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 /*
  55  * Declare a helper function to detect underflow of various space info members
  56  */
  57 #define DECLARE_SPACE_INFO_UPDATE(name)                                 \
  58 static inline void update_##name(struct btrfs_space_info *sinfo,        \
  59                                  s64 bytes)                             \
  60 {                                                                       \
  61         if (bytes < 0 && sinfo->name < -bytes) {                        \
  62                 WARN_ON(1);                                             \
  63                 sinfo->name = 0;                                        \
  64                 return;                                                 \
  65         }                                                               \
  66         sinfo->name += bytes;                                           \
  67 }
  68
  69 DECLARE_SPACE_INFO_UPDATE(bytes_may_use);
  70 DECLARE_SPACE_INFO_UPDATE(bytes_pinned);
  71
  72 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
  73                                struct btrfs_delayed_ref_node *node, u64 parent,
  74                                u64 root_objectid, u64 owner_objectid,
  75                                u64 owner_offset, int refs_to_drop,
  76                                struct btrfs_delayed_extent_op *extra_op);
  77 static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op,
  78                                     struct extent_buffer *leaf,
  79                                     struct btrfs_extent_item *ei);
  80 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
  81                                       u64 parent, u64 root_objectid,
  82                                       u64 flags, u64 owner, u64 offset,
  83                                       struct btrfs_key *ins, int ref_mod);
  84 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
  85                                      struct btrfs_delayed_ref_node *node,
  86                                      struct btrfs_delayed_extent_op *extent_op);
  87 static int do_chunk_alloc(struct btrfs_trans_handle *trans, u64 flags,
  88                           int force);
  89 static int find_next_key(struct btrfs_path *path, int level,
  90                          struct btrfs_key *key);
  91 static void dump_space_info(struct btrfs_fs_info *fs_info,
  92                             struct btrfs_space_info *info, u64 bytes,
  93                             int dump_block_groups);
  94 static int block_rsv_use_bytes(struct btrfs_block_rsv *block_rsv,
  95                                u64 num_bytes);
  96 static void space_info_add_new_bytes(struct btrfs_fs_info *fs_info,
  97                                      struct btrfs_space_info *space_info,
  98                                      u64 num_bytes);
  99 static void space_info_add_old_bytes(struct btrfs_fs_info *fs_info,
 100                                      struct btrfs_space_info *space_info,
 101                                      u64 num_bytes);
 102
 103 static noinline int
 104 block_group_cache_done(struct btrfs_block_group_cache *cache)
 105 {
 106         smp_mb();
 107         return cache->cached == BTRFS_CACHE_FINISHED ||
 108                 cache->cached == BTRFS_CACHE_ERROR;
 109 }
 110
 111 static int block_group_bits(struct btrfs_block_group_cache *cache, u64 bits)
 112 {
 113         return (cache->flags & bits) == bits;
 114 }
 115
 116 void btrfs_get_block_group(struct btrfs_block_group_cache *cache)
 117 {
 118         atomic_inc(&cache->count);
 119 }
 120
 121 void btrfs_put_block_group(struct btrfs_block_group_cache *cache)
 122 {
 123         if (atomic_dec_and_test(&cache->count)) {
 124                 WARN_ON(cache->pinned > 0);
 125                 WARN_ON(cache->reserved > 0);
 126
 127                 /*
 128                  * If not empty, someone is still holding mutex of
 129                  * full_stripe_lock, which can only be released by caller.
 130                  * And it will definitely cause use-after-free when caller
 131                  * tries to release full stripe lock.
 132                  *
 133                  * No better way to resolve, but only to warn.
 134                  */
 135                 WARN_ON(!RB_EMPTY_ROOT(&cache->full_stripe_locks_root.root));
 136                 kfree(cache->free_space_ctl);
 137                 kfree(cache);
 138         }
 139 }
 140
 141 /*
 142  * this adds the block group to the fs_info rb tree for the block group
 143  * cache
 144  */
 145 static int btrfs_add_block_group_cache(struct btrfs_fs_info *info,
 146                                 struct btrfs_block_group_cache *block_group)
 147 {
 148         struct rb_node **p;
 149         struct rb_node *parent = NULL;
 150         struct btrfs_block_group_cache *cache;
 151
 152         spin_lock(&info->block_group_cache_lock);
 153         p = &info->block_group_cache_tree.rb_node;
 154
 155         while (*p) {
 156                 parent = *p;
 157                 cache = rb_entry(parent, struct btrfs_block_group_cache,
 158                                  cache_node);
 159                 if (block_group->key.objectid < cache->key.objectid) {
 160                         p = &(*p)->rb_left;
 161                 } else if (block_group->key.objectid > cache->key.objectid) {
 162                         p = &(*p)->rb_right;
 163                 } else {
 164                         spin_unlock(&info->block_group_cache_lock);
 165                         return -EEXIST;
 166                 }
 167         }
 168
 169         rb_link_node(&block_group->cache_node, parent, p);
 170         rb_insert_color(&block_group->cache_node,
 171                         &info->block_group_cache_tree);
 172
 173         if (info->first_logical_byte > block_group->key.objectid)
 174                 info->first_logical_byte = block_group->key.objectid;
 175
 176         spin_unlock(&info->block_group_cache_lock);
 177
 178         return 0;
 179 }
 180
 181 /*
 182  * This will return the block group at or after bytenr if contains is 0, else
 183  * it will return the block group that contains the bytenr
 184  */
 185 static struct btrfs_block_group_cache *
 186 block_group_cache_tree_search(struct btrfs_fs_info *info, u64 bytenr,
 187                               int contains)
 188 {
 189         struct btrfs_block_group_cache *cache, *ret = NULL;
 190         struct rb_node *n;
 191         u64 end, start;
 192
 193         spin_lock(&info->block_group_cache_lock);
 194         n = info->block_group_cache_tree.rb_node;
 195
 196         while (n) {
 197                 cache = rb_entry(n, struct btrfs_block_group_cache,
 198                                  cache_node);
 199                 end = cache->key.objectid + cache->key.offset - 1;
 200                 start = cache->key.objectid;
 201
 202                 if (bytenr < start) {
 203                         if (!contains && (!ret || start < ret->key.objectid))
 204                                 ret = cache;
 205                         n = n->rb_left;
 206                 } else if (bytenr > start) {
 207                         if (contains && bytenr <= end) {
 208                                 ret = cache;
 209                                 break;
 210                         }
 211                         n = n->rb_right;
 212                 } else {
 213                         ret = cache;
 214                         break;
 215                 }
 216         }
 217         if (ret) {
 218                 btrfs_get_block_group(ret);
 219                 if (bytenr == 0 && info->first_logical_byte > ret->key.objectid)
 220                         info->first_logical_byte = ret->key.objectid;
 221         }
 222         spin_unlock(&info->block_group_cache_lock);
 223
 224         return ret;
 225 }
 226
 227 static int add_excluded_extent(struct btrfs_fs_info *fs_info,
 228                                u64 start, u64 num_bytes)
 229 {
 230         u64 end = start + num_bytes - 1;
 231         set_extent_bits(&fs_info->freed_extents[0],
 232                         start, end, EXTENT_UPTODATE);
 233         set_extent_bits(&fs_info->freed_extents[1],
 234                         start, end, EXTENT_UPTODATE);
 235         return 0;
 236 }
 237
 238 static void free_excluded_extents(struct btrfs_block_group_cache *cache)
 239 {
 240         struct btrfs_fs_info *fs_info = cache->fs_info;
 241         u64 start, end;
 242
 243         start = cache->key.objectid;
 244         end = start + cache->key.offset - 1;
 245
 246         clear_extent_bits(&fs_info->freed_extents[0],
 247                           start, end, EXTENT_UPTODATE);
 248         clear_extent_bits(&fs_info->freed_extents[1],
 249                           start, end, EXTENT_UPTODATE);
 250 }
 251
 252 static int exclude_super_stripes(struct btrfs_block_group_cache *cache)
 253 {
 254         struct btrfs_fs_info *fs_info = cache->fs_info;
 255         u64 bytenr;
 256         u64 *logical;
 257         int stripe_len;
 258         int i, nr, ret;
 259
 260         if (cache->key.objectid < BTRFS_SUPER_INFO_OFFSET) {
 261                 stripe_len = BTRFS_SUPER_INFO_OFFSET - cache->key.objectid;
 262                 cache->bytes_super += stripe_len;
 263                 ret = add_excluded_extent(fs_info, cache->key.objectid,
 264                                           stripe_len);
 265                 if (ret)
 266                         return ret;
 267         }
 268
 269         for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
 270                 bytenr = btrfs_sb_offset(i);
 271                 ret = btrfs_rmap_block(fs_info, cache->key.objectid,
 272                                        bytenr, &logical, &nr, &stripe_len);
 273                 if (ret)
 274                         return ret;
 275
 276                 while (nr--) {
 277                         u64 start, len;
 278
 279                         if (logical[nr] > cache->key.objectid +
 280                             cache->key.offset)
 281                                 continue;
 282
 283                         if (logical[nr] + stripe_len <= cache->key.objectid)
 284                                 continue;
 285
 286                         start = logical[nr];
 287                         if (start < cache->key.objectid) {
 288                                 start = cache->key.objectid;
 289                                 len = (logical[nr] + stripe_len) - start;
 290                         } else {
 291                                 len = min_t(u64, stripe_len,
 292                                             cache->key.objectid +
 293                                             cache->key.offset - start);
 294                         }
 295
 296                         cache->bytes_super += len;
 297                         ret = add_excluded_extent(fs_info, start, len);
 298                         if (ret) {
 299                                 kfree(logical);
 300                                 return ret;
 301                         }
 302                 }
 303
 304                 kfree(logical);
 305         }
 306         return 0;
 307 }
 308
 309 static struct btrfs_caching_control *
 310 get_caching_control(struct btrfs_block_group_cache *cache)
 311 {
 312         struct btrfs_caching_control *ctl;
 313
 314         spin_lock(&cache->lock);
 315         if (!cache->caching_ctl) {
 316                 spin_unlock(&cache->lock);
 317                 return NULL;
 318         }
 319
 320         ctl = cache->caching_ctl;
 321         refcount_inc(&ctl->count);
 322         spin_unlock(&cache->lock);
 323         return ctl;
 324 }
 325
 326 static void put_caching_control(struct btrfs_caching_control *ctl)
 327 {
 328         if (refcount_dec_and_test(&ctl->count))
 329                 kfree(ctl);
 330 }
 331
 332 #ifdef CONFIG_BTRFS_DEBUG
 333 static void fragment_free_space(struct btrfs_block_group_cache *block_group)
 334 {
 335         struct btrfs_fs_info *fs_info = block_group->fs_info;
 336         u64 start = block_group->key.objectid;
 337         u64 len = block_group->key.offset;
 338         u64 chunk = block_group->flags & BTRFS_BLOCK_GROUP_METADATA ?
 339                 fs_info->nodesize : fs_info->sectorsize;
 340         u64 step = chunk << 1;
 341
 342         while (len > chunk) {
 343                 btrfs_remove_free_space(block_group, start, chunk);
 344                 start += step;
 345                 if (len < step)
 346                         len = 0;
 347                 else
 348                         len -= step;
 349         }
 350 }
 351 #endif
 352
 353 /*
 354  * this is only called by cache_block_group, since we could have freed extents
 355  * we need to check the pinned_extents for any extents that can't be used yet
 356  * since their free space will be released as soon as the transaction commits.
 357  */
 358 u64 add_new_free_space(struct btrfs_block_group_cache *block_group,
 359                        u64 start, u64 end)
 360 {
 361         struct btrfs_fs_info *info = block_group->fs_info;
 362         u64 extent_start, extent_end, size, total_added = 0;
 363         int ret;
 364
 365         while (start < end) {
 366                 ret = find_first_extent_bit(info->pinned_extents, start,
 367                                             &extent_start, &extent_end,
 368                                             EXTENT_DIRTY | EXTENT_UPTODATE,
 369                                             NULL);
 370                 if (ret)
 371                         break;
 372
 373                 if (extent_start <= start) {
 374                         start = extent_end + 1;
 375                 } else if (extent_start > start && extent_start < end) {
 376                         size = extent_start - start;
 377                         total_added += size;
 378                         ret = btrfs_add_free_space(block_group, start,
 379                                                    size);
 380                         BUG_ON(ret); /* -ENOMEM or logic error */
 381                         start = extent_end + 1;
 382                 } else {
 383                         break;
 384                 }
 385         }
 386
 387         if (start < end) {
 388                 size = end - start;
 389                 total_added += size;
 390                 ret = btrfs_add_free_space(block_group, start, size);
 391                 BUG_ON(ret); /* -ENOMEM or logic error */
 392         }
 393
 394         return total_added;
 395 }
 396
 397 static int load_extent_tree_free(struct btrfs_caching_control *caching_ctl)
 398 {
 399         struct btrfs_block_group_cache *block_group = caching_ctl->block_group;
 400         struct btrfs_fs_info *fs_info = block_group->fs_info;
 401         struct btrfs_root *extent_root = fs_info->extent_root;
 402         struct btrfs_path *path;
 403         struct extent_buffer *leaf;
 404         struct btrfs_key key;
 405         u64 total_found = 0;
 406         u64 last = 0;
 407         u32 nritems;
 408         int ret;
 409         bool wakeup = true;
 410
 411         path = btrfs_alloc_path();
 412         if (!path)
 413                 return -ENOMEM;
 414
 415         last = max_t(u64, block_group->key.objectid, BTRFS_SUPER_INFO_OFFSET);
 416
 417 #ifdef CONFIG_BTRFS_DEBUG
 418         /*
 419          * If we're fragmenting we don't want to make anybody think we can
 420          * allocate from this block group until we've had a chance to fragment
 421          * the free space.
 422          */
 423         if (btrfs_should_fragment_free_space(block_group))
 424                 wakeup = false;
 425 #endif
 426         /*
 427          * We don't want to deadlock with somebody trying to allocate a new
 428          * extent for the extent root while also trying to search the extent
 429          * root to add free space.  So we skip locking and search the commit
 430          * root, since its read-only
 431          */
 432         path->skip_locking = 1;
 433         path->search_commit_root = 1;
 434         path->reada = READA_FORWARD;
 435
 436         key.objectid = last;
 437         key.offset = 0;
 438         key.type = BTRFS_EXTENT_ITEM_KEY;
 439
 440 next:
 441         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
 442         if (ret < 0)
 443                 goto out;
 444
 445         leaf = path->nodes[0];
 446         nritems = btrfs_header_nritems(leaf);
 447
 448         while (1) {
 449                 if (btrfs_fs_closing(fs_info) > 1) {
 450                         last = (u64)-1;
 451                         break;
 452                 }
 453
 454                 if (path->slots[0] < nritems) {
 455                         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
 456                 } else {
 457                         ret = find_next_key(path, 0, &key);
 458                         if (ret)
 459                                 break;
 460
 461                         if (need_resched() ||
 462                             rwsem_is_contended(&fs_info->commit_root_sem)) {
 463                                 if (wakeup)
 464                                         caching_ctl->progress = last;
 465                                 btrfs_release_path(path);
 466                                 up_read(&fs_info->commit_root_sem);
 467                                 mutex_unlock(&caching_ctl->mutex);
 468                                 cond_resched();
 469                                 mutex_lock(&caching_ctl->mutex);
 470                                 down_read(&fs_info->commit_root_sem);
 471                                 goto next;
 472                         }
 473
 474                         ret = btrfs_next_leaf(extent_root, path);
 475                         if (ret < 0)
 476                                 goto out;
 477                         if (ret)
 478                                 break;
 479                         leaf = path->nodes[0];
 480                         nritems = btrfs_header_nritems(leaf);
 481                         continue;
 482                 }
 483
 484                 if (key.objectid < last) {
 485                         key.objectid = last;
 486                         key.offset = 0;
 487                         key.type = BTRFS_EXTENT_ITEM_KEY;
 488
 489                         if (wakeup)
 490                                 caching_ctl->progress = last;
 491                         btrfs_release_path(path);
 492                         goto next;
 493                 }
 494
 495                 if (key.objectid < block_group->key.objectid) {
 496                         path->slots[0]++;
 497                         continue;
 498                 }
 499
 500                 if (key.objectid >= block_group->key.objectid +
 501                     block_group->key.offset)
 502                         break;
 503
 504                 if (key.type == BTRFS_EXTENT_ITEM_KEY ||
 505                     key.type == BTRFS_METADATA_ITEM_KEY) {
 506                         total_found += add_new_free_space(block_group, last,
 507                                                           key.objectid);
 508                         if (key.type == BTRFS_METADATA_ITEM_KEY)
 509                                 last = key.objectid +
 510                                         fs_info->nodesize;
 511                         else
 512                                 last = key.objectid + key.offset;
 513
 514                         if (total_found > CACHING_CTL_WAKE_UP) {
 515                                 total_found = 0;
 516                                 if (wakeup)
 517                                         wake_up(&caching_ctl->wait);
 518                         }
 519                 }
 520                 path->slots[0]++;
 521         }
 522         ret = 0;
 523
 524         total_found += add_new_free_space(block_group, last,
 525                                           block_group->key.objectid +
 526                                           block_group->key.offset);
 527         caching_ctl->progress = (u64)-1;
 528
 529 out:
 530         btrfs_free_path(path);
 531         return ret;
 532 }
 533
 534 static noinline void caching_thread(struct btrfs_work *work)
 535 {
 536         struct btrfs_block_group_cache *block_group;
 537         struct btrfs_fs_info *fs_info;
 538         struct btrfs_caching_control *caching_ctl;
 539         int ret;
 540
 541         caching_ctl = container_of(work, struct btrfs_caching_control, work);
 542         block_group = caching_ctl->block_group;
 543         fs_info = block_group->fs_info;
 544
 545         mutex_lock(&caching_ctl->mutex);
 546         down_read(&fs_info->commit_root_sem);
 547
 548         if (btrfs_fs_compat_ro(fs_info, FREE_SPACE_TREE))
 549                 ret = load_free_space_tree(caching_ctl);
 550         else
 551                 ret = load_extent_tree_free(caching_ctl);
 552
 553         spin_lock(&block_group->lock);
 554         block_group->caching_ctl = NULL;
 555         block_group->cached = ret ? BTRFS_CACHE_ERROR : BTRFS_CACHE_FINISHED;
 556         spin_unlock(&block_group->lock);
 557
 558 #ifdef CONFIG_BTRFS_DEBUG
 559         if (btrfs_should_fragment_free_space(block_group)) {
 560                 u64 bytes_used;
 561
 562                 spin_lock(&block_group->space_info->lock);
 563                 spin_lock(&block_group->lock);
 564                 bytes_used = block_group->key.offset -
 565                         btrfs_block_group_used(&block_group->item);
 566                 block_group->space_info->bytes_used += bytes_used >> 1;
 567                 spin_unlock(&block_group->lock);
 568                 spin_unlock(&block_group->space_info->lock);
 569                 fragment_free_space(block_group);
 570         }
 571 #endif
 572
 573         caching_ctl->progress = (u64)-1;
 574
 575         up_read(&fs_info->commit_root_sem);
 576         free_excluded_extents(block_group);
 577         mutex_unlock(&caching_ctl->mutex);
 578
 579         wake_up(&caching_ctl->wait);
 580
 581         put_caching_control(caching_ctl);
 582         btrfs_put_block_group(block_group);
 583 }
 584
 585 static int cache_block_group(struct btrfs_block_group_cache *cache,
 586                              int load_cache_only)
 587 {
 588         DEFINE_WAIT(wait);
 589         struct btrfs_fs_info *fs_info = cache->fs_info;
 590         struct btrfs_caching_control *caching_ctl;
 591         int ret = 0;
 592
 593         caching_ctl = kzalloc(sizeof(*caching_ctl), GFP_NOFS);
 594         if (!caching_ctl)
 595                 return -ENOMEM;
 596
 597         INIT_LIST_HEAD(&caching_ctl->list);
 598         mutex_init(&caching_ctl->mutex);
 599         init_waitqueue_head(&caching_ctl->wait);
 600         caching_ctl->block_group = cache;
 601         caching_ctl->progress = cache->key.objectid;
 602         refcount_set(&caching_ctl->count, 1);
 603         btrfs_init_work(&caching_ctl->work, btrfs_cache_helper,
 604                         caching_thread, NULL, NULL);
 605
 606         spin_lock(&cache->lock);
 607         /*
 608          * This should be a rare occasion, but this could happen I think in the
 609          * case where one thread starts to load the space cache info, and then
 610          * some other thread starts a transaction commit which tries to do an
 611          * allocation while the other thread is still loading the space cache
 612          * info.  The previous loop should have kept us from choosing this block
 613          * group, but if we've moved to the state where we will wait on caching
 614          * block groups we need to first check if we're doing a fast load here,
 615          * so we can wait for it to finish, otherwise we could end up allocating
 616          * from a block group who's cache gets evicted for one reason or
 617          * another.
 618          */
 619         while (cache->cached == BTRFS_CACHE_FAST) {
 620                 struct btrfs_caching_control *ctl;
 621
 622                 ctl = cache->caching_ctl;
 623                 refcount_inc(&ctl->count);
 624                 prepare_to_wait(&ctl->wait, &wait, TASK_UNINTERRUPTIBLE);
 625                 spin_unlock(&cache->lock);
 626
 627                 schedule();
 628
 629                 finish_wait(&ctl->wait, &wait);
 630                 put_caching_control(ctl);
 631                 spin_lock(&cache->lock);
 632         }
 633
 634         if (cache->cached != BTRFS_CACHE_NO) {
 635                 spin_unlock(&cache->lock);
 636                 kfree(caching_ctl);
 637                 return 0;
 638         }
 639         WARN_ON(cache->caching_ctl);
 640         cache->caching_ctl = caching_ctl;
 641         cache->cached = BTRFS_CACHE_FAST;
 642         spin_unlock(&cache->lock);
 643
 644         if (btrfs_test_opt(fs_info, SPACE_CACHE)) {
 645                 mutex_lock(&caching_ctl->mutex);
 646                 ret = load_free_space_cache(cache);
 647
 648                 spin_lock(&cache->lock);
 649                 if (ret == 1) {
 650                         cache->caching_ctl = NULL;
 651                         cache->cached = BTRFS_CACHE_FINISHED;
 652                         cache->last_byte_to_unpin = (u64)-1;
 653                         caching_ctl->progress = (u64)-1;
 654                 } else {
 655                         if (load_cache_only) {
 656                                 cache->caching_ctl = NULL;
 657                                 cache->cached = BTRFS_CACHE_NO;
 658                         } else {
 659                                 cache->cached = BTRFS_CACHE_STARTED;
 660                                 cache->has_caching_ctl = 1;
 661                         }
 662                 }
 663                 spin_unlock(&cache->lock);
 664 #ifdef CONFIG_BTRFS_DEBUG
 665                 if (ret == 1 &&
 666                     btrfs_should_fragment_free_space(cache)) {
 667                         u64 bytes_used;
 668
 669                         spin_lock(&cache->space_info->lock);
 670                         spin_lock(&cache->lock);
 671                         bytes_used = cache->key.offset -
 672                                 btrfs_block_group_used(&cache->item);
 673                         cache->space_info->bytes_used += bytes_used >> 1;
 674                         spin_unlock(&cache->lock);
 675                         spin_unlock(&cache->space_info->lock);
 676                         fragment_free_space(cache);
 677                 }
 678 #endif
 679                 mutex_unlock(&caching_ctl->mutex);
 680
 681                 wake_up(&caching_ctl->wait);
 682                 if (ret == 1) {
 683                         put_caching_control(caching_ctl);
 684                         free_excluded_extents(cache);
 685                         return 0;
 686                 }
 687         } else {
 688                 /*
 689                  * We're either using the free space tree or no caching at all.
 690                  * Set cached to the appropriate value and wakeup any waiters.
 691                  */
 692                 spin_lock(&cache->lock);
 693                 if (load_cache_only) {
 694                         cache->caching_ctl = NULL;
 695                         cache->cached = BTRFS_CACHE_NO;
 696                 } else {
 697                         cache->cached = BTRFS_CACHE_STARTED;
 698                         cache->has_caching_ctl = 1;
 699                 }
 700                 spin_unlock(&cache->lock);
 701                 wake_up(&caching_ctl->wait);
 702         }
 703
 704         if (load_cache_only) {
 705                 put_caching_control(caching_ctl);
 706                 return 0;
 707         }
 708
 709         down_write(&fs_info->commit_root_sem);
 710         refcount_inc(&caching_ctl->count);
 711         list_add_tail(&caching_ctl->list, &fs_info->caching_block_groups);
 712         up_write(&fs_info->commit_root_sem);
 713
 714         btrfs_get_block_group(cache);
 715
 716         btrfs_queue_work(fs_info->caching_workers, &caching_ctl->work);
 717
 718         return ret;
 719 }
 720
 721 /*
 722  * return the block group that starts at or after bytenr
 723  */
 724 static struct btrfs_block_group_cache *
 725 btrfs_lookup_first_block_group(struct btrfs_fs_info *info, u64 bytenr)
 726 {
 727         return block_group_cache_tree_search(info, bytenr, 0);
 728 }
 729
 730 /*
 731  * return the block group that contains the given bytenr
 732  */
 733 struct btrfs_block_group_cache *btrfs_lookup_block_group(
 734                                                  struct btrfs_fs_info *info,
 735                                                  u64 bytenr)
 736 {
 737         return block_group_cache_tree_search(info, bytenr, 1);
 738 }
 739
 740 static struct btrfs_space_info *__find_space_info(struct btrfs_fs_info *info,
 741                                                   u64 flags)
 742 {
 743         struct list_head *head = &info->space_info;
 744         struct btrfs_space_info *found;
 745
 746         flags &= BTRFS_BLOCK_GROUP_TYPE_MASK;
 747
 748         rcu_read_lock();
 749         list_for_each_entry_rcu(found, head, list) {
 750                 if (found->flags & flags) {
 751                         rcu_read_unlock();
 752                         return found;
 753                 }
 754         }
 755         rcu_read_unlock();
 756         return NULL;
 757 }
 758
 759 static void add_pinned_bytes(struct btrfs_fs_info *fs_info,
 760                              struct btrfs_ref *ref)
 761 {
 762         struct btrfs_space_info *space_info;
 763         s64 num_bytes = -ref->len;
 764         u64 flags;
 765
 766         if (ref->type == BTRFS_REF_METADATA) {
 767                 if (ref->tree_ref.root == BTRFS_CHUNK_TREE_OBJECTID)
 768                         flags = BTRFS_BLOCK_GROUP_SYSTEM;
 769                 else
 770                         flags = BTRFS_BLOCK_GROUP_METADATA;
 771         } else {
 772                 flags = BTRFS_BLOCK_GROUP_DATA;
 773         }
 774
 775         space_info = __find_space_info(fs_info, flags);
 776         ASSERT(space_info);
 777         percpu_counter_add_batch(&space_info->total_bytes_pinned, num_bytes,
 778                     BTRFS_TOTAL_BYTES_PINNED_BATCH);
 779 }
 780
 781 /*
 782  * after adding space to the filesystem, we need to clear the full flags
 783  * on all the space infos.
 784  */
 785 void btrfs_clear_space_info_full(struct btrfs_fs_info *info)
 786 {
 787         struct list_head *head = &info->space_info;
 788         struct btrfs_space_info *found;
 789
 790         rcu_read_lock();
 791         list_for_each_entry_rcu(found, head, list)
 792                 found->full = 0;
 793         rcu_read_unlock();
 794 }
 795
 796 /* simple helper to search for an existing data extent at a given offset */
 797 int btrfs_lookup_data_extent(struct btrfs_fs_info *fs_info, u64 start, u64 len)
 798 {
 799         int ret;
 800         struct btrfs_key key;
 801         struct btrfs_path *path;
 802
 803         path = btrfs_alloc_path();
 804         if (!path)
 805                 return -ENOMEM;
 806
 807         key.objectid = start;
 808         key.offset = len;
 809         key.type = BTRFS_EXTENT_ITEM_KEY;
 810         ret = btrfs_search_slot(NULL, fs_info->extent_root, &key, path, 0, 0);
 811         btrfs_free_path(path);
 812         return ret;
 813 }
 814
 815 /*
 816  * helper function to lookup reference count and flags of a tree block.
 817  *
 818  * the head node for delayed ref is used to store the sum of all the
 819  * reference count modifications queued up in the rbtree. the head
 820  * node may also store the extent flags to set. This way you can check
 821  * to see what the reference count and extent flags would be if all of
 822  * the delayed refs are not processed.
 823  */
 824 int btrfs_lookup_extent_info(struct btrfs_trans_handle *trans,
 825                              struct btrfs_fs_info *fs_info, u64 bytenr,
 826                              u64 offset, int metadata, u64 *refs, u64 *flags)
 827 {
 828         struct btrfs_delayed_ref_head *head;
 829         struct btrfs_delayed_ref_root *delayed_refs;
 830         struct btrfs_path *path;
 831         struct btrfs_extent_item *ei;
 832         struct extent_buffer *leaf;
 833         struct btrfs_key key;
 834         u32 item_size;
 835         u64 num_refs;
 836         u64 extent_flags;
 837         int ret;
 838
 839         /*
 840          * If we don't have skinny metadata, don't bother doing anything
 841          * different
 842          */
 843         if (metadata && !btrfs_fs_incompat(fs_info, SKINNY_METADATA)) {
 844                 offset = fs_info->nodesize;
 845                 metadata = 0;
 846         }
 847
 848         path = btrfs_alloc_path();
 849         if (!path)
 850                 return -ENOMEM;
 851
 852         if (!trans) {
 853                 path->skip_locking = 1;
 854                 path->search_commit_root = 1;
 855         }
 856
 857 search_again:
 858         key.objectid = bytenr;
 859         key.offset = offset;
 860         if (metadata)
 861                 key.type = BTRFS_METADATA_ITEM_KEY;
 862         else
 863                 key.type = BTRFS_EXTENT_ITEM_KEY;
 864
 865         ret = btrfs_search_slot(trans, fs_info->extent_root, &key, path, 0, 0);
 866         if (ret < 0)
 867                 goto out_free;
 868
 869         if (ret > 0 && metadata && key.type == BTRFS_METADATA_ITEM_KEY) {
 870                 if (path->slots[0]) {
 871                         path->slots[0]--;
 872                         btrfs_item_key_to_cpu(path->nodes[0], &key,
 873                                               path->slots[0]);
 874                         if (key.objectid == bytenr &&
 875                             key.type == BTRFS_EXTENT_ITEM_KEY &&
 876                             key.offset == fs_info->nodesize)
 877                                 ret = 0;
 878                 }
 879         }
 880
 881         if (ret == 0) {
 882                 leaf = path->nodes[0];
 883                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
 884                 if (item_size >= sizeof(*ei)) {
 885                         ei = btrfs_item_ptr(leaf, path->slots[0],
 886                                             struct btrfs_extent_item);
 887                         num_refs = btrfs_extent_refs(leaf, ei);
 888                         extent_flags = btrfs_extent_flags(leaf, ei);
 889                 } else {
 890                         ret = -EINVAL;
 891                         btrfs_print_v0_err(fs_info);
 892                         if (trans)
 893                                 btrfs_abort_transaction(trans, ret);
 894                         else
 895                                 btrfs_handle_fs_error(fs_info, ret, NULL);
 896
 897                         goto out_free;
 898                 }
 899
 900                 BUG_ON(num_refs == 0);
 901         } else {
 902                 num_refs = 0;
 903                 extent_flags = 0;
 904                 ret = 0;
 905         }
 906
 907         if (!trans)
 908                 goto out;
 909
 910         delayed_refs = &trans->transaction->delayed_refs;
 911         spin_lock(&delayed_refs->lock);
 912         head = btrfs_find_delayed_ref_head(delayed_refs, bytenr);
 913         if (head) {
 914                 if (!mutex_trylock(&head->mutex)) {
 915                         refcount_inc(&head->refs);
 916                         spin_unlock(&delayed_refs->lock);
 917
 918                         btrfs_release_path(path);
 919
 920                         /*
 921                          * Mutex was contended, block until it's released and try
 922                          * again
 923                          */
 924                         mutex_lock(&head->mutex);
 925                         mutex_unlock(&head->mutex);
 926                         btrfs_put_delayed_ref_head(head);
 927                         goto search_again;
 928                 }
 929                 spin_lock(&head->lock);
 930                 if (head->extent_op && head->extent_op->update_flags)
 931                         extent_flags |= head->extent_op->flags_to_set;
 932                 else
 933                         BUG_ON(num_refs == 0);
 934
 935                 num_refs += head->ref_mod;
 936                 spin_unlock(&head->lock);
 937                 mutex_unlock(&head->mutex);
 938         }
 939         spin_unlock(&delayed_refs->lock);
 940 out:
 941         WARN_ON(num_refs == 0);
 942         if (refs)
 943                 *refs = num_refs;
 944         if (flags)
 945                 *flags = extent_flags;
 946 out_free:
 947         btrfs_free_path(path);
 948         return ret;
 949 }
 950
 951 /*
 952  * Back reference rules.  Back refs have three main goals:
 953  *
 954  * 1) differentiate between all holders of references to an extent so that
 955  *    when a reference is dropped we can make sure it was a valid reference
 956  *    before freeing the extent.
 957  *
 958  * 2) Provide enough information to quickly find the holders of an extent
 959  *    if we notice a given block is corrupted or bad.
 960  *
 961  * 3) Make it easy to migrate blocks for FS shrinking or storage pool
 962  *    maintenance.  This is actually the same as #2, but with a slightly
 963  *    different use case.
 964  *
 965  * There are two kinds of back refs. The implicit back refs is optimized
 966  * for pointers in non-shared tree blocks. For a given pointer in a block,
 967  * back refs of this kind provide information about the block's owner tree
 968  * and the pointer's key. These information allow us to find the block by
 969  * b-tree searching. The full back refs is for pointers in tree blocks not
 970  * referenced by their owner trees. The location of tree block is recorded
 971  * in the back refs. Actually the full back refs is generic, and can be
 972  * used in all cases the implicit back refs is used. The major shortcoming
 973  * of the full back refs is its overhead. Every time a tree block gets
 974  * COWed, we have to update back refs entry for all pointers in it.
 975  *
 976  * For a newly allocated tree block, we use implicit back refs for
 977  * pointers in it. This means most tree related operations only involve
 978  * implicit back refs. For a tree block created in old transaction, the
 979  * only way to drop a reference to it is COW it. So we can detect the
 980  * event that tree block loses its owner tree's reference and do the
 981  * back refs conversion.
 982  *
 983  * When a tree block is COWed through a tree, there are four cases:
 984  *
 985  * The reference count of the block is one and the tree is the block's
 986  * owner tree. Nothing to do in this case.
 987  *
 988  * The reference count of the block is one and the tree is not the
 989  * block's owner tree. In this case, full back refs is used for pointers
 990  * in the block. Remove these full back refs, add implicit back refs for
 991  * every pointers in the new block.
 992  *
 993  * The reference count of the block is greater than one and the tree is
 994  * the block's owner tree. In this case, implicit back refs is used for
 995  * pointers in the block. Add full back refs for every pointers in the
 996  * block, increase lower level extents' reference counts. The original
 997  * implicit back refs are entailed to the new block.
 998  *
 999  * The reference count of the block is greater than one and the tree is
1000  * not the block's owner tree. Add implicit back refs for every pointer in
1001  * the new block, increase lower level extents' reference count.
1002  *
1003  * Back Reference Key composing:
1004  *
1005  * The key objectid corresponds to the first byte in the extent,
1006  * The key type is used to differentiate between types of back refs.
1007  * There are different meanings of the key offset for different types
1008  * of back refs.
1009  *
1010  * File extents can be referenced by:
1011  *
1012  * - multiple snapshots, subvolumes, or different generations in one subvol
1013  * - different files inside a single subvolume
1014  * - different offsets inside a file (bookend extents in file.c)
1015  *
1016  * The extent ref structure for the implicit back refs has fields for:
1017  *
1018  * - Objectid of the subvolume root
1019  * - objectid of the file holding the reference
1020  * - original offset in the file
1021  * - how many bookend extents
1022  *
1023  * The key offset for the implicit back refs is hash of the first
1024  * three fields.
1025  *
1026  * The extent ref structure for the full back refs has field for:
1027  *
1028  * - number of pointers in the tree leaf
1029  *
1030  * The key offset for the implicit back refs is the first byte of
1031  * the tree leaf
1032  *
1033  * When a file extent is allocated, The implicit back refs is used.
1034  * the fields are filled in:
1035  *
1036  *     (root_key.objectid, inode objectid, offset in file, 1)
1037  *
1038  * When a file extent is removed file truncation, we find the
1039  * corresponding implicit back refs and check the following fields:
1040  *
1041  *     (btrfs_header_owner(leaf), inode objectid, offset in file)
1042  *
1043  * Btree extents can be referenced by:
1044  *
1045  * - Different subvolumes
1046  *
1047  * Both the implicit back refs and the full back refs for tree blocks
1048  * only consist of key. The key offset for the implicit back refs is
1049  * objectid of block's owner tree. The key offset for the full back refs
1050  * is the first byte of parent block.
1051  *
1052  * When implicit back refs is used, information about the lowest key and
1053  * level of the tree block are required. These information are stored in
1054  * tree block info structure.
1055  */
1056
1057 /*
1058  * is_data == BTRFS_REF_TYPE_BLOCK, tree block type is required,
1059  * is_data == BTRFS_REF_TYPE_DATA, data type is requiried,
1060  * is_data == BTRFS_REF_TYPE_ANY, either type is OK.
1061  */
1062 int btrfs_get_extent_inline_ref_type(const struct extent_buffer *eb,
1063                                      struct btrfs_extent_inline_ref *iref,
1064                                      enum btrfs_inline_ref_type is_data)
1065 {
1066         int type = btrfs_extent_inline_ref_type(eb, iref);
1067         u64 offset = btrfs_extent_inline_ref_offset(eb, iref);
1068
1069         if (type == BTRFS_TREE_BLOCK_REF_KEY ||
1070             type == BTRFS_SHARED_BLOCK_REF_KEY ||
1071             type == BTRFS_SHARED_DATA_REF_KEY ||
1072             type == BTRFS_EXTENT_DATA_REF_KEY) {
1073                 if (is_data == BTRFS_REF_TYPE_BLOCK) {
1074                         if (type == BTRFS_TREE_BLOCK_REF_KEY)
1075                                 return type;
1076                         if (type == BTRFS_SHARED_BLOCK_REF_KEY) {
1077                                 ASSERT(eb->fs_info);
1078                                 /*
1079                                  * Every shared one has parent tree
1080                                  * block, which must be aligned to
1081                                  * nodesize.
1082                                  */
1083                                 if (offset &&
1084                                     IS_ALIGNED(offset, eb->fs_info->nodesize))
1085                                         return type;
1086                         }
1087                 } else if (is_data == BTRFS_REF_TYPE_DATA) {
1088                         if (type == BTRFS_EXTENT_DATA_REF_KEY)
1089                                 return type;
1090                         if (type == BTRFS_SHARED_DATA_REF_KEY) {
1091                                 ASSERT(eb->fs_info);
1092                                 /*
1093                                  * Every shared one has parent tree
1094                                  * block, which must be aligned to
1095                                  * nodesize.
1096                                  */
1097                                 if (offset &&
1098                                     IS_ALIGNED(offset, eb->fs_info->nodesize))
1099                                         return type;
1100                         }
1101                 } else {
1102                         ASSERT(is_data == BTRFS_REF_TYPE_ANY);
1103                         return type;
1104                 }
1105         }
1106
1107         btrfs_print_leaf((struct extent_buffer *)eb);
1108         btrfs_err(eb->fs_info, "eb %llu invalid extent inline ref type %d",
1109                   eb->start, type);
1110         WARN_ON(1);
1111
1112         return BTRFS_REF_TYPE_INVALID;
1113 }
1114
1115 static u64 hash_extent_data_ref(u64 root_objectid, u64 owner, u64 offset)
1116 {
1117         u32 high_crc = ~(u32)0;
1118         u32 low_crc = ~(u32)0;
1119         __le64 lenum;
1120
1121         lenum = cpu_to_le64(root_objectid);
1122         high_crc = crc32c(high_crc, &lenum, sizeof(lenum));
1123         lenum = cpu_to_le64(owner);
1124         low_crc = crc32c(low_crc, &lenum, sizeof(lenum));
1125         lenum = cpu_to_le64(offset);
1126         low_crc = crc32c(low_crc, &lenum, sizeof(lenum));
1127
1128         return ((u64)high_crc << 31) ^ (u64)low_crc;
1129 }
1130
1131 static u64 hash_extent_data_ref_item(struct extent_buffer *leaf,
1132                                      struct btrfs_extent_data_ref *ref)
1133 {
1134         return hash_extent_data_ref(btrfs_extent_data_ref_root(leaf, ref),
1135                                     btrfs_extent_data_ref_objectid(leaf, ref),
1136                                     btrfs_extent_data_ref_offset(leaf, ref));
1137 }
1138
1139 static int match_extent_data_ref(struct extent_buffer *leaf,
1140                                  struct btrfs_extent_data_ref *ref,
1141                                  u64 root_objectid, u64 owner, u64 offset)
1142 {
1143         if (btrfs_extent_data_ref_root(leaf, ref) != root_objectid ||
1144             btrfs_extent_data_ref_objectid(leaf, ref) != owner ||
1145             btrfs_extent_data_ref_offset(leaf, ref) != offset)
1146                 return 0;
1147         return 1;
1148 }
1149
1150 static noinline int lookup_extent_data_ref(struct btrfs_trans_handle *trans,
1151                                            struct btrfs_path *path,
1152                                            u64 bytenr, u64 parent,
1153                                            u64 root_objectid,
1154                                            u64 owner, u64 offset)
1155 {
1156         struct btrfs_root *root = trans->fs_info->extent_root;
1157         struct btrfs_key key;
1158         struct btrfs_extent_data_ref *ref;
1159         struct extent_buffer *leaf;
1160         u32 nritems;
1161         int ret;
1162         int recow;
1163         int err = -ENOENT;
1164
1165         key.objectid = bytenr;
1166         if (parent) {
1167                 key.type = BTRFS_SHARED_DATA_REF_KEY;
1168                 key.offset = parent;
1169         } else {
1170                 key.type = BTRFS_EXTENT_DATA_REF_KEY;
1171                 key.offset = hash_extent_data_ref(root_objectid,
1172                                                   owner, offset);
1173         }
1174 again:
1175         recow = 0;
1176         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1177         if (ret < 0) {
1178                 err = ret;
1179                 goto fail;
1180         }
1181
1182         if (parent) {
1183                 if (!ret)
1184                         return 0;
1185                 goto fail;
1186         }
1187
1188         leaf = path->nodes[0];
1189         nritems = btrfs_header_nritems(leaf);
1190         while (1) {
1191                 if (path->slots[0] >= nritems) {
1192                         ret = btrfs_next_leaf(root, path);
1193                         if (ret < 0)
1194                                 err = ret;
1195                         if (ret)
1196                                 goto fail;
1197
1198                         leaf = path->nodes[0];
1199                         nritems = btrfs_header_nritems(leaf);
1200                         recow = 1;
1201                 }
1202
1203                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1204                 if (key.objectid != bytenr ||
1205                     key.type != BTRFS_EXTENT_DATA_REF_KEY)
1206                         goto fail;
1207
1208                 ref = btrfs_item_ptr(leaf, path->slots[0],
1209                                      struct btrfs_extent_data_ref);
1210
1211                 if (match_extent_data_ref(leaf, ref, root_objectid,
1212                                           owner, offset)) {
1213                         if (recow) {
1214                                 btrfs_release_path(path);
1215                                 goto again;
1216                         }
1217                         err = 0;
1218                         break;
1219                 }
1220                 path->slots[0]++;
1221         }
1222 fail:
1223         return err;
1224 }
1225
1226 static noinline int insert_extent_data_ref(struct btrfs_trans_handle *trans,
1227                                            struct btrfs_path *path,
1228                                            u64 bytenr, u64 parent,
1229                                            u64 root_objectid, u64 owner,
1230                                            u64 offset, int refs_to_add)
1231 {
1232         struct btrfs_root *root = trans->fs_info->extent_root;
1233         struct btrfs_key key;
1234         struct extent_buffer *leaf;
1235         u32 size;
1236         u32 num_refs;
1237         int ret;
1238
1239         key.objectid = bytenr;
1240         if (parent) {
1241                 key.type = BTRFS_SHARED_DATA_REF_KEY;
1242                 key.offset = parent;
1243                 size = sizeof(struct btrfs_shared_data_ref);
1244         } else {
1245                 key.type = BTRFS_EXTENT_DATA_REF_KEY;
1246                 key.offset = hash_extent_data_ref(root_objectid,
1247                                                   owner, offset);
1248                 size = sizeof(struct btrfs_extent_data_ref);
1249         }
1250
1251         ret = btrfs_insert_empty_item(trans, root, path, &key, size);
1252         if (ret && ret != -EEXIST)
1253                 goto fail;
1254
1255         leaf = path->nodes[0];
1256         if (parent) {
1257                 struct btrfs_shared_data_ref *ref;
1258                 ref = btrfs_item_ptr(leaf, path->slots[0],
1259                                      struct btrfs_shared_data_ref);
1260                 if (ret == 0) {
1261                         btrfs_set_shared_data_ref_count(leaf, ref, refs_to_add);
1262                 } else {
1263                         num_refs = btrfs_shared_data_ref_count(leaf, ref);
1264                         num_refs += refs_to_add;
1265                         btrfs_set_shared_data_ref_count(leaf, ref, num_refs);
1266                 }
1267         } else {
1268                 struct btrfs_extent_data_ref *ref;
1269                 while (ret == -EEXIST) {
1270                         ref = btrfs_item_ptr(leaf, path->slots[0],
1271                                              struct btrfs_extent_data_ref);
1272                         if (match_extent_data_ref(leaf, ref, root_objectid,
1273                                                   owner, offset))
1274                                 break;
1275                         btrfs_release_path(path);
1276                         key.offset++;
1277                         ret = btrfs_insert_empty_item(trans, root, path, &key,
1278                                                       size);
1279                         if (ret && ret != -EEXIST)
1280                                 goto fail;
1281
1282                         leaf = path->nodes[0];
1283                 }
1284                 ref = btrfs_item_ptr(leaf, path->slots[0],
1285                                      struct btrfs_extent_data_ref);
1286                 if (ret == 0) {
1287                         btrfs_set_extent_data_ref_root(leaf, ref,
1288                                                        root_objectid);
1289                         btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
1290                         btrfs_set_extent_data_ref_offset(leaf, ref, offset);
1291                         btrfs_set_extent_data_ref_count(leaf, ref, refs_to_add);
1292                 } else {
1293                         num_refs = btrfs_extent_data_ref_count(leaf, ref);
1294                         num_refs += refs_to_add;
1295                         btrfs_set_extent_data_ref_count(leaf, ref, num_refs);
1296                 }
1297         }
1298         btrfs_mark_buffer_dirty(leaf);
1299         ret = 0;
1300 fail:
1301         btrfs_release_path(path);
1302         return ret;
1303 }
1304
1305 static noinline int remove_extent_data_ref(struct btrfs_trans_handle *trans,
1306                                            struct btrfs_path *path,
1307                                            int refs_to_drop, int *last_ref)
1308 {
1309         struct btrfs_key key;
1310         struct btrfs_extent_data_ref *ref1 = NULL;
1311         struct btrfs_shared_data_ref *ref2 = NULL;
1312         struct extent_buffer *leaf;
1313         u32 num_refs = 0;
1314         int ret = 0;
1315
1316         leaf = path->nodes[0];
1317         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1318
1319         if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
1320                 ref1 = btrfs_item_ptr(leaf, path->slots[0],
1321                                       struct btrfs_extent_data_ref);
1322                 num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1323         } else if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
1324                 ref2 = btrfs_item_ptr(leaf, path->slots[0],
1325                                       struct btrfs_shared_data_ref);
1326                 num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1327         } else if (unlikely(key.type == BTRFS_EXTENT_REF_V0_KEY)) {
1328                 btrfs_print_v0_err(trans->fs_info);
1329                 btrfs_abort_transaction(trans, -EINVAL);
1330                 return -EINVAL;
1331         } else {
1332                 BUG();
1333         }
1334
1335         BUG_ON(num_refs < refs_to_drop);
1336         num_refs -= refs_to_drop;
1337
1338         if (num_refs == 0) {
1339                 ret = btrfs_del_item(trans, trans->fs_info->extent_root, path);
1340                 *last_ref = 1;
1341         } else {
1342                 if (key.type == BTRFS_EXTENT_DATA_REF_KEY)
1343                         btrfs_set_extent_data_ref_count(leaf, ref1, num_refs);
1344                 else if (key.type == BTRFS_SHARED_DATA_REF_KEY)
1345                         btrfs_set_shared_data_ref_count(leaf, ref2, num_refs);
1346                 btrfs_mark_buffer_dirty(leaf);
1347         }
1348         return ret;
1349 }
1350
1351 static noinline u32 extent_data_ref_count(struct btrfs_path *path,
1352                                           struct btrfs_extent_inline_ref *iref)
1353 {
1354         struct btrfs_key key;
1355         struct extent_buffer *leaf;
1356         struct btrfs_extent_data_ref *ref1;
1357         struct btrfs_shared_data_ref *ref2;
1358         u32 num_refs = 0;
1359         int type;
1360
1361         leaf = path->nodes[0];
1362         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1363
1364         BUG_ON(key.type == BTRFS_EXTENT_REF_V0_KEY);
1365         if (iref) {
1366                 /*
1367                  * If type is invalid, we should have bailed out earlier than
1368                  * this call.
1369                  */
1370                 type = btrfs_get_extent_inline_ref_type(leaf, iref, BTRFS_REF_TYPE_DATA);
1371                 ASSERT(type != BTRFS_REF_TYPE_INVALID);
1372                 if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1373                         ref1 = (struct btrfs_extent_data_ref *)(&iref->offset);
1374                         num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1375                 } else {
1376                         ref2 = (struct btrfs_shared_data_ref *)(iref + 1);
1377                         num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1378                 }
1379         } else if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
1380                 ref1 = btrfs_item_ptr(leaf, path->slots[0],
1381                                       struct btrfs_extent_data_ref);
1382                 num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1383         } else if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
1384                 ref2 = btrfs_item_ptr(leaf, path->slots[0],
1385                                       struct btrfs_shared_data_ref);
1386                 num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1387         } else {
1388                 WARN_ON(1);
1389         }
1390         return num_refs;
1391 }
1392
1393 static noinline int lookup_tree_block_ref(struct btrfs_trans_handle *trans,
1394                                           struct btrfs_path *path,
1395                                           u64 bytenr, u64 parent,
1396                                           u64 root_objectid)
1397 {
1398         struct btrfs_root *root = trans->fs_info->extent_root;
1399         struct btrfs_key key;
1400         int ret;
1401
1402         key.objectid = bytenr;
1403         if (parent) {
1404                 key.type = BTRFS_SHARED_BLOCK_REF_KEY;
1405                 key.offset = parent;
1406         } else {
1407                 key.type = BTRFS_TREE_BLOCK_REF_KEY;
1408                 key.offset = root_objectid;
1409         }
1410
1411         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1412         if (ret > 0)
1413                 ret = -ENOENT;
1414         return ret;
1415 }
1416
1417 static noinline int insert_tree_block_ref(struct btrfs_trans_handle *trans,
1418                                           struct btrfs_path *path,
1419                                           u64 bytenr, u64 parent,
1420                                           u64 root_objectid)
1421 {
1422         struct btrfs_key key;
1423         int ret;
1424
1425         key.objectid = bytenr;
1426         if (parent) {
1427                 key.type = BTRFS_SHARED_BLOCK_REF_KEY;
1428                 key.offset = parent;
1429         } else {
1430                 key.type = BTRFS_TREE_BLOCK_REF_KEY;
1431                 key.offset = root_objectid;
1432         }
1433
1434         ret = btrfs_insert_empty_item(trans, trans->fs_info->extent_root,
1435                                       path, &key, 0);
1436         btrfs_release_path(path);
1437         return ret;
1438 }
1439
1440 static inline int extent_ref_type(u64 parent, u64 owner)
1441 {
1442         int type;
1443         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1444                 if (parent > 0)
1445                         type = BTRFS_SHARED_BLOCK_REF_KEY;
1446                 else
1447                         type = BTRFS_TREE_BLOCK_REF_KEY;
1448         } else {
1449                 if (parent > 0)
1450                         type = BTRFS_SHARED_DATA_REF_KEY;
1451                 else
1452                         type = BTRFS_EXTENT_DATA_REF_KEY;
1453         }
1454         return type;
1455 }
1456
1457 static int find_next_key(struct btrfs_path *path, int level,
1458                          struct btrfs_key *key)
1459
1460 {
1461         for (; level < BTRFS_MAX_LEVEL; level++) {
1462                 if (!path->nodes[level])
1463                         break;
1464                 if (path->slots[level] + 1 >=
1465                     btrfs_header_nritems(path->nodes[level]))
1466                         continue;
1467                 if (level == 0)
1468                         btrfs_item_key_to_cpu(path->nodes[level], key,
1469                                               path->slots[level] + 1);
1470                 else
1471                         btrfs_node_key_to_cpu(path->nodes[level], key,
1472                                               path->slots[level] + 1);
1473                 return 0;
1474         }
1475         return 1;
1476 }
1477
1478 /*
1479  * look for inline back ref. if back ref is found, *ref_ret is set
1480  * to the address of inline back ref, and 0 is returned.
1481  *
1482  * if back ref isn't found, *ref_ret is set to the address where it
1483  * should be inserted, and -ENOENT is returned.
1484  *
1485  * if insert is true and there are too many inline back refs, the path
1486  * points to the extent item, and -EAGAIN is returned.
1487  *
1488  * NOTE: inline back refs are ordered in the same way that back ref
1489  *       items in the tree are ordered.
1490  */
1491 static noinline_for_stack
1492 int lookup_inline_extent_backref(struct btrfs_trans_handle *trans,
1493                                  struct btrfs_path *path,
1494                                  struct btrfs_extent_inline_ref **ref_ret,
1495                                  u64 bytenr, u64 num_bytes,
1496                                  u64 parent, u64 root_objectid,
1497                                  u64 owner, u64 offset, int insert)
1498 {
1499         struct btrfs_fs_info *fs_info = trans->fs_info;
1500         struct btrfs_root *root = fs_info->extent_root;
1501         struct btrfs_key key;
1502         struct extent_buffer *leaf;
1503         struct btrfs_extent_item *ei;
1504         struct btrfs_extent_inline_ref *iref;
1505         u64 flags;
1506         u64 item_size;
1507         unsigned long ptr;
1508         unsigned long end;
1509         int extra_size;
1510         int type;
1511         int want;
1512         int ret;
1513         int err = 0;
1514         bool skinny_metadata = btrfs_fs_incompat(fs_info, SKINNY_METADATA);
1515         int needed;
1516
1517         key.objectid = bytenr;
1518         key.type = BTRFS_EXTENT_ITEM_KEY;
1519         key.offset = num_bytes;
1520
1521         want = extent_ref_type(parent, owner);
1522         if (insert) {
1523                 extra_size = btrfs_extent_inline_ref_size(want);
1524                 path->keep_locks = 1;
1525         } else
1526                 extra_size = -1;
1527
1528         /*
1529          * Owner is our level, so we can just add one to get the level for the
1530          * block we are interested in.
