fs/btrfs/extent-tree.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Copyright (C) 2007 Oracle.  All rights reserved.
   4  */
   5
   6 #include <linux/sched.h>
   7 #include <linux/sched/signal.h>
   8 #include <linux/pagemap.h>
   9 #include <linux/writeback.h>
  10 #include <linux/blkdev.h>
  11 #include <linux/sort.h>
  12 #include <linux/rcupdate.h>
  13 #include <linux/kthread.h>
  14 #include <linux/slab.h>
  15 #include <linux/ratelimit.h>
  16 #include <linux/percpu_counter.h>
  17 #include <linux/lockdep.h>
  18 #include <linux/crc32c.h>
  19 #include "tree-log.h"
  20 #include "disk-io.h"
  21 #include "print-tree.h"
  22 #include "volumes.h"
  23 #include "raid56.h"
  24 #include "locking.h"
  25 #include "free-space-cache.h"
  26 #include "free-space-tree.h"
  27 #include "math.h"
  28 #include "sysfs.h"
  29 #include "qgroup.h"
  30 #include "ref-verify.h"
  31
  32 #undef SCRAMBLE_DELAYED_REFS
  33
  34 /*
  35  * control flags for do_chunk_alloc's force field
  36  * CHUNK_ALLOC_NO_FORCE means to only allocate a chunk
  37  * if we really need one.
  38  *
  39  * CHUNK_ALLOC_LIMITED means to only try and allocate one
  40  * if we have very few chunks already allocated.  This is
  41  * used as part of the clustering code to help make sure
  42  * we have a good pool of storage to cluster in, without
  43  * filling the FS with empty chunks
  44  *
  45  * CHUNK_ALLOC_FORCE means it must try to allocate one
  46  *
  47  */
  48 enum {
  49         CHUNK_ALLOC_NO_FORCE = 0,
  50         CHUNK_ALLOC_LIMITED = 1,
  51         CHUNK_ALLOC_FORCE = 2,
  52 };
  53
  54 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
  55                                struct btrfs_fs_info *fs_info,
  56                                 struct btrfs_delayed_ref_node *node, u64 parent,
  57                                 u64 root_objectid, u64 owner_objectid,
  58                                 u64 owner_offset, int refs_to_drop,
  59                                 struct btrfs_delayed_extent_op *extra_op);
  60 static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op,
  61                                     struct extent_buffer *leaf,
  62                                     struct btrfs_extent_item *ei);
  63 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
  64                                       struct btrfs_fs_info *fs_info,
  65                                       u64 parent, u64 root_objectid,
  66                                       u64 flags, u64 owner, u64 offset,
  67                                       struct btrfs_key *ins, int ref_mod);
  68 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
  69                                      struct btrfs_delayed_ref_node *node,
  70                                      struct btrfs_delayed_extent_op *extent_op);
  71 static int do_chunk_alloc(struct btrfs_trans_handle *trans,
  72                           struct btrfs_fs_info *fs_info, u64 flags,
  73                           int force);
  74 static int find_next_key(struct btrfs_path *path, int level,
  75                          struct btrfs_key *key);
  76 static void dump_space_info(struct btrfs_fs_info *fs_info,
  77                             struct btrfs_space_info *info, u64 bytes,
  78                             int dump_block_groups);
  79 static int block_rsv_use_bytes(struct btrfs_block_rsv *block_rsv,
  80                                u64 num_bytes);
  81 static void space_info_add_new_bytes(struct btrfs_fs_info *fs_info,
  82                                      struct btrfs_space_info *space_info,
  83                                      u64 num_bytes);
  84 static void space_info_add_old_bytes(struct btrfs_fs_info *fs_info,
  85                                      struct btrfs_space_info *space_info,
  86                                      u64 num_bytes);
  87
  88 static noinline int
  89 block_group_cache_done(struct btrfs_block_group_cache *cache)
  90 {
  91         smp_mb();
  92         return cache->cached == BTRFS_CACHE_FINISHED ||
  93                 cache->cached == BTRFS_CACHE_ERROR;
  94 }
  95
  96 static int block_group_bits(struct btrfs_block_group_cache *cache, u64 bits)
  97 {
  98         return (cache->flags & bits) == bits;
  99 }
 100
 101 void btrfs_get_block_group(struct btrfs_block_group_cache *cache)
 102 {
 103         atomic_inc(&cache->count);
 104 }
 105
 106 void btrfs_put_block_group(struct btrfs_block_group_cache *cache)
 107 {
 108         if (atomic_dec_and_test(&cache->count)) {
 109                 WARN_ON(cache->pinned > 0);
 110                 WARN_ON(cache->reserved > 0);
 111
 112                 /*
 113                  * If not empty, someone is still holding mutex of
 114                  * full_stripe_lock, which can only be released by caller.
 115                  * And it will definitely cause use-after-free when caller
 116                  * tries to release full stripe lock.
 117                  *
 118                  * No better way to resolve, but only to warn.
 119                  */
 120                 WARN_ON(!RB_EMPTY_ROOT(&cache->full_stripe_locks_root.root));
 121                 kfree(cache->free_space_ctl);
 122                 kfree(cache);
 123         }
 124 }
 125
 126 /*
 127  * this adds the block group to the fs_info rb tree for the block group
 128  * cache
 129  */
 130 static int btrfs_add_block_group_cache(struct btrfs_fs_info *info,
 131                                 struct btrfs_block_group_cache *block_group)
 132 {
 133         struct rb_node **p;
 134         struct rb_node *parent = NULL;
 135         struct btrfs_block_group_cache *cache;
 136
 137         spin_lock(&info->block_group_cache_lock);
 138         p = &info->block_group_cache_tree.rb_node;
 139
 140         while (*p) {
 141                 parent = *p;
 142                 cache = rb_entry(parent, struct btrfs_block_group_cache,
 143                                  cache_node);
 144                 if (block_group->key.objectid < cache->key.objectid) {
 145                         p = &(*p)->rb_left;
 146                 } else if (block_group->key.objectid > cache->key.objectid) {
 147                         p = &(*p)->rb_right;
 148                 } else {
 149                         spin_unlock(&info->block_group_cache_lock);
 150                         return -EEXIST;
 151                 }
 152         }
 153
 154         rb_link_node(&block_group->cache_node, parent, p);
 155         rb_insert_color(&block_group->cache_node,
 156                         &info->block_group_cache_tree);
 157
 158         if (info->first_logical_byte > block_group->key.objectid)
 159                 info->first_logical_byte = block_group->key.objectid;
 160
 161         spin_unlock(&info->block_group_cache_lock);
 162
 163         return 0;
 164 }
 165
 166 /*
 167  * This will return the block group at or after bytenr if contains is 0, else
 168  * it will return the block group that contains the bytenr
 169  */
 170 static struct btrfs_block_group_cache *
 171 block_group_cache_tree_search(struct btrfs_fs_info *info, u64 bytenr,
 172                               int contains)
 173 {
 174         struct btrfs_block_group_cache *cache, *ret = NULL;
 175         struct rb_node *n;
 176         u64 end, start;
 177
 178         spin_lock(&info->block_group_cache_lock);
 179         n = info->block_group_cache_tree.rb_node;
 180
 181         while (n) {
 182                 cache = rb_entry(n, struct btrfs_block_group_cache,
 183                                  cache_node);
 184                 end = cache->key.objectid + cache->key.offset - 1;
 185                 start = cache->key.objectid;
 186
 187                 if (bytenr < start) {
 188                         if (!contains && (!ret || start < ret->key.objectid))
 189                                 ret = cache;
 190                         n = n->rb_left;
 191                 } else if (bytenr > start) {
 192                         if (contains && bytenr <= end) {
 193                                 ret = cache;
 194                                 break;
 195                         }
 196                         n = n->rb_right;
 197                 } else {
 198                         ret = cache;
 199                         break;
 200                 }
 201         }
 202         if (ret) {
 203                 btrfs_get_block_group(ret);
 204                 if (bytenr == 0 && info->first_logical_byte > ret->key.objectid)
 205                         info->first_logical_byte = ret->key.objectid;
 206         }
 207         spin_unlock(&info->block_group_cache_lock);
 208
 209         return ret;
 210 }
 211
 212 static int add_excluded_extent(struct btrfs_fs_info *fs_info,
 213                                u64 start, u64 num_bytes)
 214 {
 215         u64 end = start + num_bytes - 1;
 216         set_extent_bits(&fs_info->freed_extents[0],
 217                         start, end, EXTENT_UPTODATE);
 218         set_extent_bits(&fs_info->freed_extents[1],
 219                         start, end, EXTENT_UPTODATE);
 220         return 0;
 221 }
 222
 223 static void free_excluded_extents(struct btrfs_fs_info *fs_info,
 224                                   struct btrfs_block_group_cache *cache)
 225 {
 226         u64 start, end;
 227
 228         start = cache->key.objectid;
 229         end = start + cache->key.offset - 1;
 230
 231         clear_extent_bits(&fs_info->freed_extents[0],
 232                           start, end, EXTENT_UPTODATE);
 233         clear_extent_bits(&fs_info->freed_extents[1],
 234                           start, end, EXTENT_UPTODATE);
 235 }
 236
 237 static int exclude_super_stripes(struct btrfs_fs_info *fs_info,
 238                                  struct btrfs_block_group_cache *cache)
 239 {
 240         u64 bytenr;
 241         u64 *logical;
 242         int stripe_len;
 243         int i, nr, ret;
 244
 245         if (cache->key.objectid < BTRFS_SUPER_INFO_OFFSET) {
 246                 stripe_len = BTRFS_SUPER_INFO_OFFSET - cache->key.objectid;
 247                 cache->bytes_super += stripe_len;
 248                 ret = add_excluded_extent(fs_info, cache->key.objectid,
 249                                           stripe_len);
 250                 if (ret)
 251                         return ret;
 252         }
 253
 254         for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
 255                 bytenr = btrfs_sb_offset(i);
 256                 ret = btrfs_rmap_block(fs_info, cache->key.objectid,
 257                                        bytenr, &logical, &nr, &stripe_len);
 258                 if (ret)
 259                         return ret;
 260
 261                 while (nr--) {
 262                         u64 start, len;
 263
 264                         if (logical[nr] > cache->key.objectid +
 265                             cache->key.offset)
 266                                 continue;
 267
 268                         if (logical[nr] + stripe_len <= cache->key.objectid)
 269                                 continue;
 270
 271                         start = logical[nr];
 272                         if (start < cache->key.objectid) {
 273                                 start = cache->key.objectid;
 274                                 len = (logical[nr] + stripe_len) - start;
 275                         } else {
 276                                 len = min_t(u64, stripe_len,
 277                                             cache->key.objectid +
 278                                             cache->key.offset - start);
 279                         }
 280
 281                         cache->bytes_super += len;
 282                         ret = add_excluded_extent(fs_info, start, len);
 283                         if (ret) {
 284                                 kfree(logical);
 285                                 return ret;
 286                         }
 287                 }
 288
 289                 kfree(logical);
 290         }
 291         return 0;
 292 }
 293
 294 static struct btrfs_caching_control *
 295 get_caching_control(struct btrfs_block_group_cache *cache)
 296 {
 297         struct btrfs_caching_control *ctl;
 298
 299         spin_lock(&cache->lock);
 300         if (!cache->caching_ctl) {
 301                 spin_unlock(&cache->lock);
 302                 return NULL;
 303         }
 304
 305         ctl = cache->caching_ctl;
 306         refcount_inc(&ctl->count);
 307         spin_unlock(&cache->lock);
 308         return ctl;
 309 }
 310
 311 static void put_caching_control(struct btrfs_caching_control *ctl)
 312 {
 313         if (refcount_dec_and_test(&ctl->count))
 314                 kfree(ctl);
 315 }
 316
 317 #ifdef CONFIG_BTRFS_DEBUG
 318 static void fragment_free_space(struct btrfs_block_group_cache *block_group)
 319 {
 320         struct btrfs_fs_info *fs_info = block_group->fs_info;
 321         u64 start = block_group->key.objectid;
 322         u64 len = block_group->key.offset;
 323         u64 chunk = block_group->flags & BTRFS_BLOCK_GROUP_METADATA ?
 324                 fs_info->nodesize : fs_info->sectorsize;
 325         u64 step = chunk << 1;
 326
 327         while (len > chunk) {
 328                 btrfs_remove_free_space(block_group, start, chunk);
 329                 start += step;
 330                 if (len < step)
 331                         len = 0;
 332                 else
 333                         len -= step;
 334         }
 335 }
 336 #endif
 337
 338 /*
 339  * this is only called by cache_block_group, since we could have freed extents
 340  * we need to check the pinned_extents for any extents that can't be used yet
 341  * since their free space will be released as soon as the transaction commits.
 342  */
 343 u64 add_new_free_space(struct btrfs_block_group_cache *block_group,
 344                        u64 start, u64 end)
 345 {
 346         struct btrfs_fs_info *info = block_group->fs_info;
 347         u64 extent_start, extent_end, size, total_added = 0;
 348         int ret;
 349
 350         while (start < end) {
 351                 ret = find_first_extent_bit(info->pinned_extents, start,
 352                                             &extent_start, &extent_end,
 353                                             EXTENT_DIRTY | EXTENT_UPTODATE,
 354                                             NULL);
 355                 if (ret)
 356                         break;
 357
 358                 if (extent_start <= start) {
 359                         start = extent_end + 1;
 360                 } else if (extent_start > start && extent_start < end) {
 361                         size = extent_start - start;
 362                         total_added += size;
 363                         ret = btrfs_add_free_space(block_group, start,
 364                                                    size);
 365                         BUG_ON(ret); /* -ENOMEM or logic error */
 366                         start = extent_end + 1;
 367                 } else {
 368                         break;
 369                 }
 370         }
 371
 372         if (start < end) {
 373                 size = end - start;
 374                 total_added += size;
 375                 ret = btrfs_add_free_space(block_group, start, size);
 376                 BUG_ON(ret); /* -ENOMEM or logic error */
 377         }
 378
 379         return total_added;
 380 }
 381
 382 static int load_extent_tree_free(struct btrfs_caching_control *caching_ctl)
 383 {
 384         struct btrfs_block_group_cache *block_group = caching_ctl->block_group;
 385         struct btrfs_fs_info *fs_info = block_group->fs_info;
 386         struct btrfs_root *extent_root = fs_info->extent_root;
 387         struct btrfs_path *path;
 388         struct extent_buffer *leaf;
 389         struct btrfs_key key;
 390         u64 total_found = 0;
 391         u64 last = 0;
 392         u32 nritems;
 393         int ret;
 394         bool wakeup = true;
 395
 396         path = btrfs_alloc_path();
 397         if (!path)
 398                 return -ENOMEM;
 399
 400         last = max_t(u64, block_group->key.objectid, BTRFS_SUPER_INFO_OFFSET);
 401
 402 #ifdef CONFIG_BTRFS_DEBUG
 403         /*
 404          * If we're fragmenting we don't want to make anybody think we can
 405          * allocate from this block group until we've had a chance to fragment
 406          * the free space.
