         btrfs_header_level(buf),
5869                                         BTRFS_DROP_DELAYED_REF, NULL, 0);
5870                 BUG_ON(ret); /* -ENOMEM */
5871         }
5872
5873         if (!last_ref)
5874                 return;
5875
5876         cache = btrfs_lookup_block_group(root->fs_info, buf->start);
5877
5878         if (btrfs_header_generation(buf) == trans->transid) {
5879                 if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
5880                         ret = check_ref_cleanup(trans, root, buf->start);
5881                         if (!ret)
5882                                 goto out;
5883                 }
5884
5885                 if (btrfs_header_flag(buf, BTRFS_HEADER_FLAG_WRITTEN)) {
5886                         pin_down_extent(root, cache, buf->start, buf->len, 1);
5887                         goto out;
5888                 }
5889
5890                 WARN_ON(test_bit(EXTENT_BUFFER_DIRTY, &buf->bflags));
5891
5892                 btrfs_add_free_space(cache, buf->start, buf->len);
5893                 btrfs_update_reserved_bytes(cache, buf->len, RESERVE_FREE);
5894         }
5895 out:
5896         /*
5897          * Deleting the buffer, clear the corrupt flag since it doesn't matter
5898          * anymore.
5899          */
5900         clear_bit(EXTENT_BUFFER_CORRUPT, &buf->bflags);
5901         btrfs_put_block_group(cache);
5902 }
5903
5904 /* Can return -ENOMEM */
5905 int btrfs_free_extent(struct btrfs_trans_handle *trans, struct btrfs_root *root,
5906                       u64 bytenr, u64 num_bytes, u64 parent, u64 root_objectid,
5907                       u64 owner, u64 offset, int for_cow)
5908 {
5909         int ret;
5910         struct btrfs_fs_info *fs_info = root->fs_info;
5911
5912         /*
5913          * tree log blocks never actually go into the extent allocation
5914          * tree, just update pinning info and exit early.
5915          */
5916         if (root_objectid == BTRFS_TREE_LOG_OBJECTID) {
5917                 WARN_ON(owner >= BTRFS_FIRST_FREE_OBJECTID);
5918                 /* unlocks the pinned mutex */
5919                 btrfs_pin_extent(root, bytenr, num_bytes, 1);
5920                 ret = 0;
5921         } else if (owner < BTRFS_FIRST_FREE_OBJECTID) {
5922                 ret = btrfs_add_delayed_tree_ref(fs_info, trans, bytenr,
5923                                         num_bytes,
5924                                         parent, root_objectid, (int)owner,
5925                                         BTRFS_DROP_DELAYED_REF, NULL, for_cow);
5926         } else {
5927                 ret = btrfs_add_delayed_data_ref(fs_info, trans, bytenr,
5928                                                 num_bytes,
5929                                                 parent, root_objectid, owner,
5930                                                 offset, BTRFS_DROP_DELAYED_REF,
5931                                                 NULL, for_cow);
5932         }
5933         return ret;
5934 }
5935
5936 static u64 stripe_align(struct btrfs_root *root,
5937                         struct btrfs_block_group_cache *cache,
5938                         u64 val, u64 num_bytes)
5939 {
5940         u64 ret = ALIGN(val, root->stripesize);
5941         return ret;
5942 }
5943
5944 /*
5945  * when we wait for progress in the block group caching, its because
5946  * our allocation attempt failed at least once.  So, we must sleep
5947  * and let some progress happen before we try again.
5948  *
5949  * This function will sleep at least once waiting for new free space to
5950  * show up, and then it will check the block group free space numbers
5951  * for our min num_bytes.  Another option is to have it go ahead
5952  * and look in the rbtree for a free extent of a given size, but this
5953  * is a good start.
5954  */
5955 static noinline int
5956 wait_block_group_cache_progress(struct btrfs_block_group_cache *cache,
5957                                 u64 num_bytes)
5958 {
5959         struct btrfs_caching_control *caching_ctl;
5960
5961         caching_ctl = get_caching_control(cache);
5962         if (!caching_ctl)
5963                 return 0;
5964
5965         wait_event(caching_ctl->wait, block_group_cache_done(cache) ||
5966                    (cache->free_space_ctl->free_space >= num_bytes));
5967
5968         put_caching_control(caching_ctl);
5969         return 0;
5970 }
5971
5972 static noinline int
5973 wait_block_group_cache_done(struct btrfs_block_group_cache *cache)
5974 {
5975         struct btrfs_caching_control *caching_ctl;
5976
5977         caching_ctl = get_caching_control(cache);
5978         if (!caching_ctl)
5979                 return 0;
5980
5981         wait_event(caching_ctl->wait, block_group_cache_done(cache));
5982
5983         put_caching_control(caching_ctl);
5984         return 0;
5985 }
5986
5987 int __get_raid_index(u64 flags)
5988 {
5989         if (flags & BTRFS_BLOCK_GROUP_RAID10)
5990                 return BTRFS_RAID_RAID10;
5991         else if (flags & BTRFS_BLOCK_GROUP_RAID1)
5992                 return BTRFS_RAID_RAID1;
5993         else if (flags & BTRFS_BLOCK_GROUP_DUP)
5994                 return BTRFS_RAID_DUP;
5995         else if (flags & BTRFS_BLOCK_GROUP_RAID0)
5996                 return BTRFS_RAID_RAID0;
5997         else if (flags & BTRFS_BLOCK_GROUP_RAID5)
5998                 return BTRFS_RAID_RAID5;
5999         else if (flags & BTRFS_BLOCK_GROUP_RAID6)
6000                 return BTRFS_RAID_RAID6;
6001
6002         return BTRFS_RAID_SINGLE; /* BTRFS_BLOCK_GROUP_SINGLE */
6003 }
6004
6005 static int get_block_group_index(struct btrfs_block_group_cache *cache)
6006 {
6007         return __get_raid_index(cache->flags);
6008 }
6009
6010 enum btrfs_loop_type {
6011         LOOP_CACHING_NOWAIT = 0,
6012         LOOP_CACHING_WAIT = 1,
6013         LOOP_ALLOC_CHUNK = 2,
6014         LOOP_NO_EMPTY_SIZE = 3,
6015 };
6016
6017 /*
6018  * walks the btree of allocated extents and find a hole of a given size.
6019  * The key ins is changed to record the hole:
6020  * ins->objectid == block start
6021  * ins->flags = BTRFS_EXTENT_ITEM_KEY
6022  * ins->offset == number of blocks
6023  * Any available blocks before search_start are skipped.
6024  */
6025 static noinline int find_free_extent(struct btrfs_trans_handle *trans,
6026                                      struct btrfs_root *orig_root,
6027                                      u64 num_bytes, u64 empty_size,
6028                                      u64 hint_byte, struct btrfs_key *ins,
6029                                      u64 flags)
6030 {
6031         int ret = 0;
6032         struct btrfs_root *root = orig_root->fs_info->extent_root;
6033         struct btrfs_free_cluster *last_ptr = NULL;
6034         struct btrfs_block_group_cache *block_group = NULL;
6035         struct btrfs_block_group_cache *used_block_group;
6036         u64 search_start = 0;
6037         int empty_cluster = 2 * 1024 * 1024;
6038         struct btrfs_space_info *space_info;
6039         int loop = 0;
6040         int index = __get_raid_index(flags);
6041         int alloc_type = (flags & BTRFS_BLOCK_GROUP_DATA) ?
6042                 RESERVE_ALLOC_NO_ACCOUNT : RESERVE_ALLOC;
6043         bool found_uncached_bg = false;
6044         bool failed_cluster_refill = false;
6045         bool failed_alloc = false;
6046         bool use_cluster = true;
6047         bool have_caching_bg = false;
6048
6049         WARN_ON(num_bytes < root->sectorsize);
6050         btrfs_set_key_type(ins, BTRFS_EXTENT_ITEM_KEY);
6051         ins->objectid = 0;
6052         ins->offset = 0;
6053
6054         trace_find_free_extent(orig_root, num_bytes, empty_size, flags);
6055
6056         space_info = __find_space_info(root->fs_info, flags);
6057         if (!space_info) {
6058                 btrfs_err(root->fs_info, "No space info for %llu", flags);
6059                 return -ENOSPC;
6060         }
6061
6062         /*
6063          * If the space info is for both data and metadata it means we have a
6064          * small filesystem and we can't use the clustering stuff.
6065          */
6066         if (btrfs_mixed_space_info(space_info))
6067                 use_cluster = false;
6068
6069         if (flags & BTRFS_BLOCK_GROUP_METADATA && use_cluster) {
6070                 last_ptr = &root->fs_info->meta_alloc_cluster;
6071                 if (!btrfs_test_opt(root, SSD))
6072                         empty_cluster = 64 * 1024;
6073         }
6074
6075         if ((flags & BTRFS_BLOCK_GROUP_DATA) && use_cluster &&
6076             btrfs_test_opt(root, SSD)) {
6077                 last_ptr = &root->fs_info->data_alloc_cluster;
6078         }
6079
6080         if (last_ptr) {
6081                 spin_lock(&last_ptr->lock);
6082                 if (last_ptr->block_group)
6083                         hint_byte = last_ptr->window_start;
6084                 spin_unlock(&last_ptr->lock);
6085         }
6086
6087         search_start = max(search_start, first_logical_byte(root, 0));
6088         search_start = max(search_start, hint_byte);
6089
6090         if (!last_ptr)
6091                 empty_cluster = 0;
6092
6093         if (search_start == hint_byte) {
6094                 block_group = btrfs_lookup_block_group(root->fs_info,
6095                                                        search_start);
6096                 used_block_group = block_group;
6097                 /*
6098                  * we don't want to use the block group if it doesn't match our
6099                  * allocation bits, or if its not cached.
6100                  *
6101                  * However if we are re-searching with an ideal block group
6102                  * picked out then we don't care that the block group is cached.
6103                  */
6104                 if (block_group && block_group_bits(block_group, flags) &&
6105                     block_group->cached != BTRFS_CACHE_NO) {
6106                         down_read(&space_info->groups_sem);
6107                         if (list_empty(&block_group->list) ||
6108                             block_group->ro) {
6109                                 /*
6110                                  * someone is removing this block group,
6111                                  * we can't jump into the have_block_group
6112                                  * target because our list pointers are not
6113                                  * valid
6114                                  */
6115                                 btrfs_put_block_group(block_group);
6116                                 up_read(&space_info->groups_sem);
6117                         } else {
6118                                 index = get_block_group_index(block_group);
6119                                 goto have_block_group;
6120                         }
6121                 } else if (block_group) {
6122                         btrfs_put_block_group(block_group);
6123                 }
6124         }
6125 search:
6126         have_caching_bg = false;
6127         down_read(&space_info->groups_sem);
6128         list_for_each_entry(block_group, &space_info->block_groups[index],
6129                             list) {
6130                 u64 offset;
6131                 int cached;
6132
6133                 used_block_group = block_group;
6134                 btrfs_get_block_group(block_group);
6135                 search_start = block_group->key.objectid;
6136
6137                 /*
6138                  * this can happen if we end up cycling through all the
6139                  * raid types, but we want to make sure we only allocate
6140                  * for the proper type.
6141                  */
6142                 if (!block_group_bits(block_group, flags)) {
6143                     u64 extra = BTRFS_BLOCK_GROUP_DUP |
6144                                 BTRFS_BLOCK_GROUP_RAID1 |
6145                                 BTRFS_BLOCK_GROUP_RAID5 |
6146                                 BTRFS_BLOCK_GROUP_RAID6 |
6147                                 BTRFS_BLOCK_GROUP_RAID10;
6148
6149                         /*
6150                          * if they asked for extra copies and this block group
6151                          * doesn't provide them, bail.  This does allow us to
6152                          * fill raid0 from raid1.
6153                          */
6154                         if ((flags & extra) && !(block_group->flags & extra))
6155                                 goto loop;
6156                 }
6157
6158 have_block_group:
6159                 cached = block_group_cache_done(block_group);
6160                 if (unlikely(!cached)) {
6161                         found_uncached_bg = true;
6162                         ret = cache_block_group(block_group, 0);
6163                         BUG_ON(ret < 0);
6164                         ret = 0;
6165                 }
6166
6167                 if (unlikely(block_group->ro))
6168                         goto loop;
6169
6170                 /*
6171                  * Ok we want to try and use the cluster allocator, so
6172                  * lets look there
6173                  */
6174                 if (last_ptr) {
6175                         unsigned long aligned_cluster;
6176                         /*
6177                          * the refill lock keeps out other
6178                          * people trying to start a new cluster
6179                          */
6180                         spin_lock(&last_ptr->refill_lock);
6181                         used_block_group = last_ptr->block_group;
6182                         if (used_block_group != block_group &&
6183                             (!used_block_group ||
6184                              used_block_group->ro ||
6185                              !block_group_bits(used_block_group, flags))) {
6186                                 used_block_group = block_group;
6187                                 goto refill_cluster;
6188                         }
6189
6190                         if (used_block_group != block_group)
6191                                 btrfs_get_block_group(used_block_group);
6192
6193                         offset = btrfs_alloc_from_cluster(used_block_group,
6194                           last_ptr, num_bytes, used_block_group->key.objectid);
6195                         if (offset) {
6196                                 /* we have a block, we're done */
6197                                 spin_unlock(&last_ptr->refill_lock);
6198                                 trace_btrfs_reserve_extent_cluster(root,
6199                                         block_group, search_start, num_bytes);
6200                                 goto checks;
6201                         }
6202
6203                         WARN_ON(last_ptr->block_group != used_block_group);
6204                         if (used_block_group != block_group) {
6205                                 btrfs_put_block_group(used_block_group);
6206                                 used_block_group = block_group;
6207                         }
6208 refill_cluster:
6209                         BUG_ON(used_block_group != block_group);
6210                         /* If we are on LOOP_NO_EMPTY_SIZE, we can't
6211                          * set up a new clusters, so lets just skip it
6212                          * and let the allocator find whatever block
6213                          * it can find.  If we reach this point, we
6214                          * will have tried the cluster allocator
6215                          * plenty of times and not have found
6216                          * anything, so we are likely way too
6217                          * fragmented for the clustering stuff to find
6218                          * anything.