1531          */
1532         if (skinny_metadata && owner < BTRFS_FIRST_FREE_OBJECTID) {
1533                 key.type = BTRFS_METADATA_ITEM_KEY;
1534                 key.offset = owner;
1535         }
1536
1537 again:
1538         ret = btrfs_search_slot(trans, root, &key, path, extra_size, 1);
1539         if (ret < 0) {
1540                 err = ret;
1541                 goto out;
1542         }
1543
1544         /*
1545          * We may be a newly converted file system which still has the old fat
1546          * extent entries for metadata, so try and see if we have one of those.
1547          */
1548         if (ret > 0 && skinny_metadata) {
1549                 skinny_metadata = false;
1550                 if (path->slots[0]) {
1551                         path->slots[0]--;
1552                         btrfs_item_key_to_cpu(path->nodes[0], &key,
1553                                               path->slots[0]);
1554                         if (key.objectid == bytenr &&
1555                             key.type == BTRFS_EXTENT_ITEM_KEY &&
1556                             key.offset == num_bytes)
1557                                 ret = 0;
1558                 }
1559                 if (ret) {
1560                         key.objectid = bytenr;
1561                         key.type = BTRFS_EXTENT_ITEM_KEY;
1562                         key.offset = num_bytes;
1563                         btrfs_release_path(path);
1564                         goto again;
1565                 }
1566         }
1567
1568         if (ret && !insert) {
1569                 err = -ENOENT;
1570                 goto out;
1571         } else if (WARN_ON(ret)) {
1572                 err = -EIO;
1573                 goto out;
1574         }
1575
1576         leaf = path->nodes[0];
1577         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1578         if (unlikely(item_size < sizeof(*ei))) {
1579                 err = -EINVAL;
1580                 btrfs_print_v0_err(fs_info);
1581                 btrfs_abort_transaction(trans, err);
1582                 goto out;
1583         }
1584
1585         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1586         flags = btrfs_extent_flags(leaf, ei);
1587
1588         ptr = (unsigned long)(ei + 1);
1589         end = (unsigned long)ei + item_size;
1590
1591         if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK && !skinny_metadata) {
1592                 ptr += sizeof(struct btrfs_tree_block_info);
1593                 BUG_ON(ptr > end);
1594         }
1595
1596         if (owner >= BTRFS_FIRST_FREE_OBJECTID)
1597                 needed = BTRFS_REF_TYPE_DATA;
1598         else
1599                 needed = BTRFS_REF_TYPE_BLOCK;
1600
1601         err = -ENOENT;
1602         while (1) {
1603                 if (ptr >= end) {
1604                         WARN_ON(ptr > end);
1605                         break;
1606                 }
1607                 iref = (struct btrfs_extent_inline_ref *)ptr;
1608                 type = btrfs_get_extent_inline_ref_type(leaf, iref, needed);
1609                 if (type == BTRFS_REF_TYPE_INVALID) {
1610                         err = -EUCLEAN;
1611                         goto out;
1612                 }
1613
1614                 if (want < type)
1615                         break;
1616                 if (want > type) {
1617                         ptr += btrfs_extent_inline_ref_size(type);
1618                         continue;
1619                 }
1620
1621                 if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1622                         struct btrfs_extent_data_ref *dref;
1623                         dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1624                         if (match_extent_data_ref(leaf, dref, root_objectid,
1625                                                   owner, offset)) {
1626                                 err = 0;
1627                                 break;
1628                         }
1629                         if (hash_extent_data_ref_item(leaf, dref) <
1630                             hash_extent_data_ref(root_objectid, owner, offset))
1631                                 break;
1632                 } else {
1633                         u64 ref_offset;
1634                         ref_offset = btrfs_extent_inline_ref_offset(leaf, iref);
1635                         if (parent > 0) {
1636                                 if (parent == ref_offset) {
1637                                         err = 0;
1638                                         break;
1639                                 }
1640                                 if (ref_offset < parent)
1641                                         break;
1642                         } else {
1643                                 if (root_objectid == ref_offset) {
1644                                         err = 0;
1645                                         break;
1646                                 }
1647                                 if (ref_offset < root_objectid)
1648                                         break;
1649                         }
1650                 }
1651                 ptr += btrfs_extent_inline_ref_size(type);
1652         }
1653         if (err == -ENOENT && insert) {
1654                 if (item_size + extra_size >=
1655                     BTRFS_MAX_EXTENT_ITEM_SIZE(root)) {
1656                         err = -EAGAIN;
1657                         goto out;
1658                 }
1659                 /*
1660                  * To add new inline back ref, we have to make sure
1661                  * there is no corresponding back ref item.
1662                  * For simplicity, we just do not add new inline back
1663                  * ref if there is any kind of item for this block
1664                  */
1665                 if (find_next_key(path, 0, &key) == 0 &&
1666                     key.objectid == bytenr &&
1667                     key.type < BTRFS_BLOCK_GROUP_ITEM_KEY) {
1668                         err = -EAGAIN;
1669                         goto out;
1670                 }
1671         }
1672         *ref_ret = (struct btrfs_extent_inline_ref *)ptr;
1673 out:
1674         if (insert) {
1675                 path->keep_locks = 0;
1676                 btrfs_unlock_up_safe(path, 1);
1677         }
1678         return err;
1679 }
1680
1681 /*
1682  * helper to add new inline back ref
1683  */
1684 static noinline_for_stack
1685 void setup_inline_extent_backref(struct btrfs_fs_info *fs_info,
1686                                  struct btrfs_path *path,
1687                                  struct btrfs_extent_inline_ref *iref,
1688                                  u64 parent, u64 root_objectid,
1689                                  u64 owner, u64 offset, int refs_to_add,
1690                                  struct btrfs_delayed_extent_op *extent_op)
1691 {
1692         struct extent_buffer *leaf;
1693         struct btrfs_extent_item *ei;
1694         unsigned long ptr;
1695         unsigned long end;
1696         unsigned long item_offset;
1697         u64 refs;
1698         int size;
1699         int type;
1700
1701         leaf = path->nodes[0];
1702         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1703         item_offset = (unsigned long)iref - (unsigned long)ei;
1704
1705         type = extent_ref_type(parent, owner);
1706         size = btrfs_extent_inline_ref_size(type);
1707
1708         btrfs_extend_item(path, size);
1709
1710         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1711         refs = btrfs_extent_refs(leaf, ei);
1712         refs += refs_to_add;
1713         btrfs_set_extent_refs(leaf, ei, refs);
1714         if (extent_op)
1715                 __run_delayed_extent_op(extent_op, leaf, ei);
1716
1717         ptr = (unsigned long)ei + item_offset;
1718         end = (unsigned long)ei + btrfs_item_size_nr(leaf, path->slots[0]);
1719         if (ptr < end - size)
1720                 memmove_extent_buffer(leaf, ptr + size, ptr,
1721                                       end - size - ptr);
1722
1723         iref = (struct btrfs_extent_inline_ref *)ptr;
1724         btrfs_set_extent_inline_ref_type(leaf, iref, type);
1725         if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1726                 struct btrfs_extent_data_ref *dref;
1727                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1728                 btrfs_set_extent_data_ref_root(leaf, dref, root_objectid);
1729                 btrfs_set_extent_data_ref_objectid(leaf, dref, owner);
1730                 btrfs_set_extent_data_ref_offset(leaf, dref, offset);
1731                 btrfs_set_extent_data_ref_count(leaf, dref, refs_to_add);
1732         } else if (type == BTRFS_SHARED_DATA_REF_KEY) {
1733                 struct btrfs_shared_data_ref *sref;
1734                 sref = (struct btrfs_shared_data_ref *)(iref + 1);
1735                 btrfs_set_shared_data_ref_count(leaf, sref, refs_to_add);
1736                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
1737         } else if (type == BTRFS_SHARED_BLOCK_REF_KEY) {
1738                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
1739         } else {
1740                 btrfs_set_extent_inline_ref_offset(leaf, iref, root_objectid);
1741         }
1742         btrfs_mark_buffer_dirty(leaf);
1743 }
1744
1745 static int lookup_extent_backref(struct btrfs_trans_handle *trans,
1746                                  struct btrfs_path *path,
1747                                  struct btrfs_extent_inline_ref **ref_ret,
1748                                  u64 bytenr, u64 num_bytes, u64 parent,
1749                                  u64 root_objectid, u64 owner, u64 offset)
1750 {
1751         int ret;
1752
1753         ret = lookup_inline_extent_backref(trans, path, ref_ret, bytenr,
1754                                            num_bytes, parent, root_objectid,
1755                                            owner, offset, 0);
1756         if (ret != -ENOENT)
1757                 return ret;
1758
1759         btrfs_release_path(path);
1760         *ref_ret = NULL;
1761
1762         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1763                 ret = lookup_tree_block_ref(trans, path, bytenr, parent,
1764                                             root_objectid);
1765         } else {
1766                 ret = lookup_extent_data_ref(trans, path, bytenr, parent,
1767                                              root_objectid, owner, offset);
1768         }
1769         return ret;
1770 }
1771
1772 /*
1773  * helper to update/remove inline back ref
1774  */
1775 static noinline_for_stack
1776 void update_inline_extent_backref(struct btrfs_path *path,
1777                                   struct btrfs_extent_inline_ref *iref,
1778                                   int refs_to_mod,
1779                                   struct btrfs_delayed_extent_op *extent_op,
1780                                   int *last_ref)
1781 {
1782         struct extent_buffer *leaf = path->nodes[0];
1783         struct btrfs_extent_item *ei;
1784         struct btrfs_extent_data_ref *dref = NULL;
1785         struct btrfs_shared_data_ref *sref = NULL;
1786         unsigned long ptr;
1787         unsigned long end;
1788         u32 item_size;
1789         int size;
1790         int type;
1791         u64 refs;
1792
1793         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1794         refs = btrfs_extent_refs(leaf, ei);
1795         WARN_ON(refs_to_mod < 0 && refs + refs_to_mod <= 0);
1796         refs += refs_to_mod;
1797         btrfs_set_extent_refs(leaf, ei, refs);
1798         if (extent_op)
1799                 __run_delayed_extent_op(extent_op, leaf, ei);
1800
1801         /*
1802          * If type is invalid, we should have bailed out after
1803          * lookup_inline_extent_backref().
1804          */
1805         type = btrfs_get_extent_inline_ref_type(leaf, iref, BTRFS_REF_TYPE_ANY);
1806         ASSERT(type != BTRFS_REF_TYPE_INVALID);
1807
1808         if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1809                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1810                 refs = btrfs_extent_data_ref_count(leaf, dref);
1811         } else if (type == BTRFS_SHARED_DATA_REF_KEY) {
1812                 sref = (struct btrfs_shared_data_ref *)(iref + 1);
1813                 refs = btrfs_shared_data_ref_count(leaf, sref);
1814         } else {
1815                 refs = 1;
1816                 BUG_ON(refs_to_mod != -1);
1817         }
1818
1819         BUG_ON(refs_to_mod < 0 && refs < -refs_to_mod);
1820         refs += refs_to_mod;
1821
1822         if (refs > 0) {
1823                 if (type == BTRFS_EXTENT_DATA_REF_KEY)
1824                         btrfs_set_extent_data_ref_count(leaf, dref, refs);
1825                 else
1826                         btrfs_set_shared_data_ref_count(leaf, sref, refs);
1827         } else {
1828                 *last_ref = 1;
1829                 size =  btrfs_extent_inline_ref_size(type);
1830                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1831                 ptr = (unsigned long)iref;
1832                 end = (unsigned long)ei + item_size;
1833                 if (ptr + size < end)
1834                         memmove_extent_buffer(leaf, ptr, ptr + size,
1835                                               end - ptr - size);
1836                 item_size -= size;
1837                 btrfs_truncate_item(path, item_size, 1);
1838         }
1839         btrfs_mark_buffer_dirty(leaf);
1840 }
1841
1842 static noinline_for_stack
1843 int insert_inline_extent_backref(struct btrfs_trans_handle *trans,
1844                                  struct btrfs_path *path,
1845                                  u64 bytenr, u64 num_bytes, u64 parent,
1846                                  u64 root_objectid, u64 owner,
1847                                  u64 offset, int refs_to_add,
1848                                  struct btrfs_delayed_extent_op *extent_op)
1849 {
1850         struct btrfs_extent_inline_ref *iref;
1851         int ret;
1852
1853         ret = lookup_inline_extent_backref(trans, path, &iref, bytenr,
1854                                            num_bytes, parent, root_objectid,
1855                                            owner, offset, 1);
1856         if (ret == 0) {
1857                 BUG_ON(owner < BTRFS_FIRST_FREE_OBJECTID);
1858                 update_inline_extent_backref(path, iref, refs_to_add,
1859                                              extent_op, NULL);
1860         } else if (ret == -ENOENT) {
1861                 setup_inline_extent_backref(trans->fs_info, path, iref, parent,
1862                                             root_objectid, owner, offset,
1863                                             refs_to_add, extent_op);
1864                 ret = 0;
1865         }
1866         return ret;
1867 }
1868
1869 static int insert_extent_backref(struct btrfs_trans_handle *trans,
1870                                  struct btrfs_path *path,
1871                                  u64 bytenr, u64 parent, u64 root_objectid,
1872                                  u64 owner, u64 offset, int refs_to_add)
1873 {
1874         int ret;
1875         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1876                 BUG_ON(refs_to_add != 1);
1877                 ret = insert_tree_block_ref(trans, path, bytenr, parent,
1878                                             root_objectid);
1879         } else {
1880                 ret = insert_extent_data_ref(trans, path, bytenr, parent,
1881                                              root_objectid, owner, offset,
1882                                              refs_to_add);
1883         }
1884         return ret;
1885 }
1886
1887 static int remove_extent_backref(struct btrfs_trans_handle *trans,
1888                                  struct btrfs_path *path,
1889                                  struct btrfs_extent_inline_ref *iref,
1890                                  int refs_to_drop, int is_data, int *last_ref)
1891 {
1892         int ret = 0;
1893
1894         BUG_ON(!is_data && refs_to_drop != 1);
1895         if (iref) {
1896                 update_inline_extent_backref(path, iref, -refs_to_drop, NULL,
1897                                              last_ref);
1898         } else if (is_data) {
1899                 ret = remove_extent_data_ref(trans, path, refs_to_drop,
1900                                              last_ref);
1901         } else {
1902                 *last_ref = 1;
1903                 ret = btrfs_del_item(trans, trans->fs_info->extent_root, path);
1904         }
1905         return ret;
1906 }
1907
1908 static int btrfs_issue_discard(struct block_device *bdev, u64 start, u64 len,
1909                                u64 *discarded_bytes)
1910 {
1911         int j, ret = 0;
1912         u64 bytes_left, end;
1913         u64 aligned_start = ALIGN(start, 1 << 9);
1914
1915         if (WARN_ON(start != aligned_start)) {
1916                 len -= aligned_start - start;
1917                 len = round_down(len, 1 << 9);
1918                 start = aligned_start;
1919         }
1920
1921         *discarded_bytes = 0;
1922
1923         if (!len)
1924                 return 0;
1925
1926         end = start + len;
1927         bytes_left = len;
1928
1929         /* Skip any superblocks on this device. */
1930         for (j = 0; j < BTRFS_SUPER_MIRROR_MAX; j++) {
1931                 u64 sb_start = btrfs_sb_offset(j);
1932                 u64 sb_end = sb_start + BTRFS_SUPER_INFO_SIZE;
1933                 u64 size = sb_start - start;
1934
1935                 if (!in_range(sb_start, start, bytes_left) &&
1936                     !in_range(sb_end, start, bytes_left) &&
1937                     !in_range(start, sb_start, BTRFS_SUPER_INFO_SIZE))
1938                         continue;
1939
1940                 /*
1941                  * Superblock spans beginning of range.  Adjust start and
1942                  * try again.
1943                  */
1944                 if (sb_start <= start) {
1945                         start += sb_end - start;
1946                         if (start > end) {
1947                                 bytes_left = 0;
1948                                 break;
1949                         }
1950                         bytes_left = end - start;
1951                         continue;
1952                 }
1953
1954                 if (size) {
1955                         ret = blkdev_issue_discard(bdev, start >> 9, size >> 9,
1956                                                    GFP_NOFS, 0);
1957                         if (!ret)
1958                                 *discarded_bytes += size;
1959                         else if (ret != -EOPNOTSUPP)
1960                                 return ret;
1961                 }
1962
1963                 start = sb_end;
1964                 if (start > end) {
1965                         bytes_left = 0;
1966                         break;
1967                 }
1968                 bytes_left = end - start;
1969         }
1970
1971         if (bytes_left) {
1972                 ret = blkdev_issue_discard(bdev, start >> 9, bytes_left >> 9,
1973                                            GFP_NOFS, 0);
1974                 if (!ret)
1975                         *discarded_bytes += bytes_left;
1976         }
1977         return ret;
1978 }
1979
1980 int btrfs_discard_extent(struct btrfs_fs_info *fs_info, u64 bytenr,
1981                          u64 num_bytes, u64 *actual_bytes)
1982 {
1983         int ret;
1984         u64 discarded_bytes = 0;
1985         struct btrfs_bio *bbio = NULL;
1986
1987
1988         /*
1989          * Avoid races with device replace and make sure our bbio has devices
1990          * associated to its stripes that don't go away while we are discarding.
1991          */
1992         btrfs_bio_counter_inc_blocked(fs_info);
1993         /* Tell the block device(s) that the sectors can be discarded */
1994         ret = btrfs_map_block(fs_info, BTRFS_MAP_DISCARD, bytenr, &num_bytes,
1995                               &bbio, 0);
1996         /* Error condition is -ENOMEM */
1997         if (!ret) {
1998                 struct btrfs_bio_stripe *stripe = bbio->stripes;
1999                 int i;
2000
2001
2002                 for (i = 0; i < bbio->num_stripes; i++, stripe++) {
2003                         u64 bytes;
2004                         struct request_queue *req_q;
2005
2006                         if (!stripe->dev->bdev) {
2007                                 ASSERT(btrfs_test_opt(fs_info, DEGRADED));
2008                                 continue;
2009                         }
2010                         req_q = bdev_get_queue(stripe->dev->bdev);
2011                         if (!blk_queue_discard(req_q))
2012                                 continue;
2013
2014                         ret = btrfs_issue_discard(stripe->dev->bdev,
2015                                                   stripe->physical,
2016                                                   stripe->length,
2017                                                   &bytes);
2018                         if (!ret)
2019                                 discarded_bytes += bytes;
2020                         else if (ret != -EOPNOTSUPP)
2021                                 break; /* Logic errors or -ENOMEM, or -EIO but I don't know how that could happen JDM */
2022
2023                         /*
2024                          * Just in case we get back EOPNOTSUPP for some reason,
2025                          * just ignore the return value so we don't screw up
2026                          * people calling discard_extent.
2027                          */
2028                         ret = 0;
2029                 }
2030                 btrfs_put_bbio(bbio);
2031         }
2032         btrfs_bio_counter_dec(fs_info);
2033
2034         if (actual_bytes)
2035                 *actual_bytes = discarded_bytes;
2036
2037
2038         if (ret == -EOPNOTSUPP)
2039                 ret = 0;
2040         return ret;
2041 }
2042
2043 /* Can return -ENOMEM */
2044 int btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
2045                          struct btrfs_ref *generic_ref)
2046 {
2047         struct btrfs_fs_info *fs_info = trans->fs_info;
2048         int old_ref_mod, new_ref_mod;
2049         int ret;
2050
2051         ASSERT(generic_ref->type != BTRFS_REF_NOT_SET &&
2052                generic_ref->action);
2053         BUG_ON(generic_ref->type == BTRFS_REF_METADATA &&
2054                generic_ref->tree_ref.root == BTRFS_TREE_LOG_OBJECTID);
2055
2056         if (generic_ref->type == BTRFS_REF_METADATA)
2057                 ret = btrfs_add_delayed_tree_ref(trans, generic_ref,
2058                                 NULL, &old_ref_mod, &new_ref_mod);
2059         else
2060                 ret = btrfs_add_delayed_data_ref(trans, generic_ref, 0,
2061                                                  &old_ref_mod, &new_ref_mod);
2062
2063         btrfs_ref_tree_mod(fs_info, generic_ref);
2064
2065         if (ret == 0 && old_ref_mod < 0 && new_ref_mod >= 0)
2066                 add_pinned_bytes(fs_info, generic_ref);
2067
2068         return ret;
2069 }
2070
2071 /*
2072  * __btrfs_inc_extent_ref - insert backreference for a given extent
2073  *
2074  * @trans:          Handle of transaction
2075  *
2076  * @node:           The delayed ref node used to get the bytenr/length for
2077  *                  extent whose references are incremented.
2078  *
2079  * @parent:         If this is a shared extent (BTRFS_SHARED_DATA_REF_KEY/
2080  *                  BTRFS_SHARED_BLOCK_REF_KEY) then it holds the logical
2081  *                  bytenr of the parent block. Since new extents are always
2082  *                  created with indirect references, this will only be the case
2083  *                  when relocating a shared extent. In that case, root_objectid
2084  *                  will be BTRFS_TREE_RELOC_OBJECTID. Otheriwse, parent must
2085  *                  be 0
2086  *
2087  * @root_objectid:  The id of the root where this modification has originated,
2088  *                  this can be either one of the well-known metadata trees or
2089  *                  the subvolume id which references this extent.
2090  *
2091  * @owner:          For data extents it is the inode number of the owning file.
2092  *                  For metadata extents this parameter holds the level in the
2093  *                  tree of the extent.
2094  *
2095  * @offset:         For metadata extents the offset is ignored and is currently
2096  *                  always passed as 0. For data extents it is the fileoffset
2097  *                  this extent belongs to.
2098  *
2099  * @refs_to_add     Number of references to add
2100  *
2101  * @extent_op       Pointer to a structure, holding information necessary when
2102  *                  updating a tree block's flags
2103  *
2104  */
2105 static int __btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
2106                                   struct btrfs_delayed_ref_node *node,
2107                                   u64 parent, u64 root_objectid,
2108                                   u64 owner, u64 offset, int refs_to_add,
2109                                   struct btrfs_delayed_extent_op *extent_op)
2110 {
2111         struct btrfs_path *path;
2112         struct extent_buffer *leaf;
2113         struct btrfs_extent_item *item;
2114         struct btrfs_key key;
2115         u64 bytenr = node->bytenr;
2116         u64 num_bytes = node->num_bytes;
2117         u64 refs;
2118         int ret;
2119
2120         path = btrfs_alloc_path();
2121         if (!path)
2122                 return -ENOMEM;
2123
2124         path->reada = READA_FORWARD;
2125         path->leave_spinning = 1;
2126         /* this will setup the path even if it fails to insert the back ref */
2127         ret = insert_inline_extent_backref(trans, path, bytenr, num_bytes,
2128                                            parent, root_objectid, owner,
2129                                            offset, refs_to_add, extent_op);
2130         if ((ret < 0 && ret != -EAGAIN) || !ret)
2131                 goto out;
2132
2133         /*
2134          * Ok we had -EAGAIN which means we didn't have space to insert and
2135          * inline extent ref, so just update the reference count and add a
2136          * normal backref.
2137          */
2138         leaf = path->nodes[0];
2139         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
2140         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2141         refs = btrfs_extent_refs(leaf, item);
2142         btrfs_set_extent_refs(leaf, item, refs + refs_to_add);
2143         if (extent_op)
2144                 __run_delayed_extent_op(extent_op, leaf, item);
2145
2146         btrfs_mark_buffer_dirty(leaf);
2147         btrfs_release_path(path);
2148
2149         path->reada = READA_FORWARD;
2150         path->leave_spinning = 1;
2151         /* now insert the actual backref */
2152         ret = insert_extent_backref(trans, path, bytenr, parent, root_objectid,
2153                                     owner, offset, refs_to_add);
2154         if (ret)
2155                 btrfs_abort_transaction(trans, ret);
2156 out:
2157         btrfs_free_path(path);
2158         return ret;
2159 }
2160
2161 static int run_delayed_data_ref(struct btrfs_trans_handle *trans,
2162                                 struct btrfs_delayed_ref_node *node,
2163                                 struct btrfs_delayed_extent_op *extent_op,
2164                                 int insert_reserved)
2165 {
2166         int ret = 0;
2167         struct btrfs_delayed_data_ref *ref;
2168         struct btrfs_key ins;
2169         u64 parent = 0;
2170         u64 ref_root = 0;
2171         u64 flags = 0;
2172
2173         ins.objectid = node->bytenr;
2174         ins.offset = node->num_bytes;
2175         ins.type = BTRFS_EXTENT_ITEM_KEY;
2176
2177         ref = btrfs_delayed_node_to_data_ref(node);
2178         trace_run_delayed_data_ref(trans->fs_info, node, ref, node->action);
2179
2180         if (node->type == BTRFS_SHARED_DATA_REF_KEY)
2181                 parent = ref->parent;
2182         ref_root = ref->root;
2183
2184         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
2185                 if (extent_op)
2186                         flags |= extent_op->flags_to_set;
2187                 ret = alloc_reserved_file_extent(trans, parent, ref_root,
2188                                                  flags, ref->objectid,
2189                                                  ref->offset, &ins,
2190                                                  node->ref_mod);
2191         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
2192                 ret = __btrfs_inc_extent_ref(trans, node, parent, ref_root,
2193                                              ref->objectid, ref->offset,
2194                                              node->ref_mod, extent_op);
2195         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
2196                 ret = __btrfs_free_extent(trans, node, parent,
2197                                           ref_root, ref->objectid,
2198                                           ref->offset, node->ref_mod,
2199                                           extent_op);
2200         } else {
2201                 BUG();
2202         }
2203         return ret;
2204 }
2205
2206 static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op,
2207                                     struct extent_buffer *leaf,
2208                                     struct btrfs_extent_item *ei)
2209 {
2210         u64 flags = btrfs_extent_flags(leaf, ei);
2211         if (extent_op->update_flags) {
2212                 flags |= extent_op->flags_to_set;
2213                 btrfs_set_extent_flags(leaf, ei, flags);
2214         }
2215
2216         if (extent_op->update_key) {
2217                 struct btrfs_tree_block_info *bi;
2218                 BUG_ON(!(flags & BTRFS_EXTENT_FLAG_TREE_BLOCK));
2219                 bi = (struct btrfs_tree_block_info *)(ei + 1);
2220                 btrfs_set_tree_block_key(leaf, bi, &extent_op->key);
2221         }
2222 }
2223
2224 static int run_delayed_extent_op(struct btrfs_trans_handle *trans,
2225                                  struct btrfs_delayed_ref_head *head,
2226                                  struct btrfs_delayed_extent_op *extent_op)
2227 {
2228         struct btrfs_fs_info *fs_info = trans->fs_info;
2229         struct btrfs_key key;
2230         struct btrfs_path *path;
2231         struct btrfs_extent_item *ei;
2232         struct extent_buffer *leaf;
2233         u32 item_size;
2234         int ret;
2235         int err = 0;
2236         int metadata = !extent_op->is_data;
2237
2238         if (trans->aborted)
2239                 return 0;
2240
2241         if (metadata && !btrfs_fs_incompat(fs_info, SKINNY_METADATA))
2242                 metadata = 0;
2243
2244         path = btrfs_alloc_path();
2245         if (!path)
2246                 return -ENOMEM;
2247
2248         key.objectid = head->bytenr;
2249
2250         if (metadata) {
2251                 key.type = BTRFS_METADATA_ITEM_KEY;
2252                 key.offset = extent_op->level;
2253         } else {
2254                 key.type = BTRFS_EXTENT_ITEM_KEY;
2255                 key.offset = head->num_bytes;
2256         }
2257
2258 again:
2259         path->reada = READA_FORWARD;
2260         path->leave_spinning = 1;
2261         ret = btrfs_search_slot(trans, fs_info->extent_root, &key, path, 0, 1);
2262         if (ret < 0) {
2263                 err = ret;
2264                 goto out;
2265         }
2266         if (ret > 0) {
2267                 if (metadata) {
2268                         if (path->slots[0] > 0) {
2269                                 path->slots[0]--;
2270                                 btrfs_item_key_to_cpu(path->nodes[0], &key,
2271                                                       path->slots[0]);
2272                                 if (key.objectid == head->bytenr &&
2273                                     key.type == BTRFS_EXTENT_ITEM_KEY &&
2274                                     key.offset == head->num_bytes)
2275                                         ret = 0;
2276                         }
2277                         if (ret > 0) {
2278                                 btrfs_release_path(path);
2279                                 metadata = 0;
2280
2281                                 key.objectid = head->bytenr;
2282                                 key.offset = head->num_bytes;
2283                                 key.type = BTRFS_EXTENT_ITEM_KEY;
2284                                 goto again;
2285                         }
2286                 } else {
2287                         err = -EIO;
2288                         goto out;
2289                 }
2290         }
2291
2292         leaf = path->nodes[0];
2293         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2294
2295         if (unlikely(item_size < sizeof(*ei))) {
2296                 err = -EINVAL;
2297                 btrfs_print_v0_err(fs_info);
2298                 btrfs_abort_transaction(trans, err);
2299                 goto out;
2300         }
2301
2302         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2303         __run_delayed_extent_op(extent_op, leaf, ei);
2304
2305         btrfs_mark_buffer_dirty(leaf);
2306 out:
2307         btrfs_free_path(path);
2308         return err;
2309 }
2310
2311 static int run_delayed_tree_ref(struct btrfs_trans_handle *trans,
2312                                 struct btrfs_delayed_ref_node *node,
2313                                 struct btrfs_delayed_extent_op *extent_op,
2314                                 int insert_reserved)
2315 {
2316         int ret = 0;
2317         struct btrfs_delayed_tree_ref *ref;
2318         u64 parent = 0;
2319         u64 ref_root = 0;
2320
2321         ref = btrfs_delayed_node_to_tree_ref(node);
2322         trace_run_delayed_tree_ref(trans->fs_info, node, ref, node->action);
2323
2324         if (node->type == BTRFS_SHARED_BLOCK_REF_KEY)
2325                 parent = ref->parent;
2326         ref_root = ref->root;
2327
2328         if (node->ref_mod != 1) {
2329                 btrfs_err(trans->fs_info,
2330         "btree block(%llu) has %d references rather than 1: action %d ref_root %llu parent %llu",
2331                           node->bytenr, node->ref_mod, node->action, ref_root,
2332                           parent);
2333                 return -EIO;
2334         }
2335         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
2336                 BUG_ON(!extent_op || !extent_op->update_flags);
2337                 ret = alloc_reserved_tree_block(trans, node, extent_op);
2338         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
2339                 ret = __btrfs_inc_extent_ref(trans, node, parent, ref_root,
2340                                              ref->level, 0, 1, extent_op);
2341         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
2342                 ret = __btrfs_free_extent(trans, node, parent, ref_root,
2343                                           ref->level, 0, 1, extent_op);
2344         } else {
2345                 BUG();
2346         }
2347         return ret;
2348 }
2349
2350 /* helper function to actually process a single delayed ref entry */
2351 static int run_one_delayed_ref(struct btrfs_trans_handle *trans,
2352                                struct btrfs_delayed_ref_node *node,
2353                                struct btrfs_delayed_extent_op *extent_op,
2354                                int insert_reserved)
2355 {
2356         int ret = 0;
2357
2358         if (trans->aborted) {
2359                 if (insert_reserved)
2360                         btrfs_pin_extent(trans->fs_info, node->bytenr,
2361                                          node->num_bytes, 1);
2362                 return 0;
2363         }
2364
2365         if (node->type == BTRFS_TREE_BLOCK_REF_KEY ||
2366             node->type == BTRFS_SHARED_BLOCK_REF_KEY)
2367                 ret = run_delayed_tree_ref(trans, node, extent_op,
2368                                            insert_reserved);
2369         else if (node->type == BTRFS_EXTENT_DATA_REF_KEY ||
2370                  node->type == BTRFS_SHARED_DATA_REF_KEY)
2371                 ret = run_delayed_data_ref(trans, node, extent_op,
2372                                            insert_reserved);
2373         else
2374                 BUG();
2375         if (ret && insert_reserved)
2376                 btrfs_pin_extent(trans->fs_info, node->bytenr,
2377                                  node->num_bytes, 1);
2378         return ret;
2379 }
2380
2381 static inline struct btrfs_delayed_ref_node *
2382 select_delayed_ref(struct btrfs_delayed_ref_head *head)
2383 {
2384         struct btrfs_delayed_ref_node *ref;
2385
2386         if (RB_EMPTY_ROOT(&head->ref_tree.rb_root))
2387                 return NULL;
2388
2389         /*
2390          * Select a delayed ref of type BTRFS_ADD_DELAYED_REF first.
2391          * This is to prevent a ref count from going down to zero, which deletes
2392          * the extent item from the extent tree, when there still are references
2393          * to add, which would fail because they would not find the extent item.
2394          */
2395         if (!list_empty(&head->ref_add_list))
2396                 return list_first_entry(&head->ref_add_list,
2397                                 struct btrfs_delayed_ref_node, add_list);
2398
2399         ref = rb_entry(rb_first_cached(&head->ref_tree),
2400                        struct btrfs_delayed_ref_node, ref_node);
2401         ASSERT(list_empty(&ref->add_list));
2402         return ref;
2403 }
2404
2405 static void unselect_delayed_ref_head(struct btrfs_delayed_ref_root *delayed_refs,
2406                                       struct btrfs_delayed_ref_head *head)
2407 {
2408         spin_lock(&delayed_refs->lock);
2409         head->processing = 0;
2410         delayed_refs->num_heads_ready++;
2411         spin_unlock(&delayed_refs->lock);
2412         btrfs_delayed_ref_unlock(head);
2413 }
2414
2415 static struct btrfs_delayed_extent_op *cleanup_extent_op(
2416                                 struct btrfs_delayed_ref_head *head)
2417 {
2418         struct btrfs_delayed_extent_op *extent_op = head->extent_op;
2419
2420         if (!extent_op)
2421                 return NULL;
2422
2423         if (head->must_insert_reserved) {
2424                 head->extent_op = NULL;
2425                 btrfs_free_delayed_extent_op(extent_op);
2426                 return NULL;
2427         }
2428         return extent_op;
2429 }
2430
2431 static int run_and_cleanup_extent_op(struct btrfs_trans_handle *trans,
2432                                      struct btrfs_delayed_ref_head *head)
2433 {
2434         struct btrfs_delayed_extent_op *extent_op;
2435         int ret;
2436
2437         extent_op = cleanup_extent_op(head);
2438         if (!extent_op)
2439                 return 0;
2440         head->extent_op = NULL;
2441         spin_unlock(&head->lock);
2442         ret = run_delayed_extent_op(trans, head, extent_op);
2443         btrfs_free_delayed_extent_op(extent_op);
2444         return ret ? ret : 1;
2445 }
2446
2447 void btrfs_cleanup_ref_head_accounting(struct btrfs_fs_info *fs_info,
2448                                   struct btrfs_delayed_ref_root *delayed_refs,
2449                                   struct btrfs_delayed_ref_head *head)
2450 {
2451         int nr_items = 1;       /* Dropping this ref head update. */
2452
2453         if (head->total_ref_mod < 0) {
2454                 struct btrfs_space_info *space_info;
2455                 u64 flags;
2456
2457                 if (head->is_data)
2458                         flags = BTRFS_BLOCK_GROUP_DATA;
2459                 else if (head->is_system)
2460                         flags = BTRFS_BLOCK_GROUP_SYSTEM;
2461                 else
2462                         flags = BTRFS_BLOCK_GROUP_METADATA;
2463                 space_info = __find_space_info(fs_info, flags);
2464                 ASSERT(space_info);
2465                 percpu_counter_add_batch(&space_info->total_bytes_pinned,
2466                                    -head->num_bytes,
2467                                    BTRFS_TOTAL_BYTES_PINNED_BATCH);
2468
2469                 /*
2470                  * We had csum deletions accounted for in our delayed refs rsv,
2471                  * we need to drop the csum leaves for this update from our
2472                  * delayed_refs_rsv.
2473                  */
2474                 if (head->is_data) {
2475                         spin_lock(&delayed_refs->lock);
2476                         delayed_refs->pending_csums -= head->num_bytes;
2477                         spin_unlock(&delayed_refs->lock);
2478                         nr_items += btrfs_csum_bytes_to_leaves(fs_info,
2479                                 head->num_bytes);
2480                 }
2481         }
2482
2483         btrfs_delayed_refs_rsv_release(fs_info, nr_items);
2484 }
2485
2486 static int cleanup_ref_head(struct btrfs_trans_handle *trans,
2487                             struct btrfs_delayed_ref_head *head)
2488 {
2489
2490         struct btrfs_fs_info *fs_info = trans->fs_info;
2491         struct btrfs_delayed_ref_root *delayed_refs;
2492         int ret;
2493
2494         delayed_refs = &trans->transaction->delayed_refs;
2495
2496         ret = run_and_cleanup_extent_op(trans, head);
2497         if (ret < 0) {
2498                 unselect_delayed_ref_head(delayed_refs, head);
2499                 btrfs_debug(fs_info, "run_delayed_extent_op returned %d", ret);
2500                 return ret;
2501         } else if (ret) {
2502                 return ret;
2503         }
2504
2505         /*
2506          * Need to drop our head ref lock and re-acquire the delayed ref lock
2507          * and then re-check to make sure nobody got added.
2508          */
2509         spin_unlock(&head->lock);
2510         spin_lock(&delayed_refs->lock);
2511         spin_lock(&head->lock);
2512         if (!RB_EMPTY_ROOT(&head->ref_tree.rb_root) || head->extent_op) {
2513                 spin_unlock(&head->lock);
2514                 spin_unlock(&delayed_refs->lock);
2515                 return 1;
2516         }
2517         btrfs_delete_ref_head(delayed_refs, head);
2518         spin_unlock(&head->lock);
2519         spin_unlock(&delayed_refs->lock);
2520
2521         if (head->must_insert_reserved) {
2522                 btrfs_pin_extent(fs_info, head->bytenr,
2523                                  head->num_bytes, 1);
2524                 if (head->is_data) {
2525                         ret = btrfs_del_csums(trans, fs_info, head->bytenr,
2526                                               head->num_bytes);
2527                 }
2528         }
2529
2530         btrfs_cleanup_ref_head_accounting(fs_info, delayed_refs, head);
2531
2532         trace_run_delayed_ref_head(fs_info, head, 0);
2533         btrfs_delayed_ref_unlock(head);
2534         btrfs_put_delayed_ref_head(head);
2535         return 0;
2536 }
2537
2538 static struct btrfs_delayed_ref_head *btrfs_obtain_ref_head(
2539                                         struct btrfs_trans_handle *trans)
2540 {
2541         struct btrfs_delayed_ref_root *delayed_refs =
2542                 &trans->transaction->delayed_refs;
2543         struct btrfs_delayed_ref_head *head = NULL;
2544         int ret;
2545
2546         spin_lock(&delayed_refs->lock);
2547         head = btrfs_select_ref_head(delayed_refs);
2548         if (!head) {
2549                 spin_unlock(&delayed_refs->lock);
2550                 return head;
2551         }
2552
2553         /*
2554          * Grab the lock that says we are going to process all the refs for
2555          * this head
2556          */
2557         ret = btrfs_delayed_ref_lock(delayed_refs, head);
2558         spin_unlock(&delayed_refs->lock);
2559
2560         /*
2561          * We may have dropped the spin lock to get the head mutex lock, and
2562          * that might have given someone else time to free the head.  If that's
2563          * true, it has been removed from our list and we can move on.
2564          */
2565         if (ret == -EAGAIN)
2566                 head = ERR_PTR(-EAGAIN);
2567
2568         return head;
2569 }
2570
2571 static int btrfs_run_delayed_refs_for_head(struct btrfs_trans_handle *trans,
2572                                     struct btrfs_delayed_ref_head *locked_ref,
2573                                     unsigned long *run_refs)
2574 {
2575         struct btrfs_fs_info *fs_info = trans->fs_info;
2576         struct btrfs_delayed_ref_root *delayed_refs;
2577         struct btrfs_delayed_extent_op *extent_op;
2578         struct btrfs_delayed_ref_node *ref;
2579         int must_insert_reserved = 0;
2580         int ret;
2581
2582         delayed_refs = &trans->transaction->delayed_refs;
2583
2584         lockdep_assert_held(&locked_ref->mutex);
2585         lockdep_assert_held(&locked_ref->lock);
2586
2587         while ((ref = select_delayed_ref(locked_ref))) {
2588                 if (ref->seq &&
2589                     btrfs_check_delayed_seq(fs_info, ref->seq)) {
2590                         spin_unlock(&locked_ref->lock);
2591                         unselect_delayed_ref_head(delayed_refs, locked_ref);
2592                         return -EAGAIN;
2593                 }
2594
2595                 (*run_refs)++;
2596                 ref->in_tree = 0;
2597                 rb_erase_cached(&ref->ref_node, &locked_ref->ref_tree);
2598                 RB_CLEAR_NODE(&ref->ref_node);
2599                 if (!list_empty(&ref->add_list))
2600                         list_del(&ref->add_list);
2601                 /*
2602                  * When we play the delayed ref, also correct the ref_mod on
2603                  * head
2604                  */
2605                 switch (ref->action) {
2606                 case BTRFS_ADD_DELAYED_REF:
2607                 case BTRFS_ADD_DELAYED_EXTENT:
2608                         locked_ref->ref_mod -= ref->ref_mod;
2609                         break;
2610                 case BTRFS_DROP_DELAYED_REF:
2611                         locked_ref->ref_mod += ref->ref_mod;
2612                         break;
2613                 default:
2614                         WARN_ON(1);
2615                 }
2616                 atomic_dec(&delayed_refs->num_entries);
2617
2618                 /*
2619                  * Record the must_insert_reserved flag before we drop the
2620                  * spin lock.
2621                  */
2622                 must_insert_reserved = locked_ref->must_insert_reserved;
2623                 locked_ref->must_insert_reserved = 0;
2624
2625                 extent_op = locked_ref->extent_op;
2626                 locked_ref->extent_op = NULL;
2627                 spin_unlock(&locked_ref->lock);
2628
2629                 ret = run_one_delayed_ref(trans, ref, extent_op,
2630                                           must_insert_reserved);
2631
2632                 btrfs_free_delayed_extent_op(extent_op);
2633                 if (ret) {
2634                         unselect_delayed_ref_head(delayed_refs, locked_ref);
2635                         btrfs_put_delayed_ref(ref);
2636                         btrfs_debug(fs_info, "run_one_delayed_ref returned %d",
2637                                     ret);
2638                         return ret;
2639                 }
2640
2641                 btrfs_put_delayed_ref(ref);
2642                 cond_resched();
2643
2644                 spin_lock(&locked_ref->lock);
2645                 btrfs_merge_delayed_refs(trans, delayed_refs, locked_ref);
2646         }
2647
2648         return 0;
2649 }
2650
2651 /*
2652  * Returns 0 on success or if called with an already aborted transaction.
2653  * Returns -ENOMEM or -EIO on failure and will abort the transaction.
2654  */
2655 static noinline int __btrfs_run_delayed_refs(struct btrfs_trans_handle *trans,
2656                                              unsigned long nr)
2657 {
2658         struct btrfs_fs_info *fs_info = trans->fs_info;
2659         struct btrfs_delayed_ref_root *delayed_refs;
2660         struct btrfs_delayed_ref_head *locked_ref = NULL;
2661         ktime_t start = ktime_get();
2662         int ret;
2663         unsigned long count = 0;
2664         unsigned long actual_count = 0;
2665
2666         delayed_refs = &trans->transaction->delayed_refs;
2667         do {
2668                 if (!locked_ref) {
2669                         locked_ref = btrfs_obtain_ref_head(trans);
2670                         if (IS_ERR_OR_NULL(locked_ref)) {
2671                                 if (PTR_ERR(locked_ref) == -EAGAIN) {
2672                                         continue;
2673                                 } else {
2674                                         break;
2675                                 }
2676                         }
2677                         count++;
2678                 }
2679                 /*
2680                  * We need to try and merge add/drops of the same ref since we
2681                  * can run into issues with relocate dropping the implicit ref
2682                  * and then it being added back again before the drop can
2683                  * finish.  If we merged anything we need to re-loop so we can
2684                  * get a good ref.
2685                  * Or we can get node references of the same type that weren't
2686                  * merged when created due to bumps in the tree mod seq, and
2687                  * we need to merge them to prevent adding an inline extent
2688                  * backref before dropping it (triggering a BUG_ON at
2689                  * insert_inline_extent_backref()).
2690                  */
2691                 spin_lock(&locked_ref->lock);
2692                 btrfs_merge_delayed_refs(trans, delayed_refs, locked_ref);
2693
2694                 ret = btrfs_run_delayed_refs_for_head(trans, locked_ref,
2695                                                       &actual_count);
2696                 if (ret < 0 && ret != -EAGAIN) {
2697                         /*
2698                          * Error, btrfs_run_delayed_refs_for_head already
2699                          * unlocked everything so just bail out
2700                          */
2701                         return ret;
2702                 } else if (!ret) {
2703                         /*
2704                          * Success, perform the usual cleanup of a processed
2705                          * head
2706                          */
2707                         ret = cleanup_ref_head(trans, locked_ref);
2708                         if (ret > 0 ) {
2709                                 /* We dropped our lock, we need to loop. */
2710                                 ret = 0;
2711                                 continue;
2712                         } else if (ret) {
2713                                 return ret;
2714                         }
2715                 }
2716
2717                 /*
2718                  * Either success case or btrfs_run_delayed_refs_for_head
2719                  * returned -EAGAIN, meaning we need to select another head
2720                  */
2721
2722                 locked_ref = NULL;
2723                 cond_resched();
2724         } while ((nr != -1 && count < nr) || locked_ref);
2725
2726         /*
2727          * We don't want to include ref heads since we can have empty ref heads
2728          * and those will drastically skew our runtime down since we just do
2729          * accounting, no actual extent tree updates.
2730          */
2731         if (actual_count > 0) {
2732                 u64 runtime = ktime_to_ns(ktime_sub(ktime_get(), start));
2733                 u64 avg;
2734
2735                 /*
2736                  * We weigh the current average higher than our current runtime
2737                  * to avoid large swings in the average.
2738                  */
2739                 spin_lock(&delayed_refs->lock);
2740                 avg = fs_info->avg_delayed_ref_runtime * 3 + runtime;
2741                 fs_info->avg_delayed_ref_runtime = avg >> 2;    /* div by 4 */
2742                 spin_unlock(&delayed_refs->lock);
2743         }
2744         return 0;
2745 }
2746
2747 #ifdef SCRAMBLE_DELAYED_REFS
2748 /*
2749  * Normally delayed refs get processed in ascending bytenr order. This
2750  * correlates in most cases to the order added. To expose dependencies on this
2751  * order, we start to process the tree in the middle instead of the beginning
2752  */
2753 static u64 find_middle(struct rb_root *root)
2754 {
2755         struct rb_node *n = root->rb_node;
2756         struct btrfs_delayed_ref_node *entry;
2757         int alt = 1;
2758         u64 middle;
2759         u64 first = 0, last = 0;
2760
2761         n = rb_first(root);
2762         if (n) {
2763                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2764                 first = entry->bytenr;
2765         }
2766         n = rb_last(root);
2767         if (n) {
2768                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2769                 last = entry->bytenr;
2770         }
2771         n = root->rb_node;
2772
2773         while (n) {
2774                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2775                 WARN_ON(!entry->in_tree);
2776
2777                 middle = entry->bytenr;
2778
2779                 if (alt)
2780                         n = n->rb_left;
2781                 else
2782                         n = n->rb_right;
2783
2784                 alt = 1 - alt;
2785         }
2786         return middle;
2787 }
2788 #endif
2789
2790 static inline u64 heads_to_leaves(struct btrfs_fs_info *fs_info, u64 heads)
2791 {
2792         u64 num_bytes;
2793
2794         num_bytes = heads * (sizeof(struct btrfs_extent_item) +
2795                              sizeof(struct btrfs_extent_inline_ref));
2796         if (!btrfs_fs_incompat(fs_info, SKINNY_METADATA))
2797                 num_bytes += heads * sizeof(struct btrfs_tree_block_info);
2798
2799         /*
2800          * We don't ever fill up leaves all the way so multiply by 2 just to be
2801          * closer to what we're really going to want to use.
2802          */
2803         return div_u64(num_bytes, BTRFS_LEAF_DATA_SIZE(fs_info));
2804 }
2805
2806 /*
2807  * Takes the number of bytes to be csumm'ed and figures out how many leaves it
2808  * would require to store the csums for that many bytes.
2809  */
2810 u64 btrfs_csum_bytes_to_leaves(struct btrfs_fs_info *fs_info, u64 csum_bytes)
2811 {
2812         u64 csum_size;
2813         u64 num_csums_per_leaf;
2814         u64 num_csums;
2815
2816         csum_size = BTRFS_MAX_ITEM_SIZE(fs_info);
2817         num_csums_per_leaf = div64_u64(csum_size,
2818                         (u64)btrfs_super_csum_size(fs_info->super_copy));
2819         num_csums = div64_u64(csum_bytes, fs_info->sectorsize);
2820         num_csums += num_csums_per_leaf - 1;
2821         num_csums = div64_u64(num_csums, num_csums_per_leaf);
2822         return num_csums;
2823 }
2824
2825 bool btrfs_check_space_for_delayed_refs(struct btrfs_fs_info *fs_info)
2826 {
2827         struct btrfs_block_rsv *delayed_refs_rsv = &fs_info->delayed_refs_rsv;
2828         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
2829         bool ret = false;
2830         u64 reserved;
2831
2832         spin_lock(&global_rsv->lock);
2833         reserved = global_rsv->reserved;
2834         spin_unlock(&global_rsv->lock);
2835
2836         /*
2837          * Since the global reserve is just kind of magic we don't really want
2838          * to rely on it to save our bacon, so if our size is more than the
2839          * delayed_refs_rsv and the global rsv then it's time to think about
2840          * bailing.
2841          */
2842         spin_lock(&delayed_refs_rsv->lock);
2843         reserved += delayed_refs_rsv->reserved;
2844         if (delayed_refs_rsv->size >= reserved)
2845                 ret = true;
2846         spin_unlock(&delayed_refs_rsv->lock);
2847         return ret;
2848 }
2849
2850 int btrfs_should_throttle_delayed_refs(struct btrfs_trans_handle *trans)
2851 {
2852         u64 num_entries =
2853                 atomic_read(&trans->transaction->delayed_refs.num_entries);
2854         u64 avg_runtime;
2855         u64 val;
2856
2857         smp_mb();
2858         avg_runtime = trans->fs_info->avg_delayed_ref_runtime;
2859         val = num_entries * avg_runtime;
2860         if (val >= NSEC_PER_SEC)
2861                 return 1;
2862         if (val >= NSEC_PER_SEC / 2)
2863                 return 2;
2864
2865         return btrfs_check_space_for_delayed_refs(trans->fs_info);
2866 }
2867
2868 /*
2869  * this starts processing the delayed reference count updates and
2870  * extent insertions we have queued up so far.  count can be
2871  * 0, which means to process everything in the tree at the start
2872  * of the run (but not newly added entries), or it can be some target
2873  * number you'd like to process.
2874  *
2875  * Returns 0 on success or if called with an aborted transaction
2876  * Returns <0 on error and aborts the transaction
2877  */
2878 int btrfs_run_delayed_refs(struct btrfs_trans_handle *trans,
2879                            unsigned long count)
2880 {
2881         struct btrfs_fs_info *fs_info = trans->fs_info;
2882         struct rb_node *node;
2883         struct btrfs_delayed_ref_root *delayed_refs;
2884         struct btrfs_delayed_ref_head *head;
2885         int ret;
2886         int run_all = count == (unsigned long)-1;
2887
2888         /* We'll clean this up in btrfs_cleanup_transaction */
2889         if (trans->aborted)
2890                 return 0;
2891
2892         if (test_bit(BTRFS_FS_CREATING_FREE_SPACE_TREE, &fs_info->flags))
2893                 return 0;
2894
2895         delayed_refs = &trans->transaction->delayed_refs;
2896         if (count == 0)
2897                 count = atomic_read(&delayed_refs->num_entries) * 2;
2898
2899 again:
2900 #ifdef SCRAMBLE_DELAYED_REFS
2901         delayed_refs->run_delayed_start = find_middle(&delayed_refs->root);
2902 #endif
2903         ret = __btrfs_run_delayed_refs(trans, count);
2904         if (ret < 0) {
2905                 btrfs_abort_transaction(trans, ret);
2906                 return ret;
2907         }
2908
2909         if (run_all) {
2910                 btrfs_create_pending_block_groups(trans);
2911
2912                 spin_lock(&delayed_refs->lock);
2913                 node = rb_first_cached(&delayed_refs->href_root);
2914                 if (!node) {
2915                         spin_unlock(&delayed_refs->lock);
2916                         goto out;
2917                 }
2918                 head = rb_entry(node, struct btrfs_delayed_ref_head,
2919                                 href_node);
2920                 refcount_inc(&head->refs);
2921                 spin_unlock(&delayed_refs->lock);
2922
2923                 /* Mutex was contended, block until it's released and retry. */
2924                 mutex_lock(&head->mutex);
2925                 mutex_unlock(&head->mutex);
2926
2927                 btrfs_put_delayed_ref_head(head);
2928                 cond_resched();
2929                 goto again;
2930         }
2931 out:
2932         return 0;
2933 }
2934
2935 int btrfs_set_disk_extent_flags(struct btrfs_trans_handle *trans,
2936                                 u64 bytenr, u64 num_bytes, u64 flags,
2937                                 int level, int is_data)
2938 {
2939         struct btrfs_delayed_extent_op *extent_op;
2940         int ret;
2941
2942         extent_op = btrfs_alloc_delayed_extent_op();
2943         if (!extent_op)
2944                 return -ENOMEM;
2945
2946         extent_op->flags_to_set = flags;
2947         extent_op->update_flags = true;
2948         extent_op->update_key = false;
2949         extent_op->is_data = is_data ? true : false;
2950         extent_op->level = level;
2951
2952         ret = btrfs_add_delayed_extent_op(trans, bytenr, num_bytes, extent_op);
2953         if (ret)
2954                 btrfs_free_delayed_extent_op(extent_op);
2955         return ret;
2956 }
2957
2958 static noinline int check_delayed_ref(struct btrfs_root *root,
2959                                       struct btrfs_path *path,
2960                                       u64 objectid, u64 offset, u64 bytenr)
2961 {
2962         struct btrfs_delayed_ref_head *head;
2963         struct btrfs_delayed_ref_node *ref;
2964         struct btrfs_delayed_data_ref *data_ref;
2965         struct btrfs_delayed_ref_root *delayed_refs;
2966         struct btrfs_transaction *cur_trans;
2967         struct rb_node *node;
2968         int ret = 0;
2969
2970         spin_lock(&root->fs_info->trans_lock);
2971         cur_trans = root->fs_info->running_transaction;
2972         if (cur_trans)
2973                 refcount_inc(&cur_trans->use_count);
2974         spin_unlock(&root->fs_info->trans_lock);
2975         if (!cur_trans)
2976                 return 0;
2977
2978         delayed_refs = &cur_trans->delayed_refs;
2979         spin_lock(&delayed_refs->lock);
2980         head = btrfs_find_delayed_ref_head(delayed_refs, bytenr);
2981         if (!head) {
2982                 spin_unlock(&delayed_refs->lock);
2983                 btrfs_put_transaction(cur_trans);
2984                 return 0;
2985         }
2986
2987         if (!mutex_trylock(&head->mutex)) {
2988                 refcount_inc(&head->refs);
2989                 spin_unlock(&delayed_refs->lock);
2990
2991                 btrfs_release_path(path);
2992
2993                 /*
2994                  * Mutex was contended, block until it's released and let
2995                  * caller try again
2996                  */
2997                 mutex_lock(&head->mutex);
2998                 mutex_unlock(&head->mutex);
2999                 btrfs_put_delayed_ref_head(head);
3000                 btrfs_put_transaction(cur_trans);
3001                 return -EAGAIN;
3002         }
3003         spin_unlock(&delayed_refs->lock);
3004
3005         spin_lock(&head->lock);
3006         /*
3007          * XXX: We should replace this with a proper search function in the
3008          * future.
