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(fs_info, 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, s64 num_bytes,
 760                              bool metadata, u64 root_objectid)
 761 {
 762         struct btrfs_space_info *space_info;
 763         u64 flags;
 764
 765         if (metadata) {
 766                 if (root_objectid == BTRFS_CHUNK_TREE_OBJECTID)
 767                         flags = BTRFS_BLOCK_GROUP_SYSTEM;
 768                 else
 769                         flags = BTRFS_BLOCK_GROUP_METADATA;
 770         } else {
 771                 flags = BTRFS_BLOCK_GROUP_DATA;
 772         }
 773
 774         space_info = __find_space_info(fs_info, flags);
 775         ASSERT(space_info);
 776         percpu_counter_add_batch(&space_info->total_bytes_pinned, num_bytes,
 777                     BTRFS_TOTAL_BYTES_PINNED_BATCH);
 778 }
 779
 780 /*
 781  * after adding space to the filesystem, we need to clear the full flags
 782  * on all the space infos.
 783  */
 784 void btrfs_clear_space_info_full(struct btrfs_fs_info *info)
 785 {
 786         struct list_head *head = &info->space_info;
 787         struct btrfs_space_info *found;
 788
 789         rcu_read_lock();
 790         list_for_each_entry_rcu(found, head, list)
 791                 found->full = 0;
 792         rcu_read_unlock();
 793 }
 794
 795 /* simple helper to search for an existing data extent at a given offset */
 796 int btrfs_lookup_data_extent(struct btrfs_fs_info *fs_info, u64 start, u64 len)
 797 {
 798         int ret;
 799         struct btrfs_key key;
 800         struct btrfs_path *path;
 801
 802         path = btrfs_alloc_path();
 803         if (!path)
 804                 return -ENOMEM;
 805
 806         key.objectid = start;
 807         key.offset = len;
 808         key.type = BTRFS_EXTENT_ITEM_KEY;
 809         ret = btrfs_search_slot(NULL, fs_info->extent_root, &key, path, 0, 0);
 810         btrfs_free_path(path);
 811         return ret;
 812 }
 813
 814 /*
 815  * helper function to lookup reference count and flags of a tree block.
 816  *
 817  * the head node for delayed ref is used to store the sum of all the
 818  * reference count modifications queued up in the rbtree. the head
 819  * node may also store the extent flags to set. This way you can check
 820  * to see what the reference count and extent flags would be if all of
 821  * the delayed refs are not processed.
 822  */
 823 int btrfs_lookup_extent_info(struct btrfs_trans_handle *trans,
 824                              struct btrfs_fs_info *fs_info, u64 bytenr,
 825                              u64 offset, int metadata, u64 *refs, u64 *flags)
 826 {
 827         struct btrfs_delayed_ref_head *head;
 828         struct btrfs_delayed_ref_root *delayed_refs;
 829         struct btrfs_path *path;
 830         struct btrfs_extent_item *ei;
 831         struct extent_buffer *leaf;
 832         struct btrfs_key key;
 833         u32 item_size;
 834         u64 num_refs;
 835         u64 extent_flags;
 836         int ret;
 837
 838         /*
 839          * If we don't have skinny metadata, don't bother doing anything
 840          * different
 841          */
 842         if (metadata && !btrfs_fs_incompat(fs_info, SKINNY_METADATA)) {
 843                 offset = fs_info->nodesize;
 844                 metadata = 0;
 845         }
 846
 847         path = btrfs_alloc_path();
 848         if (!path)
 849                 return -ENOMEM;
 850
 851         if (!trans) {
 852                 path->skip_locking = 1;
 853                 path->search_commit_root = 1;
 854         }
 855
 856 search_again:
 857         key.objectid = bytenr;
 858         key.offset = offset;
 859         if (metadata)
 860                 key.type = BTRFS_METADATA_ITEM_KEY;
 861         else
 862                 key.type = BTRFS_EXTENT_ITEM_KEY;
 863
 864         ret = btrfs_search_slot(trans, fs_info->extent_root, &key, path, 0, 0);
 865         if (ret < 0)
 866                 goto out_free;
 867
 868         if (ret > 0 && metadata && key.type == BTRFS_METADATA_ITEM_KEY) {
 869                 if (path->slots[0]) {
 870                         path->slots[0]--;
 871                         btrfs_item_key_to_cpu(path->nodes[0], &key,
 872                                               path->slots[0]);
 873                         if (key.objectid == bytenr &&
 874                             key.type == BTRFS_EXTENT_ITEM_KEY &&
 875                             key.offset == fs_info->nodesize)
 876                                 ret = 0;
 877                 }
 878         }
 879
 880         if (ret == 0) {
 881                 leaf = path->nodes[0];
 882                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
 883                 if (item_size >= sizeof(*ei)) {
 884                         ei = btrfs_item_ptr(leaf, path->slots[0],
 885                                             struct btrfs_extent_item);
 886                         num_refs = btrfs_extent_refs(leaf, ei);
 887                         extent_flags = btrfs_extent_flags(leaf, ei);
 888                 } else {
 889                         ret = -EINVAL;
 890                         btrfs_print_v0_err(fs_info);
 891                         if (trans)
 892                                 btrfs_abort_transaction(trans, ret);
 893                         else
 894                                 btrfs_handle_fs_error(fs_info, ret, NULL);
 895
 896                         goto out_free;
 897                 }
 898
 899                 BUG_ON(num_refs == 0);
 900         } else {
 901                 num_refs = 0;
 902                 extent_flags = 0;
 903                 ret = 0;
 904         }
 905
 906         if (!trans)
 907                 goto out;
 908
 909         delayed_refs = &trans->transaction->delayed_refs;
 910         spin_lock(&delayed_refs->lock);
 911         head = btrfs_find_delayed_ref_head(delayed_refs, bytenr);
 912         if (head) {
 913                 if (!mutex_trylock(&head->mutex)) {
 914                         refcount_inc(&head->refs);
 915                         spin_unlock(&delayed_refs->lock);
 916
 917                         btrfs_release_path(path);
 918
 919                         /*
 920                          * Mutex was contended, block until it's released and try
 921                          * again
 922                          */
 923                         mutex_lock(&head->mutex);
 924                         mutex_unlock(&head->mutex);
 925                         btrfs_put_delayed_ref_head(head);
 926                         goto search_again;
 927                 }
 928                 spin_lock(&head->lock);
 929                 if (head->extent_op && head->extent_op->update_flags)
 930                         extent_flags |= head->extent_op->flags_to_set;
 931                 else
 932                         BUG_ON(num_refs == 0);
 933
 934                 num_refs += head->ref_mod;
 935                 spin_unlock(&head->lock);
 936                 mutex_unlock(&head->mutex);
 937         }
 938         spin_unlock(&delayed_refs->lock);
 939 out:
 940         WARN_ON(num_refs == 0);
 941         if (refs)
 942                 *refs = num_refs;
 943         if (flags)
 944                 *flags = extent_flags;
 945 out_free:
 946         btrfs_free_path(path);
 947         return ret;
 948 }
 949
 950 /*
 951  * Back reference rules.  Back refs have three main goals:
 952  *
 953  * 1) differentiate between all holders of references to an extent so that
 954  *    when a reference is dropped we can make sure it was a valid reference
 955  *    before freeing the extent.
 956  *
 957  * 2) Provide enough information to quickly find the holders of an extent
 958  *    if we notice a given block is corrupted or bad.
 959  *
 960  * 3) Make it easy to migrate blocks for FS shrinking or storage pool
 961  *    maintenance.  This is actually the same as #2, but with a slightly
 962  *    different use case.
 963  *
 964  * There are two kinds of back refs. The implicit back refs is optimized
 965  * for pointers in non-shared tree blocks. For a given pointer in a block,
 966  * back refs of this kind provide information about the block's owner tree
 967  * and the pointer's key. These information allow us to find the block by
 968  * b-tree searching. The full back refs is for pointers in tree blocks not
 969  * referenced by their owner trees. The location of tree block is recorded
 970  * in the back refs. Actually the full back refs is generic, and can be
 971  * used in all cases the implicit back refs is used. The major shortcoming
 972  * of the full back refs is its overhead. Every time a tree block gets
 973  * COWed, we have to update back refs entry for all pointers in it.
 974  *
 975  * For a newly allocated tree block, we use implicit back refs for
 976  * pointers in it. This means most tree related operations only involve
 977  * implicit back refs. For a tree block created in old transaction, the
 978  * only way to drop a reference to it is COW it. So we can detect the
 979  * event that tree block loses its owner tree's reference and do the
 980  * back refs conversion.
 981  *
 982  * When a tree block is COWed through a tree, there are four cases:
 983  *
 984  * The reference count of the block is one and the tree is the block's
 985  * owner tree. Nothing to do in this case.
 986  *
 987  * The reference count of the block is one and the tree is not the
 988  * block's owner tree. In this case, full back refs is used for pointers
 989  * in the block. Remove these full back refs, add implicit back refs for
 990  * every pointers in the new block.
 991  *
 992  * The reference count of the block is greater than one and the tree is
 993  * the block's owner tree. In this case, implicit back refs is used for
 994  * pointers in the block. Add full back refs for every pointers in the
 995  * block, increase lower level extents' reference counts. The original
 996  * implicit back refs are entailed to the new block.
 997  *
 998  * The reference count of the block is greater than one and the tree is
 999  * not the block's owner tree. Add implicit back refs for every pointer in
1000  * the new block, increase lower level extents' reference count.
1001  *
1002  * Back Reference Key composing:
1003  *
1004  * The key objectid corresponds to the first byte in the extent,
1005  * The key type is used to differentiate between types of back refs.
1006  * There are different meanings of the key offset for different types
1007  * of back refs.
1008  *
1009  * File extents can be referenced by:
1010  *
1011  * - multiple snapshots, subvolumes, or different generations in one subvol
1012  * - different files inside a single subvolume
1013  * - different offsets inside a file (bookend extents in file.c)
1014  *
1015  * The extent ref structure for the implicit back refs has fields for:
1016  *
1017  * - Objectid of the subvolume root
1018  * - objectid of the file holding the reference
1019  * - original offset in the file
1020  * - how many bookend extents
1021  *
1022  * The key offset for the implicit back refs is hash of the first
1023  * three fields.
1024  *
1025  * The extent ref structure for the full back refs has field for:
1026  *
1027  * - number of pointers in the tree leaf
1028  *
1029  * The key offset for the implicit back refs is the first byte of
1030  * the tree leaf
1031  *
1032  * When a file extent is allocated, The implicit back refs is used.
1033  * the fields are filled in:
1034  *
1035  *     (root_key.objectid, inode objectid, offset in file, 1)
1036  *
1037  * When a file extent is removed file truncation, we find the
1038  * corresponding implicit back refs and check the following fields:
1039  *
1040  *     (btrfs_header_owner(leaf), inode objectid, offset in file)
1041  *
1042  * Btree extents can be referenced by:
1043  *
1044  * - Different subvolumes
1045  *
1046  * Both the implicit back refs and the full back refs for tree blocks
1047  * only consist of key. The key offset for the implicit back refs is
1048  * objectid of block's owner tree. The key offset for the full back refs
1049  * is the first byte of parent block.
1050  *
1051  * When implicit back refs is used, information about the lowest key and
1052  * level of the tree block are required. These information are stored in
1053  * tree block info structure.
1054  */
1055
1056 /*
1057  * is_data == BTRFS_REF_TYPE_BLOCK, tree block type is required,
1058  * is_data == BTRFS_REF_TYPE_DATA, data type is requiried,
1059  * is_data == BTRFS_REF_TYPE_ANY, either type is OK.
1060  */
1061 int btrfs_get_extent_inline_ref_type(const struct extent_buffer *eb,
1062                                      struct btrfs_extent_inline_ref *iref,
1063                                      enum btrfs_inline_ref_type is_data)
1064 {
1065         int type = btrfs_extent_inline_ref_type(eb, iref);
1066         u64 offset = btrfs_extent_inline_ref_offset(eb, iref);
1067
1068         if (type == BTRFS_TREE_BLOCK_REF_KEY ||
1069             type == BTRFS_SHARED_BLOCK_REF_KEY ||
1070             type == BTRFS_SHARED_DATA_REF_KEY ||
1071             type == BTRFS_EXTENT_DATA_REF_KEY) {
1072                 if (is_data == BTRFS_REF_TYPE_BLOCK) {
1073                         if (type == BTRFS_TREE_BLOCK_REF_KEY)
1074                                 return type;
1075                         if (type == BTRFS_SHARED_BLOCK_REF_KEY) {
1076                                 ASSERT(eb->fs_info);
1077                                 /*
1078                                  * Every shared one has parent tree
1079                                  * block, which must be aligned to
1080                                  * nodesize.
1081                                  */
1082                                 if (offset &&
1083                                     IS_ALIGNED(offset, eb->fs_info->nodesize))
1084                                         return type;
1085                         }
1086                 } else if (is_data == BTRFS_REF_TYPE_DATA) {
1087                         if (type == BTRFS_EXTENT_DATA_REF_KEY)
1088                                 return type;
1089                         if (type == BTRFS_SHARED_DATA_REF_KEY) {
1090                                 ASSERT(eb->fs_info);
1091                                 /*
1092                                  * Every shared one has parent tree
1093                                  * block, which must be aligned to
1094                                  * nodesize.
1095                                  */
1096                                 if (offset &&
1097                                     IS_ALIGNED(offset, eb->fs_info->nodesize))
1098                                         return type;
1099                         }
1100                 } else {
1101                         ASSERT(is_data == BTRFS_REF_TYPE_ANY);
1102                         return type;
1103                 }
1104         }
1105
1106         btrfs_print_leaf((struct extent_buffer *)eb);
1107         btrfs_err(eb->fs_info, "eb %llu invalid extent inline ref type %d",
1108                   eb->start, type);
1109         WARN_ON(1);
1110
1111         return BTRFS_REF_TYPE_INVALID;
1112 }
1113
1114 static u64 hash_extent_data_ref(u64 root_objectid, u64 owner, u64 offset)
1115 {
1116         u32 high_crc = ~(u32)0;
1117         u32 low_crc = ~(u32)0;
1118         __le64 lenum;
1119
1120         lenum = cpu_to_le64(root_objectid);
1121         high_crc = crc32c(high_crc, &lenum, sizeof(lenum));
1122         lenum = cpu_to_le64(owner);
1123         low_crc = crc32c(low_crc, &lenum, sizeof(lenum));
1124         lenum = cpu_to_le64(offset);
1125         low_crc = crc32c(low_crc, &lenum, sizeof(lenum));
1126
1127         return ((u64)high_crc << 31) ^ (u64)low_crc;
1128 }
1129
1130 static u64 hash_extent_data_ref_item(struct extent_buffer *leaf,
1131                                      struct btrfs_extent_data_ref *ref)
1132 {
1133         return hash_extent_data_ref(btrfs_extent_data_ref_root(leaf, ref),
1134                                     btrfs_extent_data_ref_objectid(leaf, ref),
1135                                     btrfs_extent_data_ref_offset(leaf, ref));
1136 }
1137
1138 static int match_extent_data_ref(struct extent_buffer *leaf,
1139                                  struct btrfs_extent_data_ref *ref,
1140                                  u64 root_objectid, u64 owner, u64 offset)
1141 {
1142         if (btrfs_extent_data_ref_root(leaf, ref) != root_objectid ||
1143             btrfs_extent_data_ref_objectid(leaf, ref) != owner ||
1144             btrfs_extent_data_ref_offset(leaf, ref) != offset)
1145                 return 0;
1146         return 1;
1147 }
1148
1149 static noinline int lookup_extent_data_ref(struct btrfs_trans_handle *trans,
1150                                            struct btrfs_path *path,
1151                                            u64 bytenr, u64 parent,
1152                                            u64 root_objectid,
1153                                            u64 owner, u64 offset)
1154 {
1155         struct btrfs_root *root = trans->fs_info->extent_root;
1156         struct btrfs_key key;
1157         struct btrfs_extent_data_ref *ref;
1158         struct extent_buffer *leaf;
1159         u32 nritems;
1160         int ret;
1161         int recow;
1162         int err = -ENOENT;
1163
1164         key.objectid = bytenr;
1165         if (parent) {
1166                 key.type = BTRFS_SHARED_DATA_REF_KEY;
1167                 key.offset = parent;
1168         } else {
1169                 key.type = BTRFS_EXTENT_DATA_REF_KEY;
1170                 key.offset = hash_extent_data_ref(root_objectid,
1171                                                   owner, offset);
1172         }
1173 again:
1174         recow = 0;
1175         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1176         if (ret < 0) {
1177                 err = ret;
1178                 goto fail;
1179         }
1180
1181         if (parent) {
1182                 if (!ret)
1183                         return 0;
1184                 goto fail;
1185         }
1186
1187         leaf = path->nodes[0];
1188         nritems = btrfs_header_nritems(leaf);
1189         while (1) {
1190                 if (path->slots[0] >= nritems) {
1191                         ret = btrfs_next_leaf(root, path);
1192                         if (ret < 0)
1193                                 err = ret;
1194                         if (ret)
1195                                 goto fail;
1196
1197                         leaf = path->nodes[0];
1198                         nritems = btrfs_header_nritems(leaf);
1199                         recow = 1;
1200                 }
1201
1202                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1203                 if (key.objectid != bytenr ||
1204                     key.type != BTRFS_EXTENT_DATA_REF_KEY)
1205                         goto fail;
1206
1207                 ref = btrfs_item_ptr(leaf, path->slots[0],
1208                                      struct btrfs_extent_data_ref);
1209
1210                 if (match_extent_data_ref(leaf, ref, root_objectid,
1211                                           owner, offset)) {
1212                         if (recow) {
1213                                 btrfs_release_path(path);
1214                                 goto again;
1215                         }
1216                         err = 0;
1217                         break;
1218                 }
1219                 path->slots[0]++;
1220         }
1221 fail:
1222         return err;
1223 }
1224
1225 static noinline int insert_extent_data_ref(struct btrfs_trans_handle *trans,
1226                                            struct btrfs_path *path,
1227                                            u64 bytenr, u64 parent,
1228                                            u64 root_objectid, u64 owner,
1229                                            u64 offset, int refs_to_add)
1230 {
1231         struct btrfs_root *root = trans->fs_info->extent_root;
1232         struct btrfs_key key;
1233         struct extent_buffer *leaf;
1234         u32 size;
1235         u32 num_refs;
1236         int ret;
1237
1238         key.objectid = bytenr;
1239         if (parent) {
1240                 key.type = BTRFS_SHARED_DATA_REF_KEY;
1241                 key.offset = parent;
1242                 size = sizeof(struct btrfs_shared_data_ref);
1243         } else {
1244                 key.type = BTRFS_EXTENT_DATA_REF_KEY;
1245                 key.offset = hash_extent_data_ref(root_objectid,
1246                                                   owner, offset);
1247                 size = sizeof(struct btrfs_extent_data_ref);
1248         }
1249
1250         ret = btrfs_insert_empty_item(trans, root, path, &key, size);
1251         if (ret && ret != -EEXIST)
1252                 goto fail;
1253
1254         leaf = path->nodes[0];
1255         if (parent) {
1256                 struct btrfs_shared_data_ref *ref;
1257                 ref = btrfs_item_ptr(leaf, path->slots[0],
1258                                      struct btrfs_shared_data_ref);
1259                 if (ret == 0) {
1260                         btrfs_set_shared_data_ref_count(leaf, ref, refs_to_add);
1261                 } else {
1262                         num_refs = btrfs_shared_data_ref_count(leaf, ref);
1263                         num_refs += refs_to_add;
1264                         btrfs_set_shared_data_ref_count(leaf, ref, num_refs);
1265                 }
1266         } else {
1267                 struct btrfs_extent_data_ref *ref;
1268                 while (ret == -EEXIST) {
1269                         ref = btrfs_item_ptr(leaf, path->slots[0],
1270                                              struct btrfs_extent_data_ref);
1271                         if (match_extent_data_ref(leaf, ref, root_objectid,
1272                                                   owner, offset))
1273                                 break;
1274                         btrfs_release_path(path);
1275                         key.offset++;
1276                         ret = btrfs_insert_empty_item(trans, root, path, &key,
1277                                                       size);
1278                         if (ret && ret != -EEXIST)
1279                                 goto fail;
1280
1281                         leaf = path->nodes[0];
1282                 }
1283                 ref = btrfs_item_ptr(leaf, path->slots[0],
1284                                      struct btrfs_extent_data_ref);
1285                 if (ret == 0) {
1286                         btrfs_set_extent_data_ref_root(leaf, ref,
1287                                                        root_objectid);
1288                         btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
1289                         btrfs_set_extent_data_ref_offset(leaf, ref, offset);
1290                         btrfs_set_extent_data_ref_count(leaf, ref, refs_to_add);
1291                 } else {
1292                         num_refs = btrfs_extent_data_ref_count(leaf, ref);
1293                         num_refs += refs_to_add;
1294                         btrfs_set_extent_data_ref_count(leaf, ref, num_refs);
1295                 }
1296         }
1297         btrfs_mark_buffer_dirty(leaf);
1298         ret = 0;
1299 fail:
1300         btrfs_release_path(path);
1301         return ret;
1302 }
1303
1304 static noinline int remove_extent_data_ref(struct btrfs_trans_handle *trans,
1305                                            struct btrfs_path *path,
1306                                            int refs_to_drop, int *last_ref)
1307 {
1308         struct btrfs_key key;
1309         struct btrfs_extent_data_ref *ref1 = NULL;
1310         struct btrfs_shared_data_ref *ref2 = NULL;
1311         struct extent_buffer *leaf;
1312         u32 num_refs = 0;
1313         int ret = 0;
1314
1315         leaf = path->nodes[0];
1316         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1317
1318         if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
1319                 ref1 = btrfs_item_ptr(leaf, path->slots[0],
1320                                       struct btrfs_extent_data_ref);
1321                 num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1322         } else if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
1323                 ref2 = btrfs_item_ptr(leaf, path->slots[0],
1324                                       struct btrfs_shared_data_ref);
1325                 num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1326         } else if (unlikely(key.type == BTRFS_EXTENT_REF_V0_KEY)) {
1327                 btrfs_print_v0_err(trans->fs_info);
1328                 btrfs_abort_transaction(trans, -EINVAL);
1329                 return -EINVAL;
1330         } else {
1331                 BUG();
1332         }
1333
1334         BUG_ON(num_refs < refs_to_drop);
1335         num_refs -= refs_to_drop;
1336
1337         if (num_refs == 0) {
1338                 ret = btrfs_del_item(trans, trans->fs_info->extent_root, path);
1339                 *last_ref = 1;
1340         } else {
1341                 if (key.type == BTRFS_EXTENT_DATA_REF_KEY)
1342                         btrfs_set_extent_data_ref_count(leaf, ref1, num_refs);
1343                 else if (key.type == BTRFS_SHARED_DATA_REF_KEY)
1344                         btrfs_set_shared_data_ref_count(leaf, ref2, num_refs);
1345                 btrfs_mark_buffer_dirty(leaf);
1346         }
1347         return ret;
1348 }
1349
1350 static noinline u32 extent_data_ref_count(struct btrfs_path *path,
1351                                           struct btrfs_extent_inline_ref *iref)
1352 {
1353         struct btrfs_key key;
1354         struct extent_buffer *leaf;
1355         struct btrfs_extent_data_ref *ref1;
1356         struct btrfs_shared_data_ref *ref2;
1357         u32 num_refs = 0;
1358         int type;
1359
1360         leaf = path->nodes[0];
1361         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1362
1363         BUG_ON(key.type == BTRFS_EXTENT_REF_V0_KEY);
1364         if (iref) {
1365                 /*
1366                  * If type is invalid, we should have bailed out earlier than
1367                  * this call.
1368                  */
1369                 type = btrfs_get_extent_inline_ref_type(leaf, iref, BTRFS_REF_TYPE_DATA);
1370                 ASSERT(type != BTRFS_REF_TYPE_INVALID);
1371                 if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1372                         ref1 = (struct btrfs_extent_data_ref *)(&iref->offset);
1373                         num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1374                 } else {
1375                         ref2 = (struct btrfs_shared_data_ref *)(iref + 1);
1376                         num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1377                 }
1378         } else if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
1379                 ref1 = btrfs_item_ptr(leaf, path->slots[0],
1380                                       struct btrfs_extent_data_ref);
1381                 num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1382         } else if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
1383                 ref2 = btrfs_item_ptr(leaf, path->slots[0],
1384                                       struct btrfs_shared_data_ref);
1385                 num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1386         } else {
1387                 WARN_ON(1);
1388         }
1389         return num_refs;
1390 }
1391
1392 static noinline int lookup_tree_block_ref(struct btrfs_trans_handle *trans,
1393                                           struct btrfs_path *path,
1394                                           u64 bytenr, u64 parent,
1395                                           u64 root_objectid)
1396 {
1397         struct btrfs_root *root = trans->fs_info->extent_root;
1398         struct btrfs_key key;
1399         int ret;
1400
1401         key.objectid = bytenr;
1402         if (parent) {
1403                 key.type = BTRFS_SHARED_BLOCK_REF_KEY;
1404                 key.offset = parent;
1405         } else {
1406                 key.type = BTRFS_TREE_BLOCK_REF_KEY;
1407                 key.offset = root_objectid;
1408         }
1409
1410         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1411         if (ret > 0)
1412                 ret = -ENOENT;
1413         return ret;
1414 }
1415
1416 static noinline int insert_tree_block_ref(struct btrfs_trans_handle *trans,
1417                                           struct btrfs_path *path,
1418                                           u64 bytenr, u64 parent,
1419                                           u64 root_objectid)
1420 {
1421         struct btrfs_key key;
1422         int ret;
1423
1424         key.objectid = bytenr;
1425         if (parent) {
1426                 key.type = BTRFS_SHARED_BLOCK_REF_KEY;
1427                 key.offset = parent;
1428         } else {
1429                 key.type = BTRFS_TREE_BLOCK_REF_KEY;
1430                 key.offset = root_objectid;
1431         }
1432
1433         ret = btrfs_insert_empty_item(trans, trans->fs_info->extent_root,
1434                                       path, &key, 0);
1435         btrfs_release_path(path);
1436         return ret;
1437 }
1438
1439 static inline int extent_ref_type(u64 parent, u64 owner)
1440 {
1441         int type;
1442         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1443                 if (parent > 0)
1444                         type = BTRFS_SHARED_BLOCK_REF_KEY;
1445                 else
1446                         type = BTRFS_TREE_BLOCK_REF_KEY;
1447         } else {
1448                 if (parent > 0)
1449                         type = BTRFS_SHARED_DATA_REF_KEY;
1450                 else
1451                         type = BTRFS_EXTENT_DATA_REF_KEY;
1452         }
1453         return type;
1454 }
1455
1456 static int find_next_key(struct btrfs_path *path, int level,
1457                          struct btrfs_key *key)
1458
1459 {
1460         for (; level < BTRFS_MAX_LEVEL; level++) {
1461                 if (!path->nodes[level])
1462                         break;
1463                 if (path->slots[level] + 1 >=
1464                     btrfs_header_nritems(path->nodes[level]))
1465                         continue;
1466                 if (level == 0)
1467                         btrfs_item_key_to_cpu(path->nodes[level], key,
1468                                               path->slots[level] + 1);
1469                 else
1470                         btrfs_node_key_to_cpu(path->nodes[level], key,
1471                                               path->slots[level] + 1);
1472                 return 0;
1473         }
1474         return 1;
1475 }
1476
1477 /*
1478  * look for inline back ref. if back ref is found, *ref_ret is set
1479  * to the address of inline back ref, and 0 is returned.
1480  *
1481  * if back ref isn't found, *ref_ret is set to the address where it
1482  * should be inserted, and -ENOENT is returned.
1483  *
1484  * if insert is true and there are too many inline back refs, the path
1485  * points to the extent item, and -EAGAIN is returned.
1486  *
1487  * NOTE: inline back refs are ordered in the same way that back ref
1488  *       items in the tree are ordered.
1489  */
1490 static noinline_for_stack
1491 int lookup_inline_extent_backref(struct btrfs_trans_handle *trans,
1492                                  struct btrfs_path *path,
1493                                  struct btrfs_extent_inline_ref **ref_ret,
1494                                  u64 bytenr, u64 num_bytes,
1495                                  u64 parent, u64 root_objectid,
1496                                  u64 owner, u64 offset, int insert)
1497 {
1498         struct btrfs_fs_info *fs_info = trans->fs_info;
1499         struct btrfs_root *root = fs_info->extent_root;
1500         struct btrfs_key key;
1501         struct extent_buffer *leaf;
1502         struct btrfs_extent_item *ei;
1503         struct btrfs_extent_inline_ref *iref;
1504         u64 flags;
1505         u64 item_size;
1506         unsigned long ptr;
1507         unsigned long end;
1508         int extra_size;
1509         int type;
1510         int want;
1511         int ret;
1512         int err = 0;
1513         bool skinny_metadata = btrfs_fs_incompat(fs_info, SKINNY_METADATA);
1514         int needed;
1515
1516         key.objectid = bytenr;
1517         key.type = BTRFS_EXTENT_ITEM_KEY;
1518         key.offset = num_bytes;
1519
1520         want = extent_ref_type(parent, owner);
1521         if (insert) {
1522                 extra_size = btrfs_extent_inline_ref_size(want);
1523                 path->keep_locks = 1;
1524         } else
1525                 extra_size = -1;
1526
1527         /*
1528          * Owner is our level, so we can just add one to get the level for the
1529          * block we are interested in.
1530          */
1531         if (skinny_metadata && owner < BTRFS_FIRST_FREE_OBJECTID) {
1532                 key.type = BTRFS_METADATA_ITEM_KEY;
1533                 key.offset = owner;
1534         }
1535
1536 again:
1537         ret = btrfs_search_slot(trans, root, &key, path, extra_size, 1);
1538         if (ret < 0) {
1539                 err = ret;
1540                 goto out;
1541         }
1542
1543         /*
1544          * We may be a newly converted file system which still has the old fat
1545          * extent entries for metadata, so try and see if we have one of those.
1546          */
1547         if (ret > 0 && skinny_metadata) {
1548                 skinny_metadata = false;
1549                 if (path->slots[0]) {
1550                         path->slots[0]--;
1551                         btrfs_item_key_to_cpu(path->nodes[0], &key,
1552                                               path->slots[0]);
1553                         if (key.objectid == bytenr &&
1554                             key.type == BTRFS_EXTENT_ITEM_KEY &&
1555                             key.offset == num_bytes)
1556                                 ret = 0;
1557                 }
1558                 if (ret) {
1559                         key.objectid = bytenr;
1560                         key.type = BTRFS_EXTENT_ITEM_KEY;
1561                         key.offset = num_bytes;
1562                         btrfs_release_path(path);
1563                         goto again;
1564                 }
1565         }
1566
1567         if (ret && !insert) {
1568                 err = -ENOENT;
1569                 goto out;
1570         } else if (WARN_ON(ret)) {
1571                 err = -EIO;
1572                 goto out;
1573         }
1574
1575         leaf = path->nodes[0];
1576         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1577         if (unlikely(item_size < sizeof(*ei))) {
1578                 err = -EINVAL;
1579                 btrfs_print_v0_err(fs_info);
1580                 btrfs_abort_transaction(trans, err);
1581                 goto out;
1582         }
1583
1584         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1585         flags = btrfs_extent_flags(leaf, ei);
1586
1587         ptr = (unsigned long)(ei + 1);
1588         end = (unsigned long)ei + item_size;
1589
1590         if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK && !skinny_metadata) {
1591                 ptr += sizeof(struct btrfs_tree_block_info);
1592                 BUG_ON(ptr > end);
1593         }
1594
1595         if (owner >= BTRFS_FIRST_FREE_OBJECTID)
1596                 needed = BTRFS_REF_TYPE_DATA;
1597         else
1598                 needed = BTRFS_REF_TYPE_BLOCK;
1599
1600         err = -ENOENT;
1601         while (1) {
1602                 if (ptr >= end) {
1603                         WARN_ON(ptr > end);
1604                         break;
1605                 }
1606                 iref = (struct btrfs_extent_inline_ref *)ptr;
1607                 type = btrfs_get_extent_inline_ref_type(leaf, iref, needed);
1608                 if (type == BTRFS_REF_TYPE_INVALID) {
1609                         err = -EUCLEAN;
1610                         goto out;
1611                 }
1612
1613                 if (want < type)
1614                         break;
1615                 if (want > type) {
1616                         ptr += btrfs_extent_inline_ref_size(type);
1617                         continue;
1618                 }
1619
1620                 if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1621                         struct btrfs_extent_data_ref *dref;
1622                         dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1623                         if (match_extent_data_ref(leaf, dref, root_objectid,
1624                                                   owner, offset)) {
1625                                 err = 0;
1626                                 break;
1627                         }
1628                         if (hash_extent_data_ref_item(leaf, dref) <
1629                             hash_extent_data_ref(root_objectid, owner, offset))
1630                                 break;
1631                 } else {
1632                         u64 ref_offset;
1633                         ref_offset = btrfs_extent_inline_ref_offset(leaf, iref);
1634                         if (parent > 0) {
1635                                 if (parent == ref_offset) {
1636                                         err = 0;
1637                                         break;
1638                                 }
1639                                 if (ref_offset < parent)
1640                                         break;
1641                         } else {
1642                                 if (root_objectid == ref_offset) {
1643                                         err = 0;
1644                                         break;
1645                                 }
1646                                 if (ref_offset < root_objectid)
1647                                         break;
1648                         }
1649                 }
1650                 ptr += btrfs_extent_inline_ref_size(type);
1651         }
1652         if (err == -ENOENT && insert) {
1653                 if (item_size + extra_size >=
1654                     BTRFS_MAX_EXTENT_ITEM_SIZE(root)) {
1655                         err = -EAGAIN;
1656                         goto out;
1657                 }
1658                 /*
1659                  * To add new inline back ref, we have to make sure
1660                  * there is no corresponding back ref item.