6219                          *
6220                          * However, if the cluster is taken from the
6221                          * current block group, release the cluster
6222                          * first, so that we stand a better chance of
6223                          * succeeding in the unclustered
6224                          * allocation.  */
6225                         if (loop >= LOOP_NO_EMPTY_SIZE &&
6226                             last_ptr->block_group != block_group) {
6227                                 spin_unlock(&last_ptr->refill_lock);
6228                                 goto unclustered_alloc;
6229                         }
6230
6231                         /*
6232                          * this cluster didn't work out, free it and
6233                          * start over
6234                          */
6235                         btrfs_return_cluster_to_free_space(NULL, last_ptr);
6236
6237                         if (loop >= LOOP_NO_EMPTY_SIZE) {
6238                                 spin_unlock(&last_ptr->refill_lock);
6239                                 goto unclustered_alloc;
6240                         }
6241
6242                         aligned_cluster = max_t(unsigned long,
6243                                                 empty_cluster + empty_size,
6244                                               block_group->full_stripe_len);
6245
6246                         /* allocate a cluster in this block group */
6247                         ret = btrfs_find_space_cluster(trans, root,
6248                                                block_group, last_ptr,
6249                                                search_start, num_bytes,
6250                                                aligned_cluster);
6251                         if (ret == 0) {
6252                                 /*
6253                                  * now pull our allocation out of this
6254                                  * cluster
6255                                  */
6256                                 offset = btrfs_alloc_from_cluster(block_group,
6257                                                   last_ptr, num_bytes,
6258                                                   search_start);
6259                                 if (offset) {
6260                                         /* we found one, proceed */
6261                                         spin_unlock(&last_ptr->refill_lock);
6262                                         trace_btrfs_reserve_extent_cluster(root,
6263                                                 block_group, search_start,
6264                                                 num_bytes);
6265                                         goto checks;
6266                                 }
6267                         } else if (!cached && loop > LOOP_CACHING_NOWAIT
6268                                    && !failed_cluster_refill) {
6269                                 spin_unlock(&last_ptr->refill_lock);
6270
6271                                 failed_cluster_refill = true;
6272                                 wait_block_group_cache_progress(block_group,
6273                                        num_bytes + empty_cluster + empty_size);
6274                                 goto have_block_group;
6275                         }
6276
6277                         /*
6278                          * at this point we either didn't find a cluster
6279                          * or we weren't able to allocate a block from our
6280                          * cluster.  Free the cluster we've been trying
6281                          * to use, and go to the next block group
6282                          */
6283                         btrfs_return_cluster_to_free_space(NULL, last_ptr);
6284                         spin_unlock(&last_ptr->refill_lock);
6285                         goto loop;
6286                 }
6287
6288 unclustered_alloc:
6289                 spin_lock(&block_group->free_space_ctl->tree_lock);
6290                 if (cached &&
6291                     block_group->free_space_ctl->free_space <
6292                     num_bytes + empty_cluster + empty_size) {
6293                         spin_unlock(&block_group->free_space_ctl->tree_lock);
6294                         goto loop;
6295                 }
6296                 spin_unlock(&block_group->free_space_ctl->tree_lock);
6297
6298                 offset = btrfs_find_space_for_alloc(block_group, search_start,
6299                                                     num_bytes, empty_size);
6300                 /*
6301                  * If we didn't find a chunk, and we haven't failed on this
6302                  * block group before, and this block group is in the middle of
6303                  * caching and we are ok with waiting, then go ahead and wait
6304                  * for progress to be made, and set failed_alloc to true.
6305                  *
6306                  * If failed_alloc is true then we've already waited on this
6307                  * block group once and should move on to the next block group.
6308                  */
6309                 if (!offset && !failed_alloc && !cached &&
6310                     loop > LOOP_CACHING_NOWAIT) {
6311                         wait_block_group_cache_progress(block_group,
6312                                                 num_bytes + empty_size);
6313                         failed_alloc = true;
6314                         goto have_block_group;
6315                 } else if (!offset) {
6316                         if (!cached)
6317                                 have_caching_bg = true;
6318                         goto loop;
6319                 }
6320 checks:
6321                 search_start = stripe_align(root, used_block_group,
6322                                             offset, num_bytes);
6323
6324                 /* move on to the next group */
6325                 if (search_start + num_bytes >
6326                     used_block_group->key.objectid + used_block_group->key.offset) {
6327                         btrfs_add_free_space(used_block_group, offset, num_bytes);
6328                         goto loop;
6329                 }
6330
6331                 if (offset < search_start)
6332                         btrfs_add_free_space(used_block_group, offset,
6333                                              search_start - offset);
6334                 BUG_ON(offset > search_start);
6335
6336                 ret = btrfs_update_reserved_bytes(used_block_group, num_bytes,
6337                                                   alloc_type);
6338                 if (ret == -EAGAIN) {
6339                         btrfs_add_free_space(used_block_group, offset, num_bytes);
6340                         goto loop;
6341                 }
6342
6343                 /* we are all good, lets return */
6344                 ins->objectid = search_start;
6345                 ins->offset = num_bytes;
6346
6347                 trace_btrfs_reserve_extent(orig_root, block_group,
6348                                            search_start, num_bytes);
6349                 if (used_block_group != block_group)
6350                         btrfs_put_block_group(used_block_group);
6351                 btrfs_put_block_group(block_group);
6352                 break;
6353 loop:
6354                 failed_cluster_refill = false;
6355                 failed_alloc = false;
6356                 BUG_ON(index != get_block_group_index(block_group));
6357                 if (used_block_group != block_group)
6358                         btrfs_put_block_group(used_block_group);
6359                 btrfs_put_block_group(block_group);
6360         }
6361         up_read(&space_info->groups_sem);
6362
6363         if (!ins->objectid && loop >= LOOP_CACHING_WAIT && have_caching_bg)
6364                 goto search;
6365
6366         if (!ins->objectid && ++index < BTRFS_NR_RAID_TYPES)
6367                 goto search;
6368
6369         /*
6370          * LOOP_CACHING_NOWAIT, search partially cached block groups, kicking
6371          *                      caching kthreads as we move along
6372          * LOOP_CACHING_WAIT, search everything, and wait if our bg is caching
6373          * LOOP_ALLOC_CHUNK, force a chunk allocation and try again
6374          * LOOP_NO_EMPTY_SIZE, set empty_size and empty_cluster to 0 and try
6375          *                      again
6376          */
6377         if (!ins->objectid && loop < LOOP_NO_EMPTY_SIZE) {
6378                 index = 0;
6379                 loop++;
6380                 if (loop == LOOP_ALLOC_CHUNK) {
6381                         ret = do_chunk_alloc(trans, root, flags,
6382                                              CHUNK_ALLOC_FORCE);
6383                         /*
6384                          * Do not bail out on ENOSPC since we
6385                          * can do more things.
6386                          */
6387                         if (ret < 0 && ret != -ENOSPC) {
6388                                 btrfs_abort_transaction(trans,
6389                                                         root, ret);
6390                                 goto out;
6391                         }
6392                 }
6393
6394                 if (loop == LOOP_NO_EMPTY_SIZE) {
6395                         empty_size = 0;
6396                         empty_cluster = 0;
6397                 }
6398
6399                 goto search;
6400         } else if (!ins->objectid) {
6401                 ret = -ENOSPC;
6402         } else if (ins->objectid) {
6403                 ret = 0;
6404         }
6405 out:
6406
6407         return ret;
6408 }
6409
6410 static void dump_space_info(struct btrfs_space_info *info, u64 bytes,
6411                             int dump_block_groups)
6412 {
6413         struct btrfs_block_group_cache *cache;
6414         int index = 0;
6415
6416         spin_lock(&info->lock);
6417         printk(KERN_INFO "space_info %llu has %llu free, is %sfull\n",
6418                (unsigned long long)info->flags,
6419                (unsigned long long)(info->total_bytes - info->bytes_used -
6420                                     info->bytes_pinned - info->bytes_reserved -
6421                                     info->bytes_readonly),
6422                (info->full) ? "" : "not ");
6423         printk(KERN_INFO "space_info total=%llu, used=%llu, pinned=%llu, "
6424                "reserved=%llu, may_use=%llu, readonly=%llu\n",
6425                (unsigned long long)info->total_bytes,
6426                (unsigned long long)info->bytes_used,
6427                (unsigned long long)info->bytes_pinned,
6428                (unsigned long long)info->bytes_reserved,
6429                (unsigned long long)info->bytes_may_use,
6430                (unsigned long long)info->bytes_readonly);
6431         spin_unlock(&info->lock);
6432
6433         if (!dump_block_groups)
6434                 return;
6435
6436         down_read(&info->groups_sem);
6437 again:
6438         list_for_each_entry(cache, &info->block_groups[index], list) {
6439                 spin_lock(&cache->lock);
6440                 printk(KERN_INFO "block group %llu has %llu bytes, %llu used %llu pinned %llu reserved %s\n",
6441                        (unsigned long long)cache->key.objectid,
6442                        (unsigned long long)cache->key.offset,
6443                        (unsigned long long)btrfs_block_group_used(&cache->item),
6444                        (unsigned long long)cache->pinned,
6445                        (unsigned long long)cache->reserved,
6446                        cache->ro ? "[readonly]" : "");
6447                 btrfs_dump_free_space(cache, bytes);
6448                 spin_unlock(&cache->lock);
6449         }
6450         if (++index < BTRFS_NR_RAID_TYPES)
6451                 goto again;
6452         up_read(&info->groups_sem);
6453 }
6454
6455 int btrfs_reserve_extent(struct btrfs_trans_handle *trans,
6456                          struct btrfs_root *root,
6457                          u64 num_bytes, u64 min_alloc_size,
6458                          u64 empty_size, u64 hint_byte,
6459                          struct btrfs_key *ins, int is_data)
6460 {
6461         bool final_tried = false;
6462         u64 flags;
6463         int ret;
6464
6465         flags = btrfs_get_alloc_profile(root, is_data);
6466 again:
6467         WARN_ON(num_bytes < root->sectorsize);
6468         ret = find_free_extent(trans, root, num_bytes, empty_size,
6469                                hint_byte, ins, flags);
6470
6471         if (ret == -ENOSPC) {
6472                 if (!final_tried) {
6473                         num_bytes = num_bytes >> 1;
6474                         num_bytes = round_down(num_bytes, root->sectorsize);
6475                         num_bytes = max(num_bytes, min_alloc_size);
6476                         if (num_bytes == min_alloc_size)
6477                                 final_tried = true;
6478                         goto again;
6479                 } else if (btrfs_test_opt(root, ENOSPC_DEBUG)) {
6480                         struct btrfs_space_info *sinfo;
6481
6482                         sinfo = __find_space_info(root->fs_info, flags);
6483                         btrfs_err(root->fs_info, "allocation failed flags %llu, wanted %llu",
6484                                 (unsigned long long)flags,
6485                                 (unsigned long long)num_bytes);
6486                         if (sinfo)
6487                                 dump_space_info(sinfo, num_bytes, 1);
6488                 }
6489         }
6490
6491         trace_btrfs_reserved_extent_alloc(root, ins->objectid, ins->offset);
6492
6493         return ret;
6494 }
6495
6496 static int __btrfs_free_reserved_extent(struct btrfs_root *root,
6497                                         u64 start, u64 len, int pin)
6498 {
6499         struct btrfs_block_group_cache *cache;
6500         int ret = 0;
6501
6502         cache = btrfs_lookup_block_group(root->fs_info, start);
6503         if (!cache) {
6504                 btrfs_err(root->fs_info, "Unable to find block group for %llu",
6505                         (unsigned long long)start);
6506                 return -ENOSPC;