3009          */
3010         for (node = rb_first_cached(&head->ref_tree); node;
3011              node = rb_next(node)) {
3012                 ref = rb_entry(node, struct btrfs_delayed_ref_node, ref_node);
3013                 /* If it's a shared ref we know a cross reference exists */
3014                 if (ref->type != BTRFS_EXTENT_DATA_REF_KEY) {
3015                         ret = 1;
3016                         break;
3017                 }
3018
3019                 data_ref = btrfs_delayed_node_to_data_ref(ref);
3020
3021                 /*
3022                  * If our ref doesn't match the one we're currently looking at
3023                  * then we have a cross reference.
3024                  */
3025                 if (data_ref->root != root->root_key.objectid ||
3026                     data_ref->objectid != objectid ||
3027                     data_ref->offset != offset) {
3028                         ret = 1;
3029                         break;
3030                 }
3031         }
3032         spin_unlock(&head->lock);
3033         mutex_unlock(&head->mutex);
3034         btrfs_put_transaction(cur_trans);
3035         return ret;
3036 }
3037
3038 static noinline int check_committed_ref(struct btrfs_root *root,
3039                                         struct btrfs_path *path,
3040                                         u64 objectid, u64 offset, u64 bytenr)
3041 {
3042         struct btrfs_fs_info *fs_info = root->fs_info;
3043         struct btrfs_root *extent_root = fs_info->extent_root;
3044         struct extent_buffer *leaf;
3045         struct btrfs_extent_data_ref *ref;
3046         struct btrfs_extent_inline_ref *iref;
3047         struct btrfs_extent_item *ei;
3048         struct btrfs_key key;
3049         u32 item_size;
3050         int type;
3051         int ret;
3052
3053         key.objectid = bytenr;
3054         key.offset = (u64)-1;
3055         key.type = BTRFS_EXTENT_ITEM_KEY;
3056
3057         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
3058         if (ret < 0)
3059                 goto out;
3060         BUG_ON(ret == 0); /* Corruption */
3061
3062         ret = -ENOENT;
3063         if (path->slots[0] == 0)
3064                 goto out;
3065
3066         path->slots[0]--;
3067         leaf = path->nodes[0];
3068         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
3069
3070         if (key.objectid != bytenr || key.type != BTRFS_EXTENT_ITEM_KEY)
3071                 goto out;
3072
3073         ret = 1;
3074         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
3075         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
3076
3077         if (item_size != sizeof(*ei) +
3078             btrfs_extent_inline_ref_size(BTRFS_EXTENT_DATA_REF_KEY))
3079                 goto out;
3080
3081         if (btrfs_extent_generation(leaf, ei) <=
3082             btrfs_root_last_snapshot(&root->root_item))
3083                 goto out;
3084
3085         iref = (struct btrfs_extent_inline_ref *)(ei + 1);
3086
3087         type = btrfs_get_extent_inline_ref_type(leaf, iref, BTRFS_REF_TYPE_DATA);
3088         if (type != BTRFS_EXTENT_DATA_REF_KEY)
3089                 goto out;
3090
3091         ref = (struct btrfs_extent_data_ref *)(&iref->offset);
3092         if (btrfs_extent_refs(leaf, ei) !=
3093             btrfs_extent_data_ref_count(leaf, ref) ||
3094             btrfs_extent_data_ref_root(leaf, ref) !=
3095             root->root_key.objectid ||
3096             btrfs_extent_data_ref_objectid(leaf, ref) != objectid ||
3097             btrfs_extent_data_ref_offset(leaf, ref) != offset)
3098                 goto out;
3099
3100         ret = 0;
3101 out:
3102         return ret;
3103 }
3104
3105 int btrfs_cross_ref_exist(struct btrfs_root *root, u64 objectid, u64 offset,
3106                           u64 bytenr)
3107 {
3108         struct btrfs_path *path;
3109         int ret;
3110
3111         path = btrfs_alloc_path();
3112         if (!path)
3113                 return -ENOMEM;
3114
3115         do {
3116                 ret = check_committed_ref(root, path, objectid,
3117                                           offset, bytenr);
3118                 if (ret && ret != -ENOENT)
3119                         goto out;
3120
3121                 ret = check_delayed_ref(root, path, objectid, offset, bytenr);
3122         } while (ret == -EAGAIN);
3123
3124 out:
3125         btrfs_free_path(path);
3126         if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID)
3127                 WARN_ON(ret > 0);
3128         return ret;
3129 }
3130
3131 static int __btrfs_mod_ref(struct btrfs_trans_handle *trans,
3132                            struct btrfs_root *root,
3133                            struct extent_buffer *buf,
3134                            int full_backref, int inc)
3135 {
3136         struct btrfs_fs_info *fs_info = root->fs_info;
3137         u64 bytenr;
3138         u64 num_bytes;
3139         u64 parent;
3140         u64 ref_root;
3141         u32 nritems;
3142         struct btrfs_key key;
3143         struct btrfs_file_extent_item *fi;
3144         struct btrfs_ref generic_ref = { 0 };
3145         bool for_reloc = btrfs_header_flag(buf, BTRFS_HEADER_FLAG_RELOC);
3146         int i;
3147         int action;
3148         int level;
3149         int ret = 0;
3150
3151         if (btrfs_is_testing(fs_info))
3152                 return 0;
3153
3154         ref_root = btrfs_header_owner(buf);
3155         nritems = btrfs_header_nritems(buf);
3156         level = btrfs_header_level(buf);
3157
3158         if (!test_bit(BTRFS_ROOT_REF_COWS, &root->state) && level == 0)
3159                 return 0;
3160
3161         if (full_backref)
3162                 parent = buf->start;
3163         else
3164                 parent = 0;
3165         if (inc)
3166                 action = BTRFS_ADD_DELAYED_REF;
3167         else
3168                 action = BTRFS_DROP_DELAYED_REF;
3169
3170         for (i = 0; i < nritems; i++) {
3171                 if (level == 0) {
3172                         btrfs_item_key_to_cpu(buf, &key, i);
3173                         if (key.type != BTRFS_EXTENT_DATA_KEY)
3174                                 continue;
3175                         fi = btrfs_item_ptr(buf, i,
3176                                             struct btrfs_file_extent_item);
3177                         if (btrfs_file_extent_type(buf, fi) ==
3178                             BTRFS_FILE_EXTENT_INLINE)
3179                                 continue;
3180                         bytenr = btrfs_file_extent_disk_bytenr(buf, fi);
3181                         if (bytenr == 0)
3182                                 continue;
3183
3184                         num_bytes = btrfs_file_extent_disk_num_bytes(buf, fi);
3185                         key.offset -= btrfs_file_extent_offset(buf, fi);
3186                         btrfs_init_generic_ref(&generic_ref, action, bytenr,
3187                                                num_bytes, parent);
3188                         generic_ref.real_root = root->root_key.objectid;
3189                         btrfs_init_data_ref(&generic_ref, ref_root, key.objectid,
3190                                             key.offset);
3191                         generic_ref.skip_qgroup = for_reloc;
3192                         if (inc)
3193                                 ret = btrfs_inc_extent_ref(trans, &generic_ref);
3194                         else
3195                                 ret = btrfs_free_extent(trans, &generic_ref);
3196                         if (ret)
3197                                 goto fail;
3198                 } else {
3199                         bytenr = btrfs_node_blockptr(buf, i);
3200                         num_bytes = fs_info->nodesize;
3201                         btrfs_init_generic_ref(&generic_ref, action, bytenr,
3202                                                num_bytes, parent);
3203                         generic_ref.real_root = root->root_key.objectid;
3204                         btrfs_init_tree_ref(&generic_ref, level - 1, ref_root);
3205                         generic_ref.skip_qgroup = for_reloc;
3206                         if (inc)
3207                                 ret = btrfs_inc_extent_ref(trans, &generic_ref);
3208                         else
3209                                 ret = btrfs_free_extent(trans, &generic_ref);
3210                         if (ret)
3211                                 goto fail;
3212                 }
3213         }
3214         return 0;
3215 fail:
3216         return ret;
3217 }
3218
3219 int btrfs_inc_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
3220                   struct extent_buffer *buf, int full_backref)
3221 {
3222         return __btrfs_mod_ref(trans, root, buf, full_backref, 1);
3223 }
3224
3225 int btrfs_dec_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
3226                   struct extent_buffer *buf, int full_backref)
3227 {
3228         return __btrfs_mod_ref(trans, root, buf, full_backref, 0);
3229 }
3230
3231 static int write_one_cache_group(struct btrfs_trans_handle *trans,
3232                                  struct btrfs_path *path,
3233                                  struct btrfs_block_group_cache *cache)
3234 {
3235         struct btrfs_fs_info *fs_info = trans->fs_info;
3236         int ret;
3237         struct btrfs_root *extent_root = fs_info->extent_root;
3238         unsigned long bi;
3239         struct extent_buffer *leaf;
3240
3241         ret = btrfs_search_slot(trans, extent_root, &cache->key, path, 0, 1);
3242         if (ret) {
3243                 if (ret > 0)
3244                         ret = -ENOENT;
3245                 goto fail;
3246         }
3247
3248         leaf = path->nodes[0];
3249         bi = btrfs_item_ptr_offset(leaf, path->slots[0]);
3250         write_extent_buffer(leaf, &cache->item, bi, sizeof(cache->item));
3251         btrfs_mark_buffer_dirty(leaf);
3252 fail:
3253         btrfs_release_path(path);
3254         return ret;
3255
3256 }
3257
3258 static struct btrfs_block_group_cache *next_block_group(
3259                 struct btrfs_block_group_cache *cache)
3260 {
3261         struct btrfs_fs_info *fs_info = cache->fs_info;
3262         struct rb_node *node;
3263
3264         spin_lock(&fs_info->block_group_cache_lock);
3265
3266         /* If our block group was removed, we need a full search. */
3267         if (RB_EMPTY_NODE(&cache->cache_node)) {
3268                 const u64 next_bytenr = cache->key.objectid + cache->key.offset;
3269
3270                 spin_unlock(&fs_info->block_group_cache_lock);
3271                 btrfs_put_block_group(cache);
3272                 cache = btrfs_lookup_first_block_group(fs_info, next_bytenr); return cache;
3273         }
3274         node = rb_next(&cache->cache_node);
3275         btrfs_put_block_group(cache);
3276         if (node) {
3277                 cache = rb_entry(node, struct btrfs_block_group_cache,
3278                                  cache_node);
3279                 btrfs_get_block_group(cache);
3280         } else
3281                 cache = NULL;
3282         spin_unlock(&fs_info->block_group_cache_lock);
3283         return cache;
3284 }
3285
3286 static int cache_save_setup(struct btrfs_block_group_cache *block_group,
3287                             struct btrfs_trans_handle *trans,
3288                             struct btrfs_path *path)
3289 {
3290         struct btrfs_fs_info *fs_info = block_group->fs_info;
3291         struct btrfs_root *root = fs_info->tree_root;
3292         struct inode *inode = NULL;
3293         struct extent_changeset *data_reserved = NULL;
3294         u64 alloc_hint = 0;
3295         int dcs = BTRFS_DC_ERROR;
3296         u64 num_pages = 0;
3297         int retries = 0;
3298         int ret = 0;
3299
3300         /*
3301          * If this block group is smaller than 100 megs don't bother caching the
3302          * block group.
3303          */
3304         if (block_group->key.offset < (100 * SZ_1M)) {
3305                 spin_lock(&block_group->lock);
3306                 block_group->disk_cache_state = BTRFS_DC_WRITTEN;
3307                 spin_unlock(&block_group->lock);
3308                 return 0;
3309         }
3310
3311         if (trans->aborted)
3312                 return 0;
3313 again:
3314         inode = lookup_free_space_inode(block_group, path);
3315         if (IS_ERR(inode) && PTR_ERR(inode) != -ENOENT) {
3316                 ret = PTR_ERR(inode);
3317                 btrfs_release_path(path);
3318                 goto out;
3319         }
3320
3321         if (IS_ERR(inode)) {
3322                 BUG_ON(retries);
3323                 retries++;
3324
3325                 if (block_group->ro)
3326                         goto out_free;
3327
3328                 ret = create_free_space_inode(trans, block_group, path);
3329                 if (ret)
3330                         goto out_free;
3331                 goto again;
3332         }
3333
3334         /*
3335          * We want to set the generation to 0, that way if anything goes wrong
3336          * from here on out we know not to trust this cache when we load up next
3337          * time.
3338          */
3339         BTRFS_I(inode)->generation = 0;
3340         ret = btrfs_update_inode(trans, root, inode);
3341         if (ret) {
3342                 /*
3343                  * So theoretically we could recover from this, simply set the
3344                  * super cache generation to 0 so we know to invalidate the
3345                  * cache, but then we'd have to keep track of the block groups
3346                  * that fail this way so we know we _have_ to reset this cache
3347                  * before the next commit or risk reading stale cache.  So to
3348                  * limit our exposure to horrible edge cases lets just abort the
3349                  * transaction, this only happens in really bad situations
3350                  * anyway.
3351                  */
3352                 btrfs_abort_transaction(trans, ret);
3353                 goto out_put;
3354         }
3355         WARN_ON(ret);
3356
3357         /* We've already setup this transaction, go ahead and exit */
3358         if (block_group->cache_generation == trans->transid &&
3359             i_size_read(inode)) {
3360                 dcs = BTRFS_DC_SETUP;
3361                 goto out_put;
3362         }
3363
3364         if (i_size_read(inode) > 0) {
3365                 ret = btrfs_check_trunc_cache_free_space(fs_info,
3366                                         &fs_info->global_block_rsv);
3367                 if (ret)
3368                         goto out_put;
3369
3370                 ret = btrfs_truncate_free_space_cache(trans, NULL, inode);
3371                 if (ret)
3372                         goto out_put;
3373         }
3374
3375         spin_lock(&block_group->lock);
3376         if (block_group->cached != BTRFS_CACHE_FINISHED ||
3377             !btrfs_test_opt(fs_info, SPACE_CACHE)) {
3378                 /*
3379                  * don't bother trying to write stuff out _if_
3380                  * a) we're not cached,
3381                  * b) we're with nospace_cache mount option,
3382                  * c) we're with v2 space_cache (FREE_SPACE_TREE).
3383                  */
3384                 dcs = BTRFS_DC_WRITTEN;
3385                 spin_unlock(&block_group->lock);
3386                 goto out_put;
3387         }
3388         spin_unlock(&block_group->lock);
3389
3390         /*
3391          * We hit an ENOSPC when setting up the cache in this transaction, just
3392          * skip doing the setup, we've already cleared the cache so we're safe.
3393          */
3394         if (test_bit(BTRFS_TRANS_CACHE_ENOSPC, &trans->transaction->flags)) {
3395                 ret = -ENOSPC;
3396                 goto out_put;
3397         }
3398
3399         /*
3400          * Try to preallocate enough space based on how big the block group is.
3401          * Keep in mind this has to include any pinned space which could end up
3402          * taking up quite a bit since it's not folded into the other space
3403          * cache.
3404          */
3405         num_pages = div_u64(block_group->key.offset, SZ_256M);
3406         if (!num_pages)
3407                 num_pages = 1;
3408
3409         num_pages *= 16;
3410         num_pages *= PAGE_SIZE;
3411
3412         ret = btrfs_check_data_free_space(inode, &data_reserved, 0, num_pages);
3413         if (ret)
3414                 goto out_put;
3415
3416         ret = btrfs_prealloc_file_range_trans(inode, trans, 0, 0, num_pages,
3417                                               num_pages, num_pages,
3418                                               &alloc_hint);
3419         /*
3420          * Our cache requires contiguous chunks so that we don't modify a bunch
3421          * of metadata or split extents when writing the cache out, which means
3422          * we can enospc if we are heavily fragmented in addition to just normal
3423          * out of space conditions.  So if we hit this just skip setting up any
3424          * other block groups for this transaction, maybe we'll unpin enough
3425          * space the next time around.
3426          */
3427         if (!ret)
3428                 dcs = BTRFS_DC_SETUP;
3429         else if (ret == -ENOSPC)
3430                 set_bit(BTRFS_TRANS_CACHE_ENOSPC, &trans->transaction->flags);
3431
3432 out_put:
3433         iput(inode);
3434 out_free:
3435         btrfs_release_path(path);
3436 out:
3437         spin_lock(&block_group->lock);
3438         if (!ret && dcs == BTRFS_DC_SETUP)
3439                 block_group->cache_generation = trans->transid;
3440         block_group->disk_cache_state = dcs;
3441         spin_unlock(&block_group->lock);
3442
3443         extent_changeset_free(data_reserved);
3444         return ret;
3445 }
3446
3447 int btrfs_setup_space_cache(struct btrfs_trans_handle *trans)
3448 {
3449         struct btrfs_fs_info *fs_info = trans->fs_info;
3450         struct btrfs_block_group_cache *cache, *tmp;
3451         struct btrfs_transaction *cur_trans = trans->transaction;
3452         struct btrfs_path *path;
3453
3454         if (list_empty(&cur_trans->dirty_bgs) ||
3455             !btrfs_test_opt(fs_info, SPACE_CACHE))
3456                 return 0;
3457
3458         path = btrfs_alloc_path();
3459         if (!path)
3460                 return -ENOMEM;
3461
3462         /* Could add new block groups, use _safe just in case */
3463         list_for_each_entry_safe(cache, tmp, &cur_trans->dirty_bgs,
3464                                  dirty_list) {
3465                 if (cache->disk_cache_state == BTRFS_DC_CLEAR)
3466                         cache_save_setup(cache, trans, path);
3467         }
3468
3469         btrfs_free_path(path);
3470         return 0;
3471 }
3472
3473 /*
3474  * transaction commit does final block group cache writeback during a
3475  * critical section where nothing is allowed to change the FS.  This is
3476  * required in order for the cache to actually match the block group,
3477  * but can introduce a lot of latency into the commit.
3478  *
3479  * So, btrfs_start_dirty_block_groups is here to kick off block group
3480  * cache IO.  There's a chance we'll have to redo some of it if the
3481  * block group changes again during the commit, but it greatly reduces
3482  * the commit latency by getting rid of the easy block groups while
3483  * we're still allowing others to join the commit.
3484  */
3485 int btrfs_start_dirty_block_groups(struct btrfs_trans_handle *trans)
3486 {
3487         struct btrfs_fs_info *fs_info = trans->fs_info;
3488         struct btrfs_block_group_cache *cache;
3489         struct btrfs_transaction *cur_trans = trans->transaction;
3490         int ret = 0;
3491         int should_put;
3492         struct btrfs_path *path = NULL;
3493         LIST_HEAD(dirty);
3494         struct list_head *io = &cur_trans->io_bgs;
3495         int num_started = 0;
3496         int loops = 0;
3497
3498         spin_lock(&cur_trans->dirty_bgs_lock);
3499         if (list_empty(&cur_trans->dirty_bgs)) {
3500                 spin_unlock(&cur_trans->dirty_bgs_lock);
3501                 return 0;
3502         }
3503         list_splice_init(&cur_trans->dirty_bgs, &dirty);
3504         spin_unlock(&cur_trans->dirty_bgs_lock);
3505
3506 again:
3507         /*
3508          * make sure all the block groups on our dirty list actually
3509          * exist
3510          */
3511         btrfs_create_pending_block_groups(trans);
3512
3513         if (!path) {
3514                 path = btrfs_alloc_path();
3515                 if (!path)
3516                         return -ENOMEM;
3517         }
3518
3519         /*
3520          * cache_write_mutex is here only to save us from balance or automatic
3521          * removal of empty block groups deleting this block group while we are
3522          * writing out the cache
3523          */
3524         mutex_lock(&trans->transaction->cache_write_mutex);
3525         while (!list_empty(&dirty)) {
3526                 bool drop_reserve = true;
3527
3528                 cache = list_first_entry(&dirty,
3529                                          struct btrfs_block_group_cache,
3530                                          dirty_list);
3531                 /*
3532                  * this can happen if something re-dirties a block
3533                  * group that is already under IO.  Just wait for it to
3534                  * finish and then do it all again
3535                  */
3536                 if (!list_empty(&cache->io_list)) {
3537                         list_del_init(&cache->io_list);
3538                         btrfs_wait_cache_io(trans, cache, path);
3539                         btrfs_put_block_group(cache);
3540                 }
3541
3542
3543                 /*
3544                  * btrfs_wait_cache_io uses the cache->dirty_list to decide
3545                  * if it should update the cache_state.  Don't delete
3546                  * until after we wait.
3547                  *
3548                  * Since we're not running in the commit critical section
3549                  * we need the dirty_bgs_lock to protect from update_block_group
3550                  */
3551                 spin_lock(&cur_trans->dirty_bgs_lock);
3552                 list_del_init(&cache->dirty_list);
3553                 spin_unlock(&cur_trans->dirty_bgs_lock);
3554
3555                 should_put = 1;
3556
3557                 cache_save_setup(cache, trans, path);
3558
3559                 if (cache->disk_cache_state == BTRFS_DC_SETUP) {
3560                         cache->io_ctl.inode = NULL;
3561                         ret = btrfs_write_out_cache(trans, cache, path);
3562                         if (ret == 0 && cache->io_ctl.inode) {
3563                                 num_started++;
3564                                 should_put = 0;
3565
3566                                 /*
3567                                  * The cache_write_mutex is protecting the
3568                                  * io_list, also refer to the definition of
3569                                  * btrfs_transaction::io_bgs for more details
3570                                  */
3571                                 list_add_tail(&cache->io_list, io);
3572                         } else {
3573                                 /*
3574                                  * if we failed to write the cache, the
3575                                  * generation will be bad and life goes on
3576                                  */
3577                                 ret = 0;
3578                         }
3579                 }
3580                 if (!ret) {
3581                         ret = write_one_cache_group(trans, path, cache);
3582                         /*
3583                          * Our block group might still be attached to the list
3584                          * of new block groups in the transaction handle of some
3585                          * other task (struct btrfs_trans_handle->new_bgs). This
3586                          * means its block group item isn't yet in the extent
3587                          * tree. If this happens ignore the error, as we will
3588                          * try again later in the critical section of the
3589                          * transaction commit.
3590                          */
3591                         if (ret == -ENOENT) {
3592                                 ret = 0;
3593                                 spin_lock(&cur_trans->dirty_bgs_lock);
3594                                 if (list_empty(&cache->dirty_list)) {
3595                                         list_add_tail(&cache->dirty_list,
3596                                                       &cur_trans->dirty_bgs);
3597                                         btrfs_get_block_group(cache);
3598                                         drop_reserve = false;
3599                                 }
3600                                 spin_unlock(&cur_trans->dirty_bgs_lock);
3601                         } else if (ret) {
3602                                 btrfs_abort_transaction(trans, ret);
3603                         }
3604                 }
3605
3606                 /* if it's not on the io list, we need to put the block group */
3607                 if (should_put)
3608                         btrfs_put_block_group(cache);
3609                 if (drop_reserve)
3610                         btrfs_delayed_refs_rsv_release(fs_info, 1);
3611
3612                 if (ret)
3613                         break;
3614
3615                 /*
3616                  * Avoid blocking other tasks for too long. It might even save
3617                  * us from writing caches for block groups that are going to be
3618                  * removed.
3619                  */
3620                 mutex_unlock(&trans->transaction->cache_write_mutex);
3621                 mutex_lock(&trans->transaction->cache_write_mutex);
3622         }
3623         mutex_unlock(&trans->transaction->cache_write_mutex);
3624
3625         /*
3626          * go through delayed refs for all the stuff we've just kicked off
3627          * and then loop back (just once)
3628          */
3629         ret = btrfs_run_delayed_refs(trans, 0);
3630         if (!ret && loops == 0) {
3631                 loops++;
3632                 spin_lock(&cur_trans->dirty_bgs_lock);
3633                 list_splice_init(&cur_trans->dirty_bgs, &dirty);
3634                 /*
3635                  * dirty_bgs_lock protects us from concurrent block group
3636                  * deletes too (not just cache_write_mutex).
3637                  */
3638                 if (!list_empty(&dirty)) {
3639                         spin_unlock(&cur_trans->dirty_bgs_lock);
3640                         goto again;
3641                 }
3642                 spin_unlock(&cur_trans->dirty_bgs_lock);
3643         } else if (ret < 0) {
3644                 btrfs_cleanup_dirty_bgs(cur_trans, fs_info);
3645         }
3646
3647         btrfs_free_path(path);
3648         return ret;
3649 }
3650
3651 int btrfs_write_dirty_block_groups(struct btrfs_trans_handle *trans)
3652 {
3653         struct btrfs_fs_info *fs_info = trans->fs_info;
3654         struct btrfs_block_group_cache *cache;
3655         struct btrfs_transaction *cur_trans = trans->transaction;
3656         int ret = 0;
3657         int should_put;
3658         struct btrfs_path *path;
3659         struct list_head *io = &cur_trans->io_bgs;
3660         int num_started = 0;
3661
3662         path = btrfs_alloc_path();
3663         if (!path)
3664                 return -ENOMEM;
3665
3666         /*
3667          * Even though we are in the critical section of the transaction commit,
3668          * we can still have concurrent tasks adding elements to this
3669          * transaction's list of dirty block groups. These tasks correspond to
3670          * endio free space workers started when writeback finishes for a
3671          * space cache, which run inode.c:btrfs_finish_ordered_io(), and can
3672          * allocate new block groups as a result of COWing nodes of the root
3673          * tree when updating the free space inode. The writeback for the space
3674          * caches is triggered by an earlier call to
3675          * btrfs_start_dirty_block_groups() and iterations of the following
3676          * loop.
3677          * Also we want to do the cache_save_setup first and then run the
3678          * delayed refs to make sure we have the best chance at doing this all
3679          * in one shot.
3680          */
3681         spin_lock(&cur_trans->dirty_bgs_lock);
3682         while (!list_empty(&cur_trans->dirty_bgs)) {
3683                 cache = list_first_entry(&cur_trans->dirty_bgs,
3684                                          struct btrfs_block_group_cache,
3685                                          dirty_list);
3686
3687                 /*
3688                  * this can happen if cache_save_setup re-dirties a block
3689                  * group that is already under IO.  Just wait for it to
3690                  * finish and then do it all again
3691                  */
3692                 if (!list_empty(&cache->io_list)) {
3693                         spin_unlock(&cur_trans->dirty_bgs_lock);
3694                         list_del_init(&cache->io_list);
3695                         btrfs_wait_cache_io(trans, cache, path);
3696                         btrfs_put_block_group(cache);
3697                         spin_lock(&cur_trans->dirty_bgs_lock);
3698                 }
3699
3700                 /*
3701                  * don't remove from the dirty list until after we've waited
3702                  * on any pending IO
3703                  */
3704                 list_del_init(&cache->dirty_list);
3705                 spin_unlock(&cur_trans->dirty_bgs_lock);
3706                 should_put = 1;
3707
3708                 cache_save_setup(cache, trans, path);
3709
3710                 if (!ret)
3711                         ret = btrfs_run_delayed_refs(trans,
3712                                                      (unsigned long) -1);
3713
3714                 if (!ret && cache->disk_cache_state == BTRFS_DC_SETUP) {
3715                         cache->io_ctl.inode = NULL;
3716                         ret = btrfs_write_out_cache(trans, cache, path);
3717                         if (ret == 0 && cache->io_ctl.inode) {
3718                                 num_started++;
3719                                 should_put = 0;
3720                                 list_add_tail(&cache->io_list, io);
3721                         } else {
3722                                 /*
3723                                  * if we failed to write the cache, the
3724                                  * generation will be bad and life goes on
3725                                  */
3726                                 ret = 0;
3727                         }
3728                 }
3729                 if (!ret) {
3730                         ret = write_one_cache_group(trans, path, cache);
3731                         /*
3732                          * One of the free space endio workers might have
3733                          * created a new block group while updating a free space
3734                          * cache's inode (at inode.c:btrfs_finish_ordered_io())
3735                          * and hasn't released its transaction handle yet, in
3736                          * which case the new block group is still attached to
3737                          * its transaction handle and its creation has not
3738                          * finished yet (no block group item in the extent tree
3739                          * yet, etc). If this is the case, wait for all free
3740                          * space endio workers to finish and retry. This is a
3741                          * a very rare case so no need for a more efficient and
3742                          * complex approach.
3743                          */
3744                         if (ret == -ENOENT) {
3745                                 wait_event(cur_trans->writer_wait,
3746                                    atomic_read(&cur_trans->num_writers) == 1);
3747                                 ret = write_one_cache_group(trans, path, cache);
3748                         }
3749                         if (ret)
3750                                 btrfs_abort_transaction(trans, ret);
3751                 }
3752
3753                 /* if its not on the io list, we need to put the block group */
3754                 if (should_put)
3755                         btrfs_put_block_group(cache);
3756                 btrfs_delayed_refs_rsv_release(fs_info, 1);
3757                 spin_lock(&cur_trans->dirty_bgs_lock);
3758         }
3759         spin_unlock(&cur_trans->dirty_bgs_lock);
3760
3761         /*
3762          * Refer to the definition of io_bgs member for details why it's safe
3763          * to use it without any locking
3764          */
3765         while (!list_empty(io)) {
3766                 cache = list_first_entry(io, struct btrfs_block_group_cache,
3767                                          io_list);
3768                 list_del_init(&cache->io_list);
3769                 btrfs_wait_cache_io(trans, cache, path);
3770                 btrfs_put_block_group(cache);
3771         }
3772
3773         btrfs_free_path(path);
3774         return ret;
3775 }
3776
3777 int btrfs_extent_readonly(struct btrfs_fs_info *fs_info, u64 bytenr)
3778 {
3779         struct btrfs_block_group_cache *block_group;
3780         int readonly = 0;
3781
3782         block_group = btrfs_lookup_block_group(fs_info, bytenr);
3783         if (!block_group || block_group->ro)
3784                 readonly = 1;
3785         if (block_group)
3786                 btrfs_put_block_group(block_group);
3787         return readonly;
3788 }
3789
3790 bool btrfs_inc_nocow_writers(struct btrfs_fs_info *fs_info, u64 bytenr)
3791 {
3792         struct btrfs_block_group_cache *bg;
3793         bool ret = true;
3794
3795         bg = btrfs_lookup_block_group(fs_info, bytenr);
3796         if (!bg)
3797                 return false;
3798
3799         spin_lock(&bg->lock);
3800         if (bg->ro)
3801                 ret = false;
3802         else
3803                 atomic_inc(&bg->nocow_writers);
3804         spin_unlock(&bg->lock);
3805
3806         /* no put on block group, done by btrfs_dec_nocow_writers */
3807         if (!ret)
3808                 btrfs_put_block_group(bg);
3809
3810         return ret;
3811
3812 }
3813
3814 void btrfs_dec_nocow_writers(struct btrfs_fs_info *fs_info, u64 bytenr)
3815 {
3816         struct btrfs_block_group_cache *bg;
3817
3818         bg = btrfs_lookup_block_group(fs_info, bytenr);
3819         ASSERT(bg);
3820         if (atomic_dec_and_test(&bg->nocow_writers))
3821                 wake_up_var(&bg->nocow_writers);
3822         /*
3823          * Once for our lookup and once for the lookup done by a previous call
3824          * to btrfs_inc_nocow_writers()
3825          */
3826         btrfs_put_block_group(bg);
3827         btrfs_put_block_group(bg);
3828 }
3829
3830 void btrfs_wait_nocow_writers(struct btrfs_block_group_cache *bg)
3831 {
3832         wait_var_event(&bg->nocow_writers, !atomic_read(&bg->nocow_writers));
3833 }
3834
3835 static const char *alloc_name(u64 flags)
3836 {
3837         switch (flags) {
3838         case BTRFS_BLOCK_GROUP_METADATA|BTRFS_BLOCK_GROUP_DATA:
3839                 return "mixed";
3840         case BTRFS_BLOCK_GROUP_METADATA:
3841                 return "metadata";
3842         case BTRFS_BLOCK_GROUP_DATA:
3843                 return "data";
3844         case BTRFS_BLOCK_GROUP_SYSTEM:
3845                 return "system";
3846         default:
3847                 WARN_ON(1);
3848                 return "invalid-combination";
3849         };
3850 }
3851
3852 static int create_space_info(struct btrfs_fs_info *info, u64 flags)
3853 {
3854
3855         struct btrfs_space_info *space_info;
3856         int i;
3857         int ret;
3858
3859         space_info = kzalloc(sizeof(*space_info), GFP_NOFS);
3860         if (!space_info)
3861                 return -ENOMEM;
3862
3863         ret = percpu_counter_init(&space_info->total_bytes_pinned, 0,
3864                                  GFP_KERNEL);
3865         if (ret) {
3866                 kfree(space_info);
3867                 return ret;
3868         }
3869
3870         for (i = 0; i < BTRFS_NR_RAID_TYPES; i++)
3871                 INIT_LIST_HEAD(&space_info->block_groups[i]);
3872         init_rwsem(&space_info->groups_sem);
3873         spin_lock_init(&space_info->lock);
3874         space_info->flags = flags & BTRFS_BLOCK_GROUP_TYPE_MASK;
3875         space_info->force_alloc = CHUNK_ALLOC_NO_FORCE;
3876         init_waitqueue_head(&space_info->wait);
3877         INIT_LIST_HEAD(&space_info->ro_bgs);
3878         INIT_LIST_HEAD(&space_info->tickets);
3879         INIT_LIST_HEAD(&space_info->priority_tickets);
3880
3881         ret = kobject_init_and_add(&space_info->kobj, &space_info_ktype,
3882                                     info->space_info_kobj, "%s",
3883                                     alloc_name(space_info->flags));
3884         if (ret) {
3885                 percpu_counter_destroy(&space_info->total_bytes_pinned);
3886                 kfree(space_info);
3887                 return ret;
3888         }
3889
3890         list_add_rcu(&space_info->list, &info->space_info);
3891         if (flags & BTRFS_BLOCK_GROUP_DATA)
3892                 info->data_sinfo = space_info;
3893
3894         return ret;
3895 }
3896
3897 static void update_space_info(struct btrfs_fs_info *info, u64 flags,
3898                              u64 total_bytes, u64 bytes_used,
3899                              u64 bytes_readonly,
3900                              struct btrfs_space_info **space_info)
3901 {
3902         struct btrfs_space_info *found;
3903         int factor;
3904
3905         factor = btrfs_bg_type_to_factor(flags);
3906
3907         found = __find_space_info(info, flags);
3908         ASSERT(found);
3909         spin_lock(&found->lock);
3910         found->total_bytes += total_bytes;
3911         found->disk_total += total_bytes * factor;
3912         found->bytes_used += bytes_used;
3913         found->disk_used += bytes_used * factor;
3914         found->bytes_readonly += bytes_readonly;
3915         if (total_bytes > 0)
3916                 found->full = 0;
3917         space_info_add_new_bytes(info, found, total_bytes -
3918                                  bytes_used - bytes_readonly);
3919         spin_unlock(&found->lock);
3920         *space_info = found;
3921 }
3922
3923 static void set_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
3924 {
3925         u64 extra_flags = chunk_to_extended(flags) &
3926                                 BTRFS_EXTENDED_PROFILE_MASK;
3927
3928         write_seqlock(&fs_info->profiles_lock);
3929         if (flags & BTRFS_BLOCK_GROUP_DATA)
3930                 fs_info->avail_data_alloc_bits |= extra_flags;
3931         if (flags & BTRFS_BLOCK_GROUP_METADATA)
3932                 fs_info->avail_metadata_alloc_bits |= extra_flags;
3933         if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
3934                 fs_info->avail_system_alloc_bits |= extra_flags;
3935         write_sequnlock(&fs_info->profiles_lock);
3936 }
3937
3938 /*
3939  * returns target flags in extended format or 0 if restripe for this
3940  * chunk_type is not in progress
3941  *
3942  * should be called with balance_lock held
3943  */
3944 static u64 get_restripe_target(struct btrfs_fs_info *fs_info, u64 flags)
3945 {
3946         struct btrfs_balance_control *bctl = fs_info->balance_ctl;
3947         u64 target = 0;
3948
3949         if (!bctl)
3950                 return 0;
3951
3952         if (flags & BTRFS_BLOCK_GROUP_DATA &&
3953             bctl->data.flags & BTRFS_BALANCE_ARGS_CONVERT) {
3954                 target = BTRFS_BLOCK_GROUP_DATA | bctl->data.target;
3955         } else if (flags & BTRFS_BLOCK_GROUP_SYSTEM &&
3956                    bctl->sys.flags & BTRFS_BALANCE_ARGS_CONVERT) {
3957                 target = BTRFS_BLOCK_GROUP_SYSTEM | bctl->sys.target;
3958         } else if (flags & BTRFS_BLOCK_GROUP_METADATA &&
3959                    bctl->meta.flags & BTRFS_BALANCE_ARGS_CONVERT) {
3960                 target = BTRFS_BLOCK_GROUP_METADATA | bctl->meta.target;
3961         }
3962
3963         return target;
3964 }
3965
3966 /*
3967  * @flags: available profiles in extended format (see ctree.h)
3968  *
3969  * Returns reduced profile in chunk format.  If profile changing is in
3970  * progress (either running or paused) picks the target profile (if it's
3971  * already available), otherwise falls back to plain reducing.
3972  */
3973 static u64 btrfs_reduce_alloc_profile(struct btrfs_fs_info *fs_info, u64 flags)
3974 {
3975         u64 num_devices = fs_info->fs_devices->rw_devices;
3976         u64 target;
3977         u64 raid_type;
3978         u64 allowed = 0;
3979
3980         /*
3981          * see if restripe for this chunk_type is in progress, if so
3982          * try to reduce to the target profile
3983          */
3984         spin_lock(&fs_info->balance_lock);
3985         target = get_restripe_target(fs_info, flags);
3986         if (target) {
3987                 /* pick target profile only if it's already available */
3988                 if ((flags & target) & BTRFS_EXTENDED_PROFILE_MASK) {
3989                         spin_unlock(&fs_info->balance_lock);
3990                         return extended_to_chunk(target);
3991                 }
3992         }
3993         spin_unlock(&fs_info->balance_lock);
3994
3995         /* First, mask out the RAID levels which aren't possible */
3996         for (raid_type = 0; raid_type < BTRFS_NR_RAID_TYPES; raid_type++) {
3997                 if (num_devices >= btrfs_raid_array[raid_type].devs_min)
3998                         allowed |= btrfs_raid_array[raid_type].bg_flag;
3999         }
4000         allowed &= flags;
4001
4002         if (allowed & BTRFS_BLOCK_GROUP_RAID6)
4003                 allowed = BTRFS_BLOCK_GROUP_RAID6;
4004         else if (allowed & BTRFS_BLOCK_GROUP_RAID5)
4005                 allowed = BTRFS_BLOCK_GROUP_RAID5;
4006         else if (allowed & BTRFS_BLOCK_GROUP_RAID10)
4007                 allowed = BTRFS_BLOCK_GROUP_RAID10;
4008         else if (allowed & BTRFS_BLOCK_GROUP_RAID1)
4009                 allowed = BTRFS_BLOCK_GROUP_RAID1;
4010         else if (allowed & BTRFS_BLOCK_GROUP_RAID0)
4011                 allowed = BTRFS_BLOCK_GROUP_RAID0;
4012
4013         flags &= ~BTRFS_BLOCK_GROUP_PROFILE_MASK;
4014
4015         return extended_to_chunk(flags | allowed);
4016 }
4017
4018 static u64 get_alloc_profile(struct btrfs_fs_info *fs_info, u64 orig_flags)
4019 {
4020         unsigned seq;
4021         u64 flags;
4022
4023         do {
4024                 flags = orig_flags;
4025                 seq = read_seqbegin(&fs_info->profiles_lock);
4026
4027                 if (flags & BTRFS_BLOCK_GROUP_DATA)
4028                         flags |= fs_info->avail_data_alloc_bits;
4029                 else if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
4030                         flags |= fs_info->avail_system_alloc_bits;
4031                 else if (flags & BTRFS_BLOCK_GROUP_METADATA)
4032                         flags |= fs_info->avail_metadata_alloc_bits;
4033         } while (read_seqretry(&fs_info->profiles_lock, seq));
4034
4035         return btrfs_reduce_alloc_profile(fs_info, flags);
4036 }
4037
4038 static u64 get_alloc_profile_by_root(struct btrfs_root *root, int data)
4039 {
4040         struct btrfs_fs_info *fs_info = root->fs_info;
4041         u64 flags;
4042         u64 ret;
4043
4044         if (data)
4045                 flags = BTRFS_BLOCK_GROUP_DATA;
4046         else if (root == fs_info->chunk_root)
4047                 flags = BTRFS_BLOCK_GROUP_SYSTEM;
4048         else
4049                 flags = BTRFS_BLOCK_GROUP_METADATA;
4050
4051         ret = get_alloc_profile(fs_info, flags);
4052         return ret;
4053 }
4054
4055 u64 btrfs_data_alloc_profile(struct btrfs_fs_info *fs_info)
4056 {
4057         return get_alloc_profile(fs_info, BTRFS_BLOCK_GROUP_DATA);
4058 }
4059
4060 u64 btrfs_metadata_alloc_profile(struct btrfs_fs_info *fs_info)
4061 {
4062         return get_alloc_profile(fs_info, BTRFS_BLOCK_GROUP_METADATA);
4063 }
4064
4065 u64 btrfs_system_alloc_profile(struct btrfs_fs_info *fs_info)
4066 {
4067         return get_alloc_profile(fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
4068 }
4069
4070 static u64 btrfs_space_info_used(struct btrfs_space_info *s_info,
4071                                  bool may_use_included)
4072 {
4073         ASSERT(s_info);
4074         return s_info->bytes_used + s_info->bytes_reserved +
4075                 s_info->bytes_pinned + s_info->bytes_readonly +
4076                 (may_use_included ? s_info->bytes_may_use : 0);
4077 }
4078
4079 int btrfs_alloc_data_chunk_ondemand(struct btrfs_inode *inode, u64 bytes)
4080 {
4081         struct btrfs_root *root = inode->root;
4082         struct btrfs_fs_info *fs_info = root->fs_info;
4083         struct btrfs_space_info *data_sinfo = fs_info->data_sinfo;
4084         u64 used;
4085         int ret = 0;
4086         int need_commit = 2;
4087         int have_pinned_space;
4088
4089         /* make sure bytes are sectorsize aligned */
4090         bytes = ALIGN(bytes, fs_info->sectorsize);
4091
4092         if (btrfs_is_free_space_inode(inode)) {
4093                 need_commit = 0;
4094                 ASSERT(current->journal_info);
4095         }
4096
4097 again:
4098         /* make sure we have enough space to handle the data first */
4099         spin_lock(&data_sinfo->lock);
4100         used = btrfs_space_info_used(data_sinfo, true);
4101
4102         if (used + bytes > data_sinfo->total_bytes) {
4103                 struct btrfs_trans_handle *trans;
4104
4105                 /*
4106                  * if we don't have enough free bytes in this space then we need
4107                  * to alloc a new chunk.
4108                  */
4109                 if (!data_sinfo->full) {
4110                         u64 alloc_target;
4111
4112                         data_sinfo->force_alloc = CHUNK_ALLOC_FORCE;
4113                         spin_unlock(&data_sinfo->lock);
4114
4115                         alloc_target = btrfs_data_alloc_profile(fs_info);
4116                         /*
4117                          * It is ugly that we don't call nolock join
4118                          * transaction for the free space inode case here.
4119                          * But it is safe because we only do the data space
4120                          * reservation for the free space cache in the
4121                          * transaction context, the common join transaction
4122                          * just increase the counter of the current transaction
4123                          * handler, doesn't try to acquire the trans_lock of
4124                          * the fs.
4125                          */
4126                         trans = btrfs_join_transaction(root);
4127                         if (IS_ERR(trans))
4128                                 return PTR_ERR(trans);
4129
4130                         ret = do_chunk_alloc(trans, alloc_target,
4131                                              CHUNK_ALLOC_NO_FORCE);
4132                         btrfs_end_transaction(trans);
4133                         if (ret < 0) {
4134                                 if (ret != -ENOSPC)
4135                                         return ret;
4136                                 else {
4137                                         have_pinned_space = 1;
4138                                         goto commit_trans;
4139                                 }
4140                         }
4141
4142                         goto again;
4143                 }
4144
4145                 /*
4146                  * If we don't have enough pinned space to deal with this
4147                  * allocation, and no removed chunk in current transaction,
4148                  * don't bother committing the transaction.
4149                  */
4150                 have_pinned_space = __percpu_counter_compare(
4151                         &data_sinfo->total_bytes_pinned,
4152                         used + bytes - data_sinfo->total_bytes,
4153                         BTRFS_TOTAL_BYTES_PINNED_BATCH);
4154                 spin_unlock(&data_sinfo->lock);
4155
4156                 /* commit the current transaction and try again */
4157 commit_trans:
4158                 if (need_commit) {
4159                         need_commit--;
4160
4161                         if (need_commit > 0) {
4162                                 btrfs_start_delalloc_roots(fs_info, -1);
4163                                 btrfs_wait_ordered_roots(fs_info, U64_MAX, 0,
4164                                                          (u64)-1);
4165                         }
4166
4167                         trans = btrfs_join_transaction(root);
4168                         if (IS_ERR(trans))
4169                                 return PTR_ERR(trans);
4170                         if (have_pinned_space >= 0 ||
4171                             test_bit(BTRFS_TRANS_HAVE_FREE_BGS,
4172                                      &trans->transaction->flags) ||
4173                             need_commit > 0) {
4174                                 ret = btrfs_commit_transaction(trans);
4175                                 if (ret)
4176                                         return ret;
4177                                 /*
4178                                  * The cleaner kthread might still be doing iput
4179                                  * operations. Wait for it to finish so that
4180                                  * more space is released.  We don't need to
4181                                  * explicitly run the delayed iputs here because
4182                                  * the commit_transaction would have woken up
4183                                  * the cleaner.
4184                                  */
4185                                 ret = btrfs_wait_on_delayed_iputs(fs_info);
4186                                 if (ret)
4187                                         return ret;
4188                                 goto again;
4189                         } else {
4190                                 btrfs_end_transaction(trans);
4191                         }
4192                 }
4193
4194                 trace_btrfs_space_reservation(fs_info,
4195                                               "space_info:enospc",
4196                                               data_sinfo->flags, bytes, 1);
4197                 return -ENOSPC;
4198         }
4199         update_bytes_may_use(data_sinfo, bytes);
4200         trace_btrfs_space_reservation(fs_info, "space_info",
4201                                       data_sinfo->flags, bytes, 1);
4202         spin_unlock(&data_sinfo->lock);
4203
4204         return 0;
4205 }
4206
4207 int btrfs_check_data_free_space(struct inode *inode,
4208                         struct extent_changeset **reserved, u64 start, u64 len)
4209 {
4210         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
4211         int ret;
4212
4213         /* align the range */
4214         len = round_up(start + len, fs_info->sectorsize) -
4215               round_down(start, fs_info->sectorsize);
4216         start = round_down(start, fs_info->sectorsize);
4217
4218         ret = btrfs_alloc_data_chunk_ondemand(BTRFS_I(inode), len);
4219         if (ret < 0)
4220                 return ret;
4221
4222         /* Use new btrfs_qgroup_reserve_data to reserve precious data space. */
4223         ret = btrfs_qgroup_reserve_data(inode, reserved, start, len);
4224         if (ret < 0)
4225                 btrfs_free_reserved_data_space_noquota(inode, start, len);
4226         else
4227                 ret = 0;
4228         return ret;
4229 }
4230
4231 /*
4232  * Called if we need to clear a data reservation for this inode
4233  * Normally in a error case.
4234  *
4235  * This one will *NOT* use accurate qgroup reserved space API, just for case
4236  * which we can't sleep and is sure it won't affect qgroup reserved space.
4237  * Like clear_bit_hook().
4238  */
4239 void btrfs_free_reserved_data_space_noquota(struct inode *inode, u64 start,
4240                                             u64 len)
4241 {
4242         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
4243         struct btrfs_space_info *data_sinfo;
4244
4245         /* Make sure the range is aligned to sectorsize */
4246         len = round_up(start + len, fs_info->sectorsize) -
4247               round_down(start, fs_info->sectorsize);
4248         start = round_down(start, fs_info->sectorsize);
4249
4250         data_sinfo = fs_info->data_sinfo;
4251         spin_lock(&data_sinfo->lock);
4252         update_bytes_may_use(data_sinfo, -len);
4253         trace_btrfs_space_reservation(fs_info, "space_info",
4254                                       data_sinfo->flags, len, 0);
4255         spin_unlock(&data_sinfo->lock);
4256 }
4257
4258 /*
4259  * Called if we need to clear a data reservation for this inode
4260  * Normally in a error case.
4261  *
4262  * This one will handle the per-inode data rsv map for accurate reserved
4263  * space framework.
4264  */
4265 void btrfs_free_reserved_data_space(struct inode *inode,
4266                         struct extent_changeset *reserved, u64 start, u64 len)
4267 {
4268         struct btrfs_root *root = BTRFS_I(inode)->root;
4269
4270         /* Make sure the range is aligned to sectorsize */
4271         len = round_up(start + len, root->fs_info->sectorsize) -
4272               round_down(start, root->fs_info->sectorsize);
4273         start = round_down(start, root->fs_info->sectorsize);
4274
4275         btrfs_free_reserved_data_space_noquota(inode, start, len);
4276         btrfs_qgroup_free_data(inode, reserved, start, len);
4277 }
4278
4279 static void force_metadata_allocation(struct btrfs_fs_info *info)
4280 {
4281         struct list_head *head = &info->space_info;
4282         struct btrfs_space_info *found;
4283
4284         rcu_read_lock();
4285         list_for_each_entry_rcu(found, head, list) {
4286                 if (found->flags & BTRFS_BLOCK_GROUP_METADATA)
4287                         found->force_alloc = CHUNK_ALLOC_FORCE;
4288         }
4289         rcu_read_unlock();
4290 }
4291
4292 static inline u64 calc_global_rsv_need_space(struct btrfs_block_rsv *global)
4293 {
4294         return (global->size << 1);
4295 }
4296
4297 static int should_alloc_chunk(struct btrfs_fs_info *fs_info,
4298                               struct btrfs_space_info *sinfo, int force)
4299 {
4300         u64 bytes_used = btrfs_space_info_used(sinfo, false);
4301         u64 thresh;
4302
4303         if (force == CHUNK_ALLOC_FORCE)
4304                 return 1;
4305
4306         /*
4307          * in limited mode, we want to have some free space up to
4308          * about 1% of the FS size.
4309          */
4310         if (force == CHUNK_ALLOC_LIMITED) {
4311                 thresh = btrfs_super_total_bytes(fs_info->super_copy);
4312                 thresh = max_t(u64, SZ_64M, div_factor_fine(thresh, 1));
4313
4314                 if (sinfo->total_bytes - bytes_used < thresh)
4315                         return 1;
4316         }
4317
4318         if (bytes_used + SZ_2M < div_factor(sinfo->total_bytes, 8))
4319                 return 0;
4320         return 1;
4321 }
4322
4323 static u64 get_profile_num_devs(struct btrfs_fs_info *fs_info, u64 type)
4324 {
4325         u64 num_dev;
4326
4327         if (type & (BTRFS_BLOCK_GROUP_RAID10 |
4328                     BTRFS_BLOCK_GROUP_RAID0 |
4329                     BTRFS_BLOCK_GROUP_RAID5 |
4330                     BTRFS_BLOCK_GROUP_RAID6))
4331                 num_dev = fs_info->fs_devices->rw_devices;
4332         else if (type & BTRFS_BLOCK_GROUP_RAID1)
4333                 num_dev = 2;
4334         else
4335                 num_dev = 1;    /* DUP or single */
4336
4337         return num_dev;
4338 }
4339
4340 /*
4341  * If @is_allocation is true, reserve space in the system space info necessary
4342  * for allocating a chunk, otherwise if it's false, reserve space necessary for
4343  * removing a chunk.
4344  */
4345 void check_system_chunk(struct btrfs_trans_handle *trans, u64 type)
4346 {
4347         struct btrfs_fs_info *fs_info = trans->fs_info;
4348         struct btrfs_space_info *info;
4349         u64 left;
4350         u64 thresh;
4351         int ret = 0;
4352         u64 num_devs;
4353
4354         /*
4355          * Needed because we can end up allocating a system chunk and for an
4356          * atomic and race free space reservation in the chunk block reserve.
4357          */
4358         lockdep_assert_held(&fs_info->chunk_mutex);
4359
4360         info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
4361         spin_lock(&info->lock);
4362         left = info->total_bytes - btrfs_space_info_used(info, true);
4363         spin_unlock(&info->lock);
4364
4365         num_devs = get_profile_num_devs(fs_info, type);
4366
4367         /* num_devs device items to update and 1 chunk item to add or remove */
4368         thresh = btrfs_calc_trunc_metadata_size(fs_info, num_devs) +
4369                 btrfs_calc_trans_metadata_size(fs_info, 1);
4370
4371         if (left < thresh && btrfs_test_opt(fs_info, ENOSPC_DEBUG)) {
4372                 btrfs_info(fs_info, "left=%llu, need=%llu, flags=%llu",
4373                            left, thresh, type);
4374                 dump_space_info(fs_info, info, 0, 0);
4375         }
4376
4377         if (left < thresh) {
4378                 u64 flags = btrfs_system_alloc_profile(fs_info);
4379
4380                 /*
4381                  * Ignore failure to create system chunk. We might end up not
4382                  * needing it, as we might not need to COW all nodes/leafs from
4383                  * the paths we visit in the chunk tree (they were already COWed
4384                  * or created in the current transaction for example).
4385                  */
4386                 ret = btrfs_alloc_chunk(trans, flags);
4387         }
4388
4389         if (!ret) {
4390                 ret = btrfs_block_rsv_add(fs_info->chunk_root,
4391                                           &fs_info->chunk_block_rsv,
4392                                           thresh, BTRFS_RESERVE_NO_FLUSH);
4393                 if (!ret)
4394                         trans->chunk_bytes_reserved += thresh;
4395         }
4396 }
4397
4398 /*
4399  * If force is CHUNK_ALLOC_FORCE:
4400  *    - return 1 if it successfully allocates a chunk,
4401  *    - return errors including -ENOSPC otherwise.