1661                  * For simplicity, we just do not add new inline back
1662                  * ref if there is any kind of item for this block
1663                  */
1664                 if (find_next_key(path, 0, &key) == 0 &&
1665                     key.objectid == bytenr &&
1666                     key.type < BTRFS_BLOCK_GROUP_ITEM_KEY) {
1667                         err = -EAGAIN;
1668                         goto out;
1669                 }
1670         }
1671         *ref_ret = (struct btrfs_extent_inline_ref *)ptr;
1672 out:
1673         if (insert) {
1674                 path->keep_locks = 0;
1675                 btrfs_unlock_up_safe(path, 1);
1676         }
1677         return err;
1678 }
1679
1680 /*
1681  * helper to add new inline back ref
1682  */
1683 static noinline_for_stack
1684 void setup_inline_extent_backref(struct btrfs_fs_info *fs_info,
1685                                  struct btrfs_path *path,
1686                                  struct btrfs_extent_inline_ref *iref,
1687                                  u64 parent, u64 root_objectid,
1688                                  u64 owner, u64 offset, int refs_to_add,
1689                                  struct btrfs_delayed_extent_op *extent_op)
1690 {
1691         struct extent_buffer *leaf;
1692         struct btrfs_extent_item *ei;
1693         unsigned long ptr;
1694         unsigned long end;
1695         unsigned long item_offset;
1696         u64 refs;
1697         int size;
1698         int type;
1699
1700         leaf = path->nodes[0];
1701         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1702         item_offset = (unsigned long)iref - (unsigned long)ei;
1703
1704         type = extent_ref_type(parent, owner);
1705         size = btrfs_extent_inline_ref_size(type);
1706
1707         btrfs_extend_item(fs_info, path, size);
1708
1709         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1710         refs = btrfs_extent_refs(leaf, ei);
1711         refs += refs_to_add;
1712         btrfs_set_extent_refs(leaf, ei, refs);
1713         if (extent_op)
1714                 __run_delayed_extent_op(extent_op, leaf, ei);
1715
1716         ptr = (unsigned long)ei + item_offset;
1717         end = (unsigned long)ei + btrfs_item_size_nr(leaf, path->slots[0]);
1718         if (ptr < end - size)
1719                 memmove_extent_buffer(leaf, ptr + size, ptr,
1720                                       end - size - ptr);
1721
1722         iref = (struct btrfs_extent_inline_ref *)ptr;
1723         btrfs_set_extent_inline_ref_type(leaf, iref, type);
1724         if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1725                 struct btrfs_extent_data_ref *dref;
1726                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1727                 btrfs_set_extent_data_ref_root(leaf, dref, root_objectid);
1728                 btrfs_set_extent_data_ref_objectid(leaf, dref, owner);
1729                 btrfs_set_extent_data_ref_offset(leaf, dref, offset);
1730                 btrfs_set_extent_data_ref_count(leaf, dref, refs_to_add);
1731         } else if (type == BTRFS_SHARED_DATA_REF_KEY) {
1732                 struct btrfs_shared_data_ref *sref;
1733                 sref = (struct btrfs_shared_data_ref *)(iref + 1);
1734                 btrfs_set_shared_data_ref_count(leaf, sref, refs_to_add);
1735                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
1736         } else if (type == BTRFS_SHARED_BLOCK_REF_KEY) {
1737                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
1738         } else {
1739                 btrfs_set_extent_inline_ref_offset(leaf, iref, root_objectid);
1740         }
1741         btrfs_mark_buffer_dirty(leaf);
1742 }
1743
1744 static int lookup_extent_backref(struct btrfs_trans_handle *trans,
1745                                  struct btrfs_path *path,
1746                                  struct btrfs_extent_inline_ref **ref_ret,
1747                                  u64 bytenr, u64 num_bytes, u64 parent,
1748                                  u64 root_objectid, u64 owner, u64 offset)
1749 {
1750         int ret;
1751
1752         ret = lookup_inline_extent_backref(trans, path, ref_ret, bytenr,
1753                                            num_bytes, parent, root_objectid,
1754                                            owner, offset, 0);
1755         if (ret != -ENOENT)
1756                 return ret;
1757
1758         btrfs_release_path(path);
1759         *ref_ret = NULL;
1760
1761         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1762                 ret = lookup_tree_block_ref(trans, path, bytenr, parent,
1763                                             root_objectid);
1764         } else {
1765                 ret = lookup_extent_data_ref(trans, path, bytenr, parent,
1766                                              root_objectid, owner, offset);
1767         }
1768         return ret;
1769 }
1770
1771 /*
1772  * helper to update/remove inline back ref
1773  */
1774 static noinline_for_stack
1775 void update_inline_extent_backref(struct btrfs_path *path,
1776                                   struct btrfs_extent_inline_ref *iref,
1777                                   int refs_to_mod,
1778                                   struct btrfs_delayed_extent_op *extent_op,
1779                                   int *last_ref)
1780 {
1781         struct extent_buffer *leaf = path->nodes[0];
1782         struct btrfs_fs_info *fs_info = leaf->fs_info;
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(fs_info, 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 #define in_range(b, first, len)        ((b) >= (first) && (b) < (first) + (len))
1909 static int btrfs_issue_discard(struct block_device *bdev, u64 start, u64 len,
1910                                u64 *discarded_bytes)
1911 {
1912         int j, ret = 0;
1913         u64 bytes_left, end;
1914         u64 aligned_start = ALIGN(start, 1 << 9);
1915
1916         if (WARN_ON(start != aligned_start)) {
1917                 len -= aligned_start - start;
1918                 len = round_down(len, 1 << 9);
1919                 start = aligned_start;
1920         }
1921
1922         *discarded_bytes = 0;
1923
1924         if (!len)
1925                 return 0;
1926
1927         end = start + len;
1928         bytes_left = len;
1929
1930         /* Skip any superblocks on this device. */
1931         for (j = 0; j < BTRFS_SUPER_MIRROR_MAX; j++) {
1932                 u64 sb_start = btrfs_sb_offset(j);
1933                 u64 sb_end = sb_start + BTRFS_SUPER_INFO_SIZE;
1934                 u64 size = sb_start - start;
1935
1936                 if (!in_range(sb_start, start, bytes_left) &&
1937                     !in_range(sb_end, start, bytes_left) &&
1938                     !in_range(start, sb_start, BTRFS_SUPER_INFO_SIZE))
1939                         continue;
1940
1941                 /*
1942                  * Superblock spans beginning of range.  Adjust start and
1943                  * try again.
1944                  */
1945                 if (sb_start <= start) {
1946                         start += sb_end - start;
1947                         if (start > end) {
1948                                 bytes_left = 0;
1949                                 break;
1950                         }
1951                         bytes_left = end - start;
1952                         continue;
1953                 }
1954
1955                 if (size) {
1956                         ret = blkdev_issue_discard(bdev, start >> 9, size >> 9,
1957                                                    GFP_NOFS, 0);
1958                         if (!ret)
1959                                 *discarded_bytes += size;
1960                         else if (ret != -EOPNOTSUPP)
1961                                 return ret;
1962                 }
1963
1964                 start = sb_end;
1965                 if (start > end) {
1966                         bytes_left = 0;
1967                         break;
1968                 }
1969                 bytes_left = end - start;
1970         }
1971
1972         if (bytes_left) {
1973                 ret = blkdev_issue_discard(bdev, start >> 9, bytes_left >> 9,
1974                                            GFP_NOFS, 0);
1975                 if (!ret)
1976                         *discarded_bytes += bytes_left;
1977         }
1978         return ret;
1979 }
1980
1981 int btrfs_discard_extent(struct btrfs_fs_info *fs_info, u64 bytenr,
1982                          u64 num_bytes, u64 *actual_bytes)
1983 {
1984         int ret;
1985         u64 discarded_bytes = 0;
1986         struct btrfs_bio *bbio = NULL;
1987
1988
1989         /*
1990          * Avoid races with device replace and make sure our bbio has devices
1991          * associated to its stripes that don't go away while we are discarding.
1992          */
1993         btrfs_bio_counter_inc_blocked(fs_info);
1994         /* Tell the block device(s) that the sectors can be discarded */
1995         ret = btrfs_map_block(fs_info, BTRFS_MAP_DISCARD, bytenr, &num_bytes,
1996                               &bbio, 0);
1997         /* Error condition is -ENOMEM */
1998         if (!ret) {
1999                 struct btrfs_bio_stripe *stripe = bbio->stripes;
2000                 int i;
2001
2002
2003                 for (i = 0; i < bbio->num_stripes; i++, stripe++) {
2004                         u64 bytes;
2005                         struct request_queue *req_q;
2006
2007                         if (!stripe->dev->bdev) {
2008                                 ASSERT(btrfs_test_opt(fs_info, DEGRADED));
2009                                 continue;
2010                         }
2011                         req_q = bdev_get_queue(stripe->dev->bdev);
2012                         if (!blk_queue_discard(req_q))
2013                                 continue;
2014
2015                         ret = btrfs_issue_discard(stripe->dev->bdev,
2016                                                   stripe->physical,
2017                                                   stripe->length,
2018                                                   &bytes);
2019                         if (!ret)
2020                                 discarded_bytes += bytes;
2021                         else if (ret != -EOPNOTSUPP)
2022                                 break; /* Logic errors or -ENOMEM, or -EIO but I don't know how that could happen JDM */
2023
2024                         /*
2025                          * Just in case we get back EOPNOTSUPP for some reason,
2026                          * just ignore the return value so we don't screw up
2027                          * people calling discard_extent.
2028                          */
2029                         ret = 0;
2030                 }
2031                 btrfs_put_bbio(bbio);
2032         }
2033         btrfs_bio_counter_dec(fs_info);
2034
2035         if (actual_bytes)
2036                 *actual_bytes = discarded_bytes;
2037
2038
2039         if (ret == -EOPNOTSUPP)
2040                 ret = 0;
2041         return ret;
2042 }
2043
2044 /* Can return -ENOMEM */
2045 int btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
2046                          struct btrfs_root *root,
2047                          u64 bytenr, u64 num_bytes, u64 parent,
2048                          u64 root_objectid, u64 owner, u64 offset)
2049 {
2050         struct btrfs_fs_info *fs_info = root->fs_info;
2051         int old_ref_mod, new_ref_mod;
2052         int ret;
2053
2054         BUG_ON(owner < BTRFS_FIRST_FREE_OBJECTID &&
2055                root_objectid == BTRFS_TREE_LOG_OBJECTID);
2056
2057         btrfs_ref_tree_mod(root, bytenr, num_bytes, parent, root_objectid,
2058                            owner, offset, BTRFS_ADD_DELAYED_REF);
2059
2060         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
2061                 ret = btrfs_add_delayed_tree_ref(trans, bytenr,
2062                                                  num_bytes, parent,
2063                                                  root_objectid, (int)owner,
2064                                                  BTRFS_ADD_DELAYED_REF, NULL,
2065                                                  &old_ref_mod, &new_ref_mod);
2066         } else {
2067                 ret = btrfs_add_delayed_data_ref(trans, bytenr,
2068                                                  num_bytes, parent,
2069                                                  root_objectid, owner, offset,
2070                                                  0, BTRFS_ADD_DELAYED_REF,
2071                                                  &old_ref_mod, &new_ref_mod);
2072         }
2073
2074         if (ret == 0 && old_ref_mod < 0 && new_ref_mod >= 0) {
2075                 bool metadata = owner < BTRFS_FIRST_FREE_OBJECTID;
2076
2077                 add_pinned_bytes(fs_info, -num_bytes, metadata, root_objectid);
2078         }
2079
2080         return ret;
2081 }
2082
2083 /*
2084  * __btrfs_inc_extent_ref - insert backreference for a given extent
2085  *
2086  * @trans:          Handle of transaction
2087  *
2088  * @node:           The delayed ref node used to get the bytenr/length for
2089  *                  extent whose references are incremented.
2090  *
2091  * @parent:         If this is a shared extent (BTRFS_SHARED_DATA_REF_KEY/
2092  *                  BTRFS_SHARED_BLOCK_REF_KEY) then it holds the logical
2093  *                  bytenr of the parent block. Since new extents are always
2094  *                  created with indirect references, this will only be the case
2095  *                  when relocating a shared extent. In that case, root_objectid
2096  *                  will be BTRFS_TREE_RELOC_OBJECTID. Otheriwse, parent must
2097  *                  be 0
2098  *
2099  * @root_objectid:  The id of the root where this modification has originated,
2100  *                  this can be either one of the well-known metadata trees or
2101  *                  the subvolume id which references this extent.
2102  *
2103  * @owner:          For data extents it is the inode number of the owning file.
2104  *                  For metadata extents this parameter holds the level in the
2105  *                  tree of the extent.
2106  *
2107  * @offset:         For metadata extents the offset is ignored and is currently
2108  *                  always passed as 0. For data extents it is the fileoffset
2109  *                  this extent belongs to.
2110  *
2111  * @refs_to_add     Number of references to add
2112  *
2113  * @extent_op       Pointer to a structure, holding information necessary when
2114  *                  updating a tree block's flags
2115  *
2116  */
2117 static int __btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
2118                                   struct btrfs_delayed_ref_node *node,
2119                                   u64 parent, u64 root_objectid,
2120                                   u64 owner, u64 offset, int refs_to_add,
2121                                   struct btrfs_delayed_extent_op *extent_op)
2122 {
2123         struct btrfs_path *path;
2124         struct extent_buffer *leaf;
2125         struct btrfs_extent_item *item;
2126         struct btrfs_key key;
2127         u64 bytenr = node->bytenr;
2128         u64 num_bytes = node->num_bytes;
2129         u64 refs;
2130         int ret;
2131
2132         path = btrfs_alloc_path();
2133         if (!path)
2134                 return -ENOMEM;
2135
2136         path->reada = READA_FORWARD;
2137         path->leave_spinning = 1;
2138         /* this will setup the path even if it fails to insert the back ref */
2139         ret = insert_inline_extent_backref(trans, path, bytenr, num_bytes,
2140                                            parent, root_objectid, owner,
2141                                            offset, refs_to_add, extent_op);
2142         if ((ret < 0 && ret != -EAGAIN) || !ret)
2143                 goto out;
2144
2145         /*
2146          * Ok we had -EAGAIN which means we didn't have space to insert and
2147          * inline extent ref, so just update the reference count and add a
2148          * normal backref.
2149          */
2150         leaf = path->nodes[0];
2151         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
2152         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2153         refs = btrfs_extent_refs(leaf, item);
2154         btrfs_set_extent_refs(leaf, item, refs + refs_to_add);
2155         if (extent_op)
2156                 __run_delayed_extent_op(extent_op, leaf, item);
2157
2158         btrfs_mark_buffer_dirty(leaf);
2159         btrfs_release_path(path);
2160
2161         path->reada = READA_FORWARD;
2162         path->leave_spinning = 1;
2163         /* now insert the actual backref */
2164         ret = insert_extent_backref(trans, path, bytenr, parent, root_objectid,
2165                                     owner, offset, refs_to_add);
2166         if (ret)
2167                 btrfs_abort_transaction(trans, ret);
2168 out:
2169         btrfs_free_path(path);
2170         return ret;
2171 }
2172
2173 static int run_delayed_data_ref(struct btrfs_trans_handle *trans,
2174                                 struct btrfs_delayed_ref_node *node,
2175                                 struct btrfs_delayed_extent_op *extent_op,
2176                                 int insert_reserved)
2177 {
2178         int ret = 0;
2179         struct btrfs_delayed_data_ref *ref;
2180         struct btrfs_key ins;
2181         u64 parent = 0;
2182         u64 ref_root = 0;
2183         u64 flags = 0;
2184
2185         ins.objectid = node->bytenr;
2186         ins.offset = node->num_bytes;
2187         ins.type = BTRFS_EXTENT_ITEM_KEY;
2188
2189         ref = btrfs_delayed_node_to_data_ref(node);
2190         trace_run_delayed_data_ref(trans->fs_info, node, ref, node->action);
2191
2192         if (node->type == BTRFS_SHARED_DATA_REF_KEY)
2193                 parent = ref->parent;
2194         ref_root = ref->root;
2195
2196         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
2197                 if (extent_op)
2198                         flags |= extent_op->flags_to_set;
2199                 ret = alloc_reserved_file_extent(trans, parent, ref_root,
2200                                                  flags, ref->objectid,
2201                                                  ref->offset, &ins,
2202                                                  node->ref_mod);
2203         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
2204                 ret = __btrfs_inc_extent_ref(trans, node, parent, ref_root,
2205                                              ref->objectid, ref->offset,
2206                                              node->ref_mod, extent_op);
2207         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
2208                 ret = __btrfs_free_extent(trans, node, parent,
2209                                           ref_root, ref->objectid,
2210                                           ref->offset, node->ref_mod,
2211                                           extent_op);
2212         } else {
2213                 BUG();
2214         }
2215         return ret;
2216 }
2217
2218 static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op,
2219                                     struct extent_buffer *leaf,
2220                                     struct btrfs_extent_item *ei)
2221 {
2222         u64 flags = btrfs_extent_flags(leaf, ei);
2223         if (extent_op->update_flags) {
2224                 flags |= extent_op->flags_to_set;
2225                 btrfs_set_extent_flags(leaf, ei, flags);
2226         }
2227
2228         if (extent_op->update_key) {
2229                 struct btrfs_tree_block_info *bi;
2230                 BUG_ON(!(flags & BTRFS_EXTENT_FLAG_TREE_BLOCK));
2231                 bi = (struct btrfs_tree_block_info *)(ei + 1);
2232                 btrfs_set_tree_block_key(leaf, bi, &extent_op->key);
2233         }
2234 }
2235
2236 static int run_delayed_extent_op(struct btrfs_trans_handle *trans,
2237                                  struct btrfs_delayed_ref_head *head,
2238                                  struct btrfs_delayed_extent_op *extent_op)
2239 {
2240         struct btrfs_fs_info *fs_info = trans->fs_info;
2241         struct btrfs_key key;
2242         struct btrfs_path *path;
2243         struct btrfs_extent_item *ei;
2244         struct extent_buffer *leaf;
2245         u32 item_size;
2246         int ret;
2247         int err = 0;
2248         int metadata = !extent_op->is_data;
2249
2250         if (trans->aborted)
2251                 return 0;
2252
2253         if (metadata && !btrfs_fs_incompat(fs_info, SKINNY_METADATA))
2254                 metadata = 0;
2255
2256         path = btrfs_alloc_path();
2257         if (!path)
2258                 return -ENOMEM;
2259
2260         key.objectid = head->bytenr;
2261
2262         if (metadata) {
2263                 key.type = BTRFS_METADATA_ITEM_KEY;
2264                 key.offset = extent_op->level;
2265         } else {
2266                 key.type = BTRFS_EXTENT_ITEM_KEY;
2267                 key.offset = head->num_bytes;
2268         }
2269
2270 again:
2271         path->reada = READA_FORWARD;
2272         path->leave_spinning = 1;
2273         ret = btrfs_search_slot(trans, fs_info->extent_root, &key, path, 0, 1);
2274         if (ret < 0) {
2275                 err = ret;
2276                 goto out;
2277         }
2278         if (ret > 0) {
2279                 if (metadata) {
2280                         if (path->slots[0] > 0) {
2281                                 path->slots[0]--;
2282                                 btrfs_item_key_to_cpu(path->nodes[0], &key,
2283                                                       path->slots[0]);
2284                                 if (key.objectid == head->bytenr &&
2285                                     key.type == BTRFS_EXTENT_ITEM_KEY &&
2286                                     key.offset == head->num_bytes)
2287                                         ret = 0;
2288                         }
2289                         if (ret > 0) {
2290                                 btrfs_release_path(path);
2291                                 metadata = 0;
2292
2293                                 key.objectid = head->bytenr;
2294                                 key.offset = head->num_bytes;
2295                                 key.type = BTRFS_EXTENT_ITEM_KEY;
2296                                 goto again;
2297                         }
2298                 } else {
2299                         err = -EIO;
2300                         goto out;
2301                 }
2302         }
2303
2304         leaf = path->nodes[0];
2305         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2306
2307         if (unlikely(item_size < sizeof(*ei))) {
2308                 err = -EINVAL;
2309                 btrfs_print_v0_err(fs_info);
2310                 btrfs_abort_transaction(trans, err);
2311                 goto out;
2312         }
2313
2314         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2315         __run_delayed_extent_op(extent_op, leaf, ei);
2316
2317         btrfs_mark_buffer_dirty(leaf);
2318 out:
2319         btrfs_free_path(path);
2320         return err;
2321 }
2322
2323 static int run_delayed_tree_ref(struct btrfs_trans_handle *trans,
2324                                 struct btrfs_delayed_ref_node *node,
2325                                 struct btrfs_delayed_extent_op *extent_op,
2326                                 int insert_reserved)
2327 {
2328         int ret = 0;
2329         struct btrfs_delayed_tree_ref *ref;
2330         u64 parent = 0;
2331         u64 ref_root = 0;
2332
2333         ref = btrfs_delayed_node_to_tree_ref(node);
2334         trace_run_delayed_tree_ref(trans->fs_info, node, ref, node->action);
2335
2336         if (node->type == BTRFS_SHARED_BLOCK_REF_KEY)
2337                 parent = ref->parent;
2338         ref_root = ref->root;
2339
2340         if (node->ref_mod != 1) {
2341                 btrfs_err(trans->fs_info,
2342         "btree block(%llu) has %d references rather than 1: action %d ref_root %llu parent %llu",
2343                           node->bytenr, node->ref_mod, node->action, ref_root,
2344                           parent);
2345                 return -EIO;
2346         }
2347         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
2348                 BUG_ON(!extent_op || !extent_op->update_flags);
2349                 ret = alloc_reserved_tree_block(trans, node, extent_op);
2350         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
2351                 ret = __btrfs_inc_extent_ref(trans, node, parent, ref_root,
2352                                              ref->level, 0, 1, extent_op);
2353         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
2354                 ret = __btrfs_free_extent(trans, node, parent, ref_root,
2355                                           ref->level, 0, 1, extent_op);
2356         } else {
2357                 BUG();
2358         }
2359         return ret;
2360 }
2361
2362 /* helper function to actually process a single delayed ref entry */
2363 static int run_one_delayed_ref(struct btrfs_trans_handle *trans,
2364                                struct btrfs_delayed_ref_node *node,
2365                                struct btrfs_delayed_extent_op *extent_op,
2366                                int insert_reserved)
2367 {
2368         int ret = 0;
2369
2370         if (trans->aborted) {
2371                 if (insert_reserved)
2372                         btrfs_pin_extent(trans->fs_info, node->bytenr,
2373                                          node->num_bytes, 1);
2374                 return 0;
2375         }
2376
2377         if (node->type == BTRFS_TREE_BLOCK_REF_KEY ||
2378             node->type == BTRFS_SHARED_BLOCK_REF_KEY)
2379                 ret = run_delayed_tree_ref(trans, node, extent_op,
2380                                            insert_reserved);
2381         else if (node->type == BTRFS_EXTENT_DATA_REF_KEY ||
2382                  node->type == BTRFS_SHARED_DATA_REF_KEY)
2383                 ret = run_delayed_data_ref(trans, node, extent_op,
2384                                            insert_reserved);
2385         else
2386                 BUG();
2387         if (ret && insert_reserved)
2388                 btrfs_pin_extent(trans->fs_info, node->bytenr,
2389                                  node->num_bytes, 1);
2390         return ret;
2391 }
2392
2393 static inline struct btrfs_delayed_ref_node *
2394 select_delayed_ref(struct btrfs_delayed_ref_head *head)
2395 {
2396         struct btrfs_delayed_ref_node *ref;
2397
2398         if (RB_EMPTY_ROOT(&head->ref_tree.rb_root))
2399                 return NULL;
2400
2401         /*
2402          * Select a delayed ref of type BTRFS_ADD_DELAYED_REF first.
2403          * This is to prevent a ref count from going down to zero, which deletes
2404          * the extent item from the extent tree, when there still are references
2405          * to add, which would fail because they would not find the extent item.
2406          */
2407         if (!list_empty(&head->ref_add_list))
2408                 return list_first_entry(&head->ref_add_list,
2409                                 struct btrfs_delayed_ref_node, add_list);
2410
2411         ref = rb_entry(rb_first_cached(&head->ref_tree),
2412                        struct btrfs_delayed_ref_node, ref_node);
2413         ASSERT(list_empty(&ref->add_list));
2414         return ref;
2415 }
2416
2417 static void unselect_delayed_ref_head(struct btrfs_delayed_ref_root *delayed_refs,
2418                                       struct btrfs_delayed_ref_head *head)
2419 {
2420         spin_lock(&delayed_refs->lock);
2421         head->processing = 0;
2422         delayed_refs->num_heads_ready++;
2423         spin_unlock(&delayed_refs->lock);
2424         btrfs_delayed_ref_unlock(head);
2425 }
2426
2427 static struct btrfs_delayed_extent_op *cleanup_extent_op(
2428                                 struct btrfs_delayed_ref_head *head)
2429 {
2430         struct btrfs_delayed_extent_op *extent_op = head->extent_op;
2431
2432         if (!extent_op)
2433                 return NULL;
2434
2435         if (head->must_insert_reserved) {
2436                 head->extent_op = NULL;
2437                 btrfs_free_delayed_extent_op(extent_op);
2438                 return NULL;
2439         }
2440         return extent_op;
2441 }
2442
2443 static int run_and_cleanup_extent_op(struct btrfs_trans_handle *trans,
2444                                      struct btrfs_delayed_ref_head *head)
2445 {
2446         struct btrfs_delayed_extent_op *extent_op;
2447         int ret;
2448
2449         extent_op = cleanup_extent_op(head);
2450         if (!extent_op)
2451                 return 0;
2452         head->extent_op = NULL;
2453         spin_unlock(&head->lock);
2454         ret = run_delayed_extent_op(trans, head, extent_op);
2455         btrfs_free_delayed_extent_op(extent_op);
2456         return ret ? ret : 1;
2457 }
2458
2459 void btrfs_cleanup_ref_head_accounting(struct btrfs_fs_info *fs_info,
2460                                   struct btrfs_delayed_ref_root *delayed_refs,
2461                                   struct btrfs_delayed_ref_head *head)
2462 {
2463         int nr_items = 1;       /* Dropping this ref head update. */
2464
2465         if (head->total_ref_mod < 0) {
2466                 struct btrfs_space_info *space_info;
2467                 u64 flags;
2468
2469                 if (head->is_data)
2470                         flags = BTRFS_BLOCK_GROUP_DATA;
2471                 else if (head->is_system)
2472                         flags = BTRFS_BLOCK_GROUP_SYSTEM;
2473                 else
2474                         flags = BTRFS_BLOCK_GROUP_METADATA;
2475                 space_info = __find_space_info(fs_info, flags);
2476                 ASSERT(space_info);
2477                 percpu_counter_add_batch(&space_info->total_bytes_pinned,
2478                                    -head->num_bytes,
2479                                    BTRFS_TOTAL_BYTES_PINNED_BATCH);
2480
2481                 /*
2482                  * We had csum deletions accounted for in our delayed refs rsv,
2483                  * we need to drop the csum leaves for this update from our
2484                  * delayed_refs_rsv.
2485                  */
2486                 if (head->is_data) {
2487                         spin_lock(&delayed_refs->lock);
2488                         delayed_refs->pending_csums -= head->num_bytes;
2489                         spin_unlock(&delayed_refs->lock);
2490                         nr_items += btrfs_csum_bytes_to_leaves(fs_info,
2491                                 head->num_bytes);
2492                 }
2493         }
2494
2495         /* Also free its reserved qgroup space */
2496         btrfs_qgroup_free_delayed_ref(fs_info, head->qgroup_ref_root,
2497                                       head->qgroup_reserved);
2498         btrfs_delayed_refs_rsv_release(fs_info, nr_items);
2499 }
2500
2501 static int cleanup_ref_head(struct btrfs_trans_handle *trans,
2502                             struct btrfs_delayed_ref_head *head)
2503 {
2504
2505         struct btrfs_fs_info *fs_info = trans->fs_info;
2506         struct btrfs_delayed_ref_root *delayed_refs;
2507         int ret;
2508
2509         delayed_refs = &trans->transaction->delayed_refs;
2510
2511         ret = run_and_cleanup_extent_op(trans, head);
2512         if (ret < 0) {
2513                 unselect_delayed_ref_head(delayed_refs, head);
2514                 btrfs_debug(fs_info, "run_delayed_extent_op returned %d", ret);
2515                 return ret;
2516         } else if (ret) {
2517                 return ret;
2518         }
2519
2520         /*
2521          * Need to drop our head ref lock and re-acquire the delayed ref lock
2522          * and then re-check to make sure nobody got added.
2523          */
2524         spin_unlock(&head->lock);
2525         spin_lock(&delayed_refs->lock);
2526         spin_lock(&head->lock);
2527         if (!RB_EMPTY_ROOT(&head->ref_tree.rb_root) || head->extent_op) {
2528                 spin_unlock(&head->lock);
2529                 spin_unlock(&delayed_refs->lock);
2530                 return 1;
2531         }
2532         btrfs_delete_ref_head(delayed_refs, head);
2533         spin_unlock(&head->lock);
2534         spin_unlock(&delayed_refs->lock);
2535
2536         if (head->must_insert_reserved) {
2537                 btrfs_pin_extent(fs_info, head->bytenr,
2538                                  head->num_bytes, 1);
2539                 if (head->is_data) {
2540                         ret = btrfs_del_csums(trans, fs_info, head->bytenr,
2541                                               head->num_bytes);
2542                 }
2543         }
2544
2545         btrfs_cleanup_ref_head_accounting(fs_info, delayed_refs, head);
2546
2547         trace_run_delayed_ref_head(fs_info, head, 0);
2548         btrfs_delayed_ref_unlock(head);
2549         btrfs_put_delayed_ref_head(head);
2550         return 0;
2551 }
2552
2553 static struct btrfs_delayed_ref_head *btrfs_obtain_ref_head(
2554                                         struct btrfs_trans_handle *trans)
2555 {
2556         struct btrfs_delayed_ref_root *delayed_refs =
2557                 &trans->transaction->delayed_refs;
2558         struct btrfs_delayed_ref_head *head = NULL;
2559         int ret;
2560
2561         spin_lock(&delayed_refs->lock);
2562         head = btrfs_select_ref_head(delayed_refs);
2563         if (!head) {
2564                 spin_unlock(&delayed_refs->lock);
2565                 return head;
2566         }
2567
2568         /*
2569          * Grab the lock that says we are going to process all the refs for
2570          * this head
2571          */
2572         ret = btrfs_delayed_ref_lock(delayed_refs, head);
2573         spin_unlock(&delayed_refs->lock);
2574
2575         /*
2576          * We may have dropped the spin lock to get the head mutex lock, and
2577          * that might have given someone else time to free the head.  If that's
2578          * true, it has been removed from our list and we can move on.
2579          */
2580         if (ret == -EAGAIN)
2581                 head = ERR_PTR(-EAGAIN);
2582
2583         return head;
2584 }
2585
2586 static int btrfs_run_delayed_refs_for_head(struct btrfs_trans_handle *trans,
2587                                     struct btrfs_delayed_ref_head *locked_ref,
2588                                     unsigned long *run_refs)
2589 {
2590         struct btrfs_fs_info *fs_info = trans->fs_info;
2591         struct btrfs_delayed_ref_root *delayed_refs;
2592         struct btrfs_delayed_extent_op *extent_op;
2593         struct btrfs_delayed_ref_node *ref;
2594         int must_insert_reserved = 0;
2595         int ret;
2596
2597         delayed_refs = &trans->transaction->delayed_refs;
2598
2599         lockdep_assert_held(&locked_ref->mutex);
2600         lockdep_assert_held(&locked_ref->lock);
2601
2602         while ((ref = select_delayed_ref(locked_ref))) {
2603                 if (ref->seq &&
2604                     btrfs_check_delayed_seq(fs_info, ref->seq)) {
2605                         spin_unlock(&locked_ref->lock);
2606                         unselect_delayed_ref_head(delayed_refs, locked_ref);
2607                         return -EAGAIN;
2608                 }
2609
2610                 (*run_refs)++;
2611                 ref->in_tree = 0;
2612                 rb_erase_cached(&ref->ref_node, &locked_ref->ref_tree);
2613                 RB_CLEAR_NODE(&ref->ref_node);
2614                 if (!list_empty(&ref->add_list))
2615                         list_del(&ref->add_list);
2616                 /*
2617                  * When we play the delayed ref, also correct the ref_mod on
2618                  * head
2619                  */
2620                 switch (ref->action) {
2621                 case BTRFS_ADD_DELAYED_REF:
2622                 case BTRFS_ADD_DELAYED_EXTENT:
2623                         locked_ref->ref_mod -= ref->ref_mod;
2624                         break;
2625                 case BTRFS_DROP_DELAYED_REF:
2626                         locked_ref->ref_mod += ref->ref_mod;
2627                         break;
2628                 default:
2629                         WARN_ON(1);
2630                 }
2631                 atomic_dec(&delayed_refs->num_entries);
2632
2633                 /*
2634                  * Record the must_insert_reserved flag before we drop the
2635                  * spin lock.