6507         }
6508
6509         if (btrfs_test_opt(root, DISCARD))
6510                 ret = btrfs_discard_extent(root, start, len, NULL);
6511
6512         if (pin)
6513                 pin_down_extent(root, cache, start, len, 1);
6514         else {
6515                 btrfs_add_free_space(cache, start, len);
6516                 btrfs_update_reserved_bytes(cache, len, RESERVE_FREE);
6517         }
6518         btrfs_put_block_group(cache);
6519
6520         trace_btrfs_reserved_extent_free(root, start, len);
6521
6522         return ret;
6523 }
6524
6525 int btrfs_free_reserved_extent(struct btrfs_root *root,
6526                                         u64 start, u64 len)
6527 {
6528         return __btrfs_free_reserved_extent(root, start, len, 0);
6529 }
6530
6531 int btrfs_free_and_pin_reserved_extent(struct btrfs_root *root,
6532                                        u64 start, u64 len)
6533 {
6534         return __btrfs_free_reserved_extent(root, start, len, 1);
6535 }
6536
6537 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
6538                                       struct btrfs_root *root,
6539                                       u64 parent, u64 root_objectid,
6540                                       u64 flags, u64 owner, u64 offset,
6541                                       struct btrfs_key *ins, int ref_mod)
6542 {
6543         int ret;
6544         struct btrfs_fs_info *fs_info = root->fs_info;
6545         struct btrfs_extent_item *extent_item;
6546         struct btrfs_extent_inline_ref *iref;
6547         struct btrfs_path *path;
6548         struct extent_buffer *leaf;
6549         int type;
6550         u32 size;
6551
6552         if (parent > 0)
6553                 type = BTRFS_SHARED_DATA_REF_KEY;
6554         else
6555                 type = BTRFS_EXTENT_DATA_REF_KEY;
6556
6557         size = sizeof(*extent_item) + btrfs_extent_inline_ref_size(type);
6558
6559         path = btrfs_alloc_path();
6560         if (!path)
6561                 return -ENOMEM;
6562
6563         path->leave_spinning = 1;
6564         ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path,
6565                                       ins, size);
6566         if (ret) {
6567                 btrfs_free_path(path);
6568                 return ret;
6569         }
6570
6571         leaf = path->nodes[0];
6572         extent_item = btrfs_item_ptr(leaf, path->slots[0],
6573                                      struct btrfs_extent_item);
6574         btrfs_set_extent_refs(leaf, extent_item, ref_mod);
6575         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
6576         btrfs_set_extent_flags(leaf, extent_item,
6577                                flags | BTRFS_EXTENT_FLAG_DATA);
6578
6579         iref = (struct btrfs_extent_inline_ref *)(extent_item + 1);
6580         btrfs_set_extent_inline_ref_type(leaf, iref, type);
6581         if (parent > 0) {
6582                 struct btrfs_shared_data_ref *ref;
6583                 ref = (struct btrfs_shared_data_ref *)(iref + 1);
6584                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
6585                 btrfs_set_shared_data_ref_count(leaf, ref, ref_mod);
6586         } else {
6587                 struct btrfs_extent_data_ref *ref;
6588                 ref = (struct btrfs_extent_data_ref *)(&iref->offset);
6589                 btrfs_set_extent_data_ref_root(leaf, ref, root_objectid);
6590                 btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
6591                 btrfs_set_extent_data_ref_offset(leaf, ref, offset);
6592                 btrfs_set_extent_data_ref_count(leaf, ref, ref_mod);
6593         }
6594
6595         btrfs_mark_buffer_dirty(path->nodes[0]);
6596         btrfs_free_path(path);
6597
6598         ret = update_block_group(root, ins->objectid, ins->offset, 1);
6599         if (ret) { /* -ENOENT, logic error */
6600                 btrfs_err(fs_info, "update block group failed for %llu %llu",
6601                         (unsigned long long)ins->objectid,
6602                         (unsigned long long)ins->offset);
6603                 BUG();
6604         }
6605         return ret;
6606 }
6607
6608 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
6609                                      struct btrfs_root *root,
6610                                      u64 parent, u64 root_objectid,
6611                                      u64 flags, struct btrfs_disk_key *key,
6612                                      int level, struct btrfs_key *ins)
6613 {
6614         int ret;
6615         struct btrfs_fs_info *fs_info = root->fs_info;
6616         struct btrfs_extent_item *extent_item;
6617         struct btrfs_tree_block_info *block_info;
6618         struct btrfs_extent_inline_ref *iref;
6619         struct btrfs_path *path;
6620         struct extent_buffer *leaf;
6621         u32 size = sizeof(*extent_item) + sizeof(*iref);
6622         bool skinny_metadata = btrfs_fs_incompat(root->fs_info,
6623                                                  SKINNY_METADATA);
6624
6625         if (!skinny_metadata)
6626                 size += sizeof(*block_info);
6627
6628         path = btrfs_alloc_path();
6629         if (!path)
6630                 return -ENOMEM;
6631
6632         path->leave_spinning = 1;
6633         ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path,
6634                                       ins, size);
6635         if (ret) {
6636                 btrfs_free_path(path);
6637                 return ret;
6638         }
6639
6640         leaf = path->nodes[0];
6641         extent_item = btrfs_item_ptr(leaf, path->slots[0],
6642                                      struct btrfs_extent_item);
6643         btrfs_set_extent_refs(leaf, extent_item, 1);
6644         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
6645         btrfs_set_extent_flags(leaf, extent_item,
6646                                flags | BTRFS_EXTENT_FLAG_TREE_BLOCK);
6647
6648         if (skinny_metadata) {
6649                 iref = (struct btrfs_extent_inline_ref *)(extent_item + 1);
6650         } else {
6651                 block_info = (struct btrfs_tree_block_info *)(extent_item + 1);
6652                 btrfs_set_tree_block_key(leaf, block_info, key);
6653                 btrfs_set_tree_block_level(leaf, block_info, level);
6654                 iref = (struct btrfs_extent_inline_ref *)(block_info + 1);
6655         }
6656
6657         if (parent > 0) {
6658                 BUG_ON(!(flags & BTRFS_BLOCK_FLAG_FULL_BACKREF));
6659                 btrfs_set_extent_inline_ref_type(leaf, iref,
6660                                                  BTRFS_SHARED_BLOCK_REF_KEY);
6661                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
6662         } else {
6663                 btrfs_set_extent_inline_ref_type(leaf, iref,
6664                                                  BTRFS_TREE_BLOCK_REF_KEY);
6665                 btrfs_set_extent_inline_ref_offset(leaf, iref, root_objectid);
6666         }
6667
6668         btrfs_mark_buffer_dirty(leaf);
6669         btrfs_free_path(path);
6670
6671         ret = update_block_group(root, ins->objectid, root->leafsize, 1);
6672         if (ret) { /* -ENOENT, logic error */
6673                 btrfs_err(fs_info, "update block group failed for %llu %llu",
6674                         (unsigned long long)ins->objectid,
6675                         (unsigned long long)ins->offset);
6676                 BUG();
6677         }
6678         return ret;
6679 }
6680
6681 int btrfs_alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
6682                                      struct btrfs_root *root,
6683                                      u64 root_objectid, u64 owner,
6684                                      u64 offset, struct btrfs_key *ins)
6685 {
6686         int ret;
6687
6688         BUG_ON(root_objectid == BTRFS_TREE_LOG_OBJECTID);
6689
6690         ret = btrfs_add_delayed_data_ref(root->fs_info, trans, ins->objectid,
6691                                          ins->offset, 0,
6692                                          root_objectid, owner, offset,
6693                                          BTRFS_ADD_DELAYED_EXTENT, NULL, 0);
6694         return ret;
6695 }
6696
6697 /*
6698  * this is used by the tree logging recovery code.  It records that
6699  * an extent has been allocated and makes sure to clear the free
6700  * space cache bits as well
6701  */
6702 int btrfs_alloc_logged_file_extent(struct btrfs_trans_handle *trans,
6703                                    struct btrfs_root *root,
6704                                    u64 root_objectid, u64 owner, u64 offset,
6705                                    struct btrfs_key *ins)
6706 {
6707         int ret;
6708         struct btrfs_block_group_cache *block_group;
6709
6710         /*
6711          * Mixed block groups will exclude before processing the log so we only
6712          * need to do the exlude dance if this fs isn't mixed.
6713          */
6714         if (!btrfs_fs_incompat(root->fs_info, MIXED_GROUPS)) {
6715                 ret = __exclude_logged_extent(root, ins->objectid, ins->offset);
6716                 if (ret)
6717                         return ret;
6718         }
6719
6720         block_group = btrfs_lookup_block_group(root->fs_info, ins->objectid);
6721         if (!block_group)
6722                 return -EINVAL;
6723
6724         ret = btrfs_update_reserved_bytes(block_group, ins->offset,
6725                                           RESERVE_ALLOC_NO_ACCOUNT);
6726         BUG_ON(ret); /* logic error */
6727         ret = alloc_reserved_file_extent(trans, root, 0, root_objectid,
6728                                          0, owner, offset, ins, 1);
6729         btrfs_put_block_group(block_group);
6730         return ret;
6731 }
6732
6733 static struct extent_buffer *
6734 btrfs_init_new_buffer(struct btrfs_trans_handle *trans, struct btrfs_root *root,
6735                       u64 bytenr, u32 blocksize, int level)
6736 {
6737         struct extent_buffer *buf;
6738
6739         buf = btrfs_find_create_tree_block(root, bytenr, blocksize);
6740         if (!buf)
6741                 return ERR_PTR(-ENOMEM);
6742         btrfs_set_header_generation(buf, trans->transid);
6743         btrfs_set_buffer_lockdep_class(root->root_key.objectid, buf, level);
6744         btrfs_tree_lock(buf);
6745         clean_tree_block(trans, root, buf);
6746         clear_bit(EXTENT_BUFFER_STALE, &buf->bflags);
6747
6748         btrfs_set_lock_blocking(buf);
6749         btrfs_set_buffer_uptodate(buf);
6750
6751         if (root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID) {
6752                 /*
6753                  * we allow two log transactions at a time, use different
6754                  * EXENT bit to differentiate dirty pages.
6755                  */
6756                 if (root->log_transid % 2 == 0)
6757                         set_extent_dirty(&root->dirty_log_pages, buf->start,
6758                                         buf->start + buf->len - 1, GFP_NOFS);
6759                 else
6760                         set_extent_new(&root->dirty_log_pages, buf->start,
6761                                         buf->start + buf->len - 1, GFP_NOFS);
6762         } else {
6763                 set_extent_dirty(&trans->transaction->dirty_pages, buf->start,
6764                          buf->start + buf->len - 1, GFP_NOFS);
6765         }
6766         trans->blocks_used++;
6767         /* this returns a buffer locked for blocking */
6768         return buf;
6769 }
6770
6771 static struct btrfs_block_rsv *
6772 use_block_rsv(struct btrfs_trans_handle *trans,
6773               struct btrfs_root *root, u32 blocksize)
6774 {
6775         struct btrfs_block_rsv *block_rsv;
6776         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
6777         int ret;
6778         bool global_updated = false;
6779
6780         block_rsv = get_block_rsv(trans, root);
6781
6782         if (unlikely(block_rsv->size == 0))
6783                 goto try_reserve;
6784 again:
6785         ret = block_rsv_use_bytes(block_rsv, blocksize);
6786         if (!ret)
6787                 return block_rsv;
6788
6789         if (block_rsv->failfast)
6790                 return ERR_PTR(ret);
6791
6792         if (block_rsv->type == BTRFS_BLOCK_RSV_GLOBAL && !global_updated) {
6793                 global_updated = true;
6794                 update_global_block_rsv(root->fs_info);
6795                 goto again;
6796         }
6797
6798         if (btrfs_test_opt(root, ENOSPC_DEBUG)) {
6799                 static DEFINE_RATELIMIT_STATE(_rs,
6800                                 DEFAULT_RATELIMIT_INTERVAL * 10,
6801                                 /*DEFAULT_RATELIMIT_BURST*/ 1);
6802                 if (__ratelimit(&_rs))
6803                         WARN(1, KERN_DEBUG
6804                                 "btrfs: block rsv returned %d\n", ret);
6805         }
6806 try_reserve:
6807         ret = reserve_metadata_bytes(root, block_rsv, blocksize,
6808                                      BTRFS_RESERVE_NO_FLUSH);
6809         if (!ret)
6810                 return block_rsv;
6811         /*
6812          * If we couldn't reserve metadata bytes try and use some from
6813          * the global reserve if its space type is the same as the global
6814          * reservation.
6815          */
6816         if (block_rsv->type != BTRFS_BLOCK_RSV_GLOBAL &&
6817             block_rsv->space_info == global_rsv->space_info) {
6818                 ret = block_rsv_use_bytes(global_rsv, blocksize);
6819                 if (!ret)
6820                         return global_rsv;
6821         }
6822         return ERR_PTR(ret);
6823 }
6824
6825 static void unuse_block_rsv(struct btrfs_fs_info *fs_info,
6826                             struct btrfs_block_rsv *block_rsv, u32 blocksize)
6827 {
6828         block_rsv_add_bytes(block_rsv, blocksize, 0);
6829         block_rsv_release_bytes(fs_info, block_rsv, NULL, 0);
6830 }
6831
6832 /*
6833  * finds a free extent and does all the dirty work required for allocation
6834  * returns the key for the extent through ins, and a tree buffer for
6835  * the first block of the extent through buf.
6836  *
6837  * returns the tree buffer or NULL.