4402  * If force is NOT CHUNK_ALLOC_FORCE:
4403  *    - return 0 if it doesn't need to allocate a new chunk,
4404  *    - return 1 if it successfully allocates a chunk,
4405  *    - return errors including -ENOSPC otherwise.
4406  */
4407 static int do_chunk_alloc(struct btrfs_trans_handle *trans, u64 flags,
4408                           int force)
4409 {
4410         struct btrfs_fs_info *fs_info = trans->fs_info;
4411         struct btrfs_space_info *space_info;
4412         bool wait_for_alloc = false;
4413         bool should_alloc = false;
4414         int ret = 0;
4415
4416         /* Don't re-enter if we're already allocating a chunk */
4417         if (trans->allocating_chunk)
4418                 return -ENOSPC;
4419
4420         space_info = __find_space_info(fs_info, flags);
4421         ASSERT(space_info);
4422
4423         do {
4424                 spin_lock(&space_info->lock);
4425                 if (force < space_info->force_alloc)
4426                         force = space_info->force_alloc;
4427                 should_alloc = should_alloc_chunk(fs_info, space_info, force);
4428                 if (space_info->full) {
4429                         /* No more free physical space */
4430                         if (should_alloc)
4431                                 ret = -ENOSPC;
4432                         else
4433                                 ret = 0;
4434                         spin_unlock(&space_info->lock);
4435                         return ret;
4436                 } else if (!should_alloc) {
4437                         spin_unlock(&space_info->lock);
4438                         return 0;
4439                 } else if (space_info->chunk_alloc) {
4440                         /*
4441                          * Someone is already allocating, so we need to block
4442                          * until this someone is finished and then loop to
4443                          * recheck if we should continue with our allocation
4444                          * attempt.
4445                          */
4446                         wait_for_alloc = true;
4447                         spin_unlock(&space_info->lock);
4448                         mutex_lock(&fs_info->chunk_mutex);
4449                         mutex_unlock(&fs_info->chunk_mutex);
4450                 } else {
4451                         /* Proceed with allocation */
4452                         space_info->chunk_alloc = 1;
4453                         wait_for_alloc = false;
4454                         spin_unlock(&space_info->lock);
4455                 }
4456
4457                 cond_resched();
4458         } while (wait_for_alloc);
4459
4460         mutex_lock(&fs_info->chunk_mutex);
4461         trans->allocating_chunk = true;
4462
4463         /*
4464          * If we have mixed data/metadata chunks we want to make sure we keep
4465          * allocating mixed chunks instead of individual chunks.
4466          */
4467         if (btrfs_mixed_space_info(space_info))
4468                 flags |= (BTRFS_BLOCK_GROUP_DATA | BTRFS_BLOCK_GROUP_METADATA);
4469
4470         /*
4471          * if we're doing a data chunk, go ahead and make sure that
4472          * we keep a reasonable number of metadata chunks allocated in the
4473          * FS as well.
4474          */
4475         if (flags & BTRFS_BLOCK_GROUP_DATA && fs_info->metadata_ratio) {
4476                 fs_info->data_chunk_allocations++;
4477                 if (!(fs_info->data_chunk_allocations %
4478                       fs_info->metadata_ratio))
4479                         force_metadata_allocation(fs_info);
4480         }
4481
4482         /*
4483          * Check if we have enough space in SYSTEM chunk because we may need
4484          * to update devices.
4485          */
4486         check_system_chunk(trans, flags);
4487
4488         ret = btrfs_alloc_chunk(trans, flags);
4489         trans->allocating_chunk = false;
4490
4491         spin_lock(&space_info->lock);
4492         if (ret < 0) {
4493                 if (ret == -ENOSPC)
4494                         space_info->full = 1;
4495                 else
4496                         goto out;
4497         } else {
4498                 ret = 1;
4499                 space_info->max_extent_size = 0;
4500         }
4501
4502         space_info->force_alloc = CHUNK_ALLOC_NO_FORCE;
4503 out:
4504         space_info->chunk_alloc = 0;
4505         spin_unlock(&space_info->lock);
4506         mutex_unlock(&fs_info->chunk_mutex);
4507         /*
4508          * When we allocate a new chunk we reserve space in the chunk block
4509          * reserve to make sure we can COW nodes/leafs in the chunk tree or
4510          * add new nodes/leafs to it if we end up needing to do it when
4511          * inserting the chunk item and updating device items as part of the
4512          * second phase of chunk allocation, performed by
4513          * btrfs_finish_chunk_alloc(). So make sure we don't accumulate a
4514          * large number of new block groups to create in our transaction
4515          * handle's new_bgs list to avoid exhausting the chunk block reserve
4516          * in extreme cases - like having a single transaction create many new
4517          * block groups when starting to write out the free space caches of all
4518          * the block groups that were made dirty during the lifetime of the
4519          * transaction.
4520          */
4521         if (trans->chunk_bytes_reserved >= (u64)SZ_2M)
4522                 btrfs_create_pending_block_groups(trans);
4523
4524         return ret;
4525 }
4526
4527 static int can_overcommit(struct btrfs_fs_info *fs_info,
4528                           struct btrfs_space_info *space_info, u64 bytes,
4529                           enum btrfs_reserve_flush_enum flush,
4530                           bool system_chunk)
4531 {
4532         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
4533         u64 profile;
4534         u64 space_size;
4535         u64 avail;
4536         u64 used;
4537         int factor;
4538
4539         /* Don't overcommit when in mixed mode. */
4540         if (space_info->flags & BTRFS_BLOCK_GROUP_DATA)
4541                 return 0;
4542
4543         if (system_chunk)
4544                 profile = btrfs_system_alloc_profile(fs_info);
4545         else
4546                 profile = btrfs_metadata_alloc_profile(fs_info);
4547
4548         used = btrfs_space_info_used(space_info, false);
4549
4550         /*
4551          * We only want to allow over committing if we have lots of actual space
4552          * free, but if we don't have enough space to handle the global reserve
4553          * space then we could end up having a real enospc problem when trying
4554          * to allocate a chunk or some other such important allocation.
4555          */
4556         spin_lock(&global_rsv->lock);
4557         space_size = calc_global_rsv_need_space(global_rsv);
4558         spin_unlock(&global_rsv->lock);
4559         if (used + space_size >= space_info->total_bytes)
4560                 return 0;
4561
4562         used += space_info->bytes_may_use;
4563
4564         avail = atomic64_read(&fs_info->free_chunk_space);
4565
4566         /*
4567          * If we have dup, raid1 or raid10 then only half of the free
4568          * space is actually usable.  For raid56, the space info used
4569          * doesn't include the parity drive, so we don't have to
4570          * change the math
4571          */
4572         factor = btrfs_bg_type_to_factor(profile);
4573         avail = div_u64(avail, factor);
4574
4575         /*
4576          * If we aren't flushing all things, let us overcommit up to
4577          * 1/2th of the space. If we can flush, don't let us overcommit
4578          * too much, let it overcommit up to 1/8 of the space.
4579          */
4580         if (flush == BTRFS_RESERVE_FLUSH_ALL)
4581                 avail >>= 3;
4582         else
4583                 avail >>= 1;
4584
4585         if (used + bytes < space_info->total_bytes + avail)
4586                 return 1;
4587         return 0;
4588 }
4589
4590 static void btrfs_writeback_inodes_sb_nr(struct btrfs_fs_info *fs_info,
4591                                          unsigned long nr_pages, int nr_items)
4592 {
4593         struct super_block *sb = fs_info->sb;
4594
4595         if (down_read_trylock(&sb->s_umount)) {
4596                 writeback_inodes_sb_nr(sb, nr_pages, WB_REASON_FS_FREE_SPACE);
4597                 up_read(&sb->s_umount);
4598         } else {
4599                 /*
4600                  * We needn't worry the filesystem going from r/w to r/o though
4601                  * we don't acquire ->s_umount mutex, because the filesystem
4602                  * should guarantee the delalloc inodes list be empty after
4603                  * the filesystem is readonly(all dirty pages are written to
4604                  * the disk).
4605                  */
4606                 btrfs_start_delalloc_roots(fs_info, nr_items);
4607                 if (!current->journal_info)
4608                         btrfs_wait_ordered_roots(fs_info, nr_items, 0, (u64)-1);
4609         }
4610 }
4611
4612 static inline u64 calc_reclaim_items_nr(struct btrfs_fs_info *fs_info,
4613                                         u64 to_reclaim)
4614 {
4615         u64 bytes;
4616         u64 nr;
4617
4618         bytes = btrfs_calc_trans_metadata_size(fs_info, 1);
4619         nr = div64_u64(to_reclaim, bytes);
4620         if (!nr)
4621                 nr = 1;
4622         return nr;
4623 }
4624
4625 #define EXTENT_SIZE_PER_ITEM    SZ_256K
4626
4627 /*
4628  * shrink metadata reservation for delalloc
4629  */
4630 static void shrink_delalloc(struct btrfs_fs_info *fs_info, u64 to_reclaim,
4631                             u64 orig, bool wait_ordered)
4632 {
4633         struct btrfs_space_info *space_info;
4634         struct btrfs_trans_handle *trans;
4635         u64 delalloc_bytes;
4636         u64 dio_bytes;
4637         u64 async_pages;
4638         u64 items;
4639         long time_left;
4640         unsigned long nr_pages;
4641         int loops;
4642
4643         /* Calc the number of the pages we need flush for space reservation */
4644         items = calc_reclaim_items_nr(fs_info, to_reclaim);
4645         to_reclaim = items * EXTENT_SIZE_PER_ITEM;
4646
4647         trans = (struct btrfs_trans_handle *)current->journal_info;
4648         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
4649
4650         delalloc_bytes = percpu_counter_sum_positive(
4651                                                 &fs_info->delalloc_bytes);
4652         dio_bytes = percpu_counter_sum_positive(&fs_info->dio_bytes);
4653         if (delalloc_bytes == 0 && dio_bytes == 0) {
4654                 if (trans)
4655                         return;
4656                 if (wait_ordered)
4657                         btrfs_wait_ordered_roots(fs_info, items, 0, (u64)-1);
4658                 return;
4659         }
4660
4661         /*
4662          * If we are doing more ordered than delalloc we need to just wait on
4663          * ordered extents, otherwise we'll waste time trying to flush delalloc
4664          * that likely won't give us the space back we need.
4665          */
4666         if (dio_bytes > delalloc_bytes)
4667                 wait_ordered = true;
4668
4669         loops = 0;
4670         while ((delalloc_bytes || dio_bytes) && loops < 3) {
4671                 nr_pages = min(delalloc_bytes, to_reclaim) >> PAGE_SHIFT;
4672
4673                 /*
4674                  * Triggers inode writeback for up to nr_pages. This will invoke
4675                  * ->writepages callback and trigger delalloc filling
4676                  *  (btrfs_run_delalloc_range()).
4677                  */
4678                 btrfs_writeback_inodes_sb_nr(fs_info, nr_pages, items);
4679
4680                 /*
4681                  * We need to wait for the compressed pages to start before
4682                  * we continue.
4683                  */
4684                 async_pages = atomic_read(&fs_info->async_delalloc_pages);
4685                 if (!async_pages)
4686                         goto skip_async;
4687
4688                 /*
4689                  * Calculate how many compressed pages we want to be written
4690                  * before we continue. I.e if there are more async pages than we
4691                  * require wait_event will wait until nr_pages are written.
4692                  */
4693                 if (async_pages <= nr_pages)
4694                         async_pages = 0;
4695                 else
4696                         async_pages -= nr_pages;
4697
4698                 wait_event(fs_info->async_submit_wait,
4699                            atomic_read(&fs_info->async_delalloc_pages) <=
4700                            (int)async_pages);
4701 skip_async:
4702                 spin_lock(&space_info->lock);
4703                 if (list_empty(&space_info->tickets) &&
4704                     list_empty(&space_info->priority_tickets)) {
4705                         spin_unlock(&space_info->lock);
4706                         break;
4707                 }
4708                 spin_unlock(&space_info->lock);
4709
4710                 loops++;
4711                 if (wait_ordered && !trans) {
4712                         btrfs_wait_ordered_roots(fs_info, items, 0, (u64)-1);
4713                 } else {
4714                         time_left = schedule_timeout_killable(1);
4715                         if (time_left)
4716                                 break;
4717                 }
4718                 delalloc_bytes = percpu_counter_sum_positive(
4719                                                 &fs_info->delalloc_bytes);
4720                 dio_bytes = percpu_counter_sum_positive(&fs_info->dio_bytes);
4721         }
4722 }
4723
4724 struct reserve_ticket {
4725         u64 orig_bytes;
4726         u64 bytes;
4727         int error;
4728         struct list_head list;
4729         wait_queue_head_t wait;
4730 };
4731
4732 /**
4733  * maybe_commit_transaction - possibly commit the transaction if its ok to
4734  * @root - the root we're allocating for
4735  * @bytes - the number of bytes we want to reserve
4736  * @force - force the commit
4737  *
4738  * This will check to make sure that committing the transaction will actually
4739  * get us somewhere and then commit the transaction if it does.  Otherwise it
4740  * will return -ENOSPC.
4741  */
4742 static int may_commit_transaction(struct btrfs_fs_info *fs_info,
4743                                   struct btrfs_space_info *space_info)
4744 {
4745         struct reserve_ticket *ticket = NULL;
4746         struct btrfs_block_rsv *delayed_rsv = &fs_info->delayed_block_rsv;
4747         struct btrfs_block_rsv *delayed_refs_rsv = &fs_info->delayed_refs_rsv;
4748         struct btrfs_trans_handle *trans;
4749         u64 bytes_needed;
4750         u64 reclaim_bytes = 0;
4751
4752         trans = (struct btrfs_trans_handle *)current->journal_info;
4753         if (trans)
4754                 return -EAGAIN;
4755
4756         spin_lock(&space_info->lock);
4757         if (!list_empty(&space_info->priority_tickets))
4758                 ticket = list_first_entry(&space_info->priority_tickets,
4759                                           struct reserve_ticket, list);
4760         else if (!list_empty(&space_info->tickets))
4761                 ticket = list_first_entry(&space_info->tickets,
4762                                           struct reserve_ticket, list);
4763         bytes_needed = (ticket) ? ticket->bytes : 0;
4764         spin_unlock(&space_info->lock);
4765
4766         if (!bytes_needed)
4767                 return 0;
4768
4769         trans = btrfs_join_transaction(fs_info->extent_root);
4770         if (IS_ERR(trans))
4771                 return PTR_ERR(trans);
4772
4773         /*
4774          * See if there is enough pinned space to make this reservation, or if
4775          * we have block groups that are going to be freed, allowing us to
4776          * possibly do a chunk allocation the next loop through.
4777          */
4778         if (test_bit(BTRFS_TRANS_HAVE_FREE_BGS, &trans->transaction->flags) ||
4779             __percpu_counter_compare(&space_info->total_bytes_pinned,
4780                                      bytes_needed,
4781                                      BTRFS_TOTAL_BYTES_PINNED_BATCH) >= 0)
4782                 goto commit;
4783
4784         /*
4785          * See if there is some space in the delayed insertion reservation for
4786          * this reservation.
4787          */
4788         if (space_info != delayed_rsv->space_info)
4789                 goto enospc;
4790
4791         spin_lock(&delayed_rsv->lock);
4792         reclaim_bytes += delayed_rsv->reserved;
4793         spin_unlock(&delayed_rsv->lock);
4794
4795         spin_lock(&delayed_refs_rsv->lock);
4796         reclaim_bytes += delayed_refs_rsv->reserved;
4797         spin_unlock(&delayed_refs_rsv->lock);
4798         if (reclaim_bytes >= bytes_needed)
4799                 goto commit;
4800         bytes_needed -= reclaim_bytes;
4801
4802         if (__percpu_counter_compare(&space_info->total_bytes_pinned,
4803                                    bytes_needed,
4804                                    BTRFS_TOTAL_BYTES_PINNED_BATCH) < 0)
4805                 goto enospc;
4806
4807 commit:
4808         return btrfs_commit_transaction(trans);
4809 enospc:
4810         btrfs_end_transaction(trans);
4811         return -ENOSPC;
4812 }
4813
4814 /*
4815  * Try to flush some data based on policy set by @state. This is only advisory
4816  * and may fail for various reasons. The caller is supposed to examine the
4817  * state of @space_info to detect the outcome.
4818  */
4819 static void flush_space(struct btrfs_fs_info *fs_info,
4820                        struct btrfs_space_info *space_info, u64 num_bytes,
4821                        int state)
4822 {
4823         struct btrfs_root *root = fs_info->extent_root;
4824         struct btrfs_trans_handle *trans;
4825         int nr;
4826         int ret = 0;
4827
4828         switch (state) {
4829         case FLUSH_DELAYED_ITEMS_NR:
4830         case FLUSH_DELAYED_ITEMS:
4831                 if (state == FLUSH_DELAYED_ITEMS_NR)
4832                         nr = calc_reclaim_items_nr(fs_info, num_bytes) * 2;
4833                 else
4834                         nr = -1;
4835
4836                 trans = btrfs_join_transaction(root);
4837                 if (IS_ERR(trans)) {
4838                         ret = PTR_ERR(trans);
4839                         break;
4840                 }
4841                 ret = btrfs_run_delayed_items_nr(trans, nr);
4842                 btrfs_end_transaction(trans);
4843                 break;
4844         case FLUSH_DELALLOC:
4845         case FLUSH_DELALLOC_WAIT:
4846                 shrink_delalloc(fs_info, num_bytes * 2, num_bytes,
4847                                 state == FLUSH_DELALLOC_WAIT);
4848                 break;
4849         case FLUSH_DELAYED_REFS_NR:
4850         case FLUSH_DELAYED_REFS:
4851                 trans = btrfs_join_transaction(root);
4852                 if (IS_ERR(trans)) {
4853                         ret = PTR_ERR(trans);
4854                         break;
4855                 }
4856                 if (state == FLUSH_DELAYED_REFS_NR)
4857                         nr = calc_reclaim_items_nr(fs_info, num_bytes);
4858                 else
4859                         nr = 0;
4860                 btrfs_run_delayed_refs(trans, nr);
4861                 btrfs_end_transaction(trans);
4862                 break;
4863         case ALLOC_CHUNK:
4864         case ALLOC_CHUNK_FORCE:
4865                 trans = btrfs_join_transaction(root);
4866                 if (IS_ERR(trans)) {
4867                         ret = PTR_ERR(trans);
4868                         break;
4869                 }
4870                 ret = do_chunk_alloc(trans,
4871                                      btrfs_metadata_alloc_profile(fs_info),
4872                                      (state == ALLOC_CHUNK) ?
4873                                       CHUNK_ALLOC_NO_FORCE : CHUNK_ALLOC_FORCE);
4874                 btrfs_end_transaction(trans);
4875                 if (ret > 0 || ret == -ENOSPC)
4876                         ret = 0;
4877                 break;
4878         case COMMIT_TRANS:
4879                 /*
4880                  * If we have pending delayed iputs then we could free up a
4881                  * bunch of pinned space, so make sure we run the iputs before
4882                  * we do our pinned bytes check below.
4883                  */
4884                 btrfs_run_delayed_iputs(fs_info);
4885                 btrfs_wait_on_delayed_iputs(fs_info);
4886
4887                 ret = may_commit_transaction(fs_info, space_info);
4888                 break;
4889         default:
4890                 ret = -ENOSPC;
4891                 break;
4892         }
4893
4894         trace_btrfs_flush_space(fs_info, space_info->flags, num_bytes, state,
4895                                 ret);
4896         return;
4897 }
4898
4899 static inline u64
4900 btrfs_calc_reclaim_metadata_size(struct btrfs_fs_info *fs_info,
4901                                  struct btrfs_space_info *space_info,
4902                                  bool system_chunk)
4903 {
4904         struct reserve_ticket *ticket;
4905         u64 used;
4906         u64 expected;
4907         u64 to_reclaim = 0;
4908
4909         list_for_each_entry(ticket, &space_info->tickets, list)
4910                 to_reclaim += ticket->bytes;
4911         list_for_each_entry(ticket, &space_info->priority_tickets, list)
4912                 to_reclaim += ticket->bytes;
4913         if (to_reclaim)
4914                 return to_reclaim;
4915
4916         to_reclaim = min_t(u64, num_online_cpus() * SZ_1M, SZ_16M);
4917         if (can_overcommit(fs_info, space_info, to_reclaim,
4918                            BTRFS_RESERVE_FLUSH_ALL, system_chunk))
4919                 return 0;
4920
4921         used = btrfs_space_info_used(space_info, true);
4922
4923         if (can_overcommit(fs_info, space_info, SZ_1M,
4924                            BTRFS_RESERVE_FLUSH_ALL, system_chunk))
4925                 expected = div_factor_fine(space_info->total_bytes, 95);
4926         else
4927                 expected = div_factor_fine(space_info->total_bytes, 90);
4928
4929         if (used > expected)
4930                 to_reclaim = used - expected;
4931         else
4932                 to_reclaim = 0;
4933         to_reclaim = min(to_reclaim, space_info->bytes_may_use +
4934                                      space_info->bytes_reserved);
4935         return to_reclaim;
4936 }
4937
4938 static inline int need_do_async_reclaim(struct btrfs_fs_info *fs_info,
4939                                         struct btrfs_space_info *space_info,
4940                                         u64 used, bool system_chunk)
4941 {
4942         u64 thresh = div_factor_fine(space_info->total_bytes, 98);
4943
4944         /* If we're just plain full then async reclaim just slows us down. */
4945         if ((space_info->bytes_used + space_info->bytes_reserved) >= thresh)
4946                 return 0;
4947
4948         if (!btrfs_calc_reclaim_metadata_size(fs_info, space_info,
4949                                               system_chunk))
4950                 return 0;
4951
4952         return (used >= thresh && !btrfs_fs_closing(fs_info) &&
4953                 !test_bit(BTRFS_FS_STATE_REMOUNTING, &fs_info->fs_state));
4954 }
4955
4956 static bool wake_all_tickets(struct list_head *head)
4957 {
4958         struct reserve_ticket *ticket;
4959
4960         while (!list_empty(head)) {
4961                 ticket = list_first_entry(head, struct reserve_ticket, list);
4962                 list_del_init(&ticket->list);
4963                 ticket->error = -ENOSPC;
4964                 wake_up(&ticket->wait);
4965                 if (ticket->bytes != ticket->orig_bytes)
4966                         return true;
4967         }
4968         return false;
4969 }
4970
4971 /*
4972  * This is for normal flushers, we can wait all goddamned day if we want to.  We
4973  * will loop and continuously try to flush as long as we are making progress.
4974  * We count progress as clearing off tickets each time we have to loop.
4975  */
4976 static void btrfs_async_reclaim_metadata_space(struct work_struct *work)
4977 {
4978         struct btrfs_fs_info *fs_info;
4979         struct btrfs_space_info *space_info;
4980         u64 to_reclaim;
4981         int flush_state;
4982         int commit_cycles = 0;
4983         u64 last_tickets_id;
4984
4985         fs_info = container_of(work, struct btrfs_fs_info, async_reclaim_work);
4986         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
4987
4988         spin_lock(&space_info->lock);
4989         to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info, space_info,
4990                                                       false);
4991         if (!to_reclaim) {
4992                 space_info->flush = 0;
4993                 spin_unlock(&space_info->lock);
4994                 return;
4995         }
4996         last_tickets_id = space_info->tickets_id;
4997         spin_unlock(&space_info->lock);
4998
4999         flush_state = FLUSH_DELAYED_ITEMS_NR;
5000         do {
5001                 flush_space(fs_info, space_info, to_reclaim, flush_state);
5002                 spin_lock(&space_info->lock);
5003                 if (list_empty(&space_info->tickets)) {
5004                         space_info->flush = 0;
5005                         spin_unlock(&space_info->lock);
5006                         return;
5007                 }
5008                 to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info,
5009                                                               space_info,
5010                                                               false);
5011                 if (last_tickets_id == space_info->tickets_id) {
5012                         flush_state++;
5013                 } else {
5014                         last_tickets_id = space_info->tickets_id;
5015                         flush_state = FLUSH_DELAYED_ITEMS_NR;
5016                         if (commit_cycles)
5017                                 commit_cycles--;
5018                 }
5019
5020                 /*
5021                  * We don't want to force a chunk allocation until we've tried
5022                  * pretty hard to reclaim space.  Think of the case where we
5023                  * freed up a bunch of space and so have a lot of pinned space
5024                  * to reclaim.  We would rather use that than possibly create a
5025                  * underutilized metadata chunk.  So if this is our first run
5026                  * through the flushing state machine skip ALLOC_CHUNK_FORCE and
5027                  * commit the transaction.  If nothing has changed the next go
5028                  * around then we can force a chunk allocation.
5029                  */
5030                 if (flush_state == ALLOC_CHUNK_FORCE && !commit_cycles)
5031                         flush_state++;
5032
5033                 if (flush_state > COMMIT_TRANS) {
5034                         commit_cycles++;
5035                         if (commit_cycles > 2) {
5036                                 if (wake_all_tickets(&space_info->tickets)) {
5037                                         flush_state = FLUSH_DELAYED_ITEMS_NR;
5038                                         commit_cycles--;
5039                                 } else {
5040                                         space_info->flush = 0;
5041                                 }
5042                         } else {
5043                                 flush_state = FLUSH_DELAYED_ITEMS_NR;
5044                         }
5045                 }
5046                 spin_unlock(&space_info->lock);
5047         } while (flush_state <= COMMIT_TRANS);
5048 }
5049
5050 void btrfs_init_async_reclaim_work(struct work_struct *work)
5051 {
5052         INIT_WORK(work, btrfs_async_reclaim_metadata_space);
5053 }
5054
5055 static const enum btrfs_flush_state priority_flush_states[] = {
5056         FLUSH_DELAYED_ITEMS_NR,
5057         FLUSH_DELAYED_ITEMS,
5058         ALLOC_CHUNK,
5059 };
5060
5061 static void priority_reclaim_metadata_space(struct btrfs_fs_info *fs_info,
5062                                             struct btrfs_space_info *space_info,
5063                                             struct reserve_ticket *ticket)
5064 {
5065         u64 to_reclaim;
5066         int flush_state;
5067
5068         spin_lock(&space_info->lock);
5069         to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info, space_info,
5070                                                       false);
5071         if (!to_reclaim) {
5072                 spin_unlock(&space_info->lock);
5073                 return;
5074         }
5075         spin_unlock(&space_info->lock);
5076
5077         flush_state = 0;
5078         do {
5079                 flush_space(fs_info, space_info, to_reclaim,
5080                             priority_flush_states[flush_state]);
5081                 flush_state++;
5082                 spin_lock(&space_info->lock);
5083                 if (ticket->bytes == 0) {
5084                         spin_unlock(&space_info->lock);
5085                         return;
5086                 }
5087                 spin_unlock(&space_info->lock);
5088         } while (flush_state < ARRAY_SIZE(priority_flush_states));
5089 }
5090
5091 static int wait_reserve_ticket(struct btrfs_fs_info *fs_info,
5092                                struct btrfs_space_info *space_info,
5093                                struct reserve_ticket *ticket)
5094
5095 {
5096         DEFINE_WAIT(wait);
5097         u64 reclaim_bytes = 0;
5098         int ret = 0;
5099
5100         spin_lock(&space_info->lock);
5101         while (ticket->bytes > 0 && ticket->error == 0) {
5102                 ret = prepare_to_wait_event(&ticket->wait, &wait, TASK_KILLABLE);
5103                 if (ret) {
5104                         ret = -EINTR;
5105                         break;
5106                 }
5107                 spin_unlock(&space_info->lock);
5108
5109                 schedule();
5110
5111                 finish_wait(&ticket->wait, &wait);
5112                 spin_lock(&space_info->lock);
5113         }
5114         if (!ret)
5115                 ret = ticket->error;
5116         if (!list_empty(&ticket->list))
5117                 list_del_init(&ticket->list);
5118         if (ticket->bytes && ticket->bytes < ticket->orig_bytes)
5119                 reclaim_bytes = ticket->orig_bytes - ticket->bytes;
5120         spin_unlock(&space_info->lock);
5121
5122         if (reclaim_bytes)
5123                 space_info_add_old_bytes(fs_info, space_info, reclaim_bytes);
5124         return ret;
5125 }
5126
5127 /**
5128  * reserve_metadata_bytes - try to reserve bytes from the block_rsv's space
5129  * @root - the root we're allocating for
5130  * @space_info - the space info we want to allocate from
5131  * @orig_bytes - the number of bytes we want
5132  * @flush - whether or not we can flush to make our reservation
5133  *
5134  * This will reserve orig_bytes number of bytes from the space info associated
5135  * with the block_rsv.  If there is not enough space it will make an attempt to
5136  * flush out space to make room.  It will do this by flushing delalloc if
5137  * possible or committing the transaction.  If flush is 0 then no attempts to
5138  * regain reservations will be made and this will fail if there is not enough
5139  * space already.
5140  */
5141 static int __reserve_metadata_bytes(struct btrfs_fs_info *fs_info,
5142                                     struct btrfs_space_info *space_info,
5143                                     u64 orig_bytes,
5144                                     enum btrfs_reserve_flush_enum flush,
5145                                     bool system_chunk)
5146 {
5147         struct reserve_ticket ticket;
5148         u64 used;
5149         u64 reclaim_bytes = 0;
5150         int ret = 0;
5151
5152         ASSERT(orig_bytes);
5153         ASSERT(!current->journal_info || flush != BTRFS_RESERVE_FLUSH_ALL);
5154
5155         spin_lock(&space_info->lock);
5156         ret = -ENOSPC;
5157         used = btrfs_space_info_used(space_info, true);
5158
5159         /*
5160          * If we have enough space then hooray, make our reservation and carry
5161          * on.  If not see if we can overcommit, and if we can, hooray carry on.
5162          * If not things get more complicated.
5163          */
5164         if (used + orig_bytes <= space_info->total_bytes) {
5165                 update_bytes_may_use(space_info, orig_bytes);
5166                 trace_btrfs_space_reservation(fs_info, "space_info",
5167                                               space_info->flags, orig_bytes, 1);
5168                 ret = 0;
5169         } else if (can_overcommit(fs_info, space_info, orig_bytes, flush,
5170                                   system_chunk)) {
5171                 update_bytes_may_use(space_info, orig_bytes);
5172                 trace_btrfs_space_reservation(fs_info, "space_info",
5173                                               space_info->flags, orig_bytes, 1);
5174                 ret = 0;
5175         }
5176
5177         /*
5178          * If we couldn't make a reservation then setup our reservation ticket
5179          * and kick the async worker if it's not already running.
5180          *
5181          * If we are a priority flusher then we just need to add our ticket to
5182          * the list and we will do our own flushing further down.
5183          */
5184         if (ret && flush != BTRFS_RESERVE_NO_FLUSH) {
5185                 ticket.orig_bytes = orig_bytes;
5186                 ticket.bytes = orig_bytes;
5187                 ticket.error = 0;
5188                 init_waitqueue_head(&ticket.wait);
5189                 if (flush == BTRFS_RESERVE_FLUSH_ALL) {
5190                         list_add_tail(&ticket.list, &space_info->tickets);
5191                         if (!space_info->flush) {
5192                                 space_info->flush = 1;
5193                                 trace_btrfs_trigger_flush(fs_info,
5194                                                           space_info->flags,
5195                                                           orig_bytes, flush,
5196                                                           "enospc");
5197                                 queue_work(system_unbound_wq,
5198                                            &fs_info->async_reclaim_work);
5199                         }
5200                 } else {
5201                         list_add_tail(&ticket.list,
5202                                       &space_info->priority_tickets);
5203                 }
5204         } else if (!ret && space_info->flags & BTRFS_BLOCK_GROUP_METADATA) {
5205                 used += orig_bytes;
5206                 /*
5207                  * We will do the space reservation dance during log replay,
5208                  * which means we won't have fs_info->fs_root set, so don't do
5209                  * the async reclaim as we will panic.
5210                  */
5211                 if (!test_bit(BTRFS_FS_LOG_RECOVERING, &fs_info->flags) &&
5212                     need_do_async_reclaim(fs_info, space_info,
5213                                           used, system_chunk) &&
5214                     !work_busy(&fs_info->async_reclaim_work)) {
5215                         trace_btrfs_trigger_flush(fs_info, space_info->flags,
5216                                                   orig_bytes, flush, "preempt");
5217                         queue_work(system_unbound_wq,
5218                                    &fs_info->async_reclaim_work);
5219                 }
5220         }
5221         spin_unlock(&space_info->lock);
5222         if (!ret || flush == BTRFS_RESERVE_NO_FLUSH)
5223                 return ret;
5224
5225         if (flush == BTRFS_RESERVE_FLUSH_ALL)
5226                 return wait_reserve_ticket(fs_info, space_info, &ticket);
5227
5228         ret = 0;
5229         priority_reclaim_metadata_space(fs_info, space_info, &ticket);
5230         spin_lock(&space_info->lock);
5231         if (ticket.bytes) {
5232                 if (ticket.bytes < orig_bytes)
5233                         reclaim_bytes = orig_bytes - ticket.bytes;
5234                 list_del_init(&ticket.list);
5235                 ret = -ENOSPC;
5236         }
5237         spin_unlock(&space_info->lock);
5238
5239         if (reclaim_bytes)
5240                 space_info_add_old_bytes(fs_info, space_info, reclaim_bytes);
5241         ASSERT(list_empty(&ticket.list));
5242         return ret;
5243 }
5244
5245 /**
5246  * reserve_metadata_bytes - try to reserve bytes from the block_rsv's space
5247  * @root - the root we're allocating for
5248  * @block_rsv - the block_rsv we're allocating for
5249  * @orig_bytes - the number of bytes we want
5250  * @flush - whether or not we can flush to make our reservation
5251  *
5252  * This will reserve orig_bytes number of bytes from the space info associated
5253  * with the block_rsv.  If there is not enough space it will make an attempt to
5254  * flush out space to make room.  It will do this by flushing delalloc if
5255  * possible or committing the transaction.  If flush is 0 then no attempts to
5256  * regain reservations will be made and this will fail if there is not enough
5257  * space already.
5258  */
5259 static int reserve_metadata_bytes(struct btrfs_root *root,
5260                                   struct btrfs_block_rsv *block_rsv,
5261                                   u64 orig_bytes,
5262                                   enum btrfs_reserve_flush_enum flush)
5263 {
5264         struct btrfs_fs_info *fs_info = root->fs_info;
5265         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
5266         int ret;
5267         bool system_chunk = (root == fs_info->chunk_root);
5268
5269         ret = __reserve_metadata_bytes(fs_info, block_rsv->space_info,
5270                                        orig_bytes, flush, system_chunk);
5271         if (ret == -ENOSPC &&
5272             unlikely(root->orphan_cleanup_state == ORPHAN_CLEANUP_STARTED)) {
5273                 if (block_rsv != global_rsv &&
5274                     !block_rsv_use_bytes(global_rsv, orig_bytes))
5275                         ret = 0;
5276         }
5277         if (ret == -ENOSPC) {
5278                 trace_btrfs_space_reservation(fs_info, "space_info:enospc",
5279                                               block_rsv->space_info->flags,
5280                                               orig_bytes, 1);
5281
5282                 if (btrfs_test_opt(fs_info, ENOSPC_DEBUG))
5283                         dump_space_info(fs_info, block_rsv->space_info,
5284                                         orig_bytes, 0);
5285         }
5286         return ret;
5287 }
5288
5289 static struct btrfs_block_rsv *get_block_rsv(
5290                                         const struct btrfs_trans_handle *trans,
5291                                         const struct btrfs_root *root)
5292 {
5293         struct btrfs_fs_info *fs_info = root->fs_info;
5294         struct btrfs_block_rsv *block_rsv = NULL;
5295
5296         if (test_bit(BTRFS_ROOT_REF_COWS, &root->state) ||
5297             (root == fs_info->csum_root && trans->adding_csums) ||
5298             (root == fs_info->uuid_root))
5299                 block_rsv = trans->block_rsv;
5300
5301         if (!block_rsv)
5302                 block_rsv = root->block_rsv;
5303
5304         if (!block_rsv)
5305                 block_rsv = &fs_info->empty_block_rsv;
5306
5307         return block_rsv;
5308 }
5309
5310 static int block_rsv_use_bytes(struct btrfs_block_rsv *block_rsv,
5311                                u64 num_bytes)
5312 {
5313         int ret = -ENOSPC;
5314         spin_lock(&block_rsv->lock);
5315         if (block_rsv->reserved >= num_bytes) {
5316                 block_rsv->reserved -= num_bytes;
5317                 if (block_rsv->reserved < block_rsv->size)
5318                         block_rsv->full = 0;
5319                 ret = 0;
5320         }
5321         spin_unlock(&block_rsv->lock);
5322         return ret;
5323 }
5324
5325 static void block_rsv_add_bytes(struct btrfs_block_rsv *block_rsv,
5326                                 u64 num_bytes, bool update_size)
5327 {
5328         spin_lock(&block_rsv->lock);
5329         block_rsv->reserved += num_bytes;
5330         if (update_size)
5331                 block_rsv->size += num_bytes;
5332         else if (block_rsv->reserved >= block_rsv->size)
5333                 block_rsv->full = 1;
5334         spin_unlock(&block_rsv->lock);
5335 }
5336
5337 int btrfs_cond_migrate_bytes(struct btrfs_fs_info *fs_info,
5338                              struct btrfs_block_rsv *dest, u64 num_bytes,
5339                              int min_factor)
5340 {
5341         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
5342         u64 min_bytes;
5343
5344         if (global_rsv->space_info != dest->space_info)
5345                 return -ENOSPC;
5346
5347         spin_lock(&global_rsv->lock);
5348         min_bytes = div_factor(global_rsv->size, min_factor);
5349         if (global_rsv->reserved < min_bytes + num_bytes) {
5350                 spin_unlock(&global_rsv->lock);
5351                 return -ENOSPC;
5352         }
5353         global_rsv->reserved -= num_bytes;
5354         if (global_rsv->reserved < global_rsv->size)
5355                 global_rsv->full = 0;
5356         spin_unlock(&global_rsv->lock);
5357
5358         block_rsv_add_bytes(dest, num_bytes, true);
5359         return 0;
5360 }
5361
5362 /**
5363  * btrfs_migrate_to_delayed_refs_rsv - transfer bytes to our delayed refs rsv.
5364  * @fs_info - the fs info for our fs.
5365  * @src - the source block rsv to transfer from.
5366  * @num_bytes - the number of bytes to transfer.
5367  *
5368  * This transfers up to the num_bytes amount from the src rsv to the
5369  * delayed_refs_rsv.  Any extra bytes are returned to the space info.
5370  */
5371 void btrfs_migrate_to_delayed_refs_rsv(struct btrfs_fs_info *fs_info,
5372                                        struct btrfs_block_rsv *src,
5373                                        u64 num_bytes)
5374 {
5375         struct btrfs_block_rsv *delayed_refs_rsv = &fs_info->delayed_refs_rsv;
5376         u64 to_free = 0;
5377
5378         spin_lock(&src->lock);
5379         src->reserved -= num_bytes;
5380         src->size -= num_bytes;
5381         spin_unlock(&src->lock);
5382
5383         spin_lock(&delayed_refs_rsv->lock);
5384         if (delayed_refs_rsv->size > delayed_refs_rsv->reserved) {
5385                 u64 delta = delayed_refs_rsv->size -
5386                         delayed_refs_rsv->reserved;
5387                 if (num_bytes > delta) {
5388                         to_free = num_bytes - delta;
5389                         num_bytes = delta;
5390                 }
5391         } else {
5392                 to_free = num_bytes;
5393                 num_bytes = 0;
5394         }
5395
5396         if (num_bytes)
5397                 delayed_refs_rsv->reserved += num_bytes;
5398         if (delayed_refs_rsv->reserved >= delayed_refs_rsv->size)
5399                 delayed_refs_rsv->full = 1;
5400         spin_unlock(&delayed_refs_rsv->lock);
5401
5402         if (num_bytes)
5403                 trace_btrfs_space_reservation(fs_info, "delayed_refs_rsv",
5404                                               0, num_bytes, 1);
5405         if (to_free)
5406                 space_info_add_old_bytes(fs_info, delayed_refs_rsv->space_info,
5407                                          to_free);
5408 }
5409
5410 /**
5411  * btrfs_delayed_refs_rsv_refill - refill based on our delayed refs usage.
5412  * @fs_info - the fs_info for our fs.
5413  * @flush - control how we can flush for this reservation.
5414  *
5415  * This will refill the delayed block_rsv up to 1 items size worth of space and
5416  * will return -ENOSPC if we can't make the reservation.
5417  */
5418 int btrfs_delayed_refs_rsv_refill(struct btrfs_fs_info *fs_info,
5419                                   enum btrfs_reserve_flush_enum flush)
5420 {
5421         struct btrfs_block_rsv *block_rsv = &fs_info->delayed_refs_rsv;
5422         u64 limit = btrfs_calc_trans_metadata_size(fs_info, 1);
5423         u64 num_bytes = 0;
5424         int ret = -ENOSPC;
5425
5426         spin_lock(&block_rsv->lock);
5427         if (block_rsv->reserved < block_rsv->size) {
5428                 num_bytes = block_rsv->size - block_rsv->reserved;
5429                 num_bytes = min(num_bytes, limit);
5430         }
5431         spin_unlock(&block_rsv->lock);
5432
5433         if (!num_bytes)
5434                 return 0;
5435
5436         ret = reserve_metadata_bytes(fs_info->extent_root, block_rsv,
5437                                      num_bytes, flush);
5438         if (ret)
5439                 return ret;
5440         block_rsv_add_bytes(block_rsv, num_bytes, 0);
5441         trace_btrfs_space_reservation(fs_info, "delayed_refs_rsv",
5442                                       0, num_bytes, 1);
5443         return 0;
5444 }
5445
5446 /*
5447  * This is for space we already have accounted in space_info->bytes_may_use, so
5448  * basically when we're returning space from block_rsv's.
5449  */
5450 static void space_info_add_old_bytes(struct btrfs_fs_info *fs_info,
5451                                      struct btrfs_space_info *space_info,
5452                                      u64 num_bytes)
5453 {
5454         struct reserve_ticket *ticket;
5455         struct list_head *head;
5456         u64 used;
5457         enum btrfs_reserve_flush_enum flush = BTRFS_RESERVE_NO_FLUSH;
5458         bool check_overcommit = false;
5459
5460         spin_lock(&space_info->lock);
5461         head = &space_info->priority_tickets;
5462
5463         /*
5464          * If we are over our limit then we need to check and see if we can
5465          * overcommit, and if we can't then we just need to free up our space
5466          * and not satisfy any requests.
5467          */
5468         used = btrfs_space_info_used(space_info, true);
5469         if (used - num_bytes >= space_info->total_bytes)
5470                 check_overcommit = true;
5471 again:
5472         while (!list_empty(head) && num_bytes) {
5473                 ticket = list_first_entry(head, struct reserve_ticket,
5474                                           list);
5475                 /*
5476                  * We use 0 bytes because this space is already reserved, so
5477                  * adding the ticket space would be a double count.
5478                  */
5479                 if (check_overcommit &&
5480                     !can_overcommit(fs_info, space_info, 0, flush, false))
5481                         break;
5482                 if (num_bytes >= ticket->bytes) {
5483                         list_del_init(&ticket->list);
5484                         num_bytes -= ticket->bytes;
5485                         ticket->bytes = 0;
5486                         space_info->tickets_id++;
5487                         wake_up(&ticket->wait);
5488                 } else {
5489                         ticket->bytes -= num_bytes;
5490                         num_bytes = 0;
5491                 }
5492         }
5493
5494         if (num_bytes && head == &space_info->priority_tickets) {
5495                 head = &space_info->tickets;
5496                 flush = BTRFS_RESERVE_FLUSH_ALL;
5497                 goto again;
5498         }
5499         update_bytes_may_use(space_info, -num_bytes);
5500         trace_btrfs_space_reservation(fs_info, "space_info",
5501                                       space_info->flags, num_bytes, 0);
5502         spin_unlock(&space_info->lock);
5503 }
5504
5505 /*
5506  * This is for newly allocated space that isn't accounted in
5507  * space_info->bytes_may_use yet.  So if we allocate a chunk or unpin an extent
5508  * we use this helper.
5509  */
5510 static void space_info_add_new_bytes(struct btrfs_fs_info *fs_info,
5511                                      struct btrfs_space_info *space_info,
5512                                      u64 num_bytes)
5513 {
5514         struct reserve_ticket *ticket;
5515         struct list_head *head = &space_info->priority_tickets;
5516
5517 again:
5518         while (!list_empty(head) && num_bytes) {
5519                 ticket = list_first_entry(head, struct reserve_ticket,
5520                                           list);
5521                 if (num_bytes >= ticket->bytes) {
5522                         trace_btrfs_space_reservation(fs_info, "space_info",
5523                                                       space_info->flags,
5524                                                       ticket->bytes, 1);
5525                         list_del_init(&ticket->list);
5526                         num_bytes -= ticket->bytes;
5527                         update_bytes_may_use(space_info, ticket->bytes);
5528                         ticket->bytes = 0;
5529                         space_info->tickets_id++;
5530                         wake_up(&ticket->wait);
5531                 } else {
5532                         trace_btrfs_space_reservation(fs_info, "space_info",
5533                                                       space_info->flags,
5534                                                       num_bytes, 1);
5535                         update_bytes_may_use(space_info, num_bytes);
5536                         ticket->bytes -= num_bytes;
5537                         num_bytes = 0;
5538                 }
5539         }
5540
5541         if (num_bytes && head == &space_info->priority_tickets) {
5542                 head = &space_info->tickets;
5543                 goto again;
5544         }
5545 }
5546
5547 static u64 block_rsv_release_bytes(struct btrfs_fs_info *fs_info,
5548                                     struct btrfs_block_rsv *block_rsv,
5549                                     struct btrfs_block_rsv *dest, u64 num_bytes,
5550                                     u64 *qgroup_to_release_ret)
5551 {
5552         struct btrfs_space_info *space_info = block_rsv->space_info;
5553         u64 qgroup_to_release = 0;
5554         u64 ret;
5555
5556         spin_lock(&block_rsv->lock);
5557         if (num_bytes == (u64)-1) {
5558                 num_bytes = block_rsv->size;
5559                 qgroup_to_release = block_rsv->qgroup_rsv_size;
5560         }
5561         block_rsv->size -= num_bytes;
5562         if (block_rsv->reserved >= block_rsv->size) {
5563                 num_bytes = block_rsv->reserved - block_rsv->size;
5564                 block_rsv->reserved = block_rsv->size;
5565                 block_rsv->full = 1;
5566         } else {
5567                 num_bytes = 0;
5568         }
5569         if (block_rsv->qgroup_rsv_reserved >= block_rsv->qgroup_rsv_size) {
5570                 qgroup_to_release = block_rsv->qgroup_rsv_reserved -
5571                                     block_rsv->qgroup_rsv_size;
5572                 block_rsv->qgroup_rsv_reserved = block_rsv->qgroup_rsv_size;
5573         } else {
5574                 qgroup_to_release = 0;
5575         }
5576         spin_unlock(&block_rsv->lock);
5577
5578         ret = num_bytes;
5579         if (num_bytes > 0) {
5580                 if (dest) {
5581                         spin_lock(&dest->lock);
5582                         if (!dest->full) {
5583                                 u64 bytes_to_add;
5584
5585                                 bytes_to_add = dest->size - dest->reserved;
5586                                 bytes_to_add = min(num_bytes, bytes_to_add);
5587                                 dest->reserved += bytes_to_add;
5588                                 if (dest->reserved >= dest->size)
5589                                         dest->full = 1;
5590                                 num_bytes -= bytes_to_add;
5591                         }
5592                         spin_unlock(&dest->lock);
5593                 }
5594                 if (num_bytes)
5595                         space_info_add_old_bytes(fs_info, space_info,
5596                                                  num_bytes);
5597         }
5598         if (qgroup_to_release_ret)
5599                 *qgroup_to_release_ret = qgroup_to_release;
5600         return ret;
5601 }
5602
5603 int btrfs_block_rsv_migrate(struct btrfs_block_rsv *src,
5604                             struct btrfs_block_rsv *dst, u64 num_bytes,
5605                             bool update_size)
5606 {
5607         int ret;
5608
5609         ret = block_rsv_use_bytes(src, num_bytes);
5610         if (ret)
5611                 return ret;
5612
5613         block_rsv_add_bytes(dst, num_bytes, update_size);
5614         return 0;
5615 }
5616
5617 void btrfs_init_block_rsv(struct btrfs_block_rsv *rsv, unsigned short type)
5618 {
5619         memset(rsv, 0, sizeof(*rsv));
5620         spin_lock_init(&rsv->lock);
5621         rsv->type = type;
5622 }
5623
5624 void btrfs_init_metadata_block_rsv(struct btrfs_fs_info *fs_info,
5625                                    struct btrfs_block_rsv *rsv,
5626                                    unsigned short type)
5627 {
5628         btrfs_init_block_rsv(rsv, type);
5629         rsv->space_info = __find_space_info(fs_info,
5630                                             BTRFS_BLOCK_GROUP_METADATA);
5631 }
5632
5633 struct btrfs_block_rsv *btrfs_alloc_block_rsv(struct btrfs_fs_info *fs_info,
5634                                               unsigned short type)
5635 {
5636         struct btrfs_block_rsv *block_rsv;
5637
5638         block_rsv = kmalloc(sizeof(*block_rsv), GFP_NOFS);
5639         if (!block_rsv)
5640                 return NULL;
5641
5642         btrfs_init_metadata_block_rsv(fs_info, block_rsv, type);
5643         return block_rsv;
5644 }
5645
5646 void btrfs_free_block_rsv(struct btrfs_fs_info *fs_info,
5647                           struct btrfs_block_rsv *rsv)
5648 {
5649         if (!rsv)
5650                 return;
5651         btrfs_block_rsv_release(fs_info, rsv, (u64)-1);
5652         kfree(rsv);
5653 }
5654
5655 int btrfs_block_rsv_add(struct btrfs_root *root,
5656                         struct btrfs_block_rsv *block_rsv, u64 num_bytes,
5657                         enum btrfs_reserve_flush_enum flush)
5658 {
5659         int ret;
5660
5661         if (num_bytes == 0)
5662                 return 0;
5663
5664         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
5665         if (!ret)
5666                 block_rsv_add_bytes(block_rsv, num_bytes, true);
5667
5668         return ret;
5669 }
5670
5671 int btrfs_block_rsv_check(struct btrfs_block_rsv *block_rsv, int min_factor)
5672 {
5673         u64 num_bytes = 0;
5674         int ret = -ENOSPC;
5675
5676         if (!block_rsv)
5677                 return 0;
5678
5679         spin_lock(&block_rsv->lock);
5680         num_bytes = div_factor(block_rsv->size, min_factor);
5681         if (block_rsv->reserved >= num_bytes)
5682                 ret = 0;
5683         spin_unlock(&block_rsv->lock);
5684
5685         return ret;
5686 }
5687
5688 int btrfs_block_rsv_refill(struct btrfs_root *root,
5689                            struct btrfs_block_rsv *block_rsv, u64 min_reserved,
5690                            enum btrfs_reserve_flush_enum flush)
5691 {
5692         u64 num_bytes = 0;
5693         int ret = -ENOSPC;
5694
5695         if (!block_rsv)
5696                 return 0;
5697
5698         spin_lock(&block_rsv->lock);
5699         num_bytes = min_reserved;
5700         if (block_rsv->reserved >= num_bytes)
5701                 ret = 0;
5702         else
5703                 num_bytes -= block_rsv->reserved;
5704         spin_unlock(&block_rsv->lock);
5705
5706         if (!ret)
5707                 return 0;
5708
5709         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
5710         if (!ret) {
5711                 block_rsv_add_bytes(block_rsv, num_bytes, false);
5712                 return 0;
5713         }
5714
5715         return ret;
5716 }
5717
5718 static u64 __btrfs_block_rsv_release(struct btrfs_fs_info *fs_info,
5719                                      struct btrfs_block_rsv *block_rsv,
5720                                      u64 num_bytes, u64 *qgroup_to_release)
5721 {
5722         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
5723         struct btrfs_block_rsv *delayed_rsv = &fs_info->delayed_refs_rsv;
5724         struct btrfs_block_rsv *target = delayed_rsv;
5725
5726         if (target->full || target == block_rsv)
5727                 target = global_rsv;
5728
5729         if (block_rsv->space_info != target->space_info)
5730                 target = NULL;
5731
5732         return block_rsv_release_bytes(fs_info, block_rsv, target, num_bytes,
5733                                        qgroup_to_release);
5734 }
5735
5736 void btrfs_block_rsv_release(struct btrfs_fs_info *fs_info,
5737                              struct btrfs_block_rsv *block_rsv,
5738                              u64 num_bytes)
5739 {
5740         __btrfs_block_rsv_release(fs_info, block_rsv, num_bytes, NULL);
5741 }
5742
5743 /**
5744  * btrfs_inode_rsv_release - release any excessive reservation.
5745  * @inode - the inode we need to release from.
5746  * @qgroup_free - free or convert qgroup meta.
5747  *   Unlike normal operation, qgroup meta reservation needs to know if we are
5748  *   freeing qgroup reservation or just converting it into per-trans.  Normally
5749  *   @qgroup_free is true for error handling, and false for normal release.
5750  *
5751  * This is the same as btrfs_block_rsv_release, except that it handles the
5752  * tracepoint for the reservation.
5753  */
5754 static void btrfs_inode_rsv_release(struct btrfs_inode *inode, bool qgroup_free)
5755 {
5756         struct btrfs_fs_info *fs_info = inode->root->fs_info;
5757         struct btrfs_block_rsv *block_rsv = &inode->block_rsv;
5758         u64 released = 0;
5759         u64 qgroup_to_release = 0;
5760
5761         /*
5762          * Since we statically set the block_rsv->size we just want to say we
5763          * are releasing 0 bytes, and then we'll just get the reservation over
5764          * the size free'd.
5765          */
5766         released = __btrfs_block_rsv_release(fs_info, block_rsv, 0,
5767                                              &qgroup_to_release);
5768         if (released > 0)
5769                 trace_btrfs_space_reservation(fs_info, "delalloc",
5770                                               btrfs_ino(inode), released, 0);
5771         if (qgroup_free)
5772                 btrfs_qgroup_free_meta_prealloc(inode->root, qgroup_to_release);
5773         else
5774                 btrfs_qgroup_convert_reserved_meta(inode->root,
5775                                                    qgroup_to_release);
5776 }
5777
5778 /**
5779  * btrfs_delayed_refs_rsv_release - release a ref head's reservation.
5780  * @fs_info - the fs_info for our fs.
5781  * @nr - the number of items to drop.
5782  *
5783  * This drops the delayed ref head's count from the delayed refs rsv and frees
5784  * any excess reservation we had.