 407          */
 408         if (btrfs_should_fragment_free_space(block_group))
 409                 wakeup = false;
 410 #endif
 411         /*
 412          * We don't want to deadlock with somebody trying to allocate a new
 413          * extent for the extent root while also trying to search the extent
 414          * root to add free space.  So we skip locking and search the commit
 415          * root, since its read-only
 416          */
 417         path->skip_locking = 1;
 418         path->search_commit_root = 1;
 419         path->reada = READA_FORWARD;
 420
 421         key.objectid = last;
 422         key.offset = 0;
 423         key.type = BTRFS_EXTENT_ITEM_KEY;
 424
 425 next:
 426         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
 427         if (ret < 0)
 428                 goto out;
 429
 430         leaf = path->nodes[0];
 431         nritems = btrfs_header_nritems(leaf);
 432
 433         while (1) {
 434                 if (btrfs_fs_closing(fs_info) > 1) {
 435                         last = (u64)-1;
 436                         break;
 437                 }
 438
 439                 if (path->slots[0] < nritems) {
 440                         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
 441                 } else {
 442                         ret = find_next_key(path, 0, &key);
 443                         if (ret)
 444                                 break;
 445
 446                         if (need_resched() ||
 447                             rwsem_is_contended(&fs_info->commit_root_sem)) {
 448                                 if (wakeup)
 449                                         caching_ctl->progress = last;
 450                                 btrfs_release_path(path);
 451                                 up_read(&fs_info->commit_root_sem);
 452                                 mutex_unlock(&caching_ctl->mutex);
 453                                 cond_resched();
 454                                 mutex_lock(&caching_ctl->mutex);
 455                                 down_read(&fs_info->commit_root_sem);
 456                                 goto next;
 457                         }
 458
 459                         ret = btrfs_next_leaf(extent_root, path);
 460                         if (ret < 0)
 461                                 goto out;
 462                         if (ret)
 463                                 break;
 464                         leaf = path->nodes[0];
 465                         nritems = btrfs_header_nritems(leaf);
 466                         continue;
 467                 }
 468
 469                 if (key.objectid < last) {
 470                         key.objectid = last;
 471                         key.offset = 0;
 472                         key.type = BTRFS_EXTENT_ITEM_KEY;
 473
 474                         if (wakeup)
 475                                 caching_ctl->progress = last;
 476                         btrfs_release_path(path);
 477                         goto next;
 478                 }
 479
 480                 if (key.objectid < block_group->key.objectid) {
 481                         path->slots[0]++;
 482                         continue;
 483                 }
 484
 485                 if (key.objectid >= block_group->key.objectid +
 486                     block_group->key.offset)
 487                         break;
 488
 489                 if (key.type == BTRFS_EXTENT_ITEM_KEY ||
 490                     key.type == BTRFS_METADATA_ITEM_KEY) {
 491                         total_found += add_new_free_space(block_group, last,
 492                                                           key.objectid);
 493                         if (key.type == BTRFS_METADATA_ITEM_KEY)
 494                                 last = key.objectid +
 495                                         fs_info->nodesize;
 496                         else
 497                                 last = key.objectid + key.offset;
 498
 499                         if (total_found > CACHING_CTL_WAKE_UP) {
 500                                 total_found = 0;
 501                                 if (wakeup)
 502                                         wake_up(&caching_ctl->wait);
 503                         }
 504                 }
 505                 path->slots[0]++;
 506         }
 507         ret = 0;
 508
 509         total_found += add_new_free_space(block_group, last,
 510                                           block_group->key.objectid +
 511                                           block_group->key.offset);
 512         caching_ctl->progress = (u64)-1;
 513
 514 out:
 515         btrfs_free_path(path);
 516         return ret;
 517 }
 518
 519 static noinline void caching_thread(struct btrfs_work *work)
 520 {
 521         struct btrfs_block_group_cache *block_group;
 522         struct btrfs_fs_info *fs_info;
 523         struct btrfs_caching_control *caching_ctl;
 524         int ret;
 525
 526         caching_ctl = container_of(work, struct btrfs_caching_control, work);
 527         block_group = caching_ctl->block_group;
 528         fs_info = block_group->fs_info;
 529
 530         mutex_lock(&caching_ctl->mutex);
 531         down_read(&fs_info->commit_root_sem);
 532
 533         if (btrfs_fs_compat_ro(fs_info, FREE_SPACE_TREE))
 534                 ret = load_free_space_tree(caching_ctl);
 535         else
 536                 ret = load_extent_tree_free(caching_ctl);
 537
 538         spin_lock(&block_group->lock);
 539         block_group->caching_ctl = NULL;
 540         block_group->cached = ret ? BTRFS_CACHE_ERROR : BTRFS_CACHE_FINISHED;
 541         spin_unlock(&block_group->lock);
 542
 543 #ifdef CONFIG_BTRFS_DEBUG
 544         if (btrfs_should_fragment_free_space(block_group)) {
 545                 u64 bytes_used;
 546
 547                 spin_lock(&block_group->space_info->lock);
 548                 spin_lock(&block_group->lock);
 549                 bytes_used = block_group->key.offset -
 550                         btrfs_block_group_used(&block_group->item);
 551                 block_group->space_info->bytes_used += bytes_used >> 1;
 552                 spin_unlock(&block_group->lock);
 553                 spin_unlock(&block_group->space_info->lock);
 554                 fragment_free_space(block_group);
 555         }
 556 #endif
 557
 558         caching_ctl->progress = (u64)-1;
 559
 560         up_read(&fs_info->commit_root_sem);
 561         free_excluded_extents(fs_info, block_group);
 562         mutex_unlock(&caching_ctl->mutex);
 563
 564         wake_up(&caching_ctl->wait);
 565
 566         put_caching_control(caching_ctl);
 567         btrfs_put_block_group(block_group);
 568 }
 569
 570 static int cache_block_group(struct btrfs_block_group_cache *cache,
 571                              int load_cache_only)
 572 {
 573         DEFINE_WAIT(wait);
 574         struct btrfs_fs_info *fs_info = cache->fs_info;
 575         struct btrfs_caching_control *caching_ctl;
 576         int ret = 0;
 577
 578         caching_ctl = kzalloc(sizeof(*caching_ctl), GFP_NOFS);
 579         if (!caching_ctl)
 580                 return -ENOMEM;
 581
 582         INIT_LIST_HEAD(&caching_ctl->list);
 583         mutex_init(&caching_ctl->mutex);
 584         init_waitqueue_head(&caching_ctl->wait);
 585         caching_ctl->block_group = cache;
 586         caching_ctl->progress = cache->key.objectid;
 587         refcount_set(&caching_ctl->count, 1);
 588         btrfs_init_work(&caching_ctl->work, btrfs_cache_helper,
 589                         caching_thread, NULL, NULL);
 590
 591         spin_lock(&cache->lock);
 592         /*
 593          * This should be a rare occasion, but this could happen I think in the
 594          * case where one thread starts to load the space cache info, and then
 595          * some other thread starts a transaction commit which tries to do an
 596          * allocation while the other thread is still loading the space cache
 597          * info.  The previous loop should have kept us from choosing this block
 598          * group, but if we've moved to the state where we will wait on caching
 599          * block groups we need to first check if we're doing a fast load here,
 600          * so we can wait for it to finish, otherwise we could end up allocating
 601          * from a block group who's cache gets evicted for one reason or
 602          * another.
 603          */
 604         while (cache->cached == BTRFS_CACHE_FAST) {
 605                 struct btrfs_caching_control *ctl;
 606
 607                 ctl = cache->caching_ctl;
 608                 refcount_inc(&ctl->count);
 609                 prepare_to_wait(&ctl->wait, &wait, TASK_UNINTERRUPTIBLE);
 610                 spin_unlock(&cache->lock);
 611
 612                 schedule();
 613
 614                 finish_wait(&ctl->wait, &wait);
 615                 put_caching_control(ctl);
 616                 spin_lock(&cache->lock);
 617         }
 618
 619         if (cache->cached != BTRFS_CACHE_NO) {
 620                 spin_unlock(&cache->lock);
 621                 kfree(caching_ctl);
 622                 return 0;
 623         }
 624         WARN_ON(cache->caching_ctl);
 625         cache->caching_ctl = caching_ctl;
 626         cache->cached = BTRFS_CACHE_FAST;
 627         spin_unlock(&cache->lock);
 628
 629         if (btrfs_test_opt(fs_info, SPACE_CACHE)) {
 630                 mutex_lock(&caching_ctl->mutex);
 631                 ret = load_free_space_cache(fs_info, cache);
 632
 633                 spin_lock(&cache->lock);
 634                 if (ret == 1) {
 635                         cache->caching_ctl = NULL;
 636                         cache->cached = BTRFS_CACHE_FINISHED;
 637                         cache->last_byte_to_unpin = (u64)-1;
 638                         caching_ctl->progress = (u64)-1;
 639                 } else {
 640                         if (load_cache_only) {
 641                                 cache->caching_ctl = NULL;
 642                                 cache->cached = BTRFS_CACHE_NO;
 643                         } else {
 644                                 cache->cached = BTRFS_CACHE_STARTED;
 645                                 cache->has_caching_ctl = 1;
 646                         }
 647                 }
 648                 spin_unlock(&cache->lock);
 649 #ifdef CONFIG_BTRFS_DEBUG
 650                 if (ret == 1 &&
 651                     btrfs_should_fragment_free_space(cache)) {
 652                         u64 bytes_used;
 653
 654                         spin_lock(&cache->space_info->lock);
 655                         spin_lock(&cache->lock);
 656                         bytes_used = cache->key.offset -
 657                                 btrfs_block_group_used(&cache->item);
 658                         cache->space_info->bytes_used += bytes_used >> 1;
 659                         spin_unlock(&cache->lock);
 660                         spin_unlock(&cache->space_info->lock);
 661                         fragment_free_space(cache);
 662                 }
 663 #endif
 664                 mutex_unlock(&caching_ctl->mutex);
 665
 666                 wake_up(&caching_ctl->wait);
 667                 if (ret == 1) {
 668                         put_caching_control(caching_ctl);
 669                         free_excluded_extents(fs_info, cache);
 670                         return 0;
 671                 }
 672         } else {
 673                 /*
 674                  * We're either using the free space tree or no caching at all.
 675                  * Set cached to the appropriate value and wakeup any waiters.
 676                  */
 677                 spin_lock(&cache->lock);
 678                 if (load_cache_only) {
 679                         cache->caching_ctl = NULL;
 680                         cache->cached = BTRFS_CACHE_NO;
 681                 } else {
 682                         cache->cached = BTRFS_CACHE_STARTED;
 683                         cache->has_caching_ctl = 1;
 684                 }
 685                 spin_unlock(&cache->lock);
 686                 wake_up(&caching_ctl->wait);
 687         }
 688
 689         if (load_cache_only) {
 690                 put_caching_control(caching_ctl);
 691                 return 0;
 692         }
 693
 694         down_write(&fs_info->commit_root_sem);
 695         refcount_inc(&caching_ctl->count);
 696         list_add_tail(&caching_ctl->list, &fs_info->caching_block_groups);
 697         up_write(&fs_info->commit_root_sem);
 698
 699         btrfs_get_block_group(cache);
 700
 701         btrfs_queue_work(fs_info->caching_workers, &caching_ctl->work);
 702
 703         return ret;
 704 }
 705
 706 /*
 707  * return the block group that starts at or after bytenr
 708  */
 709 static struct btrfs_block_group_cache *
 710 btrfs_lookup_first_block_group(struct btrfs_fs_info *info, u64 bytenr)
 711 {
 712         return block_group_cache_tree_search(info, bytenr, 0);
 713 }
 714
 715 /*
 716  * return the block group that contains the given bytenr
 717  */
 718 struct btrfs_block_group_cache *btrfs_lookup_block_group(
 719                                                  struct btrfs_fs_info *info,
 720                                                  u64 bytenr)
 721 {
 722         return block_group_cache_tree_search(info, bytenr, 1);
 723 }
 724
 725 static struct btrfs_space_info *__find_space_info(struct btrfs_fs_info *info,
 726                                                   u64 flags)
 727 {
 728         struct list_head *head = &info->space_info;
 729         struct btrfs_space_info *found;
 730
 731         flags &= BTRFS_BLOCK_GROUP_TYPE_MASK;
 732
 733         rcu_read_lock();
 734         list_for_each_entry_rcu(found, head, list) {
 735                 if (found->flags & flags) {
 736                         rcu_read_unlock();
 737                         return found;
 738                 }
 739         }
 740         rcu_read_unlock();
 741         return NULL;
 742 }
 743
 744 static void add_pinned_bytes(struct btrfs_fs_info *fs_info, s64 num_bytes,
 745                              bool metadata, u64 root_objectid)
 746 {
 747         struct btrfs_space_info *space_info;
 748         u64 flags;
 749
 750         if (metadata) {
 751                 if (root_objectid == BTRFS_CHUNK_TREE_OBJECTID)
 752                         flags = BTRFS_BLOCK_GROUP_SYSTEM;
 753                 else
 754                         flags = BTRFS_BLOCK_GROUP_METADATA;
 755         } else {
 756                 flags = BTRFS_BLOCK_GROUP_DATA;
 757         }
 758
 759         space_info = __find_space_info(fs_info, flags);
 760         ASSERT(space_info);
 761         percpu_counter_add(&space_info->total_bytes_pinned, num_bytes);
 762 }
 763
 764 /*
 765  * after adding space to the filesystem, we need to clear the full flags
 766  * on all the space infos.
 767  */
 768 void btrfs_clear_space_info_full(struct btrfs_fs_info *info)
 769 {
 770         struct list_head *head = &info->space_info;
 771         struct btrfs_space_info *found;
 772
 773         rcu_read_lock();
 774         list_for_each_entry_rcu(found, head, list)
 775                 found->full = 0;
 776         rcu_read_unlock();
 777 }
 778
 779 /* simple helper to search for an existing data extent at a given offset */
 780 int btrfs_lookup_data_extent(struct btrfs_fs_info *fs_info, u64 start, u64 len)
 781 {
 782         int ret;
 783         struct btrfs_key key;
 784         struct btrfs_path *path;
 785
 786         path = btrfs_alloc_path();
 787         if (!path)
 788                 return -ENOMEM;
 789
 790         key.objectid = start;
 791         key.offset = len;
 792         key.type = BTRFS_EXTENT_ITEM_KEY;
 793         ret = btrfs_search_slot(NULL, fs_info->extent_root, &key, path, 0, 0);
 794         btrfs_free_path(path);
 795         return ret;
 796 }
 797
 798 /*
 799  * helper function to lookup reference count and flags of a tree block.