6838  */
6839 struct extent_buffer *btrfs_alloc_free_block(struct btrfs_trans_handle *trans,
6840                                         struct btrfs_root *root, u32 blocksize,
6841                                         u64 parent, u64 root_objectid,
6842                                         struct btrfs_disk_key *key, int level,
6843                                         u64 hint, u64 empty_size)
6844 {
6845         struct btrfs_key ins;
6846         struct btrfs_block_rsv *block_rsv;
6847         struct extent_buffer *buf;
6848         u64 flags = 0;
6849         int ret;
6850         bool skinny_metadata = btrfs_fs_incompat(root->fs_info,
6851                                                  SKINNY_METADATA);
6852
6853         block_rsv = use_block_rsv(trans, root, blocksize);
6854         if (IS_ERR(block_rsv))
6855                 return ERR_CAST(block_rsv);
6856
6857         ret = btrfs_reserve_extent(trans, root, blocksize, blocksize,
6858                                    empty_size, hint, &ins, 0);
6859         if (ret) {
6860                 unuse_block_rsv(root->fs_info, block_rsv, blocksize);
6861                 return ERR_PTR(ret);
6862         }
6863
6864         buf = btrfs_init_new_buffer(trans, root, ins.objectid,
6865                                     blocksize, level);
6866         BUG_ON(IS_ERR(buf)); /* -ENOMEM */
6867
6868         if (root_objectid == BTRFS_TREE_RELOC_OBJECTID) {
6869                 if (parent == 0)
6870                         parent = ins.objectid;
6871                 flags |= BTRFS_BLOCK_FLAG_FULL_BACKREF;
6872         } else
6873                 BUG_ON(parent > 0);
6874
6875         if (root_objectid != BTRFS_TREE_LOG_OBJECTID) {
6876                 struct btrfs_delayed_extent_op *extent_op;
6877                 extent_op = btrfs_alloc_delayed_extent_op();
6878                 BUG_ON(!extent_op); /* -ENOMEM */
6879                 if (key)
6880                         memcpy(&extent_op->key, key, sizeof(extent_op->key));
6881                 else
6882                         memset(&extent_op->key, 0, sizeof(extent_op->key));
6883                 extent_op->flags_to_set = flags;
6884                 if (skinny_metadata)
6885                         extent_op->update_key = 0;
6886                 else
6887                         extent_op->update_key = 1;
6888                 extent_op->update_flags = 1;
6889                 extent_op->is_data = 0;
6890                 extent_op->level = level;
6891
6892                 ret = btrfs_add_delayed_tree_ref(root->fs_info, trans,
6893                                         ins.objectid,
6894                                         ins.offset, parent, root_objectid,
6895                                         level, BTRFS_ADD_DELAYED_EXTENT,
6896                                         extent_op, 0);
6897                 BUG_ON(ret); /* -ENOMEM */
6898         }
6899         return buf;
6900 }
6901
6902 struct walk_control {
6903         u64 refs[BTRFS_MAX_LEVEL];
6904         u64 flags[BTRFS_MAX_LEVEL];
6905         struct btrfs_key update_progress;
6906         int stage;
6907         int level;
6908         int shared_level;
6909         int update_ref;
6910         int keep_locks;
6911         int reada_slot;
6912         int reada_count;
6913         int for_reloc;
6914 };
6915
6916 #define DROP_REFERENCE  1
6917 #define UPDATE_BACKREF  2
6918
6919 static noinline void reada_walk_down(struct btrfs_trans_handle *trans,
6920                                      struct btrfs_root *root,
6921                                      struct walk_control *wc,
6922                                      struct btrfs_path *path)
6923 {
6924         u64 bytenr;
6925         u64 generation;
6926         u64 refs;
6927         u64 flags;
6928         u32 nritems;
6929         u32 blocksize;
6930         struct btrfs_key key;
6931         struct extent_buffer *eb;
6932         int ret;
6933         int slot;
6934         int nread = 0;
6935
6936         if (path->slots[wc->level] < wc->reada_slot) {
6937                 wc->reada_count = wc->reada_count * 2 / 3;
6938                 wc->reada_count = max(wc->reada_count, 2);
6939         } else {
6940                 wc->reada_count = wc->reada_count * 3 / 2;
6941                 wc->reada_count = min_t(int, wc->reada_count,
6942                                         BTRFS_NODEPTRS_PER_BLOCK(root));
6943         }
6944
6945         eb = path->nodes[wc->level];
6946         nritems = btrfs_header_nritems(eb);
6947         blocksize = btrfs_level_size(root, wc->level - 1);
6948
6949         for (slot = path->slots[wc->level]; slot < nritems; slot++) {
6950                 if (nread >= wc->reada_count)
6951                         break;
6952
6953                 cond_resched();
6954                 bytenr = btrfs_node_blockptr(eb, slot);
6955                 generation = btrfs_node_ptr_generation(eb, slot);
6956
6957                 if (slot == path->slots[wc->level])
6958                         goto reada;
6959
6960                 if (wc->stage == UPDATE_BACKREF &&
6961                     generation <= root->root_key.offset)
6962                         continue;
6963
6964                 /* We don't lock the tree block, it's OK to be racy here */
6965                 ret = btrfs_lookup_extent_info(trans, root, bytenr,
6966                                                wc->level - 1, 1, &refs,
6967                                                &flags);
6968                 /* We don't care about errors in readahead. */
6969                 if (ret < 0)
6970                         continue;
6971                 BUG_ON(refs == 0);
6972
6973                 if (wc->stage == DROP_REFERENCE) {
6974                         if (refs == 1)
6975                                 goto reada;
6976
6977                         if (wc->level == 1 &&
6978                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
6979                                 continue;
6980                         if (!wc->update_ref ||
6981                             generation <= root->root_key.offset)
6982                                 continue;
6983                         btrfs_node_key_to_cpu(eb, &key, slot);
6984                         ret = btrfs_comp_cpu_keys(&key,
6985                                                   &wc->update_progress);
6986                         if (ret < 0)
6987                                 continue;
6988                 } else {
6989                         if (wc->level == 1 &&
6990                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
6991                                 continue;
6992                 }
6993 reada:
6994                 ret = readahead_tree_block(root, bytenr, blocksize,
6995                                            generation);
6996                 if (ret)
6997                         break;
6998                 nread++;
6999         }
7000         wc->reada_slot = slot;
7001 }
7002
7003 /*
7004  * helper to process tree block while walking down the tree.
7005  *
7006  * when wc->stage == UPDATE_BACKREF, this function updates
7007  * back refs for pointers in the block.
7008  *
7009  * NOTE: return value 1 means we should stop walking down.
7010  */
7011 static noinline int walk_down_proc(struct btrfs_trans_handle *trans,
7012                                    struct btrfs_root *root,
7013                                    struct btrfs_path *path,
7014                                    struct walk_control *wc, int lookup_info)
7015 {
7016         int level = wc->level;
7017         struct extent_buffer *eb = path->nodes[level];
7018         u64 flag = BTRFS_BLOCK_FLAG_FULL_BACKREF;
7019         int ret;
7020
7021         if (wc->stage == UPDATE_BACKREF &&
7022             btrfs_header_owner(eb) != root->root_key.objectid)
7023                 return 1;
7024
7025         /*
7026          * when reference count of tree block is 1, it won't increase
7027          * again. once full backref flag is set, we never clear it.
7028          */
7029         if (lookup_info &&
7030             ((wc->stage == DROP_REFERENCE && wc->refs[level] != 1) ||
7031              (wc->stage == UPDATE_BACKREF && !(wc->flags[level] & flag)))) {
7032                 BUG_ON(!path->locks[level]);
7033                 ret = btrfs_lookup_extent_info(trans, root,
7034                                                eb->start, level, 1,
7035                                                &wc->refs[level],
7036                                                &wc->flags[level]);
7037                 BUG_ON(ret == -ENOMEM);
7038                 if (ret)
7039                         return ret;
7040                 BUG_ON(wc->refs[level] == 0);
7041         }
7042
7043         if (wc->stage == DROP_REFERENCE) {
7044                 if (wc->refs[level] > 1)
7045                         return 1;
7046
7047                 if (path->locks[level] && !wc->keep_locks) {
7048                         btrfs_tree_unlock_rw(eb, path->locks[level]);
7049                         path->locks[level] = 0;
7050                 }
7051                 return 0;
7052         }
7053
7054         /* wc->stage == UPDATE_BACKREF */
7055         if (!(wc->flags[level] & flag)) {
7056                 BUG_ON(!path->locks[level]);
7057                 ret = btrfs_inc_ref(trans, root, eb, 1, wc->for_reloc);
7058                 BUG_ON(ret); /* -ENOMEM */
7059                 ret = btrfs_dec_ref(trans, root, eb, 0, wc->for_reloc);
7060                 BUG_ON(ret); /* -ENOMEM */
7061                 ret = btrfs_set_disk_extent_flags(trans, root, eb->start,
7062                                                   eb->len, flag,
7063                                                   btrfs_header_level(eb), 0);
7064                 BUG_ON(ret); /* -ENOMEM */
7065                 wc->flags[level] |= flag;
7066         }
7067
7068         /*
7069          * the block is shared by multiple trees, so it's not good to
7070          * keep the tree lock
7071          */
7072         if (path->locks[level] && level > 0) {
7073                 btrfs_tree_unlock_rw(eb, path->locks[level]);
7074                 path->locks[level] = 0;
7075         }
7076         return 0;
7077 }
7078
7079 /*
7080  * helper to process tree block pointer.
7081  *
7082  * when wc->stage == DROP_REFERENCE, this function checks
7083  * reference count of the block pointed to. if the block
7084  * is shared and we need update back refs for the subtree
7085  * rooted at the block, this function changes wc->stage to
7086  * UPDATE_BACKREF. if the block is shared and there is no
7087  * need to update back, this function drops the reference
7088  * to the block.
7089  *
7090  * NOTE: return value 1 means we should stop walking down.
7091  */
7092 static noinline int do_walk_down(struct btrfs_trans_handle *trans,
7093                                  struct btrfs_root *root,
7094                                  struct btrfs_path *path,
7095                                  struct walk_control *wc, int *lookup_info)
7096 {
7097         u64 bytenr;
7098         u64 generation;
7099         u64 parent;
7100         u32 blocksize;
7101         struct btrfs_key key;
7102         struct extent_buffer *next;
7103         int level = wc->level;
7104         int reada = 0;
7105         int ret = 0;
7106
7107         generation = btrfs_node_ptr_generation(path->nodes[level],
7108                                                path->slots[level]);
7109         /*
7110          * if the lower level block was created before the snapshot
7111          * was created, we know there is no need to update back refs
7112          * for the subtree
7113          */
7114         if (wc->stage == UPDATE_BACKREF &&
7115             generation <= root->root_key.offset) {
7116                 *lookup_info = 1;
7117                 return 1;
7118         }
7119
7120         bytenr = btrfs_node_blockptr(path->nodes[level], path->slots[level]);
7121         blocksize = btrfs_level_size(root, level - 1);
7122
7123         next = btrfs_find_tree_block(root, bytenr, blocksize);
7124         if (!next) {
7125                 next = btrfs_find_create_tree_block(root, bytenr, blocksize);
7126                 if (!next)
7127                         return -ENOMEM;
7128                 reada = 1;
7129         }
7130         btrfs_tree_lock(next);
7131         btrfs_set_lock_blocking(next);
7132
7133         ret = btrfs_lookup_extent_info(trans, root, bytenr, level - 1, 1,
7134                                        &wc->refs[level - 1],
7135                                        &wc->flags[level - 1]);
7136         if (ret < 0) {
7137                 btrfs_tree_unlock(next);
7138                 return ret;
7139         }
7140
7141         if (unlikely(wc->refs[level - 1] == 0)) {
7142                 btrfs_err(root->fs_info, "Missing references.");
7143                 BUG();
7144         }
7145         *lookup_info = 0;
7146
7147         if (wc->stage == DROP_REFERENCE) {
7148                 if (wc->refs[level - 1] > 1) {
7149                         if (level == 1 &&
7150                             (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
7151                                 goto skip;
7152
7153                         if (!wc->update_ref ||
7154                             generation <= root->root_key.offset)
7155                                 goto skip;
7156
7157                         btrfs_node_key_to_cpu(path->nodes[level], &key,
7158                                               path->slots[level]);
7159                         ret = btrfs_comp_cpu_keys(&key, &wc->update_progress);
7160                         if (ret < 0)
7161                                 goto skip;
7162
7163                         wc->stage = UPDATE_BACKREF;
7164                         wc->shared_level = level - 1;
7165                 }
7166         } else {
7167                 if (level == 1 &&
7168                     (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
7169                         goto skip;
7170         }
7171
7172         if (!btrfs_buffer_uptodate(next, generation, 0)) {
7173                 btrfs_tree_unlock(next);
7174                 free_extent_buffer(next);
7175                 next = NULL;
7176                 *lookup_info = 1;
7177         }
7178
7179         if (!next) {
7180                 if (reada && level == 1)
7181                         reada_walk_down(trans, root, wc, path);
7182                 next = read_tree_block(root, bytenr, blocksize, generation);
7183                 if (!next || !extent_buffer_uptodate(next)) {
7184                         free_extent_buffer(next);
7185                         return -EIO;
7186                 }
7187                 btrfs_tree_lock(next);
7188                 btrfs_set_lock_blocking(next);
7189         }
7190
7191         level--;
7192         BUG_ON(level != btrfs_header_level(next));
7193         path->nodes[level] = next;
7194         path->slots[level] = 0;
7195         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
7196         wc->level = level;
7197         if (wc->level == 1)
7198                 wc->reada_slot = 0;
7199         return 0;
7200 skip:
7201         wc->refs[level - 1] = 0;
7202         wc->flags[level - 1] = 0;
7203         if (wc->stage == DROP_REFERENCE) {
7204                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF) {
7205                         parent = path->nodes[level]->start;
7206                 } else {
7207                         BUG_ON(root->root_key.objectid !=
7208                                btrfs_header_owner(path->nodes[level]));
7209                         parent = 0;
7210                 }
7211
7212                 ret = btrfs_free_extent(trans, root, bytenr, blocksize, parent,
7213                                 root->root_key.objectid, level - 1, 0, 0);
7214                 BUG_ON(ret); /* -ENOMEM */
7215         }
7216         btrfs_tree_unlock(next);
7217         free_extent_buffer(next);
7218         *lookup_info = 1;
7219         return 1;
7220 }
7221
7222 /*
7223  * helper to process tree block while walking up the tree.
7224  *
7225  * when wc->stage == DROP_REFERENCE, this function drops
7226  * reference count on the block.
7227  *
7228  * when wc->stage == UPDATE_BACKREF, this function changes
7229  * wc->stage back to DROP_REFERENCE if we changed wc->stage
7230  * to UPDATE_BACKREF previously while processing the block.
7231  *
7232  * NOTE: return value 1 means we should stop walking up.
7233  */
7234 static noinline int walk_up_proc(struct btrfs_trans_handle *trans,
7235                                  struct btrfs_root *root,
7236                                  struct btrfs_path *path,
7237                                  struct walk_control *wc)
7238 {
7239         int ret;
7240         int level = wc->level;
7241         struct extent_buffer *eb = path->nodes[level];
7242         u64 parent = 0;
7243
7244         if (wc->stage == UPDATE_BACKREF) {
7245                 BUG_ON(wc->shared_level < level);
7246                 if (level < wc->shared_level)
7247                         goto out;
7248
7249                 ret = find_next_key(path, level + 1, &wc->update_progress);
7250                 if (ret > 0)
7251                         wc->update_ref = 0;
7252
7253                 wc->stage = DROP_REFERENCE;
7254                 wc->shared_level = -1;
7255                 path->slots[level] = 0;
7256
7257                 /*
7258                  * check reference count again if the block isn't locked.