5785  */
5786 void btrfs_delayed_refs_rsv_release(struct btrfs_fs_info *fs_info, int nr)
5787 {
5788         struct btrfs_block_rsv *block_rsv = &fs_info->delayed_refs_rsv;
5789         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
5790         u64 num_bytes = btrfs_calc_trans_metadata_size(fs_info, nr);
5791         u64 released = 0;
5792
5793         released = block_rsv_release_bytes(fs_info, block_rsv, global_rsv,
5794                                            num_bytes, NULL);
5795         if (released)
5796                 trace_btrfs_space_reservation(fs_info, "delayed_refs_rsv",
5797                                               0, released, 0);
5798 }
5799
5800 static void update_global_block_rsv(struct btrfs_fs_info *fs_info)
5801 {
5802         struct btrfs_block_rsv *block_rsv = &fs_info->global_block_rsv;
5803         struct btrfs_space_info *sinfo = block_rsv->space_info;
5804         u64 num_bytes;
5805
5806         /*
5807          * The global block rsv is based on the size of the extent tree, the
5808          * checksum tree and the root tree.  If the fs is empty we want to set
5809          * it to a minimal amount for safety.
5810          */
5811         num_bytes = btrfs_root_used(&fs_info->extent_root->root_item) +
5812                 btrfs_root_used(&fs_info->csum_root->root_item) +
5813                 btrfs_root_used(&fs_info->tree_root->root_item);
5814         num_bytes = max_t(u64, num_bytes, SZ_16M);
5815
5816         spin_lock(&sinfo->lock);
5817         spin_lock(&block_rsv->lock);
5818
5819         block_rsv->size = min_t(u64, num_bytes, SZ_512M);
5820
5821         if (block_rsv->reserved < block_rsv->size) {
5822                 num_bytes = btrfs_space_info_used(sinfo, true);
5823                 if (sinfo->total_bytes > num_bytes) {
5824                         num_bytes = sinfo->total_bytes - num_bytes;
5825                         num_bytes = min(num_bytes,
5826                                         block_rsv->size - block_rsv->reserved);
5827                         block_rsv->reserved += num_bytes;
5828                         update_bytes_may_use(sinfo, num_bytes);
5829                         trace_btrfs_space_reservation(fs_info, "space_info",
5830                                                       sinfo->flags, num_bytes,
5831                                                       1);
5832                 }
5833         } else if (block_rsv->reserved > block_rsv->size) {
5834                 num_bytes = block_rsv->reserved - block_rsv->size;
5835                 update_bytes_may_use(sinfo, -num_bytes);
5836                 trace_btrfs_space_reservation(fs_info, "space_info",
5837                                       sinfo->flags, num_bytes, 0);
5838                 block_rsv->reserved = block_rsv->size;
5839         }
5840
5841         if (block_rsv->reserved == block_rsv->size)
5842                 block_rsv->full = 1;
5843         else
5844                 block_rsv->full = 0;
5845
5846         spin_unlock(&block_rsv->lock);
5847         spin_unlock(&sinfo->lock);
5848 }
5849
5850 static void init_global_block_rsv(struct btrfs_fs_info *fs_info)
5851 {
5852         struct btrfs_space_info *space_info;
5853
5854         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
5855         fs_info->chunk_block_rsv.space_info = space_info;
5856
5857         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
5858         fs_info->global_block_rsv.space_info = space_info;
5859         fs_info->trans_block_rsv.space_info = space_info;
5860         fs_info->empty_block_rsv.space_info = space_info;
5861         fs_info->delayed_block_rsv.space_info = space_info;
5862         fs_info->delayed_refs_rsv.space_info = space_info;
5863
5864         fs_info->extent_root->block_rsv = &fs_info->delayed_refs_rsv;
5865         fs_info->csum_root->block_rsv = &fs_info->delayed_refs_rsv;
5866         fs_info->dev_root->block_rsv = &fs_info->global_block_rsv;
5867         fs_info->tree_root->block_rsv = &fs_info->global_block_rsv;
5868         if (fs_info->quota_root)
5869                 fs_info->quota_root->block_rsv = &fs_info->global_block_rsv;
5870         fs_info->chunk_root->block_rsv = &fs_info->chunk_block_rsv;
5871
5872         update_global_block_rsv(fs_info);
5873 }
5874
5875 static void release_global_block_rsv(struct btrfs_fs_info *fs_info)
5876 {
5877         block_rsv_release_bytes(fs_info, &fs_info->global_block_rsv, NULL,
5878                                 (u64)-1, NULL);
5879         WARN_ON(fs_info->trans_block_rsv.size > 0);
5880         WARN_ON(fs_info->trans_block_rsv.reserved > 0);
5881         WARN_ON(fs_info->chunk_block_rsv.size > 0);
5882         WARN_ON(fs_info->chunk_block_rsv.reserved > 0);
5883         WARN_ON(fs_info->delayed_block_rsv.size > 0);
5884         WARN_ON(fs_info->delayed_block_rsv.reserved > 0);
5885         WARN_ON(fs_info->delayed_refs_rsv.reserved > 0);
5886         WARN_ON(fs_info->delayed_refs_rsv.size > 0);
5887 }
5888
5889 /*
5890  * btrfs_update_delayed_refs_rsv - adjust the size of the delayed refs rsv
5891  * @trans - the trans that may have generated delayed refs
5892  *
5893  * This is to be called anytime we may have adjusted trans->delayed_ref_updates,
5894  * it'll calculate the additional size and add it to the delayed_refs_rsv.
5895  */
5896 void btrfs_update_delayed_refs_rsv(struct btrfs_trans_handle *trans)
5897 {
5898         struct btrfs_fs_info *fs_info = trans->fs_info;
5899         struct btrfs_block_rsv *delayed_rsv = &fs_info->delayed_refs_rsv;
5900         u64 num_bytes;
5901
5902         if (!trans->delayed_ref_updates)
5903                 return;
5904
5905         num_bytes = btrfs_calc_trans_metadata_size(fs_info,
5906                                                    trans->delayed_ref_updates);
5907         spin_lock(&delayed_rsv->lock);
5908         delayed_rsv->size += num_bytes;
5909         delayed_rsv->full = 0;
5910         spin_unlock(&delayed_rsv->lock);
5911         trans->delayed_ref_updates = 0;
5912 }
5913
5914 /*
5915  * To be called after all the new block groups attached to the transaction
5916  * handle have been created (btrfs_create_pending_block_groups()).
5917  */
5918 void btrfs_trans_release_chunk_metadata(struct btrfs_trans_handle *trans)
5919 {
5920         struct btrfs_fs_info *fs_info = trans->fs_info;
5921
5922         if (!trans->chunk_bytes_reserved)
5923                 return;
5924
5925         WARN_ON_ONCE(!list_empty(&trans->new_bgs));
5926
5927         block_rsv_release_bytes(fs_info, &fs_info->chunk_block_rsv, NULL,
5928                                 trans->chunk_bytes_reserved, NULL);
5929         trans->chunk_bytes_reserved = 0;
5930 }
5931
5932 /*
5933  * btrfs_subvolume_reserve_metadata() - reserve space for subvolume operation
5934  * root: the root of the parent directory
5935  * rsv: block reservation
5936  * items: the number of items that we need do reservation
5937  * use_global_rsv: allow fallback to the global block reservation
5938  *
5939  * This function is used to reserve the space for snapshot/subvolume
5940  * creation and deletion. Those operations are different with the
5941  * common file/directory operations, they change two fs/file trees
5942  * and root tree, the number of items that the qgroup reserves is
5943  * different with the free space reservation. So we can not use
5944  * the space reservation mechanism in start_transaction().
5945  */
5946 int btrfs_subvolume_reserve_metadata(struct btrfs_root *root,
5947                                      struct btrfs_block_rsv *rsv, int items,
5948                                      bool use_global_rsv)
5949 {
5950         u64 qgroup_num_bytes = 0;
5951         u64 num_bytes;
5952         int ret;
5953         struct btrfs_fs_info *fs_info = root->fs_info;
5954         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
5955
5956         if (test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags)) {
5957                 /* One for parent inode, two for dir entries */
5958                 qgroup_num_bytes = 3 * fs_info->nodesize;
5959                 ret = btrfs_qgroup_reserve_meta_prealloc(root,
5960                                 qgroup_num_bytes, true);
5961                 if (ret)
5962                         return ret;
5963         }
5964
5965         num_bytes = btrfs_calc_trans_metadata_size(fs_info, items);
5966         rsv->space_info = __find_space_info(fs_info,
5967                                             BTRFS_BLOCK_GROUP_METADATA);
5968         ret = btrfs_block_rsv_add(root, rsv, num_bytes,
5969                                   BTRFS_RESERVE_FLUSH_ALL);
5970
5971         if (ret == -ENOSPC && use_global_rsv)
5972                 ret = btrfs_block_rsv_migrate(global_rsv, rsv, num_bytes, true);
5973
5974         if (ret && qgroup_num_bytes)
5975                 btrfs_qgroup_free_meta_prealloc(root, qgroup_num_bytes);
5976
5977         return ret;
5978 }
5979
5980 void btrfs_subvolume_release_metadata(struct btrfs_fs_info *fs_info,
5981                                       struct btrfs_block_rsv *rsv)
5982 {
5983         btrfs_block_rsv_release(fs_info, rsv, (u64)-1);
5984 }
5985
5986 static void btrfs_calculate_inode_block_rsv_size(struct btrfs_fs_info *fs_info,
5987                                                  struct btrfs_inode *inode)
5988 {
5989         struct btrfs_block_rsv *block_rsv = &inode->block_rsv;
5990         u64 reserve_size = 0;
5991         u64 qgroup_rsv_size = 0;
5992         u64 csum_leaves;
5993         unsigned outstanding_extents;
5994
5995         lockdep_assert_held(&inode->lock);
5996         outstanding_extents = inode->outstanding_extents;
5997         if (outstanding_extents)
5998                 reserve_size = btrfs_calc_trans_metadata_size(fs_info,
5999                                                 outstanding_extents + 1);
6000         csum_leaves = btrfs_csum_bytes_to_leaves(fs_info,
6001                                                  inode->csum_bytes);
6002         reserve_size += btrfs_calc_trans_metadata_size(fs_info,
6003                                                        csum_leaves);
6004         /*
6005          * For qgroup rsv, the calculation is very simple:
6006          * account one nodesize for each outstanding extent
6007          *
6008          * This is overestimating in most cases.
6009          */
6010         qgroup_rsv_size = (u64)outstanding_extents * fs_info->nodesize;
6011
6012         spin_lock(&block_rsv->lock);
6013         block_rsv->size = reserve_size;
6014         block_rsv->qgroup_rsv_size = qgroup_rsv_size;
6015         spin_unlock(&block_rsv->lock);
6016 }
6017
6018 static void calc_inode_reservations(struct btrfs_fs_info *fs_info,
6019                                     u64 num_bytes, u64 *meta_reserve,
6020                                     u64 *qgroup_reserve)
6021 {
6022         u64 nr_extents = count_max_extents(num_bytes);
6023         u64 csum_leaves = btrfs_csum_bytes_to_leaves(fs_info, num_bytes);
6024
6025         /* We add one for the inode update at finish ordered time */
6026         *meta_reserve = btrfs_calc_trans_metadata_size(fs_info,
6027                                                 nr_extents + csum_leaves + 1);
6028         *qgroup_reserve = nr_extents * fs_info->nodesize;
6029 }
6030
6031 int btrfs_delalloc_reserve_metadata(struct btrfs_inode *inode, u64 num_bytes)
6032 {
6033         struct btrfs_root *root = inode->root;
6034         struct btrfs_fs_info *fs_info = root->fs_info;
6035         struct btrfs_block_rsv *block_rsv = &inode->block_rsv;
6036         u64 meta_reserve, qgroup_reserve;
6037         unsigned nr_extents;
6038         enum btrfs_reserve_flush_enum flush = BTRFS_RESERVE_FLUSH_ALL;
6039         int ret = 0;
6040         bool delalloc_lock = true;
6041
6042         /* If we are a free space inode we need to not flush since we will be in
6043          * the middle of a transaction commit.  We also don't need the delalloc
6044          * mutex since we won't race with anybody.  We need this mostly to make
6045          * lockdep shut its filthy mouth.
6046          *
6047          * If we have a transaction open (can happen if we call truncate_block
6048          * from truncate), then we need FLUSH_LIMIT so we don't deadlock.
6049          */
6050         if (btrfs_is_free_space_inode(inode)) {
6051                 flush = BTRFS_RESERVE_NO_FLUSH;
6052                 delalloc_lock = false;
6053         } else {
6054                 if (current->journal_info)
6055                         flush = BTRFS_RESERVE_FLUSH_LIMIT;
6056
6057                 if (btrfs_transaction_in_commit(fs_info))
6058                         schedule_timeout(1);
6059         }
6060
6061         if (delalloc_lock)
6062                 mutex_lock(&inode->delalloc_mutex);
6063
6064         num_bytes = ALIGN(num_bytes, fs_info->sectorsize);
6065
6066         /*
6067          * We always want to do it this way, every other way is wrong and ends
6068          * in tears.  Pre-reserving the amount we are going to add will always
6069          * be the right way, because otherwise if we have enough parallelism we
6070          * could end up with thousands of inodes all holding little bits of
6071          * reservations they were able to make previously and the only way to
6072          * reclaim that space is to ENOSPC out the operations and clear
6073          * everything out and try again, which is bad.  This way we just
6074          * over-reserve slightly, and clean up the mess when we are done.
6075          */
6076         calc_inode_reservations(fs_info, num_bytes, &meta_reserve,
6077                                 &qgroup_reserve);
6078         ret = btrfs_qgroup_reserve_meta_prealloc(root, qgroup_reserve, true);
6079         if (ret)
6080                 goto out_fail;
6081         ret = reserve_metadata_bytes(root, block_rsv, meta_reserve, flush);
6082         if (ret)
6083                 goto out_qgroup;
6084
6085         /*
6086          * Now we need to update our outstanding extents and csum bytes _first_
6087          * and then add the reservation to the block_rsv.  This keeps us from
6088          * racing with an ordered completion or some such that would think it
6089          * needs to free the reservation we just made.
6090          */
6091         spin_lock(&inode->lock);
6092         nr_extents = count_max_extents(num_bytes);
6093         btrfs_mod_outstanding_extents(inode, nr_extents);
6094         inode->csum_bytes += num_bytes;
6095         btrfs_calculate_inode_block_rsv_size(fs_info, inode);
6096         spin_unlock(&inode->lock);
6097
6098         /* Now we can safely add our space to our block rsv */
6099         block_rsv_add_bytes(block_rsv, meta_reserve, false);
6100         trace_btrfs_space_reservation(root->fs_info, "delalloc",
6101                                       btrfs_ino(inode), meta_reserve, 1);
6102
6103         spin_lock(&block_rsv->lock);
6104         block_rsv->qgroup_rsv_reserved += qgroup_reserve;
6105         spin_unlock(&block_rsv->lock);
6106
6107         if (delalloc_lock)
6108                 mutex_unlock(&inode->delalloc_mutex);
6109         return 0;
6110 out_qgroup:
6111         btrfs_qgroup_free_meta_prealloc(root, qgroup_reserve);
6112 out_fail:
6113         btrfs_inode_rsv_release(inode, true);
6114         if (delalloc_lock)
6115                 mutex_unlock(&inode->delalloc_mutex);
6116         return ret;
6117 }
6118
6119 /**
6120  * btrfs_delalloc_release_metadata - release a metadata reservation for an inode
6121  * @inode: the inode to release the reservation for.
6122  * @num_bytes: the number of bytes we are releasing.
6123  * @qgroup_free: free qgroup reservation or convert it to per-trans reservation
6124  *
6125  * This will release the metadata reservation for an inode.  This can be called
6126  * once we complete IO for a given set of bytes to release their metadata
6127  * reservations, or on error for the same reason.
6128  */
6129 void btrfs_delalloc_release_metadata(struct btrfs_inode *inode, u64 num_bytes,
6130                                      bool qgroup_free)
6131 {
6132         struct btrfs_fs_info *fs_info = inode->root->fs_info;
6133
6134         num_bytes = ALIGN(num_bytes, fs_info->sectorsize);
6135         spin_lock(&inode->lock);
6136         inode->csum_bytes -= num_bytes;
6137         btrfs_calculate_inode_block_rsv_size(fs_info, inode);
6138         spin_unlock(&inode->lock);
6139
6140         if (btrfs_is_testing(fs_info))
6141                 return;
6142
6143         btrfs_inode_rsv_release(inode, qgroup_free);
6144 }
6145
6146 /**
6147  * btrfs_delalloc_release_extents - release our outstanding_extents
6148  * @inode: the inode to balance the reservation for.
6149  * @num_bytes: the number of bytes we originally reserved with
6150  * @qgroup_free: do we need to free qgroup meta reservation or convert them.
6151  *
6152  * When we reserve space we increase outstanding_extents for the extents we may
6153  * add.  Once we've set the range as delalloc or created our ordered extents we
6154  * have outstanding_extents to track the real usage, so we use this to free our
6155  * temporarily tracked outstanding_extents.  This _must_ be used in conjunction
6156  * with btrfs_delalloc_reserve_metadata.
6157  */
6158 void btrfs_delalloc_release_extents(struct btrfs_inode *inode, u64 num_bytes,
6159                                     bool qgroup_free)
6160 {
6161         struct btrfs_fs_info *fs_info = inode->root->fs_info;
6162         unsigned num_extents;
6163
6164         spin_lock(&inode->lock);
6165         num_extents = count_max_extents(num_bytes);
6166         btrfs_mod_outstanding_extents(inode, -num_extents);
6167         btrfs_calculate_inode_block_rsv_size(fs_info, inode);
6168         spin_unlock(&inode->lock);
6169
6170         if (btrfs_is_testing(fs_info))
6171                 return;
6172
6173         btrfs_inode_rsv_release(inode, qgroup_free);
6174 }
6175
6176 /**
6177  * btrfs_delalloc_reserve_space - reserve data and metadata space for
6178  * delalloc
6179  * @inode: inode we're writing to
6180  * @start: start range we are writing to
6181  * @len: how long the range we are writing to
6182  * @reserved: mandatory parameter, record actually reserved qgroup ranges of
6183  *            current reservation.
6184  *
6185  * This will do the following things
6186  *
6187  * o reserve space in data space info for num bytes
6188  *   and reserve precious corresponding qgroup space
6189  *   (Done in check_data_free_space)
6190  *
6191  * o reserve space for metadata space, based on the number of outstanding
6192  *   extents and how much csums will be needed
6193  *   also reserve metadata space in a per root over-reserve method.
6194  * o add to the inodes->delalloc_bytes
6195  * o add it to the fs_info's delalloc inodes list.
6196  *   (Above 3 all done in delalloc_reserve_metadata)
6197  *
6198  * Return 0 for success
6199  * Return <0 for error(-ENOSPC or -EQUOT)
6200  */
6201 int btrfs_delalloc_reserve_space(struct inode *inode,
6202                         struct extent_changeset **reserved, u64 start, u64 len)
6203 {
6204         int ret;
6205
6206         ret = btrfs_check_data_free_space(inode, reserved, start, len);
6207         if (ret < 0)
6208                 return ret;
6209         ret = btrfs_delalloc_reserve_metadata(BTRFS_I(inode), len);
6210         if (ret < 0)
6211                 btrfs_free_reserved_data_space(inode, *reserved, start, len);
6212         return ret;
6213 }
6214
6215 /**
6216  * btrfs_delalloc_release_space - release data and metadata space for delalloc
6217  * @inode: inode we're releasing space for
6218  * @start: start position of the space already reserved
6219  * @len: the len of the space already reserved
6220  * @release_bytes: the len of the space we consumed or didn't use
6221  *
6222  * This function will release the metadata space that was not used and will
6223  * decrement ->delalloc_bytes and remove it from the fs_info delalloc_inodes
6224  * list if there are no delalloc bytes left.
6225  * Also it will handle the qgroup reserved space.
6226  */
6227 void btrfs_delalloc_release_space(struct inode *inode,
6228                                   struct extent_changeset *reserved,
6229                                   u64 start, u64 len, bool qgroup_free)
6230 {
6231         btrfs_delalloc_release_metadata(BTRFS_I(inode), len, qgroup_free);
6232         btrfs_free_reserved_data_space(inode, reserved, start, len);
6233 }
6234
6235 static int update_block_group(struct btrfs_trans_handle *trans,
6236                               u64 bytenr, u64 num_bytes, int alloc)
6237 {
6238         struct btrfs_fs_info *info = trans->fs_info;
6239         struct btrfs_block_group_cache *cache = NULL;
6240         u64 total = num_bytes;
6241         u64 old_val;
6242         u64 byte_in_group;
6243         int factor;
6244         int ret = 0;
6245
6246         /* block accounting for super block */
6247         spin_lock(&info->delalloc_root_lock);
6248         old_val = btrfs_super_bytes_used(info->super_copy);
6249         if (alloc)
6250                 old_val += num_bytes;
6251         else
6252                 old_val -= num_bytes;
6253         btrfs_set_super_bytes_used(info->super_copy, old_val);
6254         spin_unlock(&info->delalloc_root_lock);
6255
6256         while (total) {
6257                 cache = btrfs_lookup_block_group(info, bytenr);
6258                 if (!cache) {
6259                         ret = -ENOENT;
6260                         break;
6261                 }
6262                 factor = btrfs_bg_type_to_factor(cache->flags);
6263
6264                 /*
6265                  * If this block group has free space cache written out, we
6266                  * need to make sure to load it if we are removing space.  This
6267                  * is because we need the unpinning stage to actually add the
6268                  * space back to the block group, otherwise we will leak space.
6269                  */
6270                 if (!alloc && cache->cached == BTRFS_CACHE_NO)
6271                         cache_block_group(cache, 1);
6272
6273                 byte_in_group = bytenr - cache->key.objectid;
6274                 WARN_ON(byte_in_group > cache->key.offset);
6275
6276                 spin_lock(&cache->space_info->lock);
6277                 spin_lock(&cache->lock);
6278
6279                 if (btrfs_test_opt(info, SPACE_CACHE) &&
6280                     cache->disk_cache_state < BTRFS_DC_CLEAR)
6281                         cache->disk_cache_state = BTRFS_DC_CLEAR;
6282
6283                 old_val = btrfs_block_group_used(&cache->item);
6284                 num_bytes = min(total, cache->key.offset - byte_in_group);
6285                 if (alloc) {
6286                         old_val += num_bytes;
6287                         btrfs_set_block_group_used(&cache->item, old_val);
6288                         cache->reserved -= num_bytes;
6289                         cache->space_info->bytes_reserved -= num_bytes;
6290                         cache->space_info->bytes_used += num_bytes;
6291                         cache->space_info->disk_used += num_bytes * factor;
6292                         spin_unlock(&cache->lock);
6293                         spin_unlock(&cache->space_info->lock);
6294                 } else {
6295                         old_val -= num_bytes;
6296                         btrfs_set_block_group_used(&cache->item, old_val);
6297                         cache->pinned += num_bytes;
6298                         update_bytes_pinned(cache->space_info, num_bytes);
6299                         cache->space_info->bytes_used -= num_bytes;
6300                         cache->space_info->disk_used -= num_bytes * factor;
6301                         spin_unlock(&cache->lock);
6302                         spin_unlock(&cache->space_info->lock);
6303
6304                         trace_btrfs_space_reservation(info, "pinned",
6305                                                       cache->space_info->flags,
6306                                                       num_bytes, 1);
6307                         percpu_counter_add_batch(&cache->space_info->total_bytes_pinned,
6308                                            num_bytes,
6309                                            BTRFS_TOTAL_BYTES_PINNED_BATCH);
6310                         set_extent_dirty(info->pinned_extents,
6311                                          bytenr, bytenr + num_bytes - 1,
6312                                          GFP_NOFS | __GFP_NOFAIL);
6313                 }
6314
6315                 spin_lock(&trans->transaction->dirty_bgs_lock);
6316                 if (list_empty(&cache->dirty_list)) {
6317                         list_add_tail(&cache->dirty_list,
6318                                       &trans->transaction->dirty_bgs);
6319                         trans->delayed_ref_updates++;
6320                         btrfs_get_block_group(cache);
6321                 }
6322                 spin_unlock(&trans->transaction->dirty_bgs_lock);
6323
6324                 /*
6325                  * No longer have used bytes in this block group, queue it for
6326                  * deletion. We do this after adding the block group to the
6327                  * dirty list to avoid races between cleaner kthread and space
6328                  * cache writeout.
6329                  */
6330                 if (!alloc && old_val == 0)
6331                         btrfs_mark_bg_unused(cache);
6332
6333                 btrfs_put_block_group(cache);
6334                 total -= num_bytes;
6335                 bytenr += num_bytes;
6336         }
6337
6338         /* Modified block groups are accounted for in the delayed_refs_rsv. */
6339         btrfs_update_delayed_refs_rsv(trans);
6340         return ret;
6341 }
6342
6343 static u64 first_logical_byte(struct btrfs_fs_info *fs_info, u64 search_start)
6344 {
6345         struct btrfs_block_group_cache *cache;
6346         u64 bytenr;
6347
6348         spin_lock(&fs_info->block_group_cache_lock);
6349         bytenr = fs_info->first_logical_byte;
6350         spin_unlock(&fs_info->block_group_cache_lock);
6351
6352         if (bytenr < (u64)-1)
6353                 return bytenr;
6354
6355         cache = btrfs_lookup_first_block_group(fs_info, search_start);
6356         if (!cache)
6357                 return 0;
6358
6359         bytenr = cache->key.objectid;
6360         btrfs_put_block_group(cache);
6361
6362         return bytenr;
6363 }
6364
6365 static int pin_down_extent(struct btrfs_block_group_cache *cache,
6366                            u64 bytenr, u64 num_bytes, int reserved)
6367 {
6368         struct btrfs_fs_info *fs_info = cache->fs_info;
6369
6370         spin_lock(&cache->space_info->lock);
6371         spin_lock(&cache->lock);
6372         cache->pinned += num_bytes;
6373         update_bytes_pinned(cache->space_info, num_bytes);
6374         if (reserved) {
6375                 cache->reserved -= num_bytes;
6376                 cache->space_info->bytes_reserved -= num_bytes;
6377         }
6378         spin_unlock(&cache->lock);
6379         spin_unlock(&cache->space_info->lock);
6380
6381         trace_btrfs_space_reservation(fs_info, "pinned",
6382                                       cache->space_info->flags, num_bytes, 1);
6383         percpu_counter_add_batch(&cache->space_info->total_bytes_pinned,
6384                     num_bytes, BTRFS_TOTAL_BYTES_PINNED_BATCH);
6385         set_extent_dirty(fs_info->pinned_extents, bytenr,
6386                          bytenr + num_bytes - 1, GFP_NOFS | __GFP_NOFAIL);
6387         return 0;
6388 }
6389
6390 /*
6391  * this function must be called within transaction
6392  */
6393 int btrfs_pin_extent(struct btrfs_fs_info *fs_info,
6394                      u64 bytenr, u64 num_bytes, int reserved)
6395 {
6396         struct btrfs_block_group_cache *cache;
6397
6398         cache = btrfs_lookup_block_group(fs_info, bytenr);
6399         BUG_ON(!cache); /* Logic error */
6400
6401         pin_down_extent(cache, bytenr, num_bytes, reserved);
6402
6403         btrfs_put_block_group(cache);
6404         return 0;
6405 }
6406
6407 /*
6408  * this function must be called within transaction
6409  */
6410 int btrfs_pin_extent_for_log_replay(struct btrfs_fs_info *fs_info,
6411                                     u64 bytenr, u64 num_bytes)
6412 {
6413         struct btrfs_block_group_cache *cache;
6414         int ret;
6415
6416         cache = btrfs_lookup_block_group(fs_info, bytenr);
6417         if (!cache)
6418                 return -EINVAL;
6419
6420         /*
6421          * pull in the free space cache (if any) so that our pin
6422          * removes the free space from the cache.  We have load_only set
6423          * to one because the slow code to read in the free extents does check
6424          * the pinned extents.
6425          */
6426         cache_block_group(cache, 1);
6427
6428         pin_down_extent(cache, bytenr, num_bytes, 0);
6429
6430         /* remove us from the free space cache (if we're there at all) */
6431         ret = btrfs_remove_free_space(cache, bytenr, num_bytes);
6432         btrfs_put_block_group(cache);
6433         return ret;
6434 }
6435
6436 static int __exclude_logged_extent(struct btrfs_fs_info *fs_info,
6437                                    u64 start, u64 num_bytes)
6438 {
6439         int ret;
6440         struct btrfs_block_group_cache *block_group;
6441         struct btrfs_caching_control *caching_ctl;
6442
6443         block_group = btrfs_lookup_block_group(fs_info, start);
6444         if (!block_group)
6445                 return -EINVAL;
6446
6447         cache_block_group(block_group, 0);
6448         caching_ctl = get_caching_control(block_group);
6449
6450         if (!caching_ctl) {
6451                 /* Logic error */
6452                 BUG_ON(!block_group_cache_done(block_group));
6453                 ret = btrfs_remove_free_space(block_group, start, num_bytes);
6454         } else {
6455                 mutex_lock(&caching_ctl->mutex);
6456
6457                 if (start >= caching_ctl->progress) {
6458                         ret = add_excluded_extent(fs_info, start, num_bytes);
6459                 } else if (start + num_bytes <= caching_ctl->progress) {
6460                         ret = btrfs_remove_free_space(block_group,
6461                                                       start, num_bytes);
6462                 } else {
6463                         num_bytes = caching_ctl->progress - start;
6464                         ret = btrfs_remove_free_space(block_group,
6465                                                       start, num_bytes);
6466                         if (ret)
6467                                 goto out_lock;
6468
6469                         num_bytes = (start + num_bytes) -
6470                                 caching_ctl->progress;
6471                         start = caching_ctl->progress;
6472                         ret = add_excluded_extent(fs_info, start, num_bytes);
6473                 }
6474 out_lock:
6475                 mutex_unlock(&caching_ctl->mutex);
6476                 put_caching_control(caching_ctl);
6477         }
6478         btrfs_put_block_group(block_group);
6479         return ret;
6480 }
6481
6482 int btrfs_exclude_logged_extents(struct extent_buffer *eb)
6483 {
6484         struct btrfs_fs_info *fs_info = eb->fs_info;
6485         struct btrfs_file_extent_item *item;
6486         struct btrfs_key key;
6487         int found_type;
6488         int i;
6489         int ret = 0;
6490
6491         if (!btrfs_fs_incompat(fs_info, MIXED_GROUPS))
6492                 return 0;
6493
6494         for (i = 0; i < btrfs_header_nritems(eb); i++) {
6495                 btrfs_item_key_to_cpu(eb, &key, i);
6496                 if (key.type != BTRFS_EXTENT_DATA_KEY)
6497                         continue;
6498                 item = btrfs_item_ptr(eb, i, struct btrfs_file_extent_item);
6499                 found_type = btrfs_file_extent_type(eb, item);
6500                 if (found_type == BTRFS_FILE_EXTENT_INLINE)
6501                         continue;
6502                 if (btrfs_file_extent_disk_bytenr(eb, item) == 0)
6503                         continue;
6504                 key.objectid = btrfs_file_extent_disk_bytenr(eb, item);
6505                 key.offset = btrfs_file_extent_disk_num_bytes(eb, item);
6506                 ret = __exclude_logged_extent(fs_info, key.objectid, key.offset);
6507                 if (ret)
6508                         break;
6509         }
6510
6511         return ret;
6512 }
6513
6514 static void
6515 btrfs_inc_block_group_reservations(struct btrfs_block_group_cache *bg)
6516 {
6517         atomic_inc(&bg->reservations);
6518 }
6519
6520 void btrfs_dec_block_group_reservations(struct btrfs_fs_info *fs_info,
6521                                         const u64 start)
6522 {
6523         struct btrfs_block_group_cache *bg;
6524
6525         bg = btrfs_lookup_block_group(fs_info, start);
6526         ASSERT(bg);
6527         if (atomic_dec_and_test(&bg->reservations))
6528                 wake_up_var(&bg->reservations);
6529         btrfs_put_block_group(bg);
6530 }
6531
6532 void btrfs_wait_block_group_reservations(struct btrfs_block_group_cache *bg)
6533 {
6534         struct btrfs_space_info *space_info = bg->space_info;
6535
6536         ASSERT(bg->ro);
6537
6538         if (!(bg->flags & BTRFS_BLOCK_GROUP_DATA))
6539                 return;
6540
6541         /*
6542          * Our block group is read only but before we set it to read only,
6543          * some task might have had allocated an extent from it already, but it
6544          * has not yet created a respective ordered extent (and added it to a
6545          * root's list of ordered extents).
6546          * Therefore wait for any task currently allocating extents, since the
6547          * block group's reservations counter is incremented while a read lock
6548          * on the groups' semaphore is held and decremented after releasing
6549          * the read access on that semaphore and creating the ordered extent.
6550          */
6551         down_write(&space_info->groups_sem);
6552         up_write(&space_info->groups_sem);
6553
6554         wait_var_event(&bg->reservations, !atomic_read(&bg->reservations));
6555 }
6556
6557 /**
6558  * btrfs_add_reserved_bytes - update the block_group and space info counters
6559  * @cache:      The cache we are manipulating
6560  * @ram_bytes:  The number of bytes of file content, and will be same to
6561  *              @num_bytes except for the compress path.
6562  * @num_bytes:  The number of bytes in question
6563  * @delalloc:   The blocks are allocated for the delalloc write
6564  *
6565  * This is called by the allocator when it reserves space. If this is a
6566  * reservation and the block group has become read only we cannot make the
6567  * reservation and return -EAGAIN, otherwise this function always succeeds.
6568  */
6569 static int btrfs_add_reserved_bytes(struct btrfs_block_group_cache *cache,
6570                                     u64 ram_bytes, u64 num_bytes, int delalloc)
6571 {
6572         struct btrfs_space_info *space_info = cache->space_info;
6573         int ret = 0;
6574
6575         spin_lock(&space_info->lock);
6576         spin_lock(&cache->lock);
6577         if (cache->ro) {
6578                 ret = -EAGAIN;
6579         } else {
6580                 cache->reserved += num_bytes;
6581                 space_info->bytes_reserved += num_bytes;
6582                 update_bytes_may_use(space_info, -ram_bytes);
6583                 if (delalloc)
6584                         cache->delalloc_bytes += num_bytes;
6585         }
6586         spin_unlock(&cache->lock);
6587         spin_unlock(&space_info->lock);
6588         return ret;
6589 }
6590
6591 /**
6592  * btrfs_free_reserved_bytes - update the block_group and space info counters
6593  * @cache:      The cache we are manipulating
6594  * @num_bytes:  The number of bytes in question
6595  * @delalloc:   The blocks are allocated for the delalloc write
6596  *
6597  * This is called by somebody who is freeing space that was never actually used
6598  * on disk.  For example if you reserve some space for a new leaf in transaction
6599  * A and before transaction A commits you free that leaf, you call this with
6600  * reserve set to 0 in order to clear the reservation.
6601  */
6602
6603 static void btrfs_free_reserved_bytes(struct btrfs_block_group_cache *cache,
6604                                       u64 num_bytes, int delalloc)
6605 {
6606         struct btrfs_space_info *space_info = cache->space_info;
6607
6608         spin_lock(&space_info->lock);
6609         spin_lock(&cache->lock);
6610         if (cache->ro)
6611                 space_info->bytes_readonly += num_bytes;
6612         cache->reserved -= num_bytes;
6613         space_info->bytes_reserved -= num_bytes;
6614         space_info->max_extent_size = 0;
6615
6616         if (delalloc)
6617                 cache->delalloc_bytes -= num_bytes;
6618         spin_unlock(&cache->lock);
6619         spin_unlock(&space_info->lock);
6620 }
6621 void btrfs_prepare_extent_commit(struct btrfs_fs_info *fs_info)
6622 {
6623         struct btrfs_caching_control *next;
6624         struct btrfs_caching_control *caching_ctl;
6625         struct btrfs_block_group_cache *cache;
6626
6627         down_write(&fs_info->commit_root_sem);
6628
6629         list_for_each_entry_safe(caching_ctl, next,
6630                                  &fs_info->caching_block_groups, list) {
6631                 cache = caching_ctl->block_group;
6632                 if (block_group_cache_done(cache)) {
6633                         cache->last_byte_to_unpin = (u64)-1;
6634                         list_del_init(&caching_ctl->list);
6635                         put_caching_control(caching_ctl);
6636                 } else {
6637                         cache->last_byte_to_unpin = caching_ctl->progress;
6638                 }
6639         }
6640
6641         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
6642                 fs_info->pinned_extents = &fs_info->freed_extents[1];
6643         else
6644                 fs_info->pinned_extents = &fs_info->freed_extents[0];
6645
6646         up_write(&fs_info->commit_root_sem);
6647
6648         update_global_block_rsv(fs_info);
6649 }
6650
6651 /*
6652  * Returns the free cluster for the given space info and sets empty_cluster to
6653  * what it should be based on the mount options.
6654  */
6655 static struct btrfs_free_cluster *
6656 fetch_cluster_info(struct btrfs_fs_info *fs_info,
6657                    struct btrfs_space_info *space_info, u64 *empty_cluster)
6658 {
6659         struct btrfs_free_cluster *ret = NULL;
6660
6661         *empty_cluster = 0;
6662         if (btrfs_mixed_space_info(space_info))
6663                 return ret;
6664
6665         if (space_info->flags & BTRFS_BLOCK_GROUP_METADATA) {
6666                 ret = &fs_info->meta_alloc_cluster;
6667                 if (btrfs_test_opt(fs_info, SSD))
6668                         *empty_cluster = SZ_2M;
6669                 else
6670                         *empty_cluster = SZ_64K;
6671         } else if ((space_info->flags & BTRFS_BLOCK_GROUP_DATA) &&
6672                    btrfs_test_opt(fs_info, SSD_SPREAD)) {
6673                 *empty_cluster = SZ_2M;
6674                 ret = &fs_info->data_alloc_cluster;
6675         }
6676
6677         return ret;
6678 }
6679
6680 static int unpin_extent_range(struct btrfs_fs_info *fs_info,
6681                               u64 start, u64 end,
6682                               const bool return_free_space)
6683 {
6684         struct btrfs_block_group_cache *cache = NULL;
6685         struct btrfs_space_info *space_info;
6686         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
6687         struct btrfs_free_cluster *cluster = NULL;
6688         u64 len;
6689         u64 total_unpinned = 0;
6690         u64 empty_cluster = 0;
6691         bool readonly;
6692
6693         while (start <= end) {
6694                 readonly = false;
6695                 if (!cache ||
6696                     start >= cache->key.objectid + cache->key.offset) {
6697                         if (cache)
6698                                 btrfs_put_block_group(cache);
6699                         total_unpinned = 0;
6700                         cache = btrfs_lookup_block_group(fs_info, start);
6701                         BUG_ON(!cache); /* Logic error */
6702
6703                         cluster = fetch_cluster_info(fs_info,
6704                                                      cache->space_info,
6705                                                      &empty_cluster);
6706                         empty_cluster <<= 1;
6707                 }
6708
6709                 len = cache->key.objectid + cache->key.offset - start;
6710                 len = min(len, end + 1 - start);
6711
6712                 if (start < cache->last_byte_to_unpin) {
6713                         len = min(len, cache->last_byte_to_unpin - start);
6714                         if (return_free_space)
6715                                 btrfs_add_free_space(cache, start, len);
6716                 }
6717
6718                 start += len;
6719                 total_unpinned += len;
6720                 space_info = cache->space_info;
6721
6722                 /*
6723                  * If this space cluster has been marked as fragmented and we've
6724                  * unpinned enough in this block group to potentially allow a
6725                  * cluster to be created inside of it go ahead and clear the
6726                  * fragmented check.
6727                  */
6728                 if (cluster && cluster->fragmented &&
6729                     total_unpinned > empty_cluster) {
6730                         spin_lock(&cluster->lock);
6731                         cluster->fragmented = 0;
6732                         spin_unlock(&cluster->lock);
6733                 }
6734
6735                 spin_lock(&space_info->lock);
6736                 spin_lock(&cache->lock);
6737                 cache->pinned -= len;
6738                 update_bytes_pinned(space_info, -len);
6739
6740                 trace_btrfs_space_reservation(fs_info, "pinned",
6741                                               space_info->flags, len, 0);
6742                 space_info->max_extent_size = 0;
6743                 percpu_counter_add_batch(&space_info->total_bytes_pinned,
6744                             -len, BTRFS_TOTAL_BYTES_PINNED_BATCH);
6745                 if (cache->ro) {
6746                         space_info->bytes_readonly += len;
6747                         readonly = true;
6748                 }
6749                 spin_unlock(&cache->lock);
6750                 if (!readonly && return_free_space &&
6751                     global_rsv->space_info == space_info) {
6752                         u64 to_add = len;
6753
6754                         spin_lock(&global_rsv->lock);
6755                         if (!global_rsv->full) {
6756                                 to_add = min(len, global_rsv->size -
6757                                              global_rsv->reserved);
6758                                 global_rsv->reserved += to_add;
6759                                 update_bytes_may_use(space_info, to_add);
6760                                 if (global_rsv->reserved >= global_rsv->size)
6761                                         global_rsv->full = 1;
6762                                 trace_btrfs_space_reservation(fs_info,
6763                                                               "space_info",
6764                                                               space_info->flags,
6765                                                               to_add, 1);
6766                                 len -= to_add;
6767                         }
6768                         spin_unlock(&global_rsv->lock);
6769                         /* Add to any tickets we may have */
6770                         if (len)
6771                                 space_info_add_new_bytes(fs_info, space_info,
6772                                                          len);
6773                 }
6774                 spin_unlock(&space_info->lock);
6775         }
6776
6777         if (cache)
6778                 btrfs_put_block_group(cache);
6779         return 0;
6780 }
6781
6782 int btrfs_finish_extent_commit(struct btrfs_trans_handle *trans)
6783 {
6784         struct btrfs_fs_info *fs_info = trans->fs_info;
6785         struct btrfs_block_group_cache *block_group, *tmp;
6786         struct list_head *deleted_bgs;
6787         struct extent_io_tree *unpin;
6788         u64 start;
6789         u64 end;
6790         int ret;
6791
6792         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
6793                 unpin = &fs_info->freed_extents[1];
6794         else
6795                 unpin = &fs_info->freed_extents[0];
6796
6797         while (!trans->aborted) {
6798                 struct extent_state *cached_state = NULL;
6799
6800                 mutex_lock(&fs_info->unused_bg_unpin_mutex);
6801                 ret = find_first_extent_bit(unpin, 0, &start, &end,
6802                                             EXTENT_DIRTY, &cached_state);
6803                 if (ret) {
6804                         mutex_unlock(&fs_info->unused_bg_unpin_mutex);
6805                         break;
6806                 }
6807
6808                 if (btrfs_test_opt(fs_info, DISCARD))
6809                         ret = btrfs_discard_extent(fs_info, start,
6810                                                    end + 1 - start, NULL);
6811
6812                 clear_extent_dirty(unpin, start, end, &cached_state);
6813                 unpin_extent_range(fs_info, start, end, true);
6814                 mutex_unlock(&fs_info->unused_bg_unpin_mutex);
6815                 free_extent_state(cached_state);
6816                 cond_resched();
6817         }
6818
6819         /*
6820          * Transaction is finished.  We don't need the lock anymore.  We
6821          * do need to clean up the block groups in case of a transaction
6822          * abort.
6823          */
6824         deleted_bgs = &trans->transaction->deleted_bgs;
6825         list_for_each_entry_safe(block_group, tmp, deleted_bgs, bg_list) {
6826                 u64 trimmed = 0;
6827
6828                 ret = -EROFS;
6829                 if (!trans->aborted)
6830                         ret = btrfs_discard_extent(fs_info,
6831                                                    block_group->key.objectid,
6832                                                    block_group->key.offset,
6833                                                    &trimmed);
6834
6835                 list_del_init(&block_group->bg_list);
6836                 btrfs_put_block_group_trimming(block_group);
6837                 btrfs_put_block_group(block_group);
6838
6839                 if (ret) {
6840                         const char *errstr = btrfs_decode_error(ret);
6841                         btrfs_warn(fs_info,
6842                            "discard failed while removing blockgroup: errno=%d %s",
6843                                    ret, errstr);
6844                 }
6845         }
6846
6847         return 0;
6848 }
6849
6850 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
6851                                struct btrfs_delayed_ref_node *node, u64 parent,
6852                                u64 root_objectid, u64 owner_objectid,
6853                                u64 owner_offset, int refs_to_drop,
6854                                struct btrfs_delayed_extent_op *extent_op)
6855 {
6856         struct btrfs_fs_info *info = trans->fs_info;
6857         struct btrfs_key key;
6858         struct btrfs_path *path;
6859         struct btrfs_root *extent_root = info->extent_root;
6860         struct extent_buffer *leaf;
6861         struct btrfs_extent_item *ei;
6862         struct btrfs_extent_inline_ref *iref;
6863         int ret;
6864         int is_data;
6865         int extent_slot = 0;
6866         int found_extent = 0;
6867         int num_to_del = 1;
6868         u32 item_size;
6869         u64 refs;
6870         u64 bytenr = node->bytenr;
6871         u64 num_bytes = node->num_bytes;
6872         int last_ref = 0;
6873         bool skinny_metadata = btrfs_fs_incompat(info, SKINNY_METADATA);
6874
6875         path = btrfs_alloc_path();
6876         if (!path)
6877                 return -ENOMEM;
6878
6879         path->reada = READA_FORWARD;
6880         path->leave_spinning = 1;
6881
6882         is_data = owner_objectid >= BTRFS_FIRST_FREE_OBJECTID;
6883         BUG_ON(!is_data && refs_to_drop != 1);
6884
6885         if (is_data)
6886                 skinny_metadata = false;
6887
6888         ret = lookup_extent_backref(trans, path, &iref, bytenr, num_bytes,
6889                                     parent, root_objectid, owner_objectid,
6890                                     owner_offset);
6891         if (ret == 0) {
6892                 extent_slot = path->slots[0];
6893                 while (extent_slot >= 0) {
6894                         btrfs_item_key_to_cpu(path->nodes[0], &key,
6895                                               extent_slot);
6896                         if (key.objectid != bytenr)
6897                                 break;
6898                         if (key.type == BTRFS_EXTENT_ITEM_KEY &&
6899                             key.offset == num_bytes) {
6900                                 found_extent = 1;
6901                                 break;
6902                         }
6903                         if (key.type == BTRFS_METADATA_ITEM_KEY &&
6904                             key.offset == owner_objectid) {
6905                                 found_extent = 1;
6906                                 break;
6907                         }
6908                         if (path->slots[0] - extent_slot > 5)
6909                                 break;
6910                         extent_slot--;
6911                 }
6912
6913                 if (!found_extent) {
6914                         BUG_ON(iref);
6915                         ret = remove_extent_backref(trans, path, NULL,
6916                                                     refs_to_drop,
6917                                                     is_data, &last_ref);
6918                         if (ret) {
6919                                 btrfs_abort_transaction(trans, ret);
6920                                 goto out;
6921                         }
6922                         btrfs_release_path(path);
6923                         path->leave_spinning = 1;
6924
6925                         key.objectid = bytenr;
6926                         key.type = BTRFS_EXTENT_ITEM_KEY;
6927                         key.offset = num_bytes;
6928
6929                         if (!is_data && skinny_metadata) {
6930                                 key.type = BTRFS_METADATA_ITEM_KEY;
6931                                 key.offset = owner_objectid;
6932                         }
6933
6934                         ret = btrfs_search_slot(trans, extent_root,
6935                                                 &key, path, -1, 1);
6936                         if (ret > 0 && skinny_metadata && path->slots[0]) {
6937                                 /*
6938                                  * Couldn't find our skinny metadata item,
6939                                  * see if we have ye olde extent item.
6940                                  */
6941                                 path->slots[0]--;
6942                                 btrfs_item_key_to_cpu(path->nodes[0], &key,
6943                                                       path->slots[0]);
6944                                 if (key.objectid == bytenr &&
6945                                     key.type == BTRFS_EXTENT_ITEM_KEY &&
6946                                     key.offset == num_bytes)
6947                                         ret = 0;
6948                         }
6949
6950                         if (ret > 0 && skinny_metadata) {
6951                                 skinny_metadata = false;
6952                                 key.objectid = bytenr;
6953                                 key.type = BTRFS_EXTENT_ITEM_KEY;
6954                                 key.offset = num_bytes;
6955                                 btrfs_release_path(path);
6956                                 ret = btrfs_search_slot(trans, extent_root,
6957                                                         &key, path, -1, 1);
6958                         }
6959
6960                         if (ret) {
6961                                 btrfs_err(info,
6962                                           "umm, got %d back from search, was looking for %llu",
6963                                           ret, bytenr);
6964                                 if (ret > 0)
6965                                         btrfs_print_leaf(path->nodes[0]);
6966                         }
6967                         if (ret < 0) {
6968                                 btrfs_abort_transaction(trans, ret);
6969                                 goto out;
6970                         }
6971                         extent_slot = path->slots[0];
6972                 }
6973         } else if (WARN_ON(ret == -ENOENT)) {
6974                 btrfs_print_leaf(path->nodes[0]);
6975                 btrfs_err(info,
6976                         "unable to find ref byte nr %llu parent %llu root %llu  owner %llu offset %llu",
6977                         bytenr, parent, root_objectid, owner_objectid,
6978                         owner_offset);
6979                 btrfs_abort_transaction(trans, ret);
6980                 goto out;
6981         } else {
6982                 btrfs_abort_transaction(trans, ret);
6983                 goto out;
6984         }
6985
6986         leaf = path->nodes[0];
6987         item_size = btrfs_item_size_nr(leaf, extent_slot);
6988         if (unlikely(item_size < sizeof(*ei))) {
6989                 ret = -EINVAL;
6990                 btrfs_print_v0_err(info);
6991                 btrfs_abort_transaction(trans, ret);
6992                 goto out;
6993         }
6994         ei = btrfs_item_ptr(leaf, extent_slot,
6995                             struct btrfs_extent_item);
6996         if (owner_objectid < BTRFS_FIRST_FREE_OBJECTID &&
6997             key.type == BTRFS_EXTENT_ITEM_KEY) {
6998                 struct btrfs_tree_block_info *bi;
6999                 BUG_ON(item_size < sizeof(*ei) + sizeof(*bi));
7000                 bi = (struct btrfs_tree_block_info *)(ei + 1);
7001                 WARN_ON(owner_objectid != btrfs_tree_block_level(leaf, bi));
7002         }
7003
7004         refs = btrfs_extent_refs(leaf, ei);
7005         if (refs < refs_to_drop) {
7006                 btrfs_err(info,
7007                           "trying to drop %d refs but we only have %Lu for bytenr %Lu",
7008                           refs_to_drop, refs, bytenr);
7009                 ret = -EINVAL;
7010                 btrfs_abort_transaction(trans, ret);
7011                 goto out;
7012         }
7013         refs -= refs_to_drop;
7014
7015         if (refs > 0) {
7016                 if (extent_op)
7017                         __run_delayed_extent_op(extent_op, leaf, ei);
7018                 /*
7019                  * In the case of inline back ref, reference count will
7020                  * be updated by remove_extent_backref
7021                  */
7022                 if (iref) {
7023                         BUG_ON(!found_extent);
7024                 } else {
7025                         btrfs_set_extent_refs(leaf, ei, refs);
7026                         btrfs_mark_buffer_dirty(leaf);
7027                 }
7028                 if (found_extent) {
7029                         ret = remove_extent_backref(trans, path, iref,
7030                                                     refs_to_drop, is_data,
7031                                                     &last_ref);
7032                         if (ret) {
7033                                 btrfs_abort_transaction(trans, ret);
7034                                 goto out;
7035                         }
7036                 }
7037         } else {
7038                 if (found_extent) {
7039                         BUG_ON(is_data && refs_to_drop !=
7040                                extent_data_ref_count(path, iref));
7041                         if (iref) {
7042                                 BUG_ON(path->slots[0] != extent_slot);
7043                         } else {
7044                                 BUG_ON(path->slots[0] != extent_slot + 1);
7045                                 path->slots[0] = extent_slot;
7046                                 num_to_del = 2;
7047                         }
7048                 }
7049
7050                 last_ref = 1;
7051                 ret = btrfs_del_items(trans, extent_root, path, path->slots[0],
7052                                       num_to_del);
7053                 if (ret) {
7054                         btrfs_abort_transaction(trans, ret);
7055                         goto out;
7056                 }
7057                 btrfs_release_path(path);
7058
7059                 if (is_data) {
7060                         ret = btrfs_del_csums(trans, info, bytenr, num_bytes);
7061                         if (ret) {
7062                                 btrfs_abort_transaction(trans, ret);
7063                                 goto out;
7064                         }
7065                 }
7066
7067                 ret = add_to_free_space_tree(trans, bytenr, num_bytes);
7068                 if (ret) {
7069                         btrfs_abort_transaction(trans, ret);
7070                         goto out;
7071                 }
7072
7073                 ret = update_block_group(trans, bytenr, num_bytes, 0);
7074                 if (ret) {
7075                         btrfs_abort_transaction(trans, ret);
7076                         goto out;
7077                 }
7078         }
7079         btrfs_release_path(path);
7080
7081 out:
7082         btrfs_free_path(path);
7083         return ret;
7084 }
7085
7086 /*
7087  * when we free an block, it is possible (and likely) that we free the last
7088  * delayed ref for that extent as well.  This searches the delayed ref tree for
7089  * a given extent, and if there are no other delayed refs to be processed, it
7090  * removes it from the tree.