2636                  */
2637                 must_insert_reserved = locked_ref->must_insert_reserved;
2638                 locked_ref->must_insert_reserved = 0;
2639
2640                 extent_op = locked_ref->extent_op;
2641                 locked_ref->extent_op = NULL;
2642                 spin_unlock(&locked_ref->lock);
2643
2644                 ret = run_one_delayed_ref(trans, ref, extent_op,
2645                                           must_insert_reserved);
2646
2647                 btrfs_free_delayed_extent_op(extent_op);
2648                 if (ret) {
2649                         unselect_delayed_ref_head(delayed_refs, locked_ref);
2650                         btrfs_put_delayed_ref(ref);
2651                         btrfs_debug(fs_info, "run_one_delayed_ref returned %d",
2652                                     ret);
2653                         return ret;
2654                 }
2655
2656                 btrfs_put_delayed_ref(ref);
2657                 cond_resched();
2658
2659                 spin_lock(&locked_ref->lock);
2660                 btrfs_merge_delayed_refs(trans, delayed_refs, locked_ref);
2661         }
2662
2663         return 0;
2664 }
2665
2666 /*
2667  * Returns 0 on success or if called with an already aborted transaction.
2668  * Returns -ENOMEM or -EIO on failure and will abort the transaction.
2669  */
2670 static noinline int __btrfs_run_delayed_refs(struct btrfs_trans_handle *trans,
2671                                              unsigned long nr)
2672 {
2673         struct btrfs_fs_info *fs_info = trans->fs_info;
2674         struct btrfs_delayed_ref_root *delayed_refs;
2675         struct btrfs_delayed_ref_head *locked_ref = NULL;
2676         ktime_t start = ktime_get();
2677         int ret;
2678         unsigned long count = 0;
2679         unsigned long actual_count = 0;
2680
2681         delayed_refs = &trans->transaction->delayed_refs;
2682         do {
2683                 if (!locked_ref) {
2684                         locked_ref = btrfs_obtain_ref_head(trans);
2685                         if (IS_ERR_OR_NULL(locked_ref)) {
2686                                 if (PTR_ERR(locked_ref) == -EAGAIN) {
2687                                         continue;
2688                                 } else {
2689                                         break;
2690                                 }
2691                         }
2692                         count++;
2693                 }
2694                 /*
2695                  * We need to try and merge add/drops of the same ref since we
2696                  * can run into issues with relocate dropping the implicit ref
2697                  * and then it being added back again before the drop can
2698                  * finish.  If we merged anything we need to re-loop so we can
2699                  * get a good ref.
2700                  * Or we can get node references of the same type that weren't
2701                  * merged when created due to bumps in the tree mod seq, and
2702                  * we need to merge them to prevent adding an inline extent
2703                  * backref before dropping it (triggering a BUG_ON at
2704                  * insert_inline_extent_backref()).
2705                  */
2706                 spin_lock(&locked_ref->lock);
2707                 btrfs_merge_delayed_refs(trans, delayed_refs, locked_ref);
2708
2709                 ret = btrfs_run_delayed_refs_for_head(trans, locked_ref,
2710                                                       &actual_count);
2711                 if (ret < 0 && ret != -EAGAIN) {
2712                         /*
2713                          * Error, btrfs_run_delayed_refs_for_head already
2714                          * unlocked everything so just bail out
2715                          */
2716                         return ret;
2717                 } else if (!ret) {
2718                         /*
2719                          * Success, perform the usual cleanup of a processed
2720                          * head
2721                          */
2722                         ret = cleanup_ref_head(trans, locked_ref);
2723                         if (ret > 0 ) {
2724                                 /* We dropped our lock, we need to loop. */
2725                                 ret = 0;
2726                                 continue;
2727                         } else if (ret) {
2728                                 return ret;
2729                         }
2730                 }
2731
2732                 /*
2733                  * Either success case or btrfs_run_delayed_refs_for_head
2734                  * returned -EAGAIN, meaning we need to select another head
2735                  */
2736
2737                 locked_ref = NULL;
2738                 cond_resched();
2739         } while ((nr != -1 && count < nr) || locked_ref);
2740
2741         /*
2742          * We don't want to include ref heads since we can have empty ref heads
2743          * and those will drastically skew our runtime down since we just do
2744          * accounting, no actual extent tree updates.
2745          */
2746         if (actual_count > 0) {
2747                 u64 runtime = ktime_to_ns(ktime_sub(ktime_get(), start));
2748                 u64 avg;
2749
2750                 /*
2751                  * We weigh the current average higher than our current runtime
2752                  * to avoid large swings in the average.
2753                  */
2754                 spin_lock(&delayed_refs->lock);
2755                 avg = fs_info->avg_delayed_ref_runtime * 3 + runtime;
2756                 fs_info->avg_delayed_ref_runtime = avg >> 2;    /* div by 4 */
2757                 spin_unlock(&delayed_refs->lock);
2758         }
2759         return 0;
2760 }
2761
2762 #ifdef SCRAMBLE_DELAYED_REFS
2763 /*
2764  * Normally delayed refs get processed in ascending bytenr order. This
2765  * correlates in most cases to the order added. To expose dependencies on this
2766  * order, we start to process the tree in the middle instead of the beginning
2767  */
2768 static u64 find_middle(struct rb_root *root)
2769 {
2770         struct rb_node *n = root->rb_node;
2771         struct btrfs_delayed_ref_node *entry;
2772         int alt = 1;
2773         u64 middle;
2774         u64 first = 0, last = 0;
2775
2776         n = rb_first(root);
2777         if (n) {
2778                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2779                 first = entry->bytenr;
2780         }
2781         n = rb_last(root);
2782         if (n) {
2783                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2784                 last = entry->bytenr;
2785         }
2786         n = root->rb_node;
2787
2788         while (n) {
2789                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2790                 WARN_ON(!entry->in_tree);
2791
2792                 middle = entry->bytenr;
2793
2794                 if (alt)
2795                         n = n->rb_left;
2796                 else
2797                         n = n->rb_right;
2798
2799                 alt = 1 - alt;
2800         }
2801         return middle;
2802 }
2803 #endif
2804
2805 static inline u64 heads_to_leaves(struct btrfs_fs_info *fs_info, u64 heads)
2806 {
2807         u64 num_bytes;
2808
2809         num_bytes = heads * (sizeof(struct btrfs_extent_item) +
2810                              sizeof(struct btrfs_extent_inline_ref));
2811         if (!btrfs_fs_incompat(fs_info, SKINNY_METADATA))
2812                 num_bytes += heads * sizeof(struct btrfs_tree_block_info);
2813
2814         /*
2815          * We don't ever fill up leaves all the way so multiply by 2 just to be
2816          * closer to what we're really going to want to use.
2817          */
2818         return div_u64(num_bytes, BTRFS_LEAF_DATA_SIZE(fs_info));
2819 }
2820
2821 /*
2822  * Takes the number of bytes to be csumm'ed and figures out how many leaves it
2823  * would require to store the csums for that many bytes.
2824  */
2825 u64 btrfs_csum_bytes_to_leaves(struct btrfs_fs_info *fs_info, u64 csum_bytes)
2826 {
2827         u64 csum_size;
2828         u64 num_csums_per_leaf;
2829         u64 num_csums;
2830
2831         csum_size = BTRFS_MAX_ITEM_SIZE(fs_info);
2832         num_csums_per_leaf = div64_u64(csum_size,
2833                         (u64)btrfs_super_csum_size(fs_info->super_copy));
2834         num_csums = div64_u64(csum_bytes, fs_info->sectorsize);
2835         num_csums += num_csums_per_leaf - 1;
2836         num_csums = div64_u64(num_csums, num_csums_per_leaf);
2837         return num_csums;
2838 }
2839
2840 bool btrfs_check_space_for_delayed_refs(struct btrfs_fs_info *fs_info)
2841 {
2842         struct btrfs_block_rsv *delayed_refs_rsv = &fs_info->delayed_refs_rsv;
2843         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
2844         bool ret = false;
2845         u64 reserved;
2846
2847         spin_lock(&global_rsv->lock);
2848         reserved = global_rsv->reserved;
2849         spin_unlock(&global_rsv->lock);
2850
2851         /*
2852          * Since the global reserve is just kind of magic we don't really want
2853          * to rely on it to save our bacon, so if our size is more than the
2854          * delayed_refs_rsv and the global rsv then it's time to think about
2855          * bailing.
2856          */
2857         spin_lock(&delayed_refs_rsv->lock);
2858         reserved += delayed_refs_rsv->reserved;
2859         if (delayed_refs_rsv->size >= reserved)
2860                 ret = true;
2861         spin_unlock(&delayed_refs_rsv->lock);
2862         return ret;
2863 }
2864
2865 int btrfs_should_throttle_delayed_refs(struct btrfs_trans_handle *trans)
2866 {
2867         u64 num_entries =
2868                 atomic_read(&trans->transaction->delayed_refs.num_entries);
2869         u64 avg_runtime;
2870         u64 val;
2871
2872         smp_mb();
2873         avg_runtime = trans->fs_info->avg_delayed_ref_runtime;
2874         val = num_entries * avg_runtime;
2875         if (val >= NSEC_PER_SEC)
2876                 return 1;
2877         if (val >= NSEC_PER_SEC / 2)
2878                 return 2;
2879
2880         return btrfs_check_space_for_delayed_refs(trans->fs_info);
2881 }
2882
2883 struct async_delayed_refs {
2884         struct btrfs_root *root;
2885         u64 transid;
2886         int count;
2887         int error;
2888         int sync;
2889         struct completion wait;
2890         struct btrfs_work work;
2891 };
2892
2893 static inline struct async_delayed_refs *
2894 to_async_delayed_refs(struct btrfs_work *work)
2895 {
2896         return container_of(work, struct async_delayed_refs, work);
2897 }
2898
2899 static void delayed_ref_async_start(struct btrfs_work *work)
2900 {
2901         struct async_delayed_refs *async = to_async_delayed_refs(work);
2902         struct btrfs_trans_handle *trans;
2903         struct btrfs_fs_info *fs_info = async->root->fs_info;
2904         int ret;
2905
2906         /* if the commit is already started, we don't need to wait here */
2907         if (btrfs_transaction_blocked(fs_info))
2908                 goto done;
2909
2910         trans = btrfs_join_transaction(async->root);
2911         if (IS_ERR(trans)) {
2912                 async->error = PTR_ERR(trans);
2913                 goto done;
2914         }
2915
2916         /*
2917          * trans->sync means that when we call end_transaction, we won't
2918          * wait on delayed refs
2919          */
2920         trans->sync = true;
2921
2922         /* Don't bother flushing if we got into a different transaction */
2923         if (trans->transid > async->transid)
2924                 goto end;
2925
2926         ret = btrfs_run_delayed_refs(trans, async->count);
2927         if (ret)
2928                 async->error = ret;
2929 end:
2930         ret = btrfs_end_transaction(trans);
2931         if (ret && !async->error)
2932                 async->error = ret;
2933 done:
2934         if (async->sync)
2935                 complete(&async->wait);
2936         else
2937                 kfree(async);
2938 }
2939
2940 int btrfs_async_run_delayed_refs(struct btrfs_fs_info *fs_info,
2941                                  unsigned long count, u64 transid, int wait)
2942 {
2943         struct async_delayed_refs *async;
2944         int ret;
2945
2946         async = kmalloc(sizeof(*async), GFP_NOFS);
2947         if (!async)
2948                 return -ENOMEM;
2949
2950         async->root = fs_info->tree_root;
2951         async->count = count;
2952         async->error = 0;
2953         async->transid = transid;
2954         if (wait)
2955                 async->sync = 1;
2956         else
2957                 async->sync = 0;
2958         init_completion(&async->wait);
2959
2960         btrfs_init_work(&async->work, btrfs_extent_refs_helper,
2961                         delayed_ref_async_start, NULL, NULL);
2962
2963         btrfs_queue_work(fs_info->extent_workers, &async->work);
2964
2965         if (wait) {
2966                 wait_for_completion(&async->wait);
2967                 ret = async->error;
2968                 kfree(async);
2969                 return ret;
2970         }
2971         return 0;
2972 }
2973
2974 /*
2975  * this starts processing the delayed reference count updates and
2976  * extent insertions we have queued up so far.  count can be
2977  * 0, which means to process everything in the tree at the start
2978  * of the run (but not newly added entries), or it can be some target
2979  * number you'd like to process.
2980  *
2981  * Returns 0 on success or if called with an aborted transaction
2982  * Returns <0 on error and aborts the transaction
2983  */
2984 int btrfs_run_delayed_refs(struct btrfs_trans_handle *trans,
2985                            unsigned long count)
2986 {
2987         struct btrfs_fs_info *fs_info = trans->fs_info;
2988         struct rb_node *node;
2989         struct btrfs_delayed_ref_root *delayed_refs;
2990         struct btrfs_delayed_ref_head *head;
2991         int ret;
2992         int run_all = count == (unsigned long)-1;
2993
2994         /* We'll clean this up in btrfs_cleanup_transaction */
2995         if (trans->aborted)
2996                 return 0;
2997
2998         if (test_bit(BTRFS_FS_CREATING_FREE_SPACE_TREE, &fs_info->flags))
2999                 return 0;
3000
3001         delayed_refs = &trans->transaction->delayed_refs;
3002         if (count == 0)
3003                 count = atomic_read(&delayed_refs->num_entries) * 2;
3004
3005 again:
3006 #ifdef SCRAMBLE_DELAYED_REFS
3007         delayed_refs->run_delayed_start = find_middle(&delayed_refs->root);
3008 #endif
3009         ret = __btrfs_run_delayed_refs(trans, count);
3010         if (ret < 0) {
3011                 btrfs_abort_transaction(trans, ret);
3012                 return ret;
3013         }
3014
3015         if (run_all) {
3016                 btrfs_create_pending_block_groups(trans);
3017
3018                 spin_lock(&delayed_refs->lock);
3019                 node = rb_first_cached(&delayed_refs->href_root);
3020                 if (!node) {
3021                         spin_unlock(&delayed_refs->lock);
3022                         goto out;
3023                 }
3024                 head = rb_entry(node, struct btrfs_delayed_ref_head,
3025                                 href_node);
3026                 refcount_inc(&head->refs);
3027                 spin_unlock(&delayed_refs->lock);
3028
3029                 /* Mutex was contended, block until it's released and retry. */
3030                 mutex_lock(&head->mutex);
3031                 mutex_unlock(&head->mutex);
3032
3033                 btrfs_put_delayed_ref_head(head);
3034                 cond_resched();
3035                 goto again;
3036         }
3037 out:
3038         return 0;
3039 }
3040
3041 int btrfs_set_disk_extent_flags(struct btrfs_trans_handle *trans,
3042                                 struct btrfs_fs_info *fs_info,
3043                                 u64 bytenr, u64 num_bytes, u64 flags,
3044                                 int level, int is_data)
3045 {
3046         struct btrfs_delayed_extent_op *extent_op;
3047         int ret;
3048
3049         extent_op = btrfs_alloc_delayed_extent_op();
3050         if (!extent_op)
3051                 return -ENOMEM;
3052
3053         extent_op->flags_to_set = flags;
3054         extent_op->update_flags = true;
3055         extent_op->update_key = false;
3056         extent_op->is_data = is_data ? true : false;
3057         extent_op->level = level;
3058
3059         ret = btrfs_add_delayed_extent_op(fs_info, trans, bytenr,
3060                                           num_bytes, extent_op);
3061         if (ret)
3062                 btrfs_free_delayed_extent_op(extent_op);
3063         return ret;
3064 }
3065
3066 static noinline int check_delayed_ref(struct btrfs_root *root,
3067                                       struct btrfs_path *path,
3068                                       u64 objectid, u64 offset, u64 bytenr)
3069 {
3070         struct btrfs_delayed_ref_head *head;
3071         struct btrfs_delayed_ref_node *ref;
3072         struct btrfs_delayed_data_ref *data_ref;
3073         struct btrfs_delayed_ref_root *delayed_refs;
3074         struct btrfs_transaction *cur_trans;
3075         struct rb_node *node;
3076         int ret = 0;
3077
3078         spin_lock(&root->fs_info->trans_lock);
3079         cur_trans = root->fs_info->running_transaction;
3080         if (cur_trans)
3081                 refcount_inc(&cur_trans->use_count);
3082         spin_unlock(&root->fs_info->trans_lock);
3083         if (!cur_trans)
3084                 return 0;
3085
3086         delayed_refs = &cur_trans->delayed_refs;
3087         spin_lock(&delayed_refs->lock);
3088         head = btrfs_find_delayed_ref_head(delayed_refs, bytenr);
3089         if (!head) {
3090                 spin_unlock(&delayed_refs->lock);
3091                 btrfs_put_transaction(cur_trans);
3092                 return 0;
3093         }
3094
3095         if (!mutex_trylock(&head->mutex)) {
3096                 refcount_inc(&head->refs);
3097                 spin_unlock(&delayed_refs->lock);
3098
3099                 btrfs_release_path(path);
3100
3101                 /*
3102                  * Mutex was contended, block until it's released and let
3103                  * caller try again
3104                  */
3105                 mutex_lock(&head->mutex);
3106                 mutex_unlock(&head->mutex);
3107                 btrfs_put_delayed_ref_head(head);
3108                 btrfs_put_transaction(cur_trans);
3109                 return -EAGAIN;
3110         }
3111         spin_unlock(&delayed_refs->lock);
3112
3113         spin_lock(&head->lock);
3114         /*
3115          * XXX: We should replace this with a proper search function in the
3116          * future.
3117          */
3118         for (node = rb_first_cached(&head->ref_tree); node;
3119              node = rb_next(node)) {
3120                 ref = rb_entry(node, struct btrfs_delayed_ref_node, ref_node);
3121                 /* If it's a shared ref we know a cross reference exists */
3122                 if (ref->type != BTRFS_EXTENT_DATA_REF_KEY) {
3123                         ret = 1;
3124                         break;
3125                 }
3126
3127                 data_ref = btrfs_delayed_node_to_data_ref(ref);
3128
3129                 /*
3130                  * If our ref doesn't match the one we're currently looking at
3131                  * then we have a cross reference.
3132                  */
3133                 if (data_ref->root != root->root_key.objectid ||
3134                     data_ref->objectid != objectid ||
3135                     data_ref->offset != offset) {
3136                         ret = 1;
3137                         break;
3138                 }
3139         }
3140         spin_unlock(&head->lock);
3141         mutex_unlock(&head->mutex);
3142         btrfs_put_transaction(cur_trans);
3143         return ret;
3144 }
3145
3146 static noinline int check_committed_ref(struct btrfs_root *root,
3147                                         struct btrfs_path *path,
3148                                         u64 objectid, u64 offset, u64 bytenr)
3149 {
3150         struct btrfs_fs_info *fs_info = root->fs_info;
3151         struct btrfs_root *extent_root = fs_info->extent_root;
3152         struct extent_buffer *leaf;
3153         struct btrfs_extent_data_ref *ref;
3154         struct btrfs_extent_inline_ref *iref;
3155         struct btrfs_extent_item *ei;
3156         struct btrfs_key key;
3157         u32 item_size;
3158         int type;
3159         int ret;
3160
3161         key.objectid = bytenr;
3162         key.offset = (u64)-1;
3163         key.type = BTRFS_EXTENT_ITEM_KEY;
3164
3165         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
3166         if (ret < 0)
3167                 goto out;
3168         BUG_ON(ret == 0); /* Corruption */
3169
3170         ret = -ENOENT;
3171         if (path->slots[0] == 0)
3172                 goto out;
3173
3174         path->slots[0]--;
3175         leaf = path->nodes[0];
3176         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
3177
3178         if (key.objectid != bytenr || key.type != BTRFS_EXTENT_ITEM_KEY)
3179                 goto out;
3180
3181         ret = 1;
3182         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
3183         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
3184
3185         if (item_size != sizeof(*ei) +
3186             btrfs_extent_inline_ref_size(BTRFS_EXTENT_DATA_REF_KEY))
3187                 goto out;
3188
3189         if (btrfs_extent_generation(leaf, ei) <=
3190             btrfs_root_last_snapshot(&root->root_item))
3191                 goto out;
3192
3193         iref = (struct btrfs_extent_inline_ref *)(ei + 1);
3194
3195         type = btrfs_get_extent_inline_ref_type(leaf, iref, BTRFS_REF_TYPE_DATA);
3196         if (type != BTRFS_EXTENT_DATA_REF_KEY)
3197                 goto out;
3198
3199         ref = (struct btrfs_extent_data_ref *)(&iref->offset);
3200         if (btrfs_extent_refs(leaf, ei) !=
3201             btrfs_extent_data_ref_count(leaf, ref) ||
3202             btrfs_extent_data_ref_root(leaf, ref) !=
3203             root->root_key.objectid ||
3204             btrfs_extent_data_ref_objectid(leaf, ref) != objectid ||
3205             btrfs_extent_data_ref_offset(leaf, ref) != offset)
3206                 goto out;
3207
3208         ret = 0;
3209 out:
3210         return ret;
3211 }
3212
3213 int btrfs_cross_ref_exist(struct btrfs_root *root, u64 objectid, u64 offset,
3214                           u64 bytenr)
3215 {
3216         struct btrfs_path *path;
3217         int ret;
3218
3219         path = btrfs_alloc_path();
3220         if (!path)
3221                 return -ENOMEM;
3222
3223         do {
3224                 ret = check_committed_ref(root, path, objectid,
3225                                           offset, bytenr);
3226                 if (ret && ret != -ENOENT)
3227                         goto out;
3228
3229                 ret = check_delayed_ref(root, path, objectid, offset, bytenr);
3230         } while (ret == -EAGAIN);
3231
3232 out:
3233         btrfs_free_path(path);
3234         if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID)
3235                 WARN_ON(ret > 0);
3236         return ret;
3237 }
3238
3239 static int __btrfs_mod_ref(struct btrfs_trans_handle *trans,
3240                            struct btrfs_root *root,
3241                            struct extent_buffer *buf,
3242                            int full_backref, int inc)
3243 {
3244         struct btrfs_fs_info *fs_info = root->fs_info;
3245         u64 bytenr;
3246         u64 num_bytes;
3247         u64 parent;
3248         u64 ref_root;
3249         u32 nritems;
3250         struct btrfs_key key;
3251         struct btrfs_file_extent_item *fi;
3252         int i;
3253         int level;
3254         int ret = 0;
3255         int (*process_func)(struct btrfs_trans_handle *,
3256                             struct btrfs_root *,
3257                             u64, u64, u64, u64, u64, u64);
3258
3259
3260         if (btrfs_is_testing(fs_info))
3261                 return 0;
3262
3263         ref_root = btrfs_header_owner(buf);
3264         nritems = btrfs_header_nritems(buf);
3265         level = btrfs_header_level(buf);
3266
3267         if (!test_bit(BTRFS_ROOT_REF_COWS, &root->state) && level == 0)
3268                 return 0;
3269
3270         if (inc)
3271                 process_func = btrfs_inc_extent_ref;
3272         else
3273                 process_func = btrfs_free_extent;
3274
3275         if (full_backref)
3276                 parent = buf->start;
3277         else
3278                 parent = 0;
3279
3280         for (i = 0; i < nritems; i++) {
3281                 if (level == 0) {
3282                         btrfs_item_key_to_cpu(buf, &key, i);
3283                         if (key.type != BTRFS_EXTENT_DATA_KEY)
3284                                 continue;
3285                         fi = btrfs_item_ptr(buf, i,
3286                                             struct btrfs_file_extent_item);
3287                         if (btrfs_file_extent_type(buf, fi) ==
3288                             BTRFS_FILE_EXTENT_INLINE)
3289                                 continue;
3290                         bytenr = btrfs_file_extent_disk_bytenr(buf, fi);
3291                         if (bytenr == 0)
3292                                 continue;
3293
3294                         num_bytes = btrfs_file_extent_disk_num_bytes(buf, fi);
3295                         key.offset -= btrfs_file_extent_offset(buf, fi);
3296                         ret = process_func(trans, root, bytenr, num_bytes,
3297                                            parent, ref_root, key.objectid,
3298                                            key.offset);
3299                         if (ret)
3300                                 goto fail;
3301                 } else {
3302                         bytenr = btrfs_node_blockptr(buf, i);
3303                         num_bytes = fs_info->nodesize;
3304                         ret = process_func(trans, root, bytenr, num_bytes,
3305                                            parent, ref_root, level - 1, 0);
3306                         if (ret)
3307                                 goto fail;
3308                 }
3309         }
3310         return 0;
3311 fail:
3312         return ret;
3313 }
3314
3315 int btrfs_inc_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
3316                   struct extent_buffer *buf, int full_backref)
3317 {
3318         return __btrfs_mod_ref(trans, root, buf, full_backref, 1);
3319 }
3320
3321 int btrfs_dec_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
3322                   struct extent_buffer *buf, int full_backref)
3323 {
3324         return __btrfs_mod_ref(trans, root, buf, full_backref, 0);
3325 }
3326
3327 static int write_one_cache_group(struct btrfs_trans_handle *trans,
3328                                  struct btrfs_fs_info *fs_info,
3329                                  struct btrfs_path *path,
3330                                  struct btrfs_block_group_cache *cache)
3331 {
3332         int ret;
3333         struct btrfs_root *extent_root = fs_info->extent_root;
3334         unsigned long bi;
3335         struct extent_buffer *leaf;
3336
3337         ret = btrfs_search_slot(trans, extent_root, &cache->key, path, 0, 1);
3338         if (ret) {
3339                 if (ret > 0)
3340                         ret = -ENOENT;
3341                 goto fail;
3342         }
3343
3344         leaf = path->nodes[0];
3345         bi = btrfs_item_ptr_offset(leaf, path->slots[0]);
3346         write_extent_buffer(leaf, &cache->item, bi, sizeof(cache->item));
3347         btrfs_mark_buffer_dirty(leaf);
3348 fail:
3349         btrfs_release_path(path);
3350         return ret;
3351
3352 }
3353
3354 static struct btrfs_block_group_cache *
3355 next_block_group(struct btrfs_fs_info *fs_info,
3356                  struct btrfs_block_group_cache *cache)
3357 {
3358         struct rb_node *node;
3359
3360         spin_lock(&fs_info->block_group_cache_lock);
3361
3362         /* If our block group was removed, we need a full search. */
3363         if (RB_EMPTY_NODE(&cache->cache_node)) {
3364                 const u64 next_bytenr = cache->key.objectid + cache->key.offset;
3365
3366                 spin_unlock(&fs_info->block_group_cache_lock);
3367                 btrfs_put_block_group(cache);
3368                 cache = btrfs_lookup_first_block_group(fs_info, next_bytenr); return cache;
3369         }
3370         node = rb_next(&cache->cache_node);
3371         btrfs_put_block_group(cache);
3372         if (node) {
3373                 cache = rb_entry(node, struct btrfs_block_group_cache,
3374                                  cache_node);
3375                 btrfs_get_block_group(cache);
3376         } else
3377                 cache = NULL;
3378         spin_unlock(&fs_info->block_group_cache_lock);
3379         return cache;
3380 }
3381
3382 static int cache_save_setup(struct btrfs_block_group_cache *block_group,
3383                             struct btrfs_trans_handle *trans,
3384                             struct btrfs_path *path)
3385 {
3386         struct btrfs_fs_info *fs_info = block_group->fs_info;
3387         struct btrfs_root *root = fs_info->tree_root;
3388         struct inode *inode = NULL;
3389         struct extent_changeset *data_reserved = NULL;
3390         u64 alloc_hint = 0;
3391         int dcs = BTRFS_DC_ERROR;
3392         u64 num_pages = 0;
3393         int retries = 0;
3394         int ret = 0;
3395
3396         /*
3397          * If this block group is smaller than 100 megs don't bother caching the
3398          * block group.
3399          */
3400         if (block_group->key.offset < (100 * SZ_1M)) {
3401                 spin_lock(&block_group->lock);
3402                 block_group->disk_cache_state = BTRFS_DC_WRITTEN;
3403                 spin_unlock(&block_group->lock);
3404                 return 0;
3405         }
3406
3407         if (trans->aborted)
3408                 return 0;
3409 again:
3410         inode = lookup_free_space_inode(fs_info, block_group, path);
3411         if (IS_ERR(inode) && PTR_ERR(inode) != -ENOENT) {
3412                 ret = PTR_ERR(inode);
3413                 btrfs_release_path(path);
3414                 goto out;
3415         }
3416
3417         if (IS_ERR(inode)) {
3418                 BUG_ON(retries);
3419                 retries++;
3420
3421                 if (block_group->ro)
3422                         goto out_free;
3423
3424                 ret = create_free_space_inode(fs_info, trans, block_group,
3425                                               path);
3426                 if (ret)
3427                         goto out_free;
3428                 goto again;
3429         }
3430
3431         /*
3432          * We want to set the generation to 0, that way if anything goes wrong
3433          * from here on out we know not to trust this cache when we load up next
3434          * time.
3435          */
3436         BTRFS_I(inode)->generation = 0;
3437         ret = btrfs_update_inode(trans, root, inode);
3438         if (ret) {
3439                 /*
3440                  * So theoretically we could recover from this, simply set the
3441                  * super cache generation to 0 so we know to invalidate the
3442                  * cache, but then we'd have to keep track of the block groups
3443                  * that fail this way so we know we _have_ to reset this cache
3444                  * before the next commit or risk reading stale cache.  So to
3445                  * limit our exposure to horrible edge cases lets just abort the
3446                  * transaction, this only happens in really bad situations
3447                  * anyway.
3448                  */
3449                 btrfs_abort_transaction(trans, ret);
3450                 goto out_put;
3451         }
3452         WARN_ON(ret);
3453
3454         /* We've already setup this transaction, go ahead and exit */
3455         if (block_group->cache_generation == trans->transid &&
3456             i_size_read(inode)) {
3457                 dcs = BTRFS_DC_SETUP;
3458                 goto out_put;
3459         }
3460
3461         if (i_size_read(inode) > 0) {
3462                 ret = btrfs_check_trunc_cache_free_space(fs_info,
3463                                         &fs_info->global_block_rsv);
3464                 if (ret)
3465                         goto out_put;
3466
3467                 ret = btrfs_truncate_free_space_cache(trans, NULL, inode);
3468                 if (ret)
3469                         goto out_put;
3470         }
3471
3472         spin_lock(&block_group->lock);
3473         if (block_group->cached != BTRFS_CACHE_FINISHED ||
3474             !btrfs_test_opt(fs_info, SPACE_CACHE)) {
3475                 /*
3476                  * don't bother trying to write stuff out _if_
3477                  * a) we're not cached,
3478                  * b) we're with nospace_cache mount option,
3479                  * c) we're with v2 space_cache (FREE_SPACE_TREE).
3480                  */
3481                 dcs = BTRFS_DC_WRITTEN;
3482                 spin_unlock(&block_group->lock);
3483                 goto out_put;
3484         }
3485         spin_unlock(&block_group->lock);
3486
3487         /*
3488          * We hit an ENOSPC when setting up the cache in this transaction, just
3489          * skip doing the setup, we've already cleared the cache so we're safe.
3490          */
3491         if (test_bit(BTRFS_TRANS_CACHE_ENOSPC, &trans->transaction->flags)) {
3492                 ret = -ENOSPC;
3493                 goto out_put;
3494         }
3495
3496         /*
3497          * Try to preallocate enough space based on how big the block group is.
3498          * Keep in mind this has to include any pinned space which could end up
3499          * taking up quite a bit since it's not folded into the other space
3500          * cache.
3501          */
3502         num_pages = div_u64(block_group->key.offset, SZ_256M);
3503         if (!num_pages)
3504                 num_pages = 1;
3505
3506         num_pages *= 16;
3507         num_pages *= PAGE_SIZE;
3508
3509         ret = btrfs_check_data_free_space(inode, &data_reserved, 0, num_pages);
3510         if (ret)
3511                 goto out_put;
3512
3513         ret = btrfs_prealloc_file_range_trans(inode, trans, 0, 0, num_pages,
3514                                               num_pages, num_pages,
3515                                               &alloc_hint);
3516         /*
3517          * Our cache requires contiguous chunks so that we don't modify a bunch
3518          * of metadata or split extents when writing the cache out, which means
3519          * we can enospc if we are heavily fragmented in addition to just normal
3520          * out of space conditions.  So if we hit this just skip setting up any
3521          * other block groups for this transaction, maybe we'll unpin enough
3522          * space the next time around.
3523          */
3524         if (!ret)
3525                 dcs = BTRFS_DC_SETUP;
3526         else if (ret == -ENOSPC)
3527                 set_bit(BTRFS_TRANS_CACHE_ENOSPC, &trans->transaction->flags);
3528
3529 out_put:
3530         iput(inode);
3531 out_free:
3532         btrfs_release_path(path);
3533 out:
3534         spin_lock(&block_group->lock);
3535         if (!ret && dcs == BTRFS_DC_SETUP)
3536                 block_group->cache_generation = trans->transid;
3537         block_group->disk_cache_state = dcs;
3538         spin_unlock(&block_group->lock);
3539
3540         extent_changeset_free(data_reserved);
3541         return ret;
3542 }
3543
3544 int btrfs_setup_space_cache(struct btrfs_trans_handle *trans,
3545                             struct btrfs_fs_info *fs_info)
3546 {
3547         struct btrfs_block_group_cache *cache, *tmp;
3548         struct btrfs_transaction *cur_trans = trans->transaction;
3549         struct btrfs_path *path;
3550
3551         if (list_empty(&cur_trans->dirty_bgs) ||
3552             !btrfs_test_opt(fs_info, SPACE_CACHE))
3553                 return 0;
3554
3555         path = btrfs_alloc_path();
3556         if (!path)
3557                 return -ENOMEM;
3558
3559         /* Could add new block groups, use _safe just in case */
3560         list_for_each_entry_safe(cache, tmp, &cur_trans->dirty_bgs,
3561                                  dirty_list) {
3562                 if (cache->disk_cache_state == BTRFS_DC_CLEAR)
3563                         cache_save_setup(cache, trans, path);
3564         }
3565
3566         btrfs_free_path(path);
3567         return 0;
3568 }
3569
3570 /*
3571  * transaction commit does final block group cache writeback during a
3572  * critical section where nothing is allowed to change the FS.  This is
3573  * required in order for the cache to actually match the block group,
3574  * but can introduce a lot of latency into the commit.