7259                  * we should start walking down the tree again if reference
7260                  * count is one.
7261                  */
7262                 if (!path->locks[level]) {
7263                         BUG_ON(level == 0);
7264                         btrfs_tree_lock(eb);
7265                         btrfs_set_lock_blocking(eb);
7266                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
7267
7268                         ret = btrfs_lookup_extent_info(trans, root,
7269                                                        eb->start, level, 1,
7270                                                        &wc->refs[level],
7271                                                        &wc->flags[level]);
7272                         if (ret < 0) {
7273                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
7274                                 path->locks[level] = 0;
7275                                 return ret;
7276                         }
7277                         BUG_ON(wc->refs[level] == 0);
7278                         if (wc->refs[level] == 1) {
7279                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
7280                                 path->locks[level] = 0;
7281                                 return 1;
7282                         }
7283                 }
7284         }
7285
7286         /* wc->stage == DROP_REFERENCE */
7287         BUG_ON(wc->refs[level] > 1 && !path->locks[level]);
7288
7289         if (wc->refs[level] == 1) {
7290                 if (level == 0) {
7291                         if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
7292                                 ret = btrfs_dec_ref(trans, root, eb, 1,
7293                                                     wc->for_reloc);
7294                         else
7295                                 ret = btrfs_dec_ref(trans, root, eb, 0,
7296                                                     wc->for_reloc);
7297                         BUG_ON(ret); /* -ENOMEM */
7298                 }
7299                 /* make block locked assertion in clean_tree_block happy */
7300                 if (!path->locks[level] &&
7301                     btrfs_header_generation(eb) == trans->transid) {
7302                         btrfs_tree_lock(eb);
7303                         btrfs_set_lock_blocking(eb);
7304                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
7305                 }
7306                 clean_tree_block(trans, root, eb);
7307         }
7308
7309         if (eb == root->node) {
7310                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
7311                         parent = eb->start;
7312                 else
7313                         BUG_ON(root->root_key.objectid !=
7314                                btrfs_header_owner(eb));
7315         } else {
7316                 if (wc->flags[level + 1] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
7317                         parent = path->nodes[level + 1]->start;
7318                 else
7319                         BUG_ON(root->root_key.objectid !=
7320                                btrfs_header_owner(path->nodes[level + 1]));
7321         }
7322
7323         btrfs_free_tree_block(trans, root, eb, parent, wc->refs[level] == 1);
7324 out:
7325         wc->refs[level] = 0;
7326         wc->flags[level] = 0;
7327         return 0;
7328 }
7329
7330 static noinline int walk_down_tree(struct btrfs_trans_handle *trans,
7331                                    struct btrfs_root *root,
7332                                    struct btrfs_path *path,
7333                                    struct walk_control *wc)
7334 {
7335         int level = wc->level;
7336         int lookup_info = 1;
7337         int ret;
7338
7339         while (level >= 0) {
7340                 ret = walk_down_proc(trans, root, path, wc, lookup_info);
7341                 if (ret > 0)
7342                         break;
7343
7344                 if (level == 0)
7345                         break;
7346
7347                 if (path->slots[level] >=
7348                     btrfs_header_nritems(path->nodes[level]))
7349                         break;
7350
7351                 ret = do_walk_down(trans, root, path, wc, &lookup_info);
7352                 if (ret > 0) {
7353                         path->slots[level]++;
7354                         continue;
7355                 } else if (ret < 0)
7356                         return ret;
7357                 level = wc->level;
7358         }
7359         return 0;
7360 }
7361
7362 static noinline int walk_up_tree(struct btrfs_trans_handle *trans,
7363                                  struct btrfs_root *root,
7364                                  struct btrfs_path *path,
7365                                  struct walk_control *wc, int max_level)
7366 {
7367         int level = wc->level;
7368         int ret;
7369
7370         path->slots[level] = btrfs_header_nritems(path->nodes[level]);
7371         while (level < max_level && path->nodes[level]) {
7372                 wc->level = level;
7373                 if (path->slots[level] + 1 <
7374                     btrfs_header_nritems(path->nodes[level])) {
7375                         path->slots[level]++;
7376                         return 0;
7377                 } else {
7378                         ret = walk_up_proc(trans, root, path, wc);
7379                         if (ret > 0)
7380                                 return 0;
7381
7382                         if (path->locks[level]) {
7383                                 btrfs_tree_unlock_rw(path->nodes[level],
7384                                                      path->locks[level]);
7385                                 path->locks[level] = 0;
7386                         }
7387                         free_extent_buffer(path->nodes[level]);
7388                         path->nodes[level] = NULL;
7389                         level++;
7390                 }
7391         }
7392         return 1;
7393 }
7394
7395 /*
7396  * drop a subvolume tree.
7397  *
7398  * this function traverses the tree freeing any blocks that only
7399  * referenced by the tree.
7400  *
7401  * when a shared tree block is found. this function decreases its
7402  * reference count by one. if update_ref is true, this function
7403  * also make sure backrefs for the shared block and all lower level
7404  * blocks are properly updated.
7405  *
7406  * If called with for_reloc == 0, may exit early with -EAGAIN
7407  */
7408 int btrfs_drop_snapshot(struct btrfs_root *root,
7409                          struct btrfs_block_rsv *block_rsv, int update_ref,
7410                          int for_reloc)
7411 {
7412         struct btrfs_path *path;
7413         struct btrfs_trans_handle *trans;
7414         struct btrfs_root *tree_root = root->fs_info->tree_root;
7415         struct btrfs_root_item *root_item = &root->root_item;
7416         struct walk_control *wc;
7417         struct btrfs_key key;
7418         int err = 0;
7419         int ret;
7420         int level;
7421
7422         path = btrfs_alloc_path();
7423         if (!path) {
7424                 err = -ENOMEM;
7425                 goto out;
7426         }
7427
7428         wc = kzalloc(sizeof(*wc), GFP_NOFS);
7429         if (!wc) {
7430                 btrfs_free_path(path);
7431                 err = -ENOMEM;
7432                 goto out;
7433         }
7434
7435         trans = btrfs_start_transaction(tree_root, 0);
7436         if (IS_ERR(trans)) {
7437                 err = PTR_ERR(trans);
7438                 goto out_free;
7439         }
7440
7441         if (block_rsv)
7442                 trans->block_rsv = block_rsv;
7443
7444         if (btrfs_disk_key_objectid(&root_item->drop_progress) == 0) {
7445                 level = btrfs_header_level(root->node);
7446                 path->nodes[level] = btrfs_lock_root_node(root);
7447                 btrfs_set_lock_blocking(path->nodes[level]);
7448                 path->slots[level] = 0;
7449                 path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
7450                 memset(&wc->update_progress, 0,
7451                        sizeof(wc->update_progress));
7452         } else {
7453                 btrfs_disk_key_to_cpu(&key, &root_item->drop_progress);
7454                 memcpy(&wc->update_progress, &key,
7455                        sizeof(wc->update_progress));
7456
7457                 level = root_item->drop_level;
7458                 BUG_ON(level == 0);
7459                 path->lowest_level = level;
7460                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
7461                 path->lowest_level = 0;
7462                 if (ret < 0) {
7463                         err = ret;
7464                         goto out_end_trans;
7465                 }
7466                 WARN_ON(ret > 0);
7467
7468                 /*
7469                  * unlock our path, this is safe because only this
7470                  * function is allowed to delete this snapshot
7471                  */
7472                 btrfs_unlock_up_safe(path, 0);
7473
7474                 level = btrfs_header_level(root->node);
7475                 while (1) {
7476                         btrfs_tree_lock(path->nodes[level]);
7477                         btrfs_set_lock_blocking(path->nodes[level]);
7478
7479                         ret = btrfs_lookup_extent_info(trans, root,
7480                                                 path->nodes[level]->start,
7481                                                 level, 1, &wc->refs[level],
7482                                                 &wc->flags[level]);
7483                         if (ret < 0) {
7484                                 err = ret;
7485                                 goto out_end_trans;
7486                         }
7487                         BUG_ON(wc->refs[level] == 0);
7488
7489                         if (level == root_item->drop_level)
7490                                 break;
7491
7492                         btrfs_tree_unlock(path->nodes[level]);
7493                         WARN_ON(wc->refs[level] != 1);
7494                         level--;
7495                 }
7496         }
7497
7498         wc->level = level;
7499         wc->shared_level = -1;
7500         wc->stage = DROP_REFERENCE;
7501         wc->update_ref = update_ref;
7502         wc->keep_locks = 0;
7503         wc->for_reloc = for_reloc;
7504         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(root);
7505
7506         while (1) {
7507                 if (!for_reloc && btrfs_need_cleaner_sleep(root)) {
7508                         pr_debug("btrfs: drop snapshot early exit\n");
7509                         err = -EAGAIN;
7510                         goto out_end_trans;
7511                 }
7512
7513                 ret = walk_down_tree(trans, root, path, wc);
7514                 if (ret < 0) {
7515                         err = ret;
7516                         break;
7517                 }
7518
7519                 ret = walk_up_tree(trans, root, path, wc, BTRFS_MAX_LEVEL);
7520                 if (ret < 0) {
7521                         err = ret;
7522                         break;
7523                 }
7524
7525                 if (ret > 0) {
7526                         BUG_ON(wc->stage != DROP_REFERENCE);
7527                         break;
7528                 }
7529
7530                 if (wc->stage == DROP_REFERENCE) {
7531                         level = wc->level;
7532                         btrfs_node_key(path->nodes[level],
7533                                        &root_item->drop_progress,
7534                                        path->slots[level]);
7535                         root_item->drop_level = level;
7536                 }
7537
7538                 BUG_ON(wc->level == 0);
7539                 if (btrfs_should_end_transaction(trans, tree_root)) {
7540                         ret = btrfs_update_root(trans, tree_root,
7541                                                 &root->root_key,
7542                                                 root_item);
7543                         if (ret) {
7544                                 btrfs_abort_transaction(trans, tree_root, ret);
7545                                 err = ret;
7546                                 goto out_end_trans;
7547                         }
7548
7549                         btrfs_end_transaction_throttle(trans, tree_root);
7550                         trans = btrfs_start_transaction(tree_root, 0);
7551                         if (IS_ERR(trans)) {
7552                                 err = PTR_ERR(trans);
7553                                 goto out_free;
7554                         }
7555                         if (block_rsv)
7556                                 trans->block_rsv = block_rsv;
7557                 }
7558         }
7559         btrfs_release_path(path);
7560         if (err)
7561                 goto out_end_trans;
7562
7563         ret = btrfs_del_root(trans, tree_root, &root->root_key);
7564         if (ret) {
7565                 btrfs_abort_transaction(trans, tree_root, ret);
7566                 goto out_end_trans;
7567         }
7568
7569         if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) {
7570                 ret = btrfs_find_root(tree_root, &root->root_key, path,
7571                                       NULL, NULL);
7572                 if (ret < 0) {
7573                         btrfs_abort_transaction(trans, tree_root, ret);
7574                         err = ret;
7575                         goto out_end_trans;
7576                 } else if (ret > 0) {
7577                         /* if we fail to delete the orphan item this time
7578                          * around, it'll get picked up the next time.
7579                          *
7580                          * The most common failure here is just -ENOENT.
7581                          */
7582                         btrfs_del_orphan_item(trans, tree_root,
7583                                               root->root_key.objectid);
7584                 }
7585         }
7586
7587         if (root->in_radix) {
7588                 btrfs_drop_and_free_fs_root(tree_root->fs_info, root);
7589         } else {
7590                 free_extent_buffer(root->node);
7591                 free_extent_buffer(root->commit_root);
7592                 btrfs_put_fs_root(root);
7593         }
7594 out_end_trans:
7595         btrfs_end_transaction_throttle(trans, tree_root);
7596 out_free:
7597         kfree(wc);
7598         btrfs_free_path(path);
7599 out:
7600         if (err)
7601                 btrfs_std_error(root->fs_info, err);
7602         return err;
7603 }
7604
7605 /*
7606  * drop subtree rooted at tree block 'node'.
7607  *
7608  * NOTE: this function will unlock and release tree block 'node'
7609  * only used by relocation code
7610  */
7611 int btrfs_drop_subtree(struct btrfs_trans_handle *trans,
7612                         struct btrfs_root *root,
7613                         struct extent_buffer *node,
7614                         struct extent_buffer *parent)
7615 {
7616         struct btrfs_path *path;
7617         struct walk_control *wc;
7618         int level;
7619         int parent_level;
7620         int ret = 0;
7621         int wret;
7622
7623         BUG_ON(root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID);
7624
7625         path = btrfs_alloc_path();
7626         if (!path)
7627                 return -ENOMEM;
7628
7629         wc = kzalloc(sizeof(*wc), GFP_NOFS);
7630         if (!wc) {
7631                 btrfs_free_path(path);
7632                 return -ENOMEM;
7633         }
7634
7635         btrfs_assert_tree_locked(parent);
7636         parent_level = btrfs_header_level(parent);
7637         extent_buffer_get(parent);
7638         path->nodes[parent_level] = parent;
7639         path->slots[parent_level] = btrfs_header_nritems(parent);
7640
7641         btrfs_assert_tree_locked(node);
7642         level = btrfs_header_level(node);
7643         path->nodes[level] = node;
7644         path->slots[level] = 0;
7645         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
7646
7647         wc->refs[parent_level] = 1;
7648         wc->flags[parent_level] = BTRFS_BLOCK_FLAG_FULL_BACKREF;
7649         wc->level = level;
7650         wc->shared_level = -1;
7651         wc->stage = DROP_REFERENCE;
7652         wc->update_ref = 0;
7653         wc->keep_locks = 1;
7654         wc->for_reloc = 1;
7655         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(root);
7656
7657         while (1) {
7658                 wret = walk_down_tree(trans, root, path, wc);
7659                 if (wret < 0) {
7660                         ret = wret;
7661                         break;
7662                 }
7663
7664                 wret = walk_up_tree(trans, root, path, wc, parent_level);
7665                 if (wret < 0)
7666                         ret = wret;
7667                 if (wret != 0)
7668                         break;
7669         }
7670
7671         kfree(wc);
7672         btrfs_free_path(path);
7673         return ret;
7674 }
7675
7676 static u64 update_block_group_flags(struct btrfs_root *root, u64 flags)
7677 {
7678         u64 num_devices;
7679         u64 stripped;
7680
7681         /*
7682          * if restripe for this chunk_type is on pick target profile and
7683          * return, otherwise do the usual balance
7684          */
7685         stripped = get_restripe_target(root->fs_info, flags);
7686         if (stripped)
7687                 return extended_to_chunk(stripped);
7688
7689         /*
7690          * we add in the count of missing devices because we want
7691          * to make sure that any RAID levels on a degraded FS
7692          * continue to be honored.