 800  *
 801  * the head node for delayed ref is used to store the sum of all the
 802  * reference count modifications queued up in the rbtree. the head
 803  * node may also store the extent flags to set. This way you can check
 804  * to see what the reference count and extent flags would be if all of
 805  * the delayed refs are not processed.
 806  */
 807 int btrfs_lookup_extent_info(struct btrfs_trans_handle *trans,
 808                              struct btrfs_fs_info *fs_info, u64 bytenr,
 809                              u64 offset, int metadata, u64 *refs, u64 *flags)
 810 {
 811         struct btrfs_delayed_ref_head *head;
 812         struct btrfs_delayed_ref_root *delayed_refs;
 813         struct btrfs_path *path;
 814         struct btrfs_extent_item *ei;
 815         struct extent_buffer *leaf;
 816         struct btrfs_key key;
 817         u32 item_size;
 818         u64 num_refs;
 819         u64 extent_flags;
 820         int ret;
 821
 822         /*
 823          * If we don't have skinny metadata, don't bother doing anything
 824          * different
 825          */
 826         if (metadata && !btrfs_fs_incompat(fs_info, SKINNY_METADATA)) {
 827                 offset = fs_info->nodesize;
 828                 metadata = 0;
 829         }
 830
 831         path = btrfs_alloc_path();
 832         if (!path)
 833                 return -ENOMEM;
 834
 835         if (!trans) {
 836                 path->skip_locking = 1;
 837                 path->search_commit_root = 1;
 838         }
 839
 840 search_again:
 841         key.objectid = bytenr;
 842         key.offset = offset;
 843         if (metadata)
 844                 key.type = BTRFS_METADATA_ITEM_KEY;
 845         else
 846                 key.type = BTRFS_EXTENT_ITEM_KEY;
 847
 848         ret = btrfs_search_slot(trans, fs_info->extent_root, &key, path, 0, 0);
 849         if (ret < 0)
 850                 goto out_free;
 851
 852         if (ret > 0 && metadata && key.type == BTRFS_METADATA_ITEM_KEY) {
 853                 if (path->slots[0]) {
 854                         path->slots[0]--;
 855                         btrfs_item_key_to_cpu(path->nodes[0], &key,
 856                                               path->slots[0]);
 857                         if (key.objectid == bytenr &&
 858                             key.type == BTRFS_EXTENT_ITEM_KEY &&
 859                             key.offset == fs_info->nodesize)
 860                                 ret = 0;
 861                 }
 862         }
 863
 864         if (ret == 0) {
 865                 leaf = path->nodes[0];
 866                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
 867                 if (item_size >= sizeof(*ei)) {
 868                         ei = btrfs_item_ptr(leaf, path->slots[0],
 869                                             struct btrfs_extent_item);
 870                         num_refs = btrfs_extent_refs(leaf, ei);
 871                         extent_flags = btrfs_extent_flags(leaf, ei);
 872                 } else {
 873 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
 874                         struct btrfs_extent_item_v0 *ei0;
 875                         BUG_ON(item_size != sizeof(*ei0));
 876                         ei0 = btrfs_item_ptr(leaf, path->slots[0],
 877                                              struct btrfs_extent_item_v0);
 878                         num_refs = btrfs_extent_refs_v0(leaf, ei0);
 879                         /* FIXME: this isn't correct for data */
 880                         extent_flags = BTRFS_BLOCK_FLAG_FULL_BACKREF;
 881 #else
 882                         BUG();
 883 #endif
 884                 }
 885                 BUG_ON(num_refs == 0);
 886         } else {
 887                 num_refs = 0;
 888                 extent_flags = 0;
 889                 ret = 0;
 890         }
 891
 892         if (!trans)
 893                 goto out;
 894
 895         delayed_refs = &trans->transaction->delayed_refs;
 896         spin_lock(&delayed_refs->lock);
 897         head = btrfs_find_delayed_ref_head(delayed_refs, bytenr);
 898         if (head) {
 899                 if (!mutex_trylock(&head->mutex)) {
 900                         refcount_inc(&head->refs);
 901                         spin_unlock(&delayed_refs->lock);
 902
 903                         btrfs_release_path(path);
 904
 905                         /*
 906                          * Mutex was contended, block until it's released and try
 907                          * again
 908                          */
 909                         mutex_lock(&head->mutex);
 910                         mutex_unlock(&head->mutex);
 911                         btrfs_put_delayed_ref_head(head);
 912                         goto search_again;
 913                 }
 914                 spin_lock(&head->lock);
 915                 if (head->extent_op && head->extent_op->update_flags)
 916                         extent_flags |= head->extent_op->flags_to_set;
 917                 else
 918                         BUG_ON(num_refs == 0);
 919
 920                 num_refs += head->ref_mod;
 921                 spin_unlock(&head->lock);
 922                 mutex_unlock(&head->mutex);
 923         }
 924         spin_unlock(&delayed_refs->lock);
 925 out:
 926         WARN_ON(num_refs == 0);
 927         if (refs)
 928                 *refs = num_refs;
 929         if (flags)
 930                 *flags = extent_flags;
 931 out_free:
 932         btrfs_free_path(path);
 933         return ret;
 934 }
 935
 936 /*
 937  * Back reference rules.  Back refs have three main goals:
 938  *
 939  * 1) differentiate between all holders of references to an extent so that
 940  *    when a reference is dropped we can make sure it was a valid reference
 941  *    before freeing the extent.
 942  *
 943  * 2) Provide enough information to quickly find the holders of an extent
 944  *    if we notice a given block is corrupted or bad.
 945  *
 946  * 3) Make it easy to migrate blocks for FS shrinking or storage pool
 947  *    maintenance.  This is actually the same as #2, but with a slightly
 948  *    different use case.
 949  *
 950  * There are two kinds of back refs. The implicit back refs is optimized
 951  * for pointers in non-shared tree blocks. For a given pointer in a block,
 952  * back refs of this kind provide information about the block's owner tree
 953  * and the pointer's key. These information allow us to find the block by
 954  * b-tree searching. The full back refs is for pointers in tree blocks not
 955  * referenced by their owner trees. The location of tree block is recorded
 956  * in the back refs. Actually the full back refs is generic, and can be
 957  * used in all cases the implicit back refs is used. The major shortcoming
 958  * of the full back refs is its overhead. Every time a tree block gets
 959  * COWed, we have to update back refs entry for all pointers in it.
 960  *
 961  * For a newly allocated tree block, we use implicit back refs for
 962  * pointers in it. This means most tree related operations only involve
 963  * implicit back refs. For a tree block created in old transaction, the
 964  * only way to drop a reference to it is COW it. So we can detect the
 965  * event that tree block loses its owner tree's reference and do the
 966  * back refs conversion.
 967  *
 968  * When a tree block is COWed through a tree, there are four cases:
 969  *
 970  * The reference count of the block is one and the tree is the block's
 971  * owner tree. Nothing to do in this case.
 972  *
 973  * The reference count of the block is one and the tree is not the
 974  * block's owner tree. In this case, full back refs is used for pointers
 975  * in the block. Remove these full back refs, add implicit back refs for
 976  * every pointers in the new block.
 977  *
 978  * The reference count of the block is greater than one and the tree is
 979  * the block's owner tree. In this case, implicit back refs is used for
 980  * pointers in the block. Add full back refs for every pointers in the
 981  * block, increase lower level extents' reference counts. The original
 982  * implicit back refs are entailed to the new block.
 983  *
 984  * The reference count of the block is greater than one and the tree is
 985  * not the block's owner tree. Add implicit back refs for every pointer in
 986  * the new block, increase lower level extents' reference count.
 987  *
 988  * Back Reference Key composing:
 989  *
 990  * The key objectid corresponds to the first byte in the extent,
 991  * The key type is used to differentiate between types of back refs.
 992  * There are different meanings of the key offset for different types
 993  * of back refs.
 994  *
 995  * File extents can be referenced by:
 996  *
 997  * - multiple snapshots, subvolumes, or different generations in one subvol
 998  * - different files inside a single subvolume
 999  * - different offsets inside a file (bookend extents in file.c)
1000  *
1001  * The extent ref structure for the implicit back refs has fields for:
1002  *
1003  * - Objectid of the subvolume root
1004  * - objectid of the file holding the reference
1005  * - original offset in the file
1006  * - how many bookend extents
1007  *
1008  * The key offset for the implicit back refs is hash of the first
1009  * three fields.
1010  *
1011  * The extent ref structure for the full back refs has field for:
1012  *
1013  * - number of pointers in the tree leaf
1014  *
1015  * The key offset for the implicit back refs is the first byte of
1016  * the tree leaf
1017  *
1018  * When a file extent is allocated, The implicit back refs is used.
1019  * the fields are filled in:
1020  *
1021  *     (root_key.objectid, inode objectid, offset in file, 1)
1022  *
1023  * When a file extent is removed file truncation, we find the
1024  * corresponding implicit back refs and check the following fields:
1025  *
1026  *     (btrfs_header_owner(leaf), inode objectid, offset in file)
1027  *
1028  * Btree extents can be referenced by:
1029  *
1030  * - Different subvolumes
1031  *
1032  * Both the implicit back refs and the full back refs for tree blocks
1033  * only consist of key. The key offset for the implicit back refs is
1034  * objectid of block's owner tree. The key offset for the full back refs
1035  * is the first byte of parent block.
1036  *
1037  * When implicit back refs is used, information about the lowest key and
1038  * level of the tree block are required. These information are stored in
1039  * tree block info structure.
1040  */
1041
1042 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1043 static int convert_extent_item_v0(struct btrfs_trans_handle *trans,
1044                                   struct btrfs_fs_info *fs_info,
1045                                   struct btrfs_path *path,
1046                                   u64 owner, u32 extra_size)
1047 {
1048         struct btrfs_root *root = fs_info->extent_root;
1049         struct btrfs_extent_item *item;
1050         struct btrfs_extent_item_v0 *ei0;
1051         struct btrfs_extent_ref_v0 *ref0;
1052         struct btrfs_tree_block_info *bi;
1053         struct extent_buffer *leaf;
1054         struct btrfs_key key;
1055         struct btrfs_key found_key;
1056         u32 new_size = sizeof(*item);
1057         u64 refs;
1058         int ret;
1059
1060         leaf = path->nodes[0];
1061         BUG_ON(btrfs_item_size_nr(leaf, path->slots[0]) != sizeof(*ei0));
1062
1063         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1064         ei0 = btrfs_item_ptr(leaf, path->slots[0],
1065                              struct btrfs_extent_item_v0);
1066         refs = btrfs_extent_refs_v0(leaf, ei0);
1067
1068         if (owner == (u64)-1) {
1069                 while (1) {
1070                         if (path->slots[0] >= btrfs_header_nritems(leaf)) {
1071                                 ret = btrfs_next_leaf(root, path);
1072                                 if (ret < 0)
1073                                         return ret;
1074                                 BUG_ON(ret > 0); /* Corruption */
1075                                 leaf = path->nodes[0];
1076                         }
1077                         btrfs_item_key_to_cpu(leaf, &found_key,
1078                                               path->slots[0]);
1079                         BUG_ON(key.objectid != found_key.objectid);
1080                         if (found_key.type != BTRFS_EXTENT_REF_V0_KEY) {
1081                                 path->slots[0]++;
1082                                 continue;
1083                         }
1084                         ref0 = btrfs_item_ptr(leaf, path->slots[0],
1085                                               struct btrfs_extent_ref_v0);
1086                         owner = btrfs_ref_objectid_v0(leaf, ref0);
1087                         break;
1088                 }
1089         }
1090         btrfs_release_path(path);
1091
1092         if (owner < BTRFS_FIRST_FREE_OBJECTID)
1093                 new_size += sizeof(*bi);
1094
1095         new_size -= sizeof(*ei0);
1096         ret = btrfs_search_slot(trans, root, &key, path,
1097                                 new_size + extra_size, 1);
1098         if (ret < 0)
1099                 return ret;
1100         BUG_ON(ret); /* Corruption */
1101
1102         btrfs_extend_item(fs_info, path, new_size);
1103
1104         leaf = path->nodes[0];
1105         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1106         btrfs_set_extent_refs(leaf, item, refs);
1107         /* FIXME: get real generation */
1108         btrfs_set_extent_generation(leaf, item, 0);
1109         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1110                 btrfs_set_extent_flags(leaf, item,
1111                                        BTRFS_EXTENT_FLAG_TREE_BLOCK |
1112                                        BTRFS_BLOCK_FLAG_FULL_BACKREF);
1113                 bi = (struct btrfs_tree_block_info *)(item + 1);
1114                 /* FIXME: get first key of the block */
1115                 memzero_extent_buffer(leaf, (unsigned long)bi, sizeof(*bi));
1116                 btrfs_set_tree_block_level(leaf, bi, (int)owner);
1117         } else {
1118                 btrfs_set_extent_flags(leaf, item, BTRFS_EXTENT_FLAG_DATA);
1119         }
1120         btrfs_mark_buffer_dirty(leaf);
1121         return 0;
1122 }
1123 #endif
1124
1125 /*
1126  * is_data == BTRFS_REF_TYPE_BLOCK, tree block type is required,
1127  * is_data == BTRFS_REF_TYPE_DATA, data type is requried,
1128  * is_data == BTRFS_REF_TYPE_ANY, either type is OK.
1129  */
1130 int btrfs_get_extent_inline_ref_type(const struct extent_buffer *eb,
1131                                      struct btrfs_extent_inline_ref *iref,
1132                                      enum btrfs_inline_ref_type is_data)
1133 {
1134         int type = btrfs_extent_inline_ref_type(eb, iref);
1135         u64 offset = btrfs_extent_inline_ref_offset(eb, iref);
1136
1137         if (type == BTRFS_TREE_BLOCK_REF_KEY ||
1138             type == BTRFS_SHARED_BLOCK_REF_KEY ||
1139             type == BTRFS_SHARED_DATA_REF_KEY ||
1140             type == BTRFS_EXTENT_DATA_REF_KEY) {
1141                 if (is_data == BTRFS_REF_TYPE_BLOCK) {
1142                         if (type == BTRFS_TREE_BLOCK_REF_KEY)
1143                                 return type;
1144                         if (type == BTRFS_SHARED_BLOCK_REF_KEY) {
1145                                 ASSERT(eb->fs_info);
1146                                 /*
1147                                  * Every shared one has parent tree
1148                                  * block, which must be aligned to
1149                                  * nodesize.