3575  *
3576  * So, btrfs_start_dirty_block_groups is here to kick off block group
3577  * cache IO.  There's a chance we'll have to redo some of it if the
3578  * block group changes again during the commit, but it greatly reduces
3579  * the commit latency by getting rid of the easy block groups while
3580  * we're still allowing others to join the commit.
3581  */
3582 int btrfs_start_dirty_block_groups(struct btrfs_trans_handle *trans)
3583 {
3584         struct btrfs_fs_info *fs_info = trans->fs_info;
3585         struct btrfs_block_group_cache *cache;
3586         struct btrfs_transaction *cur_trans = trans->transaction;
3587         int ret = 0;
3588         int should_put;
3589         struct btrfs_path *path = NULL;
3590         LIST_HEAD(dirty);
3591         struct list_head *io = &cur_trans->io_bgs;
3592         int num_started = 0;
3593         int loops = 0;
3594
3595         spin_lock(&cur_trans->dirty_bgs_lock);
3596         if (list_empty(&cur_trans->dirty_bgs)) {
3597                 spin_unlock(&cur_trans->dirty_bgs_lock);
3598                 return 0;
3599         }
3600         list_splice_init(&cur_trans->dirty_bgs, &dirty);
3601         spin_unlock(&cur_trans->dirty_bgs_lock);
3602
3603 again:
3604         /*
3605          * make sure all the block groups on our dirty list actually
3606          * exist
3607          */
3608         btrfs_create_pending_block_groups(trans);
3609
3610         if (!path) {
3611                 path = btrfs_alloc_path();
3612                 if (!path)
3613                         return -ENOMEM;
3614         }
3615
3616         /*
3617          * cache_write_mutex is here only to save us from balance or automatic
3618          * removal of empty block groups deleting this block group while we are
3619          * writing out the cache
3620          */
3621         mutex_lock(&trans->transaction->cache_write_mutex);
3622         while (!list_empty(&dirty)) {
3623                 bool drop_reserve = true;
3624
3625                 cache = list_first_entry(&dirty,
3626                                          struct btrfs_block_group_cache,
3627                                          dirty_list);
3628                 /*
3629                  * this can happen if something re-dirties a block
3630                  * group that is already under IO.  Just wait for it to
3631                  * finish and then do it all again
3632                  */
3633                 if (!list_empty(&cache->io_list)) {
3634                         list_del_init(&cache->io_list);
3635                         btrfs_wait_cache_io(trans, cache, path);
3636                         btrfs_put_block_group(cache);
3637                 }
3638
3639
3640                 /*
3641                  * btrfs_wait_cache_io uses the cache->dirty_list to decide
3642                  * if it should update the cache_state.  Don't delete
3643                  * until after we wait.
3644                  *
3645                  * Since we're not running in the commit critical section
3646                  * we need the dirty_bgs_lock to protect from update_block_group
3647                  */
3648                 spin_lock(&cur_trans->dirty_bgs_lock);
3649                 list_del_init(&cache->dirty_list);
3650                 spin_unlock(&cur_trans->dirty_bgs_lock);
3651
3652                 should_put = 1;
3653
3654                 cache_save_setup(cache, trans, path);
3655
3656                 if (cache->disk_cache_state == BTRFS_DC_SETUP) {
3657                         cache->io_ctl.inode = NULL;
3658                         ret = btrfs_write_out_cache(fs_info, trans,
3659                                                     cache, path);
3660                         if (ret == 0 && cache->io_ctl.inode) {
3661                                 num_started++;
3662                                 should_put = 0;
3663
3664                                 /*
3665                                  * The cache_write_mutex is protecting the
3666                                  * io_list, also refer to the definition of
3667                                  * btrfs_transaction::io_bgs for more details
3668                                  */
3669                                 list_add_tail(&cache->io_list, io);
3670                         } else {
3671                                 /*
3672                                  * if we failed to write the cache, the
3673                                  * generation will be bad and life goes on
3674                                  */
3675                                 ret = 0;
3676                         }
3677                 }
3678                 if (!ret) {
3679                         ret = write_one_cache_group(trans, fs_info,
3680                                                     path, cache);
3681                         /*
3682                          * Our block group might still be attached to the list
3683                          * of new block groups in the transaction handle of some
3684                          * other task (struct btrfs_trans_handle->new_bgs). This
3685                          * means its block group item isn't yet in the extent
3686                          * tree. If this happens ignore the error, as we will
3687                          * try again later in the critical section of the
3688                          * transaction commit.
3689                          */
3690                         if (ret == -ENOENT) {
3691                                 ret = 0;
3692                                 spin_lock(&cur_trans->dirty_bgs_lock);
3693                                 if (list_empty(&cache->dirty_list)) {
3694                                         list_add_tail(&cache->dirty_list,
3695                                                       &cur_trans->dirty_bgs);
3696                                         btrfs_get_block_group(cache);
3697                                         drop_reserve = false;
3698                                 }
3699                                 spin_unlock(&cur_trans->dirty_bgs_lock);
3700                         } else if (ret) {
3701                                 btrfs_abort_transaction(trans, ret);
3702                         }
3703                 }
3704
3705                 /* if it's not on the io list, we need to put the block group */
3706                 if (should_put)
3707                         btrfs_put_block_group(cache);
3708                 if (drop_reserve)
3709                         btrfs_delayed_refs_rsv_release(fs_info, 1);
3710
3711                 if (ret)
3712                         break;
3713
3714                 /*
3715                  * Avoid blocking other tasks for too long. It might even save
3716                  * us from writing caches for block groups that are going to be
3717                  * removed.
3718                  */
3719                 mutex_unlock(&trans->transaction->cache_write_mutex);
3720                 mutex_lock(&trans->transaction->cache_write_mutex);
3721         }
3722         mutex_unlock(&trans->transaction->cache_write_mutex);
3723
3724         /*
3725          * go through delayed refs for all the stuff we've just kicked off
3726          * and then loop back (just once)
3727          */
3728         ret = btrfs_run_delayed_refs(trans, 0);
3729         if (!ret && loops == 0) {
3730                 loops++;
3731                 spin_lock(&cur_trans->dirty_bgs_lock);
3732                 list_splice_init(&cur_trans->dirty_bgs, &dirty);
3733                 /*
3734                  * dirty_bgs_lock protects us from concurrent block group
3735                  * deletes too (not just cache_write_mutex).
3736                  */
3737                 if (!list_empty(&dirty)) {
3738                         spin_unlock(&cur_trans->dirty_bgs_lock);
3739                         goto again;
3740                 }
3741                 spin_unlock(&cur_trans->dirty_bgs_lock);
3742         } else if (ret < 0) {
3743                 btrfs_cleanup_dirty_bgs(cur_trans, fs_info);
3744         }
3745
3746         btrfs_free_path(path);
3747         return ret;
3748 }
3749
3750 int btrfs_write_dirty_block_groups(struct btrfs_trans_handle *trans,
3751                                    struct btrfs_fs_info *fs_info)
3752 {
3753         struct btrfs_block_group_cache *cache;
3754         struct btrfs_transaction *cur_trans = trans->transaction;
3755         int ret = 0;
3756         int should_put;
3757         struct btrfs_path *path;
3758         struct list_head *io = &cur_trans->io_bgs;
3759         int num_started = 0;
3760
3761         path = btrfs_alloc_path();
3762         if (!path)
3763                 return -ENOMEM;
3764
3765         /*
3766          * Even though we are in the critical section of the transaction commit,
3767          * we can still have concurrent tasks adding elements to this
3768          * transaction's list of dirty block groups. These tasks correspond to
3769          * endio free space workers started when writeback finishes for a
3770          * space cache, which run inode.c:btrfs_finish_ordered_io(), and can
3771          * allocate new block groups as a result of COWing nodes of the root
3772          * tree when updating the free space inode. The writeback for the space
3773          * caches is triggered by an earlier call to
3774          * btrfs_start_dirty_block_groups() and iterations of the following
3775          * loop.
3776          * Also we want to do the cache_save_setup first and then run the
3777          * delayed refs to make sure we have the best chance at doing this all
3778          * in one shot.
3779          */
3780         spin_lock(&cur_trans->dirty_bgs_lock);
3781         while (!list_empty(&cur_trans->dirty_bgs)) {
3782                 cache = list_first_entry(&cur_trans->dirty_bgs,
3783                                          struct btrfs_block_group_cache,
3784                                          dirty_list);
3785
3786                 /*
3787                  * this can happen if cache_save_setup re-dirties a block
3788                  * group that is already under IO.  Just wait for it to
3789                  * finish and then do it all again
3790                  */
3791                 if (!list_empty(&cache->io_list)) {
3792                         spin_unlock(&cur_trans->dirty_bgs_lock);
3793                         list_del_init(&cache->io_list);
3794                         btrfs_wait_cache_io(trans, cache, path);
3795                         btrfs_put_block_group(cache);
3796                         spin_lock(&cur_trans->dirty_bgs_lock);
3797                 }
3798
3799                 /*
3800                  * don't remove from the dirty list until after we've waited
3801                  * on any pending IO
3802                  */
3803                 list_del_init(&cache->dirty_list);
3804                 spin_unlock(&cur_trans->dirty_bgs_lock);
3805                 should_put = 1;
3806
3807                 cache_save_setup(cache, trans, path);
3808
3809                 if (!ret)
3810                         ret = btrfs_run_delayed_refs(trans,
3811                                                      (unsigned long) -1);
3812
3813                 if (!ret && cache->disk_cache_state == BTRFS_DC_SETUP) {
3814                         cache->io_ctl.inode = NULL;
3815                         ret = btrfs_write_out_cache(fs_info, trans,
3816                                                     cache, path);
3817                         if (ret == 0 && cache->io_ctl.inode) {
3818                                 num_started++;
3819                                 should_put = 0;
3820                                 list_add_tail(&cache->io_list, io);
3821                         } else {
3822                                 /*
3823                                  * if we failed to write the cache, the
3824                                  * generation will be bad and life goes on
3825                                  */
3826                                 ret = 0;
3827                         }
3828                 }
3829                 if (!ret) {
3830                         ret = write_one_cache_group(trans, fs_info,
3831                                                     path, cache);
3832                         /*
3833                          * One of the free space endio workers might have
3834                          * created a new block group while updating a free space
3835                          * cache's inode (at inode.c:btrfs_finish_ordered_io())
3836                          * and hasn't released its transaction handle yet, in
3837                          * which case the new block group is still attached to
3838                          * its transaction handle and its creation has not
3839                          * finished yet (no block group item in the extent tree
3840                          * yet, etc). If this is the case, wait for all free
3841                          * space endio workers to finish and retry. This is a
3842                          * a very rare case so no need for a more efficient and
3843                          * complex approach.
3844                          */
3845                         if (ret == -ENOENT) {
3846                                 wait_event(cur_trans->writer_wait,
3847                                    atomic_read(&cur_trans->num_writers) == 1);
3848                                 ret = write_one_cache_group(trans, fs_info,
3849                                                             path, cache);
3850                         }
3851                         if (ret)
3852                                 btrfs_abort_transaction(trans, ret);
3853                 }
3854
3855                 /* if its not on the io list, we need to put the block group */
3856                 if (should_put)
3857                         btrfs_put_block_group(cache);
3858                 btrfs_delayed_refs_rsv_release(fs_info, 1);
3859                 spin_lock(&cur_trans->dirty_bgs_lock);
3860         }
3861         spin_unlock(&cur_trans->dirty_bgs_lock);
3862
3863         /*
3864          * Refer to the definition of io_bgs member for details why it's safe
3865          * to use it without any locking
3866          */
3867         while (!list_empty(io)) {
3868                 cache = list_first_entry(io, struct btrfs_block_group_cache,
3869                                          io_list);
3870                 list_del_init(&cache->io_list);
3871                 btrfs_wait_cache_io(trans, cache, path);
3872                 btrfs_put_block_group(cache);
3873         }
3874
3875         btrfs_free_path(path);
3876         return ret;
3877 }
3878
3879 int btrfs_extent_readonly(struct btrfs_fs_info *fs_info, u64 bytenr)
3880 {
3881         struct btrfs_block_group_cache *block_group;
3882         int readonly = 0;
3883
3884         block_group = btrfs_lookup_block_group(fs_info, bytenr);
3885         if (!block_group || block_group->ro)
3886                 readonly = 1;
3887         if (block_group)
3888                 btrfs_put_block_group(block_group);
3889         return readonly;
3890 }
3891
3892 bool btrfs_inc_nocow_writers(struct btrfs_fs_info *fs_info, u64 bytenr)
3893 {
3894         struct btrfs_block_group_cache *bg;
3895         bool ret = true;
3896
3897         bg = btrfs_lookup_block_group(fs_info, bytenr);
3898         if (!bg)
3899                 return false;
3900
3901         spin_lock(&bg->lock);
3902         if (bg->ro)
3903                 ret = false;
3904         else
3905                 atomic_inc(&bg->nocow_writers);
3906         spin_unlock(&bg->lock);
3907
3908         /* no put on block group, done by btrfs_dec_nocow_writers */
3909         if (!ret)
3910                 btrfs_put_block_group(bg);
3911
3912         return ret;
3913
3914 }
3915
3916 void btrfs_dec_nocow_writers(struct btrfs_fs_info *fs_info, u64 bytenr)
3917 {
3918         struct btrfs_block_group_cache *bg;
3919
3920         bg = btrfs_lookup_block_group(fs_info, bytenr);
3921         ASSERT(bg);
3922         if (atomic_dec_and_test(&bg->nocow_writers))
3923                 wake_up_var(&bg->nocow_writers);
3924         /*
3925          * Once for our lookup and once for the lookup done by a previous call
3926          * to btrfs_inc_nocow_writers()
3927          */
3928         btrfs_put_block_group(bg);
3929         btrfs_put_block_group(bg);
3930 }
3931
3932 void btrfs_wait_nocow_writers(struct btrfs_block_group_cache *bg)
3933 {
3934         wait_var_event(&bg->nocow_writers, !atomic_read(&bg->nocow_writers));
3935 }
3936
3937 static const char *alloc_name(u64 flags)
3938 {
3939         switch (flags) {
3940         case BTRFS_BLOCK_GROUP_METADATA|BTRFS_BLOCK_GROUP_DATA:
3941                 return "mixed";
3942         case BTRFS_BLOCK_GROUP_METADATA:
3943                 return "metadata";
3944         case BTRFS_BLOCK_GROUP_DATA:
3945                 return "data";
3946         case BTRFS_BLOCK_GROUP_SYSTEM:
3947                 return "system";
3948         default:
3949                 WARN_ON(1);
3950                 return "invalid-combination";
3951         };
3952 }
3953
3954 static int create_space_info(struct btrfs_fs_info *info, u64 flags)
3955 {
3956
3957         struct btrfs_space_info *space_info;
3958         int i;
3959         int ret;
3960
3961         space_info = kzalloc(sizeof(*space_info), GFP_NOFS);
3962         if (!space_info)
3963                 return -ENOMEM;
3964
3965         ret = percpu_counter_init(&space_info->total_bytes_pinned, 0,
3966                                  GFP_KERNEL);
3967         if (ret) {
3968                 kfree(space_info);
3969                 return ret;
3970         }
3971
3972         for (i = 0; i < BTRFS_NR_RAID_TYPES; i++)
3973                 INIT_LIST_HEAD(&space_info->block_groups[i]);
3974         init_rwsem(&space_info->groups_sem);
3975         spin_lock_init(&space_info->lock);
3976         space_info->flags = flags & BTRFS_BLOCK_GROUP_TYPE_MASK;
3977         space_info->force_alloc = CHUNK_ALLOC_NO_FORCE;
3978         init_waitqueue_head(&space_info->wait);
3979         INIT_LIST_HEAD(&space_info->ro_bgs);
3980         INIT_LIST_HEAD(&space_info->tickets);
3981         INIT_LIST_HEAD(&space_info->priority_tickets);
3982
3983         ret = kobject_init_and_add(&space_info->kobj, &space_info_ktype,
3984                                     info->space_info_kobj, "%s",
3985                                     alloc_name(space_info->flags));
3986         if (ret) {
3987                 percpu_counter_destroy(&space_info->total_bytes_pinned);
3988                 kfree(space_info);
3989                 return ret;
3990         }
3991
3992         list_add_rcu(&space_info->list, &info->space_info);
3993         if (flags & BTRFS_BLOCK_GROUP_DATA)
3994                 info->data_sinfo = space_info;
3995
3996         return ret;
3997 }
3998
3999 static void update_space_info(struct btrfs_fs_info *info, u64 flags,
4000                              u64 total_bytes, u64 bytes_used,
4001                              u64 bytes_readonly,
4002                              struct btrfs_space_info **space_info)
4003 {
4004         struct btrfs_space_info *found;
4005         int factor;
4006
4007         factor = btrfs_bg_type_to_factor(flags);
4008
4009         found = __find_space_info(info, flags);
4010         ASSERT(found);
4011         spin_lock(&found->lock);
4012         found->total_bytes += total_bytes;
4013         found->disk_total += total_bytes * factor;
4014         found->bytes_used += bytes_used;
4015         found->disk_used += bytes_used * factor;
4016         found->bytes_readonly += bytes_readonly;
4017         if (total_bytes > 0)
4018                 found->full = 0;
4019         space_info_add_new_bytes(info, found, total_bytes -
4020                                  bytes_used - bytes_readonly);
4021         spin_unlock(&found->lock);
4022         *space_info = found;
4023 }
4024
4025 static void set_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
4026 {
4027         u64 extra_flags = chunk_to_extended(flags) &
4028                                 BTRFS_EXTENDED_PROFILE_MASK;
4029
4030         write_seqlock(&fs_info->profiles_lock);
4031         if (flags & BTRFS_BLOCK_GROUP_DATA)
4032                 fs_info->avail_data_alloc_bits |= extra_flags;
4033         if (flags & BTRFS_BLOCK_GROUP_METADATA)
4034                 fs_info->avail_metadata_alloc_bits |= extra_flags;
4035         if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
4036                 fs_info->avail_system_alloc_bits |= extra_flags;
4037         write_sequnlock(&fs_info->profiles_lock);
4038 }
4039
4040 /*
4041  * returns target flags in extended format or 0 if restripe for this
4042  * chunk_type is not in progress
4043  *
4044  * should be called with balance_lock held
4045  */
4046 static u64 get_restripe_target(struct btrfs_fs_info *fs_info, u64 flags)
4047 {
4048         struct btrfs_balance_control *bctl = fs_info->balance_ctl;
4049         u64 target = 0;
4050
4051         if (!bctl)
4052                 return 0;
4053
4054         if (flags & BTRFS_BLOCK_GROUP_DATA &&
4055             bctl->data.flags & BTRFS_BALANCE_ARGS_CONVERT) {
4056                 target = BTRFS_BLOCK_GROUP_DATA | bctl->data.target;
4057         } else if (flags & BTRFS_BLOCK_GROUP_SYSTEM &&
4058                    bctl->sys.flags & BTRFS_BALANCE_ARGS_CONVERT) {
4059                 target = BTRFS_BLOCK_GROUP_SYSTEM | bctl->sys.target;
4060         } else if (flags & BTRFS_BLOCK_GROUP_METADATA &&
4061                    bctl->meta.flags & BTRFS_BALANCE_ARGS_CONVERT) {
4062                 target = BTRFS_BLOCK_GROUP_METADATA | bctl->meta.target;
4063         }
4064
4065         return target;
4066 }
4067
4068 /*
4069  * @flags: available profiles in extended format (see ctree.h)
4070  *
4071  * Returns reduced profile in chunk format.  If profile changing is in
4072  * progress (either running or paused) picks the target profile (if it's
4073  * already available), otherwise falls back to plain reducing.
4074  */
4075 static u64 btrfs_reduce_alloc_profile(struct btrfs_fs_info *fs_info, u64 flags)
4076 {
4077         u64 num_devices = fs_info->fs_devices->rw_devices;
4078         u64 target;
4079         u64 raid_type;
4080         u64 allowed = 0;
4081
4082         /*
4083          * see if restripe for this chunk_type is in progress, if so
4084          * try to reduce to the target profile
4085          */
4086         spin_lock(&fs_info->balance_lock);
4087         target = get_restripe_target(fs_info, flags);
4088         if (target) {
4089                 /* pick target profile only if it's already available */
4090                 if ((flags & target) & BTRFS_EXTENDED_PROFILE_MASK) {
4091                         spin_unlock(&fs_info->balance_lock);
4092                         return extended_to_chunk(target);
4093                 }
4094         }
4095         spin_unlock(&fs_info->balance_lock);
4096
4097         /* First, mask out the RAID levels which aren't possible */
4098         for (raid_type = 0; raid_type < BTRFS_NR_RAID_TYPES; raid_type++) {
4099                 if (num_devices >= btrfs_raid_array[raid_type].devs_min)
4100                         allowed |= btrfs_raid_array[raid_type].bg_flag;
4101         }
4102         allowed &= flags;
4103
4104         if (allowed & BTRFS_BLOCK_GROUP_RAID6)
4105                 allowed = BTRFS_BLOCK_GROUP_RAID6;
4106         else if (allowed & BTRFS_BLOCK_GROUP_RAID5)
4107                 allowed = BTRFS_BLOCK_GROUP_RAID5;
4108         else if (allowed & BTRFS_BLOCK_GROUP_RAID10)
4109                 allowed = BTRFS_BLOCK_GROUP_RAID10;
4110         else if (allowed & BTRFS_BLOCK_GROUP_RAID1)
4111                 allowed = BTRFS_BLOCK_GROUP_RAID1;
4112         else if (allowed & BTRFS_BLOCK_GROUP_RAID0)
4113                 allowed = BTRFS_BLOCK_GROUP_RAID0;
4114
4115         flags &= ~BTRFS_BLOCK_GROUP_PROFILE_MASK;
4116
4117         return extended_to_chunk(flags | allowed);
4118 }
4119
4120 static u64 get_alloc_profile(struct btrfs_fs_info *fs_info, u64 orig_flags)
4121 {
4122         unsigned seq;
4123         u64 flags;
4124
4125         do {
4126                 flags = orig_flags;
4127                 seq = read_seqbegin(&fs_info->profiles_lock);
4128
4129                 if (flags & BTRFS_BLOCK_GROUP_DATA)
4130                         flags |= fs_info->avail_data_alloc_bits;
4131                 else if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
4132                         flags |= fs_info->avail_system_alloc_bits;
4133                 else if (flags & BTRFS_BLOCK_GROUP_METADATA)
4134                         flags |= fs_info->avail_metadata_alloc_bits;
4135         } while (read_seqretry(&fs_info->profiles_lock, seq));
4136
4137         return btrfs_reduce_alloc_profile(fs_info, flags);
4138 }
4139
4140 static u64 get_alloc_profile_by_root(struct btrfs_root *root, int data)
4141 {
4142         struct btrfs_fs_info *fs_info = root->fs_info;
4143         u64 flags;
4144         u64 ret;
4145
4146         if (data)
4147                 flags = BTRFS_BLOCK_GROUP_DATA;
4148         else if (root == fs_info->chunk_root)
4149                 flags = BTRFS_BLOCK_GROUP_SYSTEM;
4150         else
4151                 flags = BTRFS_BLOCK_GROUP_METADATA;
4152
4153         ret = get_alloc_profile(fs_info, flags);
4154         return ret;
4155 }
4156
4157 u64 btrfs_data_alloc_profile(struct btrfs_fs_info *fs_info)
4158 {
4159         return get_alloc_profile(fs_info, BTRFS_BLOCK_GROUP_DATA);
4160 }
4161
4162 u64 btrfs_metadata_alloc_profile(struct btrfs_fs_info *fs_info)
4163 {
4164         return get_alloc_profile(fs_info, BTRFS_BLOCK_GROUP_METADATA);
4165 }
4166
4167 u64 btrfs_system_alloc_profile(struct btrfs_fs_info *fs_info)
4168 {
4169         return get_alloc_profile(fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
4170 }
4171
4172 static u64 btrfs_space_info_used(struct btrfs_space_info *s_info,
4173                                  bool may_use_included)
4174 {
4175         ASSERT(s_info);
4176         return s_info->bytes_used + s_info->bytes_reserved +
4177                 s_info->bytes_pinned + s_info->bytes_readonly +
4178                 (may_use_included ? s_info->bytes_may_use : 0);
4179 }
4180
4181 int btrfs_alloc_data_chunk_ondemand(struct btrfs_inode *inode, u64 bytes)
4182 {
4183         struct btrfs_root *root = inode->root;
4184         struct btrfs_fs_info *fs_info = root->fs_info;
4185         struct btrfs_space_info *data_sinfo = fs_info->data_sinfo;
4186         u64 used;
4187         int ret = 0;
4188         int need_commit = 2;
4189         int have_pinned_space;
4190
4191         /* make sure bytes are sectorsize aligned */
4192         bytes = ALIGN(bytes, fs_info->sectorsize);
4193
4194         if (btrfs_is_free_space_inode(inode)) {
4195                 need_commit = 0;
4196                 ASSERT(current->journal_info);
4197         }
4198
4199 again:
4200         /* make sure we have enough space to handle the data first */
4201         spin_lock(&data_sinfo->lock);
4202         used = btrfs_space_info_used(data_sinfo, true);
4203
4204         if (used + bytes > data_sinfo->total_bytes) {
4205                 struct btrfs_trans_handle *trans;
4206
4207                 /*
4208                  * if we don't have enough free bytes in this space then we need
4209                  * to alloc a new chunk.
4210                  */
4211                 if (!data_sinfo->full) {
4212                         u64 alloc_target;
4213
4214                         data_sinfo->force_alloc = CHUNK_ALLOC_FORCE;
4215                         spin_unlock(&data_sinfo->lock);
4216
4217                         alloc_target = btrfs_data_alloc_profile(fs_info);
4218                         /*
4219                          * It is ugly that we don't call nolock join
4220                          * transaction for the free space inode case here.
4221                          * But it is safe because we only do the data space
4222                          * reservation for the free space cache in the
4223                          * transaction context, the common join transaction
4224                          * just increase the counter of the current transaction
4225                          * handler, doesn't try to acquire the trans_lock of
4226                          * the fs.
4227                          */
4228                         trans = btrfs_join_transaction(root);
4229                         if (IS_ERR(trans))
4230                                 return PTR_ERR(trans);
4231
4232                         ret = do_chunk_alloc(trans, alloc_target,
4233                                              CHUNK_ALLOC_NO_FORCE);
4234                         btrfs_end_transaction(trans);
4235                         if (ret < 0) {
4236                                 if (ret != -ENOSPC)
4237                                         return ret;
4238                                 else {
4239                                         have_pinned_space = 1;
4240                                         goto commit_trans;
4241                                 }
4242                         }
4243
4244                         goto again;
4245                 }
4246
4247                 /*
4248                  * If we don't have enough pinned space to deal with this
4249                  * allocation, and no removed chunk in current transaction,
4250                  * don't bother committing the transaction.
4251                  */
4252                 have_pinned_space = __percpu_counter_compare(
4253                         &data_sinfo->total_bytes_pinned,
4254                         used + bytes - data_sinfo->total_bytes,
4255                         BTRFS_TOTAL_BYTES_PINNED_BATCH);
4256                 spin_unlock(&data_sinfo->lock);
4257
4258                 /* commit the current transaction and try again */
4259 commit_trans:
4260                 if (need_commit) {
4261                         need_commit--;
4262
4263                         if (need_commit > 0) {
4264                                 btrfs_start_delalloc_roots(fs_info, -1);
4265                                 btrfs_wait_ordered_roots(fs_info, U64_MAX, 0,
4266                                                          (u64)-1);
4267                         }
4268
4269                         trans = btrfs_join_transaction(root);
4270                         if (IS_ERR(trans))
4271                                 return PTR_ERR(trans);
4272                         if (have_pinned_space >= 0 ||
4273                             test_bit(BTRFS_TRANS_HAVE_FREE_BGS,
4274                                      &trans->transaction->flags) ||
4275                             need_commit > 0) {
4276                                 ret = btrfs_commit_transaction(trans);
4277                                 if (ret)
4278                                         return ret;
4279                                 /*
4280                                  * The cleaner kthread might still be doing iput
4281                                  * operations. Wait for it to finish so that
4282                                  * more space is released.  We don't need to
4283                                  * explicitly run the delayed iputs here because
4284                                  * the commit_transaction would have woken up
4285                                  * the cleaner.
4286                                  */
4287                                 ret = btrfs_wait_on_delayed_iputs(fs_info);
4288                                 if (ret)
4289                                         return ret;
4290                                 goto again;
4291                         } else {
4292                                 btrfs_end_transaction(trans);
4293                         }
4294                 }
4295
4296                 trace_btrfs_space_reservation(fs_info,
4297                                               "space_info:enospc",
4298                                               data_sinfo->flags, bytes, 1);
4299                 return -ENOSPC;
4300         }
4301         update_bytes_may_use(data_sinfo, bytes);
4302         trace_btrfs_space_reservation(fs_info, "space_info",
4303                                       data_sinfo->flags, bytes, 1);
4304         spin_unlock(&data_sinfo->lock);
4305
4306         return 0;
4307 }
4308
4309 int btrfs_check_data_free_space(struct inode *inode,
4310                         struct extent_changeset **reserved, u64 start, u64 len)
4311 {
4312         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
4313         int ret;
4314
4315         /* align the range */
4316         len = round_up(start + len, fs_info->sectorsize) -
4317               round_down(start, fs_info->sectorsize);
4318         start = round_down(start, fs_info->sectorsize);
4319
4320         ret = btrfs_alloc_data_chunk_ondemand(BTRFS_I(inode), len);
4321         if (ret < 0)
4322                 return ret;
4323
4324         /* Use new btrfs_qgroup_reserve_data to reserve precious data space. */
4325         ret = btrfs_qgroup_reserve_data(inode, reserved, start, len);
4326         if (ret < 0)
4327                 btrfs_free_reserved_data_space_noquota(inode, start, len);
4328         else
4329                 ret = 0;
4330         return ret;
4331 }
4332
4333 /*
4334  * Called if we need to clear a data reservation for this inode
4335  * Normally in a error case.
4336  *
4337  * This one will *NOT* use accurate qgroup reserved space API, just for case
4338  * which we can't sleep and is sure it won't affect qgroup reserved space.
4339  * Like clear_bit_hook().
4340  */
4341 void btrfs_free_reserved_data_space_noquota(struct inode *inode, u64 start,
4342                                             u64 len)
4343 {
4344         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
4345         struct btrfs_space_info *data_sinfo;
4346
4347         /* Make sure the range is aligned to sectorsize */
4348         len = round_up(start + len, fs_info->sectorsize) -
4349               round_down(start, fs_info->sectorsize);
4350         start = round_down(start, fs_info->sectorsize);
4351
4352         data_sinfo = fs_info->data_sinfo;
4353         spin_lock(&data_sinfo->lock);
4354         update_bytes_may_use(data_sinfo, -len);
4355         trace_btrfs_space_reservation(fs_info, "space_info",
4356                                       data_sinfo->flags, len, 0);
4357         spin_unlock(&data_sinfo->lock);
4358 }
4359
4360 /*
4361  * Called if we need to clear a data reservation for this inode
4362  * Normally in a error case.
4363  *
4364  * This one will handle the per-inode data rsv map for accurate reserved
4365  * space framework.
4366  */
4367 void btrfs_free_reserved_data_space(struct inode *inode,
4368                         struct extent_changeset *reserved, u64 start, u64 len)
4369 {
4370         struct btrfs_root *root = BTRFS_I(inode)->root;
4371
4372         /* Make sure the range is aligned to sectorsize */
4373         len = round_up(start + len, root->fs_info->sectorsize) -
4374               round_down(start, root->fs_info->sectorsize);
4375         start = round_down(start, root->fs_info->sectorsize);
4376
4377         btrfs_free_reserved_data_space_noquota(inode, start, len);
4378         btrfs_qgroup_free_data(inode, reserved, start, len);
4379 }
4380
4381 static void force_metadata_allocation(struct btrfs_fs_info *info)
4382 {
4383         struct list_head *head = &info->space_info;
4384         struct btrfs_space_info *found;
4385
4386         rcu_read_lock();
4387         list_for_each_entry_rcu(found, head, list) {
4388                 if (found->flags & BTRFS_BLOCK_GROUP_METADATA)
4389                         found->force_alloc = CHUNK_ALLOC_FORCE;
4390         }
4391         rcu_read_unlock();
4392 }
4393
4394 static inline u64 calc_global_rsv_need_space(struct btrfs_block_rsv *global)
4395 {
4396         return (global->size << 1);
4397 }
4398
4399 static int should_alloc_chunk(struct btrfs_fs_info *fs_info,
4400                               struct btrfs_space_info *sinfo, int force)
4401 {
4402         u64 bytes_used = btrfs_space_info_used(sinfo, false);
4403         u64 thresh;
4404
4405         if (force == CHUNK_ALLOC_FORCE)
4406                 return 1;
4407
4408         /*
4409          * in limited mode, we want to have some free space up to
4410          * about 1% of the FS size.