7091  */
7092 static noinline int check_ref_cleanup(struct btrfs_trans_handle *trans,
7093                                       u64 bytenr)
7094 {
7095         struct btrfs_delayed_ref_head *head;
7096         struct btrfs_delayed_ref_root *delayed_refs;
7097         int ret = 0;
7098
7099         delayed_refs = &trans->transaction->delayed_refs;
7100         spin_lock(&delayed_refs->lock);
7101         head = btrfs_find_delayed_ref_head(delayed_refs, bytenr);
7102         if (!head)
7103                 goto out_delayed_unlock;
7104
7105         spin_lock(&head->lock);
7106         if (!RB_EMPTY_ROOT(&head->ref_tree.rb_root))
7107                 goto out;
7108
7109         if (cleanup_extent_op(head) != NULL)
7110                 goto out;
7111
7112         /*
7113          * waiting for the lock here would deadlock.  If someone else has it
7114          * locked they are already in the process of dropping it anyway
7115          */
7116         if (!mutex_trylock(&head->mutex))
7117                 goto out;
7118
7119         btrfs_delete_ref_head(delayed_refs, head);
7120         head->processing = 0;
7121
7122         spin_unlock(&head->lock);
7123         spin_unlock(&delayed_refs->lock);
7124
7125         BUG_ON(head->extent_op);
7126         if (head->must_insert_reserved)
7127                 ret = 1;
7128
7129         btrfs_cleanup_ref_head_accounting(trans->fs_info, delayed_refs, head);
7130         mutex_unlock(&head->mutex);
7131         btrfs_put_delayed_ref_head(head);
7132         return ret;
7133 out:
7134         spin_unlock(&head->lock);
7135
7136 out_delayed_unlock:
7137         spin_unlock(&delayed_refs->lock);
7138         return 0;
7139 }
7140
7141 void btrfs_free_tree_block(struct btrfs_trans_handle *trans,
7142                            struct btrfs_root *root,
7143                            struct extent_buffer *buf,
7144                            u64 parent, int last_ref)
7145 {
7146         struct btrfs_fs_info *fs_info = root->fs_info;
7147         struct btrfs_ref generic_ref = { 0 };
7148         int pin = 1;
7149         int ret;
7150
7151         btrfs_init_generic_ref(&generic_ref, BTRFS_DROP_DELAYED_REF,
7152                                buf->start, buf->len, parent);
7153         btrfs_init_tree_ref(&generic_ref, btrfs_header_level(buf),
7154                             root->root_key.objectid);
7155
7156         if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
7157                 int old_ref_mod, new_ref_mod;
7158
7159                 btrfs_ref_tree_mod(fs_info, &generic_ref);
7160                 ret = btrfs_add_delayed_tree_ref(trans, &generic_ref, NULL,
7161                                                  &old_ref_mod, &new_ref_mod);
7162                 BUG_ON(ret); /* -ENOMEM */
7163                 pin = old_ref_mod >= 0 && new_ref_mod < 0;
7164         }
7165
7166         if (last_ref && btrfs_header_generation(buf) == trans->transid) {
7167                 struct btrfs_block_group_cache *cache;
7168
7169                 if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
7170                         ret = check_ref_cleanup(trans, buf->start);
7171                         if (!ret)
7172                                 goto out;
7173                 }
7174
7175                 pin = 0;
7176                 cache = btrfs_lookup_block_group(fs_info, buf->start);
7177
7178                 if (btrfs_header_flag(buf, BTRFS_HEADER_FLAG_WRITTEN)) {
7179                         pin_down_extent(cache, buf->start, buf->len, 1);
7180                         btrfs_put_block_group(cache);
7181                         goto out;
7182                 }
7183
7184                 WARN_ON(test_bit(EXTENT_BUFFER_DIRTY, &buf->bflags));
7185
7186                 btrfs_add_free_space(cache, buf->start, buf->len);
7187                 btrfs_free_reserved_bytes(cache, buf->len, 0);
7188                 btrfs_put_block_group(cache);
7189                 trace_btrfs_reserved_extent_free(fs_info, buf->start, buf->len);
7190         }
7191 out:
7192         if (pin)
7193                 add_pinned_bytes(fs_info, &generic_ref);
7194
7195         if (last_ref) {
7196                 /*
7197                  * Deleting the buffer, clear the corrupt flag since it doesn't
7198                  * matter anymore.
7199                  */
7200                 clear_bit(EXTENT_BUFFER_CORRUPT, &buf->bflags);
7201         }
7202 }
7203
7204 /* Can return -ENOMEM */
7205 int btrfs_free_extent(struct btrfs_trans_handle *trans, struct btrfs_ref *ref)
7206 {
7207         struct btrfs_fs_info *fs_info = trans->fs_info;
7208         int old_ref_mod, new_ref_mod;
7209         int ret;
7210
7211         if (btrfs_is_testing(fs_info))
7212                 return 0;
7213
7214         /*
7215          * tree log blocks never actually go into the extent allocation
7216          * tree, just update pinning info and exit early.
7217          */
7218         if ((ref->type == BTRFS_REF_METADATA &&
7219              ref->tree_ref.root == BTRFS_TREE_LOG_OBJECTID) ||
7220             (ref->type == BTRFS_REF_DATA &&
7221              ref->data_ref.ref_root == BTRFS_TREE_LOG_OBJECTID)) {
7222                 /* unlocks the pinned mutex */
7223                 btrfs_pin_extent(fs_info, ref->bytenr, ref->len, 1);
7224                 old_ref_mod = new_ref_mod = 0;
7225                 ret = 0;
7226         } else if (ref->type == BTRFS_REF_METADATA) {
7227                 ret = btrfs_add_delayed_tree_ref(trans, ref, NULL,
7228                                                  &old_ref_mod, &new_ref_mod);
7229         } else {
7230                 ret = btrfs_add_delayed_data_ref(trans, ref, 0,
7231                                                  &old_ref_mod, &new_ref_mod);
7232         }
7233
7234         if (!((ref->type == BTRFS_REF_METADATA &&
7235                ref->tree_ref.root == BTRFS_TREE_LOG_OBJECTID) ||
7236               (ref->type == BTRFS_REF_DATA &&
7237                ref->data_ref.ref_root == BTRFS_TREE_LOG_OBJECTID)))
7238                 btrfs_ref_tree_mod(fs_info, ref);
7239
7240         if (ret == 0 && old_ref_mod >= 0 && new_ref_mod < 0)
7241                 add_pinned_bytes(fs_info, ref);
7242
7243         return ret;
7244 }
7245
7246 /*
7247  * when we wait for progress in the block group caching, its because
7248  * our allocation attempt failed at least once.  So, we must sleep
7249  * and let some progress happen before we try again.
7250  *
7251  * This function will sleep at least once waiting for new free space to
7252  * show up, and then it will check the block group free space numbers
7253  * for our min num_bytes.  Another option is to have it go ahead
7254  * and look in the rbtree for a free extent of a given size, but this
7255  * is a good start.
7256  *
7257  * Callers of this must check if cache->cached == BTRFS_CACHE_ERROR before using
7258  * any of the information in this block group.
7259  */
7260 static noinline void
7261 wait_block_group_cache_progress(struct btrfs_block_group_cache *cache,
7262                                 u64 num_bytes)
7263 {
7264         struct btrfs_caching_control *caching_ctl;
7265
7266         caching_ctl = get_caching_control(cache);
7267         if (!caching_ctl)
7268                 return;
7269
7270         wait_event(caching_ctl->wait, block_group_cache_done(cache) ||
7271                    (cache->free_space_ctl->free_space >= num_bytes));
7272
7273         put_caching_control(caching_ctl);
7274 }
7275
7276 static noinline int
7277 wait_block_group_cache_done(struct btrfs_block_group_cache *cache)
7278 {
7279         struct btrfs_caching_control *caching_ctl;
7280         int ret = 0;
7281
7282         caching_ctl = get_caching_control(cache);
7283         if (!caching_ctl)
7284                 return (cache->cached == BTRFS_CACHE_ERROR) ? -EIO : 0;
7285
7286         wait_event(caching_ctl->wait, block_group_cache_done(cache));
7287         if (cache->cached == BTRFS_CACHE_ERROR)
7288                 ret = -EIO;
7289         put_caching_control(caching_ctl);
7290         return ret;
7291 }
7292
7293 enum btrfs_loop_type {
7294         LOOP_CACHING_NOWAIT = 0,
7295         LOOP_CACHING_WAIT = 1,
7296         LOOP_ALLOC_CHUNK = 2,
7297         LOOP_NO_EMPTY_SIZE = 3,
7298 };
7299
7300 static inline void
7301 btrfs_lock_block_group(struct btrfs_block_group_cache *cache,
7302                        int delalloc)
7303 {
7304         if (delalloc)
7305                 down_read(&cache->data_rwsem);
7306 }
7307
7308 static inline void
7309 btrfs_grab_block_group(struct btrfs_block_group_cache *cache,
7310                        int delalloc)
7311 {
7312         btrfs_get_block_group(cache);
7313         if (delalloc)
7314                 down_read(&cache->data_rwsem);
7315 }
7316
7317 static struct btrfs_block_group_cache *
7318 btrfs_lock_cluster(struct btrfs_block_group_cache *block_group,
7319                    struct btrfs_free_cluster *cluster,
7320                    int delalloc)
7321 {
7322         struct btrfs_block_group_cache *used_bg = NULL;
7323
7324         spin_lock(&cluster->refill_lock);
7325         while (1) {
7326                 used_bg = cluster->block_group;
7327                 if (!used_bg)
7328                         return NULL;
7329
7330                 if (used_bg == block_group)
7331                         return used_bg;
7332
7333                 btrfs_get_block_group(used_bg);
7334
7335                 if (!delalloc)
7336                         return used_bg;
7337
7338                 if (down_read_trylock(&used_bg->data_rwsem))
7339                         return used_bg;
7340
7341                 spin_unlock(&cluster->refill_lock);
7342
7343                 /* We should only have one-level nested. */
7344                 down_read_nested(&used_bg->data_rwsem, SINGLE_DEPTH_NESTING);
7345
7346                 spin_lock(&cluster->refill_lock);
7347                 if (used_bg == cluster->block_group)
7348                         return used_bg;
7349
7350                 up_read(&used_bg->data_rwsem);
7351                 btrfs_put_block_group(used_bg);
7352         }
7353 }
7354
7355 static inline void
7356 btrfs_release_block_group(struct btrfs_block_group_cache *cache,
7357                          int delalloc)
7358 {
7359         if (delalloc)
7360                 up_read(&cache->data_rwsem);
7361         btrfs_put_block_group(cache);
7362 }
7363
7364 /*
7365  * Structure used internally for find_free_extent() function.  Wraps needed
7366  * parameters.
7367  */
7368 struct find_free_extent_ctl {
7369         /* Basic allocation info */
7370         u64 ram_bytes;
7371         u64 num_bytes;
7372         u64 empty_size;
7373         u64 flags;
7374         int delalloc;
7375
7376         /* Where to start the search inside the bg */
7377         u64 search_start;
7378
7379         /* For clustered allocation */
7380         u64 empty_cluster;
7381
7382         bool have_caching_bg;
7383         bool orig_have_caching_bg;
7384
7385         /* RAID index, converted from flags */
7386         int index;
7387
7388         /*
7389          * Current loop number, check find_free_extent_update_loop() for details
7390          */
7391         int loop;
7392
7393         /*
7394          * Whether we're refilling a cluster, if true we need to re-search
7395          * current block group but don't try to refill the cluster again.
7396          */
7397         bool retry_clustered;
7398
7399         /*
7400          * Whether we're updating free space cache, if true we need to re-search
7401          * current block group but don't try updating free space cache again.
7402          */
7403         bool retry_unclustered;
7404
7405         /* If current block group is cached */
7406         int cached;
7407
7408         /* Max contiguous hole found */
7409         u64 max_extent_size;
7410
7411         /* Total free space from free space cache, not always contiguous */
7412         u64 total_free_space;
7413
7414         /* Found result */
7415         u64 found_offset;
7416 };
7417
7418
7419 /*
7420  * Helper function for find_free_extent().
7421  *
7422  * Return -ENOENT to inform caller that we need fallback to unclustered mode.
7423  * Return -EAGAIN to inform caller that we need to re-search this block group
7424  * Return >0 to inform caller that we find nothing
7425  * Return 0 means we have found a location and set ffe_ctl->found_offset.
7426  */
7427 static int find_free_extent_clustered(struct btrfs_block_group_cache *bg,
7428                 struct btrfs_free_cluster *last_ptr,
7429                 struct find_free_extent_ctl *ffe_ctl,
7430                 struct btrfs_block_group_cache **cluster_bg_ret)
7431 {
7432         struct btrfs_block_group_cache *cluster_bg;
7433         u64 aligned_cluster;
7434         u64 offset;
7435         int ret;
7436
7437         cluster_bg = btrfs_lock_cluster(bg, last_ptr, ffe_ctl->delalloc);
7438         if (!cluster_bg)
7439                 goto refill_cluster;
7440         if (cluster_bg != bg && (cluster_bg->ro ||
7441             !block_group_bits(cluster_bg, ffe_ctl->flags)))
7442                 goto release_cluster;
7443
7444         offset = btrfs_alloc_from_cluster(cluster_bg, last_ptr,
7445                         ffe_ctl->num_bytes, cluster_bg->key.objectid,
7446                         &ffe_ctl->max_extent_size);
7447         if (offset) {
7448                 /* We have a block, we're done */
7449                 spin_unlock(&last_ptr->refill_lock);
7450                 trace_btrfs_reserve_extent_cluster(cluster_bg,
7451                                 ffe_ctl->search_start, ffe_ctl->num_bytes);
7452                 *cluster_bg_ret = cluster_bg;
7453                 ffe_ctl->found_offset = offset;
7454                 return 0;
7455         }
7456         WARN_ON(last_ptr->block_group != cluster_bg);
7457
7458 release_cluster:
7459         /*
7460          * If we are on LOOP_NO_EMPTY_SIZE, we can't set up a new clusters, so
7461          * lets just skip it and let the allocator find whatever block it can
7462          * find. If we reach this point, we will have tried the cluster
7463          * allocator plenty of times and not have found anything, so we are
7464          * likely way too fragmented for the clustering stuff to find anything.
7465          *
7466          * However, if the cluster is taken from the current block group,
7467          * release the cluster first, so that we stand a better chance of
7468          * succeeding in the unclustered allocation.
7469          */
7470         if (ffe_ctl->loop >= LOOP_NO_EMPTY_SIZE && cluster_bg != bg) {
7471                 spin_unlock(&last_ptr->refill_lock);
7472                 btrfs_release_block_group(cluster_bg, ffe_ctl->delalloc);
7473                 return -ENOENT;
7474         }
7475
7476         /* This cluster didn't work out, free it and start over */
7477         btrfs_return_cluster_to_free_space(NULL, last_ptr);
7478
7479         if (cluster_bg != bg)
7480                 btrfs_release_block_group(cluster_bg, ffe_ctl->delalloc);
7481
7482 refill_cluster:
7483         if (ffe_ctl->loop >= LOOP_NO_EMPTY_SIZE) {
7484                 spin_unlock(&last_ptr->refill_lock);
7485                 return -ENOENT;
7486         }
7487
7488         aligned_cluster = max_t(u64,
7489                         ffe_ctl->empty_cluster + ffe_ctl->empty_size,
7490                         bg->full_stripe_len);
7491         ret = btrfs_find_space_cluster(bg, last_ptr, ffe_ctl->search_start,
7492                         ffe_ctl->num_bytes, aligned_cluster);
7493         if (ret == 0) {
7494                 /* Now pull our allocation out of this cluster */
7495                 offset = btrfs_alloc_from_cluster(bg, last_ptr,
7496                                 ffe_ctl->num_bytes, ffe_ctl->search_start,
7497                                 &ffe_ctl->max_extent_size);
7498                 if (offset) {
7499                         /* We found one, proceed */
7500                         spin_unlock(&last_ptr->refill_lock);
7501                         trace_btrfs_reserve_extent_cluster(bg,
7502                                         ffe_ctl->search_start,
7503                                         ffe_ctl->num_bytes);
7504                         ffe_ctl->found_offset = offset;
7505                         return 0;
7506                 }
7507         } else if (!ffe_ctl->cached && ffe_ctl->loop > LOOP_CACHING_NOWAIT &&
7508                    !ffe_ctl->retry_clustered) {
7509                 spin_unlock(&last_ptr->refill_lock);
7510
7511                 ffe_ctl->retry_clustered = true;
7512                 wait_block_group_cache_progress(bg, ffe_ctl->num_bytes +
7513                                 ffe_ctl->empty_cluster + ffe_ctl->empty_size);
7514                 return -EAGAIN;
7515         }
7516         /*
7517          * At this point we either didn't find a cluster or we weren't able to
7518          * allocate a block from our cluster.  Free the cluster we've been
7519          * trying to use, and go to the next block group.
7520          */
7521         btrfs_return_cluster_to_free_space(NULL, last_ptr);
7522         spin_unlock(&last_ptr->refill_lock);
7523         return 1;
7524 }
7525
7526 /*
7527  * Return >0 to inform caller that we find nothing
7528  * Return 0 when we found an free extent and set ffe_ctrl->found_offset
7529  * Return -EAGAIN to inform caller that we need to re-search this block group
7530  */
7531 static int find_free_extent_unclustered(struct btrfs_block_group_cache *bg,
7532                 struct btrfs_free_cluster *last_ptr,
7533                 struct find_free_extent_ctl *ffe_ctl)
7534 {
7535         u64 offset;
7536
7537         /*
7538          * We are doing an unclustered allocation, set the fragmented flag so
7539          * we don't bother trying to setup a cluster again until we get more
7540          * space.
7541          */
7542         if (unlikely(last_ptr)) {
7543                 spin_lock(&last_ptr->lock);
7544                 last_ptr->fragmented = 1;
7545                 spin_unlock(&last_ptr->lock);
7546         }
7547         if (ffe_ctl->cached) {
7548                 struct btrfs_free_space_ctl *free_space_ctl;
7549
7550                 free_space_ctl = bg->free_space_ctl;
7551                 spin_lock(&free_space_ctl->tree_lock);
7552                 if (free_space_ctl->free_space <
7553                     ffe_ctl->num_bytes + ffe_ctl->empty_cluster +
7554                     ffe_ctl->empty_size) {
7555                         ffe_ctl->total_free_space = max_t(u64,
7556                                         ffe_ctl->total_free_space,
7557                                         free_space_ctl->free_space);
7558                         spin_unlock(&free_space_ctl->tree_lock);
7559                         return 1;
7560                 }
7561                 spin_unlock(&free_space_ctl->tree_lock);
7562         }
7563
7564         offset = btrfs_find_space_for_alloc(bg, ffe_ctl->search_start,
7565                         ffe_ctl->num_bytes, ffe_ctl->empty_size,
7566                         &ffe_ctl->max_extent_size);
7567
7568         /*
7569          * If we didn't find a chunk, and we haven't failed on this block group
7570          * before, and this block group is in the middle of caching and we are
7571          * ok with waiting, then go ahead and wait for progress to be made, and
7572          * set @retry_unclustered to true.
7573          *
7574          * If @retry_unclustered is true then we've already waited on this
7575          * block group once and should move on to the next block group.
7576          */
7577         if (!offset && !ffe_ctl->retry_unclustered && !ffe_ctl->cached &&
7578             ffe_ctl->loop > LOOP_CACHING_NOWAIT) {
7579                 wait_block_group_cache_progress(bg, ffe_ctl->num_bytes +
7580                                                 ffe_ctl->empty_size);
7581                 ffe_ctl->retry_unclustered = true;
7582                 return -EAGAIN;
7583         } else if (!offset) {
7584                 return 1;
7585         }
7586         ffe_ctl->found_offset = offset;
7587         return 0;
7588 }
7589
7590 /*
7591  * Return >0 means caller needs to re-search for free extent
7592  * Return 0 means we have the needed free extent.
7593  * Return <0 means we failed to locate any free extent.
7594  */
7595 static int find_free_extent_update_loop(struct btrfs_fs_info *fs_info,
7596                                         struct btrfs_free_cluster *last_ptr,
7597                                         struct btrfs_key *ins,
7598                                         struct find_free_extent_ctl *ffe_ctl,
7599                                         int full_search, bool use_cluster)
7600 {
7601         struct btrfs_root *root = fs_info->extent_root;
7602         int ret;
7603
7604         if ((ffe_ctl->loop == LOOP_CACHING_NOWAIT) &&
7605             ffe_ctl->have_caching_bg && !ffe_ctl->orig_have_caching_bg)
7606                 ffe_ctl->orig_have_caching_bg = true;
7607
7608         if (!ins->objectid && ffe_ctl->loop >= LOOP_CACHING_WAIT &&
7609             ffe_ctl->have_caching_bg)
7610                 return 1;
7611
7612         if (!ins->objectid && ++(ffe_ctl->index) < BTRFS_NR_RAID_TYPES)
7613                 return 1;
7614
7615         if (ins->objectid) {
7616                 if (!use_cluster && last_ptr) {
7617                         spin_lock(&last_ptr->lock);
7618                         last_ptr->window_start = ins->objectid;
7619                         spin_unlock(&last_ptr->lock);
7620                 }
7621                 return 0;
7622         }
7623
7624         /*
7625          * LOOP_CACHING_NOWAIT, search partially cached block groups, kicking
7626          *                      caching kthreads as we move along
7627          * LOOP_CACHING_WAIT, search everything, and wait if our bg is caching
7628          * LOOP_ALLOC_CHUNK, force a chunk allocation and try again
7629          * LOOP_NO_EMPTY_SIZE, set empty_size and empty_cluster to 0 and try
7630          *                     again
7631          */
7632         if (ffe_ctl->loop < LOOP_NO_EMPTY_SIZE) {
7633                 ffe_ctl->index = 0;
7634                 if (ffe_ctl->loop == LOOP_CACHING_NOWAIT) {
7635                         /*
7636                          * We want to skip the LOOP_CACHING_WAIT step if we
7637                          * don't have any uncached bgs and we've already done a
7638                          * full search through.
7639                          */
7640                         if (ffe_ctl->orig_have_caching_bg || !full_search)
7641                                 ffe_ctl->loop = LOOP_CACHING_WAIT;
7642                         else
7643                                 ffe_ctl->loop = LOOP_ALLOC_CHUNK;
7644                 } else {
7645                         ffe_ctl->loop++;
7646                 }
7647
7648                 if (ffe_ctl->loop == LOOP_ALLOC_CHUNK) {
7649                         struct btrfs_trans_handle *trans;
7650                         int exist = 0;
7651
7652                         trans = current->journal_info;
7653                         if (trans)
7654                                 exist = 1;
7655                         else
7656                                 trans = btrfs_join_transaction(root);
7657
7658                         if (IS_ERR(trans)) {
7659                                 ret = PTR_ERR(trans);
7660                                 return ret;
7661                         }
7662
7663                         ret = do_chunk_alloc(trans, ffe_ctl->flags,
7664                                              CHUNK_ALLOC_FORCE);
7665
7666                         /*
7667                          * If we can't allocate a new chunk we've already looped
7668                          * through at least once, move on to the NO_EMPTY_SIZE
7669                          * case.
7670                          */
7671                         if (ret == -ENOSPC)
7672                                 ffe_ctl->loop = LOOP_NO_EMPTY_SIZE;
7673
7674                         /* Do not bail out on ENOSPC since we can do more. */
7675                         if (ret < 0 && ret != -ENOSPC)
7676                                 btrfs_abort_transaction(trans, ret);
7677                         else
7678                                 ret = 0;
7679                         if (!exist)
7680                                 btrfs_end_transaction(trans);
7681                         if (ret)
7682                                 return ret;
7683                 }
7684
7685                 if (ffe_ctl->loop == LOOP_NO_EMPTY_SIZE) {
7686                         /*
7687                          * Don't loop again if we already have no empty_size and
7688                          * no empty_cluster.
7689                          */
7690                         if (ffe_ctl->empty_size == 0 &&
7691                             ffe_ctl->empty_cluster == 0)
7692                                 return -ENOSPC;
7693                         ffe_ctl->empty_size = 0;
7694                         ffe_ctl->empty_cluster = 0;
7695                 }
7696                 return 1;
7697         }
7698         return -ENOSPC;
7699 }
7700
7701 /*
7702  * walks the btree of allocated extents and find a hole of a given size.
7703  * The key ins is changed to record the hole:
7704  * ins->objectid == start position
7705  * ins->flags = BTRFS_EXTENT_ITEM_KEY
7706  * ins->offset == the size of the hole.
7707  * Any available blocks before search_start are skipped.
7708  *
7709  * If there is no suitable free space, we will record the max size of
7710  * the free space extent currently.
7711  *
7712  * The overall logic and call chain:
7713  *
7714  * find_free_extent()
7715  * |- Iterate through all block groups
7716  * |  |- Get a valid block group
7717  * |  |- Try to do clustered allocation in that block group
7718  * |  |- Try to do unclustered allocation in that block group
7719  * |  |- Check if the result is valid
7720  * |  |  |- If valid, then exit
7721  * |  |- Jump to next block group
7722  * |
7723  * |- Push harder to find free extents
7724  *    |- If not found, re-iterate all block groups
7725  */
7726 static noinline int find_free_extent(struct btrfs_fs_info *fs_info,
7727                                 u64 ram_bytes, u64 num_bytes, u64 empty_size,
7728                                 u64 hint_byte, struct btrfs_key *ins,
7729                                 u64 flags, int delalloc)
7730 {
7731         int ret = 0;
7732         struct btrfs_free_cluster *last_ptr = NULL;
7733         struct btrfs_block_group_cache *block_group = NULL;
7734         struct find_free_extent_ctl ffe_ctl = {0};
7735         struct btrfs_space_info *space_info;
7736         bool use_cluster = true;
7737         bool full_search = false;
7738
7739         WARN_ON(num_bytes < fs_info->sectorsize);
7740
7741         ffe_ctl.ram_bytes = ram_bytes;
7742         ffe_ctl.num_bytes = num_bytes;
7743         ffe_ctl.empty_size = empty_size;
7744         ffe_ctl.flags = flags;
7745         ffe_ctl.search_start = 0;
7746         ffe_ctl.retry_clustered = false;
7747         ffe_ctl.retry_unclustered = false;
7748         ffe_ctl.delalloc = delalloc;
7749         ffe_ctl.index = btrfs_bg_flags_to_raid_index(flags);
7750         ffe_ctl.have_caching_bg = false;
7751         ffe_ctl.orig_have_caching_bg = false;
7752         ffe_ctl.found_offset = 0;
7753
7754         ins->type = BTRFS_EXTENT_ITEM_KEY;
7755         ins->objectid = 0;
7756         ins->offset = 0;
7757
7758         trace_find_free_extent(fs_info, num_bytes, empty_size, flags);
7759
7760         space_info = __find_space_info(fs_info, flags);
7761         if (!space_info) {
7762                 btrfs_err(fs_info, "No space info for %llu", flags);
7763                 return -ENOSPC;
7764         }
7765
7766         /*
7767          * If our free space is heavily fragmented we may not be able to make
7768          * big contiguous allocations, so instead of doing the expensive search
7769          * for free space, simply return ENOSPC with our max_extent_size so we
7770          * can go ahead and search for a more manageable chunk.
7771          *
7772          * If our max_extent_size is large enough for our allocation simply
7773          * disable clustering since we will likely not be able to find enough
7774          * space to create a cluster and induce latency trying.
7775          */
7776         if (unlikely(space_info->max_extent_size)) {
7777                 spin_lock(&space_info->lock);
7778                 if (space_info->max_extent_size &&
7779                     num_bytes > space_info->max_extent_size) {
7780                         ins->offset = space_info->max_extent_size;
7781                         spin_unlock(&space_info->lock);
7782                         return -ENOSPC;
7783                 } else if (space_info->max_extent_size) {
7784                         use_cluster = false;
7785                 }
7786                 spin_unlock(&space_info->lock);
7787         }
7788
7789         last_ptr = fetch_cluster_info(fs_info, space_info,
7790                                       &ffe_ctl.empty_cluster);
7791         if (last_ptr) {
7792                 spin_lock(&last_ptr->lock);
7793                 if (last_ptr->block_group)
7794                         hint_byte = last_ptr->window_start;
7795                 if (last_ptr->fragmented) {
7796                         /*
7797                          * We still set window_start so we can keep track of the
7798                          * last place we found an allocation to try and save
7799                          * some time.
7800                          */
7801                         hint_byte = last_ptr->window_start;
7802                         use_cluster = false;
7803                 }
7804                 spin_unlock(&last_ptr->lock);
7805         }
7806
7807         ffe_ctl.search_start = max(ffe_ctl.search_start,
7808                                    first_logical_byte(fs_info, 0));
7809         ffe_ctl.search_start = max(ffe_ctl.search_start, hint_byte);
7810         if (ffe_ctl.search_start == hint_byte) {
7811                 block_group = btrfs_lookup_block_group(fs_info,
7812                                                        ffe_ctl.search_start);
7813                 /*
7814                  * we don't want to use the block group if it doesn't match our
7815                  * allocation bits, or if its not cached.
7816                  *
7817                  * However if we are re-searching with an ideal block group
7818                  * picked out then we don't care that the block group is cached.
7819                  */
7820                 if (block_group && block_group_bits(block_group, flags) &&
7821                     block_group->cached != BTRFS_CACHE_NO) {
7822                         down_read(&space_info->groups_sem);
7823                         if (list_empty(&block_group->list) ||
7824                             block_group->ro) {
7825                                 /*
7826                                  * someone is removing this block group,
7827                                  * we can't jump into the have_block_group
7828                                  * target because our list pointers are not
7829                                  * valid
7830                                  */
7831                                 btrfs_put_block_group(block_group);
7832                                 up_read(&space_info->groups_sem);
7833                         } else {
7834                                 ffe_ctl.index = btrfs_bg_flags_to_raid_index(
7835                                                 block_group->flags);
7836                                 btrfs_lock_block_group(block_group, delalloc);
7837                                 goto have_block_group;
7838                         }
7839                 } else if (block_group) {
7840                         btrfs_put_block_group(block_group);
7841                 }
7842         }
7843 search:
7844         ffe_ctl.have_caching_bg = false;
7845         if (ffe_ctl.index == btrfs_bg_flags_to_raid_index(flags) ||
7846             ffe_ctl.index == 0)
7847                 full_search = true;
7848         down_read(&space_info->groups_sem);
7849         list_for_each_entry(block_group,
7850                             &space_info->block_groups[ffe_ctl.index], list) {
7851                 /* If the block group is read-only, we can skip it entirely. */
7852                 if (unlikely(block_group->ro))
7853                         continue;
7854
7855                 btrfs_grab_block_group(block_group, delalloc);
7856                 ffe_ctl.search_start = block_group->key.objectid;
7857
7858                 /*
7859                  * this can happen if we end up cycling through all the
7860                  * raid types, but we want to make sure we only allocate
7861                  * for the proper type.
7862                  */
7863                 if (!block_group_bits(block_group, flags)) {
7864                         u64 extra = BTRFS_BLOCK_GROUP_DUP |
7865                                 BTRFS_BLOCK_GROUP_RAID1 |
7866                                 BTRFS_BLOCK_GROUP_RAID5 |
7867                                 BTRFS_BLOCK_GROUP_RAID6 |
7868                                 BTRFS_BLOCK_GROUP_RAID10;
7869
7870                         /*
7871                          * if they asked for extra copies and this block group
7872                          * doesn't provide them, bail.  This does allow us to
7873                          * fill raid0 from raid1.
7874                          */
7875                         if ((flags & extra) && !(block_group->flags & extra))
7876                                 goto loop;
7877                 }
7878
7879 have_block_group:
7880                 ffe_ctl.cached = block_group_cache_done(block_group);
7881                 if (unlikely(!ffe_ctl.cached)) {
7882                         ffe_ctl.have_caching_bg = true;
7883                         ret = cache_block_group(block_group, 0);
7884                         BUG_ON(ret < 0);
7885                         ret = 0;
7886                 }
7887
7888                 if (unlikely(block_group->cached == BTRFS_CACHE_ERROR))
7889                         goto loop;
7890
7891                 /*
7892                  * Ok we want to try and use the cluster allocator, so
7893                  * lets look there
7894                  */
7895                 if (last_ptr && use_cluster) {
7896                         struct btrfs_block_group_cache *cluster_bg = NULL;
7897
7898                         ret = find_free_extent_clustered(block_group, last_ptr,
7899                                                          &ffe_ctl, &cluster_bg);
7900
7901                         if (ret == 0) {
7902                                 if (cluster_bg && cluster_bg != block_group) {
7903                                         btrfs_release_block_group(block_group,
7904                                                                   delalloc);
7905                                         block_group = cluster_bg;
7906                                 }
7907                                 goto checks;
7908                         } else if (ret == -EAGAIN) {
7909                                 goto have_block_group;
7910                         } else if (ret > 0) {
7911                                 goto loop;
7912                         }
7913                         /* ret == -ENOENT case falls through */
7914                 }
7915
7916                 ret = find_free_extent_unclustered(block_group, last_ptr,
7917                                                    &ffe_ctl);
7918                 if (ret == -EAGAIN)
7919                         goto have_block_group;
7920                 else if (ret > 0)
7921                         goto loop;
7922                 /* ret == 0 case falls through */
7923 checks:
7924                 ffe_ctl.search_start = round_up(ffe_ctl.found_offset,
7925                                              fs_info->stripesize);
7926
7927                 /* move on to the next group */
7928                 if (ffe_ctl.search_start + num_bytes >
7929                     block_group->key.objectid + block_group->key.offset) {
7930                         btrfs_add_free_space(block_group, ffe_ctl.found_offset,
7931                                              num_bytes);
7932                         goto loop;
7933                 }
7934
7935                 if (ffe_ctl.found_offset < ffe_ctl.search_start)
7936                         btrfs_add_free_space(block_group, ffe_ctl.found_offset,
7937                                 ffe_ctl.search_start - ffe_ctl.found_offset);
7938
7939                 ret = btrfs_add_reserved_bytes(block_group, ram_bytes,
7940                                 num_bytes, delalloc);
7941                 if (ret == -EAGAIN) {
7942                         btrfs_add_free_space(block_group, ffe_ctl.found_offset,
7943                                              num_bytes);
7944                         goto loop;
7945                 }
7946                 btrfs_inc_block_group_reservations(block_group);
7947
7948                 /* we are all good, lets return */
7949                 ins->objectid = ffe_ctl.search_start;
7950                 ins->offset = num_bytes;
7951
7952                 trace_btrfs_reserve_extent(block_group, ffe_ctl.search_start,
7953                                            num_bytes);
7954                 btrfs_release_block_group(block_group, delalloc);
7955                 break;
7956 loop:
7957                 ffe_ctl.retry_clustered = false;
7958                 ffe_ctl.retry_unclustered = false;
7959                 BUG_ON(btrfs_bg_flags_to_raid_index(block_group->flags) !=
7960                        ffe_ctl.index);
7961                 btrfs_release_block_group(block_group, delalloc);
7962                 cond_resched();
7963         }
7964         up_read(&space_info->groups_sem);
7965
7966         ret = find_free_extent_update_loop(fs_info, last_ptr, ins, &ffe_ctl,
7967                                            full_search, use_cluster);
7968         if (ret > 0)
7969                 goto search;
7970
7971         if (ret == -ENOSPC) {
7972                 /*
7973                  * Use ffe_ctl->total_free_space as fallback if we can't find
7974                  * any contiguous hole.
7975                  */
7976                 if (!ffe_ctl.max_extent_size)
7977                         ffe_ctl.max_extent_size = ffe_ctl.total_free_space;
7978                 spin_lock(&space_info->lock);
7979                 space_info->max_extent_size = ffe_ctl.max_extent_size;
7980                 spin_unlock(&space_info->lock);
7981                 ins->offset = ffe_ctl.max_extent_size;
7982         }
7983         return ret;
7984 }
7985
7986 #define DUMP_BLOCK_RSV(fs_info, rsv_name)                               \
7987 do {                                                                    \
7988         struct btrfs_block_rsv *__rsv = &(fs_info)->rsv_name;           \
7989         spin_lock(&__rsv->lock);                                        \
7990         btrfs_info(fs_info, #rsv_name ": size %llu reserved %llu",      \
7991                    __rsv->size, __rsv->reserved);                       \
7992         spin_unlock(&__rsv->lock);                                      \
7993 } while (0)
7994
7995 static void dump_space_info(struct btrfs_fs_info *fs_info,
7996                             struct btrfs_space_info *info, u64 bytes,
7997                             int dump_block_groups)
7998 {
7999         struct btrfs_block_group_cache *cache;
8000         int index = 0;
8001
8002         spin_lock(&info->lock);
8003         btrfs_info(fs_info, "space_info %llu has %llu free, is %sfull",
8004                    info->flags,
8005                    info->total_bytes - btrfs_space_info_used(info, true),
8006                    info->full ? "" : "not ");
8007         btrfs_info(fs_info,
8008                 "space_info total=%llu, used=%llu, pinned=%llu, reserved=%llu, may_use=%llu, readonly=%llu",
8009                 info->total_bytes, info->bytes_used, info->bytes_pinned,
8010                 info->bytes_reserved, info->bytes_may_use,
8011                 info->bytes_readonly);
8012         spin_unlock(&info->lock);
8013
8014         DUMP_BLOCK_RSV(fs_info, global_block_rsv);
8015         DUMP_BLOCK_RSV(fs_info, trans_block_rsv);
8016         DUMP_BLOCK_RSV(fs_info, chunk_block_rsv);
8017         DUMP_BLOCK_RSV(fs_info, delayed_block_rsv);
8018         DUMP_BLOCK_RSV(fs_info, delayed_refs_rsv);
8019
8020         if (!dump_block_groups)
8021                 return;
8022
8023         down_read(&info->groups_sem);
8024 again:
8025         list_for_each_entry(cache, &info->block_groups[index], list) {
8026                 spin_lock(&cache->lock);
8027                 btrfs_info(fs_info,
8028                         "block group %llu has %llu bytes, %llu used %llu pinned %llu reserved %s",
8029                         cache->key.objectid, cache->key.offset,
8030                         btrfs_block_group_used(&cache->item), cache->pinned,
8031                         cache->reserved, cache->ro ? "[readonly]" : "");
8032                 btrfs_dump_free_space(cache, bytes);
8033                 spin_unlock(&cache->lock);
8034         }
8035         if (++index < BTRFS_NR_RAID_TYPES)
8036                 goto again;
8037         up_read(&info->groups_sem);
8038 }
8039
8040 /*
8041  * btrfs_reserve_extent - entry point to the extent allocator. Tries to find a
8042  *                        hole that is at least as big as @num_bytes.
8043  *
8044  * @root           -    The root that will contain this extent
8045  *
8046  * @ram_bytes      -    The amount of space in ram that @num_bytes take. This
8047  *                      is used for accounting purposes. This value differs
8048  *                      from @num_bytes only in the case of compressed extents.
8049  *
8050  * @num_bytes      -    Number of bytes to allocate on-disk.
8051  *
8052  * @min_alloc_size -    Indicates the minimum amount of space that the
8053  *                      allocator should try to satisfy. In some cases
8054  *                      @num_bytes may be larger than what is required and if
8055  *                      the filesystem is fragmented then allocation fails.
8056  *                      However, the presence of @min_alloc_size gives a
8057  *                      chance to try and satisfy the smaller allocation.
8058  *
8059  * @empty_size     -    A hint that you plan on doing more COW. This is the
8060  *                      size in bytes the allocator should try to find free
8061  *                      next to the block it returns.  This is just a hint and
8062  *                      may be ignored by the allocator.
8063  *
8064  * @hint_byte      -    Hint to the allocator to start searching above the byte
8065  *                      address passed. It might be ignored.
8066  *
8067  * @ins            -    This key is modified to record the found hole. It will
8068  *                      have the following values:
8069  *                      ins->objectid == start position
8070  *                      ins->flags = BTRFS_EXTENT_ITEM_KEY
8071  *                      ins->offset == the size of the hole.
8072  *
8073  * @is_data        -    Boolean flag indicating whether an extent is
8074  *                      allocated for data (true) or metadata (false)
8075  *
8076  * @delalloc       -    Boolean flag indicating whether this allocation is for
8077  *                      delalloc or not. If 'true' data_rwsem of block groups
8078  *                      is going to be acquired.
8079  *
8080  *
8081  * Returns 0 when an allocation succeeded or < 0 when an error occurred. In
8082  * case -ENOSPC is returned then @ins->offset will contain the size of the
8083  * largest available hole the allocator managed to find.
8084  */
8085 int btrfs_reserve_extent(struct btrfs_root *root, u64 ram_bytes,
8086                          u64 num_bytes, u64 min_alloc_size,
8087                          u64 empty_size, u64 hint_byte,
8088                          struct btrfs_key *ins, int is_data, int delalloc)
8089 {
8090         struct btrfs_fs_info *fs_info = root->fs_info;
8091         bool final_tried = num_bytes == min_alloc_size;
8092         u64 flags;
8093         int ret;
8094
8095         flags = get_alloc_profile_by_root(root, is_data);
8096 again:
8097         WARN_ON(num_bytes < fs_info->sectorsize);
8098         ret = find_free_extent(fs_info, ram_bytes, num_bytes, empty_size,
8099                                hint_byte, ins, flags, delalloc);
8100         if (!ret && !is_data) {
8101                 btrfs_dec_block_group_reservations(fs_info, ins->objectid);
8102         } else if (ret == -ENOSPC) {
8103                 if (!final_tried && ins->offset) {
8104                         num_bytes = min(num_bytes >> 1, ins->offset);
8105                         num_bytes = round_down(num_bytes,
8106                                                fs_info->sectorsize);
8107                         num_bytes = max(num_bytes, min_alloc_size);
8108                         ram_bytes = num_bytes;
8109                         if (num_bytes == min_alloc_size)
8110                                 final_tried = true;
8111                         goto again;
8112                 } else if (btrfs_test_opt(fs_info, ENOSPC_DEBUG)) {
8113                         struct btrfs_space_info *sinfo;
8114
8115                         sinfo = __find_space_info(fs_info, flags);
8116                         btrfs_err(fs_info,
8117                                   "allocation failed flags %llu, wanted %llu",
8118                                   flags, num_bytes);
8119                         if (sinfo)
8120                                 dump_space_info(fs_info, sinfo, num_bytes, 1);
8121                 }
8122         }
8123
8124         return ret;
8125 }
8126
8127 static int __btrfs_free_reserved_extent(struct btrfs_fs_info *fs_info,
8128                                         u64 start, u64 len,
8129                                         int pin, int delalloc)
8130 {
8131         struct btrfs_block_group_cache *cache;
8132         int ret = 0;
8133
8134         cache = btrfs_lookup_block_group(fs_info, start);
8135         if (!cache) {
8136                 btrfs_err(fs_info, "Unable to find block group for %llu",
8137                           start);
8138                 return -ENOSPC;
8139         }
8140
8141         if (pin)
8142                 pin_down_extent(cache, start, len, 1);
8143         else {
8144                 if (btrfs_test_opt(fs_info, DISCARD))
8145                         ret = btrfs_discard_extent(fs_info, start, len, NULL);
8146                 btrfs_add_free_space(cache, start, len);
8147                 btrfs_free_reserved_bytes(cache, len, delalloc);
8148                 trace_btrfs_reserved_extent_free(fs_info, start, len);
8149         }
8150
8151         btrfs_put_block_group(cache);
8152         return ret;
8153 }
8154
8155 int btrfs_free_reserved_extent(struct btrfs_fs_info *fs_info,
8156                                u64 start, u64 len, int delalloc)
8157 {
8158         return __btrfs_free_reserved_extent(fs_info, start, len, 0, delalloc);
8159 }
8160
8161 int btrfs_free_and_pin_reserved_extent(struct btrfs_fs_info *fs_info,
8162                                        u64 start, u64 len)
8163 {
8164         return __btrfs_free_reserved_extent(fs_info, start, len, 1, 0);
8165 }
8166
8167 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
8168                                       u64 parent, u64 root_objectid,
8169                                       u64 flags, u64 owner, u64 offset,
8170                                       struct btrfs_key *ins, int ref_mod)
8171 {
8172         struct btrfs_fs_info *fs_info = trans->fs_info;
8173         int ret;
8174         struct btrfs_extent_item *extent_item;
8175         struct btrfs_extent_inline_ref *iref;
8176         struct btrfs_path *path;
8177         struct extent_buffer *leaf;
8178         int type;
8179         u32 size;
8180
8181         if (parent > 0)
8182                 type = BTRFS_SHARED_DATA_REF_KEY;
8183         else
8184                 type = BTRFS_EXTENT_DATA_REF_KEY;
8185
8186         size = sizeof(*extent_item) + btrfs_extent_inline_ref_size(type);
8187
8188         path = btrfs_alloc_path();
8189         if (!path)
8190                 return -ENOMEM;
8191
8192         path->leave_spinning = 1;
8193         ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path,
8194                                       ins, size);
8195         if (ret) {
8196                 btrfs_free_path(path);
8197                 return ret;
8198         }
8199
8200         leaf = path->nodes[0];
8201         extent_item = btrfs_item_ptr(leaf, path->slots[0],
8202                                      struct btrfs_extent_item);
8203         btrfs_set_extent_refs(leaf, extent_item, ref_mod);
8204         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
8205         btrfs_set_extent_flags(leaf, extent_item,
8206                                flags | BTRFS_EXTENT_FLAG_DATA);
8207
8208         iref = (struct btrfs_extent_inline_ref *)(extent_item + 1);
8209         btrfs_set_extent_inline_ref_type(leaf, iref, type);
8210         if (parent > 0) {
8211                 struct btrfs_shared_data_ref *ref;
8212                 ref = (struct btrfs_shared_data_ref *)(iref + 1);
8213                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
8214                 btrfs_set_shared_data_ref_count(leaf, ref, ref_mod);
8215         } else {
8216                 struct btrfs_extent_data_ref *ref;
8217                 ref = (struct btrfs_extent_data_ref *)(&iref->offset);
8218                 btrfs_set_extent_data_ref_root(leaf, ref, root_objectid);
8219                 btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
8220                 btrfs_set_extent_data_ref_offset(leaf, ref, offset);
8221                 btrfs_set_extent_data_ref_count(leaf, ref, ref_mod);
8222         }
8223
8224         btrfs_mark_buffer_dirty(path->nodes[0]);
8225         btrfs_free_path(path);
8226
8227         ret = remove_from_free_space_tree(trans, ins->objectid, ins->offset);
8228         if (ret)
8229                 return ret;
8230
8231         ret = update_block_group(trans, ins->objectid, ins->offset, 1);
8232         if (ret) { /* -ENOENT, logic error */
8233                 btrfs_err(fs_info, "update block group failed for %llu %llu",
8234                         ins->objectid, ins->offset);
8235                 BUG();
8236         }
8237         trace_btrfs_reserved_extent_alloc(fs_info, ins->objectid, ins->offset);
8238         return ret;
8239 }
8240
8241 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
8242                                      struct btrfs_delayed_ref_node *node,
8243                                      struct btrfs_delayed_extent_op *extent_op)
8244 {
8245         struct btrfs_fs_info *fs_info = trans->fs_info;
8246         int ret;
8247         struct btrfs_extent_item *extent_item;
8248         struct btrfs_key extent_key;
8249         struct btrfs_tree_block_info *block_info;
8250         struct btrfs_extent_inline_ref *iref;
8251         struct btrfs_path *path;
8252         struct extent_buffer *leaf;
8253         struct btrfs_delayed_tree_ref *ref;
8254         u32 size = sizeof(*extent_item) + sizeof(*iref);
8255         u64 num_bytes;
8256         u64 flags = extent_op->flags_to_set;
8257         bool skinny_metadata = btrfs_fs_incompat(fs_info, SKINNY_METADATA);
8258
8259         ref = btrfs_delayed_node_to_tree_ref(node);
8260
8261         extent_key.objectid = node->bytenr;
8262         if (skinny_metadata) {
8263                 extent_key.offset = ref->level;
8264                 extent_key.type = BTRFS_METADATA_ITEM_KEY;
8265                 num_bytes = fs_info->nodesize;
8266         } else {
8267                 extent_key.offset = node->num_bytes;
8268                 extent_key.type = BTRFS_EXTENT_ITEM_KEY;
8269                 size += sizeof(*block_info);
8270                 num_bytes = node->num_bytes;
8271         }
8272
8273         path = btrfs_alloc_path();
8274         if (!path)
8275                 return -ENOMEM;
8276
8277         path->leave_spinning = 1;
8278         ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path,
8279                                       &extent_key, size);
8280         if (ret) {
8281                 btrfs_free_path(path);
8282                 return ret;
8283         }
8284
8285         leaf = path->nodes[0];
8286         extent_item = btrfs_item_ptr(leaf, path->slots[0],
8287                                      struct btrfs_extent_item);
8288         btrfs_set_extent_refs(leaf, extent_item, 1);
8289         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
8290         btrfs_set_extent_flags(leaf, extent_item,
8291                                flags | BTRFS_EXTENT_FLAG_TREE_BLOCK);
8292
8293         if (skinny_metadata) {
8294                 iref = (struct btrfs_extent_inline_ref *)(extent_item + 1);
8295         } else {
8296                 block_info = (struct btrfs_tree_block_info *)(extent_item + 1);
8297                 btrfs_set_tree_block_key(leaf, block_info, &extent_op->key);
8298                 btrfs_set_tree_block_level(leaf, block_info, ref->level);
8299                 iref = (struct btrfs_extent_inline_ref *)(block_info + 1);
8300         }
8301
8302         if (node->type == BTRFS_SHARED_BLOCK_REF_KEY) {
8303                 BUG_ON(!(flags & BTRFS_BLOCK_FLAG_FULL_BACKREF));
8304                 btrfs_set_extent_inline_ref_type(leaf, iref,
8305                                                  BTRFS_SHARED_BLOCK_REF_KEY);
8306                 btrfs_set_extent_inline_ref_offset(leaf, iref, ref->parent);
8307         } else {
8308                 btrfs_set_extent_inline_ref_type(leaf, iref,
8309                                                  BTRFS_TREE_BLOCK_REF_KEY);
8310                 btrfs_set_extent_inline_ref_offset(leaf, iref, ref->root);
8311         }
8312
8313         btrfs_mark_buffer_dirty(leaf);
8314         btrfs_free_path(path);
8315
8316         ret = remove_from_free_space_tree(trans, extent_key.objectid,
8317                                           num_bytes);
8318         if (ret)
8319                 return ret;
8320
8321         ret = update_block_group(trans, extent_key.objectid,
8322                                  fs_info->nodesize, 1);
8323         if (ret) { /* -ENOENT, logic error */
8324                 btrfs_err(fs_info, "update block group failed for %llu %llu",
8325                         extent_key.objectid, extent_key.offset);
8326                 BUG();
8327         }
8328
8329         trace_btrfs_reserved_extent_alloc(fs_info, extent_key.objectid,
8330                                           fs_info->nodesize);
8331         return ret;
8332 }
8333
8334 int btrfs_alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
8335                                      struct btrfs_root *root, u64 owner,
8336                                      u64 offset, u64 ram_bytes,
8337                                      struct btrfs_key *ins)
8338 {
8339         struct btrfs_ref generic_ref = { 0 };
8340         int ret;
8341
8342         BUG_ON(root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID);
8343
8344         btrfs_init_generic_ref(&generic_ref, BTRFS_ADD_DELAYED_EXTENT,
8345                                ins->objectid, ins->offset, 0);
8346         btrfs_init_data_ref(&generic_ref, root->root_key.objectid, owner, offset);
8347         btrfs_ref_tree_mod(root->fs_info, &generic_ref);
8348         ret = btrfs_add_delayed_data_ref(trans, &generic_ref,
8349                                          ram_bytes, NULL, NULL);
8350         return ret;
8351 }
8352
8353 /*
8354  * this is used by the tree logging recovery code.  It records that
8355  * an extent has been allocated and makes sure to clear the free
8356  * space cache bits as well
8357  */
8358 int btrfs_alloc_logged_file_extent(struct btrfs_trans_handle *trans,
8359                                    u64 root_objectid, u64 owner, u64 offset,
8360                                    struct btrfs_key *ins)
8361 {
8362         struct btrfs_fs_info *fs_info = trans->fs_info;
8363         int ret;
8364         struct btrfs_block_group_cache *block_group;
8365         struct btrfs_space_info *space_info;
8366
8367         /*
8368          * Mixed block groups will exclude before processing the log so we only
8369          * need to do the exclude dance if this fs isn't mixed.
8370          */
8371         if (!btrfs_fs_incompat(fs_info, MIXED_GROUPS)) {
8372                 ret = __exclude_logged_extent(fs_info, ins->objectid,
8373                                               ins->offset);
8374                 if (ret)
8375                         return ret;
8376         }
8377
8378         block_group = btrfs_lookup_block_group(fs_info, ins->objectid);
8379         if (!block_group)
8380                 return -EINVAL;
8381
8382         space_info = block_group->space_info;
8383         spin_lock(&space_info->lock);
8384         spin_lock(&block_group->lock);
8385         space_info->bytes_reserved += ins->offset;
8386         block_group->reserved += ins->offset;
8387         spin_unlock(&block_group->lock);
8388         spin_unlock(&space_info->lock);
8389
8390         ret = alloc_reserved_file_extent(trans, 0, root_objectid, 0, owner,
8391                                          offset, ins, 1);
8392         btrfs_put_block_group(block_group);
8393         return ret;
8394 }
8395
8396 static struct extent_buffer *
8397 btrfs_init_new_buffer(struct btrfs_trans_handle *trans, struct btrfs_root *root,
8398                       u64 bytenr, int level, u64 owner)
8399 {
8400         struct btrfs_fs_info *fs_info = root->fs_info;
8401         struct extent_buffer *buf;
8402
8403         buf = btrfs_find_create_tree_block(fs_info, bytenr);
8404         if (IS_ERR(buf))
8405                 return buf;
8406
8407         /*
8408          * Extra safety check in case the extent tree is corrupted and extent
8409          * allocator chooses to use a tree block which is already used and
8410          * locked.
8411          */
8412         if (buf->lock_owner == current->pid) {
8413                 btrfs_err_rl(fs_info,
8414 "tree block %llu owner %llu already locked by pid=%d, extent tree corruption detected",
8415                         buf->start, btrfs_header_owner(buf), current->pid);
8416                 free_extent_buffer(buf);
8417                 return ERR_PTR(-EUCLEAN);
8418         }
8419
8420         btrfs_set_buffer_lockdep_class(root->root_key.objectid, buf, level);
8421         btrfs_tree_lock(buf);
8422         btrfs_clean_tree_block(buf);
8423         clear_bit(EXTENT_BUFFER_STALE, &buf->bflags);
8424
8425         btrfs_set_lock_blocking_write(buf);
8426         set_extent_buffer_uptodate(buf);
8427
8428         memzero_extent_buffer(buf, 0, sizeof(struct btrfs_header));
8429         btrfs_set_header_level(buf, level);
8430         btrfs_set_header_bytenr(buf, buf->start);
8431         btrfs_set_header_generation(buf, trans->transid);
8432         btrfs_set_header_backref_rev(buf, BTRFS_MIXED_BACKREF_REV);
8433         btrfs_set_header_owner(buf, owner);
8434         write_extent_buffer_fsid(buf, fs_info->fs_devices->metadata_uuid);
8435         write_extent_buffer_chunk_tree_uuid(buf, fs_info->chunk_tree_uuid);
8436         if (root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID) {
8437                 buf->log_index = root->log_transid % 2;
8438                 /*
8439                  * we allow two log transactions at a time, use different
8440                  * EXTENT bit to differentiate dirty pages.
8441                  */
8442                 if (buf->log_index == 0)
8443                         set_extent_dirty(&root->dirty_log_pages, buf->start,
8444                                         buf->start + buf->len - 1, GFP_NOFS);
8445                 else
8446                         set_extent_new(&root->dirty_log_pages, buf->start,
8447                                         buf->start + buf->len - 1);
8448         } else {
8449                 buf->log_index = -1;
8450                 set_extent_dirty(&trans->transaction->dirty_pages, buf->start,
8451                          buf->start + buf->len - 1, GFP_NOFS);
8452         }
8453         trans->dirty = true;
8454         /* this returns a buffer locked for blocking */
8455         return buf;
8456 }
8457
8458 static struct btrfs_block_rsv *
8459 use_block_rsv(struct btrfs_trans_handle *trans,
8460               struct btrfs_root *root, u32 blocksize)
8461 {
8462         struct btrfs_fs_info *fs_info = root->fs_info;
8463         struct btrfs_block_rsv *block_rsv;
8464         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
8465         int ret;
8466         bool global_updated = false;
8467
8468         block_rsv = get_block_rsv(trans, root);
8469
8470         if (unlikely(block_rsv->size == 0))
8471                 goto try_reserve;
8472 again:
8473         ret = block_rsv_use_bytes(block_rsv, blocksize);
8474         if (!ret)
8475                 return block_rsv;
8476
8477         if (block_rsv->failfast)
8478                 return ERR_PTR(ret);
8479
8480         if (block_rsv->type == BTRFS_BLOCK_RSV_GLOBAL && !global_updated) {
8481                 global_updated = true;
8482                 update_global_block_rsv(fs_info);
8483                 goto again;
8484         }
8485
8486         /*
8487          * The global reserve still exists to save us from ourselves, so don't
8488          * warn_on if we are short on our delayed refs reserve.