1150                                  */
1151                                 if (offset &&
1152                                     IS_ALIGNED(offset, eb->fs_info->nodesize))
1153                                         return type;
1154                         }
1155                 } else if (is_data == BTRFS_REF_TYPE_DATA) {
1156                         if (type == BTRFS_EXTENT_DATA_REF_KEY)
1157                                 return type;
1158                         if (type == BTRFS_SHARED_DATA_REF_KEY) {
1159                                 ASSERT(eb->fs_info);
1160                                 /*
1161                                  * Every shared one has parent tree
1162                                  * block, which must be aligned to
1163                                  * nodesize.
1164                                  */
1165                                 if (offset &&
1166                                     IS_ALIGNED(offset, eb->fs_info->nodesize))
1167                                         return type;
1168                         }
1169                 } else {
1170                         ASSERT(is_data == BTRFS_REF_TYPE_ANY);
1171                         return type;
1172                 }
1173         }
1174
1175         btrfs_print_leaf((struct extent_buffer *)eb);
1176         btrfs_err(eb->fs_info, "eb %llu invalid extent inline ref type %d",
1177                   eb->start, type);
1178         WARN_ON(1);
1179
1180         return BTRFS_REF_TYPE_INVALID;
1181 }
1182
1183 static u64 hash_extent_data_ref(u64 root_objectid, u64 owner, u64 offset)
1184 {
1185         u32 high_crc = ~(u32)0;
1186         u32 low_crc = ~(u32)0;
1187         __le64 lenum;
1188
1189         lenum = cpu_to_le64(root_objectid);
1190         high_crc = crc32c(high_crc, &lenum, sizeof(lenum));
1191         lenum = cpu_to_le64(owner);
1192         low_crc = crc32c(low_crc, &lenum, sizeof(lenum));
1193         lenum = cpu_to_le64(offset);
1194         low_crc = crc32c(low_crc, &lenum, sizeof(lenum));
1195
1196         return ((u64)high_crc << 31) ^ (u64)low_crc;
1197 }
1198
1199 static u64 hash_extent_data_ref_item(struct extent_buffer *leaf,
1200                                      struct btrfs_extent_data_ref *ref)
1201 {
1202         return hash_extent_data_ref(btrfs_extent_data_ref_root(leaf, ref),
1203                                     btrfs_extent_data_ref_objectid(leaf, ref),
1204                                     btrfs_extent_data_ref_offset(leaf, ref));
1205 }
1206
1207 static int match_extent_data_ref(struct extent_buffer *leaf,
1208                                  struct btrfs_extent_data_ref *ref,
1209                                  u64 root_objectid, u64 owner, u64 offset)
1210 {
1211         if (btrfs_extent_data_ref_root(leaf, ref) != root_objectid ||
1212             btrfs_extent_data_ref_objectid(leaf, ref) != owner ||
1213             btrfs_extent_data_ref_offset(leaf, ref) != offset)
1214                 return 0;
1215         return 1;
1216 }
1217
1218 static noinline int lookup_extent_data_ref(struct btrfs_trans_handle *trans,
1219                                            struct btrfs_fs_info *fs_info,
1220                                            struct btrfs_path *path,
1221                                            u64 bytenr, u64 parent,
1222                                            u64 root_objectid,
1223                                            u64 owner, u64 offset)
1224 {
1225         struct btrfs_root *root = fs_info->extent_root;
1226         struct btrfs_key key;
1227         struct btrfs_extent_data_ref *ref;
1228         struct extent_buffer *leaf;
1229         u32 nritems;
1230         int ret;
1231         int recow;
1232         int err = -ENOENT;
1233
1234         key.objectid = bytenr;
1235         if (parent) {
1236                 key.type = BTRFS_SHARED_DATA_REF_KEY;
1237                 key.offset = parent;
1238         } else {
1239                 key.type = BTRFS_EXTENT_DATA_REF_KEY;
1240                 key.offset = hash_extent_data_ref(root_objectid,
1241                                                   owner, offset);
1242         }
1243 again:
1244         recow = 0;
1245         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1246         if (ret < 0) {
1247                 err = ret;
1248                 goto fail;
1249         }
1250
1251         if (parent) {
1252                 if (!ret)
1253                         return 0;
1254 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1255                 key.type = BTRFS_EXTENT_REF_V0_KEY;
1256                 btrfs_release_path(path);
1257                 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1258                 if (ret < 0) {
1259                         err = ret;
1260                         goto fail;
1261                 }
1262                 if (!ret)
1263                         return 0;
1264 #endif
1265                 goto fail;
1266         }
1267
1268         leaf = path->nodes[0];
1269         nritems = btrfs_header_nritems(leaf);
1270         while (1) {
1271                 if (path->slots[0] >= nritems) {
1272                         ret = btrfs_next_leaf(root, path);
1273                         if (ret < 0)
1274                                 err = ret;
1275                         if (ret)
1276                                 goto fail;
1277
1278                         leaf = path->nodes[0];
1279                         nritems = btrfs_header_nritems(leaf);
1280                         recow = 1;
1281                 }
1282
1283                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1284                 if (key.objectid != bytenr ||
1285                     key.type != BTRFS_EXTENT_DATA_REF_KEY)
1286                         goto fail;
1287
1288                 ref = btrfs_item_ptr(leaf, path->slots[0],
1289                                      struct btrfs_extent_data_ref);
1290
1291                 if (match_extent_data_ref(leaf, ref, root_objectid,
1292                                           owner, offset)) {
1293                         if (recow) {
1294                                 btrfs_release_path(path);
1295                                 goto again;
1296                         }
1297                         err = 0;
1298                         break;
1299                 }
1300                 path->slots[0]++;
1301         }
1302 fail:
1303         return err;
1304 }
1305
1306 static noinline int insert_extent_data_ref(struct btrfs_trans_handle *trans,
1307                                            struct btrfs_path *path,
1308                                            u64 bytenr, u64 parent,
1309                                            u64 root_objectid, u64 owner,
1310                                            u64 offset, int refs_to_add)
1311 {
1312         struct btrfs_root *root = trans->fs_info->extent_root;
1313         struct btrfs_key key;
1314         struct extent_buffer *leaf;
1315         u32 size;
1316         u32 num_refs;
1317         int ret;
1318
1319         key.objectid = bytenr;
1320         if (parent) {
1321                 key.type = BTRFS_SHARED_DATA_REF_KEY;
1322                 key.offset = parent;
1323                 size = sizeof(struct btrfs_shared_data_ref);
1324         } else {
1325                 key.type = BTRFS_EXTENT_DATA_REF_KEY;
1326                 key.offset = hash_extent_data_ref(root_objectid,
1327                                                   owner, offset);
1328                 size = sizeof(struct btrfs_extent_data_ref);
1329         }
1330
1331         ret = btrfs_insert_empty_item(trans, root, path, &key, size);
1332         if (ret && ret != -EEXIST)
1333                 goto fail;
1334
1335         leaf = path->nodes[0];
1336         if (parent) {
1337                 struct btrfs_shared_data_ref *ref;
1338                 ref = btrfs_item_ptr(leaf, path->slots[0],
1339                                      struct btrfs_shared_data_ref);
1340                 if (ret == 0) {
1341                         btrfs_set_shared_data_ref_count(leaf, ref, refs_to_add);
1342                 } else {
1343                         num_refs = btrfs_shared_data_ref_count(leaf, ref);
1344                         num_refs += refs_to_add;
1345                         btrfs_set_shared_data_ref_count(leaf, ref, num_refs);
1346                 }
1347         } else {
1348                 struct btrfs_extent_data_ref *ref;
1349                 while (ret == -EEXIST) {
1350                         ref = btrfs_item_ptr(leaf, path->slots[0],
1351                                              struct btrfs_extent_data_ref);
1352                         if (match_extent_data_ref(leaf, ref, root_objectid,
1353                                                   owner, offset))
1354                                 break;
1355                         btrfs_release_path(path);
1356                         key.offset++;
1357                         ret = btrfs_insert_empty_item(trans, root, path, &key,
1358                                                       size);
1359                         if (ret && ret != -EEXIST)
1360                                 goto fail;
1361
1362                         leaf = path->nodes[0];
1363                 }
1364                 ref = btrfs_item_ptr(leaf, path->slots[0],
1365                                      struct btrfs_extent_data_ref);
1366                 if (ret == 0) {
1367                         btrfs_set_extent_data_ref_root(leaf, ref,
1368                                                        root_objectid);
1369                         btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
1370                         btrfs_set_extent_data_ref_offset(leaf, ref, offset);
1371                         btrfs_set_extent_data_ref_count(leaf, ref, refs_to_add);
1372                 } else {
1373                         num_refs = btrfs_extent_data_ref_count(leaf, ref);
1374                         num_refs += refs_to_add;
1375                         btrfs_set_extent_data_ref_count(leaf, ref, num_refs);
1376                 }
1377         }
1378         btrfs_mark_buffer_dirty(leaf);
1379         ret = 0;
1380 fail:
1381         btrfs_release_path(path);
1382         return ret;
1383 }
1384
1385 static noinline int remove_extent_data_ref(struct btrfs_trans_handle *trans,
1386                                            struct btrfs_fs_info *fs_info,
1387                                            struct btrfs_path *path,
1388                                            int refs_to_drop, int *last_ref)
1389 {
1390         struct btrfs_key key;
1391         struct btrfs_extent_data_ref *ref1 = NULL;
1392         struct btrfs_shared_data_ref *ref2 = NULL;
1393         struct extent_buffer *leaf;
1394         u32 num_refs = 0;
1395         int ret = 0;
1396
1397         leaf = path->nodes[0];
1398         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1399
1400         if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
1401                 ref1 = btrfs_item_ptr(leaf, path->slots[0],
1402                                       struct btrfs_extent_data_ref);
1403                 num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1404         } else if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
1405                 ref2 = btrfs_item_ptr(leaf, path->slots[0],
1406                                       struct btrfs_shared_data_ref);
1407                 num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1408 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1409         } else if (key.type == BTRFS_EXTENT_REF_V0_KEY) {
1410                 struct btrfs_extent_ref_v0 *ref0;
1411                 ref0 = btrfs_item_ptr(leaf, path->slots[0],
1412                                       struct btrfs_extent_ref_v0);
1413                 num_refs = btrfs_ref_count_v0(leaf, ref0);
1414 #endif
1415         } else {
1416                 BUG();
1417         }
1418
1419         BUG_ON(num_refs < refs_to_drop);
1420         num_refs -= refs_to_drop;
1421
1422         if (num_refs == 0) {
1423                 ret = btrfs_del_item(trans, fs_info->extent_root, path);
1424                 *last_ref = 1;
1425         } else {
1426                 if (key.type == BTRFS_EXTENT_DATA_REF_KEY)
1427                         btrfs_set_extent_data_ref_count(leaf, ref1, num_refs);
1428                 else if (key.type == BTRFS_SHARED_DATA_REF_KEY)
1429                         btrfs_set_shared_data_ref_count(leaf, ref2, num_refs);
1430 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1431                 else {
1432                         struct btrfs_extent_ref_v0 *ref0;
1433                         ref0 = btrfs_item_ptr(leaf, path->slots[0],
1434                                         struct btrfs_extent_ref_v0);
1435                         btrfs_set_ref_count_v0(leaf, ref0, num_refs);
1436                 }
1437 #endif
1438                 btrfs_mark_buffer_dirty(leaf);
1439         }
1440         return ret;
1441 }
1442
1443 static noinline u32 extent_data_ref_count(struct btrfs_path *path,
1444                                           struct btrfs_extent_inline_ref *iref)
1445 {
1446         struct btrfs_key key;
1447         struct extent_buffer *leaf;
1448         struct btrfs_extent_data_ref *ref1;
1449         struct btrfs_shared_data_ref *ref2;
1450         u32 num_refs = 0;
1451         int type;
1452
1453         leaf = path->nodes[0];
1454         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1455         if (iref) {
1456                 /*
1457                  * If type is invalid, we should have bailed out earlier than
1458                  * this call.
1459                  */
1460                 type = btrfs_get_extent_inline_ref_type(leaf, iref, BTRFS_REF_TYPE_DATA);
1461                 ASSERT(type != BTRFS_REF_TYPE_INVALID);
1462                 if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1463                         ref1 = (struct btrfs_extent_data_ref *)(&iref->offset);
1464                         num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1465                 } else {
1466                         ref2 = (struct btrfs_shared_data_ref *)(iref + 1);
1467                         num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1468                 }
1469         } else if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
1470                 ref1 = btrfs_item_ptr(leaf, path->slots[0],
1471                                       struct btrfs_extent_data_ref);
1472                 num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1473         } else if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
1474                 ref2 = btrfs_item_ptr(leaf, path->slots[0],
1475                                       struct btrfs_shared_data_ref);
1476                 num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1477 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1478         } else if (key.type == BTRFS_EXTENT_REF_V0_KEY) {
1479                 struct btrfs_extent_ref_v0 *ref0;
1480                 ref0 = btrfs_item_ptr(leaf, path->slots[0],
1481                                       struct btrfs_extent_ref_v0);
1482                 num_refs = btrfs_ref_count_v0(leaf, ref0);
1483 #endif
1484         } else {
1485                 WARN_ON(1);
1486         }
1487         return num_refs;
1488 }
1489
1490 static noinline int lookup_tree_block_ref(struct btrfs_trans_handle *trans,
1491                                           struct btrfs_fs_info *fs_info,
1492                                           struct btrfs_path *path,
1493                                           u64 bytenr, u64 parent,
1494                                           u64 root_objectid)
1495 {
1496         struct btrfs_root *root = fs_info->extent_root;
1497         struct btrfs_key key;
1498         int ret;
1499
1500         key.objectid = bytenr;
1501         if (parent) {
1502                 key.type = BTRFS_SHARED_BLOCK_REF_KEY;
1503                 key.offset = parent;
1504         } else {
1505                 key.type = BTRFS_TREE_BLOCK_REF_KEY;
1506                 key.offset = root_objectid;
1507         }
1508
1509         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1510         if (ret > 0)
1511                 ret = -ENOENT;
1512 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1513         if (ret == -ENOENT && parent) {
1514                 btrfs_release_path(path);
1515                 key.type = BTRFS_EXTENT_REF_V0_KEY;
1516                 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1517                 if (ret > 0)
1518                         ret = -ENOENT;
1519         }
1520 #endif
1521         return ret;
1522 }
1523
1524 static noinline int insert_tree_block_ref(struct btrfs_trans_handle *trans,
1525                                           struct btrfs_path *path,
1526                                           u64 bytenr, u64 parent,
1527                                           u64 root_objectid)
1528 {
1529         struct btrfs_key key;
1530         int ret;
1531
1532         key.objectid = bytenr;
1533         if (parent) {
1534                 key.type = BTRFS_SHARED_BLOCK_REF_KEY;
1535                 key.offset = parent;
1536         } else {
1537                 key.type = BTRFS_TREE_BLOCK_REF_KEY;
1538                 key.offset = root_objectid;
1539         }
1540
1541         ret = btrfs_insert_empty_item(trans, trans->fs_info->extent_root,
1542                                       path, &key, 0);
1543         btrfs_release_path(path);
1544         return ret;
1545 }
1546
1547 static inline int extent_ref_type(u64 parent, u64 owner)
1548 {
1549         int type;
1550         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1551                 if (parent > 0)
1552                         type = BTRFS_SHARED_BLOCK_REF_KEY;
1553                 else
1554                         type = BTRFS_TREE_BLOCK_REF_KEY;
1555         } else {
1556                 if (parent > 0)
1557                         type = BTRFS_SHARED_DATA_REF_KEY;
1558                 else
1559                         type = BTRFS_EXTENT_DATA_REF_KEY;
1560         }
1561         return type;
1562 }
1563
1564 static int find_next_key(struct btrfs_path *path, int level,
1565                          struct btrfs_key *key)
1566
1567 {
1568         for (; level < BTRFS_MAX_LEVEL; level++) {
1569                 if (!path->nodes[level])
1570                         break;
1571                 if (path->slots[level] + 1 >=
1572                     btrfs_header_nritems(path->nodes[level]))
1573                         continue;
1574                 if (level == 0)
1575                         btrfs_item_key_to_cpu(path->nodes[level], key,
1576                                               path->slots[level] + 1);
1577                 else
1578                         btrfs_node_key_to_cpu(path->nodes[level], key,
1579                                               path->slots[level] + 1);
1580                 return 0;
1581         }
1582         return 1;
1583 }
1584
1585 /*
1586  * look for inline back ref. if back ref is found, *ref_ret is set
1587  * to the address of inline back ref, and 0 is returned.