4411          */
4412         if (force == CHUNK_ALLOC_LIMITED) {
4413                 thresh = btrfs_super_total_bytes(fs_info->super_copy);
4414                 thresh = max_t(u64, SZ_64M, div_factor_fine(thresh, 1));
4415
4416                 if (sinfo->total_bytes - bytes_used < thresh)
4417                         return 1;
4418         }
4419
4420         if (bytes_used + SZ_2M < div_factor(sinfo->total_bytes, 8))
4421                 return 0;
4422         return 1;
4423 }
4424
4425 static u64 get_profile_num_devs(struct btrfs_fs_info *fs_info, u64 type)
4426 {
4427         u64 num_dev;
4428
4429         if (type & (BTRFS_BLOCK_GROUP_RAID10 |
4430                     BTRFS_BLOCK_GROUP_RAID0 |
4431                     BTRFS_BLOCK_GROUP_RAID5 |
4432                     BTRFS_BLOCK_GROUP_RAID6))
4433                 num_dev = fs_info->fs_devices->rw_devices;
4434         else if (type & BTRFS_BLOCK_GROUP_RAID1)
4435                 num_dev = 2;
4436         else
4437                 num_dev = 1;    /* DUP or single */
4438
4439         return num_dev;
4440 }
4441
4442 /*
4443  * If @is_allocation is true, reserve space in the system space info necessary
4444  * for allocating a chunk, otherwise if it's false, reserve space necessary for
4445  * removing a chunk.
4446  */
4447 void check_system_chunk(struct btrfs_trans_handle *trans, u64 type)
4448 {
4449         struct btrfs_fs_info *fs_info = trans->fs_info;
4450         struct btrfs_space_info *info;
4451         u64 left;
4452         u64 thresh;
4453         int ret = 0;
4454         u64 num_devs;
4455
4456         /*
4457          * Needed because we can end up allocating a system chunk and for an
4458          * atomic and race free space reservation in the chunk block reserve.
4459          */
4460         lockdep_assert_held(&fs_info->chunk_mutex);
4461
4462         info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
4463         spin_lock(&info->lock);
4464         left = info->total_bytes - btrfs_space_info_used(info, true);
4465         spin_unlock(&info->lock);
4466
4467         num_devs = get_profile_num_devs(fs_info, type);
4468
4469         /* num_devs device items to update and 1 chunk item to add or remove */
4470         thresh = btrfs_calc_trunc_metadata_size(fs_info, num_devs) +
4471                 btrfs_calc_trans_metadata_size(fs_info, 1);
4472
4473         if (left < thresh && btrfs_test_opt(fs_info, ENOSPC_DEBUG)) {
4474                 btrfs_info(fs_info, "left=%llu, need=%llu, flags=%llu",
4475                            left, thresh, type);
4476                 dump_space_info(fs_info, info, 0, 0);
4477         }
4478
4479         if (left < thresh) {
4480                 u64 flags = btrfs_system_alloc_profile(fs_info);
4481
4482                 /*
4483                  * Ignore failure to create system chunk. We might end up not
4484                  * needing it, as we might not need to COW all nodes/leafs from
4485                  * the paths we visit in the chunk tree (they were already COWed
4486                  * or created in the current transaction for example).
4487                  */
4488                 ret = btrfs_alloc_chunk(trans, flags);
4489         }
4490
4491         if (!ret) {
4492                 ret = btrfs_block_rsv_add(fs_info->chunk_root,
4493                                           &fs_info->chunk_block_rsv,
4494                                           thresh, BTRFS_RESERVE_NO_FLUSH);
4495                 if (!ret)
4496                         trans->chunk_bytes_reserved += thresh;
4497         }
4498 }
4499
4500 /*
4501  * If force is CHUNK_ALLOC_FORCE:
4502  *    - return 1 if it successfully allocates a chunk,
4503  *    - return errors including -ENOSPC otherwise.
4504  * If force is NOT CHUNK_ALLOC_FORCE:
4505  *    - return 0 if it doesn't need to allocate a new chunk,
4506  *    - return 1 if it successfully allocates a chunk,
4507  *    - return errors including -ENOSPC otherwise.
4508  */
4509 static int do_chunk_alloc(struct btrfs_trans_handle *trans, u64 flags,
4510                           int force)
4511 {
4512         struct btrfs_fs_info *fs_info = trans->fs_info;
4513         struct btrfs_space_info *space_info;
4514         bool wait_for_alloc = false;
4515         bool should_alloc = false;
4516         int ret = 0;
4517
4518         /* Don't re-enter if we're already allocating a chunk */
4519         if (trans->allocating_chunk)
4520                 return -ENOSPC;
4521
4522         space_info = __find_space_info(fs_info, flags);
4523         ASSERT(space_info);
4524
4525         do {
4526                 spin_lock(&space_info->lock);
4527                 if (force < space_info->force_alloc)
4528                         force = space_info->force_alloc;
4529                 should_alloc = should_alloc_chunk(fs_info, space_info, force);
4530                 if (space_info->full) {
4531                         /* No more free physical space */
4532                         if (should_alloc)
4533                                 ret = -ENOSPC;
4534                         else
4535                                 ret = 0;
4536                         spin_unlock(&space_info->lock);
4537                         return ret;
4538                 } else if (!should_alloc) {
4539                         spin_unlock(&space_info->lock);
4540                         return 0;
4541                 } else if (space_info->chunk_alloc) {
4542                         /*
4543                          * Someone is already allocating, so we need to block
4544                          * until this someone is finished and then loop to
4545                          * recheck if we should continue with our allocation
4546                          * attempt.
4547                          */
4548                         wait_for_alloc = true;
4549                         spin_unlock(&space_info->lock);
4550                         mutex_lock(&fs_info->chunk_mutex);
4551                         mutex_unlock(&fs_info->chunk_mutex);
4552                 } else {
4553                         /* Proceed with allocation */
4554                         space_info->chunk_alloc = 1;
4555                         wait_for_alloc = false;
4556                         spin_unlock(&space_info->lock);
4557                 }
4558
4559                 cond_resched();
4560         } while (wait_for_alloc);
4561
4562         mutex_lock(&fs_info->chunk_mutex);
4563         trans->allocating_chunk = true;
4564
4565         /*
4566          * If we have mixed data/metadata chunks we want to make sure we keep
4567          * allocating mixed chunks instead of individual chunks.
4568          */
4569         if (btrfs_mixed_space_info(space_info))
4570                 flags |= (BTRFS_BLOCK_GROUP_DATA | BTRFS_BLOCK_GROUP_METADATA);
4571
4572         /*
4573          * if we're doing a data chunk, go ahead and make sure that
4574          * we keep a reasonable number of metadata chunks allocated in the
4575          * FS as well.
4576          */
4577         if (flags & BTRFS_BLOCK_GROUP_DATA && fs_info->metadata_ratio) {
4578                 fs_info->data_chunk_allocations++;
4579                 if (!(fs_info->data_chunk_allocations %
4580                       fs_info->metadata_ratio))
4581                         force_metadata_allocation(fs_info);
4582         }
4583
4584         /*
4585          * Check if we have enough space in SYSTEM chunk because we may need
4586          * to update devices.
4587          */
4588         check_system_chunk(trans, flags);
4589
4590         ret = btrfs_alloc_chunk(trans, flags);
4591         trans->allocating_chunk = false;
4592
4593         spin_lock(&space_info->lock);
4594         if (ret < 0) {
4595                 if (ret == -ENOSPC)
4596                         space_info->full = 1;
4597                 else
4598                         goto out;
4599         } else {
4600                 ret = 1;
4601                 space_info->max_extent_size = 0;
4602         }
4603
4604         space_info->force_alloc = CHUNK_ALLOC_NO_FORCE;
4605 out:
4606         space_info->chunk_alloc = 0;
4607         spin_unlock(&space_info->lock);
4608         mutex_unlock(&fs_info->chunk_mutex);
4609         /*
4610          * When we allocate a new chunk we reserve space in the chunk block
4611          * reserve to make sure we can COW nodes/leafs in the chunk tree or
4612          * add new nodes/leafs to it if we end up needing to do it when
4613          * inserting the chunk item and updating device items as part of the
4614          * second phase of chunk allocation, performed by
4615          * btrfs_finish_chunk_alloc(). So make sure we don't accumulate a
4616          * large number of new block groups to create in our transaction
4617          * handle's new_bgs list to avoid exhausting the chunk block reserve
4618          * in extreme cases - like having a single transaction create many new
4619          * block groups when starting to write out the free space caches of all
4620          * the block groups that were made dirty during the lifetime of the
4621          * transaction.
4622          */
4623         if (trans->chunk_bytes_reserved >= (u64)SZ_2M)
4624                 btrfs_create_pending_block_groups(trans);
4625
4626         return ret;
4627 }
4628
4629 static int can_overcommit(struct btrfs_fs_info *fs_info,
4630                           struct btrfs_space_info *space_info, u64 bytes,
4631                           enum btrfs_reserve_flush_enum flush,
4632                           bool system_chunk)
4633 {
4634         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
4635         u64 profile;
4636         u64 space_size;
4637         u64 avail;
4638         u64 used;
4639         int factor;
4640
4641         /* Don't overcommit when in mixed mode. */
4642         if (space_info->flags & BTRFS_BLOCK_GROUP_DATA)
4643                 return 0;
4644
4645         if (system_chunk)
4646                 profile = btrfs_system_alloc_profile(fs_info);
4647         else
4648                 profile = btrfs_metadata_alloc_profile(fs_info);
4649
4650         used = btrfs_space_info_used(space_info, false);
4651
4652         /*
4653          * We only want to allow over committing if we have lots of actual space
4654          * free, but if we don't have enough space to handle the global reserve
4655          * space then we could end up having a real enospc problem when trying
4656          * to allocate a chunk or some other such important allocation.
4657          */
4658         spin_lock(&global_rsv->lock);
4659         space_size = calc_global_rsv_need_space(global_rsv);
4660         spin_unlock(&global_rsv->lock);
4661         if (used + space_size >= space_info->total_bytes)
4662                 return 0;
4663
4664         used += space_info->bytes_may_use;
4665
4666         avail = atomic64_read(&fs_info->free_chunk_space);
4667
4668         /*
4669          * If we have dup, raid1 or raid10 then only half of the free
4670          * space is actually usable.  For raid56, the space info used
4671          * doesn't include the parity drive, so we don't have to
4672          * change the math
4673          */
4674         factor = btrfs_bg_type_to_factor(profile);
4675         avail = div_u64(avail, factor);
4676
4677         /*
4678          * If we aren't flushing all things, let us overcommit up to
4679          * 1/2th of the space. If we can flush, don't let us overcommit
4680          * too much, let it overcommit up to 1/8 of the space.
4681          */
4682         if (flush == BTRFS_RESERVE_FLUSH_ALL)
4683                 avail >>= 3;
4684         else
4685                 avail >>= 1;
4686
4687         if (used + bytes < space_info->total_bytes + avail)
4688                 return 1;
4689         return 0;
4690 }
4691
4692 static void btrfs_writeback_inodes_sb_nr(struct btrfs_fs_info *fs_info,
4693                                          unsigned long nr_pages, int nr_items)
4694 {
4695         struct super_block *sb = fs_info->sb;
4696
4697         if (down_read_trylock(&sb->s_umount)) {
4698                 writeback_inodes_sb_nr(sb, nr_pages, WB_REASON_FS_FREE_SPACE);
4699                 up_read(&sb->s_umount);
4700         } else {
4701                 /*
4702                  * We needn't worry the filesystem going from r/w to r/o though
4703                  * we don't acquire ->s_umount mutex, because the filesystem
4704                  * should guarantee the delalloc inodes list be empty after
4705                  * the filesystem is readonly(all dirty pages are written to
4706                  * the disk).
4707                  */
4708                 btrfs_start_delalloc_roots(fs_info, nr_items);
4709                 if (!current->journal_info)
4710                         btrfs_wait_ordered_roots(fs_info, nr_items, 0, (u64)-1);
4711         }
4712 }
4713
4714 static inline u64 calc_reclaim_items_nr(struct btrfs_fs_info *fs_info,
4715                                         u64 to_reclaim)
4716 {
4717         u64 bytes;
4718         u64 nr;
4719
4720         bytes = btrfs_calc_trans_metadata_size(fs_info, 1);
4721         nr = div64_u64(to_reclaim, bytes);
4722         if (!nr)
4723                 nr = 1;
4724         return nr;
4725 }
4726
4727 #define EXTENT_SIZE_PER_ITEM    SZ_256K
4728
4729 /*
4730  * shrink metadata reservation for delalloc
4731  */
4732 static void shrink_delalloc(struct btrfs_fs_info *fs_info, u64 to_reclaim,
4733                             u64 orig, bool wait_ordered)
4734 {
4735         struct btrfs_space_info *space_info;
4736         struct btrfs_trans_handle *trans;
4737         u64 delalloc_bytes;
4738         u64 async_pages;
4739         u64 items;
4740         long time_left;
4741         unsigned long nr_pages;
4742         int loops;
4743
4744         /* Calc the number of the pages we need flush for space reservation */
4745         items = calc_reclaim_items_nr(fs_info, to_reclaim);
4746         to_reclaim = items * EXTENT_SIZE_PER_ITEM;
4747
4748         trans = (struct btrfs_trans_handle *)current->journal_info;
4749         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
4750
4751         delalloc_bytes = percpu_counter_sum_positive(
4752                                                 &fs_info->delalloc_bytes);
4753         if (delalloc_bytes == 0) {
4754                 if (trans)
4755                         return;
4756                 if (wait_ordered)
4757                         btrfs_wait_ordered_roots(fs_info, items, 0, (u64)-1);
4758                 return;
4759         }
4760
4761         loops = 0;
4762         while (delalloc_bytes && loops < 3) {
4763                 nr_pages = min(delalloc_bytes, to_reclaim) >> PAGE_SHIFT;
4764
4765                 /*
4766                  * Triggers inode writeback for up to nr_pages. This will invoke
4767                  * ->writepages callback and trigger delalloc filling
4768                  *  (btrfs_run_delalloc_range()).
4769                  */
4770                 btrfs_writeback_inodes_sb_nr(fs_info, nr_pages, items);
4771
4772                 /*
4773                  * We need to wait for the compressed pages to start before
4774                  * we continue.
4775                  */
4776                 async_pages = atomic_read(&fs_info->async_delalloc_pages);
4777                 if (!async_pages)
4778                         goto skip_async;
4779
4780                 /*
4781                  * Calculate how many compressed pages we want to be written
4782                  * before we continue. I.e if there are more async pages than we
4783                  * require wait_event will wait until nr_pages are written.
4784                  */
4785                 if (async_pages <= nr_pages)
4786                         async_pages = 0;
4787                 else
4788                         async_pages -= nr_pages;
4789
4790                 wait_event(fs_info->async_submit_wait,
4791                            atomic_read(&fs_info->async_delalloc_pages) <=
4792                            (int)async_pages);
4793 skip_async:
4794                 spin_lock(&space_info->lock);
4795                 if (list_empty(&space_info->tickets) &&
4796                     list_empty(&space_info->priority_tickets)) {
4797                         spin_unlock(&space_info->lock);
4798                         break;
4799                 }
4800                 spin_unlock(&space_info->lock);
4801
4802                 loops++;
4803                 if (wait_ordered && !trans) {
4804                         btrfs_wait_ordered_roots(fs_info, items, 0, (u64)-1);
4805                 } else {
4806                         time_left = schedule_timeout_killable(1);
4807                         if (time_left)
4808                                 break;
4809                 }
4810                 delalloc_bytes = percpu_counter_sum_positive(
4811                                                 &fs_info->delalloc_bytes);
4812         }
4813 }
4814
4815 struct reserve_ticket {
4816         u64 orig_bytes;
4817         u64 bytes;
4818         int error;
4819         struct list_head list;
4820         wait_queue_head_t wait;
4821 };
4822
4823 /**
4824  * maybe_commit_transaction - possibly commit the transaction if its ok to
4825  * @root - the root we're allocating for
4826  * @bytes - the number of bytes we want to reserve
4827  * @force - force the commit
4828  *
4829  * This will check to make sure that committing the transaction will actually
4830  * get us somewhere and then commit the transaction if it does.  Otherwise it
4831  * will return -ENOSPC.
4832  */
4833 static int may_commit_transaction(struct btrfs_fs_info *fs_info,
4834                                   struct btrfs_space_info *space_info)
4835 {
4836         struct reserve_ticket *ticket = NULL;
4837         struct btrfs_block_rsv *delayed_rsv = &fs_info->delayed_block_rsv;
4838         struct btrfs_block_rsv *delayed_refs_rsv = &fs_info->delayed_refs_rsv;
4839         struct btrfs_trans_handle *trans;
4840         u64 bytes_needed;
4841         u64 reclaim_bytes = 0;
4842
4843         trans = (struct btrfs_trans_handle *)current->journal_info;
4844         if (trans)
4845                 return -EAGAIN;
4846
4847         spin_lock(&space_info->lock);
4848         if (!list_empty(&space_info->priority_tickets))
4849                 ticket = list_first_entry(&space_info->priority_tickets,
4850                                           struct reserve_ticket, list);
4851         else if (!list_empty(&space_info->tickets))
4852                 ticket = list_first_entry(&space_info->tickets,
4853                                           struct reserve_ticket, list);
4854         bytes_needed = (ticket) ? ticket->bytes : 0;
4855         spin_unlock(&space_info->lock);
4856
4857         if (!bytes_needed)
4858                 return 0;
4859
4860         trans = btrfs_join_transaction(fs_info->extent_root);
4861         if (IS_ERR(trans))
4862                 return PTR_ERR(trans);
4863
4864         /*
4865          * See if there is enough pinned space to make this reservation, or if
4866          * we have block groups that are going to be freed, allowing us to
4867          * possibly do a chunk allocation the next loop through.
4868          */
4869         if (test_bit(BTRFS_TRANS_HAVE_FREE_BGS, &trans->transaction->flags) ||
4870             __percpu_counter_compare(&space_info->total_bytes_pinned,
4871                                      bytes_needed,
4872                                      BTRFS_TOTAL_BYTES_PINNED_BATCH) >= 0)
4873                 goto commit;
4874
4875         /*
4876          * See if there is some space in the delayed insertion reservation for
4877          * this reservation.
4878          */
4879         if (space_info != delayed_rsv->space_info)
4880                 goto enospc;
4881
4882         spin_lock(&delayed_rsv->lock);
4883         reclaim_bytes += delayed_rsv->reserved;
4884         spin_unlock(&delayed_rsv->lock);
4885
4886         spin_lock(&delayed_refs_rsv->lock);
4887         reclaim_bytes += delayed_refs_rsv->reserved;
4888         spin_unlock(&delayed_refs_rsv->lock);
4889         if (reclaim_bytes >= bytes_needed)
4890                 goto commit;
4891         bytes_needed -= reclaim_bytes;
4892
4893         if (__percpu_counter_compare(&space_info->total_bytes_pinned,
4894                                    bytes_needed,
4895                                    BTRFS_TOTAL_BYTES_PINNED_BATCH) < 0)
4896                 goto enospc;
4897
4898 commit:
4899         return btrfs_commit_transaction(trans);
4900 enospc:
4901         btrfs_end_transaction(trans);
4902         return -ENOSPC;
4903 }
4904
4905 /*
4906  * Try to flush some data based on policy set by @state. This is only advisory
4907  * and may fail for various reasons. The caller is supposed to examine the
4908  * state of @space_info to detect the outcome.
4909  */
4910 static void flush_space(struct btrfs_fs_info *fs_info,
4911                        struct btrfs_space_info *space_info, u64 num_bytes,
4912                        int state)
4913 {
4914         struct btrfs_root *root = fs_info->extent_root;
4915         struct btrfs_trans_handle *trans;
4916         int nr;
4917         int ret = 0;
4918
4919         switch (state) {
4920         case FLUSH_DELAYED_ITEMS_NR:
4921         case FLUSH_DELAYED_ITEMS:
4922                 if (state == FLUSH_DELAYED_ITEMS_NR)
4923                         nr = calc_reclaim_items_nr(fs_info, num_bytes) * 2;
4924                 else
4925                         nr = -1;
4926
4927                 trans = btrfs_join_transaction(root);
4928                 if (IS_ERR(trans)) {
4929                         ret = PTR_ERR(trans);
4930                         break;
4931                 }
4932                 ret = btrfs_run_delayed_items_nr(trans, nr);
4933                 btrfs_end_transaction(trans);
4934                 break;
4935         case FLUSH_DELALLOC:
4936         case FLUSH_DELALLOC_WAIT:
4937                 shrink_delalloc(fs_info, num_bytes * 2, num_bytes,
4938                                 state == FLUSH_DELALLOC_WAIT);
4939                 break;
4940         case FLUSH_DELAYED_REFS_NR:
4941         case FLUSH_DELAYED_REFS:
4942                 trans = btrfs_join_transaction(root);
4943                 if (IS_ERR(trans)) {
4944                         ret = PTR_ERR(trans);
4945                         break;
4946                 }
4947                 if (state == FLUSH_DELAYED_REFS_NR)
4948                         nr = calc_reclaim_items_nr(fs_info, num_bytes);
4949                 else
4950                         nr = 0;
4951                 btrfs_run_delayed_refs(trans, nr);
4952                 btrfs_end_transaction(trans);
4953                 break;
4954         case ALLOC_CHUNK:
4955         case ALLOC_CHUNK_FORCE:
4956                 trans = btrfs_join_transaction(root);
4957                 if (IS_ERR(trans)) {
4958                         ret = PTR_ERR(trans);
4959                         break;
4960                 }
4961                 ret = do_chunk_alloc(trans,
4962                                      btrfs_metadata_alloc_profile(fs_info),
4963                                      (state == ALLOC_CHUNK) ?
4964                                       CHUNK_ALLOC_NO_FORCE : CHUNK_ALLOC_FORCE);
4965                 btrfs_end_transaction(trans);
4966                 if (ret > 0 || ret == -ENOSPC)
4967                         ret = 0;
4968                 break;
4969         case COMMIT_TRANS:
4970                 /*
4971                  * If we have pending delayed iputs then we could free up a
4972                  * bunch of pinned space, so make sure we run the iputs before
4973                  * we do our pinned bytes check below.
4974                  */
4975                 btrfs_run_delayed_iputs(fs_info);
4976                 btrfs_wait_on_delayed_iputs(fs_info);
4977
4978                 ret = may_commit_transaction(fs_info, space_info);
4979                 break;
4980         default:
4981                 ret = -ENOSPC;
4982                 break;
4983         }
4984
4985         trace_btrfs_flush_space(fs_info, space_info->flags, num_bytes, state,
4986                                 ret);
4987         return;
4988 }
4989
4990 static inline u64
4991 btrfs_calc_reclaim_metadata_size(struct btrfs_fs_info *fs_info,
4992                                  struct btrfs_space_info *space_info,
4993                                  bool system_chunk)
4994 {
4995         struct reserve_ticket *ticket;
4996         u64 used;
4997         u64 expected;
4998         u64 to_reclaim = 0;
4999
5000         list_for_each_entry(ticket, &space_info->tickets, list)
5001                 to_reclaim += ticket->bytes;
5002         list_for_each_entry(ticket, &space_info->priority_tickets, list)
5003                 to_reclaim += ticket->bytes;
5004         if (to_reclaim)
5005                 return to_reclaim;
5006
5007         to_reclaim = min_t(u64, num_online_cpus() * SZ_1M, SZ_16M);
5008         if (can_overcommit(fs_info, space_info, to_reclaim,
5009                            BTRFS_RESERVE_FLUSH_ALL, system_chunk))
5010                 return 0;
5011
5012         used = btrfs_space_info_used(space_info, true);
5013
5014         if (can_overcommit(fs_info, space_info, SZ_1M,
5015                            BTRFS_RESERVE_FLUSH_ALL, system_chunk))
5016                 expected = div_factor_fine(space_info->total_bytes, 95);
5017         else
5018                 expected = div_factor_fine(space_info->total_bytes, 90);
5019
5020         if (used > expected)
5021                 to_reclaim = used - expected;
5022         else
5023                 to_reclaim = 0;
5024         to_reclaim = min(to_reclaim, space_info->bytes_may_use +
5025                                      space_info->bytes_reserved);
5026         return to_reclaim;
5027 }
5028
5029 static inline int need_do_async_reclaim(struct btrfs_fs_info *fs_info,
5030                                         struct btrfs_space_info *space_info,
5031                                         u64 used, bool system_chunk)
5032 {
5033         u64 thresh = div_factor_fine(space_info->total_bytes, 98);
5034
5035         /* If we're just plain full then async reclaim just slows us down. */
5036         if ((space_info->bytes_used + space_info->bytes_reserved) >= thresh)
5037                 return 0;
5038
5039         if (!btrfs_calc_reclaim_metadata_size(fs_info, space_info,
5040                                               system_chunk))
5041                 return 0;
5042
5043         return (used >= thresh && !btrfs_fs_closing(fs_info) &&
5044                 !test_bit(BTRFS_FS_STATE_REMOUNTING, &fs_info->fs_state));
5045 }
5046
5047 static bool wake_all_tickets(struct list_head *head)
5048 {
5049         struct reserve_ticket *ticket;
5050
5051         while (!list_empty(head)) {
5052                 ticket = list_first_entry(head, struct reserve_ticket, list);
5053                 list_del_init(&ticket->list);
5054                 ticket->error = -ENOSPC;
5055                 wake_up(&ticket->wait);
5056                 if (ticket->bytes != ticket->orig_bytes)
5057                         return true;
5058         }
5059         return false;
5060 }
5061
5062 /*
5063  * This is for normal flushers, we can wait all goddamned day if we want to.  We
5064  * will loop and continuously try to flush as long as we are making progress.
5065  * We count progress as clearing off tickets each time we have to loop.
5066  */
5067 static void btrfs_async_reclaim_metadata_space(struct work_struct *work)
5068 {
5069         struct btrfs_fs_info *fs_info;
5070         struct btrfs_space_info *space_info;
5071         u64 to_reclaim;
5072         int flush_state;
5073         int commit_cycles = 0;
5074         u64 last_tickets_id;
5075
5076         fs_info = container_of(work, struct btrfs_fs_info, async_reclaim_work);
5077         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
5078
5079         spin_lock(&space_info->lock);
5080         to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info, space_info,
5081                                                       false);
5082         if (!to_reclaim) {
5083                 space_info->flush = 0;
5084                 spin_unlock(&space_info->lock);
5085                 return;
5086         }
5087         last_tickets_id = space_info->tickets_id;
5088         spin_unlock(&space_info->lock);
5089
5090         flush_state = FLUSH_DELAYED_ITEMS_NR;
5091         do {
5092                 flush_space(fs_info, space_info, to_reclaim, flush_state);
5093                 spin_lock(&space_info->lock);
5094                 if (list_empty(&space_info->tickets)) {
5095                         space_info->flush = 0;
5096                         spin_unlock(&space_info->lock);
5097                         return;
5098                 }
5099                 to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info,
5100                                                               space_info,
5101                                                               false);
5102                 if (last_tickets_id == space_info->tickets_id) {
5103                         flush_state++;
5104                 } else {
5105                         last_tickets_id = space_info->tickets_id;
5106                         flush_state = FLUSH_DELAYED_ITEMS_NR;
5107                         if (commit_cycles)
5108                                 commit_cycles--;
5109                 }
5110
5111                 /*
5112                  * We don't want to force a chunk allocation until we've tried
5113                  * pretty hard to reclaim space.  Think of the case where we
5114                  * freed up a bunch of space and so have a lot of pinned space
5115                  * to reclaim.  We would rather use that than possibly create a
5116                  * underutilized metadata chunk.  So if this is our first run
5117                  * through the flushing state machine skip ALLOC_CHUNK_FORCE and
5118                  * commit the transaction.  If nothing has changed the next go
5119                  * around then we can force a chunk allocation.
5120                  */
5121                 if (flush_state == ALLOC_CHUNK_FORCE && !commit_cycles)
5122                         flush_state++;
5123
5124                 if (flush_state > COMMIT_TRANS) {
5125                         commit_cycles++;
5126                         if (commit_cycles > 2) {
5127                                 if (wake_all_tickets(&space_info->tickets)) {
5128                                         flush_state = FLUSH_DELAYED_ITEMS_NR;
5129                                         commit_cycles--;
5130                                 } else {
5131                                         space_info->flush = 0;
5132                                 }
5133                         } else {
5134                                 flush_state = FLUSH_DELAYED_ITEMS_NR;
5135                         }
5136                 }
5137                 spin_unlock(&space_info->lock);
5138         } while (flush_state <= COMMIT_TRANS);
5139 }
5140
5141 void btrfs_init_async_reclaim_work(struct work_struct *work)
5142 {
5143         INIT_WORK(work, btrfs_async_reclaim_metadata_space);
5144 }
5145
5146 static void priority_reclaim_metadata_space(struct btrfs_fs_info *fs_info,
5147                                             struct btrfs_space_info *space_info,
5148                                             struct reserve_ticket *ticket)
5149 {
5150         u64 to_reclaim;
5151         int flush_state = FLUSH_DELAYED_ITEMS_NR;
5152
5153         spin_lock(&space_info->lock);
5154         to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info, space_info,
5155                                                       false);
5156         if (!to_reclaim) {
5157                 spin_unlock(&space_info->lock);
5158                 return;
5159         }
5160         spin_unlock(&space_info->lock);
5161
5162         do {
5163                 flush_space(fs_info, space_info, to_reclaim, flush_state);
5164                 flush_state++;
5165                 spin_lock(&space_info->lock);
5166                 if (ticket->bytes == 0) {
5167                         spin_unlock(&space_info->lock);
5168                         return;
5169                 }
5170                 spin_unlock(&space_info->lock);
5171
5172                 /*
5173                  * Priority flushers can't wait on delalloc without
5174                  * deadlocking.
5175                  */
5176                 if (flush_state == FLUSH_DELALLOC ||
5177                     flush_state == FLUSH_DELALLOC_WAIT)
5178                         flush_state = ALLOC_CHUNK;
5179         } while (flush_state < COMMIT_TRANS);
5180 }
5181
5182 static int wait_reserve_ticket(struct btrfs_fs_info *fs_info,
5183                                struct btrfs_space_info *space_info,
5184                                struct reserve_ticket *ticket)
5185
5186 {
5187         DEFINE_WAIT(wait);
5188         u64 reclaim_bytes = 0;
5189         int ret = 0;
5190
5191         spin_lock(&space_info->lock);
5192         while (ticket->bytes > 0 && ticket->error == 0) {
5193                 ret = prepare_to_wait_event(&ticket->wait, &wait, TASK_KILLABLE);
5194                 if (ret) {
5195                         ret = -EINTR;
5196                         break;
5197                 }
5198                 spin_unlock(&space_info->lock);
5199
5200                 schedule();
5201
5202                 finish_wait(&ticket->wait, &wait);
5203                 spin_lock(&space_info->lock);
5204         }
5205         if (!ret)
5206                 ret = ticket->error;
5207         if (!list_empty(&ticket->list))
5208                 list_del_init(&ticket->list);
5209         if (ticket->bytes && ticket->bytes < ticket->orig_bytes)
5210                 reclaim_bytes = ticket->orig_bytes - ticket->bytes;
5211         spin_unlock(&space_info->lock);
5212
5213         if (reclaim_bytes)
5214                 space_info_add_old_bytes(fs_info, space_info, reclaim_bytes);
5215         return ret;
5216 }
5217
5218 /**
5219  * reserve_metadata_bytes - try to reserve bytes from the block_rsv's space
5220  * @root - the root we're allocating for
5221  * @space_info - the space info we want to allocate from
5222  * @orig_bytes - the number of bytes we want
5223  * @flush - whether or not we can flush to make our reservation
5224  *
5225  * This will reserve orig_bytes number of bytes from the space info associated
5226  * with the block_rsv.  If there is not enough space it will make an attempt to
5227  * flush out space to make room.  It will do this by flushing delalloc if
5228  * possible or committing the transaction.  If flush is 0 then no attempts to
5229  * regain reservations will be made and this will fail if there is not enough
5230  * space already.
5231  */
5232 static int __reserve_metadata_bytes(struct btrfs_fs_info *fs_info,
5233                                     struct btrfs_space_info *space_info,
5234                                     u64 orig_bytes,
5235                                     enum btrfs_reserve_flush_enum flush,
5236                                     bool system_chunk)
5237 {
5238         struct reserve_ticket ticket;
5239         u64 used;
5240         u64 reclaim_bytes = 0;
5241         int ret = 0;
5242
5243         ASSERT(orig_bytes);
5244         ASSERT(!current->journal_info || flush != BTRFS_RESERVE_FLUSH_ALL);
5245
5246         spin_lock(&space_info->lock);
5247         ret = -ENOSPC;
5248         used = btrfs_space_info_used(space_info, true);
5249
5250         /*
5251          * If we have enough space then hooray, make our reservation and carry
5252          * on.  If not see if we can overcommit, and if we can, hooray carry on.
5253          * If not things get more complicated.
5254          */
5255         if (used + orig_bytes <= space_info->total_bytes) {
5256                 update_bytes_may_use(space_info, orig_bytes);
5257                 trace_btrfs_space_reservation(fs_info, "space_info",
5258                                               space_info->flags, orig_bytes, 1);
5259                 ret = 0;
5260         } else if (can_overcommit(fs_info, space_info, orig_bytes, flush,
5261                                   system_chunk)) {
5262                 update_bytes_may_use(space_info, orig_bytes);
5263                 trace_btrfs_space_reservation(fs_info, "space_info",
5264                                               space_info->flags, orig_bytes, 1);
5265                 ret = 0;
5266         }
5267
5268         /*
5269          * If we couldn't make a reservation then setup our reservation ticket
5270          * and kick the async worker if it's not already running.