7693          */
7694         num_devices = root->fs_info->fs_devices->rw_devices +
7695                 root->fs_info->fs_devices->missing_devices;
7696
7697         stripped = BTRFS_BLOCK_GROUP_RAID0 |
7698                 BTRFS_BLOCK_GROUP_RAID5 | BTRFS_BLOCK_GROUP_RAID6 |
7699                 BTRFS_BLOCK_GROUP_RAID1 | BTRFS_BLOCK_GROUP_RAID10;
7700
7701         if (num_devices == 1) {
7702                 stripped |= BTRFS_BLOCK_GROUP_DUP;
7703                 stripped = flags & ~stripped;
7704
7705                 /* turn raid0 into single device chunks */
7706                 if (flags & BTRFS_BLOCK_GROUP_RAID0)
7707                         return stripped;
7708
7709                 /* turn mirroring into duplication */
7710                 if (flags & (BTRFS_BLOCK_GROUP_RAID1 |
7711                              BTRFS_BLOCK_GROUP_RAID10))
7712                         return stripped | BTRFS_BLOCK_GROUP_DUP;
7713         } else {
7714                 /* they already had raid on here, just return */
7715                 if (flags & stripped)
7716                         return flags;
7717
7718                 stripped |= BTRFS_BLOCK_GROUP_DUP;
7719                 stripped = flags & ~stripped;
7720
7721                 /* switch duplicated blocks with raid1 */
7722                 if (flags & BTRFS_BLOCK_GROUP_DUP)
7723                         return stripped | BTRFS_BLOCK_GROUP_RAID1;
7724
7725                 /* this is drive concat, leave it alone */
7726         }
7727
7728         return flags;
7729 }
7730
7731 static int set_block_group_ro(struct btrfs_block_group_cache *cache, int force)
7732 {
7733         struct btrfs_space_info *sinfo = cache->space_info;
7734         u64 num_bytes;
7735         u64 min_allocable_bytes;
7736         int ret = -ENOSPC;
7737
7738
7739         /*
7740          * We need some metadata space and system metadata space for
7741          * allocating chunks in some corner cases until we force to set
7742          * it to be readonly.
7743          */
7744         if ((sinfo->flags &
7745              (BTRFS_BLOCK_GROUP_SYSTEM | BTRFS_BLOCK_GROUP_METADATA)) &&
7746             !force)
7747                 min_allocable_bytes = 1 * 1024 * 1024;
7748         else
7749                 min_allocable_bytes = 0;
7750
7751         spin_lock(&sinfo->lock);
7752         spin_lock(&cache->lock);
7753
7754         if (cache->ro) {
7755                 ret = 0;
7756                 goto out;
7757         }
7758
7759         num_bytes = cache->key.offset - cache->reserved - cache->pinned -
7760                     cache->bytes_super - btrfs_block_group_used(&cache->item);
7761
7762         if (sinfo->bytes_used + sinfo->bytes_reserved + sinfo->bytes_pinned +
7763             sinfo->bytes_may_use + sinfo->bytes_readonly + num_bytes +
7764             min_allocable_bytes <= sinfo->total_bytes) {
7765                 sinfo->bytes_readonly += num_bytes;
7766                 cache->ro = 1;
7767                 ret = 0;
7768         }
7769 out:
7770         spin_unlock(&cache->lock);
7771         spin_unlock(&sinfo->lock);
7772         return ret;
7773 }
7774
7775 int btrfs_set_block_group_ro(struct btrfs_root *root,
7776                              struct btrfs_block_group_cache *cache)
7777
7778 {
7779         struct btrfs_trans_handle *trans;
7780         u64 alloc_flags;
7781         int ret;
7782
7783         BUG_ON(cache->ro);
7784
7785         trans = btrfs_join_transaction(root);
7786         if (IS_ERR(trans))
7787                 return PTR_ERR(trans);
7788
7789         alloc_flags = update_block_group_flags(root, cache->flags);
7790         if (alloc_flags != cache->flags) {
7791                 ret = do_chunk_alloc(trans, root, alloc_flags,
7792                                      CHUNK_ALLOC_FORCE);
7793                 if (ret < 0)
7794                         goto out;
7795         }
7796
7797         ret = set_block_group_ro(cache, 0);
7798         if (!ret)
7799                 goto out;
7800         alloc_flags = get_alloc_profile(root, cache->space_info->flags);
7801         ret = do_chunk_alloc(trans, root, alloc_flags,
7802                              CHUNK_ALLOC_FORCE);
7803         if (ret < 0)
7804                 goto out;
7805         ret = set_block_group_ro(cache, 0);
7806 out:
7807         btrfs_end_transaction(trans, root);
7808         return ret;
7809 }
7810
7811 int btrfs_force_chunk_alloc(struct btrfs_trans_handle *trans,
7812                             struct btrfs_root *root, u64 type)
7813 {
7814         u64 alloc_flags = get_alloc_profile(root, type);
7815         return do_chunk_alloc(trans, root, alloc_flags,
7816                               CHUNK_ALLOC_FORCE);
7817 }
7818
7819 /*
7820  * helper to account the unused space of all the readonly block group in the
7821  * list. takes mirrors into account.
7822  */
7823 static u64 __btrfs_get_ro_block_group_free_space(struct list_head *groups_list)
7824 {
7825         struct btrfs_block_group_cache *block_group;
7826         u64 free_bytes = 0;
7827         int factor;
7828
7829         list_for_each_entry(block_group, groups_list, list) {
7830                 spin_lock(&block_group->lock);
7831
7832                 if (!block_group->ro) {
7833                         spin_unlock(&block_group->lock);
7834                         continue;
7835                 }
7836
7837                 if (block_group->flags & (BTRFS_BLOCK_GROUP_RAID1 |
7838                                           BTRFS_BLOCK_GROUP_RAID10 |
7839                                           BTRFS_BLOCK_GROUP_DUP))
7840                         factor = 2;
7841                 else
7842                         factor = 1;
7843
7844                 free_bytes += (block_group->key.offset -
7845                                btrfs_block_group_used(&block_group->item)) *
7846                                factor;
7847
7848                 spin_unlock(&block_group->lock);
7849         }
7850
7851         return free_bytes;
7852 }
7853
7854 /*
7855  * helper to account the unused space of all the readonly block group in the
7856  * space_info. takes mirrors into account.
7857  */
7858 u64 btrfs_account_ro_block_groups_free_space(struct btrfs_space_info *sinfo)
7859 {
7860         int i;
7861         u64 free_bytes = 0;
7862
7863         spin_lock(&sinfo->lock);
7864
7865         for(i = 0; i < BTRFS_NR_RAID_TYPES; i++)
7866                 if (!list_empty(&sinfo->block_groups[i]))
7867                         free_bytes += __btrfs_get_ro_block_group_free_space(
7868                                                 &sinfo->block_groups[i]);
7869
7870         spin_unlock(&sinfo->lock);
7871
7872         return free_bytes;
7873 }
7874
7875 void btrfs_set_block_group_rw(struct btrfs_root *root,
7876                               struct btrfs_block_group_cache *cache)
7877 {
7878         struct btrfs_space_info *sinfo = cache->space_info;
7879         u64 num_bytes;
7880
7881         BUG_ON(!cache->ro);
7882
7883         spin_lock(&sinfo->lock);
7884         spin_lock(&cache->lock);
7885         num_bytes = cache->key.offset - cache->reserved - cache->pinned -
7886                     cache->bytes_super - btrfs_block_group_used(&cache->item);
7887         sinfo->bytes_readonly -= num_bytes;
7888         cache->ro = 0;
7889         spin_unlock(&cache->lock);
7890         spin_unlock(&sinfo->lock);
7891 }
7892
7893 /*
7894  * checks to see if its even possible to relocate this block group.
7895  *
7896  * @return - -1 if it's not a good idea to relocate this block group, 0 if its
7897  * ok to go ahead and try.
7898  */
7899 int btrfs_can_relocate(struct btrfs_root *root, u64 bytenr)
7900 {
7901         struct btrfs_block_group_cache *block_group;
7902         struct btrfs_space_info *space_info;
7903         struct btrfs_fs_devices *fs_devices = root->fs_info->fs_devices;
7904         struct btrfs_device *device;
7905         u64 min_free;
7906         u64 dev_min = 1;
7907         u64 dev_nr = 0;
7908         u64 target;
7909         int index;
7910         int full = 0;
7911         int ret = 0;
7912
7913         block_group = btrfs_lookup_block_group(root->fs_info, bytenr);
7914
7915         /* odd, couldn't find the block group, leave it alone */
7916         if (!block_group)
7917                 return -1;
7918
7919         min_free = btrfs_block_group_used(&block_group->item);
7920
7921         /* no bytes used, we're good */
7922         if (!min_free)
7923                 goto out;
7924
7925         space_info = block_group->space_info;
7926         spin_lock(&space_info->lock);
7927
7928         full = space_info->full;
7929
7930         /*
7931          * if this is the last block group we have in this space, we can't
7932          * relocate it unless we're able to allocate a new chunk below.
7933          *
7934          * Otherwise, we need to make sure we have room in the space to handle
7935          * all of the extents from this block group.  If we can, we're good
7936          */
7937         if ((space_info->total_bytes != block_group->key.offset) &&
7938             (space_info->bytes_used + space_info->bytes_reserved +
7939              space_info->bytes_pinned + space_info->bytes_readonly +
7940              min_free < space_info->total_bytes)) {
7941                 spin_unlock(&space_info->lock);
7942                 goto out;
7943         }
7944         spin_unlock(&space_info->lock);
7945
7946         /*
7947          * ok we don't have enough space, but maybe we have free space on our
7948          * devices to allocate new chunks for relocation, so loop through our
7949          * alloc devices and guess if we have enough space.  if this block
7950          * group is going to be restriped, run checks against the target
7951          * profile instead of the current one.
7952          */
7953         ret = -1;
7954
7955         /*
7956          * index:
7957          *      0: raid10
7958          *      1: raid1
7959          *      2: dup
7960          *      3: raid0
7961          *      4: single
7962          */
7963         target = get_restripe_target(root->fs_info, block_group->flags);
7964         if (target) {
7965                 index = __get_raid_index(extended_to_chunk(target));
7966         } else {
7967                 /*
7968                  * this is just a balance, so if we were marked as full
7969                  * we know there is no space for a new chunk
7970                  */
7971                 if (full)
7972                         goto out;
7973
7974                 index = get_block_group_index(block_group);
7975         }
7976
7977         if (index == BTRFS_RAID_RAID10) {
7978                 dev_min = 4;
7979                 /* Divide by 2 */
7980                 min_free >>= 1;
7981         } else if (index == BTRFS_RAID_RAID1) {
7982                 dev_min = 2;
7983         } else if (index == BTRFS_RAID_DUP) {
7984                 /* Multiply by 2 */
7985                 min_free <<= 1;
7986         } else if (index == BTRFS_RAID_RAID0) {
7987                 dev_min = fs_devices->rw_devices;
7988                 do_div(min_free, dev_min);
7989         }
7990
7991         mutex_lock(&root->fs_info->chunk_mutex);
7992         list_for_each_entry(device, &fs_devices->alloc_list, dev_alloc_list) {
7993                 u64 dev_offset;
7994
7995                 /*
7996                  * check to make sure we can actually find a chunk with enough
7997                  * space to fit our block group in.