1588  *
1589  * if back ref isn't found, *ref_ret is set to the address where it
1590  * should be inserted, and -ENOENT is returned.
1591  *
1592  * if insert is true and there are too many inline back refs, the path
1593  * points to the extent item, and -EAGAIN is returned.
1594  *
1595  * NOTE: inline back refs are ordered in the same way that back ref
1596  *       items in the tree are ordered.
1597  */
1598 static noinline_for_stack
1599 int lookup_inline_extent_backref(struct btrfs_trans_handle *trans,
1600                                  struct btrfs_fs_info *fs_info,
1601                                  struct btrfs_path *path,
1602                                  struct btrfs_extent_inline_ref **ref_ret,
1603                                  u64 bytenr, u64 num_bytes,
1604                                  u64 parent, u64 root_objectid,
1605                                  u64 owner, u64 offset, int insert)
1606 {
1607         struct btrfs_root *root = fs_info->extent_root;
1608         struct btrfs_key key;
1609         struct extent_buffer *leaf;
1610         struct btrfs_extent_item *ei;
1611         struct btrfs_extent_inline_ref *iref;
1612         u64 flags;
1613         u64 item_size;
1614         unsigned long ptr;
1615         unsigned long end;
1616         int extra_size;
1617         int type;
1618         int want;
1619         int ret;
1620         int err = 0;
1621         bool skinny_metadata = btrfs_fs_incompat(fs_info, SKINNY_METADATA);
1622         int needed;
1623
1624         key.objectid = bytenr;
1625         key.type = BTRFS_EXTENT_ITEM_KEY;
1626         key.offset = num_bytes;
1627
1628         want = extent_ref_type(parent, owner);
1629         if (insert) {
1630                 extra_size = btrfs_extent_inline_ref_size(want);
1631                 path->keep_locks = 1;
1632         } else
1633                 extra_size = -1;
1634
1635         /*
1636          * Owner is our level, so we can just add one to get the level for the
1637          * block we are interested in.
1638          */
1639         if (skinny_metadata && owner < BTRFS_FIRST_FREE_OBJECTID) {
1640                 key.type = BTRFS_METADATA_ITEM_KEY;
1641                 key.offset = owner;
1642         }
1643
1644 again:
1645         ret = btrfs_search_slot(trans, root, &key, path, extra_size, 1);
1646         if (ret < 0) {
1647                 err = ret;
1648                 goto out;
1649         }
1650
1651         /*
1652          * We may be a newly converted file system which still has the old fat
1653          * extent entries for metadata, so try and see if we have one of those.
1654          */
1655         if (ret > 0 && skinny_metadata) {
1656                 skinny_metadata = false;
1657                 if (path->slots[0]) {
1658                         path->slots[0]--;
1659                         btrfs_item_key_to_cpu(path->nodes[0], &key,
1660                                               path->slots[0]);
1661                         if (key.objectid == bytenr &&
1662                             key.type == BTRFS_EXTENT_ITEM_KEY &&
1663                             key.offset == num_bytes)
1664                                 ret = 0;
1665                 }
1666                 if (ret) {
1667                         key.objectid = bytenr;
1668                         key.type = BTRFS_EXTENT_ITEM_KEY;
1669                         key.offset = num_bytes;
1670                         btrfs_release_path(path);
1671                         goto again;
1672                 }
1673         }
1674
1675         if (ret && !insert) {
1676                 err = -ENOENT;
1677                 goto out;
1678         } else if (WARN_ON(ret)) {
1679                 err = -EIO;
1680                 goto out;
1681         }
1682
1683         leaf = path->nodes[0];
1684         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1685 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1686         if (item_size < sizeof(*ei)) {
1687                 if (!insert) {
1688                         err = -ENOENT;
1689                         goto out;
1690                 }
1691                 ret = convert_extent_item_v0(trans, fs_info, path, owner,
1692                                              extra_size);
1693                 if (ret < 0) {
1694                         err = ret;
1695                         goto out;
1696                 }
1697                 leaf = path->nodes[0];
1698                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1699         }
1700 #endif
1701         BUG_ON(item_size < sizeof(*ei));
1702
1703         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1704         flags = btrfs_extent_flags(leaf, ei);
1705
1706         ptr = (unsigned long)(ei + 1);
1707         end = (unsigned long)ei + item_size;
1708
1709         if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK && !skinny_metadata) {
1710                 ptr += sizeof(struct btrfs_tree_block_info);
1711                 BUG_ON(ptr > end);
1712         }
1713
1714         if (owner >= BTRFS_FIRST_FREE_OBJECTID)
1715                 needed = BTRFS_REF_TYPE_DATA;
1716         else
1717                 needed = BTRFS_REF_TYPE_BLOCK;
1718
1719         err = -ENOENT;
1720         while (1) {
1721                 if (ptr >= end) {
1722                         WARN_ON(ptr > end);
1723                         break;
1724                 }
1725                 iref = (struct btrfs_extent_inline_ref *)ptr;
1726                 type = btrfs_get_extent_inline_ref_type(leaf, iref, needed);
1727                 if (type == BTRFS_REF_TYPE_INVALID) {
1728                         err = -EINVAL;
1729                         goto out;
1730                 }
1731
1732                 if (want < type)
1733                         break;
1734                 if (want > type) {
1735                         ptr += btrfs_extent_inline_ref_size(type);
1736                         continue;
1737                 }
1738
1739                 if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1740                         struct btrfs_extent_data_ref *dref;
1741                         dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1742                         if (match_extent_data_ref(leaf, dref, root_objectid,
1743                                                   owner, offset)) {
1744                                 err = 0;
1745                                 break;
1746                         }
1747                         if (hash_extent_data_ref_item(leaf, dref) <
1748                             hash_extent_data_ref(root_objectid, owner, offset))
1749                                 break;
1750                 } else {
1751                         u64 ref_offset;
1752                         ref_offset = btrfs_extent_inline_ref_offset(leaf, iref);
1753                         if (parent > 0) {
1754                                 if (parent == ref_offset) {
1755                                         err = 0;
1756                                         break;
1757                                 }
1758                                 if (ref_offset < parent)
1759                                         break;
1760                         } else {
1761                                 if (root_objectid == ref_offset) {
1762                                         err = 0;
1763                                         break;
1764                                 }
1765                                 if (ref_offset < root_objectid)
1766                                         break;
1767                         }
1768                 }
1769                 ptr += btrfs_extent_inline_ref_size(type);
1770         }
1771         if (err == -ENOENT && insert) {
1772                 if (item_size + extra_size >=
1773                     BTRFS_MAX_EXTENT_ITEM_SIZE(root)) {
1774                         err = -EAGAIN;
1775                         goto out;
1776                 }
1777                 /*
1778                  * To add new inline back ref, we have to make sure
1779                  * there is no corresponding back ref item.
1780                  * For simplicity, we just do not add new inline back
1781                  * ref if there is any kind of item for this block
1782                  */
1783                 if (find_next_key(path, 0, &key) == 0 &&
1784                     key.objectid == bytenr &&
1785                     key.type < BTRFS_BLOCK_GROUP_ITEM_KEY) {
1786                         err = -EAGAIN;
1787                         goto out;
1788                 }
1789         }
1790         *ref_ret = (struct btrfs_extent_inline_ref *)ptr;
1791 out:
1792         if (insert) {
1793                 path->keep_locks = 0;
1794                 btrfs_unlock_up_safe(path, 1);
1795         }
1796         return err;
1797 }
1798
1799 /*
1800  * helper to add new inline back ref
1801  */
1802 static noinline_for_stack
1803 void setup_inline_extent_backref(struct btrfs_fs_info *fs_info,
1804                                  struct btrfs_path *path,
1805                                  struct btrfs_extent_inline_ref *iref,
1806                                  u64 parent, u64 root_objectid,
1807                                  u64 owner, u64 offset, int refs_to_add,
1808                                  struct btrfs_delayed_extent_op *extent_op)
1809 {
1810         struct extent_buffer *leaf;
1811         struct btrfs_extent_item *ei;
1812         unsigned long ptr;
1813         unsigned long end;
1814         unsigned long item_offset;
1815         u64 refs;
1816         int size;
1817         int type;
1818
1819         leaf = path->nodes[0];
1820         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1821         item_offset = (unsigned long)iref - (unsigned long)ei;
1822
1823         type = extent_ref_type(parent, owner);
1824         size = btrfs_extent_inline_ref_size(type);
1825
1826         btrfs_extend_item(fs_info, path, size);
1827
1828         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1829         refs = btrfs_extent_refs(leaf, ei);
1830         refs += refs_to_add;
1831         btrfs_set_extent_refs(leaf, ei, refs);
1832         if (extent_op)
1833                 __run_delayed_extent_op(extent_op, leaf, ei);
1834
1835         ptr = (unsigned long)ei + item_offset;
1836         end = (unsigned long)ei + btrfs_item_size_nr(leaf, path->slots[0]);
1837         if (ptr < end - size)
1838                 memmove_extent_buffer(leaf, ptr + size, ptr,
1839                                       end - size - ptr);
1840
1841         iref = (struct btrfs_extent_inline_ref *)ptr;
1842         btrfs_set_extent_inline_ref_type(leaf, iref, type);
1843         if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1844                 struct btrfs_extent_data_ref *dref;
1845                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1846                 btrfs_set_extent_data_ref_root(leaf, dref, root_objectid);
1847                 btrfs_set_extent_data_ref_objectid(leaf, dref, owner);
1848                 btrfs_set_extent_data_ref_offset(leaf, dref, offset);
1849                 btrfs_set_extent_data_ref_count(leaf, dref, refs_to_add);
1850         } else if (type == BTRFS_SHARED_DATA_REF_KEY) {
1851                 struct btrfs_shared_data_ref *sref;
1852                 sref = (struct btrfs_shared_data_ref *)(iref + 1);
1853                 btrfs_set_shared_data_ref_count(leaf, sref, refs_to_add);
1854                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
1855         } else if (type == BTRFS_SHARED_BLOCK_REF_KEY) {
1856                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
1857         } else {
1858                 btrfs_set_extent_inline_ref_offset(leaf, iref, root_objectid);
1859         }
1860         btrfs_mark_buffer_dirty(leaf);
1861 }
1862
1863 static int lookup_extent_backref(struct btrfs_trans_handle *trans,
1864                                  struct btrfs_fs_info *fs_info,
1865                                  struct btrfs_path *path,
1866                                  struct btrfs_extent_inline_ref **ref_ret,
1867                                  u64 bytenr, u64 num_bytes, u64 parent,
1868                                  u64 root_objectid, u64 owner, u64 offset)
1869 {
1870         int ret;
1871
1872         ret = lookup_inline_extent_backref(trans, fs_info, path, ref_ret,
1873                                            bytenr, num_bytes, parent,
1874                                            root_objectid, owner, offset, 0);
1875         if (ret != -ENOENT)
1876                 return ret;
1877
1878         btrfs_release_path(path);
1879         *ref_ret = NULL;
1880
1881         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1882                 ret = lookup_tree_block_ref(trans, fs_info, path, bytenr,
1883                                             parent, root_objectid);
1884         } else {
1885                 ret = lookup_extent_data_ref(trans, fs_info, path, bytenr,
1886                                              parent, root_objectid, owner,
1887                                              offset);
1888         }
1889         return ret;
1890 }
1891
1892 /*
1893  * helper to update/remove inline back ref
1894  */
1895 static noinline_for_stack
1896 void update_inline_extent_backref(struct btrfs_fs_info *fs_info,
1897                                   struct btrfs_path *path,
1898                                   struct btrfs_extent_inline_ref *iref,
1899                                   int refs_to_mod,
1900                                   struct btrfs_delayed_extent_op *extent_op,
1901                                   int *last_ref)
1902 {
1903         struct extent_buffer *leaf;
1904         struct btrfs_extent_item *ei;
1905         struct btrfs_extent_data_ref *dref = NULL;
1906         struct btrfs_shared_data_ref *sref = NULL;
1907         unsigned long ptr;
1908         unsigned long end;
1909         u32 item_size;
1910         int size;
1911         int type;
1912         u64 refs;
1913
1914         leaf = path->nodes[0];
1915         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1916         refs = btrfs_extent_refs(leaf, ei);
1917         WARN_ON(refs_to_mod < 0 && refs + refs_to_mod <= 0);
1918         refs += refs_to_mod;
1919         btrfs_set_extent_refs(leaf, ei, refs);
1920         if (extent_op)
1921                 __run_delayed_extent_op(extent_op, leaf, ei);
1922
1923         /*
1924          * If type is invalid, we should have bailed out after
1925          * lookup_inline_extent_backref().