5271          *
5272          * If we are a priority flusher then we just need to add our ticket to
5273          * the list and we will do our own flushing further down.
5274          */
5275         if (ret && flush != BTRFS_RESERVE_NO_FLUSH) {
5276                 ticket.orig_bytes = orig_bytes;
5277                 ticket.bytes = orig_bytes;
5278                 ticket.error = 0;
5279                 init_waitqueue_head(&ticket.wait);
5280                 if (flush == BTRFS_RESERVE_FLUSH_ALL) {
5281                         list_add_tail(&ticket.list, &space_info->tickets);
5282                         if (!space_info->flush) {
5283                                 space_info->flush = 1;
5284                                 trace_btrfs_trigger_flush(fs_info,
5285                                                           space_info->flags,
5286                                                           orig_bytes, flush,
5287                                                           "enospc");
5288                                 queue_work(system_unbound_wq,
5289                                            &fs_info->async_reclaim_work);
5290                         }
5291                 } else {
5292                         list_add_tail(&ticket.list,
5293                                       &space_info->priority_tickets);
5294                 }
5295         } else if (!ret && space_info->flags & BTRFS_BLOCK_GROUP_METADATA) {
5296                 used += orig_bytes;
5297                 /*
5298                  * We will do the space reservation dance during log replay,
5299                  * which means we won't have fs_info->fs_root set, so don't do
5300                  * the async reclaim as we will panic.
5301                  */
5302                 if (!test_bit(BTRFS_FS_LOG_RECOVERING, &fs_info->flags) &&
5303                     need_do_async_reclaim(fs_info, space_info,
5304                                           used, system_chunk) &&
5305                     !work_busy(&fs_info->async_reclaim_work)) {
5306                         trace_btrfs_trigger_flush(fs_info, space_info->flags,
5307                                                   orig_bytes, flush, "preempt");
5308                         queue_work(system_unbound_wq,
5309                                    &fs_info->async_reclaim_work);
5310                 }
5311         }
5312         spin_unlock(&space_info->lock);
5313         if (!ret || flush == BTRFS_RESERVE_NO_FLUSH)
5314                 return ret;
5315
5316         if (flush == BTRFS_RESERVE_FLUSH_ALL)
5317                 return wait_reserve_ticket(fs_info, space_info, &ticket);
5318
5319         ret = 0;
5320         priority_reclaim_metadata_space(fs_info, space_info, &ticket);
5321         spin_lock(&space_info->lock);
5322         if (ticket.bytes) {
5323                 if (ticket.bytes < orig_bytes)
5324                         reclaim_bytes = orig_bytes - ticket.bytes;
5325                 list_del_init(&ticket.list);
5326                 ret = -ENOSPC;
5327         }
5328         spin_unlock(&space_info->lock);
5329
5330         if (reclaim_bytes)
5331                 space_info_add_old_bytes(fs_info, space_info, reclaim_bytes);
5332         ASSERT(list_empty(&ticket.list));
5333         return ret;
5334 }
5335
5336 /**
5337  * reserve_metadata_bytes - try to reserve bytes from the block_rsv's space
5338  * @root - the root we're allocating for
5339  * @block_rsv - the block_rsv we're allocating for
5340  * @orig_bytes - the number of bytes we want
5341  * @flush - whether or not we can flush to make our reservation
5342  *
5343  * This will reserve orig_bytes number of bytes from the space info associated
5344  * with the block_rsv.  If there is not enough space it will make an attempt to
5345  * flush out space to make room.  It will do this by flushing delalloc if
5346  * possible or committing the transaction.  If flush is 0 then no attempts to
5347  * regain reservations will be made and this will fail if there is not enough
5348  * space already.
5349  */
5350 static int reserve_metadata_bytes(struct btrfs_root *root,
5351                                   struct btrfs_block_rsv *block_rsv,
5352                                   u64 orig_bytes,
5353                                   enum btrfs_reserve_flush_enum flush)
5354 {
5355         struct btrfs_fs_info *fs_info = root->fs_info;
5356         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
5357         int ret;
5358         bool system_chunk = (root == fs_info->chunk_root);
5359
5360         ret = __reserve_metadata_bytes(fs_info, block_rsv->space_info,
5361                                        orig_bytes, flush, system_chunk);
5362         if (ret == -ENOSPC &&
5363             unlikely(root->orphan_cleanup_state == ORPHAN_CLEANUP_STARTED)) {
5364                 if (block_rsv != global_rsv &&
5365                     !block_rsv_use_bytes(global_rsv, orig_bytes))
5366                         ret = 0;
5367         }
5368         if (ret == -ENOSPC) {
5369                 trace_btrfs_space_reservation(fs_info, "space_info:enospc",
5370                                               block_rsv->space_info->flags,
5371                                               orig_bytes, 1);
5372
5373                 if (btrfs_test_opt(fs_info, ENOSPC_DEBUG))
5374                         dump_space_info(fs_info, block_rsv->space_info,
5375                                         orig_bytes, 0);
5376         }
5377         return ret;
5378 }
5379
5380 static struct btrfs_block_rsv *get_block_rsv(
5381                                         const struct btrfs_trans_handle *trans,
5382                                         const struct btrfs_root *root)
5383 {
5384         struct btrfs_fs_info *fs_info = root->fs_info;
5385         struct btrfs_block_rsv *block_rsv = NULL;
5386
5387         if (test_bit(BTRFS_ROOT_REF_COWS, &root->state) ||
5388             (root == fs_info->csum_root && trans->adding_csums) ||
5389             (root == fs_info->uuid_root))
5390                 block_rsv = trans->block_rsv;
5391
5392         if (!block_rsv)
5393                 block_rsv = root->block_rsv;
5394
5395         if (!block_rsv)
5396                 block_rsv = &fs_info->empty_block_rsv;
5397
5398         return block_rsv;
5399 }
5400
5401 static int block_rsv_use_bytes(struct btrfs_block_rsv *block_rsv,
5402                                u64 num_bytes)
5403 {
5404         int ret = -ENOSPC;
5405         spin_lock(&block_rsv->lock);
5406         if (block_rsv->reserved >= num_bytes) {
5407                 block_rsv->reserved -= num_bytes;
5408                 if (block_rsv->reserved < block_rsv->size)
5409                         block_rsv->full = 0;
5410                 ret = 0;
5411         }
5412         spin_unlock(&block_rsv->lock);
5413         return ret;
5414 }
5415
5416 static void block_rsv_add_bytes(struct btrfs_block_rsv *block_rsv,
5417                                 u64 num_bytes, bool update_size)
5418 {
5419         spin_lock(&block_rsv->lock);
5420         block_rsv->reserved += num_bytes;
5421         if (update_size)
5422                 block_rsv->size += num_bytes;
5423         else if (block_rsv->reserved >= block_rsv->size)
5424                 block_rsv->full = 1;
5425         spin_unlock(&block_rsv->lock);
5426 }
5427
5428 int btrfs_cond_migrate_bytes(struct btrfs_fs_info *fs_info,
5429                              struct btrfs_block_rsv *dest, u64 num_bytes,
5430                              int min_factor)
5431 {
5432         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
5433         u64 min_bytes;
5434
5435         if (global_rsv->space_info != dest->space_info)
5436                 return -ENOSPC;
5437
5438         spin_lock(&global_rsv->lock);
5439         min_bytes = div_factor(global_rsv->size, min_factor);
5440         if (global_rsv->reserved < min_bytes + num_bytes) {
5441                 spin_unlock(&global_rsv->lock);
5442                 return -ENOSPC;
5443         }
5444         global_rsv->reserved -= num_bytes;
5445         if (global_rsv->reserved < global_rsv->size)
5446                 global_rsv->full = 0;
5447         spin_unlock(&global_rsv->lock);
5448
5449         block_rsv_add_bytes(dest, num_bytes, true);
5450         return 0;
5451 }
5452
5453 /**
5454  * btrfs_migrate_to_delayed_refs_rsv - transfer bytes to our delayed refs rsv.
5455  * @fs_info - the fs info for our fs.
5456  * @src - the source block rsv to transfer from.
5457  * @num_bytes - the number of bytes to transfer.
5458  *
5459  * This transfers up to the num_bytes amount from the src rsv to the
5460  * delayed_refs_rsv.  Any extra bytes are returned to the space info.
5461  */
5462 void btrfs_migrate_to_delayed_refs_rsv(struct btrfs_fs_info *fs_info,
5463                                        struct btrfs_block_rsv *src,
5464                                        u64 num_bytes)
5465 {
5466         struct btrfs_block_rsv *delayed_refs_rsv = &fs_info->delayed_refs_rsv;
5467         u64 to_free = 0;
5468
5469         spin_lock(&src->lock);
5470         src->reserved -= num_bytes;
5471         src->size -= num_bytes;
5472         spin_unlock(&src->lock);
5473
5474         spin_lock(&delayed_refs_rsv->lock);
5475         if (delayed_refs_rsv->size > delayed_refs_rsv->reserved) {
5476                 u64 delta = delayed_refs_rsv->size -
5477                         delayed_refs_rsv->reserved;
5478                 if (num_bytes > delta) {
5479                         to_free = num_bytes - delta;
5480                         num_bytes = delta;
5481                 }
5482         } else {
5483                 to_free = num_bytes;
5484                 num_bytes = 0;
5485         }
5486
5487         if (num_bytes)
5488                 delayed_refs_rsv->reserved += num_bytes;
5489         if (delayed_refs_rsv->reserved >= delayed_refs_rsv->size)
5490                 delayed_refs_rsv->full = 1;
5491         spin_unlock(&delayed_refs_rsv->lock);
5492
5493         if (num_bytes)
5494                 trace_btrfs_space_reservation(fs_info, "delayed_refs_rsv",
5495                                               0, num_bytes, 1);
5496         if (to_free)
5497                 space_info_add_old_bytes(fs_info, delayed_refs_rsv->space_info,
5498                                          to_free);
5499 }
5500
5501 /**
5502  * btrfs_delayed_refs_rsv_refill - refill based on our delayed refs usage.
5503  * @fs_info - the fs_info for our fs.
5504  * @flush - control how we can flush for this reservation.
5505  *
5506  * This will refill the delayed block_rsv up to 1 items size worth of space and
5507  * will return -ENOSPC if we can't make the reservation.
5508  */
5509 int btrfs_delayed_refs_rsv_refill(struct btrfs_fs_info *fs_info,
5510                                   enum btrfs_reserve_flush_enum flush)
5511 {
5512         struct btrfs_block_rsv *block_rsv = &fs_info->delayed_refs_rsv;
5513         u64 limit = btrfs_calc_trans_metadata_size(fs_info, 1);
5514         u64 num_bytes = 0;
5515         int ret = -ENOSPC;
5516
5517         spin_lock(&block_rsv->lock);
5518         if (block_rsv->reserved < block_rsv->size) {
5519                 num_bytes = block_rsv->size - block_rsv->reserved;
5520                 num_bytes = min(num_bytes, limit);
5521         }
5522         spin_unlock(&block_rsv->lock);
5523
5524         if (!num_bytes)
5525                 return 0;
5526
5527         ret = reserve_metadata_bytes(fs_info->extent_root, block_rsv,
5528                                      num_bytes, flush);
5529         if (ret)
5530                 return ret;
5531         block_rsv_add_bytes(block_rsv, num_bytes, 0);
5532         trace_btrfs_space_reservation(fs_info, "delayed_refs_rsv",
5533                                       0, num_bytes, 1);
5534         return 0;
5535 }
5536
5537 /*
5538  * This is for space we already have accounted in space_info->bytes_may_use, so
5539  * basically when we're returning space from block_rsv's.
5540  */
5541 static void space_info_add_old_bytes(struct btrfs_fs_info *fs_info,
5542                                      struct btrfs_space_info *space_info,
5543                                      u64 num_bytes)
5544 {
5545         struct reserve_ticket *ticket;
5546         struct list_head *head;
5547         u64 used;
5548         enum btrfs_reserve_flush_enum flush = BTRFS_RESERVE_NO_FLUSH;
5549         bool check_overcommit = false;
5550
5551         spin_lock(&space_info->lock);
5552         head = &space_info->priority_tickets;
5553
5554         /*
5555          * If we are over our limit then we need to check and see if we can
5556          * overcommit, and if we can't then we just need to free up our space
5557          * and not satisfy any requests.
5558          */
5559         used = btrfs_space_info_used(space_info, true);
5560         if (used - num_bytes >= space_info->total_bytes)
5561                 check_overcommit = true;
5562 again:
5563         while (!list_empty(head) && num_bytes) {
5564                 ticket = list_first_entry(head, struct reserve_ticket,
5565                                           list);
5566                 /*
5567                  * We use 0 bytes because this space is already reserved, so
5568                  * adding the ticket space would be a double count.
5569                  */
5570                 if (check_overcommit &&
5571                     !can_overcommit(fs_info, space_info, 0, flush, false))
5572                         break;
5573                 if (num_bytes >= ticket->bytes) {
5574                         list_del_init(&ticket->list);
5575                         num_bytes -= ticket->bytes;
5576                         ticket->bytes = 0;
5577                         space_info->tickets_id++;
5578                         wake_up(&ticket->wait);
5579                 } else {
5580                         ticket->bytes -= num_bytes;
5581                         num_bytes = 0;
5582                 }
5583         }
5584
5585         if (num_bytes && head == &space_info->priority_tickets) {
5586                 head = &space_info->tickets;
5587                 flush = BTRFS_RESERVE_FLUSH_ALL;
5588                 goto again;
5589         }
5590         update_bytes_may_use(space_info, -num_bytes);
5591         trace_btrfs_space_reservation(fs_info, "space_info",
5592                                       space_info->flags, num_bytes, 0);
5593         spin_unlock(&space_info->lock);
5594 }
5595
5596 /*
5597  * This is for newly allocated space that isn't accounted in
5598  * space_info->bytes_may_use yet.  So if we allocate a chunk or unpin an extent
5599  * we use this helper.
5600  */
5601 static void space_info_add_new_bytes(struct btrfs_fs_info *fs_info,
5602                                      struct btrfs_space_info *space_info,
5603                                      u64 num_bytes)
5604 {
5605         struct reserve_ticket *ticket;
5606         struct list_head *head = &space_info->priority_tickets;
5607
5608 again:
5609         while (!list_empty(head) && num_bytes) {
5610                 ticket = list_first_entry(head, struct reserve_ticket,
5611                                           list);
5612                 if (num_bytes >= ticket->bytes) {
5613                         trace_btrfs_space_reservation(fs_info, "space_info",
5614                                                       space_info->flags,
5615                                                       ticket->bytes, 1);
5616                         list_del_init(&ticket->list);
5617                         num_bytes -= ticket->bytes;
5618                         update_bytes_may_use(space_info, ticket->bytes);
5619                         ticket->bytes = 0;
5620                         space_info->tickets_id++;
5621                         wake_up(&ticket->wait);
5622                 } else {
5623                         trace_btrfs_space_reservation(fs_info, "space_info",
5624                                                       space_info->flags,
5625                                                       num_bytes, 1);
5626                         update_bytes_may_use(space_info, num_bytes);
5627                         ticket->bytes -= num_bytes;
5628                         num_bytes = 0;
5629                 }
5630         }
5631
5632         if (num_bytes && head == &space_info->priority_tickets) {
5633                 head = &space_info->tickets;
5634                 goto again;
5635         }
5636 }
5637
5638 static u64 block_rsv_release_bytes(struct btrfs_fs_info *fs_info,
5639                                     struct btrfs_block_rsv *block_rsv,
5640                                     struct btrfs_block_rsv *dest, u64 num_bytes,
5641                                     u64 *qgroup_to_release_ret)
5642 {
5643         struct btrfs_space_info *space_info = block_rsv->space_info;
5644         u64 qgroup_to_release = 0;
5645         u64 ret;
5646
5647         spin_lock(&block_rsv->lock);
5648         if (num_bytes == (u64)-1) {
5649                 num_bytes = block_rsv->size;
5650                 qgroup_to_release = block_rsv->qgroup_rsv_size;
5651         }
5652         block_rsv->size -= num_bytes;
5653         if (block_rsv->reserved >= block_rsv->size) {
5654                 num_bytes = block_rsv->reserved - block_rsv->size;
5655                 block_rsv->reserved = block_rsv->size;
5656                 block_rsv->full = 1;
5657         } else {
5658                 num_bytes = 0;
5659         }
5660         if (block_rsv->qgroup_rsv_reserved >= block_rsv->qgroup_rsv_size) {
5661                 qgroup_to_release = block_rsv->qgroup_rsv_reserved -
5662                                     block_rsv->qgroup_rsv_size;
5663                 block_rsv->qgroup_rsv_reserved = block_rsv->qgroup_rsv_size;
5664         } else {
5665                 qgroup_to_release = 0;
5666         }
5667         spin_unlock(&block_rsv->lock);
5668
5669         ret = num_bytes;
5670         if (num_bytes > 0) {
5671                 if (dest) {
5672                         spin_lock(&dest->lock);
5673                         if (!dest->full) {
5674                                 u64 bytes_to_add;
5675
5676                                 bytes_to_add = dest->size - dest->reserved;
5677                                 bytes_to_add = min(num_bytes, bytes_to_add);
5678                                 dest->reserved += bytes_to_add;
5679                                 if (dest->reserved >= dest->size)
5680                                         dest->full = 1;
5681                                 num_bytes -= bytes_to_add;
5682                         }
5683                         spin_unlock(&dest->lock);
5684                 }
5685                 if (num_bytes)
5686                         space_info_add_old_bytes(fs_info, space_info,
5687                                                  num_bytes);
5688         }
5689         if (qgroup_to_release_ret)
5690                 *qgroup_to_release_ret = qgroup_to_release;
5691         return ret;
5692 }
5693
5694 int btrfs_block_rsv_migrate(struct btrfs_block_rsv *src,
5695                             struct btrfs_block_rsv *dst, u64 num_bytes,
5696                             bool update_size)
5697 {
5698         int ret;
5699
5700         ret = block_rsv_use_bytes(src, num_bytes);
5701         if (ret)
5702                 return ret;
5703
5704         block_rsv_add_bytes(dst, num_bytes, update_size);
5705         return 0;
5706 }
5707
5708 void btrfs_init_block_rsv(struct btrfs_block_rsv *rsv, unsigned short type)
5709 {
5710         memset(rsv, 0, sizeof(*rsv));
5711         spin_lock_init(&rsv->lock);
5712         rsv->type = type;
5713 }
5714
5715 void btrfs_init_metadata_block_rsv(struct btrfs_fs_info *fs_info,
5716                                    struct btrfs_block_rsv *rsv,
5717                                    unsigned short type)
5718 {
5719         btrfs_init_block_rsv(rsv, type);
5720         rsv->space_info = __find_space_info(fs_info,
5721                                             BTRFS_BLOCK_GROUP_METADATA);
5722 }
5723
5724 struct btrfs_block_rsv *btrfs_alloc_block_rsv(struct btrfs_fs_info *fs_info,
5725                                               unsigned short type)
5726 {
5727         struct btrfs_block_rsv *block_rsv;
5728
5729         block_rsv = kmalloc(sizeof(*block_rsv), GFP_NOFS);
5730         if (!block_rsv)
5731                 return NULL;
5732
5733         btrfs_init_metadata_block_rsv(fs_info, block_rsv, type);
5734         return block_rsv;
5735 }
5736
5737 void btrfs_free_block_rsv(struct btrfs_fs_info *fs_info,
5738                           struct btrfs_block_rsv *rsv)
5739 {
5740         if (!rsv)
5741                 return;
5742         btrfs_block_rsv_release(fs_info, rsv, (u64)-1);
5743         kfree(rsv);
5744 }
5745
5746 int btrfs_block_rsv_add(struct btrfs_root *root,
5747                         struct btrfs_block_rsv *block_rsv, u64 num_bytes,
5748                         enum btrfs_reserve_flush_enum flush)
5749 {
5750         int ret;
5751
5752         if (num_bytes == 0)
5753                 return 0;
5754
5755         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
5756         if (!ret)
5757                 block_rsv_add_bytes(block_rsv, num_bytes, true);
5758
5759         return ret;
5760 }
5761
5762 int btrfs_block_rsv_check(struct btrfs_block_rsv *block_rsv, int min_factor)
5763 {
5764         u64 num_bytes = 0;
5765         int ret = -ENOSPC;
5766
5767         if (!block_rsv)
5768                 return 0;
5769
5770         spin_lock(&block_rsv->lock);
5771         num_bytes = div_factor(block_rsv->size, min_factor);
5772         if (block_rsv->reserved >= num_bytes)
5773                 ret = 0;
5774         spin_unlock(&block_rsv->lock);
5775
5776         return ret;
5777 }
5778
5779 int btrfs_block_rsv_refill(struct btrfs_root *root,
5780                            struct btrfs_block_rsv *block_rsv, u64 min_reserved,
5781                            enum btrfs_reserve_flush_enum flush)
5782 {
5783         u64 num_bytes = 0;
5784         int ret = -ENOSPC;
5785
5786         if (!block_rsv)
5787                 return 0;
5788
5789         spin_lock(&block_rsv->lock);
5790         num_bytes = min_reserved;
5791         if (block_rsv->reserved >= num_bytes)
5792                 ret = 0;
5793         else
5794                 num_bytes -= block_rsv->reserved;
5795         spin_unlock(&block_rsv->lock);
5796
5797         if (!ret)
5798                 return 0;
5799
5800         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
5801         if (!ret) {
5802                 block_rsv_add_bytes(block_rsv, num_bytes, false);
5803                 return 0;
5804         }
5805
5806         return ret;
5807 }
5808
5809 static void calc_refill_bytes(struct btrfs_block_rsv *block_rsv,
5810                                 u64 *metadata_bytes, u64 *qgroup_bytes)
5811 {
5812         *metadata_bytes = 0;
5813         *qgroup_bytes = 0;
5814
5815         spin_lock(&block_rsv->lock);
5816         if (block_rsv->reserved < block_rsv->size)
5817                 *metadata_bytes = block_rsv->size - block_rsv->reserved;
5818         if (block_rsv->qgroup_rsv_reserved < block_rsv->qgroup_rsv_size)
5819                 *qgroup_bytes = block_rsv->qgroup_rsv_size -
5820                         block_rsv->qgroup_rsv_reserved;
5821         spin_unlock(&block_rsv->lock);
5822 }
5823
5824 /**
5825  * btrfs_inode_rsv_refill - refill the inode block rsv.
5826  * @inode - the inode we are refilling.
5827  * @flush - the flushing restriction.
5828  *
5829  * Essentially the same as btrfs_block_rsv_refill, except it uses the
5830  * block_rsv->size as the minimum size.  We'll either refill the missing amount
5831  * or return if we already have enough space.  This will also handle the reserve
5832  * tracepoint for the reserved amount.
5833  */
5834 static int btrfs_inode_rsv_refill(struct btrfs_inode *inode,
5835                                   enum btrfs_reserve_flush_enum flush)
5836 {
5837         struct btrfs_root *root = inode->root;
5838         struct btrfs_block_rsv *block_rsv = &inode->block_rsv;
5839         u64 num_bytes, last = 0;
5840         u64 qgroup_num_bytes;
5841         int ret = -ENOSPC;
5842
5843         calc_refill_bytes(block_rsv, &num_bytes, &qgroup_num_bytes);
5844         if (num_bytes == 0)
5845                 return 0;
5846
5847         do {
5848                 ret = btrfs_qgroup_reserve_meta_prealloc(root, qgroup_num_bytes,
5849                                                          true);
5850                 if (ret)
5851                         return ret;
5852                 ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
5853                 if (ret) {
5854                         btrfs_qgroup_free_meta_prealloc(root, qgroup_num_bytes);
5855                         last = num_bytes;
5856                         /*
5857                          * If we are fragmented we can end up with a lot of
5858                          * outstanding extents which will make our size be much
5859                          * larger than our reserved amount.
5860                          *
5861                          * If the reservation happens here, it might be very
5862                          * big though not needed in the end, if the delalloc
5863                          * flushing happens.
5864                          *
5865                          * If this is the case try and do the reserve again.
5866                          */
5867                         if (flush == BTRFS_RESERVE_FLUSH_ALL)
5868                                 calc_refill_bytes(block_rsv, &num_bytes,
5869                                                    &qgroup_num_bytes);
5870                         if (num_bytes == 0)
5871                                 return 0;
5872                 }
5873         } while (ret && last != num_bytes);
5874
5875         if (!ret) {
5876                 block_rsv_add_bytes(block_rsv, num_bytes, false);
5877                 trace_btrfs_space_reservation(root->fs_info, "delalloc",
5878                                               btrfs_ino(inode), num_bytes, 1);
5879
5880                 /* Don't forget to increase qgroup_rsv_reserved */
5881                 spin_lock(&block_rsv->lock);
5882                 block_rsv->qgroup_rsv_reserved += qgroup_num_bytes;
5883                 spin_unlock(&block_rsv->lock);
5884         }
5885         return ret;
5886 }
5887
5888 static u64 __btrfs_block_rsv_release(struct btrfs_fs_info *fs_info,
5889                                      struct btrfs_block_rsv *block_rsv,
5890                                      u64 num_bytes, u64 *qgroup_to_release)
5891 {
5892         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
5893         struct btrfs_block_rsv *delayed_rsv = &fs_info->delayed_refs_rsv;
5894         struct btrfs_block_rsv *target = delayed_rsv;
5895
5896         if (target->full || target == block_rsv)
5897                 target = global_rsv;
5898
5899         if (block_rsv->space_info != target->space_info)
5900                 target = NULL;
5901
5902         return block_rsv_release_bytes(fs_info, block_rsv, target, num_bytes,
5903                                        qgroup_to_release);
5904 }
5905
5906 void btrfs_block_rsv_release(struct btrfs_fs_info *fs_info,
5907                              struct btrfs_block_rsv *block_rsv,
5908                              u64 num_bytes)
5909 {
5910         __btrfs_block_rsv_release(fs_info, block_rsv, num_bytes, NULL);
5911 }
5912
5913 /**
5914  * btrfs_inode_rsv_release - release any excessive reservation.
5915  * @inode - the inode we need to release from.
5916  * @qgroup_free - free or convert qgroup meta.
5917  *   Unlike normal operation, qgroup meta reservation needs to know if we are
5918  *   freeing qgroup reservation or just converting it into per-trans.  Normally
5919  *   @qgroup_free is true for error handling, and false for normal release.
5920  *
5921  * This is the same as btrfs_block_rsv_release, except that it handles the
5922  * tracepoint for the reservation.
5923  */
5924 static void btrfs_inode_rsv_release(struct btrfs_inode *inode, bool qgroup_free)
5925 {
5926         struct btrfs_fs_info *fs_info = inode->root->fs_info;
5927         struct btrfs_block_rsv *block_rsv = &inode->block_rsv;
5928         u64 released = 0;
5929         u64 qgroup_to_release = 0;
5930
5931         /*
5932          * Since we statically set the block_rsv->size we just want to say we
5933          * are releasing 0 bytes, and then we'll just get the reservation over
5934          * the size free'd.
5935          */
5936         released = __btrfs_block_rsv_release(fs_info, block_rsv, 0,
5937                                              &qgroup_to_release);
5938         if (released > 0)
5939                 trace_btrfs_space_reservation(fs_info, "delalloc",
5940                                               btrfs_ino(inode), released, 0);
5941         if (qgroup_free)
5942                 btrfs_qgroup_free_meta_prealloc(inode->root, qgroup_to_release);
5943         else
5944                 btrfs_qgroup_convert_reserved_meta(inode->root,
5945                                                    qgroup_to_release);
5946 }
5947
5948 /**
5949  * btrfs_delayed_refs_rsv_release - release a ref head's reservation.
5950  * @fs_info - the fs_info for our fs.
5951  * @nr - the number of items to drop.
5952  *
5953  * This drops the delayed ref head's count from the delayed refs rsv and frees
5954  * any excess reservation we had.
5955  */
5956 void btrfs_delayed_refs_rsv_release(struct btrfs_fs_info *fs_info, int nr)
5957 {
5958         struct btrfs_block_rsv *block_rsv = &fs_info->delayed_refs_rsv;
5959         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
5960         u64 num_bytes = btrfs_calc_trans_metadata_size(fs_info, nr);
5961         u64 released = 0;
5962
5963         released = block_rsv_release_bytes(fs_info, block_rsv, global_rsv,
5964                                            num_bytes, NULL);
5965         if (released)
5966                 trace_btrfs_space_reservation(fs_info, "delayed_refs_rsv",
5967                                               0, released, 0);
5968 }
5969
5970 static void update_global_block_rsv(struct btrfs_fs_info *fs_info)
5971 {
5972         struct btrfs_block_rsv *block_rsv = &fs_info->global_block_rsv;
5973         struct btrfs_space_info *sinfo = block_rsv->space_info;
5974         u64 num_bytes;
5975
5976         /*
5977          * The global block rsv is based on the size of the extent tree, the
5978          * checksum tree and the root tree.  If the fs is empty we want to set
5979          * it to a minimal amount for safety.
5980          */
5981         num_bytes = btrfs_root_used(&fs_info->extent_root->root_item) +
5982                 btrfs_root_used(&fs_info->csum_root->root_item) +
5983                 btrfs_root_used(&fs_info->tree_root->root_item);
5984         num_bytes = max_t(u64, num_bytes, SZ_16M);
5985
5986         spin_lock(&sinfo->lock);
5987         spin_lock(&block_rsv->lock);
5988
5989         block_rsv->size = min_t(u64, num_bytes, SZ_512M);
5990
5991         if (block_rsv->reserved < block_rsv->size) {
5992                 num_bytes = btrfs_space_info_used(sinfo, true);
5993                 if (sinfo->total_bytes > num_bytes) {
5994                         num_bytes = sinfo->total_bytes - num_bytes;
5995                         num_bytes = min(num_bytes,
5996                                         block_rsv->size - block_rsv->reserved);
5997                         block_rsv->reserved += num_bytes;
5998                         update_bytes_may_use(sinfo, num_bytes);
5999                         trace_btrfs_space_reservation(fs_info, "space_info",
6000                                                       sinfo->flags, num_bytes,
6001                                                       1);
6002                 }
6003         } else if (block_rsv->reserved > block_rsv->size) {
6004                 num_bytes = block_rsv->reserved - block_rsv->size;
6005                 update_bytes_may_use(sinfo, -num_bytes);
6006                 trace_btrfs_space_reservation(fs_info, "space_info",
6007                                       sinfo->flags, num_bytes, 0);
6008                 block_rsv->reserved = block_rsv->size;
6009         }
6010
6011         if (block_rsv->reserved == block_rsv->size)
6012                 block_rsv->full = 1;
6013         else
6014                 block_rsv->full = 0;
6015
6016         spin_unlock(&block_rsv->lock);
6017         spin_unlock(&sinfo->lock);
6018 }
6019
6020 static void init_global_block_rsv(struct btrfs_fs_info *fs_info)
6021 {
6022         struct btrfs_space_info *space_info;
6023
6024         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
6025         fs_info->chunk_block_rsv.space_info = space_info;
6026
6027         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
6028         fs_info->global_block_rsv.space_info = space_info;
6029         fs_info->trans_block_rsv.space_info = space_info;
6030         fs_info->empty_block_rsv.space_info = space_info;
6031         fs_info->delayed_block_rsv.space_info = space_info;
6032         fs_info->delayed_refs_rsv.space_info = space_info;
6033
6034         fs_info->extent_root->block_rsv = &fs_info->delayed_refs_rsv;
6035         fs_info->csum_root->block_rsv = &fs_info->delayed_refs_rsv;
6036         fs_info->dev_root->block_rsv = &fs_info->global_block_rsv;
6037         fs_info->tree_root->block_rsv = &fs_info->global_block_rsv;
6038         if (fs_info->quota_root)
6039                 fs_info->quota_root->block_rsv = &fs_info->global_block_rsv;
6040         fs_info->chunk_root->block_rsv = &fs_info->chunk_block_rsv;
6041
6042         update_global_block_rsv(fs_info);
6043 }
6044
6045 static void release_global_block_rsv(struct btrfs_fs_info *fs_info)
6046 {
6047         block_rsv_release_bytes(fs_info, &fs_info->global_block_rsv, NULL,
6048                                 (u64)-1, NULL);
6049         WARN_ON(fs_info->trans_block_rsv.size > 0);
6050         WARN_ON(fs_info->trans_block_rsv.reserved > 0);
6051         WARN_ON(fs_info->chunk_block_rsv.size > 0);
6052         WARN_ON(fs_info->chunk_block_rsv.reserved > 0);
6053         WARN_ON(fs_info->delayed_block_rsv.size > 0);
6054         WARN_ON(fs_info->delayed_block_rsv.reserved > 0);
6055         WARN_ON(fs_info->delayed_refs_rsv.reserved > 0);
6056         WARN_ON(fs_info->delayed_refs_rsv.size > 0);
6057 }
6058
6059 /*
6060  * btrfs_update_delayed_refs_rsv - adjust the size of the delayed refs rsv
6061  * @trans - the trans that may have generated delayed refs
6062  *
6063  * This is to be called anytime we may have adjusted trans->delayed_ref_updates,
6064  * it'll calculate the additional size and add it to the delayed_refs_rsv.