8489          */
8490         if (block_rsv->type != BTRFS_BLOCK_RSV_DELREFS &&
8491             btrfs_test_opt(fs_info, ENOSPC_DEBUG)) {
8492                 static DEFINE_RATELIMIT_STATE(_rs,
8493                                 DEFAULT_RATELIMIT_INTERVAL * 10,
8494                                 /*DEFAULT_RATELIMIT_BURST*/ 1);
8495                 if (__ratelimit(&_rs))
8496                         WARN(1, KERN_DEBUG
8497                                 "BTRFS: block rsv returned %d\n", ret);
8498         }
8499 try_reserve:
8500         ret = reserve_metadata_bytes(root, block_rsv, blocksize,
8501                                      BTRFS_RESERVE_NO_FLUSH);
8502         if (!ret)
8503                 return block_rsv;
8504         /*
8505          * If we couldn't reserve metadata bytes try and use some from
8506          * the global reserve if its space type is the same as the global
8507          * reservation.
8508          */
8509         if (block_rsv->type != BTRFS_BLOCK_RSV_GLOBAL &&
8510             block_rsv->space_info == global_rsv->space_info) {
8511                 ret = block_rsv_use_bytes(global_rsv, blocksize);
8512                 if (!ret)
8513                         return global_rsv;
8514         }
8515         return ERR_PTR(ret);
8516 }
8517
8518 static void unuse_block_rsv(struct btrfs_fs_info *fs_info,
8519                             struct btrfs_block_rsv *block_rsv, u32 blocksize)
8520 {
8521         block_rsv_add_bytes(block_rsv, blocksize, false);
8522         block_rsv_release_bytes(fs_info, block_rsv, NULL, 0, NULL);
8523 }
8524
8525 /*
8526  * finds a free extent and does all the dirty work required for allocation
8527  * returns the tree buffer or an ERR_PTR on error.
8528  */
8529 struct extent_buffer *btrfs_alloc_tree_block(struct btrfs_trans_handle *trans,
8530                                              struct btrfs_root *root,
8531                                              u64 parent, u64 root_objectid,
8532                                              const struct btrfs_disk_key *key,
8533                                              int level, u64 hint,
8534                                              u64 empty_size)
8535 {
8536         struct btrfs_fs_info *fs_info = root->fs_info;
8537         struct btrfs_key ins;
8538         struct btrfs_block_rsv *block_rsv;
8539         struct extent_buffer *buf;
8540         struct btrfs_delayed_extent_op *extent_op;
8541         struct btrfs_ref generic_ref = { 0 };
8542         u64 flags = 0;
8543         int ret;
8544         u32 blocksize = fs_info->nodesize;
8545         bool skinny_metadata = btrfs_fs_incompat(fs_info, SKINNY_METADATA);
8546
8547 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
8548         if (btrfs_is_testing(fs_info)) {
8549                 buf = btrfs_init_new_buffer(trans, root, root->alloc_bytenr,
8550                                             level, root_objectid);
8551                 if (!IS_ERR(buf))
8552                         root->alloc_bytenr += blocksize;
8553                 return buf;
8554         }
8555 #endif
8556
8557         block_rsv = use_block_rsv(trans, root, blocksize);
8558         if (IS_ERR(block_rsv))
8559                 return ERR_CAST(block_rsv);
8560
8561         ret = btrfs_reserve_extent(root, blocksize, blocksize, blocksize,
8562                                    empty_size, hint, &ins, 0, 0);
8563         if (ret)
8564                 goto out_unuse;
8565
8566         buf = btrfs_init_new_buffer(trans, root, ins.objectid, level,
8567                                     root_objectid);
8568         if (IS_ERR(buf)) {
8569                 ret = PTR_ERR(buf);
8570                 goto out_free_reserved;
8571         }
8572
8573         if (root_objectid == BTRFS_TREE_RELOC_OBJECTID) {
8574                 if (parent == 0)
8575                         parent = ins.objectid;
8576                 flags |= BTRFS_BLOCK_FLAG_FULL_BACKREF;
8577         } else
8578                 BUG_ON(parent > 0);
8579
8580         if (root_objectid != BTRFS_TREE_LOG_OBJECTID) {
8581                 extent_op = btrfs_alloc_delayed_extent_op();
8582                 if (!extent_op) {
8583                         ret = -ENOMEM;
8584                         goto out_free_buf;
8585                 }
8586                 if (key)
8587                         memcpy(&extent_op->key, key, sizeof(extent_op->key));
8588                 else
8589                         memset(&extent_op->key, 0, sizeof(extent_op->key));
8590                 extent_op->flags_to_set = flags;
8591                 extent_op->update_key = skinny_metadata ? false : true;
8592                 extent_op->update_flags = true;
8593                 extent_op->is_data = false;
8594                 extent_op->level = level;
8595
8596                 btrfs_init_generic_ref(&generic_ref, BTRFS_ADD_DELAYED_EXTENT,
8597                                        ins.objectid, ins.offset, parent);
8598                 generic_ref.real_root = root->root_key.objectid;
8599                 btrfs_init_tree_ref(&generic_ref, level, root_objectid);
8600                 btrfs_ref_tree_mod(fs_info, &generic_ref);
8601                 ret = btrfs_add_delayed_tree_ref(trans, &generic_ref,
8602                                                  extent_op, NULL, NULL);
8603                 if (ret)
8604                         goto out_free_delayed;
8605         }
8606         return buf;
8607
8608 out_free_delayed:
8609         btrfs_free_delayed_extent_op(extent_op);
8610 out_free_buf:
8611         free_extent_buffer(buf);
8612 out_free_reserved:
8613         btrfs_free_reserved_extent(fs_info, ins.objectid, ins.offset, 0);
8614 out_unuse:
8615         unuse_block_rsv(fs_info, block_rsv, blocksize);
8616         return ERR_PTR(ret);
8617 }
8618
8619 struct walk_control {
8620         u64 refs[BTRFS_MAX_LEVEL];
8621         u64 flags[BTRFS_MAX_LEVEL];
8622         struct btrfs_key update_progress;
8623         struct btrfs_key drop_progress;
8624         int drop_level;
8625         int stage;
8626         int level;
8627         int shared_level;
8628         int update_ref;
8629         int keep_locks;
8630         int reada_slot;
8631         int reada_count;
8632         int restarted;
8633 };
8634
8635 #define DROP_REFERENCE  1
8636 #define UPDATE_BACKREF  2
8637
8638 static noinline void reada_walk_down(struct btrfs_trans_handle *trans,
8639                                      struct btrfs_root *root,
8640                                      struct walk_control *wc,
8641                                      struct btrfs_path *path)
8642 {
8643         struct btrfs_fs_info *fs_info = root->fs_info;
8644         u64 bytenr;
8645         u64 generation;
8646         u64 refs;
8647         u64 flags;
8648         u32 nritems;
8649         struct btrfs_key key;
8650         struct extent_buffer *eb;
8651         int ret;
8652         int slot;
8653         int nread = 0;
8654
8655         if (path->slots[wc->level] < wc->reada_slot) {
8656                 wc->reada_count = wc->reada_count * 2 / 3;
8657                 wc->reada_count = max(wc->reada_count, 2);
8658         } else {
8659                 wc->reada_count = wc->reada_count * 3 / 2;
8660                 wc->reada_count = min_t(int, wc->reada_count,
8661                                         BTRFS_NODEPTRS_PER_BLOCK(fs_info));
8662         }
8663
8664         eb = path->nodes[wc->level];
8665         nritems = btrfs_header_nritems(eb);
8666
8667         for (slot = path->slots[wc->level]; slot < nritems; slot++) {
8668                 if (nread >= wc->reada_count)
8669                         break;
8670
8671                 cond_resched();
8672                 bytenr = btrfs_node_blockptr(eb, slot);
8673                 generation = btrfs_node_ptr_generation(eb, slot);
8674
8675                 if (slot == path->slots[wc->level])
8676                         goto reada;
8677
8678                 if (wc->stage == UPDATE_BACKREF &&
8679                     generation <= root->root_key.offset)
8680                         continue;
8681
8682                 /* We don't lock the tree block, it's OK to be racy here */
8683                 ret = btrfs_lookup_extent_info(trans, fs_info, bytenr,
8684                                                wc->level - 1, 1, &refs,
8685                                                &flags);
8686                 /* We don't care about errors in readahead. */
8687                 if (ret < 0)
8688                         continue;
8689                 BUG_ON(refs == 0);
8690
8691                 if (wc->stage == DROP_REFERENCE) {
8692                         if (refs == 1)
8693                                 goto reada;
8694
8695                         if (wc->level == 1 &&
8696                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
8697                                 continue;
8698                         if (!wc->update_ref ||
8699                             generation <= root->root_key.offset)
8700                                 continue;
8701                         btrfs_node_key_to_cpu(eb, &key, slot);
8702                         ret = btrfs_comp_cpu_keys(&key,
8703                                                   &wc->update_progress);
8704                         if (ret < 0)
8705                                 continue;
8706                 } else {
8707                         if (wc->level == 1 &&
8708                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
8709                                 continue;
8710                 }
8711 reada:
8712                 readahead_tree_block(fs_info, bytenr);
8713                 nread++;
8714         }
8715         wc->reada_slot = slot;
8716 }
8717
8718 /*
8719  * helper to process tree block while walking down the tree.
8720  *
8721  * when wc->stage == UPDATE_BACKREF, this function updates
8722  * back refs for pointers in the block.
8723  *
8724  * NOTE: return value 1 means we should stop walking down.
8725  */
8726 static noinline int walk_down_proc(struct btrfs_trans_handle *trans,
8727                                    struct btrfs_root *root,
8728                                    struct btrfs_path *path,
8729                                    struct walk_control *wc, int lookup_info)
8730 {
8731         struct btrfs_fs_info *fs_info = root->fs_info;
8732         int level = wc->level;
8733         struct extent_buffer *eb = path->nodes[level];
8734         u64 flag = BTRFS_BLOCK_FLAG_FULL_BACKREF;
8735         int ret;
8736
8737         if (wc->stage == UPDATE_BACKREF &&
8738             btrfs_header_owner(eb) != root->root_key.objectid)
8739                 return 1;
8740
8741         /*
8742          * when reference count of tree block is 1, it won't increase
8743          * again. once full backref flag is set, we never clear it.
8744          */
8745         if (lookup_info &&
8746             ((wc->stage == DROP_REFERENCE && wc->refs[level] != 1) ||
8747              (wc->stage == UPDATE_BACKREF && !(wc->flags[level] & flag)))) {
8748                 BUG_ON(!path->locks[level]);
8749                 ret = btrfs_lookup_extent_info(trans, fs_info,
8750                                                eb->start, level, 1,
8751                                                &wc->refs[level],
8752                                                &wc->flags[level]);
8753                 BUG_ON(ret == -ENOMEM);
8754                 if (ret)
8755                         return ret;
8756                 BUG_ON(wc->refs[level] == 0);
8757         }
8758
8759         if (wc->stage == DROP_REFERENCE) {
8760                 if (wc->refs[level] > 1)
8761                         return 1;
8762
8763                 if (path->locks[level] && !wc->keep_locks) {
8764                         btrfs_tree_unlock_rw(eb, path->locks[level]);
8765                         path->locks[level] = 0;
8766                 }
8767                 return 0;
8768         }
8769
8770         /* wc->stage == UPDATE_BACKREF */
8771         if (!(wc->flags[level] & flag)) {
8772                 BUG_ON(!path->locks[level]);
8773                 ret = btrfs_inc_ref(trans, root, eb, 1);
8774                 BUG_ON(ret); /* -ENOMEM */
8775                 ret = btrfs_dec_ref(trans, root, eb, 0);
8776                 BUG_ON(ret); /* -ENOMEM */
8777                 ret = btrfs_set_disk_extent_flags(trans, eb->start,
8778                                                   eb->len, flag,
8779                                                   btrfs_header_level(eb), 0);
8780                 BUG_ON(ret); /* -ENOMEM */
8781                 wc->flags[level] |= flag;
8782         }
8783
8784         /*
8785          * the block is shared by multiple trees, so it's not good to
8786          * keep the tree lock
8787          */
8788         if (path->locks[level] && level > 0) {
8789                 btrfs_tree_unlock_rw(eb, path->locks[level]);
8790                 path->locks[level] = 0;
8791         }
8792         return 0;
8793 }
8794
8795 /*
8796  * This is used to verify a ref exists for this root to deal with a bug where we
8797  * would have a drop_progress key that hadn't been updated properly.
8798  */
8799 static int check_ref_exists(struct btrfs_trans_handle *trans,
8800                             struct btrfs_root *root, u64 bytenr, u64 parent,
8801                             int level)
8802 {
8803         struct btrfs_path *path;
8804         struct btrfs_extent_inline_ref *iref;
8805         int ret;
8806
8807         path = btrfs_alloc_path();
8808         if (!path)
8809                 return -ENOMEM;
8810
8811         ret = lookup_extent_backref(trans, path, &iref, bytenr,
8812                                     root->fs_info->nodesize, parent,
8813                                     root->root_key.objectid, level, 0);
8814         btrfs_free_path(path);
8815         if (ret == -ENOENT)
8816                 return 0;
8817         if (ret < 0)
8818                 return ret;
8819         return 1;
8820 }
8821
8822 /*
8823  * helper to process tree block pointer.
8824  *
8825  * when wc->stage == DROP_REFERENCE, this function checks
8826  * reference count of the block pointed to. if the block
8827  * is shared and we need update back refs for the subtree
8828  * rooted at the block, this function changes wc->stage to
8829  * UPDATE_BACKREF. if the block is shared and there is no
8830  * need to update back, this function drops the reference
8831  * to the block.
8832  *
8833  * NOTE: return value 1 means we should stop walking down.
8834  */
8835 static noinline int do_walk_down(struct btrfs_trans_handle *trans,
8836                                  struct btrfs_root *root,
8837                                  struct btrfs_path *path,
8838                                  struct walk_control *wc, int *lookup_info)
8839 {
8840         struct btrfs_fs_info *fs_info = root->fs_info;
8841         u64 bytenr;
8842         u64 generation;
8843         u64 parent;
8844         struct btrfs_key key;
8845         struct btrfs_key first_key;
8846         struct btrfs_ref ref = { 0 };
8847         struct extent_buffer *next;
8848         int level = wc->level;
8849         int reada = 0;
8850         int ret = 0;
8851         bool need_account = false;
8852
8853         generation = btrfs_node_ptr_generation(path->nodes[level],
8854                                                path->slots[level]);
8855         /*
8856          * if the lower level block was created before the snapshot
8857          * was created, we know there is no need to update back refs
8858          * for the subtree
8859          */
8860         if (wc->stage == UPDATE_BACKREF &&
8861             generation <= root->root_key.offset) {
8862                 *lookup_info = 1;
8863                 return 1;
8864         }
8865
8866         bytenr = btrfs_node_blockptr(path->nodes[level], path->slots[level]);
8867         btrfs_node_key_to_cpu(path->nodes[level], &first_key,
8868                               path->slots[level]);
8869
8870         next = find_extent_buffer(fs_info, bytenr);
8871         if (!next) {
8872                 next = btrfs_find_create_tree_block(fs_info, bytenr);
8873                 if (IS_ERR(next))
8874                         return PTR_ERR(next);
8875
8876                 btrfs_set_buffer_lockdep_class(root->root_key.objectid, next,
8877                                                level - 1);
8878                 reada = 1;
8879         }
8880         btrfs_tree_lock(next);
8881         btrfs_set_lock_blocking_write(next);
8882
8883         ret = btrfs_lookup_extent_info(trans, fs_info, bytenr, level - 1, 1,
8884                                        &wc->refs[level - 1],
8885                                        &wc->flags[level - 1]);
8886         if (ret < 0)
8887                 goto out_unlock;
8888
8889         if (unlikely(wc->refs[level - 1] == 0)) {
8890                 btrfs_err(fs_info, "Missing references.");
8891                 ret = -EIO;
8892                 goto out_unlock;
8893         }
8894         *lookup_info = 0;
8895
8896         if (wc->stage == DROP_REFERENCE) {
8897                 if (wc->refs[level - 1] > 1) {
8898                         need_account = true;
8899                         if (level == 1 &&
8900                             (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
8901                                 goto skip;
8902
8903                         if (!wc->update_ref ||
8904                             generation <= root->root_key.offset)
8905                                 goto skip;
8906
8907                         btrfs_node_key_to_cpu(path->nodes[level], &key,
8908                                               path->slots[level]);
8909                         ret = btrfs_comp_cpu_keys(&key, &wc->update_progress);
8910                         if (ret < 0)
8911                                 goto skip;
8912
8913                         wc->stage = UPDATE_BACKREF;
8914                         wc->shared_level = level - 1;
8915                 }
8916         } else {
8917                 if (level == 1 &&
8918                     (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
8919                         goto skip;
8920         }
8921
8922         if (!btrfs_buffer_uptodate(next, generation, 0)) {
8923                 btrfs_tree_unlock(next);
8924                 free_extent_buffer(next);
8925                 next = NULL;
8926                 *lookup_info = 1;
8927         }
8928
8929         if (!next) {
8930                 if (reada && level == 1)
8931                         reada_walk_down(trans, root, wc, path);
8932                 next = read_tree_block(fs_info, bytenr, generation, level - 1,
8933                                        &first_key);
8934                 if (IS_ERR(next)) {
8935                         return PTR_ERR(next);
8936                 } else if (!extent_buffer_uptodate(next)) {
8937                         free_extent_buffer(next);
8938                         return -EIO;
8939                 }
8940                 btrfs_tree_lock(next);
8941                 btrfs_set_lock_blocking_write(next);
8942         }
8943
8944         level--;
8945         ASSERT(level == btrfs_header_level(next));
8946         if (level != btrfs_header_level(next)) {
8947                 btrfs_err(root->fs_info, "mismatched level");
8948                 ret = -EIO;
8949                 goto out_unlock;
8950         }
8951         path->nodes[level] = next;
8952         path->slots[level] = 0;
8953         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
8954         wc->level = level;
8955         if (wc->level == 1)
8956                 wc->reada_slot = 0;
8957         return 0;
8958 skip:
8959         wc->refs[level - 1] = 0;
8960         wc->flags[level - 1] = 0;
8961         if (wc->stage == DROP_REFERENCE) {
8962                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF) {
8963                         parent = path->nodes[level]->start;
8964                 } else {
8965                         ASSERT(root->root_key.objectid ==
8966                                btrfs_header_owner(path->nodes[level]));
8967                         if (root->root_key.objectid !=
8968                             btrfs_header_owner(path->nodes[level])) {
8969                                 btrfs_err(root->fs_info,
8970                                                 "mismatched block owner");
8971                                 ret = -EIO;
8972                                 goto out_unlock;
8973                         }
8974                         parent = 0;
8975                 }
8976
8977                 /*
8978                  * If we had a drop_progress we need to verify the refs are set
8979                  * as expected.  If we find our ref then we know that from here
8980                  * on out everything should be correct, and we can clear the
8981                  * ->restarted flag.
8982                  */
8983                 if (wc->restarted) {
8984                         ret = check_ref_exists(trans, root, bytenr, parent,
8985                                                level - 1);
8986                         if (ret < 0)
8987                                 goto out_unlock;
8988                         if (ret == 0)
8989                                 goto no_delete;
8990                         ret = 0;
8991                         wc->restarted = 0;
8992                 }
8993
8994                 /*
8995                  * Reloc tree doesn't contribute to qgroup numbers, and we have
8996                  * already accounted them at merge time (replace_path),
8997                  * thus we could skip expensive subtree trace here.
8998                  */
8999                 if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID &&
9000                     need_account) {
9001                         ret = btrfs_qgroup_trace_subtree(trans, next,
9002                                                          generation, level - 1);
9003                         if (ret) {
9004                                 btrfs_err_rl(fs_info,
9005                                              "Error %d accounting shared subtree. Quota is out of sync, rescan required.",
9006                                              ret);
9007                         }
9008                 }
9009
9010                 /*
9011                  * We need to update the next key in our walk control so we can
9012                  * update the drop_progress key accordingly.  We don't care if
9013                  * find_next_key doesn't find a key because that means we're at
9014                  * the end and are going to clean up now.
9015                  */
9016                 wc->drop_level = level;
9017                 find_next_key(path, level, &wc->drop_progress);
9018
9019                 btrfs_init_generic_ref(&ref, BTRFS_DROP_DELAYED_REF, bytenr,
9020                                        fs_info->nodesize, parent);
9021                 btrfs_init_tree_ref(&ref, level - 1, root->root_key.objectid);
9022                 ret = btrfs_free_extent(trans, &ref);
9023                 if (ret)
9024                         goto out_unlock;
9025         }
9026 no_delete:
9027         *lookup_info = 1;
9028         ret = 1;
9029
9030 out_unlock:
9031         btrfs_tree_unlock(next);
9032         free_extent_buffer(next);
9033
9034         return ret;
9035 }
9036
9037 /*
9038  * helper to process tree block while walking up the tree.
9039  *
9040  * when wc->stage == DROP_REFERENCE, this function drops
9041  * reference count on the block.
9042  *
9043  * when wc->stage == UPDATE_BACKREF, this function changes
9044  * wc->stage back to DROP_REFERENCE if we changed wc->stage
9045  * to UPDATE_BACKREF previously while processing the block.
9046  *
9047  * NOTE: return value 1 means we should stop walking up.
9048  */
9049 static noinline int walk_up_proc(struct btrfs_trans_handle *trans,
9050                                  struct btrfs_root *root,
9051                                  struct btrfs_path *path,
9052                                  struct walk_control *wc)
9053 {
9054         struct btrfs_fs_info *fs_info = root->fs_info;
9055         int ret;
9056         int level = wc->level;
9057         struct extent_buffer *eb = path->nodes[level];
9058         u64 parent = 0;
9059
9060         if (wc->stage == UPDATE_BACKREF) {
9061                 BUG_ON(wc->shared_level < level);
9062                 if (level < wc->shared_level)
9063                         goto out;
9064
9065                 ret = find_next_key(path, level + 1, &wc->update_progress);
9066                 if (ret > 0)
9067                         wc->update_ref = 0;
9068
9069                 wc->stage = DROP_REFERENCE;
9070                 wc->shared_level = -1;
9071                 path->slots[level] = 0;
9072
9073                 /*
9074                  * check reference count again if the block isn't locked.
9075                  * we should start walking down the tree again if reference
9076                  * count is one.
9077                  */
9078                 if (!path->locks[level]) {
9079                         BUG_ON(level == 0);
9080                         btrfs_tree_lock(eb);
9081                         btrfs_set_lock_blocking_write(eb);
9082                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
9083
9084                         ret = btrfs_lookup_extent_info(trans, fs_info,
9085                                                        eb->start, level, 1,
9086                                                        &wc->refs[level],
9087                                                        &wc->flags[level]);
9088                         if (ret < 0) {
9089                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
9090                                 path->locks[level] = 0;
9091                                 return ret;
9092                         }
9093                         BUG_ON(wc->refs[level] == 0);
9094                         if (wc->refs[level] == 1) {
9095                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
9096                                 path->locks[level] = 0;
9097                                 return 1;
9098                         }
9099                 }
9100         }
9101
9102         /* wc->stage == DROP_REFERENCE */
9103         BUG_ON(wc->refs[level] > 1 && !path->locks[level]);
9104
9105         if (wc->refs[level] == 1) {
9106                 if (level == 0) {
9107                         if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
9108                                 ret = btrfs_dec_ref(trans, root, eb, 1);
9109                         else
9110                                 ret = btrfs_dec_ref(trans, root, eb, 0);
9111                         BUG_ON(ret); /* -ENOMEM */
9112                         if (is_fstree(root->root_key.objectid)) {
9113                                 ret = btrfs_qgroup_trace_leaf_items(trans, eb);
9114                                 if (ret) {
9115                                         btrfs_err_rl(fs_info,
9116         "error %d accounting leaf items, quota is out of sync, rescan required",
9117                                              ret);
9118                                 }
9119                         }
9120                 }
9121                 /* make block locked assertion in btrfs_clean_tree_block happy */
9122                 if (!path->locks[level] &&
9123                     btrfs_header_generation(eb) == trans->transid) {
9124                         btrfs_tree_lock(eb);
9125                         btrfs_set_lock_blocking_write(eb);
9126                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
9127                 }
9128                 btrfs_clean_tree_block(eb);
9129         }
9130
9131         if (eb == root->node) {
9132                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
9133                         parent = eb->start;
9134                 else if (root->root_key.objectid != btrfs_header_owner(eb))
9135                         goto owner_mismatch;
9136         } else {
9137                 if (wc->flags[level + 1] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
9138                         parent = path->nodes[level + 1]->start;
9139                 else if (root->root_key.objectid !=
9140                          btrfs_header_owner(path->nodes[level + 1]))
9141                         goto owner_mismatch;
9142         }
9143
9144         btrfs_free_tree_block(trans, root, eb, parent, wc->refs[level] == 1);
9145 out:
9146         wc->refs[level] = 0;
9147         wc->flags[level] = 0;
9148         return 0;
9149
9150 owner_mismatch:
9151         btrfs_err_rl(fs_info, "unexpected tree owner, have %llu expect %llu",
9152                      btrfs_header_owner(eb), root->root_key.objectid);
9153         return -EUCLEAN;
9154 }
9155
9156 static noinline int walk_down_tree(struct btrfs_trans_handle *trans,
9157                                    struct btrfs_root *root,
9158                                    struct btrfs_path *path,
9159                                    struct walk_control *wc)
9160 {
9161         int level = wc->level;
9162         int lookup_info = 1;
9163         int ret;
9164
9165         while (level >= 0) {
9166                 ret = walk_down_proc(trans, root, path, wc, lookup_info);
9167                 if (ret > 0)
9168                         break;
9169
9170                 if (level == 0)
9171                         break;
9172
9173                 if (path->slots[level] >=
9174                     btrfs_header_nritems(path->nodes[level]))
9175                         break;
9176
9177                 ret = do_walk_down(trans, root, path, wc, &lookup_info);
9178                 if (ret > 0) {
9179                         path->slots[level]++;
9180                         continue;
9181                 } else if (ret < 0)
9182                         return ret;
9183                 level = wc->level;
9184         }
9185         return 0;
9186 }
9187
9188 static noinline int walk_up_tree(struct btrfs_trans_handle *trans,
9189                                  struct btrfs_root *root,
9190                                  struct btrfs_path *path,
9191                                  struct walk_control *wc, int max_level)
9192 {
9193         int level = wc->level;
9194         int ret;
9195
9196         path->slots[level] = btrfs_header_nritems(path->nodes[level]);
9197         while (level < max_level && path->nodes[level]) {
9198                 wc->level = level;
9199                 if (path->slots[level] + 1 <
9200                     btrfs_header_nritems(path->nodes[level])) {
9201                         path->slots[level]++;
9202                         return 0;
9203                 } else {
9204                         ret = walk_up_proc(trans, root, path, wc);
9205                         if (ret > 0)
9206                                 return 0;
9207                         if (ret < 0)
9208                                 return ret;
9209
9210                         if (path->locks[level]) {
9211                                 btrfs_tree_unlock_rw(path->nodes[level],
9212                                                      path->locks[level]);
9213                                 path->locks[level] = 0;
9214                         }
9215                         free_extent_buffer(path->nodes[level]);
9216                         path->nodes[level] = NULL;
9217                         level++;
9218                 }
9219         }
9220         return 1;
9221 }
9222
9223 /*
9224  * drop a subvolume tree.
9225  *
9226  * this function traverses the tree freeing any blocks that only
9227  * referenced by the tree.
9228  *
9229  * when a shared tree block is found. this function decreases its
9230  * reference count by one. if update_ref is true, this function
9231  * also make sure backrefs for the shared block and all lower level
9232  * blocks are properly updated.
9233  *
9234  * If called with for_reloc == 0, may exit early with -EAGAIN
9235  */
9236 int btrfs_drop_snapshot(struct btrfs_root *root,
9237                          struct btrfs_block_rsv *block_rsv, int update_ref,
9238                          int for_reloc)
9239 {
9240         struct btrfs_fs_info *fs_info = root->fs_info;
9241         struct btrfs_path *path;
9242         struct btrfs_trans_handle *trans;
9243         struct btrfs_root *tree_root = fs_info->tree_root;
9244         struct btrfs_root_item *root_item = &root->root_item;
9245         struct walk_control *wc;
9246         struct btrfs_key key;
9247         int err = 0;
9248         int ret;
9249         int level;
9250         bool root_dropped = false;
9251
9252         btrfs_debug(fs_info, "Drop subvolume %llu", root->root_key.objectid);
9253
9254         path = btrfs_alloc_path();
9255         if (!path) {
9256                 err = -ENOMEM;
9257                 goto out;
9258         }
9259
9260         wc = kzalloc(sizeof(*wc), GFP_NOFS);
9261         if (!wc) {
9262                 btrfs_free_path(path);
9263                 err = -ENOMEM;
9264                 goto out;
9265         }
9266
9267         trans = btrfs_start_transaction(tree_root, 0);
9268         if (IS_ERR(trans)) {
9269                 err = PTR_ERR(trans);
9270                 goto out_free;
9271         }
9272
9273         err = btrfs_run_delayed_items(trans);
9274         if (err)
9275                 goto out_end_trans;
9276
9277         if (block_rsv)
9278                 trans->block_rsv = block_rsv;
9279
9280         /*
9281          * This will help us catch people modifying the fs tree while we're
9282          * dropping it.  It is unsafe to mess with the fs tree while it's being
9283          * dropped as we unlock the root node and parent nodes as we walk down
9284          * the tree, assuming nothing will change.  If something does change
9285          * then we'll have stale information and drop references to blocks we've
9286          * already dropped.
9287          */
9288         set_bit(BTRFS_ROOT_DELETING, &root->state);
9289         if (btrfs_disk_key_objectid(&root_item->drop_progress) == 0) {
9290                 level = btrfs_header_level(root->node);
9291                 path->nodes[level] = btrfs_lock_root_node(root);
9292                 btrfs_set_lock_blocking_write(path->nodes[level]);
9293                 path->slots[level] = 0;
9294                 path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
9295                 memset(&wc->update_progress, 0,
9296                        sizeof(wc->update_progress));
9297         } else {
9298                 btrfs_disk_key_to_cpu(&key, &root_item->drop_progress);
9299                 memcpy(&wc->update_progress, &key,
9300                        sizeof(wc->update_progress));
9301
9302                 level = root_item->drop_level;
9303                 BUG_ON(level == 0);
9304                 path->lowest_level = level;
9305                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
9306                 path->lowest_level = 0;
9307                 if (ret < 0) {
9308                         err = ret;
9309                         goto out_end_trans;
9310                 }
9311                 WARN_ON(ret > 0);
9312
9313                 /*
9314                  * unlock our path, this is safe because only this
9315                  * function is allowed to delete this snapshot
9316                  */
9317                 btrfs_unlock_up_safe(path, 0);
9318
9319                 level = btrfs_header_level(root->node);
9320                 while (1) {
9321                         btrfs_tree_lock(path->nodes[level]);
9322                         btrfs_set_lock_blocking_write(path->nodes[level]);
9323                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
9324
9325                         ret = btrfs_lookup_extent_info(trans, fs_info,
9326                                                 path->nodes[level]->start,
9327                                                 level, 1, &wc->refs[level],
9328                                                 &wc->flags[level]);
9329                         if (ret < 0) {
9330                                 err = ret;
9331                                 goto out_end_trans;
9332                         }
9333                         BUG_ON(wc->refs[level] == 0);
9334
9335                         if (level == root_item->drop_level)
9336                                 break;
9337
9338                         btrfs_tree_unlock(path->nodes[level]);
9339                         path->locks[level] = 0;
9340                         WARN_ON(wc->refs[level] != 1);
9341                         level--;
9342                 }
9343         }
9344
9345         wc->restarted = test_bit(BTRFS_ROOT_DEAD_TREE, &root->state);
9346         wc->level = level;
9347         wc->shared_level = -1;
9348         wc->stage = DROP_REFERENCE;
9349         wc->update_ref = update_ref;
9350         wc->keep_locks = 0;
9351         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(fs_info);
9352
9353         while (1) {
9354
9355                 ret = walk_down_tree(trans, root, path, wc);
9356                 if (ret < 0) {
9357                         err = ret;
9358                         break;
9359                 }
9360
9361                 ret = walk_up_tree(trans, root, path, wc, BTRFS_MAX_LEVEL);
9362                 if (ret < 0) {
9363                         err = ret;
9364                         break;
9365                 }
9366
9367                 if (ret > 0) {
9368                         BUG_ON(wc->stage != DROP_REFERENCE);
9369                         break;
9370                 }
9371
9372                 if (wc->stage == DROP_REFERENCE) {
9373                         wc->drop_level = wc->level;
9374                         btrfs_node_key_to_cpu(path->nodes[wc->drop_level],
9375                                               &wc->drop_progress,
9376                                               path->slots[wc->drop_level]);
9377                 }
9378                 btrfs_cpu_key_to_disk(&root_item->drop_progress,
9379                                       &wc->drop_progress);
9380                 root_item->drop_level = wc->drop_level;
9381
9382                 BUG_ON(wc->level == 0);
9383                 if (btrfs_should_end_transaction(trans) ||
9384                     (!for_reloc && btrfs_need_cleaner_sleep(fs_info))) {
9385                         ret = btrfs_update_root(trans, tree_root,
9386                                                 &root->root_key,
9387                                                 root_item);
9388                         if (ret) {
9389                                 btrfs_abort_transaction(trans, ret);
9390                                 err = ret;
9391                                 goto out_end_trans;
9392                         }
9393
9394                         btrfs_end_transaction_throttle(trans);
9395                         if (!for_reloc && btrfs_need_cleaner_sleep(fs_info)) {
9396                                 btrfs_debug(fs_info,
9397                                             "drop snapshot early exit");
9398                                 err = -EAGAIN;
9399                                 goto out_free;
9400                         }
9401
9402                         trans = btrfs_start_transaction(tree_root, 0);
9403                         if (IS_ERR(trans)) {
9404                                 err = PTR_ERR(trans);
9405                                 goto out_free;
9406                         }
9407                         if (block_rsv)
9408                                 trans->block_rsv = block_rsv;
9409                 }
9410         }
9411         btrfs_release_path(path);
9412         if (err)
9413                 goto out_end_trans;
9414
9415         ret = btrfs_del_root(trans, &root->root_key);
9416         if (ret) {
9417                 btrfs_abort_transaction(trans, ret);
9418                 err = ret;
9419                 goto out_end_trans;
9420         }
9421
9422         if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) {
9423                 ret = btrfs_find_root(tree_root, &root->root_key, path,
9424                                       NULL, NULL);
9425                 if (ret < 0) {
9426                         btrfs_abort_transaction(trans, ret);
9427                         err = ret;
9428                         goto out_end_trans;
9429                 } else if (ret > 0) {
9430                         /* if we fail to delete the orphan item this time
9431                          * around, it'll get picked up the next time.
9432                          *
9433                          * The most common failure here is just -ENOENT.
9434                          */
9435                         btrfs_del_orphan_item(trans, tree_root,
9436                                               root->root_key.objectid);
9437                 }
9438         }
9439
9440         if (test_bit(BTRFS_ROOT_IN_RADIX, &root->state)) {
9441                 btrfs_add_dropped_root(trans, root);
9442         } else {
9443                 free_extent_buffer(root->node);
9444                 free_extent_buffer(root->commit_root);
9445                 btrfs_put_fs_root(root);
9446         }
9447         root_dropped = true;
9448 out_end_trans:
9449         btrfs_end_transaction_throttle(trans);
9450 out_free:
9451         kfree(wc);
9452         btrfs_free_path(path);
9453 out:
9454         /*
9455          * So if we need to stop dropping the snapshot for whatever reason we
9456          * need to make sure to add it back to the dead root list so that we
9457          * keep trying to do the work later.  This also cleans up roots if we
9458          * don't have it in the radix (like when we recover after a power fail
9459          * or unmount) so we don't leak memory.
9460          */
9461         if (!for_reloc && !root_dropped)
9462                 btrfs_add_dead_root(root);
9463         if (err && err != -EAGAIN)
9464                 btrfs_handle_fs_error(fs_info, err, NULL);
9465         return err;
9466 }
9467
9468 /*
9469  * drop subtree rooted at tree block 'node'.
9470  *
9471  * NOTE: this function will unlock and release tree block 'node'
9472  * only used by relocation code
9473  */
9474 int btrfs_drop_subtree(struct btrfs_trans_handle *trans,
9475                         struct btrfs_root *root,
9476                         struct extent_buffer *node,
9477                         struct extent_buffer *parent)
9478 {
9479         struct btrfs_fs_info *fs_info = root->fs_info;
9480         struct btrfs_path *path;
9481         struct walk_control *wc;
9482         int level;
9483         int parent_level;
9484         int ret = 0;
9485         int wret;
9486
9487         BUG_ON(root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID);
9488
9489         path = btrfs_alloc_path();
9490         if (!path)
9491                 return -ENOMEM;
9492
9493         wc = kzalloc(sizeof(*wc), GFP_NOFS);
9494         if (!wc) {
9495                 btrfs_free_path(path);
9496                 return -ENOMEM;
9497         }
9498
9499         btrfs_assert_tree_locked(parent);
9500         parent_level = btrfs_header_level(parent);
9501         extent_buffer_get(parent);
9502         path->nodes[parent_level] = parent;
9503         path->slots[parent_level] = btrfs_header_nritems(parent);
9504
9505         btrfs_assert_tree_locked(node);
9506         level = btrfs_header_level(node);
9507         path->nodes[level] = node;
9508         path->slots[level] = 0;
9509         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
9510
9511         wc->refs[parent_level] = 1;
9512         wc->flags[parent_level] = BTRFS_BLOCK_FLAG_FULL_BACKREF;
9513         wc->level = level;
9514         wc->shared_level = -1;
9515         wc->stage = DROP_REFERENCE;
9516         wc->update_ref = 0;
9517         wc->keep_locks = 1;
9518         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(fs_info);
9519
9520         while (1) {
9521                 wret = walk_down_tree(trans, root, path, wc);
9522                 if (wret < 0) {
9523                         ret = wret;
9524                         break;
9525                 }
9526
9527                 wret = walk_up_tree(trans, root, path, wc, parent_level);
9528                 if (wret < 0)
9529                         ret = wret;
9530                 if (wret != 0)
9531                         break;
9532         }
9533
9534         kfree(wc);
9535         btrfs_free_path(path);
9536         return ret;
9537 }
9538
9539 static u64 update_block_group_flags(struct btrfs_fs_info *fs_info, u64 flags)
9540 {
9541         u64 num_devices;
9542         u64 stripped;
9543
9544         /*
9545          * if restripe for this chunk_type is on pick target profile and
9546          * return, otherwise do the usual balance
9547          */
9548         stripped = get_restripe_target(fs_info, flags);
9549         if (stripped)
9550                 return extended_to_chunk(stripped);
9551
9552         num_devices = fs_info->fs_devices->rw_devices;
9553
9554         stripped = BTRFS_BLOCK_GROUP_RAID0 |
9555                 BTRFS_BLOCK_GROUP_RAID5 | BTRFS_BLOCK_GROUP_RAID6 |
9556                 BTRFS_BLOCK_GROUP_RAID1 | BTRFS_BLOCK_GROUP_RAID10;
9557
9558         if (num_devices == 1) {
9559                 stripped |= BTRFS_BLOCK_GROUP_DUP;
9560                 stripped = flags & ~stripped;
9561
9562                 /* turn raid0 into single device chunks */
9563                 if (flags & BTRFS_BLOCK_GROUP_RAID0)
9564                         return stripped;
9565
9566                 /* turn mirroring into duplication */
9567                 if (flags & (BTRFS_BLOCK_GROUP_RAID1 |
9568                              BTRFS_BLOCK_GROUP_RAID10))
9569                         return stripped | BTRFS_BLOCK_GROUP_DUP;
9570         } else {
9571                 /* they already had raid on here, just return */
9572                 if (flags & stripped)
9573                         return flags;
9574
9575                 stripped |= BTRFS_BLOCK_GROUP_DUP;
9576                 stripped = flags & ~stripped;
9577
9578                 /* switch duplicated blocks with raid1 */
9579                 if (flags & BTRFS_BLOCK_GROUP_DUP)
9580                         return stripped | BTRFS_BLOCK_GROUP_RAID1;
9581
9582                 /* this is drive concat, leave it alone */
9583         }
9584
9585         return flags;
9586 }
9587
9588 static int inc_block_group_ro(struct btrfs_block_group_cache *cache, int force)
9589 {
9590         struct btrfs_space_info *sinfo = cache->space_info;
9591         u64 num_bytes;
9592         u64 sinfo_used;
9593         u64 min_allocable_bytes;
9594         int ret = -ENOSPC;
9595
9596         /*
9597          * We need some metadata space and system metadata space for
9598          * allocating chunks in some corner cases until we force to set
9599          * it to be readonly.
9600          */
9601         if ((sinfo->flags &
9602              (BTRFS_BLOCK_GROUP_SYSTEM | BTRFS_BLOCK_GROUP_METADATA)) &&
9603             !force)
9604                 min_allocable_bytes = SZ_1M;
9605         else
9606                 min_allocable_bytes = 0;
9607
9608         spin_lock(&sinfo->lock);
9609         spin_lock(&cache->lock);
9610
9611         if (cache->ro) {
9612                 cache->ro++;
9613                 ret = 0;
9614                 goto out;
9615         }
9616
9617         num_bytes = cache->key.offset - cache->reserved - cache->pinned -
9618                     cache->bytes_super - btrfs_block_group_used(&cache->item);
9619         sinfo_used = btrfs_space_info_used(sinfo, true);
9620
9621         if (sinfo_used + num_bytes + min_allocable_bytes <=
9622             sinfo->total_bytes) {
9623                 sinfo->bytes_readonly += num_bytes;
9624                 cache->ro++;
9625                 list_add_tail(&cache->ro_list, &sinfo->ro_bgs);
9626                 ret = 0;
9627         }
9628 out:
9629         spin_unlock(&cache->lock);
9630         spin_unlock(&sinfo->lock);
9631         if (ret == -ENOSPC && btrfs_test_opt(cache->fs_info, ENOSPC_DEBUG)) {
9632                 btrfs_info(cache->fs_info,
9633                         "unable to make block group %llu ro",
9634                         cache->key.objectid);
9635                 btrfs_info(cache->fs_info,
9636                         "sinfo_used=%llu bg_num_bytes=%llu min_allocable=%llu",
9637                         sinfo_used, num_bytes, min_allocable_bytes);
9638                 dump_space_info(cache->fs_info, cache->space_info, 0, 0);
9639         }
9640         return ret;
9641 }
9642
9643 int btrfs_inc_block_group_ro(struct btrfs_block_group_cache *cache)
9644
9645 {
9646         struct btrfs_fs_info *fs_info = cache->fs_info;
9647         struct btrfs_trans_handle *trans;
9648         u64 alloc_flags;
9649         int ret;
9650
9651 again:
9652         trans = btrfs_join_transaction(fs_info->extent_root);
9653         if (IS_ERR(trans))
9654                 return PTR_ERR(trans);
9655
9656         /*
9657          * we're not allowed to set block groups readonly after the dirty
9658          * block groups cache has started writing.  If it already started,
9659          * back off and let this transaction commit
9660          */
9661         mutex_lock(&fs_info->ro_block_group_mutex);
9662         if (test_bit(BTRFS_TRANS_DIRTY_BG_RUN, &trans->transaction->flags)) {
9663                 u64 transid = trans->transid;
9664
9665                 mutex_unlock(&fs_info->ro_block_group_mutex);
9666                 btrfs_end_transaction(trans);
9667
9668                 ret = btrfs_wait_for_commit(fs_info, transid);
9669                 if (ret)
9670                         return ret;
9671                 goto again;
9672         }
9673
9674         /*
9675          * if we are changing raid levels, try to allocate a corresponding
9676          * block group with the new raid level.
9677          */
9678         alloc_flags = update_block_group_flags(fs_info, cache->flags);
9679         if (alloc_flags != cache->flags) {
9680                 ret = do_chunk_alloc(trans, alloc_flags,
9681                                      CHUNK_ALLOC_FORCE);
9682                 /*
9683                  * ENOSPC is allowed here, we may have enough space
9684                  * already allocated at the new raid level to
9685                  * carry on
9686                  */
9687                 if (ret == -ENOSPC)
9688                         ret = 0;
9689                 if (ret < 0)
9690                         goto out;
9691         }
9692
9693         ret = inc_block_group_ro(cache, 0);
9694         if (!ret)
9695                 goto out;
9696         alloc_flags = get_alloc_profile(fs_info, cache->space_info->flags);
9697         ret = do_chunk_alloc(trans, alloc_flags, CHUNK_ALLOC_FORCE);
9698         if (ret < 0)
9699                 goto out;
9700         ret = inc_block_group_ro(cache, 0);
9701 out:
9702         if (cache->flags & BTRFS_BLOCK_GROUP_SYSTEM) {
9703                 alloc_flags = update_block_group_flags(fs_info, cache->flags);
9704                 mutex_lock(&fs_info->chunk_mutex);
9705                 check_system_chunk(trans, alloc_flags);
9706                 mutex_unlock(&fs_info->chunk_mutex);
9707         }
9708         mutex_unlock(&fs_info->ro_block_group_mutex);
9709
9710         btrfs_end_transaction(trans);
9711         return ret;
9712 }
9713
9714 int btrfs_force_chunk_alloc(struct btrfs_trans_handle *trans, u64 type)
9715 {
9716         u64 alloc_flags = get_alloc_profile(trans->fs_info, type);
9717
9718         return do_chunk_alloc(trans, alloc_flags, CHUNK_ALLOC_FORCE);
9719 }
9720
9721 /*
9722  * helper to account the unused space of all the readonly block group in the
9723  * space_info. takes mirrors into account.
9724  */
9725 u64 btrfs_account_ro_block_groups_free_space(struct btrfs_space_info *sinfo)
9726 {
9727         struct btrfs_block_group_cache *block_group;
9728         u64 free_bytes = 0;
9729         int factor;
9730
9731         /* It's df, we don't care if it's racy */
9732         if (list_empty(&sinfo->ro_bgs))
9733                 return 0;
9734
9735         spin_lock(&sinfo->lock);
9736         list_for_each_entry(block_group, &sinfo->ro_bgs, ro_list) {
9737                 spin_lock(&block_group->lock);
9738
9739                 if (!block_group->ro) {
9740                         spin_unlock(&block_group->lock);
9741                         continue;
9742                 }
9743
9744                 factor = btrfs_bg_type_to_factor(block_group->flags);
9745                 free_bytes += (block_group->key.offset -
9746                                btrfs_block_group_used(&block_group->item)) *
9747                                factor;
9748
9749                 spin_unlock(&block_group->lock);
9750         }
9751         spin_unlock(&sinfo->lock);
9752
9753         return free_bytes;
9754 }
9755
9756 void btrfs_dec_block_group_ro(struct btrfs_block_group_cache *cache)
9757 {
9758         struct btrfs_space_info *sinfo = cache->space_info;
9759         u64 num_bytes;
9760
9761         BUG_ON(!cache->ro);
9762
9763         spin_lock(&sinfo->lock);
9764         spin_lock(&cache->lock);
9765         if (!--cache->ro) {
9766                 num_bytes = cache->key.offset - cache->reserved -
9767                             cache->pinned - cache->bytes_super -
9768                             btrfs_block_group_used(&cache->item);
9769                 sinfo->bytes_readonly -= num_bytes;
9770                 list_del_init(&cache->ro_list);
9771         }
9772         spin_unlock(&cache->lock);
9773         spin_unlock(&sinfo->lock);
9774 }
9775
9776 /*
9777  * Checks to see if it's even possible to relocate this block group.
9778  *
9779  * @return - -1 if it's not a good idea to relocate this block group, 0 if its
9780  * ok to go ahead and try.
9781  */
9782 int btrfs_can_relocate(struct btrfs_fs_info *fs_info, u64 bytenr)
9783 {
9784         struct btrfs_block_group_cache *block_group;
9785         struct btrfs_space_info *space_info;
9786         struct btrfs_fs_devices *fs_devices = fs_info->fs_devices;
9787         struct btrfs_device *device;
9788         u64 min_free;
9789         u64 dev_min = 1;
9790         u64 dev_nr = 0;
9791         u64 target;
9792         int debug;
9793         int index;
9794         int full = 0;
9795         int ret = 0;
9796
9797         debug = btrfs_test_opt(fs_info, ENOSPC_DEBUG);
9798
9799         block_group = btrfs_lookup_block_group(fs_info, bytenr);
9800
9801         /* odd, couldn't find the block group, leave it alone */
9802         if (!block_group) {
9803                 if (debug)
9804                         btrfs_warn(fs_info,
9805                                    "can't find block group for bytenr %llu",
9806                                    bytenr);
9807                 return -1;
9808         }
9809
9810         min_free = btrfs_block_group_used(&block_group->item);
9811
9812         /* no bytes used, we're good */
9813         if (!min_free)
9814                 goto out;
9815
9816         space_info = block_group->space_info;
9817         spin_lock(&space_info->lock);
9818
9819         full = space_info->full;
9820
9821         /*
9822          * if this is the last block group we have in this space, we can't
9823          * relocate it unless we're able to allocate a new chunk below.
9824          *
9825          * Otherwise, we need to make sure we have room in the space to handle
9826          * all of the extents from this block group.  If we can, we're good
9827          */
9828         if ((space_info->total_bytes != block_group->key.offset) &&
9829             (btrfs_space_info_used(space_info, false) + min_free <
9830              space_info->total_bytes)) {
9831                 spin_unlock(&space_info->lock);
9832                 goto out;
9833         }
9834         spin_unlock(&space_info->lock);
9835
9836         /*
9837          * ok we don't have enough space, but maybe we have free space on our
9838          * devices to allocate new chunks for relocation, so loop through our
9839          * alloc devices and guess if we have enough space.  if this block
9840          * group is going to be restriped, run checks against the target
9841          * profile instead of the current one.
9842          */
9843         ret = -1;
9844
9845         /*
9846          * index:
9847          *      0: raid10
9848          *      1: raid1
9849          *      2: dup
9850          *      3: raid0
9851          *      4: single
9852          */
9853         target = get_restripe_target(fs_info, block_group->flags);
9854         if (target) {
9855                 index = btrfs_bg_flags_to_raid_index(extended_to_chunk(target));
9856         } else {
9857                 /*
9858                  * this is just a balance, so if we were marked as full
9859                  * we know there is no space for a new chunk
9860                  */
9861                 if (full) {
9862                         if (debug)
9863                                 btrfs_warn(fs_info,
9864                                            "no space to alloc new chunk for block group %llu",
9865                                            block_group->key.objectid);
9866                         goto out;
9867                 }
9868
9869                 index = btrfs_bg_flags_to_raid_index(block_group->flags);
9870         }
9871
9872         if (index == BTRFS_RAID_RAID10) {
9873                 dev_min = 4;
9874                 /* Divide by 2 */
9875                 min_free >>= 1;
9876         } else if (index == BTRFS_RAID_RAID1) {
9877                 dev_min = 2;
9878         } else if (index == BTRFS_RAID_DUP) {
9879                 /* Multiply by 2 */
9880                 min_free <<= 1;
9881         } else if (index == BTRFS_RAID_RAID0) {
9882                 dev_min = fs_devices->rw_devices;
9883                 min_free = div64_u64(min_free, dev_min);
9884         }
9885
9886         mutex_lock(&fs_info->chunk_mutex);
9887         list_for_each_entry(device, &fs_devices->alloc_list, dev_alloc_list) {
9888                 u64 dev_offset;
9889
9890                 /*
9891                  * check to make sure we can actually find a chunk with enough
9892                  * space to fit our block group in.
9893                  */
9894                 if (device->total_bytes > device->bytes_used + min_free &&
9895                     !test_bit(BTRFS_DEV_STATE_REPLACE_TGT, &device->dev_state)) {
9896                         ret = find_free_dev_extent(device, min_free,
9897                                                    &dev_offset, NULL);
9898                         if (!ret)
9899                                 dev_nr++;
9900
9901                         if (dev_nr >= dev_min)
9902                                 break;
9903
9904                         ret = -1;
9905                 }
9906         }
9907         if (debug && ret == -1)
9908                 btrfs_warn(fs_info,
9909                            "no space to allocate a new chunk for block group %llu",
9910                            block_group->key.objectid);
9911         mutex_unlock(&fs_info->chunk_mutex);
9912 out:
9913         btrfs_put_block_group(block_group);
9914         return ret;
9915 }
9916
9917 static int find_first_block_group(struct btrfs_fs_info *fs_info,
9918                                   struct btrfs_path *path,
9919                                   struct btrfs_key *key)
9920 {
9921         struct btrfs_root *root = fs_info->extent_root;
9922         int ret = 0;
9923         struct btrfs_key found_key;
9924         struct extent_buffer *leaf;
9925         struct btrfs_block_group_item bg;
9926         u64 flags;
9927         int slot;
9928
9929         ret = btrfs_search_slot(NULL, root, key, path, 0, 0);
9930         if (ret < 0)
9931                 goto out;
9932
9933         while (1) {
9934                 slot = path->slots[0];
9935                 leaf = path->nodes[0];
9936                 if (slot >= btrfs_header_nritems(leaf)) {
9937                         ret = btrfs_next_leaf(root, path);
9938                         if (ret == 0)
9939                                 continue;
9940                         if (ret < 0)
9941                                 goto out;
9942                         break;
9943                 }
9944                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
9945
9946                 if (found_key.objectid >= key->objectid &&
9947                     found_key.type == BTRFS_BLOCK_GROUP_ITEM_KEY) {
9948                         struct extent_map_tree *em_tree;
9949                         struct extent_map *em;
9950
9951                         em_tree = &root->fs_info->mapping_tree.map_tree;
9952                         read_lock(&em_tree->lock);
9953                         em = lookup_extent_mapping(em_tree, found_key.objectid,
9954                                                    found_key.offset);
9955                         read_unlock(&em_tree->lock);
9956                         if (!em) {
9957                                 btrfs_err(fs_info,
9958                         "logical %llu len %llu found bg but no related chunk",
9959                                           found_key.objectid, found_key.offset);
9960                                 ret = -ENOENT;
9961                         } else if (em->start != found_key.objectid ||
9962                                    em->len != found_key.offset) {
9963                                 btrfs_err(fs_info,
9964                 "block group %llu len %llu mismatch with chunk %llu len %llu",
9965                                           found_key.objectid, found_key.offset,
9966                                           em->start, em->len);
9967                                 ret = -EUCLEAN;
9968                         } else {
9969                                 read_extent_buffer(leaf, &bg,
9970                                         btrfs_item_ptr_offset(leaf, slot),
9971                                         sizeof(bg));
9972                                 flags = btrfs_block_group_flags(&bg) &
9973                                         BTRFS_BLOCK_GROUP_TYPE_MASK;
9974
9975                                 if (flags != (em->map_lookup->type &
9976                                               BTRFS_BLOCK_GROUP_TYPE_MASK)) {
9977                                         btrfs_err(fs_info,
9978 "block group %llu len %llu type flags 0x%llx mismatch with chunk type flags 0x%llx",
9979                                                 found_key.objectid,
9980                                                 found_key.offset, flags,
9981                                                 (BTRFS_BLOCK_GROUP_TYPE_MASK &
9982                                                  em->map_lookup->type));
9983                                         ret = -EUCLEAN;
9984                                 } else {
9985                                         ret = 0;
9986                                 }
9987                         }
9988                         free_extent_map(em);
9989                         goto out;
9990                 }
9991                 path->slots[0]++;
9992         }
9993 out:
9994         return ret;
9995 }
9996
9997 void btrfs_put_block_group_cache(struct btrfs_fs_info *info)
9998 {
9999         struct btrfs_block_group_cache *block_group;
10000         u64 last = 0;
10001
10002         while (1) {
10003                 struct inode *inode;
10004
10005                 block_group = btrfs_lookup_first_block_group(info, last);
10006                 while (block_group) {
10007                         wait_block_group_cache_done(block_group);
10008                         spin_lock(&block_group->lock);
10009                         if (block_group->iref)
10010                                 break;
10011                         spin_unlock(&block_group->lock);
10012                         block_group = next_block_group(block_group);
10013                 }
10014                 if (!block_group) {
10015                         if (last == 0)
10016                                 break;
10017                         last = 0;
10018                         continue;
10019                 }
10020
10021                 inode = block_group->inode;
10022                 block_group->iref = 0;
10023                 block_group->inode = NULL;
10024                 spin_unlock(&block_group->lock);
10025                 ASSERT(block_group->io_ctl.inode == NULL);
10026                 iput(inode);
10027                 last = block_group->key.objectid + block_group->key.offset;
10028                 btrfs_put_block_group(block_group);
10029         }
10030 }
10031
10032 /*
10033  * Must be called only after stopping all workers, since we could have block
10034  * group caching kthreads running, and therefore they could race with us if we
10035  * freed the block groups before stopping them.