1926          */
1927         type = btrfs_get_extent_inline_ref_type(leaf, iref, BTRFS_REF_TYPE_ANY);
1928         ASSERT(type != BTRFS_REF_TYPE_INVALID);
1929
1930         if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1931                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1932                 refs = btrfs_extent_data_ref_count(leaf, dref);
1933         } else if (type == BTRFS_SHARED_DATA_REF_KEY) {
1934                 sref = (struct btrfs_shared_data_ref *)(iref + 1);
1935                 refs = btrfs_shared_data_ref_count(leaf, sref);
1936         } else {
1937                 refs = 1;
1938                 BUG_ON(refs_to_mod != -1);
1939         }
1940
1941         BUG_ON(refs_to_mod < 0 && refs < -refs_to_mod);
1942         refs += refs_to_mod;
1943
1944         if (refs > 0) {
1945                 if (type == BTRFS_EXTENT_DATA_REF_KEY)
1946                         btrfs_set_extent_data_ref_count(leaf, dref, refs);
1947                 else
1948                         btrfs_set_shared_data_ref_count(leaf, sref, refs);
1949         } else {
1950                 *last_ref = 1;
1951                 size =  btrfs_extent_inline_ref_size(type);
1952                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1953                 ptr = (unsigned long)iref;
1954                 end = (unsigned long)ei + item_size;
1955                 if (ptr + size < end)
1956                         memmove_extent_buffer(leaf, ptr, ptr + size,
1957                                               end - ptr - size);
1958                 item_size -= size;
1959                 btrfs_truncate_item(fs_info, path, item_size, 1);
1960         }
1961         btrfs_mark_buffer_dirty(leaf);
1962 }
1963
1964 static noinline_for_stack
1965 int insert_inline_extent_backref(struct btrfs_trans_handle *trans,
1966                                  struct btrfs_fs_info *fs_info,
1967                                  struct btrfs_path *path,
1968                                  u64 bytenr, u64 num_bytes, u64 parent,
1969                                  u64 root_objectid, u64 owner,
1970                                  u64 offset, int refs_to_add,
1971                                  struct btrfs_delayed_extent_op *extent_op)
1972 {
1973         struct btrfs_extent_inline_ref *iref;
1974         int ret;
1975
1976         ret = lookup_inline_extent_backref(trans, fs_info, path, &iref,
1977                                            bytenr, num_bytes, parent,
1978                                            root_objectid, owner, offset, 1);
1979         if (ret == 0) {
1980                 BUG_ON(owner < BTRFS_FIRST_FREE_OBJECTID);
1981                 update_inline_extent_backref(fs_info, path, iref,
1982                                              refs_to_add, extent_op, NULL);
1983         } else if (ret == -ENOENT) {
1984                 setup_inline_extent_backref(fs_info, path, iref, parent,
1985                                             root_objectid, owner, offset,
1986                                             refs_to_add, extent_op);
1987                 ret = 0;
1988         }
1989         return ret;
1990 }
1991
1992 static int insert_extent_backref(struct btrfs_trans_handle *trans,
1993                                  struct btrfs_fs_info *fs_info,
1994                                  struct btrfs_path *path,
1995                                  u64 bytenr, u64 parent, u64 root_objectid,
1996                                  u64 owner, u64 offset, int refs_to_add)
1997 {
1998         int ret;
1999         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
2000                 BUG_ON(refs_to_add != 1);
2001                 ret = insert_tree_block_ref(trans, path, bytenr, parent,
2002                                             root_objectid);
2003         } else {
2004                 ret = insert_extent_data_ref(trans, path, bytenr, parent,
2005                                              root_objectid, owner, offset,
2006                                              refs_to_add);
2007         }
2008         return ret;
2009 }
2010
2011 static int remove_extent_backref(struct btrfs_trans_handle *trans,
2012                                  struct btrfs_fs_info *fs_info,
2013                                  struct btrfs_path *path,
2014                                  struct btrfs_extent_inline_ref *iref,
2015                                  int refs_to_drop, int is_data, int *last_ref)
2016 {
2017         int ret = 0;
2018
2019         BUG_ON(!is_data && refs_to_drop != 1);
2020         if (iref) {
2021                 update_inline_extent_backref(fs_info, path, iref,
2022                                              -refs_to_drop, NULL, last_ref);
2023         } else if (is_data) {
2024                 ret = remove_extent_data_ref(trans, fs_info, path, refs_to_drop,
2025                                              last_ref);
2026         } else {
2027                 *last_ref = 1;
2028                 ret = btrfs_del_item(trans, fs_info->extent_root, path);
2029         }
2030         return ret;
2031 }
2032
2033 #define in_range(b, first, len)        ((b) >= (first) && (b) < (first) + (len))
2034 static int btrfs_issue_discard(struct block_device *bdev, u64 start, u64 len,
2035                                u64 *discarded_bytes)
2036 {
2037         int j, ret = 0;
2038         u64 bytes_left, end;
2039         u64 aligned_start = ALIGN(start, 1 << 9);
2040
2041         if (WARN_ON(start != aligned_start)) {
2042                 len -= aligned_start - start;
2043                 len = round_down(len, 1 << 9);
2044                 start = aligned_start;
2045         }
2046
2047         *discarded_bytes = 0;
2048
2049         if (!len)
2050                 return 0;
2051
2052         end = start + len;
2053         bytes_left = len;
2054
2055         /* Skip any superblocks on this device. */
2056         for (j = 0; j < BTRFS_SUPER_MIRROR_MAX; j++) {
2057                 u64 sb_start = btrfs_sb_offset(j);
2058                 u64 sb_end = sb_start + BTRFS_SUPER_INFO_SIZE;
2059                 u64 size = sb_start - start;
2060
2061                 if (!in_range(sb_start, start, bytes_left) &&
2062                     !in_range(sb_end, start, bytes_left) &&
2063                     !in_range(start, sb_start, BTRFS_SUPER_INFO_SIZE))
2064                         continue;
2065
2066                 /*
2067                  * Superblock spans beginning of range.  Adjust start and
2068                  * try again.
2069                  */
2070                 if (sb_start <= start) {
2071                         start += sb_end - start;
2072                         if (start > end) {
2073                                 bytes_left = 0;
2074                                 break;
2075                         }
2076                         bytes_left = end - start;
2077                         continue;
2078                 }
2079
2080                 if (size) {
2081                         ret = blkdev_issue_discard(bdev, start >> 9, size >> 9,
2082                                                    GFP_NOFS, 0);
2083                         if (!ret)
2084                                 *discarded_bytes += size;
2085                         else if (ret != -EOPNOTSUPP)
2086                                 return ret;
2087                 }
2088
2089                 start = sb_end;
2090                 if (start > end) {
2091                         bytes_left = 0;
2092                         break;
2093                 }
2094                 bytes_left = end - start;
2095         }
2096
2097         if (bytes_left) {
2098                 ret = blkdev_issue_discard(bdev, start >> 9, bytes_left >> 9,
2099                                            GFP_NOFS, 0);
2100                 if (!ret)
2101                         *discarded_bytes += bytes_left;
2102         }
2103         return ret;
2104 }
2105
2106 int btrfs_discard_extent(struct btrfs_fs_info *fs_info, u64 bytenr,
2107                          u64 num_bytes, u64 *actual_bytes)
2108 {
2109         int ret;
2110         u64 discarded_bytes = 0;
2111         struct btrfs_bio *bbio = NULL;
2112
2113
2114         /*
2115          * Avoid races with device replace and make sure our bbio has devices
2116          * associated to its stripes that don't go away while we are discarding.
2117          */
2118         btrfs_bio_counter_inc_blocked(fs_info);
2119         /* Tell the block device(s) that the sectors can be discarded */
2120         ret = btrfs_map_block(fs_info, BTRFS_MAP_DISCARD, bytenr, &num_bytes,
2121                               &bbio, 0);
2122         /* Error condition is -ENOMEM */
2123         if (!ret) {
2124                 struct btrfs_bio_stripe *stripe = bbio->stripes;
2125                 int i;
2126
2127
2128                 for (i = 0; i < bbio->num_stripes; i++, stripe++) {
2129                         u64 bytes;
2130                         struct request_queue *req_q;
2131
2132                         if (!stripe->dev->bdev) {
2133                                 ASSERT(btrfs_test_opt(fs_info, DEGRADED));
2134                                 continue;
2135                         }
2136                         req_q = bdev_get_queue(stripe->dev->bdev);
2137                         if (!blk_queue_discard(req_q))
2138                                 continue;
2139
2140                         ret = btrfs_issue_discard(stripe->dev->bdev,
2141                                                   stripe->physical,
2142                                                   stripe->length,
2143                                                   &bytes);
2144                         if (!ret)
2145                                 discarded_bytes += bytes;
2146                         else if (ret != -EOPNOTSUPP)
2147                                 break; /* Logic errors or -ENOMEM, or -EIO but I don't know how that could happen JDM */
2148
2149                         /*
2150                          * Just in case we get back EOPNOTSUPP for some reason,
2151                          * just ignore the return value so we don't screw up
2152                          * people calling discard_extent.
2153                          */
2154                         ret = 0;
2155                 }
2156                 btrfs_put_bbio(bbio);
2157         }
2158         btrfs_bio_counter_dec(fs_info);
2159
2160         if (actual_bytes)
2161                 *actual_bytes = discarded_bytes;
2162
2163
2164         if (ret == -EOPNOTSUPP)
2165                 ret = 0;
2166         return ret;
2167 }
2168
2169 /* Can return -ENOMEM */
2170 int btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
2171                          struct btrfs_root *root,
2172                          u64 bytenr, u64 num_bytes, u64 parent,
2173                          u64 root_objectid, u64 owner, u64 offset)
2174 {
2175         struct btrfs_fs_info *fs_info = root->fs_info;
2176         int old_ref_mod, new_ref_mod;
2177         int ret;
2178
2179         BUG_ON(owner < BTRFS_FIRST_FREE_OBJECTID &&
2180                root_objectid == BTRFS_TREE_LOG_OBJECTID);
2181
2182         btrfs_ref_tree_mod(root, bytenr, num_bytes, parent, root_objectid,
2183                            owner, offset, BTRFS_ADD_DELAYED_REF);
2184
2185         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
2186                 ret = btrfs_add_delayed_tree_ref(fs_info, trans, bytenr,
2187                                                  num_bytes, parent,
2188                                                  root_objectid, (int)owner,
2189                                                  BTRFS_ADD_DELAYED_REF, NULL,
2190                                                  &old_ref_mod, &new_ref_mod);
2191         } else {
2192                 ret = btrfs_add_delayed_data_ref(fs_info, trans, bytenr,
2193                                                  num_bytes, parent,
2194                                                  root_objectid, owner, offset,
2195                                                  0, BTRFS_ADD_DELAYED_REF,
2196                                                  &old_ref_mod, &new_ref_mod);
2197         }
2198
2199         if (ret == 0 && old_ref_mod < 0 && new_ref_mod >= 0) {
2200                 bool metadata = owner < BTRFS_FIRST_FREE_OBJECTID;
2201
2202                 add_pinned_bytes(fs_info, -num_bytes, metadata, root_objectid);
2203         }
2204
2205         return ret;
2206 }
2207
2208 /*
2209  * __btrfs_inc_extent_ref - insert backreference for a given extent
2210  *
2211  * @trans:          Handle of transaction
2212  *
2213  * @node:           The delayed ref node used to get the bytenr/length for
2214  *                  extent whose references are incremented.
2215  *
2216  * @parent:         If this is a shared extent (BTRFS_SHARED_DATA_REF_KEY/
2217  *                  BTRFS_SHARED_BLOCK_REF_KEY) then it holds the logical
2218  *                  bytenr of the parent block. Since new extents are always
2219  *                  created with indirect references, this will only be the case
2220  *                  when relocating a shared extent. In that case, root_objectid
2221  *                  will be BTRFS_TREE_RELOC_OBJECTID. Otheriwse, parent must
2222  *                  be 0
2223  *
2224  * @root_objectid:  The id of the root where this modification has originated,
2225  *                  this can be either one of the well-known metadata trees or
2226  *                  the subvolume id which references this extent.
2227  *
2228  * @owner:          For data extents it is the inode number of the owning file.
2229  *                  For metadata extents this parameter holds the level in the
2230  *                  tree of the extent.
2231  *
2232  * @offset:         For metadata extents the offset is ignored and is currently
2233  *                  always passed as 0. For data extents it is the fileoffset
2234  *                  this extent belongs to.
2235  *
2236  * @refs_to_add     Number of references to add
2237  *
2238  * @extent_op       Pointer to a structure, holding information necessary when
2239  *                  updating a tree block's flags
2240  *
2241  */
2242 static int __btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
2243                                   struct btrfs_fs_info *fs_info,
2244                                   struct btrfs_delayed_ref_node *node,
2245                                   u64 parent, u64 root_objectid,
2246                                   u64 owner, u64 offset, int refs_to_add,
2247                                   struct btrfs_delayed_extent_op *extent_op)
2248 {
2249         struct btrfs_path *path;
2250         struct extent_buffer *leaf;
2251         struct btrfs_extent_item *item;
2252         struct btrfs_key key;
2253         u64 bytenr = node->bytenr;
2254         u64 num_bytes = node->num_bytes;
2255         u64 refs;
2256         int ret;
2257
2258         path = btrfs_alloc_path();
2259         if (!path)
2260                 return -ENOMEM;
2261
2262         path->reada = READA_FORWARD;
2263         path->leave_spinning = 1;
2264         /* this will setup the path even if it fails to insert the back ref */
2265         ret = insert_inline_extent_backref(trans, fs_info, path, bytenr,
2266                                            num_bytes, parent, root_objectid,
2267                                            owner, offset,
2268                                            refs_to_add, extent_op);
2269         if ((ret < 0 && ret != -EAGAIN) || !ret)
2270                 goto out;
2271
2272         /*
2273          * Ok we had -EAGAIN which means we didn't have space to insert and
2274          * inline extent ref, so just update the reference count and add a
2275          * normal backref.