6065  */
6066 void btrfs_update_delayed_refs_rsv(struct btrfs_trans_handle *trans)
6067 {
6068         struct btrfs_fs_info *fs_info = trans->fs_info;
6069         struct btrfs_block_rsv *delayed_rsv = &fs_info->delayed_refs_rsv;
6070         u64 num_bytes;
6071
6072         if (!trans->delayed_ref_updates)
6073                 return;
6074
6075         num_bytes = btrfs_calc_trans_metadata_size(fs_info,
6076                                                    trans->delayed_ref_updates);
6077         spin_lock(&delayed_rsv->lock);
6078         delayed_rsv->size += num_bytes;
6079         delayed_rsv->full = 0;
6080         spin_unlock(&delayed_rsv->lock);
6081         trans->delayed_ref_updates = 0;
6082 }
6083
6084 /*
6085  * To be called after all the new block groups attached to the transaction
6086  * handle have been created (btrfs_create_pending_block_groups()).
6087  */
6088 void btrfs_trans_release_chunk_metadata(struct btrfs_trans_handle *trans)
6089 {
6090         struct btrfs_fs_info *fs_info = trans->fs_info;
6091
6092         if (!trans->chunk_bytes_reserved)
6093                 return;
6094
6095         WARN_ON_ONCE(!list_empty(&trans->new_bgs));
6096
6097         block_rsv_release_bytes(fs_info, &fs_info->chunk_block_rsv, NULL,
6098                                 trans->chunk_bytes_reserved, NULL);
6099         trans->chunk_bytes_reserved = 0;
6100 }
6101
6102 /*
6103  * btrfs_subvolume_reserve_metadata() - reserve space for subvolume operation
6104  * root: the root of the parent directory
6105  * rsv: block reservation
6106  * items: the number of items that we need do reservation
6107  * use_global_rsv: allow fallback to the global block reservation
6108  *
6109  * This function is used to reserve the space for snapshot/subvolume
6110  * creation and deletion. Those operations are different with the
6111  * common file/directory operations, they change two fs/file trees
6112  * and root tree, the number of items that the qgroup reserves is
6113  * different with the free space reservation. So we can not use
6114  * the space reservation mechanism in start_transaction().
6115  */
6116 int btrfs_subvolume_reserve_metadata(struct btrfs_root *root,
6117                                      struct btrfs_block_rsv *rsv, int items,
6118                                      bool use_global_rsv)
6119 {
6120         u64 qgroup_num_bytes = 0;
6121         u64 num_bytes;
6122         int ret;
6123         struct btrfs_fs_info *fs_info = root->fs_info;
6124         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
6125
6126         if (test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags)) {
6127                 /* One for parent inode, two for dir entries */
6128                 qgroup_num_bytes = 3 * fs_info->nodesize;
6129                 ret = btrfs_qgroup_reserve_meta_prealloc(root,
6130                                 qgroup_num_bytes, true);
6131                 if (ret)
6132                         return ret;
6133         }
6134
6135         num_bytes = btrfs_calc_trans_metadata_size(fs_info, items);
6136         rsv->space_info = __find_space_info(fs_info,
6137                                             BTRFS_BLOCK_GROUP_METADATA);
6138         ret = btrfs_block_rsv_add(root, rsv, num_bytes,
6139                                   BTRFS_RESERVE_FLUSH_ALL);
6140
6141         if (ret == -ENOSPC && use_global_rsv)
6142                 ret = btrfs_block_rsv_migrate(global_rsv, rsv, num_bytes, true);
6143
6144         if (ret && qgroup_num_bytes)
6145                 btrfs_qgroup_free_meta_prealloc(root, qgroup_num_bytes);
6146
6147         return ret;
6148 }
6149
6150 void btrfs_subvolume_release_metadata(struct btrfs_fs_info *fs_info,
6151                                       struct btrfs_block_rsv *rsv)
6152 {
6153         btrfs_block_rsv_release(fs_info, rsv, (u64)-1);
6154 }
6155
6156 static void btrfs_calculate_inode_block_rsv_size(struct btrfs_fs_info *fs_info,
6157                                                  struct btrfs_inode *inode)
6158 {
6159         struct btrfs_block_rsv *block_rsv = &inode->block_rsv;
6160         u64 reserve_size = 0;
6161         u64 qgroup_rsv_size = 0;
6162         u64 csum_leaves;
6163         unsigned outstanding_extents;
6164
6165         lockdep_assert_held(&inode->lock);
6166         outstanding_extents = inode->outstanding_extents;
6167         if (outstanding_extents)
6168                 reserve_size = btrfs_calc_trans_metadata_size(fs_info,
6169                                                 outstanding_extents + 1);
6170         csum_leaves = btrfs_csum_bytes_to_leaves(fs_info,
6171                                                  inode->csum_bytes);
6172         reserve_size += btrfs_calc_trans_metadata_size(fs_info,
6173                                                        csum_leaves);
6174         /*
6175          * For qgroup rsv, the calculation is very simple:
6176          * account one nodesize for each outstanding extent
6177          *
6178          * This is overestimating in most cases.
6179          */
6180         qgroup_rsv_size = outstanding_extents * fs_info->nodesize;
6181
6182         spin_lock(&block_rsv->lock);
6183         block_rsv->size = reserve_size;
6184         block_rsv->qgroup_rsv_size = qgroup_rsv_size;
6185         spin_unlock(&block_rsv->lock);
6186 }
6187
6188 int btrfs_delalloc_reserve_metadata(struct btrfs_inode *inode, u64 num_bytes)
6189 {
6190         struct btrfs_fs_info *fs_info = inode->root->fs_info;
6191         unsigned nr_extents;
6192         enum btrfs_reserve_flush_enum flush = BTRFS_RESERVE_FLUSH_ALL;
6193         int ret = 0;
6194         bool delalloc_lock = true;
6195
6196         /* If we are a free space inode we need to not flush since we will be in
6197          * the middle of a transaction commit.  We also don't need the delalloc
6198          * mutex since we won't race with anybody.  We need this mostly to make
6199          * lockdep shut its filthy mouth.
6200          *
6201          * If we have a transaction open (can happen if we call truncate_block
6202          * from truncate), then we need FLUSH_LIMIT so we don't deadlock.
6203          */
6204         if (btrfs_is_free_space_inode(inode)) {
6205                 flush = BTRFS_RESERVE_NO_FLUSH;
6206                 delalloc_lock = false;
6207         } else {
6208                 if (current->journal_info)
6209                         flush = BTRFS_RESERVE_FLUSH_LIMIT;
6210
6211                 if (btrfs_transaction_in_commit(fs_info))
6212                         schedule_timeout(1);
6213         }
6214
6215         if (delalloc_lock)
6216                 mutex_lock(&inode->delalloc_mutex);
6217
6218         num_bytes = ALIGN(num_bytes, fs_info->sectorsize);
6219
6220         /* Add our new extents and calculate the new rsv size. */
6221         spin_lock(&inode->lock);
6222         nr_extents = count_max_extents(num_bytes);
6223         btrfs_mod_outstanding_extents(inode, nr_extents);
6224         inode->csum_bytes += num_bytes;
6225         btrfs_calculate_inode_block_rsv_size(fs_info, inode);
6226         spin_unlock(&inode->lock);
6227
6228         ret = btrfs_inode_rsv_refill(inode, flush);
6229         if (unlikely(ret))
6230                 goto out_fail;
6231
6232         if (delalloc_lock)
6233                 mutex_unlock(&inode->delalloc_mutex);
6234         return 0;
6235
6236 out_fail:
6237         spin_lock(&inode->lock);
6238         nr_extents = count_max_extents(num_bytes);
6239         btrfs_mod_outstanding_extents(inode, -nr_extents);
6240         inode->csum_bytes -= num_bytes;
6241         btrfs_calculate_inode_block_rsv_size(fs_info, inode);
6242         spin_unlock(&inode->lock);
6243
6244         btrfs_inode_rsv_release(inode, true);
6245         if (delalloc_lock)
6246                 mutex_unlock(&inode->delalloc_mutex);
6247         return ret;
6248 }
6249
6250 /**
6251  * btrfs_delalloc_release_metadata - release a metadata reservation for an inode
6252  * @inode: the inode to release the reservation for.
6253  * @num_bytes: the number of bytes we are releasing.
6254  * @qgroup_free: free qgroup reservation or convert it to per-trans reservation
6255  *
6256  * This will release the metadata reservation for an inode.  This can be called
6257  * once we complete IO for a given set of bytes to release their metadata
6258  * reservations, or on error for the same reason.
6259  */
6260 void btrfs_delalloc_release_metadata(struct btrfs_inode *inode, u64 num_bytes,
6261                                      bool qgroup_free)
6262 {
6263         struct btrfs_fs_info *fs_info = inode->root->fs_info;
6264
6265         num_bytes = ALIGN(num_bytes, fs_info->sectorsize);
6266         spin_lock(&inode->lock);
6267         inode->csum_bytes -= num_bytes;
6268         btrfs_calculate_inode_block_rsv_size(fs_info, inode);
6269         spin_unlock(&inode->lock);
6270
6271         if (btrfs_is_testing(fs_info))
6272                 return;
6273
6274         btrfs_inode_rsv_release(inode, qgroup_free);
6275 }
6276
6277 /**
6278  * btrfs_delalloc_release_extents - release our outstanding_extents
6279  * @inode: the inode to balance the reservation for.
6280  * @num_bytes: the number of bytes we originally reserved with
6281  * @qgroup_free: do we need to free qgroup meta reservation or convert them.
6282  *
6283  * When we reserve space we increase outstanding_extents for the extents we may
6284  * add.  Once we've set the range as delalloc or created our ordered extents we
6285  * have outstanding_extents to track the real usage, so we use this to free our
6286  * temporarily tracked outstanding_extents.  This _must_ be used in conjunction
6287  * with btrfs_delalloc_reserve_metadata.
6288  */
6289 void btrfs_delalloc_release_extents(struct btrfs_inode *inode, u64 num_bytes,
6290                                     bool qgroup_free)
6291 {
6292         struct btrfs_fs_info *fs_info = inode->root->fs_info;
6293         unsigned num_extents;
6294
6295         spin_lock(&inode->lock);
6296         num_extents = count_max_extents(num_bytes);
6297         btrfs_mod_outstanding_extents(inode, -num_extents);
6298         btrfs_calculate_inode_block_rsv_size(fs_info, inode);
6299         spin_unlock(&inode->lock);
6300
6301         if (btrfs_is_testing(fs_info))
6302                 return;
6303
6304         btrfs_inode_rsv_release(inode, qgroup_free);
6305 }
6306
6307 /**
6308  * btrfs_delalloc_reserve_space - reserve data and metadata space for
6309  * delalloc
6310  * @inode: inode we're writing to
6311  * @start: start range we are writing to
6312  * @len: how long the range we are writing to
6313  * @reserved: mandatory parameter, record actually reserved qgroup ranges of
6314  *            current reservation.
6315  *
6316  * This will do the following things
6317  *
6318  * o reserve space in data space info for num bytes
6319  *   and reserve precious corresponding qgroup space
6320  *   (Done in check_data_free_space)
6321  *
6322  * o reserve space for metadata space, based on the number of outstanding
6323  *   extents and how much csums will be needed
6324  *   also reserve metadata space in a per root over-reserve method.
6325  * o add to the inodes->delalloc_bytes
6326  * o add it to the fs_info's delalloc inodes list.
6327  *   (Above 3 all done in delalloc_reserve_metadata)
6328  *
6329  * Return 0 for success
6330  * Return <0 for error(-ENOSPC or -EQUOT)
6331  */
6332 int btrfs_delalloc_reserve_space(struct inode *inode,
6333                         struct extent_changeset **reserved, u64 start, u64 len)
6334 {
6335         int ret;
6336
6337         ret = btrfs_check_data_free_space(inode, reserved, start, len);
6338         if (ret < 0)
6339                 return ret;
6340         ret = btrfs_delalloc_reserve_metadata(BTRFS_I(inode), len);
6341         if (ret < 0)
6342                 btrfs_free_reserved_data_space(inode, *reserved, start, len);
6343         return ret;
6344 }
6345
6346 /**
6347  * btrfs_delalloc_release_space - release data and metadata space for delalloc
6348  * @inode: inode we're releasing space for
6349  * @start: start position of the space already reserved
6350  * @len: the len of the space already reserved
6351  * @release_bytes: the len of the space we consumed or didn't use
6352  *
6353  * This function will release the metadata space that was not used and will
6354  * decrement ->delalloc_bytes and remove it from the fs_info delalloc_inodes
6355  * list if there are no delalloc bytes left.
6356  * Also it will handle the qgroup reserved space.
6357  */
6358 void btrfs_delalloc_release_space(struct inode *inode,
6359                                   struct extent_changeset *reserved,
6360                                   u64 start, u64 len, bool qgroup_free)
6361 {
6362         btrfs_delalloc_release_metadata(BTRFS_I(inode), len, qgroup_free);
6363         btrfs_free_reserved_data_space(inode, reserved, start, len);
6364 }
6365
6366 static int update_block_group(struct btrfs_trans_handle *trans,
6367                               struct btrfs_fs_info *info, u64 bytenr,
6368                               u64 num_bytes, int alloc)
6369 {
6370         struct btrfs_block_group_cache *cache = NULL;
6371         u64 total = num_bytes;
6372         u64 old_val;
6373         u64 byte_in_group;
6374         int factor;
6375         int ret = 0;
6376
6377         /* block accounting for super block */
6378         spin_lock(&info->delalloc_root_lock);
6379         old_val = btrfs_super_bytes_used(info->super_copy);
6380         if (alloc)
6381                 old_val += num_bytes;
6382         else
6383                 old_val -= num_bytes;
6384         btrfs_set_super_bytes_used(info->super_copy, old_val);
6385         spin_unlock(&info->delalloc_root_lock);
6386
6387         while (total) {
6388                 cache = btrfs_lookup_block_group(info, bytenr);
6389                 if (!cache) {
6390                         ret = -ENOENT;
6391                         break;
6392                 }
6393                 factor = btrfs_bg_type_to_factor(cache->flags);
6394
6395                 /*
6396                  * If this block group has free space cache written out, we
6397                  * need to make sure to load it if we are removing space.  This
6398                  * is because we need the unpinning stage to actually add the
6399                  * space back to the block group, otherwise we will leak space.
6400                  */
6401                 if (!alloc && cache->cached == BTRFS_CACHE_NO)
6402                         cache_block_group(cache, 1);
6403
6404                 byte_in_group = bytenr - cache->key.objectid;
6405                 WARN_ON(byte_in_group > cache->key.offset);
6406
6407                 spin_lock(&cache->space_info->lock);
6408                 spin_lock(&cache->lock);
6409
6410                 if (btrfs_test_opt(info, SPACE_CACHE) &&
6411                     cache->disk_cache_state < BTRFS_DC_CLEAR)
6412                         cache->disk_cache_state = BTRFS_DC_CLEAR;
6413
6414                 old_val = btrfs_block_group_used(&cache->item);
6415                 num_bytes = min(total, cache->key.offset - byte_in_group);
6416                 if (alloc) {
6417                         old_val += num_bytes;
6418                         btrfs_set_block_group_used(&cache->item, old_val);
6419                         cache->reserved -= num_bytes;
6420                         cache->space_info->bytes_reserved -= num_bytes;
6421                         cache->space_info->bytes_used += num_bytes;
6422                         cache->space_info->disk_used += num_bytes * factor;
6423                         spin_unlock(&cache->lock);
6424                         spin_unlock(&cache->space_info->lock);
6425                 } else {
6426                         old_val -= num_bytes;
6427                         btrfs_set_block_group_used(&cache->item, old_val);
6428                         cache->pinned += num_bytes;
6429                         update_bytes_pinned(cache->space_info, num_bytes);
6430                         cache->space_info->bytes_used -= num_bytes;
6431                         cache->space_info->disk_used -= num_bytes * factor;
6432                         spin_unlock(&cache->lock);
6433                         spin_unlock(&cache->space_info->lock);
6434
6435                         trace_btrfs_space_reservation(info, "pinned",
6436                                                       cache->space_info->flags,
6437                                                       num_bytes, 1);
6438                         percpu_counter_add_batch(&cache->space_info->total_bytes_pinned,
6439                                            num_bytes,
6440                                            BTRFS_TOTAL_BYTES_PINNED_BATCH);
6441                         set_extent_dirty(info->pinned_extents,
6442                                          bytenr, bytenr + num_bytes - 1,
6443                                          GFP_NOFS | __GFP_NOFAIL);
6444                 }
6445
6446                 spin_lock(&trans->transaction->dirty_bgs_lock);
6447                 if (list_empty(&cache->dirty_list)) {
6448                         list_add_tail(&cache->dirty_list,
6449                                       &trans->transaction->dirty_bgs);
6450                         trans->transaction->num_dirty_bgs++;
6451                         trans->delayed_ref_updates++;
6452                         btrfs_get_block_group(cache);
6453                 }
6454                 spin_unlock(&trans->transaction->dirty_bgs_lock);
6455
6456                 /*
6457                  * No longer have used bytes in this block group, queue it for
6458                  * deletion. We do this after adding the block group to the
6459                  * dirty list to avoid races between cleaner kthread and space
6460                  * cache writeout.
6461                  */
6462                 if (!alloc && old_val == 0)
6463                         btrfs_mark_bg_unused(cache);
6464
6465                 btrfs_put_block_group(cache);
6466                 total -= num_bytes;
6467                 bytenr += num_bytes;
6468         }
6469
6470         /* Modified block groups are accounted for in the delayed_refs_rsv. */
6471         btrfs_update_delayed_refs_rsv(trans);
6472         return ret;
6473 }
6474
6475 static u64 first_logical_byte(struct btrfs_fs_info *fs_info, u64 search_start)
6476 {
6477         struct btrfs_block_group_cache *cache;
6478         u64 bytenr;
6479
6480         spin_lock(&fs_info->block_group_cache_lock);
6481         bytenr = fs_info->first_logical_byte;
6482         spin_unlock(&fs_info->block_group_cache_lock);
6483
6484         if (bytenr < (u64)-1)
6485                 return bytenr;
6486
6487         cache = btrfs_lookup_first_block_group(fs_info, search_start);
6488         if (!cache)
6489                 return 0;
6490
6491         bytenr = cache->key.objectid;
6492         btrfs_put_block_group(cache);
6493
6494         return bytenr;
6495 }
6496
6497 static int pin_down_extent(struct btrfs_fs_info *fs_info,
6498                            struct btrfs_block_group_cache *cache,
6499                            u64 bytenr, u64 num_bytes, int reserved)
6500 {
6501         spin_lock(&cache->space_info->lock);
6502         spin_lock(&cache->lock);
6503         cache->pinned += num_bytes;
6504         update_bytes_pinned(cache->space_info, num_bytes);
6505         if (reserved) {
6506                 cache->reserved -= num_bytes;
6507                 cache->space_info->bytes_reserved -= num_bytes;
6508         }
6509         spin_unlock(&cache->lock);
6510         spin_unlock(&cache->space_info->lock);
6511
6512         trace_btrfs_space_reservation(fs_info, "pinned",
6513                                       cache->space_info->flags, num_bytes, 1);
6514         percpu_counter_add_batch(&cache->space_info->total_bytes_pinned,
6515                     num_bytes, BTRFS_TOTAL_BYTES_PINNED_BATCH);
6516         set_extent_dirty(fs_info->pinned_extents, bytenr,
6517                          bytenr + num_bytes - 1, GFP_NOFS | __GFP_NOFAIL);
6518         return 0;
6519 }
6520
6521 /*
6522  * this function must be called within transaction
6523  */
6524 int btrfs_pin_extent(struct btrfs_fs_info *fs_info,
6525                      u64 bytenr, u64 num_bytes, int reserved)
6526 {
6527         struct btrfs_block_group_cache *cache;
6528
6529         cache = btrfs_lookup_block_group(fs_info, bytenr);
6530         BUG_ON(!cache); /* Logic error */
6531
6532         pin_down_extent(fs_info, cache, bytenr, num_bytes, reserved);
6533
6534         btrfs_put_block_group(cache);
6535         return 0;
6536 }
6537
6538 /*
6539  * this function must be called within transaction
6540  */
6541 int btrfs_pin_extent_for_log_replay(struct btrfs_fs_info *fs_info,
6542                                     u64 bytenr, u64 num_bytes)
6543 {
6544         struct btrfs_block_group_cache *cache;
6545         int ret;
6546
6547         cache = btrfs_lookup_block_group(fs_info, bytenr);
6548         if (!cache)
6549                 return -EINVAL;
6550
6551         /*
6552          * pull in the free space cache (if any) so that our pin
6553          * removes the free space from the cache.  We have load_only set
6554          * to one because the slow code to read in the free extents does check
6555          * the pinned extents.
6556          */
6557         cache_block_group(cache, 1);
6558
6559         pin_down_extent(fs_info, cache, bytenr, num_bytes, 0);
6560
6561         /* remove us from the free space cache (if we're there at all) */
6562         ret = btrfs_remove_free_space(cache, bytenr, num_bytes);
6563         btrfs_put_block_group(cache);
6564         return ret;
6565 }
6566
6567 static int __exclude_logged_extent(struct btrfs_fs_info *fs_info,
6568                                    u64 start, u64 num_bytes)
6569 {
6570         int ret;
6571         struct btrfs_block_group_cache *block_group;
6572         struct btrfs_caching_control *caching_ctl;
6573
6574         block_group = btrfs_lookup_block_group(fs_info, start);
6575         if (!block_group)
6576                 return -EINVAL;
6577
6578         cache_block_group(block_group, 0);
6579         caching_ctl = get_caching_control(block_group);
6580
6581         if (!caching_ctl) {
6582                 /* Logic error */
6583                 BUG_ON(!block_group_cache_done(block_group));
6584                 ret = btrfs_remove_free_space(block_group, start, num_bytes);
6585         } else {
6586                 mutex_lock(&caching_ctl->mutex);
6587
6588                 if (start >= caching_ctl->progress) {
6589                         ret = add_excluded_extent(fs_info, start, num_bytes);
6590                 } else if (start + num_bytes <= caching_ctl->progress) {
6591                         ret = btrfs_remove_free_space(block_group,
6592                                                       start, num_bytes);
6593                 } else {
6594                         num_bytes = caching_ctl->progress - start;
6595                         ret = btrfs_remove_free_space(block_group,
6596                                                       start, num_bytes);
6597                         if (ret)
6598                                 goto out_lock;
6599
6600                         num_bytes = (start + num_bytes) -
6601                                 caching_ctl->progress;
6602                         start = caching_ctl->progress;
6603                         ret = add_excluded_extent(fs_info, start, num_bytes);
6604                 }
6605 out_lock:
6606                 mutex_unlock(&caching_ctl->mutex);
6607                 put_caching_control(caching_ctl);
6608         }
6609         btrfs_put_block_group(block_group);
6610         return ret;
6611 }
6612
6613 int btrfs_exclude_logged_extents(struct btrfs_fs_info *fs_info,
6614                                  struct extent_buffer *eb)
6615 {
6616         struct btrfs_file_extent_item *item;
6617         struct btrfs_key key;
6618         int found_type;
6619         int i;
6620         int ret = 0;
6621
6622         if (!btrfs_fs_incompat(fs_info, MIXED_GROUPS))
6623                 return 0;
6624
6625         for (i = 0; i < btrfs_header_nritems(eb); i++) {
6626                 btrfs_item_key_to_cpu(eb, &key, i);
6627                 if (key.type != BTRFS_EXTENT_DATA_KEY)
6628                         continue;
6629                 item = btrfs_item_ptr(eb, i, struct btrfs_file_extent_item);
6630                 found_type = btrfs_file_extent_type(eb, item);
6631                 if (found_type == BTRFS_FILE_EXTENT_INLINE)
6632                         continue;
6633                 if (btrfs_file_extent_disk_bytenr(eb, item) == 0)
6634                         continue;
6635                 key.objectid = btrfs_file_extent_disk_bytenr(eb, item);
6636                 key.offset = btrfs_file_extent_disk_num_bytes(eb, item);
6637                 ret = __exclude_logged_extent(fs_info, key.objectid, key.offset);
6638                 if (ret)
6639                         break;
6640         }
6641
6642         return ret;
6643 }
6644
6645 static void
6646 btrfs_inc_block_group_reservations(struct btrfs_block_group_cache *bg)
6647 {
6648         atomic_inc(&bg->reservations);
6649 }
6650
6651 void btrfs_dec_block_group_reservations(struct btrfs_fs_info *fs_info,
6652                                         const u64 start)
6653 {
6654         struct btrfs_block_group_cache *bg;
6655
6656         bg = btrfs_lookup_block_group(fs_info, start);
6657         ASSERT(bg);
6658         if (atomic_dec_and_test(&bg->reservations))
6659                 wake_up_var(&bg->reservations);
6660         btrfs_put_block_group(bg);
6661 }
6662
6663 void btrfs_wait_block_group_reservations(struct btrfs_block_group_cache *bg)
6664 {
6665         struct btrfs_space_info *space_info = bg->space_info;
6666
6667         ASSERT(bg->ro);
6668
6669         if (!(bg->flags & BTRFS_BLOCK_GROUP_DATA))
6670                 return;
6671
6672         /*
6673          * Our block group is read only but before we set it to read only,
6674          * some task might have had allocated an extent from it already, but it
6675          * has not yet created a respective ordered extent (and added it to a
6676          * root's list of ordered extents).
6677          * Therefore wait for any task currently allocating extents, since the
6678          * block group's reservations counter is incremented while a read lock
6679          * on the groups' semaphore is held and decremented after releasing
6680          * the read access on that semaphore and creating the ordered extent.
6681          */
6682         down_write(&space_info->groups_sem);
6683         up_write(&space_info->groups_sem);
6684
6685         wait_var_event(&bg->reservations, !atomic_read(&bg->reservations));
6686 }
6687
6688 /**
6689  * btrfs_add_reserved_bytes - update the block_group and space info counters
6690  * @cache:      The cache we are manipulating
6691  * @ram_bytes:  The number of bytes of file content, and will be same to
6692  *              @num_bytes except for the compress path.
6693  * @num_bytes:  The number of bytes in question
6694  * @delalloc:   The blocks are allocated for the delalloc write
6695  *
6696  * This is called by the allocator when it reserves space. If this is a
6697  * reservation and the block group has become read only we cannot make the
6698  * reservation and return -EAGAIN, otherwise this function always succeeds.
6699  */
6700 static int btrfs_add_reserved_bytes(struct btrfs_block_group_cache *cache,
6701                                     u64 ram_bytes, u64 num_bytes, int delalloc)
6702 {
6703         struct btrfs_space_info *space_info = cache->space_info;
6704         int ret = 0;
6705
6706         spin_lock(&space_info->lock);
6707         spin_lock(&cache->lock);
6708         if (cache->ro) {
6709                 ret = -EAGAIN;
6710         } else {
6711                 cache->reserved += num_bytes;
6712                 space_info->bytes_reserved += num_bytes;
6713                 update_bytes_may_use(space_info, -ram_bytes);
6714                 if (delalloc)
6715                         cache->delalloc_bytes += num_bytes;
6716         }
6717         spin_unlock(&cache->lock);
6718         spin_unlock(&space_info->lock);
6719         return ret;
6720 }
6721
6722 /**
6723  * btrfs_free_reserved_bytes - update the block_group and space info counters
6724  * @cache:      The cache we are manipulating
6725  * @num_bytes:  The number of bytes in question
6726  * @delalloc:   The blocks are allocated for the delalloc write
6727  *
6728  * This is called by somebody who is freeing space that was never actually used
6729  * on disk.  For example if you reserve some space for a new leaf in transaction
6730  * A and before transaction A commits you free that leaf, you call this with
6731  * reserve set to 0 in order to clear the reservation.
6732  */
6733
6734 static void btrfs_free_reserved_bytes(struct btrfs_block_group_cache *cache,
6735                                       u64 num_bytes, int delalloc)
6736 {
6737         struct btrfs_space_info *space_info = cache->space_info;
6738
6739         spin_lock(&space_info->lock);
6740         spin_lock(&cache->lock);
6741         if (cache->ro)
6742                 space_info->bytes_readonly += num_bytes;
6743         cache->reserved -= num_bytes;
6744         space_info->bytes_reserved -= num_bytes;
6745         space_info->max_extent_size = 0;
6746
6747         if (delalloc)
6748                 cache->delalloc_bytes -= num_bytes;
6749         spin_unlock(&cache->lock);
6750         spin_unlock(&space_info->lock);
6751 }
6752 void btrfs_prepare_extent_commit(struct btrfs_fs_info *fs_info)
6753 {
6754         struct btrfs_caching_control *next;
6755         struct btrfs_caching_control *caching_ctl;
6756         struct btrfs_block_group_cache *cache;
6757
6758         down_write(&fs_info->commit_root_sem);
6759
6760         list_for_each_entry_safe(caching_ctl, next,
6761                                  &fs_info->caching_block_groups, list) {
6762                 cache = caching_ctl->block_group;
6763                 if (block_group_cache_done(cache)) {
6764                         cache->last_byte_to_unpin = (u64)-1;
6765                         list_del_init(&caching_ctl->list);
6766                         put_caching_control(caching_ctl);
6767                 } else {
6768                         cache->last_byte_to_unpin = caching_ctl->progress;
6769                 }
6770         }
6771
6772         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
6773                 fs_info->pinned_extents = &fs_info->freed_extents[1];
6774         else
6775                 fs_info->pinned_extents = &fs_info->freed_extents[0];
6776
6777         up_write(&fs_info->commit_root_sem);
6778
6779         update_global_block_rsv(fs_info);
6780 }
6781
6782 /*
6783  * Returns the free cluster for the given space info and sets empty_cluster to
6784  * what it should be based on the mount options.
6785  */
6786 static struct btrfs_free_cluster *
6787 fetch_cluster_info(struct btrfs_fs_info *fs_info,
6788                    struct btrfs_space_info *space_info, u64 *empty_cluster)
6789 {
6790         struct btrfs_free_cluster *ret = NULL;
6791
6792         *empty_cluster = 0;
6793         if (btrfs_mixed_space_info(space_info))
6794                 return ret;
6795
6796         if (space_info->flags & BTRFS_BLOCK_GROUP_METADATA) {
6797                 ret = &fs_info->meta_alloc_cluster;
6798                 if (btrfs_test_opt(fs_info, SSD))
6799                         *empty_cluster = SZ_2M;
6800                 else
6801                         *empty_cluster = SZ_64K;
6802         } else if ((space_info->flags & BTRFS_BLOCK_GROUP_DATA) &&
6803                    btrfs_test_opt(fs_info, SSD_SPREAD)) {
6804                 *empty_cluster = SZ_2M;
6805                 ret = &fs_info->data_alloc_cluster;
6806         }
6807
6808         return ret;
6809 }
6810
6811 static int unpin_extent_range(struct btrfs_fs_info *fs_info,
6812                               u64 start, u64 end,
6813                               const bool return_free_space)
6814 {
6815         struct btrfs_block_group_cache *cache = NULL;
6816         struct btrfs_space_info *space_info;
6817         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
6818         struct btrfs_free_cluster *cluster = NULL;
6819         u64 len;
6820         u64 total_unpinned = 0;
6821         u64 empty_cluster = 0;
6822         bool readonly;
6823
6824         while (start <= end) {
6825                 readonly = false;
6826                 if (!cache ||
6827                     start >= cache->key.objectid + cache->key.offset) {
6828                         if (cache)
6829                                 btrfs_put_block_group(cache);
6830                         total_unpinned = 0;
6831                         cache = btrfs_lookup_block_group(fs_info, start);
6832                         BUG_ON(!cache); /* Logic error */
6833
6834                         cluster = fetch_cluster_info(fs_info,
6835                                                      cache->space_info,
6836                                                      &empty_cluster);
6837                         empty_cluster <<= 1;
6838                 }
6839
6840                 len = cache->key.objectid + cache->key.offset - start;
6841                 len = min(len, end + 1 - start);
6842
6843                 if (start < cache->last_byte_to_unpin) {
6844                         len = min(len, cache->last_byte_to_unpin - start);
6845                         if (return_free_space)
6846                                 btrfs_add_free_space(cache, start, len);
6847                 }
6848
6849                 start += len;
6850                 total_unpinned += len;
6851                 space_info = cache->space_info;
6852
6853                 /*
6854                  * If this space cluster has been marked as fragmented and we've
6855                  * unpinned enough in this block group to potentially allow a
6856                  * cluster to be created inside of it go ahead and clear the
6857                  * fragmented check.