7998                  */
7999                 if (device->total_bytes > device->bytes_used + min_free &&
8000                     !device->is_tgtdev_for_dev_replace) {
8001                         ret = find_free_dev_extent(device, min_free,
8002                                                    &dev_offset, NULL);
8003                         if (!ret)
8004                                 dev_nr++;
8005
8006                         if (dev_nr >= dev_min)
8007                                 break;
8008
8009                         ret = -1;
8010                 }
8011         }
8012         mutex_unlock(&root->fs_info->chunk_mutex);
8013 out:
8014         btrfs_put_block_group(block_group);
8015         return ret;
8016 }
8017
8018 static int find_first_block_group(struct btrfs_root *root,
8019                 struct btrfs_path *path, struct btrfs_key *key)
8020 {
8021         int ret = 0;
8022         struct btrfs_key found_key;
8023         struct extent_buffer *leaf;
8024         int slot;
8025
8026         ret = btrfs_search_slot(NULL, root, key, path, 0, 0);
8027         if (ret < 0)
8028                 goto out;
8029
8030         while (1) {
8031                 slot = path->slots[0];
8032                 leaf = path->nodes[0];
8033                 if (slot >= btrfs_header_nritems(leaf)) {
8034                         ret = btrfs_next_leaf(root, path);
8035                         if (ret == 0)
8036                                 continue;
8037                         if (ret < 0)
8038                                 goto out;
8039                         break;
8040                 }
8041                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
8042
8043                 if (found_key.objectid >= key->objectid &&
8044                     found_key.type == BTRFS_BLOCK_GROUP_ITEM_KEY) {
8045                         ret = 0;
8046                         goto out;
8047                 }
8048                 path->slots[0]++;
8049         }
8050 out:
8051         return ret;
8052 }
8053
8054 void btrfs_put_block_group_cache(struct btrfs_fs_info *info)
8055 {
8056         struct btrfs_block_group_cache *block_group;
8057         u64 last = 0;
8058
8059         while (1) {
8060                 struct inode *inode;
8061
8062                 block_group = btrfs_lookup_first_block_group(info, last);
8063                 while (block_group) {
8064                         spin_lock(&block_group->lock);
8065                         if (block_group->iref)
8066                                 break;
8067                         spin_unlock(&block_group->lock);
8068                         block_group = next_block_group(info->tree_root,
8069                                                        block_group);
8070                 }
8071                 if (!block_group) {
8072                         if (last == 0)
8073                                 break;
8074                         last = 0;
8075                         continue;
8076                 }
8077
8078                 inode = block_group->inode;
8079                 block_group->iref = 0;
8080                 block_group->inode = NULL;
8081                 spin_unlock(&block_group->lock);
8082                 iput(inode);
8083                 last = block_group->key.objectid + block_group->key.offset;
8084                 btrfs_put_block_group(block_group);
8085         }
8086 }
8087
8088 int btrfs_free_block_groups(struct btrfs_fs_info *info)
8089 {
8090         struct btrfs_block_group_cache *block_group;
8091         struct btrfs_space_info *space_info;
8092         struct btrfs_caching_control *caching_ctl;
8093         struct rb_node *n;
8094
8095         down_write(&info->extent_commit_sem);
8096         while (!list_empty(&info->caching_block_groups)) {
8097                 caching_ctl = list_entry(info->caching_block_groups.next,
8098                                          struct btrfs_caching_control, list);
8099                 list_del(&caching_ctl->list);
8100                 put_caching_control(caching_ctl);
8101         }
8102         up_write(&info->extent_commit_sem);
8103
8104         spin_lock(&info->block_group_cache_lock);
8105         while ((n = rb_last(&info->block_group_cache_tree)) != NULL) {
8106                 block_group = rb_entry(n, struct btrfs_block_group_cache,
8107                                        cache_node);
8108                 rb_erase(&block_group->cache_node,
8109                          &info->block_group_cache_tree);
8110                 spin_unlock(&info->block_group_cache_lock);
8111
8112                 down_write(&block_group->space_info->groups_sem);
8113                 list_del(&block_group->list);
8114                 up_write(&block_group->space_info->groups_sem);
8115
8116                 if (block_group->cached == BTRFS_CACHE_STARTED)
8117                         wait_block_group_cache_done(block_group);
8118
8119                 /*
8120                  * We haven't cached this block group, which means we could
8121                  * possibly have excluded extents on this block group.
8122                  */
8123                 if (block_group->cached == BTRFS_CACHE_NO)
8124                         free_excluded_extents(info->extent_root, block_group);
8125
8126                 btrfs_remove_free_space_cache(block_group);
8127                 btrfs_put_block_group(block_group);
8128
8129                 spin_lock(&info->block_group_cache_lock);
8130         }
8131         spin_unlock(&info->block_group_cache_lock);
8132
8133         /* now that all the block groups are freed, go through and
8134          * free all the space_info structs.  This is only called during
8135          * the final stages of unmount, and so we know nobody is
8136          * using them.  We call synchronize_rcu() once before we start,
8137          * just to be on the safe side.
8138          */
8139         synchronize_rcu();
8140
8141         release_global_block_rsv(info);
8142
8143         while(!list_empty(&info->space_info)) {
8144                 space_info = list_entry(info->space_info.next,
8145                                         struct btrfs_space_info,
8146                                         list);
8147                 if (btrfs_test_opt(info->tree_root, ENOSPC_DEBUG)) {
8148                         if (space_info->bytes_pinned > 0 ||
8149                             space_info->bytes_reserved > 0 ||
8150                             space_info->bytes_may_use > 0) {
8151                                 WARN_ON(1);
8152                                 dump_space_info(space_info, 0, 0);
8153                         }
8154                 }
8155                 list_del(&space_info->list);
8156                 kfree(space_info);
8157         }
8158         return 0;
8159 }
8160
8161 static void __link_block_group(struct btrfs_space_info *space_info,
8162                                struct btrfs_block_group_cache *cache)
8163 {
8164         int index = get_block_group_index(cache);
8165
8166         down_write(&space_info->groups_sem);
8167         list_add_tail(&cache->list, &space_info->block_groups[index]);
8168         up_write(&space_info->groups_sem);
8169 }
8170
8171 int btrfs_read_block_groups(struct btrfs_root *root)
8172 {
8173         struct btrfs_path *path;
8174         int ret;
8175         struct btrfs_block_group_cache *cache;
8176         struct btrfs_fs_info *info = root->fs_info;
8177         struct btrfs_space_info *space_info;
8178         struct btrfs_key key;
8179         struct btrfs_key found_key;
8180         struct extent_buffer *leaf;
8181         int need_clear = 0;
8182         u64 cache_gen;
8183
8184         root = info->extent_root;
8185         key.objectid = 0;
8186         key.offset = 0;
8187         btrfs_set_key_type(&key, BTRFS_BLOCK_GROUP_ITEM_KEY);
8188         path = btrfs_alloc_path();
8189         if (!path)
8190                 return -ENOMEM;
8191         path->reada = 1;
8192
8193         cache_gen = btrfs_super_cache_generation(root->fs_info->super_copy);
8194         if (btrfs_test_opt(root, SPACE_CACHE) &&
8195             btrfs_super_generation(root->fs_info->super_copy) != cache_gen)
8196                 need_clear = 1;
8197         if (btrfs_test_opt(root, CLEAR_CACHE))
8198                 need_clear = 1;
8199
8200         while (1) {
8201                 ret = find_first_block_group(root, path, &key);
8202                 if (ret > 0)
8203                         break;
8204                 if (ret != 0)
8205                         goto error;
8206                 leaf = path->nodes[0];
8207                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
8208                 cache = kzalloc(sizeof(*cache), GFP_NOFS);
8209                 if (!cache) {
8210                         ret = -ENOMEM;
8211                         goto error;
8212                 }
8213                 cache->free_space_ctl = kzalloc(sizeof(*cache->free_space_ctl),
8214                                                 GFP_NOFS);
8215                 if (!cache->free_space_ctl) {
8216                         kfree(cache);
8217                         ret = -ENOMEM;
8218                         goto error;
8219                 }
8220
8221                 atomic_set(&cache->count, 1);
8222                 spin_lock_init(&cache->lock);
8223                 cache->fs_info = info;
8224                 INIT_LIST_HEAD(&cache->list);
8225                 INIT_LIST_HEAD(&cache->cluster_list);
8226
8227                 if (need_clear) {
8228                         /*
8229                          * When we mount with old space cache, we need to
8230                          * set BTRFS_DC_CLEAR and set dirty flag.
8231                          *
8232                          * a) Setting 'BTRFS_DC_CLEAR' makes sure that we
8233                          *    truncate the old free space cache inode and
8234                          *    setup a new one.
8235                          * b) Setting 'dirty flag' makes sure that we flush
8236                          *    the new space cache info onto disk.
8237                          */
8238                         cache->disk_cache_state = BTRFS_DC_CLEAR;
8239                         if (btrfs_test_opt(root, SPACE_CACHE))
8240                                 cache->dirty = 1;
8241                 }
8242
8243                 read_extent_buffer(leaf, &cache->item,
8244                                    btrfs_item_ptr_offset(leaf, path->slots[0]),
8245                                    sizeof(cache->item));
8246                 memcpy(&cache->key, &found_key, sizeof(found_key));
8247
8248                 key.objectid = found_key.objectid + found_key.offset;
8249                 btrfs_release_path(path);
8250                 cache->flags = btrfs_block_group_flags(&cache->item);
8251                 cache->sectorsize = root->sectorsize;
8252                 cache->full_stripe_len = btrfs_full_stripe_len(root,
8253                                                &root->fs_info->mapping_tree,
8254                                                found_key.objectid);
8255                 btrfs_init_free_space_ctl(cache);
8256
8257                 /*
8258                  * We need to exclude the super stripes now so that the space
8259                  * info has super bytes accounted for, otherwise we'll think
8260                  * we have more space than we actually do.
8261                  */
8262                 ret = exclude_super_stripes(root, cache);
8263                 if (ret) {
8264                         /*
8265                          * We may have excluded something, so call this just in
8266                          * case.
8267                          */
8268                         free_excluded_extents(root, cache);
8269                         kfree(cache->free_space_ctl);
8270                         kfree(cache);
8271                         goto error;
8272                 }
8273
8274                 /*
8275                  * check for two cases, either we are full, and therefore
8276                  * don't need to bother with the caching work since we won't
8277                  * find any space, or we are empty, and we can just add all
8278                  * the space in and be done with it.  This saves us _alot_ of
8279                  * time, particularly in the full case.
8280                  */
8281                 if (found_key.offset == btrfs_block_group_used(&cache->item)) {
8282                         cache->last_byte_to_unpin = (u64)-1;
8283                         cache->cached = BTRFS_CACHE_FINISHED;
8284                         free_excluded_extents(root, cache);
8285                 } else if (btrfs_block_group_used(&cache->item) == 0) {
8286                         cache->last_byte_to_unpin = (u64)-1;
8287                         cache->cached = BTRFS_CACHE_FINISHED;
8288                         add_new_free_space(cache, root->fs_info,
8289                                            found_key.objectid,
8290                                            found_key.objectid +
8291                                            found_key.offset);
8292                         free_excluded_extents(root, cache);
8293                 }
8294
8295                 ret = btrfs_add_block_group_cache(root->fs_info, cache);
8296                 if (ret) {
8297                         btrfs_remove_free_space_cache(cache);
8298                         btrfs_put_block_group(cache);
8299                         goto error;
8300                 }
8301
8302                 ret = update_space_info(info, cache->flags, found_key.offset,
8303                                         btrfs_block_group_used(&cache->item),
8304                                         &space_info);
8305                 if (ret) {
8306                         btrfs_remove_free_space_cache(cache);
8307                         spin_lock(&info->block_group_cache_lock);
8308                         rb_erase(&cache->cache_node,
8309                                  &info->block_group_cache_tree);
8310                         spin_unlock(&info->block_group_cache_lock);
8311                         btrfs_put_block_group(cache);
8312                         goto error;
8313                 }
8314
8315                 cache->space_info = space_info;
8316                 spin_lock(&cache->space_info->lock);
8317                 cache->space_info->bytes_readonly += cache->bytes_super;
8318                 spin_unlock(&cache->space_info->lock);
8319
8320                 __link_block_group(space_info, cache);
8321
8322                 set_avail_alloc_bits(root->fs_info, cache->flags);
8323                 if (btrfs_chunk_readonly(root, cache->key.objectid))
8324                         set_block_group_ro(cache, 1);
8325         }
8326
8327         list_for_each_entry_rcu(space_info, &root->fs_info->space_info, list) {
8328                 if (!(get_alloc_profile(root, space_info->flags) &
8329                       (BTRFS_BLOCK_GROUP_RAID10 |
8330                        BTRFS_BLOCK_GROUP_RAID1 |
8331                        BTRFS_BLOCK_GROUP_RAID5 |
8332                        BTRFS_BLOCK_GROUP_RAID6 |
8333                        BTRFS_BLOCK_GROUP_DUP)))
8334                         continue;
8335                 /*
8336                  * avoid allocating from un-mirrored block group if there are
8337                  * mirrored block groups.