2276          */
2277         leaf = path->nodes[0];
2278         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
2279         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2280         refs = btrfs_extent_refs(leaf, item);
2281         btrfs_set_extent_refs(leaf, item, refs + refs_to_add);
2282         if (extent_op)
2283                 __run_delayed_extent_op(extent_op, leaf, item);
2284
2285         btrfs_mark_buffer_dirty(leaf);
2286         btrfs_release_path(path);
2287
2288         path->reada = READA_FORWARD;
2289         path->leave_spinning = 1;
2290         /* now insert the actual backref */
2291         ret = insert_extent_backref(trans, fs_info, path, bytenr, parent,
2292                                     root_objectid, owner, offset, refs_to_add);
2293         if (ret)
2294                 btrfs_abort_transaction(trans, ret);
2295 out:
2296         btrfs_free_path(path);
2297         return ret;
2298 }
2299
2300 static int run_delayed_data_ref(struct btrfs_trans_handle *trans,
2301                                 struct btrfs_fs_info *fs_info,
2302                                 struct btrfs_delayed_ref_node *node,
2303                                 struct btrfs_delayed_extent_op *extent_op,
2304                                 int insert_reserved)
2305 {
2306         int ret = 0;
2307         struct btrfs_delayed_data_ref *ref;
2308         struct btrfs_key ins;
2309         u64 parent = 0;
2310         u64 ref_root = 0;
2311         u64 flags = 0;
2312
2313         ins.objectid = node->bytenr;
2314         ins.offset = node->num_bytes;
2315         ins.type = BTRFS_EXTENT_ITEM_KEY;
2316
2317         ref = btrfs_delayed_node_to_data_ref(node);
2318         trace_run_delayed_data_ref(fs_info, node, ref, node->action);
2319
2320         if (node->type == BTRFS_SHARED_DATA_REF_KEY)
2321                 parent = ref->parent;
2322         ref_root = ref->root;
2323
2324         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
2325                 if (extent_op)
2326                         flags |= extent_op->flags_to_set;
2327                 ret = alloc_reserved_file_extent(trans, fs_info,
2328                                                  parent, ref_root, flags,
2329                                                  ref->objectid, ref->offset,
2330                                                  &ins, node->ref_mod);
2331         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
2332                 ret = __btrfs_inc_extent_ref(trans, fs_info, node, parent,
2333                                              ref_root, ref->objectid,
2334                                              ref->offset, node->ref_mod,
2335                                              extent_op);
2336         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
2337                 ret = __btrfs_free_extent(trans, fs_info, node, parent,
2338                                           ref_root, ref->objectid,
2339                                           ref->offset, node->ref_mod,
2340                                           extent_op);
2341         } else {
2342                 BUG();
2343         }
2344         return ret;
2345 }
2346
2347 static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op,
2348                                     struct extent_buffer *leaf,
2349                                     struct btrfs_extent_item *ei)
2350 {
2351         u64 flags = btrfs_extent_flags(leaf, ei);
2352         if (extent_op->update_flags) {
2353                 flags |= extent_op->flags_to_set;
2354                 btrfs_set_extent_flags(leaf, ei, flags);
2355         }
2356
2357         if (extent_op->update_key) {
2358                 struct btrfs_tree_block_info *bi;
2359                 BUG_ON(!(flags & BTRFS_EXTENT_FLAG_TREE_BLOCK));
2360                 bi = (struct btrfs_tree_block_info *)(ei + 1);
2361                 btrfs_set_tree_block_key(leaf, bi, &extent_op->key);
2362         }
2363 }
2364
2365 static int run_delayed_extent_op(struct btrfs_trans_handle *trans,
2366                                  struct btrfs_fs_info *fs_info,
2367                                  struct btrfs_delayed_ref_head *head,
2368                                  struct btrfs_delayed_extent_op *extent_op)
2369 {
2370         struct btrfs_key key;
2371         struct btrfs_path *path;
2372         struct btrfs_extent_item *ei;
2373         struct extent_buffer *leaf;
2374         u32 item_size;
2375         int ret;
2376         int err = 0;
2377         int metadata = !extent_op->is_data;
2378
2379         if (trans->aborted)
2380                 return 0;
2381
2382         if (metadata && !btrfs_fs_incompat(fs_info, SKINNY_METADATA))
2383                 metadata = 0;
2384
2385         path = btrfs_alloc_path();
2386         if (!path)
2387                 return -ENOMEM;
2388
2389         key.objectid = head->bytenr;
2390
2391         if (metadata) {
2392                 key.type = BTRFS_METADATA_ITEM_KEY;
2393                 key.offset = extent_op->level;
2394         } else {
2395                 key.type = BTRFS_EXTENT_ITEM_KEY;
2396                 key.offset = head->num_bytes;
2397         }
2398
2399 again:
2400         path->reada = READA_FORWARD;
2401         path->leave_spinning = 1;
2402         ret = btrfs_search_slot(trans, fs_info->extent_root, &key, path, 0, 1);
2403         if (ret < 0) {
2404                 err = ret;
2405                 goto out;
2406         }
2407         if (ret > 0) {
2408                 if (metadata) {
2409                         if (path->slots[0] > 0) {
2410                                 path->slots[0]--;
2411                                 btrfs_item_key_to_cpu(path->nodes[0], &key,
2412                                                       path->slots[0]);
2413                                 if (key.objectid == head->bytenr &&
2414                                     key.type == BTRFS_EXTENT_ITEM_KEY &&
2415                                     key.offset == head->num_bytes)
2416                                         ret = 0;
2417                         }
2418                         if (ret > 0) {
2419                                 btrfs_release_path(path);
2420                                 metadata = 0;
2421
2422                                 key.objectid = head->bytenr;
2423                                 key.offset = head->num_bytes;
2424                                 key.type = BTRFS_EXTENT_ITEM_KEY;
2425                                 goto again;
2426                         }
2427                 } else {
2428                         err = -EIO;
2429                         goto out;
2430                 }
2431         }
2432
2433         leaf = path->nodes[0];
2434         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2435 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
2436         if (item_size < sizeof(*ei)) {
2437                 ret = convert_extent_item_v0(trans, fs_info, path, (u64)-1, 0);
2438                 if (ret < 0) {
2439                         err = ret;
2440                         goto out;
2441                 }
2442                 leaf = path->nodes[0];
2443                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2444         }
2445 #endif
2446         BUG_ON(item_size < sizeof(*ei));
2447         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2448         __run_delayed_extent_op(extent_op, leaf, ei);
2449
2450         btrfs_mark_buffer_dirty(leaf);
2451 out:
2452         btrfs_free_path(path);
2453         return err;
2454 }
2455
2456 static int run_delayed_tree_ref(struct btrfs_trans_handle *trans,
2457                                 struct btrfs_fs_info *fs_info,
2458                                 struct btrfs_delayed_ref_node *node,
2459                                 struct btrfs_delayed_extent_op *extent_op,
2460                                 int insert_reserved)
2461 {
2462         int ret = 0;
2463         struct btrfs_delayed_tree_ref *ref;
2464         u64 parent = 0;
2465         u64 ref_root = 0;
2466
2467         ref = btrfs_delayed_node_to_tree_ref(node);
2468         trace_run_delayed_tree_ref(fs_info, node, ref, node->action);
2469
2470         if (node->type == BTRFS_SHARED_BLOCK_REF_KEY)
2471                 parent = ref->parent;
2472         ref_root = ref->root;
2473
2474         if (node->ref_mod != 1) {
2475                 btrfs_err(fs_info,
2476         "btree block(%llu) has %d references rather than 1: action %d ref_root %llu parent %llu",
2477                           node->bytenr, node->ref_mod, node->action, ref_root,
2478                           parent);
2479                 return -EIO;
2480         }
2481         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
2482                 BUG_ON(!extent_op || !extent_op->update_flags);
2483                 ret = alloc_reserved_tree_block(trans, node, extent_op);
2484         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
2485                 ret = __btrfs_inc_extent_ref(trans, fs_info, node,
2486                                              parent, ref_root,
2487                                              ref->level, 0, 1,
2488                                              extent_op);
2489         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
2490                 ret = __btrfs_free_extent(trans, fs_info, node,
2491                                           parent, ref_root,
2492                                           ref->level, 0, 1, extent_op);
2493         } else {
2494                 BUG();
2495         }
2496         return ret;
2497 }
2498
2499 /* helper function to actually process a single delayed ref entry */
2500 static int run_one_delayed_ref(struct btrfs_trans_handle *trans,
2501                                struct btrfs_fs_info *fs_info,
2502                                struct btrfs_delayed_ref_node *node,
2503                                struct btrfs_delayed_extent_op *extent_op,
2504                                int insert_reserved)
2505 {
2506         int ret = 0;
2507
2508         if (trans->aborted) {
2509                 if (insert_reserved)
2510                         btrfs_pin_extent(fs_info, node->bytenr,
2511                                          node->num_bytes, 1);
2512                 return 0;
2513         }
2514
2515         if (node->type == BTRFS_TREE_BLOCK_REF_KEY ||
2516             node->type == BTRFS_SHARED_BLOCK_REF_KEY)
2517                 ret = run_delayed_tree_ref(trans, fs_info, node, extent_op,
2518                                            insert_reserved);
2519         else if (node->type == BTRFS_EXTENT_DATA_REF_KEY ||
2520                  node->type == BTRFS_SHARED_DATA_REF_KEY)
2521                 ret = run_delayed_data_ref(trans, fs_info, node, extent_op,
2522                                            insert_reserved);
2523         else
2524                 BUG();
2525         return ret;
2526 }
2527
2528 static inline struct btrfs_delayed_ref_node *
2529 select_delayed_ref(struct btrfs_delayed_ref_head *head)
2530 {
2531         struct btrfs_delayed_ref_node *ref;
2532
2533         if (RB_EMPTY_ROOT(&head->ref_tree))
2534                 return NULL;
2535
2536         /*
2537          * Select a delayed ref of type BTRFS_ADD_DELAYED_REF first.
2538          * This is to prevent a ref count from going down to zero, which deletes
2539          * the extent item from the extent tree, when there still are references
2540          * to add, which would fail because they would not find the extent item.
2541          */
2542         if (!list_empty(&head->ref_add_list))
2543                 return list_first_entry(&head->ref_add_list,
2544                                 struct btrfs_delayed_ref_node, add_list);
2545
2546         ref = rb_entry(rb_first(&head->ref_tree),
2547                        struct btrfs_delayed_ref_node, ref_node);
2548         ASSERT(list_empty(&ref->add_list));
2549         return ref;
2550 }
2551
2552 static void unselect_delayed_ref_head(struct btrfs_delayed_ref_root *delayed_refs,
2553                                       struct btrfs_delayed_ref_head *head)
2554 {
2555         spin_lock(&delayed_refs->lock);
2556         head->processing = 0;
2557         delayed_refs->num_heads_ready++;
2558         spin_unlock(&delayed_refs->lock);
2559         btrfs_delayed_ref_unlock(head);
2560 }
2561
2562 static int cleanup_extent_op(struct btrfs_trans_handle *trans,
2563                              struct btrfs_fs_info *fs_info,
2564                              struct btrfs_delayed_ref_head *head)
2565 {
2566         struct btrfs_delayed_extent_op *extent_op = head->extent_op;
2567         int ret;
2568
2569         if (!extent_op)
2570                 return 0;
2571         head->extent_op = NULL;
2572         if (head->must_insert_reserved) {
2573                 btrfs_free_delayed_extent_op(extent_op);
2574                 return 0;
2575         }
2576         spin_unlock(&head->lock);
2577         ret = run_delayed_extent_op(trans, fs_info, head, extent_op);
2578         btrfs_free_delayed_extent_op(extent_op);
2579         return ret ? ret : 1;
2580 }
2581
2582 static int cleanup_ref_head(struct btrfs_trans_handle *trans,
2583                             struct btrfs_fs_info *fs_info,
2584                             struct btrfs_delayed_ref_head *head)
2585 {
2586         struct btrfs_delayed_ref_root *delayed_refs;
2587         int ret;
2588
2589         delayed_refs = &trans->transaction->delayed_refs;
2590
2591         ret = cleanup_extent_op(trans, fs_info, head);
2592         if (ret < 0) {
2593                 unselect_delayed_ref_head(delayed_refs, head);
2594                 btrfs_debug(fs_info, "run_delayed_extent_op returned %d", ret);
2595                 return ret;
2596         } else if (ret) {
2597                 return ret;
2598         }
2599
2600         /*
2601          * Need to drop our head ref lock and re-acquire the delayed ref lock
2602          * and then re-check to make sure nobody got added.
2603          */
2604         spin_unlock(&head->lock);
2605         spin_lock(&delayed_refs->lock);
2606         spin_lock(&head->lock);
2607         if (!RB_EMPTY_ROOT(&head->ref_tree) || head->extent_op) {
2608                 spin_unlock(&head->lock);
2609                 spin_unlock(&delayed_refs->lock);
2610                 return 1;
2611         }
2612         delayed_refs->num_heads--;
2613         rb_erase(&head->href_node, &delayed_refs->href_root);
2614         RB_CLEAR_NODE(&head->href_node);
2615         spin_unlock(&head->lock);
2616         spin_unlock(&delayed_refs->lock);
2617         atomic_dec(&delayed_refs->num_entries);
2618
2619         trace_run_delayed_ref_head(fs_info, head, 0);
2620
2621         if (head->total_ref_mod < 0) {
2622                 struct btrfs_space_info *space_info;
2623                 u64 flags;
2624
2625                 if (head->is_data)
2626                         flags = BTRFS_BLOCK_GROUP_DATA;
2627                 else if (head->is_system)
2628                         flags = BTRFS_BLOCK_GROUP_SYSTEM;
2629                 else
2630                         flags = BTRFS_BLOCK_GROUP_METADATA;
2631                 space_info = __find_space_info(fs_info, flags);
2632                 ASSERT(space_info);
2633                 percpu_counter_add(&space_info->total_bytes_pinned,
2634                                    -head->num_bytes);
2635
2636                 if (head->is_data) {
2637                         spin_lock(&delayed_refs->lock);
2638                         delayed_refs->pending_csums -= head->num_bytes;
2639                         spin_unlock(&delayed_refs->lock);
2640                 }
2641         }
2642
2643         if (head->must_insert_reserved) {
2644                 btrfs_pin_extent(fs_info, head->bytenr,
2645                                  head->num_bytes, 1);
2646                 if (head->is_data) {
2647                         ret = btrfs_del_csums(trans, fs_info, head->bytenr,
2648                                               head->num_bytes);
2649                 }
2650         }
2651
2652         /* Also free its reserved qgroup space */
2653         btrfs_qgroup_free_delayed_ref(fs_info, head->qgroup_ref_root,
2654                                       head->qgroup_reserved);
2655         btrfs_delayed_ref_unlock(head);
2656         btrfs_put_delayed_ref_head(head);
2657         return 0;
2658 }
2659
2660 /*
2661  * Returns 0 on success or if called with an already aborted transaction.
2662  * Returns -ENOMEM or -EIO on failure and will abort the transaction.