10036  */
10037 int btrfs_free_block_groups(struct btrfs_fs_info *info)
10038 {
10039         struct btrfs_block_group_cache *block_group;
10040         struct btrfs_space_info *space_info;
10041         struct btrfs_caching_control *caching_ctl;
10042         struct rb_node *n;
10043
10044         down_write(&info->commit_root_sem);
10045         while (!list_empty(&info->caching_block_groups)) {
10046                 caching_ctl = list_entry(info->caching_block_groups.next,
10047                                          struct btrfs_caching_control, list);
10048                 list_del(&caching_ctl->list);
10049                 put_caching_control(caching_ctl);
10050         }
10051         up_write(&info->commit_root_sem);
10052
10053         spin_lock(&info->unused_bgs_lock);
10054         while (!list_empty(&info->unused_bgs)) {
10055                 block_group = list_first_entry(&info->unused_bgs,
10056                                                struct btrfs_block_group_cache,
10057                                                bg_list);
10058                 list_del_init(&block_group->bg_list);
10059                 btrfs_put_block_group(block_group);
10060         }
10061         spin_unlock(&info->unused_bgs_lock);
10062
10063         spin_lock(&info->block_group_cache_lock);
10064         while ((n = rb_last(&info->block_group_cache_tree)) != NULL) {
10065                 block_group = rb_entry(n, struct btrfs_block_group_cache,
10066                                        cache_node);
10067                 rb_erase(&block_group->cache_node,
10068                          &info->block_group_cache_tree);
10069                 RB_CLEAR_NODE(&block_group->cache_node);
10070                 spin_unlock(&info->block_group_cache_lock);
10071
10072                 down_write(&block_group->space_info->groups_sem);
10073                 list_del(&block_group->list);
10074                 up_write(&block_group->space_info->groups_sem);
10075
10076                 /*
10077                  * We haven't cached this block group, which means we could
10078                  * possibly have excluded extents on this block group.
10079                  */
10080                 if (block_group->cached == BTRFS_CACHE_NO ||
10081                     block_group->cached == BTRFS_CACHE_ERROR)
10082                         free_excluded_extents(block_group);
10083
10084                 btrfs_remove_free_space_cache(block_group);
10085                 ASSERT(block_group->cached != BTRFS_CACHE_STARTED);
10086                 ASSERT(list_empty(&block_group->dirty_list));
10087                 ASSERT(list_empty(&block_group->io_list));
10088                 ASSERT(list_empty(&block_group->bg_list));
10089                 ASSERT(atomic_read(&block_group->count) == 1);
10090                 btrfs_put_block_group(block_group);
10091
10092                 spin_lock(&info->block_group_cache_lock);
10093         }
10094         spin_unlock(&info->block_group_cache_lock);
10095
10096         /* now that all the block groups are freed, go through and
10097          * free all the space_info structs.  This is only called during
10098          * the final stages of unmount, and so we know nobody is
10099          * using them.  We call synchronize_rcu() once before we start,
10100          * just to be on the safe side.
10101          */
10102         synchronize_rcu();
10103
10104         release_global_block_rsv(info);
10105
10106         while (!list_empty(&info->space_info)) {
10107                 int i;
10108
10109                 space_info = list_entry(info->space_info.next,
10110                                         struct btrfs_space_info,
10111                                         list);
10112
10113                 /*
10114                  * Do not hide this behind enospc_debug, this is actually
10115                  * important and indicates a real bug if this happens.
10116                  */
10117                 if (WARN_ON(space_info->bytes_pinned > 0 ||
10118                             space_info->bytes_reserved > 0 ||
10119                             space_info->bytes_may_use > 0))
10120                         dump_space_info(info, space_info, 0, 0);
10121                 list_del(&space_info->list);
10122                 for (i = 0; i < BTRFS_NR_RAID_TYPES; i++) {
10123                         struct kobject *kobj;
10124                         kobj = space_info->block_group_kobjs[i];
10125                         space_info->block_group_kobjs[i] = NULL;
10126                         if (kobj) {
10127                                 kobject_del(kobj);
10128                                 kobject_put(kobj);
10129                         }
10130                 }
10131                 kobject_del(&space_info->kobj);
10132                 kobject_put(&space_info->kobj);
10133         }
10134         return 0;
10135 }
10136
10137 /* link_block_group will queue up kobjects to add when we're reclaim-safe */
10138 void btrfs_add_raid_kobjects(struct btrfs_fs_info *fs_info)
10139 {
10140         struct btrfs_space_info *space_info;
10141         struct raid_kobject *rkobj;
10142         LIST_HEAD(list);
10143         int index;
10144         int ret = 0;
10145
10146         spin_lock(&fs_info->pending_raid_kobjs_lock);
10147         list_splice_init(&fs_info->pending_raid_kobjs, &list);
10148         spin_unlock(&fs_info->pending_raid_kobjs_lock);
10149
10150         list_for_each_entry(rkobj, &list, list) {
10151                 space_info = __find_space_info(fs_info, rkobj->flags);
10152                 index = btrfs_bg_flags_to_raid_index(rkobj->flags);
10153
10154                 ret = kobject_add(&rkobj->kobj, &space_info->kobj,
10155                                   "%s", get_raid_name(index));
10156                 if (ret) {
10157                         kobject_put(&rkobj->kobj);
10158                         break;
10159                 }
10160         }
10161         if (ret)
10162                 btrfs_warn(fs_info,
10163                            "failed to add kobject for block cache, ignoring");
10164 }
10165
10166 static void link_block_group(struct btrfs_block_group_cache *cache)
10167 {
10168         struct btrfs_space_info *space_info = cache->space_info;
10169         struct btrfs_fs_info *fs_info = cache->fs_info;
10170         int index = btrfs_bg_flags_to_raid_index(cache->flags);
10171         bool first = false;
10172
10173         down_write(&space_info->groups_sem);
10174         if (list_empty(&space_info->block_groups[index]))
10175                 first = true;
10176         list_add_tail(&cache->list, &space_info->block_groups[index]);
10177         up_write(&space_info->groups_sem);
10178
10179         if (first) {
10180                 struct raid_kobject *rkobj = kzalloc(sizeof(*rkobj), GFP_NOFS);
10181                 if (!rkobj) {
10182                         btrfs_warn(cache->fs_info,
10183                                 "couldn't alloc memory for raid level kobject");
10184                         return;
10185                 }
10186                 rkobj->flags = cache->flags;
10187                 kobject_init(&rkobj->kobj, &btrfs_raid_ktype);
10188
10189                 spin_lock(&fs_info->pending_raid_kobjs_lock);
10190                 list_add_tail(&rkobj->list, &fs_info->pending_raid_kobjs);
10191                 spin_unlock(&fs_info->pending_raid_kobjs_lock);
10192                 space_info->block_group_kobjs[index] = &rkobj->kobj;
10193         }
10194 }
10195
10196 static struct btrfs_block_group_cache *
10197 btrfs_create_block_group_cache(struct btrfs_fs_info *fs_info,
10198                                u64 start, u64 size)
10199 {
10200         struct btrfs_block_group_cache *cache;
10201
10202         cache = kzalloc(sizeof(*cache), GFP_NOFS);
10203         if (!cache)
10204                 return NULL;
10205
10206         cache->free_space_ctl = kzalloc(sizeof(*cache->free_space_ctl),
10207                                         GFP_NOFS);
10208         if (!cache->free_space_ctl) {
10209                 kfree(cache);
10210                 return NULL;
10211         }
10212
10213         cache->key.objectid = start;
10214         cache->key.offset = size;
10215         cache->key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
10216
10217         cache->fs_info = fs_info;
10218         cache->full_stripe_len = btrfs_full_stripe_len(fs_info, start);
10219         set_free_space_tree_thresholds(cache);
10220
10221         atomic_set(&cache->count, 1);
10222         spin_lock_init(&cache->lock);
10223         init_rwsem(&cache->data_rwsem);
10224         INIT_LIST_HEAD(&cache->list);
10225         INIT_LIST_HEAD(&cache->cluster_list);
10226         INIT_LIST_HEAD(&cache->bg_list);
10227         INIT_LIST_HEAD(&cache->ro_list);
10228         INIT_LIST_HEAD(&cache->dirty_list);
10229         INIT_LIST_HEAD(&cache->io_list);
10230         btrfs_init_free_space_ctl(cache);
10231         atomic_set(&cache->trimming, 0);
10232         mutex_init(&cache->free_space_lock);
10233         btrfs_init_full_stripe_locks_tree(&cache->full_stripe_locks_root);
10234
10235         return cache;
10236 }
10237
10238
10239 /*
10240  * Iterate all chunks and verify that each of them has the corresponding block
10241  * group
10242  */
10243 static int check_chunk_block_group_mappings(struct btrfs_fs_info *fs_info)
10244 {
10245         struct btrfs_mapping_tree *map_tree = &fs_info->mapping_tree;
10246         struct extent_map *em;
10247         struct btrfs_block_group_cache *bg;
10248         u64 start = 0;
10249         int ret = 0;
10250
10251         while (1) {
10252                 read_lock(&map_tree->map_tree.lock);
10253                 /*
10254                  * lookup_extent_mapping will return the first extent map
10255                  * intersecting the range, so setting @len to 1 is enough to
10256                  * get the first chunk.
10257                  */
10258                 em = lookup_extent_mapping(&map_tree->map_tree, start, 1);
10259                 read_unlock(&map_tree->map_tree.lock);
10260                 if (!em)
10261                         break;
10262
10263                 bg = btrfs_lookup_block_group(fs_info, em->start);
10264                 if (!bg) {
10265                         btrfs_err(fs_info,
10266         "chunk start=%llu len=%llu doesn't have corresponding block group",
10267                                      em->start, em->len);
10268                         ret = -EUCLEAN;
10269                         free_extent_map(em);
10270                         break;
10271                 }
10272                 if (bg->key.objectid != em->start ||
10273                     bg->key.offset != em->len ||
10274                     (bg->flags & BTRFS_BLOCK_GROUP_TYPE_MASK) !=
10275                     (em->map_lookup->type & BTRFS_BLOCK_GROUP_TYPE_MASK)) {
10276                         btrfs_err(fs_info,
10277 "chunk start=%llu len=%llu flags=0x%llx doesn't match block group start=%llu len=%llu flags=0x%llx",
10278                                 em->start, em->len,
10279                                 em->map_lookup->type & BTRFS_BLOCK_GROUP_TYPE_MASK,
10280                                 bg->key.objectid, bg->key.offset,
10281                                 bg->flags & BTRFS_BLOCK_GROUP_TYPE_MASK);
10282                         ret = -EUCLEAN;
10283                         free_extent_map(em);
10284                         btrfs_put_block_group(bg);
10285                         break;
10286                 }
10287                 start = em->start + em->len;
10288                 free_extent_map(em);
10289                 btrfs_put_block_group(bg);
10290         }
10291         return ret;
10292 }
10293
10294 int btrfs_read_block_groups(struct btrfs_fs_info *info)
10295 {
10296         struct btrfs_path *path;
10297         int ret;
10298         struct btrfs_block_group_cache *cache;
10299         struct btrfs_space_info *space_info;
10300         struct btrfs_key key;
10301         struct btrfs_key found_key;
10302         struct extent_buffer *leaf;
10303         int need_clear = 0;
10304         u64 cache_gen;
10305         u64 feature;
10306         int mixed;
10307
10308         feature = btrfs_super_incompat_flags(info->super_copy);
10309         mixed = !!(feature & BTRFS_FEATURE_INCOMPAT_MIXED_GROUPS);
10310
10311         key.objectid = 0;
10312         key.offset = 0;
10313         key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
10314         path = btrfs_alloc_path();
10315         if (!path)
10316                 return -ENOMEM;
10317         path->reada = READA_FORWARD;
10318
10319         cache_gen = btrfs_super_cache_generation(info->super_copy);
10320         if (btrfs_test_opt(info, SPACE_CACHE) &&
10321             btrfs_super_generation(info->super_copy) != cache_gen)
10322                 need_clear = 1;
10323         if (btrfs_test_opt(info, CLEAR_CACHE))
10324                 need_clear = 1;
10325
10326         while (1) {
10327                 ret = find_first_block_group(info, path, &key);
10328                 if (ret > 0)
10329                         break;
10330                 if (ret != 0)
10331                         goto error;
10332
10333                 leaf = path->nodes[0];
10334                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
10335
10336                 cache = btrfs_create_block_group_cache(info, found_key.objectid,
10337                                                        found_key.offset);
10338                 if (!cache) {
10339                         ret = -ENOMEM;
10340                         goto error;
10341                 }
10342
10343                 if (need_clear) {
10344                         /*
10345                          * When we mount with old space cache, we need to
10346                          * set BTRFS_DC_CLEAR and set dirty flag.
10347                          *
10348                          * a) Setting 'BTRFS_DC_CLEAR' makes sure that we
10349                          *    truncate the old free space cache inode and
10350                          *    setup a new one.
10351                          * b) Setting 'dirty flag' makes sure that we flush
10352                          *    the new space cache info onto disk.
10353                          */
10354                         if (btrfs_test_opt(info, SPACE_CACHE))
10355                                 cache->disk_cache_state = BTRFS_DC_CLEAR;
10356                 }
10357
10358                 read_extent_buffer(leaf, &cache->item,
10359                                    btrfs_item_ptr_offset(leaf, path->slots[0]),
10360                                    sizeof(cache->item));
10361                 cache->flags = btrfs_block_group_flags(&cache->item);
10362                 if (!mixed &&
10363                     ((cache->flags & BTRFS_BLOCK_GROUP_METADATA) &&
10364                     (cache->flags & BTRFS_BLOCK_GROUP_DATA))) {
10365                         btrfs_err(info,
10366 "bg %llu is a mixed block group but filesystem hasn't enabled mixed block groups",
10367                                   cache->key.objectid);
10368                         ret = -EINVAL;
10369                         goto error;
10370                 }
10371
10372                 key.objectid = found_key.objectid + found_key.offset;
10373                 btrfs_release_path(path);
10374
10375                 /*
10376                  * We need to exclude the super stripes now so that the space
10377                  * info has super bytes accounted for, otherwise we'll think
10378                  * we have more space than we actually do.
10379                  */
10380                 ret = exclude_super_stripes(cache);
10381                 if (ret) {
10382                         /*
10383                          * We may have excluded something, so call this just in
10384                          * case.
10385                          */
10386                         free_excluded_extents(cache);
10387                         btrfs_put_block_group(cache);
10388                         goto error;
10389                 }
10390
10391                 /*
10392                  * check for two cases, either we are full, and therefore
10393                  * don't need to bother with the caching work since we won't
10394                  * find any space, or we are empty, and we can just add all
10395                  * the space in and be done with it.  This saves us _a_lot_ of
10396                  * time, particularly in the full case.
10397                  */
10398                 if (found_key.offset == btrfs_block_group_used(&cache->item)) {
10399                         cache->last_byte_to_unpin = (u64)-1;
10400                         cache->cached = BTRFS_CACHE_FINISHED;
10401                         free_excluded_extents(cache);
10402                 } else if (btrfs_block_group_used(&cache->item) == 0) {
10403                         cache->last_byte_to_unpin = (u64)-1;
10404                         cache->cached = BTRFS_CACHE_FINISHED;
10405                         add_new_free_space(cache, found_key.objectid,
10406                                            found_key.objectid +
10407                                            found_key.offset);
10408                         free_excluded_extents(cache);
10409                 }
10410
10411                 ret = btrfs_add_block_group_cache(info, cache);
10412                 if (ret) {
10413                         btrfs_remove_free_space_cache(cache);
10414                         btrfs_put_block_group(cache);
10415                         goto error;
10416                 }
10417
10418                 trace_btrfs_add_block_group(info, cache, 0);
10419                 update_space_info(info, cache->flags, found_key.offset,
10420                                   btrfs_block_group_used(&cache->item),
10421                                   cache->bytes_super, &space_info);
10422
10423                 cache->space_info = space_info;
10424
10425                 link_block_group(cache);
10426
10427                 set_avail_alloc_bits(info, cache->flags);
10428                 if (btrfs_chunk_readonly(info, cache->key.objectid)) {
10429                         inc_block_group_ro(cache, 1);
10430                 } else if (btrfs_block_group_used(&cache->item) == 0) {
10431                         ASSERT(list_empty(&cache->bg_list));
10432                         btrfs_mark_bg_unused(cache);
10433                 }
10434         }
10435
10436         list_for_each_entry_rcu(space_info, &info->space_info, list) {
10437                 if (!(get_alloc_profile(info, space_info->flags) &
10438                       (BTRFS_BLOCK_GROUP_RAID10 |
10439                        BTRFS_BLOCK_GROUP_RAID1 |
10440                        BTRFS_BLOCK_GROUP_RAID5 |
10441                        BTRFS_BLOCK_GROUP_RAID6 |
10442                        BTRFS_BLOCK_GROUP_DUP)))
10443                         continue;
10444                 /*
10445                  * avoid allocating from un-mirrored block group if there are
10446                  * mirrored block groups.
10447                  */
10448                 list_for_each_entry(cache,
10449                                 &space_info->block_groups[BTRFS_RAID_RAID0],
10450                                 list)
10451                         inc_block_group_ro(cache, 1);
10452                 list_for_each_entry(cache,
10453                                 &space_info->block_groups[BTRFS_RAID_SINGLE],
10454                                 list)
10455                         inc_block_group_ro(cache, 1);
10456         }
10457
10458         btrfs_add_raid_kobjects(info);
10459         init_global_block_rsv(info);
10460         ret = check_chunk_block_group_mappings(info);
10461 error:
10462         btrfs_free_path(path);
10463         return ret;
10464 }
10465
10466 void btrfs_create_pending_block_groups(struct btrfs_trans_handle *trans)
10467 {
10468         struct btrfs_fs_info *fs_info = trans->fs_info;
10469         struct btrfs_block_group_cache *block_group;
10470         struct btrfs_root *extent_root = fs_info->extent_root;
10471         struct btrfs_block_group_item item;
10472         struct btrfs_key key;
10473         int ret = 0;
10474
10475         if (!trans->can_flush_pending_bgs)
10476                 return;
10477
10478         while (!list_empty(&trans->new_bgs)) {
10479                 block_group = list_first_entry(&trans->new_bgs,
10480                                                struct btrfs_block_group_cache,
10481                                                bg_list);
10482                 if (ret)
10483                         goto next;
10484
10485                 spin_lock(&block_group->lock);
10486                 memcpy(&item, &block_group->item, sizeof(item));
10487                 memcpy(&key, &block_group->key, sizeof(key));
10488                 spin_unlock(&block_group->lock);
10489
10490                 ret = btrfs_insert_item(trans, extent_root, &key, &item,
10491                                         sizeof(item));
10492                 if (ret)
10493                         btrfs_abort_transaction(trans, ret);
10494                 ret = btrfs_finish_chunk_alloc(trans, key.objectid, key.offset);
10495                 if (ret)
10496                         btrfs_abort_transaction(trans, ret);
10497                 add_block_group_free_space(trans, block_group);
10498                 /* already aborted the transaction if it failed. */
10499 next:
10500                 btrfs_delayed_refs_rsv_release(fs_info, 1);
10501                 list_del_init(&block_group->bg_list);
10502         }
10503         btrfs_trans_release_chunk_metadata(trans);
10504 }
10505
10506 int btrfs_make_block_group(struct btrfs_trans_handle *trans, u64 bytes_used,
10507                            u64 type, u64 chunk_offset, u64 size)
10508 {
10509         struct btrfs_fs_info *fs_info = trans->fs_info;
10510         struct btrfs_block_group_cache *cache;
10511         int ret;
10512
10513         btrfs_set_log_full_commit(trans);
10514
10515         cache = btrfs_create_block_group_cache(fs_info, chunk_offset, size);
10516         if (!cache)
10517                 return -ENOMEM;
10518
10519         btrfs_set_block_group_used(&cache->item, bytes_used);
10520         btrfs_set_block_group_chunk_objectid(&cache->item,
10521                                              BTRFS_FIRST_CHUNK_TREE_OBJECTID);
10522         btrfs_set_block_group_flags(&cache->item, type);
10523
10524         cache->flags = type;
10525         cache->last_byte_to_unpin = (u64)-1;
10526         cache->cached = BTRFS_CACHE_FINISHED;
10527         cache->needs_free_space = 1;
10528         ret = exclude_super_stripes(cache);
10529         if (ret) {
10530                 /*
10531                  * We may have excluded something, so call this just in
10532                  * case.
10533                  */
10534                 free_excluded_extents(cache);
10535                 btrfs_put_block_group(cache);
10536                 return ret;
10537         }
10538
10539         add_new_free_space(cache, chunk_offset, chunk_offset + size);
10540
10541         free_excluded_extents(cache);
10542
10543 #ifdef CONFIG_BTRFS_DEBUG
10544         if (btrfs_should_fragment_free_space(cache)) {
10545                 u64 new_bytes_used = size - bytes_used;
10546
10547                 bytes_used += new_bytes_used >> 1;
10548                 fragment_free_space(cache);
10549         }
10550 #endif
10551         /*
10552          * Ensure the corresponding space_info object is created and
10553          * assigned to our block group. We want our bg to be added to the rbtree
10554          * with its ->space_info set.
10555          */
10556         cache->space_info = __find_space_info(fs_info, cache->flags);
10557         ASSERT(cache->space_info);
10558
10559         ret = btrfs_add_block_group_cache(fs_info, cache);
10560         if (ret) {
10561                 btrfs_remove_free_space_cache(cache);
10562                 btrfs_put_block_group(cache);
10563                 return ret;
10564         }
10565
10566         /*
10567          * Now that our block group has its ->space_info set and is inserted in
10568          * the rbtree, update the space info's counters.
10569          */
10570         trace_btrfs_add_block_group(fs_info, cache, 1);
10571         update_space_info(fs_info, cache->flags, size, bytes_used,
10572                                 cache->bytes_super, &cache->space_info);
10573         update_global_block_rsv(fs_info);
10574
10575         link_block_group(cache);
10576
10577         list_add_tail(&cache->bg_list, &trans->new_bgs);
10578         trans->delayed_ref_updates++;
10579         btrfs_update_delayed_refs_rsv(trans);
10580
10581         set_avail_alloc_bits(fs_info, type);
10582         return 0;
10583 }
10584
10585 static void clear_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
10586 {
10587         u64 extra_flags = chunk_to_extended(flags) &
10588                                 BTRFS_EXTENDED_PROFILE_MASK;
10589
10590         write_seqlock(&fs_info->profiles_lock);
10591         if (flags & BTRFS_BLOCK_GROUP_DATA)
10592                 fs_info->avail_data_alloc_bits &= ~extra_flags;
10593         if (flags & BTRFS_BLOCK_GROUP_METADATA)
10594                 fs_info->avail_metadata_alloc_bits &= ~extra_flags;
10595         if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
10596                 fs_info->avail_system_alloc_bits &= ~extra_flags;
10597         write_sequnlock(&fs_info->profiles_lock);
10598 }
10599
10600 int btrfs_remove_block_group(struct btrfs_trans_handle *trans,
10601                              u64 group_start, struct extent_map *em)
10602 {
10603         struct btrfs_fs_info *fs_info = trans->fs_info;
10604         struct btrfs_root *root = fs_info->extent_root;
10605         struct btrfs_path *path;
10606         struct btrfs_block_group_cache *block_group;
10607         struct btrfs_free_cluster *cluster;
10608         struct btrfs_root *tree_root = fs_info->tree_root;
10609         struct btrfs_key key;
10610         struct inode *inode;
10611         struct kobject *kobj = NULL;
10612         int ret;
10613         int index;
10614         int factor;
10615         struct btrfs_caching_control *caching_ctl = NULL;
10616         bool remove_em;
10617         bool remove_rsv = false;
10618
10619         block_group = btrfs_lookup_block_group(fs_info, group_start);
10620         BUG_ON(!block_group);
10621         BUG_ON(!block_group->ro);
10622
10623         trace_btrfs_remove_block_group(block_group);
10624         /*
10625          * Free the reserved super bytes from this block group before
10626          * remove it.
10627          */
10628         free_excluded_extents(block_group);
10629         btrfs_free_ref_tree_range(fs_info, block_group->key.objectid,
10630                                   block_group->key.offset);
10631
10632         memcpy(&key, &block_group->key, sizeof(key));
10633         index = btrfs_bg_flags_to_raid_index(block_group->flags);
10634         factor = btrfs_bg_type_to_factor(block_group->flags);
10635
10636         /* make sure this block group isn't part of an allocation cluster */
10637         cluster = &fs_info->data_alloc_cluster;
10638         spin_lock(&cluster->refill_lock);
10639         btrfs_return_cluster_to_free_space(block_group, cluster);
10640         spin_unlock(&cluster->refill_lock);
10641
10642         /*
10643          * make sure this block group isn't part of a metadata
10644          * allocation cluster
10645          */
10646         cluster = &fs_info->meta_alloc_cluster;
10647         spin_lock(&cluster->refill_lock);
10648         btrfs_return_cluster_to_free_space(block_group, cluster);
10649         spin_unlock(&cluster->refill_lock);
10650
10651         path = btrfs_alloc_path();
10652         if (!path) {
10653                 ret = -ENOMEM;
10654                 goto out;
10655         }
10656
10657         /*
10658          * get the inode first so any iput calls done for the io_list
10659          * aren't the final iput (no unlinks allowed now)
10660          */
10661         inode = lookup_free_space_inode(block_group, path);
10662
10663         mutex_lock(&trans->transaction->cache_write_mutex);
10664         /*
10665          * Make sure our free space cache IO is done before removing the
10666          * free space inode
10667          */
10668         spin_lock(&trans->transaction->dirty_bgs_lock);
10669         if (!list_empty(&block_group->io_list)) {
10670                 list_del_init(&block_group->io_list);
10671
10672                 WARN_ON(!IS_ERR(inode) && inode != block_group->io_ctl.inode);
10673
10674                 spin_unlock(&trans->transaction->dirty_bgs_lock);
10675                 btrfs_wait_cache_io(trans, block_group, path);
10676                 btrfs_put_block_group(block_group);
10677                 spin_lock(&trans->transaction->dirty_bgs_lock);
10678         }
10679
10680         if (!list_empty(&block_group->dirty_list)) {
10681                 list_del_init(&block_group->dirty_list);
10682                 remove_rsv = true;
10683                 btrfs_put_block_group(block_group);
10684         }
10685         spin_unlock(&trans->transaction->dirty_bgs_lock);
10686         mutex_unlock(&trans->transaction->cache_write_mutex);
10687
10688         if (!IS_ERR(inode)) {
10689                 ret = btrfs_orphan_add(trans, BTRFS_I(inode));
10690                 if (ret) {
10691                         btrfs_add_delayed_iput(inode);
10692                         goto out;
10693                 }
10694                 clear_nlink(inode);
10695                 /* One for the block groups ref */
10696                 spin_lock(&block_group->lock);
10697                 if (block_group->iref) {
10698                         block_group->iref = 0;
10699                         block_group->inode = NULL;
10700                         spin_unlock(&block_group->lock);
10701                         iput(inode);
10702                 } else {
10703                         spin_unlock(&block_group->lock);
10704                 }
10705                 /* One for our lookup ref */
10706                 btrfs_add_delayed_iput(inode);
10707         }
10708
10709         key.objectid = BTRFS_FREE_SPACE_OBJECTID;
10710         key.offset = block_group->key.objectid;
10711         key.type = 0;
10712
10713         ret = btrfs_search_slot(trans, tree_root, &key, path, -1, 1);
10714         if (ret < 0)
10715                 goto out;
10716         if (ret > 0)
10717                 btrfs_release_path(path);
10718         if (ret == 0) {
10719                 ret = btrfs_del_item(trans, tree_root, path);
10720                 if (ret)
10721                         goto out;
10722                 btrfs_release_path(path);
10723         }
10724
10725         spin_lock(&fs_info->block_group_cache_lock);
10726         rb_erase(&block_group->cache_node,
10727                  &fs_info->block_group_cache_tree);
10728         RB_CLEAR_NODE(&block_group->cache_node);
10729
10730         if (fs_info->first_logical_byte == block_group->key.objectid)
10731                 fs_info->first_logical_byte = (u64)-1;
10732         spin_unlock(&fs_info->block_group_cache_lock);
10733
10734         down_write(&block_group->space_info->groups_sem);
10735         /*
10736          * we must use list_del_init so people can check to see if they
10737          * are still on the list after taking the semaphore
10738          */
10739         list_del_init(&block_group->list);
10740         if (list_empty(&block_group->space_info->block_groups[index])) {
10741                 kobj = block_group->space_info->block_group_kobjs[index];
10742                 block_group->space_info->block_group_kobjs[index] = NULL;
10743                 clear_avail_alloc_bits(fs_info, block_group->flags);
10744         }
10745         up_write(&block_group->space_info->groups_sem);
10746         if (kobj) {
10747                 kobject_del(kobj);
10748                 kobject_put(kobj);
10749         }
10750
10751         if (block_group->has_caching_ctl)
10752                 caching_ctl = get_caching_control(block_group);
10753         if (block_group->cached == BTRFS_CACHE_STARTED)
10754                 wait_block_group_cache_done(block_group);
10755         if (block_group->has_caching_ctl) {
10756                 down_write(&fs_info->commit_root_sem);
10757                 if (!caching_ctl) {
10758                         struct btrfs_caching_control *ctl;
10759
10760                         list_for_each_entry(ctl,
10761                                     &fs_info->caching_block_groups, list)
10762                                 if (ctl->block_group == block_group) {
10763                                         caching_ctl = ctl;
10764                                         refcount_inc(&caching_ctl->count);
10765                                         break;
10766                                 }
10767                 }
10768                 if (caching_ctl)
10769                         list_del_init(&caching_ctl->list);
10770                 up_write(&fs_info->commit_root_sem);
10771                 if (caching_ctl) {
10772                         /* Once for the caching bgs list and once for us. */
10773                         put_caching_control(caching_ctl);
10774                         put_caching_control(caching_ctl);
10775                 }
10776         }
10777
10778         spin_lock(&trans->transaction->dirty_bgs_lock);
10779         WARN_ON(!list_empty(&block_group->dirty_list));
10780         WARN_ON(!list_empty(&block_group->io_list));
10781         spin_unlock(&trans->transaction->dirty_bgs_lock);
10782
10783         btrfs_remove_free_space_cache(block_group);
10784
10785         spin_lock(&block_group->space_info->lock);
10786         list_del_init(&block_group->ro_list);
10787
10788         if (btrfs_test_opt(fs_info, ENOSPC_DEBUG)) {
10789                 WARN_ON(block_group->space_info->total_bytes
10790                         < block_group->key.offset);
10791                 WARN_ON(block_group->space_info->bytes_readonly
10792                         < block_group->key.offset);
10793                 WARN_ON(block_group->space_info->disk_total
10794                         < block_group->key.offset * factor);
10795         }
10796         block_group->space_info->total_bytes -= block_group->key.offset;
10797         block_group->space_info->bytes_readonly -= block_group->key.offset;
10798         block_group->space_info->disk_total -= block_group->key.offset * factor;
10799
10800         spin_unlock(&block_group->space_info->lock);
10801
10802         memcpy(&key, &block_group->key, sizeof(key));
10803
10804         mutex_lock(&fs_info->chunk_mutex);
10805         spin_lock(&block_group->lock);
10806         block_group->removed = 1;
10807         /*
10808          * At this point trimming can't start on this block group, because we
10809          * removed the block group from the tree fs_info->block_group_cache_tree
10810          * so no one can't find it anymore and even if someone already got this
10811          * block group before we removed it from the rbtree, they have already
10812          * incremented block_group->trimming - if they didn't, they won't find
10813          * any free space entries because we already removed them all when we
10814          * called btrfs_remove_free_space_cache().
10815          *
10816          * And we must not remove the extent map from the fs_info->mapping_tree
10817          * to prevent the same logical address range and physical device space
10818          * ranges from being reused for a new block group. This is because our
10819          * fs trim operation (btrfs_trim_fs() / btrfs_ioctl_fitrim()) is
10820          * completely transactionless, so while it is trimming a range the
10821          * currently running transaction might finish and a new one start,
10822          * allowing for new block groups to be created that can reuse the same
10823          * physical device locations unless we take this special care.
10824          *
10825          * There may also be an implicit trim operation if the file system
10826          * is mounted with -odiscard. The same protections must remain
10827          * in place until the extents have been discarded completely when
10828          * the transaction commit has completed.
10829          */
10830         remove_em = (atomic_read(&block_group->trimming) == 0);
10831         spin_unlock(&block_group->lock);
10832
10833         if (remove_em) {
10834                 struct extent_map_tree *em_tree;
10835
10836                 em_tree = &fs_info->mapping_tree.map_tree;
10837                 write_lock(&em_tree->lock);
10838                 remove_extent_mapping(em_tree, em);
10839                 write_unlock(&em_tree->lock);
10840                 /* once for the tree */
10841                 free_extent_map(em);
10842         }
10843
10844         mutex_unlock(&fs_info->chunk_mutex);
10845
10846         ret = remove_block_group_free_space(trans, block_group);
10847         if (ret)
10848                 goto out;
10849
10850         btrfs_put_block_group(block_group);
10851         btrfs_put_block_group(block_group);
10852
10853         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
10854         if (ret > 0)
10855                 ret = -EIO;
10856         if (ret < 0)
10857                 goto out;
10858
10859         ret = btrfs_del_item(trans, root, path);
10860 out:
10861         if (remove_rsv)
10862                 btrfs_delayed_refs_rsv_release(fs_info, 1);
10863         btrfs_free_path(path);
10864         return ret;
10865 }
10866
10867 struct btrfs_trans_handle *
10868 btrfs_start_trans_remove_block_group(struct btrfs_fs_info *fs_info,
10869                                      const u64 chunk_offset)
10870 {
10871         struct extent_map_tree *em_tree = &fs_info->mapping_tree.map_tree;
10872         struct extent_map *em;
10873         struct map_lookup *map;
10874         unsigned int num_items;
10875
10876         read_lock(&em_tree->lock);
10877         em = lookup_extent_mapping(em_tree, chunk_offset, 1);
10878         read_unlock(&em_tree->lock);
10879         ASSERT(em && em->start == chunk_offset);
10880
10881         /*
10882          * We need to reserve 3 + N units from the metadata space info in order
10883          * to remove a block group (done at btrfs_remove_chunk() and at
10884          * btrfs_remove_block_group()), which are used for:
10885          *
10886          * 1 unit for adding the free space inode's orphan (located in the tree
10887          * of tree roots).
10888          * 1 unit for deleting the block group item (located in the extent
10889          * tree).
10890          * 1 unit for deleting the free space item (located in tree of tree
10891          * roots).
10892          * N units for deleting N device extent items corresponding to each
10893          * stripe (located in the device tree).
10894          *
10895          * In order to remove a block group we also need to reserve units in the
10896          * system space info in order to update the chunk tree (update one or
10897          * more device items and remove one chunk item), but this is done at
10898          * btrfs_remove_chunk() through a call to check_system_chunk().
10899          */
10900         map = em->map_lookup;
10901         num_items = 3 + map->num_stripes;
10902         free_extent_map(em);
10903
10904         return btrfs_start_transaction_fallback_global_rsv(fs_info->extent_root,
10905                                                            num_items, 1);
10906 }
10907
10908 /*
10909  * Process the unused_bgs list and remove any that don't have any allocated
10910  * space inside of them.
10911  */
10912 void btrfs_delete_unused_bgs(struct btrfs_fs_info *fs_info)
10913 {
10914         struct btrfs_block_group_cache *block_group;
10915         struct btrfs_space_info *space_info;
10916         struct btrfs_trans_handle *trans;
10917         int ret = 0;
10918
10919         if (!test_bit(BTRFS_FS_OPEN, &fs_info->flags))
10920                 return;
10921
10922         spin_lock(&fs_info->unused_bgs_lock);
10923         while (!list_empty(&fs_info->unused_bgs)) {
10924                 u64 start, end;
10925                 int trimming;
10926
10927                 block_group = list_first_entry(&fs_info->unused_bgs,
10928                                                struct btrfs_block_group_cache,
10929                                                bg_list);
10930                 list_del_init(&block_group->bg_list);
10931
10932                 space_info = block_group->space_info;
10933
10934                 if (ret || btrfs_mixed_space_info(space_info)) {
10935                         btrfs_put_block_group(block_group);
10936                         continue;
10937                 }
10938                 spin_unlock(&fs_info->unused_bgs_lock);
10939
10940                 mutex_lock(&fs_info->delete_unused_bgs_mutex);
10941
10942                 /* Don't want to race with allocators so take the groups_sem */
10943                 down_write(&space_info->groups_sem);
10944                 spin_lock(&block_group->lock);
10945                 if (block_group->reserved || block_group->pinned ||
10946                     btrfs_block_group_used(&block_group->item) ||
10947                     block_group->ro ||
10948                     list_is_singular(&block_group->list)) {
10949                         /*
10950                          * We want to bail if we made new allocations or have
10951                          * outstanding allocations in this block group.  We do
10952                          * the ro check in case balance is currently acting on
10953                          * this block group.
10954                          */
10955                         trace_btrfs_skip_unused_block_group(block_group);
10956                         spin_unlock(&block_group->lock);
10957                         up_write(&space_info->groups_sem);
10958                         goto next;
10959                 }
10960                 spin_unlock(&block_group->lock);
10961
10962                 /* We don't want to force the issue, only flip if it's ok. */
10963                 ret = inc_block_group_ro(block_group, 0);
10964                 up_write(&space_info->groups_sem);
10965                 if (ret < 0) {
10966                         ret = 0;
10967                         goto next;
10968                 }
10969
10970                 /*
10971                  * Want to do this before we do anything else so we can recover
10972                  * properly if we fail to join the transaction.
10973                  */
10974                 trans = btrfs_start_trans_remove_block_group(fs_info,
10975                                                      block_group->key.objectid);
10976                 if (IS_ERR(trans)) {
10977                         btrfs_dec_block_group_ro(block_group);
10978                         ret = PTR_ERR(trans);
10979                         goto next;
10980                 }
10981
10982                 /*
10983                  * We could have pending pinned extents for this block group,
10984                  * just delete them, we don't care about them anymore.
10985                  */
10986                 start = block_group->key.objectid;
10987                 end = start + block_group->key.offset - 1;
10988                 /*
10989                  * Hold the unused_bg_unpin_mutex lock to avoid racing with
10990                  * btrfs_finish_extent_commit(). If we are at transaction N,
10991                  * another task might be running finish_extent_commit() for the
10992                  * previous transaction N - 1, and have seen a range belonging
10993                  * to the block group in freed_extents[] before we were able to
10994                  * clear the whole block group range from freed_extents[]. This
10995                  * means that task can lookup for the block group after we
10996                  * unpinned it from freed_extents[] and removed it, leading to
10997                  * a BUG_ON() at btrfs_unpin_extent_range().
10998                  */
10999                 mutex_lock(&fs_info->unused_bg_unpin_mutex);
11000                 ret = clear_extent_bits(&fs_info->freed_extents[0], start, end,
11001                                   EXTENT_DIRTY);
11002                 if (ret) {
11003                         mutex_unlock(&fs_info->unused_bg_unpin_mutex);
11004                         btrfs_dec_block_group_ro(block_group);
11005                         goto end_trans;
11006                 }
11007                 ret = clear_extent_bits(&fs_info->freed_extents[1], start, end,
11008                                   EXTENT_DIRTY);
11009                 if (ret) {
11010                         mutex_unlock(&fs_info->unused_bg_unpin_mutex);
11011                         btrfs_dec_block_group_ro(block_group);
11012                         goto end_trans;
11013                 }
11014                 mutex_unlock(&fs_info->unused_bg_unpin_mutex);
11015
11016                 /* Reset pinned so btrfs_put_block_group doesn't complain */
11017                 spin_lock(&space_info->lock);
11018                 spin_lock(&block_group->lock);
11019
11020                 update_bytes_pinned(space_info, -block_group->pinned);
11021                 space_info->bytes_readonly += block_group->pinned;
11022                 percpu_counter_add_batch(&space_info->total_bytes_pinned,
11023                                    -block_group->pinned,
11024                                    BTRFS_TOTAL_BYTES_PINNED_BATCH);
11025                 block_group->pinned = 0;
11026
11027                 spin_unlock(&block_group->lock);
11028                 spin_unlock(&space_info->lock);
11029
11030                 /* DISCARD can flip during remount */
11031                 trimming = btrfs_test_opt(fs_info, DISCARD);
11032
11033                 /* Implicit trim during transaction commit. */
11034                 if (trimming)
11035                         btrfs_get_block_group_trimming(block_group);
11036
11037                 /*
11038                  * Btrfs_remove_chunk will abort the transaction if things go
11039                  * horribly wrong.
11040                  */
11041                 ret = btrfs_remove_chunk(trans, block_group->key.objectid);
11042
11043                 if (ret) {
11044                         if (trimming)
11045                                 btrfs_put_block_group_trimming(block_group);
11046                         goto end_trans;
11047                 }
11048
11049                 /*
11050                  * If we're not mounted with -odiscard, we can just forget
11051                  * about this block group. Otherwise we'll need to wait
11052                  * until transaction commit to do the actual discard.
11053                  */
11054                 if (trimming) {
11055                         spin_lock(&fs_info->unused_bgs_lock);
11056                         /*
11057                          * A concurrent scrub might have added us to the list
11058                          * fs_info->unused_bgs, so use a list_move operation
11059                          * to add the block group to the deleted_bgs list.
11060                          */
11061                         list_move(&block_group->bg_list,
11062                                   &trans->transaction->deleted_bgs);
11063                         spin_unlock(&fs_info->unused_bgs_lock);
11064                         btrfs_get_block_group(block_group);
11065                 }
11066 end_trans:
11067                 btrfs_end_transaction(trans);
11068 next:
11069                 mutex_unlock(&fs_info->delete_unused_bgs_mutex);
11070                 btrfs_put_block_group(block_group);
11071                 spin_lock(&fs_info->unused_bgs_lock);
11072         }
11073         spin_unlock(&fs_info->unused_bgs_lock);
11074 }
11075
11076 int btrfs_init_space_info(struct btrfs_fs_info *fs_info)
11077 {
11078         struct btrfs_super_block *disk_super;
11079         u64 features;
11080         u64 flags;
11081         int mixed = 0;
11082         int ret;
11083
11084         disk_super = fs_info->super_copy;
11085         if (!btrfs_super_root(disk_super))
11086                 return -EINVAL;
11087
11088         features = btrfs_super_incompat_flags(disk_super);
11089         if (features & BTRFS_FEATURE_INCOMPAT_MIXED_GROUPS)
11090                 mixed = 1;
11091
11092         flags = BTRFS_BLOCK_GROUP_SYSTEM;
11093         ret = create_space_info(fs_info, flags);
11094         if (ret)
11095                 goto out;
11096
11097         if (mixed) {
11098                 flags = BTRFS_BLOCK_GROUP_METADATA | BTRFS_BLOCK_GROUP_DATA;
11099                 ret = create_space_info(fs_info, flags);
11100         } else {
11101                 flags = BTRFS_BLOCK_GROUP_METADATA;
11102                 ret = create_space_info(fs_info, flags);
11103                 if (ret)
11104                         goto out;
11105
11106                 flags = BTRFS_BLOCK_GROUP_DATA;
11107                 ret = create_space_info(fs_info, flags);
11108         }
11109 out:
11110         return ret;
11111 }
11112
11113 int btrfs_error_unpin_extent_range(struct btrfs_fs_info *fs_info,
11114                                    u64 start, u64 end)
11115 {
11116         return unpin_extent_range(fs_info, start, end, false);
11117 }
11118
11119 /*
11120  * It used to be that old block groups would be left around forever.
11121  * Iterating over them would be enough to trim unused space.  Since we
11122  * now automatically remove them, we also need to iterate over unallocated
11123  * space.
11124  *
11125  * We don't want a transaction for this since the discard may take a
11126  * substantial amount of time.  We don't require that a transaction be
11127  * running, but we do need to take a running transaction into account
11128  * to ensure that we're not discarding chunks that were released or
11129  * allocated in the current transaction.
11130  *
11131  * Holding the chunks lock will prevent other threads from allocating
11132  * or releasing chunks, but it won't prevent a running transaction
11133  * from committing and releasing the memory that the pending chunks
11134  * list head uses.  For that, we need to take a reference to the
11135  * transaction and hold the commit root sem.  We only need to hold
11136  * it while performing the free space search since we have already
11137  * held back allocations.
11138  */
11139 static int btrfs_trim_free_extents(struct btrfs_device *device,
11140                                    struct fstrim_range *range, u64 *trimmed)
11141 {
11142         u64 start, len = 0, end = 0;
11143         int ret;
11144
11145         start = max_t(u64, range->start, SZ_1M);
11146         *trimmed = 0;
11147
11148         /* Discard not supported = nothing to do. */
11149         if (!blk_queue_discard(bdev_get_queue(device->bdev)))
11150                 return 0;
11151
11152         /* Not writable = nothing to do. */
11153         if (!test_bit(BTRFS_DEV_STATE_WRITEABLE, &device->dev_state))
11154                 return 0;
11155
11156         /* No free space = nothing to do. */
11157         if (device->total_bytes <= device->bytes_used)
11158                 return 0;
11159
11160         ret = 0;
11161
11162         while (1) {
11163                 struct btrfs_fs_info *fs_info = device->fs_info;
11164                 u64 bytes;
11165
11166                 ret = mutex_lock_interruptible(&fs_info->chunk_mutex);
11167                 if (ret)
11168                         break;
11169
11170                 find_first_clear_extent_bit(&device->alloc_state, start,
11171                                             &start, &end,
11172                                             CHUNK_TRIMMED | CHUNK_ALLOCATED);
11173                 /*
11174                  * If find_first_clear_extent_bit find a range that spans the
11175                  * end of the device it will set end to -1, in this case it's up
11176                  * to the caller to trim the value to the size of the device.
11177                  */
11178                 end = min(end, device->total_bytes - 1);
11179                 len = end - start + 1;
11180
11181                 /* We didn't find any extents */
11182                 if (!len) {
11183                         mutex_unlock(&fs_info->chunk_mutex);
11184                         ret = 0;
11185                         break;
11186                 }
11187
11188                 /* Keep going until we satisfy minlen or reach end of space */
11189                 if (len < range->minlen) {
11190                         mutex_unlock(&fs_info->chunk_mutex);
11191                         start += len;
11192                         continue;
11193                 }
11194
11195                 /* If we are out of the passed range break */
11196                 if (start > range->start + range->len - 1) {
11197                         mutex_unlock(&fs_info->chunk_mutex);
11198                         break;
11199                 }
11200
11201                 start = max(range->start, start);
11202                 len = min(range->len, len);
11203
11204                 ret = btrfs_issue_discard(device->bdev, start, len,
11205                                           &bytes);
11206                 if (!ret)
11207                         set_extent_bits(&device->alloc_state, start,
11208                                         start + bytes - 1,
11209                                         CHUNK_TRIMMED);
11210                 mutex_unlock(&fs_info->chunk_mutex);
11211
11212                 if (ret)
11213                         break;
11214
11215                 start += len;
11216                 *trimmed += bytes;
11217
11218                 /* We've trimmed enough */
11219                 if (*trimmed >= range->len)
11220                         break;
11221
11222                 if (fatal_signal_pending(current)) {
11223                         ret = -ERESTARTSYS;
11224                         break;
11225                 }
11226
11227                 cond_resched();
11228         }
11229
11230         return ret;
11231 }
11232
11233 /*
11234  * Trim the whole filesystem by:
11235  * 1) trimming the free space in each block group
11236  * 2) trimming the unallocated space on each device
11237  *
11238  * This will also continue trimming even if a block group or device encounters
11239  * an error.  The return value will be the last error, or 0 if nothing bad
11240  * happens.
11241  */
11242 int btrfs_trim_fs(struct btrfs_fs_info *fs_info, struct fstrim_range *range)
11243 {
11244         struct btrfs_block_group_cache *cache = NULL;
11245         struct btrfs_device *device;
11246         struct list_head *devices;
11247         u64 group_trimmed;
11248         u64 start;
11249         u64 end;
11250         u64 trimmed = 0;
11251         u64 bg_failed = 0;
11252         u64 dev_failed = 0;
11253         int bg_ret = 0;
11254         int dev_ret = 0;
11255         int ret = 0;
11256
11257         cache = btrfs_lookup_first_block_group(fs_info, range->start);
11258         for (; cache; cache = next_block_group(cache)) {
11259                 if (cache->key.objectid >= (range->start + range->len)) {
11260                         btrfs_put_block_group(cache);
11261                         break;
11262                 }
11263
11264                 start = max(range->start, cache->key.objectid);
11265                 end = min(range->start + range->len,
11266                                 cache->key.objectid + cache->key.offset);
11267
11268                 if (end - start >= range->minlen) {
11269                         if (!block_group_cache_done(cache)) {
11270                                 ret = cache_block_group(cache, 0);
11271                                 if (ret) {
11272                                         bg_failed++;
11273                                         bg_ret = ret;
11274                                         continue;
11275                                 }
11276                                 ret = wait_block_group_cache_done(cache);
11277                                 if (ret) {
11278                                         bg_failed++;
11279                                         bg_ret = ret;
11280                                         continue;
11281                                 }
11282                         }
11283                         ret = btrfs_trim_block_group(cache,
11284                                                      &group_trimmed,
11285                                                      start,
11286                                                      end,
11287                                                      range->minlen);
11288
11289                         trimmed += group_trimmed;
11290                         if (ret) {
11291                                 bg_failed++;
11292                                 bg_ret = ret;
11293                                 continue;
11294                         }
11295                 }
11296         }
11297
11298         if (bg_failed)
11299                 btrfs_warn(fs_info,
11300                         "failed to trim %llu block group(s), last error %d",
11301                         bg_failed, bg_ret);
11302         mutex_lock(&fs_info->fs_devices->device_list_mutex);
11303         devices = &fs_info->fs_devices->devices;
11304         list_for_each_entry(device, devices, dev_list) {
11305                 ret = btrfs_trim_free_extents(device, range, &group_trimmed);
11306                 if (ret) {
11307                         dev_failed++;
11308                         dev_ret = ret;
11309                         break;
11310                 }
11311
11312                 trimmed += group_trimmed;
11313         }
11314         mutex_unlock(&fs_info->fs_devices->device_list_mutex);
11315
11316         if (dev_failed)
11317                 btrfs_warn(fs_info,
11318                         "failed to trim %llu device(s), last error %d",
11319                         dev_failed, dev_ret);
11320         range->len = trimmed;
11321         if (bg_ret)
11322                 return bg_ret;
11323         return dev_ret;
11324 }
11325
11326 /*
11327  * btrfs_{start,end}_write_no_snapshotting() are similar to
11328  * mnt_{want,drop}_write(), they are used to prevent some tasks from writing
11329  * data into the page cache through nocow before the subvolume is snapshoted,
11330  * but flush the data into disk after the snapshot creation, or to prevent
11331  * operations while snapshotting is ongoing and that cause the snapshot to be
11332  * inconsistent (writes followed by expanding truncates for example).
11333  */
11334 void btrfs_end_write_no_snapshotting(struct btrfs_root *root)
11335 {
11336         percpu_counter_dec(&root->subv_writers->counter);
11337         cond_wake_up(&root->subv_writers->wait);
11338 }
11339
11340 int btrfs_start_write_no_snapshotting(struct btrfs_root *root)
11341 {
11342         if (atomic_read(&root->will_be_snapshotted))
11343                 return 0;
11344
11345         percpu_counter_inc(&root->subv_writers->counter);
11346         /*
11347          * Make sure counter is updated before we check for snapshot creation.
11348          */
11349         smp_mb();
11350         if (atomic_read(&root->will_be_snapshotted)) {
11351                 btrfs_end_write_no_snapshotting(root);
11352                 return 0;
11353         }
11354         return 1;
11355 }
11356
11357 void btrfs_wait_for_snapshot_creation(struct btrfs_root *root)
11358 {
11359         while (true) {
11360                 int ret;
11361
11362                 ret = btrfs_start_write_no_snapshotting(root);
11363                 if (ret)
11364                         break;
11365                 wait_var_event(&root->will_be_snapshotted,
11366                                !atomic_read(&root->will_be_snapshotted));
11367         }
11368 }
11369
11370 void btrfs_mark_bg_unused(struct btrfs_block_group_cache *bg)
11371 {
11372         struct btrfs_fs_info *fs_info = bg->fs_info;
11373
11374         spin_lock(&fs_info->unused_bgs_lock);
11375         if (list_empty(&bg->bg_list)) {
11376                 btrfs_get_block_group(bg);
11377                 trace_btrfs_add_unused_block_group(bg);
11378                 list_add_tail(&bg->bg_list, &fs_info->unused_bgs);
11379         }
11380         spin_unlock(&fs_info->unused_bgs_lock);
11381 }