8338                  */
8339                 list_for_each_entry(cache, &space_info->block_groups[3], list)
8340                         set_block_group_ro(cache, 1);
8341                 list_for_each_entry(cache, &space_info->block_groups[4], list)
8342                         set_block_group_ro(cache, 1);
8343         }
8344
8345         init_global_block_rsv(info);
8346         ret = 0;
8347 error:
8348         btrfs_free_path(path);
8349         return ret;
8350 }
8351
8352 void btrfs_create_pending_block_groups(struct btrfs_trans_handle *trans,
8353                                        struct btrfs_root *root)
8354 {
8355         struct btrfs_block_group_cache *block_group, *tmp;
8356         struct btrfs_root *extent_root = root->fs_info->extent_root;
8357         struct btrfs_block_group_item item;
8358         struct btrfs_key key;
8359         int ret = 0;
8360
8361         list_for_each_entry_safe(block_group, tmp, &trans->new_bgs,
8362                                  new_bg_list) {
8363                 list_del_init(&block_group->new_bg_list);
8364
8365                 if (ret)
8366                         continue;
8367
8368                 spin_lock(&block_group->lock);
8369                 memcpy(&item, &block_group->item, sizeof(item));
8370                 memcpy(&key, &block_group->key, sizeof(key));
8371                 spin_unlock(&block_group->lock);
8372
8373                 ret = btrfs_insert_item(trans, extent_root, &key, &item,
8374                                         sizeof(item));
8375                 if (ret)
8376                         btrfs_abort_transaction(trans, extent_root, ret);
8377         }
8378 }
8379
8380 int btrfs_make_block_group(struct btrfs_trans_handle *trans,
8381                            struct btrfs_root *root, u64 bytes_used,
8382                            u64 type, u64 chunk_objectid, u64 chunk_offset,
8383                            u64 size)
8384 {
8385         int ret;
8386         struct btrfs_root *extent_root;
8387         struct btrfs_block_group_cache *cache;
8388
8389         extent_root = root->fs_info->extent_root;
8390
8391         root->fs_info->last_trans_log_full_commit = trans->transid;
8392
8393         cache = kzalloc(sizeof(*cache), GFP_NOFS);
8394         if (!cache)
8395                 return -ENOMEM;
8396         cache->free_space_ctl = kzalloc(sizeof(*cache->free_space_ctl),
8397                                         GFP_NOFS);
8398         if (!cache->free_space_ctl) {
8399                 kfree(cache);
8400                 return -ENOMEM;
8401         }
8402
8403         cache->key.objectid = chunk_offset;
8404         cache->key.offset = size;
8405         cache->key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
8406         cache->sectorsize = root->sectorsize;
8407         cache->fs_info = root->fs_info;
8408         cache->full_stripe_len = btrfs_full_stripe_len(root,
8409                                                &root->fs_info->mapping_tree,
8410                                                chunk_offset);
8411
8412         atomic_set(&cache->count, 1);
8413         spin_lock_init(&cache->lock);
8414         INIT_LIST_HEAD(&cache->list);
8415         INIT_LIST_HEAD(&cache->cluster_list);
8416         INIT_LIST_HEAD(&cache->new_bg_list);
8417
8418         btrfs_init_free_space_ctl(cache);
8419
8420         btrfs_set_block_group_used(&cache->item, bytes_used);
8421         btrfs_set_block_group_chunk_objectid(&cache->item, chunk_objectid);
8422         cache->flags = type;
8423         btrfs_set_block_group_flags(&cache->item, type);
8424
8425         cache->last_byte_to_unpin = (u64)-1;
8426         cache->cached = BTRFS_CACHE_FINISHED;
8427         ret = exclude_super_stripes(root, cache);
8428         if (ret) {
8429                 /*
8430                  * We may have excluded something, so call this just in
8431                  * case.
8432                  */
8433                 free_excluded_extents(root, cache);
8434                 kfree(cache->free_space_ctl);
8435                 kfree(cache);
8436                 return ret;
8437         }
8438
8439         add_new_free_space(cache, root->fs_info, chunk_offset,
8440                            chunk_offset + size);
8441
8442         free_excluded_extents(root, cache);
8443
8444         ret = btrfs_add_block_group_cache(root->fs_info, cache);
8445         if (ret) {
8446                 btrfs_remove_free_space_cache(cache);
8447                 btrfs_put_block_group(cache);
8448                 return ret;
8449         }
8450
8451         ret = update_space_info(root->fs_info, cache->flags, size, bytes_used,
8452                                 &cache->space_info);
8453         if (ret) {
8454                 btrfs_remove_free_space_cache(cache);
8455                 spin_lock(&root->fs_info->block_group_cache_lock);
8456                 rb_erase(&cache->cache_node,
8457                          &root->fs_info->block_group_cache_tree);
8458                 spin_unlock(&root->fs_info->block_group_cache_lock);
8459                 btrfs_put_block_group(cache);
8460                 return ret;
8461         }
8462         update_global_block_rsv(root->fs_info);
8463
8464         spin_lock(&cache->space_info->lock);
8465         cache->space_info->bytes_readonly += cache->bytes_super;
8466         spin_unlock(&cache->space_info->lock);
8467
8468         __link_block_group(cache->space_info, cache);
8469
8470         list_add_tail(&cache->new_bg_list, &trans->new_bgs);
8471
8472         set_avail_alloc_bits(extent_root->fs_info, type);
8473
8474         return 0;
8475 }
8476
8477 static void clear_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
8478 {
8479         u64 extra_flags = chunk_to_extended(flags) &
8480                                 BTRFS_EXTENDED_PROFILE_MASK;
8481
8482         write_seqlock(&fs_info->profiles_lock);
8483         if (flags & BTRFS_BLOCK_GROUP_DATA)
8484                 fs_info->avail_data_alloc_bits &= ~extra_flags;
8485         if (flags & BTRFS_BLOCK_GROUP_METADATA)
8486                 fs_info->avail_metadata_alloc_bits &= ~extra_flags;
8487         if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
8488                 fs_info->avail_system_alloc_bits &= ~extra_flags;
8489         write_sequnlock(&fs_info->profiles_lock);
8490 }
8491
8492 int btrfs_remove_block_group(struct btrfs_trans_handle *trans,
8493                              struct btrfs_root *root, u64 group_start)
8494 {
8495         struct btrfs_path *path;
8496         struct btrfs_block_group_cache *block_group;
8497         struct btrfs_free_cluster *cluster;
8498         struct btrfs_root *tree_root = root->fs_info->tree_root;
8499         struct btrfs_key key;
8500         struct inode *inode;
8501         int ret;
8502         int index;
8503         int factor;
8504
8505         root = root->fs_info->extent_root;
8506
8507         block_group = btrfs_lookup_block_group(root->fs_info, group_start);
8508         BUG_ON(!block_group);
8509         BUG_ON(!block_group->ro);
8510
8511         /*
8512          * Free the reserved super bytes from this block group before
8513          * remove it.
8514          */
8515         free_excluded_extents(root, block_group);
8516
8517         memcpy(&key, &block_group->key, sizeof(key));
8518         index = get_block_group_index(block_group);
8519         if (block_group->flags & (BTRFS_BLOCK_GROUP_DUP |
8520                                   BTRFS_BLOCK_GROUP_RAID1 |
8521                                   BTRFS_BLOCK_GROUP_RAID10))
8522                 factor = 2;
8523         else
8524                 factor = 1;
8525
8526         /* make sure this block group isn't part of an allocation cluster */
8527         cluster = &root->fs_info->data_alloc_cluster;
8528         spin_lock(&cluster->refill_lock);
8529         btrfs_return_cluster_to_free_space(block_group, cluster);
8530         spin_unlock(&cluster->refill_lock);
8531
8532         /*
8533          * make sure this block group isn't part of a metadata
8534          * allocation cluster
8535          */
8536         cluster = &root->fs_info->meta_alloc_cluster;
8537         spin_lock(&cluster->refill_lock);
8538         btrfs_return_cluster_to_free_space(block_group, cluster);
8539         spin_unlock(&cluster->refill_lock);
8540
8541         path = btrfs_alloc_path();
8542         if (!path) {
8543                 ret = -ENOMEM;
8544                 goto out;
8545         }
8546
8547         inode = lookup_free_space_inode(tree_root, block_group, path);
8548         if (!IS_ERR(inode)) {
8549                 ret = btrfs_orphan_add(trans, inode);
8550                 if (ret) {
8551                         btrfs_add_delayed_iput(inode);
8552                         goto out;
8553                 }
8554                 clear_nlink(inode);
8555                 /* One for the block groups ref */
8556                 spin_lock(&block_group->lock);
8557                 if (block_group->iref) {
8558                         block_group->iref = 0;
8559                         block_group->inode = NULL;
8560                         spin_unlock(&block_group->lock);
8561                         iput(inode);
8562                 } else {
8563                         spin_unlock(&block_group->lock);
8564                 }
8565                 /* One for our lookup ref */
8566                 btrfs_add_delayed_iput(inode);
8567         }
8568
8569         key.objectid = BTRFS_FREE_SPACE_OBJECTID;
8570         key.offset = block_group->key.objectid;
8571         key.type = 0;
8572
8573         ret = btrfs_search_slot(trans, tree_root, &key, path, -1, 1);
8574         if (ret < 0)
8575                 goto out;
8576         if (ret > 0)
8577                 btrfs_release_path(path);
8578         if (ret == 0) {
8579                 ret = btrfs_del_item(trans, tree_root, path);
8580                 if (ret)
8581                         goto out;
8582                 btrfs_release_path(path);
8583         }
8584
8585         spin_lock(&root->fs_info->block_group_cache_lock);
8586         rb_erase(&block_group->cache_node,
8587                  &root->fs_info->block_group_cache_tree);
8588
8589         if (root->fs_info->first_logical_byte == block_group->key.objectid)
8590                 root->fs_info->first_logical_byte = (u64)-1;
8591         spin_unlock(&root->fs_info->block_group_cache_lock);
8592
8593         down_write(&block_group->space_info->groups_sem);
8594         /*
8595          * we must use list_del_init so people can check to see if they
8596          * are still on the list after taking the semaphore
8597          */
8598         list_del_init(&block_group->list);
8599         if (list_empty(&block_group->space_info->block_groups[index]))
8600                 clear_avail_alloc_bits(root->fs_info, block_group->flags);
8601         up_write(&block_group->space_info->groups_sem);
8602
8603         if (block_group->cached == BTRFS_CACHE_STARTED)
8604                 wait_block_group_cache_done(block_group);
8605
8606         btrfs_remove_free_space_cache(block_group);
8607
8608         spin_lock(&block_group->space_info->lock);
8609         block_group->space_info->total_bytes -= block_group->key.offset;
8610         block_group->space_info->bytes_readonly -= block_group->key.offset;
8611         block_group->space_info->disk_total -= block_group->key.offset * factor;
8612         spin_unlock(&block_group->space_info->lock);
8613
8614         memcpy(&key, &block_group->key, sizeof(key));
8615
8616         btrfs_clear_space_info_full(root->fs_info);
8617
8618         btrfs_put_block_group(block_group);
8619         btrfs_put_block_group(block_group);
8620
8621         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
8622         if (ret > 0)
8623                 ret = -EIO;
8624         if (ret < 0)
8625                 goto out;
8626
8627         ret = btrfs_del_item(trans, root, path);
8628 out:
8629         btrfs_free_path(path);
8630         return ret;
8631 }
8632
8633 int btrfs_init_space_info(struct btrfs_fs_info *fs_info)
8634 {
8635         struct btrfs_space_info *space_info;
8636         struct btrfs_super_block *disk_super;
8637         u64 features;
8638         u64 flags;
8639         int mixed = 0;
8640         int ret;
8641
8642         disk_super = fs_info->super_copy;
8643         if (!btrfs_super_root(disk_super))
8644                 return 1;
8645
8646         features = btrfs_super_incompat_flags(disk_super);
8647         if (features & BTRFS_FEATURE_INCOMPAT_MIXED_GROUPS)
8648                 mixed = 1;
8649
8650         flags = BTRFS_BLOCK_GROUP_SYSTEM;
8651         ret = update_space_info(fs_info, flags, 0, 0, &space_info);
8652         if (ret)
8653                 goto out;
8654
8655         if (mixed) {
8656                 flags = BTRFS_BLOCK_GROUP_METADATA | BTRFS_BLOCK_GROUP_DATA;
8657                 ret = update_space_info(fs_info, flags, 0, 0, &space_info);
8658         } else {
8659                 flags = BTRFS_BLOCK_GROUP_METADATA;
8660                 ret = update_space_info(fs_info, flags, 0, 0, &space_info);
8661                 if (ret)
8662                         goto out;
8663
8664                 flags = BTRFS_BLOCK_GROUP_DATA;
8665                 ret = update_space_info(fs_info, flags, 0, 0, &space_info);
8666         }
8667 out:
8668         return ret;
8669 }
8670
8671 int btrfs_error_unpin_extent_range(struct btrfs_root *root, u64 start, u64 end)
8672 {
8673         return unpin_extent_range(root, start, end);
8674 }
8675
8676 int btrfs_error_discard_extent(struct btrfs_root *root, u64 bytenr,
8677                                u64 num_bytes, u64 *actual_bytes)
8678 {
8679         return btrfs_discard_extent(root, bytenr, num_bytes, actual_bytes);
8680 }
8681
8682 int btrfs_trim_fs(struct btrfs_root *root, struct fstrim_range *range)
8683 {
8684         struct btrfs_fs_info *fs_info = root->fs_info;
8685         struct btrfs_block_group_cache *cache = NULL;
8686         u64 group_trimmed;
8687         u64 start;
8688         u64 end;
8689         u64 trimmed = 0;
8690         u64 total_bytes = btrfs_super_total_bytes(fs_info->super_copy);
8691         int ret = 0;
8692
8693         /*
8694          * try to trim all FS space, our block group may start from non-zero.
8695          */
8696         if (range->len == total_bytes)
8697                 cache = btrfs_lookup_first_block_group(fs_info, range->start);
8698         else
8699                 cache = btrfs_lookup_block_group(fs_info, range->start);
8700
8701         while (cache) {
8702                 if (cache->key.objectid >= (range->start + range->len)) {
8703                         btrfs_put_block_group(cache);
8704                         break;
8705                 }
8706
8707                 start = max(range->start, cache->key.objectid);
8708                 end = min(range->start + range->len,
8709                                 cache->key.objectid + cache->key.offset);
8710
8711                 if (end - start >= range->minlen) {
8712                         if (!block_group_cache_done(cache)) {
8713                                 ret = cache_block_group(cache, 0);
8714                                 if (ret) {
8715                                         btrfs_put_block_group(cache);
8716                                         break;
8717                                 }
8718                                 ret = wait_block_group_cache_done(cache);
8719                                 if (ret) {
8720                                         btrfs_put_block_group(cache);
8721                                         break;
8722                                 }
8723                         }
8724                         ret = btrfs_trim_block_group(cache,
8725                                                      &group_trimmed,
8726                                                      start,
8727                                                      end,
8728                                                      range->minlen);
8729
8730                         trimmed += group_trimmed;
8731                         if (ret) {
8732                                 btrfs_put_block_group(cache);
8733                                 break;
8734                         }
8735                 }
8736
8737                 cache = next_block_group(fs_info->tree_root, cache);
8738         }
8739
8740         range->len = trimmed;
8741         return ret;
8742 }