2663  */
2664 static noinline int __btrfs_run_delayed_refs(struct btrfs_trans_handle *trans,
2665                                              unsigned long nr)
2666 {
2667         struct btrfs_fs_info *fs_info = trans->fs_info;
2668         struct btrfs_delayed_ref_root *delayed_refs;
2669         struct btrfs_delayed_ref_node *ref;
2670         struct btrfs_delayed_ref_head *locked_ref = NULL;
2671         struct btrfs_delayed_extent_op *extent_op;
2672         ktime_t start = ktime_get();
2673         int ret;
2674         unsigned long count = 0;
2675         unsigned long actual_count = 0;
2676         int must_insert_reserved = 0;
2677
2678         delayed_refs = &trans->transaction->delayed_refs;
2679         while (1) {
2680                 if (!locked_ref) {
2681                         if (count >= nr)
2682                                 break;
2683
2684                         spin_lock(&delayed_refs->lock);
2685                         locked_ref = btrfs_select_ref_head(trans);
2686                         if (!locked_ref) {
2687                                 spin_unlock(&delayed_refs->lock);
2688                                 break;
2689                         }
2690
2691                         /* grab the lock that says we are going to process
2692                          * all the refs for this head */
2693                         ret = btrfs_delayed_ref_lock(trans, locked_ref);
2694                         spin_unlock(&delayed_refs->lock);
2695                         /*
2696                          * we may have dropped the spin lock to get the head
2697                          * mutex lock, and that might have given someone else
2698                          * time to free the head.  If that's true, it has been
2699                          * removed from our list and we can move on.
2700                          */
2701                         if (ret == -EAGAIN) {
2702                                 locked_ref = NULL;
2703                                 count++;
2704                                 continue;
2705                         }
2706                 }
2707
2708                 /*
2709                  * We need to try and merge add/drops of the same ref since we
2710                  * can run into issues with relocate dropping the implicit ref
2711                  * and then it being added back again before the drop can
2712                  * finish.  If we merged anything we need to re-loop so we can
2713                  * get a good ref.
2714                  * Or we can get node references of the same type that weren't
2715                  * merged when created due to bumps in the tree mod seq, and
2716                  * we need to merge them to prevent adding an inline extent
2717                  * backref before dropping it (triggering a BUG_ON at
2718                  * insert_inline_extent_backref()).
2719                  */
2720                 spin_lock(&locked_ref->lock);
2721                 btrfs_merge_delayed_refs(trans, delayed_refs, locked_ref);
2722
2723                 ref = select_delayed_ref(locked_ref);
2724
2725                 if (ref && ref->seq &&
2726                     btrfs_check_delayed_seq(fs_info, ref->seq)) {
2727                         spin_unlock(&locked_ref->lock);
2728                         unselect_delayed_ref_head(delayed_refs, locked_ref);
2729                         locked_ref = NULL;
2730                         cond_resched();
2731                         count++;
2732                         continue;
2733                 }
2734
2735                 /*
2736                  * We're done processing refs in this ref_head, clean everything
2737                  * up and move on to the next ref_head.
2738                  */
2739                 if (!ref) {
2740                         ret = cleanup_ref_head(trans, fs_info, locked_ref);
2741                         if (ret > 0 ) {
2742                                 /* We dropped our lock, we need to loop. */
2743                                 ret = 0;
2744                                 continue;
2745                         } else if (ret) {
2746                                 return ret;
2747                         }
2748                         locked_ref = NULL;
2749                         count++;
2750                         continue;
2751                 }
2752
2753                 actual_count++;
2754                 ref->in_tree = 0;
2755                 rb_erase(&ref->ref_node, &locked_ref->ref_tree);
2756                 RB_CLEAR_NODE(&ref->ref_node);
2757                 if (!list_empty(&ref->add_list))
2758                         list_del(&ref->add_list);
2759                 /*
2760                  * When we play the delayed ref, also correct the ref_mod on
2761                  * head
2762                  */
2763                 switch (ref->action) {
2764                 case BTRFS_ADD_DELAYED_REF:
2765                 case BTRFS_ADD_DELAYED_EXTENT:
2766                         locked_ref->ref_mod -= ref->ref_mod;
2767                         break;
2768                 case BTRFS_DROP_DELAYED_REF:
2769                         locked_ref->ref_mod += ref->ref_mod;
2770                         break;
2771                 default:
2772                         WARN_ON(1);
2773                 }
2774                 atomic_dec(&delayed_refs->num_entries);
2775
2776                 /*
2777                  * Record the must-insert_reserved flag before we drop the spin
2778                  * lock.
2779                  */
2780                 must_insert_reserved = locked_ref->must_insert_reserved;
2781                 locked_ref->must_insert_reserved = 0;
2782
2783                 extent_op = locked_ref->extent_op;
2784                 locked_ref->extent_op = NULL;
2785                 spin_unlock(&locked_ref->lock);
2786
2787                 ret = run_one_delayed_ref(trans, fs_info, ref, extent_op,
2788                                           must_insert_reserved);
2789
2790                 btrfs_free_delayed_extent_op(extent_op);
2791                 if (ret) {
2792                         unselect_delayed_ref_head(delayed_refs, locked_ref);
2793                         btrfs_put_delayed_ref(ref);
2794                         btrfs_debug(fs_info, "run_one_delayed_ref returned %d",
2795                                     ret);
2796                         return ret;
2797                 }
2798
2799                 btrfs_put_delayed_ref(ref);
2800                 count++;
2801                 cond_resched();
2802         }
2803
2804         /*
2805          * We don't want to include ref heads since we can have empty ref heads
2806          * and those will drastically skew our runtime down since we just do
2807          * accounting, no actual extent tree updates.
2808          */
2809         if (actual_count > 0) {
2810                 u64 runtime = ktime_to_ns(ktime_sub(ktime_get(), start));
2811                 u64 avg;
2812
2813                 /*
2814                  * We weigh the current average higher than our current runtime
2815                  * to avoid large swings in the average.
2816                  */
2817                 spin_lock(&delayed_refs->lock);
2818                 avg = fs_info->avg_delayed_ref_runtime * 3 + runtime;
2819                 fs_info->avg_delayed_ref_runtime = avg >> 2;    /* div by 4 */
2820                 spin_unlock(&delayed_refs->lock);
2821         }
2822         return 0;
2823 }
2824
2825 #ifdef SCRAMBLE_DELAYED_REFS
2826 /*
2827  * Normally delayed refs get processed in ascending bytenr order. This
2828  * correlates in most cases to the order added. To expose dependencies on this
2829  * order, we start to process the tree in the middle instead of the beginning
2830  */
2831 static u64 find_middle(struct rb_root *root)
2832 {
2833         struct rb_node *n = root->rb_node;
2834         struct btrfs_delayed_ref_node *entry;
2835         int alt = 1;
2836         u64 middle;
2837         u64 first = 0, last = 0;
2838
2839         n = rb_first(root);
2840         if (n) {
2841                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2842                 first = entry->bytenr;
2843         }
2844         n = rb_last(root);
2845         if (n) {
2846                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2847                 last = entry->bytenr;
2848         }
2849         n = root->rb_node;
2850
2851         while (n) {
2852                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2853                 WARN_ON(!entry->in_tree);
2854
2855                 middle = entry->bytenr;
2856
2857                 if (alt)
2858                         n = n->rb_left;
2859                 else
2860                         n = n->rb_right;
2861
2862                 alt = 1 - alt;
2863         }
2864         return middle;
2865 }
2866 #endif
2867
2868 static inline u64 heads_to_leaves(struct btrfs_fs_info *fs_info, u64 heads)
2869 {
2870         u64 num_bytes;
2871
2872         num_bytes = heads * (sizeof(struct btrfs_extent_item) +
2873                              sizeof(struct btrfs_extent_inline_ref));
2874         if (!btrfs_fs_incompat(fs_info, SKINNY_METADATA))
2875                 num_bytes += heads * sizeof(struct btrfs_tree_block_info);
2876
2877         /*
2878          * We don't ever fill up leaves all the way so multiply by 2 just to be
2879          * closer to what we're really going to want to use.
2880          */
2881         return div_u64(num_bytes, BTRFS_LEAF_DATA_SIZE(fs_info));
2882 }
2883
2884 /*
2885  * Takes the number of bytes to be csumm'ed and figures out how many leaves it
2886  * would require to store the csums for that many bytes.
2887  */
2888 u64 btrfs_csum_bytes_to_leaves(struct btrfs_fs_info *fs_info, u64 csum_bytes)
2889 {
2890         u64 csum_size;
2891         u64 num_csums_per_leaf;
2892         u64 num_csums;
2893
2894         csum_size = BTRFS_MAX_ITEM_SIZE(fs_info);
2895         num_csums_per_leaf = div64_u64(csum_size,
2896                         (u64)btrfs_super_csum_size(fs_info->super_copy));
2897         num_csums = div64_u64(csum_bytes, fs_info->sectorsize);
2898         num_csums += num_csums_per_leaf - 1;
2899         num_csums = div64_u64(num_csums, num_csums_per_leaf);
2900         return num_csums;
2901 }
2902
2903 int btrfs_check_space_for_delayed_refs(struct btrfs_trans_handle *trans,
2904                                        struct btrfs_fs_info *fs_info)
2905 {
2906         struct btrfs_block_rsv *global_rsv;
2907         u64 num_heads = trans->transaction->delayed_refs.num_heads_ready;
2908         u64 csum_bytes = trans->transaction->delayed_refs.pending_csums;
2909         unsigned int num_dirty_bgs = trans->transaction->num_dirty_bgs;
2910         u64 num_bytes, num_dirty_bgs_bytes;
2911         int ret = 0;
2912
2913         num_bytes = btrfs_calc_trans_metadata_size(fs_info, 1);
2914         num_heads = heads_to_leaves(fs_info, num_heads);
2915         if (num_heads > 1)
2916                 num_bytes += (num_heads - 1) * fs_info->nodesize;
2917         num_bytes <<= 1;
2918         num_bytes += btrfs_csum_bytes_to_leaves(fs_info, csum_bytes) *
2919                                                         fs_info->nodesize;
2920         num_dirty_bgs_bytes = btrfs_calc_trans_metadata_size(fs_info,
2921                                                              num_dirty_bgs);
2922         global_rsv = &fs_info->global_block_rsv;
2923
2924         /*
2925          * If we can't allocate any more chunks lets make sure we have _lots_ of
2926          * wiggle room since running delayed refs can create more delayed refs.
2927          */
2928         if (global_rsv->space_info->full) {
2929                 num_dirty_bgs_bytes <<= 1;
2930                 num_bytes <<= 1;
2931         }
2932
2933         spin_lock(&global_rsv->lock);
2934         if (global_rsv->reserved <= num_bytes + num_dirty_bgs_bytes)
2935                 ret = 1;
2936         spin_unlock(&global_rsv->lock);
2937         return ret;
2938 }
2939
2940 int btrfs_should_throttle_delayed_refs(struct btrfs_trans_handle *trans,
2941                                        struct btrfs_fs_info *fs_info)
2942 {
2943         u64 num_entries =
2944                 atomic_read(&trans->transaction->delayed_refs.num_entries);
2945         u64 avg_runtime;
2946         u64 val;
2947
2948         smp_mb();
2949         avg_runtime = fs_info->avg_delayed_ref_runtime;
2950         val = num_entries * avg_runtime;
2951         if (val >= NSEC_PER_SEC)
2952                 return 1;
2953         if (val >= NSEC_PER_SEC / 2)
2954                 return 2;
2955
2956         return btrfs_check_space_for_delayed_refs(trans, fs_info);
2957 }
2958
2959 struct async_delayed_refs {
2960         struct btrfs_root *root;
2961         u64 transid;
2962         int count;
2963         int error;
2964         int sync;
2965         struct completion wait;
2966         struct btrfs_work work;
2967 };
2968
2969 static inline struct async_delayed_refs *
2970 to_async_delayed_refs(struct btrfs_work *work)
2971 {
2972         return container_of(work, struct async_delayed_refs, work);
2973 }
2974
2975 static void delayed_ref_async_start(struct btrfs_work *work)
2976 {
2977         struct async_delayed_refs *async = to_async_delayed_refs(work);
2978         struct btrfs_trans_handle *trans;
2979         struct btrfs_fs_info *fs_info = async->root->fs_info;
2980         int ret;
2981
2982         /* if the commit is already started, we don't need to wait here */
2983         if (btrfs_transaction_blocked(fs_info))
2984                 goto done;
2985
2986         trans = btrfs_join_transaction(async->root);
2987         if (IS_ERR(trans)) {
2988                 async->error = PTR_ERR(trans);
2989                 goto done;
2990         }
2991
2992         /*
2993          * trans->sync means that when we call end_transaction, we won't
2994          * wait on delayed refs
2995          */
2996         trans->sync = true;
2997
2998         /* Don't bother flushing if we got into a different transaction */
2999         if (trans->transid > async->transid)
3000                 goto end;
3001
3002         ret = btrfs_run_delayed_refs(trans, async->count);
3003         if (ret)
3004                 async->error = ret;
3005 end:
3006         ret = btrfs_end_transaction(trans);
3007         if (ret && !async->error)
3008                 async->error = ret;
3009 done:
3010         if (async->sync)
3011                 complete(&async->wait);
3012         else
3013                 kfree(async);
3014 }
3015
3016 int btrfs_async_run_delayed_refs(struct btrfs_fs_info *fs_info,
3017                                  unsigned long count, u64 transid, int wait)
3018 {
3019         struct async_delayed_refs *async;
3020         int ret;
3021
3022         async = kmalloc(sizeof(*async), GFP_NOFS);
3023         if (!async)
3024                 return -ENOMEM;
3025
3026         async->root = fs_info->tree_root;
3027         async->count = count;
3028         async->error = 0;
3029         async->transid = transid;
3030         if (wait)
3031                 async->sync = 1;
3032         else
3033                 async->sync = 0;
3034         init_completion(&async->wait);
3035
3036         btrfs_init_work(&async->work, btrfs_extent_refs_helper,
3037                         delayed_ref_async_start, NULL, NULL);
3038
3039         btrfs_queue_work(fs_info->extent_workers, &async->work);
3040
3041         if (wait) {
3042                 wait_for_completion(&async->wait);
3043                 ret = async->error;
3044                 kfree(async);
3045                 return ret;
3046         }
3047         return 0;
3048 }
3049
3050 /*
3051  * this starts processing the delayed reference count updates and
3052  * extent insertions we have queued up so far.  count can be
3053  * 0, which means to process everything in the tree at the start
3054  * of the run (but not newly added entries), or it can be some target
3055  * number you'd like to process.
3056  *
3057  * Returns 0 on success or if called with an aborted transaction
3058  * Returns <0 on error and aborts the transaction
3059  */
3060 int btrfs_run_delayed_refs(struct btrfs_trans_handle *trans,
3061                            unsigned long count)
3062 {
3063         struct btrfs_fs_info *fs_info = trans->fs_info;
3064         struct rb_node *node;
3065         struct btrfs_delayed_ref_root *delayed_refs;
3066     &nb