6858                  */
6859                 if (cluster && cluster->fragmented &&
6860                     total_unpinned > empty_cluster) {
6861                         spin_lock(&cluster->lock);
6862                         cluster->fragmented = 0;
6863                         spin_unlock(&cluster->lock);
6864                 }
6865
6866                 spin_lock(&space_info->lock);
6867                 spin_lock(&cache->lock);
6868                 cache->pinned -= len;
6869                 update_bytes_pinned(space_info, -len);
6870
6871                 trace_btrfs_space_reservation(fs_info, "pinned",
6872                                               space_info->flags, len, 0);
6873                 space_info->max_extent_size = 0;
6874                 percpu_counter_add_batch(&space_info->total_bytes_pinned,
6875                             -len, BTRFS_TOTAL_BYTES_PINNED_BATCH);
6876                 if (cache->ro) {
6877                         space_info->bytes_readonly += len;
6878                         readonly = true;
6879                 }
6880                 spin_unlock(&cache->lock);
6881                 if (!readonly && return_free_space &&
6882                     global_rsv->space_info == space_info) {
6883                         u64 to_add = len;
6884
6885                         spin_lock(&global_rsv->lock);
6886                         if (!global_rsv->full) {
6887                                 to_add = min(len, global_rsv->size -
6888                                              global_rsv->reserved);
6889                                 global_rsv->reserved += to_add;
6890                                 update_bytes_may_use(space_info, to_add);
6891                                 if (global_rsv->reserved >= global_rsv->size)
6892                                         global_rsv->full = 1;
6893                                 trace_btrfs_space_reservation(fs_info,
6894                                                               "space_info",
6895                                                               space_info->flags,
6896                                                               to_add, 1);
6897                                 len -= to_add;
6898                         }
6899                         spin_unlock(&global_rsv->lock);
6900                         /* Add to any tickets we may have */
6901                         if (len)
6902                                 space_info_add_new_bytes(fs_info, space_info,
6903                                                          len);
6904                 }
6905                 spin_unlock(&space_info->lock);
6906         }
6907
6908         if (cache)
6909                 btrfs_put_block_group(cache);
6910         return 0;
6911 }
6912
6913 int btrfs_finish_extent_commit(struct btrfs_trans_handle *trans)
6914 {
6915         struct btrfs_fs_info *fs_info = trans->fs_info;
6916         struct btrfs_block_group_cache *block_group, *tmp;
6917         struct list_head *deleted_bgs;
6918         struct extent_io_tree *unpin;
6919         u64 start;
6920         u64 end;
6921         int ret;
6922
6923         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
6924                 unpin = &fs_info->freed_extents[1];
6925         else
6926                 unpin = &fs_info->freed_extents[0];
6927
6928         while (!trans->aborted) {
6929                 struct extent_state *cached_state = NULL;
6930
6931                 mutex_lock(&fs_info->unused_bg_unpin_mutex);
6932                 ret = find_first_extent_bit(unpin, 0, &start, &end,
6933                                             EXTENT_DIRTY, &cached_state);
6934                 if (ret) {
6935                         mutex_unlock(&fs_info->unused_bg_unpin_mutex);
6936                         break;
6937                 }
6938
6939                 if (btrfs_test_opt(fs_info, DISCARD))
6940                         ret = btrfs_discard_extent(fs_info, start,
6941                                                    end + 1 - start, NULL);
6942
6943                 clear_extent_dirty(unpin, start, end, &cached_state);
6944                 unpin_extent_range(fs_info, start, end, true);
6945                 mutex_unlock(&fs_info->unused_bg_unpin_mutex);
6946                 free_extent_state(cached_state);
6947                 cond_resched();
6948         }
6949
6950         /*
6951          * Transaction is finished.  We don't need the lock anymore.  We
6952          * do need to clean up the block groups in case of a transaction
6953          * abort.
6954          */
6955         deleted_bgs = &trans->transaction->deleted_bgs;
6956         list_for_each_entry_safe(block_group, tmp, deleted_bgs, bg_list) {
6957                 u64 trimmed = 0;
6958
6959                 ret = -EROFS;
6960                 if (!trans->aborted)
6961                         ret = btrfs_discard_extent(fs_info,
6962                                                    block_group->key.objectid,
6963                                                    block_group->key.offset,
6964                                                    &trimmed);
6965
6966                 list_del_init(&block_group->bg_list);
6967                 btrfs_put_block_group_trimming(block_group);
6968                 btrfs_put_block_group(block_group);
6969
6970                 if (ret) {
6971                         const char *errstr = btrfs_decode_error(ret);
6972                         btrfs_warn(fs_info,
6973                            "discard failed while removing blockgroup: errno=%d %s",
6974                                    ret, errstr);
6975                 }
6976         }
6977
6978         return 0;
6979 }
6980
6981 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
6982                                struct btrfs_delayed_ref_node *node, u64 parent,
6983                                u64 root_objectid, u64 owner_objectid,
6984                                u64 owner_offset, int refs_to_drop,
6985                                struct btrfs_delayed_extent_op *extent_op)
6986 {
6987         struct btrfs_fs_info *info = trans->fs_info;
6988         struct btrfs_key key;
6989         struct btrfs_path *path;
6990         struct btrfs_root *extent_root = info->extent_root;
6991         struct extent_buffer *leaf;
6992         struct btrfs_extent_item *ei;
6993         struct btrfs_extent_inline_ref *iref;
6994         int ret;
6995         int is_data;
6996         int extent_slot = 0;
6997         int found_extent = 0;
6998         int num_to_del = 1;
6999         u32 item_size;
7000         u64 refs;
7001         u64 bytenr = node->bytenr;
7002         u64 num_bytes = node->num_bytes;
7003         int last_ref = 0;
7004         bool skinny_metadata = btrfs_fs_incompat(info, SKINNY_METADATA);
7005
7006         path = btrfs_alloc_path();
7007         if (!path)
7008                 return -ENOMEM;
7009
7010         path->reada = READA_FORWARD;
7011         path->leave_spinning = 1;
7012
7013         is_data = owner_objectid >= BTRFS_FIRST_FREE_OBJECTID;
7014         BUG_ON(!is_data && refs_to_drop != 1);
7015
7016         if (is_data)
7017                 skinny_metadata = false;
7018
7019         ret = lookup_extent_backref(trans, path, &iref, bytenr, num_bytes,
7020                                     parent, root_objectid, owner_objectid,
7021                                     owner_offset);
7022         if (ret == 0) {
7023                 extent_slot = path->slots[0];
7024                 while (extent_slot >= 0) {
7025                         btrfs_item_key_to_cpu(path->nodes[0], &key,
7026                                               extent_slot);
7027                         if (key.objectid != bytenr)
7028                                 break;
7029                         if (key.type == BTRFS_EXTENT_ITEM_KEY &&
7030                             key.offset == num_bytes) {
7031                                 found_extent = 1;
7032                                 break;
7033                         }
7034                         if (key.type == BTRFS_METADATA_ITEM_KEY &&
7035                             key.offset == owner_objectid) {
7036                                 found_extent = 1;
7037                                 break;
7038                         }
7039                         if (path->slots[0] - extent_slot > 5)
7040                                 break;
7041                         extent_slot--;
7042                 }
7043
7044                 if (!found_extent) {
7045                         BUG_ON(iref);
7046                         ret = remove_extent_backref(trans, path, NULL,
7047                                                     refs_to_drop,
7048                                                     is_data, &last_ref);
7049                         if (ret) {
7050                                 btrfs_abort_transaction(trans, ret);
7051                                 goto out;
7052                         }
7053                         btrfs_release_path(path);
7054                         path->leave_spinning = 1;
7055
7056                         key.objectid = bytenr;
7057                         key.type = BTRFS_EXTENT_ITEM_KEY;
7058                         key.offset = num_bytes;
7059
7060                         if (!is_data && skinny_metadata) {
7061                                 key.type = BTRFS_METADATA_ITEM_KEY;
7062                                 key.offset = owner_objectid;
7063                         }
7064
7065                         ret = btrfs_search_slot(trans, extent_root,
7066                                                 &key, path, -1, 1);
7067                         if (ret > 0 && skinny_metadata && path->slots[0]) {
7068                                 /*
7069                                  * Couldn't find our skinny metadata item,
7070                                  * see if we have ye olde extent item.
7071                                  */
7072                                 path->slots[0]--;
7073                                 btrfs_item_key_to_cpu(path->nodes[0], &key,
7074                                                       path->slots[0]);
7075                                 if (key.objectid == bytenr &&
7076                                     key.type == BTRFS_EXTENT_ITEM_KEY &&
7077                                     key.offset == num_bytes)
7078                                         ret = 0;
7079                         }
7080
7081                         if (ret > 0 && skinny_metadata) {
7082                                 skinny_metadata = false;
7083                                 key.objectid = bytenr;
7084                                 key.type = BTRFS_EXTENT_ITEM_KEY;
7085                                 key.offset = num_bytes;
7086                                 btrfs_release_path(path);
7087                                 ret = btrfs_search_slot(trans, extent_root,
7088                                                         &key, path, -1, 1);
7089                         }
7090
7091                         if (ret) {
7092                                 btrfs_err(info,
7093                                           "umm, got %d back from search, was looking for %llu",
7094                                           ret, bytenr);
7095                                 if (ret > 0)
7096                                         btrfs_print_leaf(path->nodes[0]);
7097                         }
7098                         if (ret < 0) {
7099                                 btrfs_abort_transaction(trans, ret);
7100                                 goto out;
7101                         }
7102                         extent_slot = path->slots[0];
7103                 }
7104         } else if (WARN_ON(ret == -ENOENT)) {
7105                 btrfs_print_leaf(path->nodes[0]);
7106                 btrfs_err(info,
7107                         "unable to find ref byte nr %llu parent %llu root %llu  owner %llu offset %llu",
7108                         bytenr, parent, root_objectid, owner_objectid,
7109                         owner_offset);
7110                 btrfs_abort_transaction(trans, ret);
7111                 goto out;
7112         } else {
7113                 btrfs_abort_transaction(trans, ret);
7114                 goto out;
7115         }
7116
7117         leaf = path->nodes[0];
7118         item_size = btrfs_item_size_nr(leaf, extent_slot);
7119         if (unlikely(item_size < sizeof(*ei))) {
7120                 ret = -EINVAL;
7121                 btrfs_print_v0_err(info);
7122                 btrfs_abort_transaction(trans, ret);
7123                 goto out;
7124         }
7125         ei = btrfs_item_ptr(leaf, extent_slot,
7126                             struct btrfs_extent_item);
7127         if (owner_objectid < BTRFS_FIRST_FREE_OBJECTID &&
7128             key.type == BTRFS_EXTENT_ITEM_KEY) {
7129                 struct btrfs_tree_block_info *bi;
7130                 BUG_ON(item_size < sizeof(*ei) + sizeof(*bi));
7131                 bi = (struct btrfs_tree_block_info *)(ei + 1);
7132                 WARN_ON(owner_objectid != btrfs_tree_block_level(leaf, bi));
7133         }
7134
7135         refs = btrfs_extent_refs(leaf, ei);
7136         if (refs < refs_to_drop) {
7137                 btrfs_err(info,
7138                           "trying to drop %d refs but we only have %Lu for bytenr %Lu",
7139                           refs_to_drop, refs, bytenr);
7140                 ret = -EINVAL;
7141                 btrfs_abort_transaction(trans, ret);
7142                 goto out;
7143         }
7144         refs -= refs_to_drop;
7145
7146         if (refs > 0) {
7147                 if (extent_op)
7148                         __run_delayed_extent_op(extent_op, leaf, ei);
7149                 /*
7150                  * In the case of inline back ref, reference count will
7151                  * be updated by remove_extent_backref
7152                  */
7153                 if (iref) {
7154                         BUG_ON(!found_extent);
7155                 } else {
7156                         btrfs_set_extent_refs(leaf, ei, refs);
7157                         btrfs_mark_buffer_dirty(leaf);
7158                 }
7159                 if (found_extent) {
7160                         ret = remove_extent_backref(trans, path, iref,
7161                                                     refs_to_drop, is_data,
7162                                                     &last_ref);
7163                         if (ret) {
7164                                 btrfs_abort_transaction(trans, ret);
7165                                 goto out;
7166                         }
7167                 }
7168         } else {
7169                 if (found_extent) {
7170                         BUG_ON(is_data && refs_to_drop !=
7171                                extent_data_ref_count(path, iref));
7172                         if (iref) {
7173                                 BUG_ON(path->slots[0] != extent_slot);
7174                         } else {
7175                                 BUG_ON(path->slots[0] != extent_slot + 1);
7176                                 path->slots[0] = extent_slot;
7177                                 num_to_del = 2;
7178                         }
7179                 }
7180
7181                 last_ref = 1;
7182                 ret = btrfs_del_items(trans, extent_root, path, path->slots[0],
7183                                       num_to_del);
7184                 if (ret) {
7185                         btrfs_abort_transaction(trans, ret);
7186                         goto out;
7187                 }
7188                 btrfs_release_path(path);
7189
7190                 if (is_data) {
7191                         ret = btrfs_del_csums(trans, info, bytenr, num_bytes);
7192                         if (ret) {
7193                                 btrfs_abort_transaction(trans, ret);
7194                                 goto out;
7195                         }
7196                 }
7197
7198                 ret = add_to_free_space_tree(trans, bytenr, num_bytes);
7199                 if (ret) {
7200                         btrfs_abort_transaction(trans, ret);
7201                         goto out;
7202                 }
7203
7204                 ret = update_block_group(trans, info, bytenr, num_bytes, 0);
7205                 if (ret) {
7206                         btrfs_abort_transaction(trans, ret);
7207                         goto out;
7208                 }
7209         }
7210         btrfs_release_path(path);
7211
7212 out:
7213         btrfs_free_path(path);
7214         return ret;
7215 }
7216
7217 /*
7218  * when we free an block, it is possible (and likely) that we free the last
7219  * delayed ref for that extent as well.  This searches the delayed ref tree for
7220  * a given extent, and if there are no other delayed refs to be processed, it
7221  * removes it from the tree.
7222  */
7223 static noinline int check_ref_cleanup(struct btrfs_trans_handle *trans,
7224                                       u64 bytenr)
7225 {
7226         struct btrfs_delayed_ref_head *head;
7227         struct btrfs_delayed_ref_root *delayed_refs;
7228         int ret = 0;
7229
7230         delayed_refs = &trans->transaction->delayed_refs;
7231         spin_lock(&delayed_refs->lock);
7232         head = btrfs_find_delayed_ref_head(delayed_refs, bytenr);
7233         if (!head)
7234                 goto out_delayed_unlock;
7235
7236         spin_lock(&head->lock);
7237         if (!RB_EMPTY_ROOT(&head->ref_tree.rb_root))
7238                 goto out;
7239
7240         if (cleanup_extent_op(head) != NULL)
7241                 goto out;
7242
7243         /*
7244          * waiting for the lock here would deadlock.  If someone else has it
7245          * locked they are already in the process of dropping it anyway
7246          */
7247         if (!mutex_trylock(&head->mutex))
7248                 goto out;
7249
7250         btrfs_delete_ref_head(delayed_refs, head);
7251         head->processing = 0;
7252
7253         spin_unlock(&head->lock);
7254         spin_unlock(&delayed_refs->lock);
7255
7256         BUG_ON(head->extent_op);
7257         if (head->must_insert_reserved)
7258                 ret = 1;
7259
7260         btrfs_cleanup_ref_head_accounting(trans->fs_info, delayed_refs, head);
7261         mutex_unlock(&head->mutex);
7262         btrfs_put_delayed_ref_head(head);
7263         return ret;
7264 out:
7265         spin_unlock(&head->lock);
7266
7267 out_delayed_unlock:
7268         spin_unlock(&delayed_refs->lock);
7269         return 0;
7270 }
7271
7272 void btrfs_free_tree_block(struct btrfs_trans_handle *trans,
7273                            struct btrfs_root *root,
7274                            struct extent_buffer *buf,
7275                            u64 parent, int last_ref)
7276 {
7277         struct btrfs_fs_info *fs_info = root->fs_info;
7278         int pin = 1;
7279         int ret;
7280
7281         if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
7282                 int old_ref_mod, new_ref_mod;
7283
7284                 btrfs_ref_tree_mod(root, buf->start, buf->len, parent,
7285                                    root->root_key.objectid,
7286                                    btrfs_header_level(buf), 0,
7287                                    BTRFS_DROP_DELAYED_REF);
7288                 ret = btrfs_add_delayed_tree_ref(trans, buf->start,
7289                                                  buf->len, parent,
7290                                                  root->root_key.objectid,
7291                                                  btrfs_header_level(buf),
7292                                                  BTRFS_DROP_DELAYED_REF, NULL,
7293                                                  &old_ref_mod, &new_ref_mod);
7294                 BUG_ON(ret); /* -ENOMEM */
7295                 pin = old_ref_mod >= 0 && new_ref_mod < 0;
7296         }
7297
7298         if (last_ref && btrfs_header_generation(buf) == trans->transid) {
7299                 struct btrfs_block_group_cache *cache;
7300
7301                 if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
7302                         ret = check_ref_cleanup(trans, buf->start);
7303                         if (!ret)
7304                                 goto out;
7305                 }
7306
7307                 pin = 0;
7308                 cache = btrfs_lookup_block_group(fs_info, buf->start);
7309
7310                 if (btrfs_header_flag(buf, BTRFS_HEADER_FLAG_WRITTEN)) {
7311                         pin_down_extent(fs_info, cache, buf->start,
7312                                         buf->len, 1);
7313                         btrfs_put_block_group(cache);
7314                         goto out;
7315                 }
7316
7317                 WARN_ON(test_bit(EXTENT_BUFFER_DIRTY, &buf->bflags));
7318
7319                 btrfs_add_free_space(cache, buf->start, buf->len);
7320                 btrfs_free_reserved_bytes(cache, buf->len, 0);
7321                 btrfs_put_block_group(cache);
7322                 trace_btrfs_reserved_extent_free(fs_info, buf->start, buf->len);
7323         }
7324 out:
7325         if (pin)
7326                 add_pinned_bytes(fs_info, buf->len, true,
7327                                  root->root_key.objectid);
7328
7329         if (last_ref) {
7330                 /*
7331                  * Deleting the buffer, clear the corrupt flag since it doesn't
7332                  * matter anymore.
7333                  */
7334                 clear_bit(EXTENT_BUFFER_CORRUPT, &buf->bflags);
7335         }
7336 }
7337
7338 /* Can return -ENOMEM */
7339 int btrfs_free_extent(struct btrfs_trans_handle *trans,
7340                       struct btrfs_root *root,
7341                       u64 bytenr, u64 num_bytes, u64 parent, u64 root_objectid,
7342                       u64 owner, u64 offset)
7343 {
7344         struct btrfs_fs_info *fs_info = root->fs_info;
7345         int old_ref_mod, new_ref_mod;
7346         int ret;
7347
7348         if (btrfs_is_testing(fs_info))
7349                 return 0;
7350
7351         if (root_objectid != BTRFS_TREE_LOG_OBJECTID)
7352                 btrfs_ref_tree_mod(root, bytenr, num_bytes, parent,
7353                                    root_objectid, owner, offset,
7354                                    BTRFS_DROP_DELAYED_REF);
7355
7356         /*
7357          * tree log blocks never actually go into the extent allocation
7358          * tree, just update pinning info and exit early.
7359          */
7360         if (root_objectid == BTRFS_TREE_LOG_OBJECTID) {
7361                 WARN_ON(owner >= BTRFS_FIRST_FREE_OBJECTID);
7362                 /* unlocks the pinned mutex */
7363                 btrfs_pin_extent(fs_info, bytenr, num_bytes, 1);
7364                 old_ref_mod = new_ref_mod = 0;
7365                 ret = 0;
7366         } else if (owner < BTRFS_FIRST_FREE_OBJECTID) {
7367                 ret = btrfs_add_delayed_tree_ref(trans, bytenr,
7368                                                  num_bytes, parent,
7369                                                  root_objectid, (int)owner,
7370                                                  BTRFS_DROP_DELAYED_REF, NULL,
7371                                                  &old_ref_mod, &new_ref_mod);
7372         } else {
7373                 ret = btrfs_add_delayed_data_ref(trans, bytenr,
7374                                                  num_bytes, parent,
7375                                                  root_objectid, owner, offset,
7376                                                  0, BTRFS_DROP_DELAYED_REF,
7377                                                  &old_ref_mod, &new_ref_mod);
7378         }
7379
7380         if (ret == 0 && old_ref_mod >= 0 && new_ref_mod < 0) {
7381                 bool metadata = owner < BTRFS_FIRST_FREE_OBJECTID;
7382
7383                 add_pinned_bytes(fs_info, num_bytes, metadata, root_objectid);
7384         }
7385
7386         return ret;
7387 }
7388
7389 /*
7390  * when we wait for progress in the block group caching, its because
7391  * our allocation attempt failed at least once.  So, we must sleep
7392  * and let some progress happen before we try again.
7393  *
7394  * This function will sleep at least once waiting for new free space to
7395  * show up, and then it will check the block group free space numbers
7396  * for our min num_bytes.  Another option is to have it go ahead
7397  * and look in the rbtree for a free extent of a given size, but this
7398  * is a good start.
7399  *
7400  * Callers of this must check if cache->cached == BTRFS_CACHE_ERROR before using
7401  * any of the information in this block group.
7402  */
7403 static noinline void
7404 wait_block_group_cache_progress(struct btrfs_block_group_cache *cache,
7405                                 u64 num_bytes)
7406 {
7407         struct btrfs_caching_control *caching_ctl;
7408
7409         caching_ctl = get_caching_control(cache);
7410         if (!caching_ctl)
7411                 return;
7412
7413         wait_event(caching_ctl->wait, block_group_cache_done(cache) ||
7414                    (cache->free_space_ctl->free_space >= num_bytes));
7415
7416         put_caching_control(caching_ctl);
7417 }
7418
7419 static noinline int
7420 wait_block_group_cache_done(struct btrfs_block_group_cache *cache)
7421 {
7422         struct btrfs_caching_control *caching_ctl;
7423         int ret = 0;
7424
7425         caching_ctl = get_caching_control(cache);
7426         if (!caching_ctl)
7427                 return (cache->cached == BTRFS_CACHE_ERROR) ? -EIO : 0;
7428
7429         wait_event(caching_ctl->wait, block_group_cache_done(cache));
7430         if (cache->cached == BTRFS_CACHE_ERROR)
7431                 ret = -EIO;
7432         put_caching_control(caching_ctl);
7433         return ret;
7434 }
7435
7436 enum btrfs_loop_type {
7437         LOOP_CACHING_NOWAIT = 0,
7438         LOOP_CACHING_WAIT = 1,
7439         LOOP_ALLOC_CHUNK = 2,
7440         LOOP_NO_EMPTY_SIZE = 3,
7441 };
7442
7443 static inline void
7444 btrfs_lock_block_group(struct btrfs_block_group_cache *cache,
7445                        int delalloc)
7446 {
7447         if (delalloc)
7448                 down_read(&cache->data_rwsem);
7449 }
7450
7451 static inline void
7452 btrfs_grab_block_group(struct btrfs_block_group_cache *cache,
7453                        int delalloc)
7454 {
7455         btrfs_get_block_group(cache);
7456         if (delalloc)
7457                 down_read(&cache->data_rwsem);
7458 }
7459
7460 static struct btrfs_block_group_cache *
7461 btrfs_lock_cluster(struct btrfs_block_group_cache *block_group,
7462                    struct btrfs_free_cluster *cluster,
7463                    int delalloc)
7464 {
7465         struct btrfs_block_group_cache *used_bg = NULL;
7466
7467         spin_lock(&cluster->refill_lock);
7468         while (1) {
7469                 used_bg = cluster->block_group;
7470                 if (!used_bg)
7471                         return NULL;
7472
7473                 if (used_bg == block_group)
7474                         return used_bg;
7475
7476                 btrfs_get_block_group(used_bg);
7477
7478                 if (!delalloc)
7479                         return used_bg;
7480
7481                 if (down_read_trylock(&used_bg->data_rwsem))
7482                         return used_bg;
7483
7484                 spin_unlock(&cluster->refill_lock);
7485
7486                 /* We should only have one-level nested. */
7487                 down_read_nested(&used_bg->data_rwsem, SINGLE_DEPTH_NESTING);
7488
7489                 spin_lock(&cluster->refill_lock);
7490                 if (used_bg == cluster->block_group)
7491                         return used_bg;
7492
7493                 up_read(&used_bg->data_rwsem);
7494                 btrfs_put_block_group(used_bg);
7495         }
7496 }
7497
7498 static inline void
7499 btrfs_release_block_group(struct btrfs_block_group_cache *cache,
7500                          int delalloc)
7501 {
7502         if (delalloc)
7503                 up_read(&cache->data_rwsem);
7504         btrfs_put_block_group(cache);
7505 }
7506
7507 /*
7508  * Structure used internally for find_free_extent() function.  Wraps needed
7509  * parameters.
7510  */
7511 struct find_free_extent_ctl {
7512         /* Basic allocation info */
7513         u64 ram_bytes;
7514         u64 num_bytes;
7515         u64 empty_size;
7516         u64 flags;
7517         int delalloc;
7518
7519         /* Where to start the search inside the bg */
7520         u64 search_start;
7521
7522         /* For clustered allocation */
7523         u64 empty_cluster;
7524
7525         bool have_caching_bg;
7526         bool orig_have_caching_bg;
7527
7528         /* RAID index, converted from flags */
7529         int index;
7530
7531         /*
7532          * Current loop number, check find_free_extent_update_loop() for details
7533          */
7534         int loop;
7535
7536         /*
7537          * Whether we're refilling a cluster, if true we need to re-search
7538          * current block group but don't try to refill the cluster again.
7539          */
7540         bool retry_clustered;
7541
7542         /*
7543          * Whether we're updating free space cache, if true we need to re-search
7544          * current block group but don't try updating free space cache again.
7545          */
7546         bool retry_unclustered;
7547
7548         /* If current block group is cached */
7549         int cached;
7550
7551         /* Max contiguous hole found */
7552         u64 max_extent_size;
7553
7554         /* Total free space from free space cache, not always contiguous */
7555         u64 total_free_space;
7556
7557         /* Found result */
7558         u64 found_offset;
7559 };
7560
7561
7562 /*
7563  * Helper function for find_free_extent().
7564  *
7565  * Return -ENOENT to inform caller that we need fallback to unclustered mode.
7566  * Return -EAGAIN to inform caller that we need to re-search this block group
7567  * Return >0 to inform caller that we find nothing
7568  * Return 0 means we have found a location and set ffe_ctl->found_offset.
7569  */
7570 static int find_free_extent_clustered(struct btrfs_block_group_cache *bg,
7571                 struct btrfs_free_cluster *last_ptr,
7572                 struct find_free_extent_ctl *ffe_ctl,
7573                 struct btrfs_block_group_cache **cluster_bg_ret)
7574 {
7575         struct btrfs_fs_info *fs_info = bg->fs_info;
7576         struct btrfs_block_group_cache *cluster_bg;
7577         u64 aligned_cluster;
7578         u64 offset;
7579         int ret;
7580
7581         cluster_bg = btrfs_lock_cluster(bg, last_ptr, ffe_ctl->delalloc);
7582         if (!cluster_bg)
7583                 goto refill_cluster;
7584         if (cluster_bg != bg && (cluster_bg->ro ||
7585             !block_group_bits(cluster_bg, ffe_ctl->flags)))
7586                 goto release_cluster;
7587
7588         offset = btrfs_alloc_from_cluster(cluster_bg, last_ptr,
7589                         ffe_ctl->num_bytes, cluster_bg->key.objectid,
7590                         &ffe_ctl->max_extent_size);
7591         if (offset) {
7592                 /* We have a block, we're done */
7593                 spin_unlock(&last_ptr->refill_lock);
7594                 trace_btrfs_reserve_extent_cluster(cluster_bg,
7595                                 ffe_ctl->search_start, ffe_ctl->num_bytes);
7596                 *cluster_bg_ret = cluster_bg;
7597                 ffe_ctl->found_offset = offset;
7598                 return 0;
7599         }
7600         WARN_ON(last_ptr->block_group != cluster_bg);
7601
7602 release_cluster:
7603         /*
7604          * If we are on LOOP_NO_EMPTY_SIZE, we can't set up a new clusters, so
7605          * lets just skip it and let the allocator find whatever block it can
7606          * find. If we reach this point, we will have tried the cluster
7607          * allocator plenty of times and not have found anything, so we are
7608          * likely way too fragmented for the clustering stuff to find anything.
7609          *
7610          * However, if the cluster is taken from the current block group,
7611          * release the cluster first, so that we stand a better chance of
7612          * succeeding in the unclustered allocation.
7613          */
7614         if (ffe_ctl->loop >= LOOP_NO_EMPTY_SIZE && cluster_bg != bg) {
7615                 spin_unlock(&last_ptr->refill_lock);
7616                 btrfs_release_block_group(cluster_bg, ffe_ctl->delalloc);
7617                 return -ENOENT;
7618         }
7619
7620         /* This cluster didn't work out, free it and start over */
7621         btrfs_return_cluster_to_free_space(NULL, last_ptr);
7622
7623         if (cluster_bg != bg)
7624                 btrfs_release_block_group(cluster_bg, ffe_ctl->delalloc);
7625
7626 refill_cluster:
7627         if (ffe_ctl->loop >= LOOP_NO_EMPTY_SIZE) {
7628                 spin_unlock(&last_ptr->refill_lock);
7629                 return -ENOENT;
7630         }
7631
7632         aligned_cluster = max_t(u64,
7633                         ffe_ctl->empty_cluster + ffe_ctl->empty_size,
7634                         bg->full_stripe_len);
7635         ret = btrfs_find_space_cluster(fs_info, bg, last_ptr,
7636                         ffe_ctl->search_start, ffe_ctl->num_bytes,
7637                         aligned_cluster);
7638         if (ret == 0) {
7639                 /* Now pull our allocation out of this cluster */
7640                 offset = btrfs_alloc_from_cluster(bg, last_ptr,
7641                                 ffe_ctl->num_bytes, ffe_ctl->search_start,
7642                                 &ffe_ctl->max_extent_size);
7643                 if (offset) {
7644                         /* We found one, proceed */
7645                         spin_unlock(&last_ptr->refill_lock);
7646                         trace_btrfs_reserve_extent_cluster(bg,
7647                                         ffe_ctl->search_start,
7648                                         ffe_ctl->num_bytes);
7649                         ffe_ctl->found_offset = offset;
7650                         return 0;
7651                 }
7652         } else if (!ffe_ctl->cached && ffe_ctl->loop > LOOP_CACHING_NOWAIT &&
7653                    !ffe_ctl->retry_clustered) {
7654                 spin_unlock(&last_ptr->refill_lock);
7655
7656                 ffe_ctl->retry_clustered = true;
7657                 wait_block_group_cache_progress(bg, ffe_ctl->num_bytes +
7658                                 ffe_ctl->empty_cluster + ffe_ctl->empty_size);
7659                 return -EAGAIN;
7660         }
7661         /*
7662          * At this point we either didn't find a cluster or we weren't able to
7663          * allocate a block from our cluster.  Free the cluster we've been
7664          * trying to use, and go to the next block group.
7665          */
7666         btrfs_return_cluster_to_free_space(NULL, last_ptr);
7667         spin_unlock(&last_ptr->refill_lock);
7668         return 1;
7669 }
7670
7671 /*
7672  * Return >0 to inform caller that we find nothing
7673  * Return 0 when we found an free extent and set ffe_ctrl->found_offset
7674  * Return -EAGAIN to inform caller that we need to re-search this block group
7675  */
7676 static int find_free_extent_unclustered(struct btrfs_block_group_cache *bg,
7677                 struct btrfs_free_cluster *last_ptr,
7678                 struct find_free_extent_ctl *ffe_ctl)
7679 {
7680         u64 offset;
7681
7682         /*
7683          * We are doing an unclustered allocation, set the fragmented flag so
7684          * we don't bother trying to setup a cluster again until we get more
7685          * space.
7686          */
7687         if (unlikely(last_ptr)) {
7688                 spin_lock(&last_ptr->lock);
7689                 last_ptr->fragmented = 1;
7690                 spin_unlock(&last_ptr->lock);
7691         }
7692         if (ffe_ctl->cached) {
7693                 struct btrfs_free_space_ctl *free_space_ctl;
7694
7695                 free_space_ctl = bg->free_space_ctl;
7696                 spin_lock(&free_space_ctl->tree_lock);
7697                 if (free_space_ctl->free_space <
7698                     ffe_ctl->num_bytes + ffe_ctl->empty_cluster +
7699                     ffe_ctl->empty_size) {
7700                         ffe_ctl->total_free_space = max_t(u64,
7701                                         ffe_ctl->total_free_space,
7702                                         free_space_ctl->free_space);
7703                         spin_unlock(&free_space_ctl->tree_lock);
7704                         return 1;
7705                 }
7706                 spin_unlock(&free_space_ctl->tree_lock);
7707         }
7708
7709         offset = btrfs_find_space_for_alloc(bg, ffe_ctl->search_start,
7710                         ffe_ctl->num_bytes, ffe_ctl->empty_size,
7711                         &ffe_ctl->max_extent_size);
7712
7713         /*
7714          * If we didn't find a chunk, and we haven't failed on this block group
7715          * before, and this block group is in the middle of caching and we are
7716          * ok with waiting, then go ahead and wait for progress to be made, and
7717          * set @retry_unclustered to true.
7718          *
7719          * If @retry_unclustered is true then we've already waited on this
7720          * block group once and should move on to the next block group.
7721          */
7722         if (!offset && !ffe_ctl->retry_unclustered && !ffe_ctl->cached &&
7723             ffe_ctl->loop > LOOP_CACHING_NOWAIT) {
7724                 wait_block_group_cache_progress(bg, ffe_ctl->num_bytes +
7725                                                 ffe_ctl->empty_size);
7726                 ffe_ctl->retry_unclustered = true;
7727                 return -EAGAIN;
7728         } else if (!offset) {
7729                 return 1;
7730         }
7731         ffe_ctl->found_offset = offset;
7732         return 0;
7733 }
7734
7735 /*
7736  * Return >0 means caller needs to re-search for free extent
7737  * Return 0 means we have the needed free extent.
7738  * Return <0 means we failed to locate any free extent.
7739  */
7740 static int find_free_extent_update_loop(struct btrfs_fs_info *fs_info,
7741                                         struct btrfs_free_cluster *last_ptr,
7742                                         struct btrfs_key *ins,
7743                                         struct find_free_extent_ctl *ffe_ctl,
7744                                         int full_search, bool use_cluster)
7745 {
7746         struct btrfs_root *root = fs_info->extent_root;
7747         int ret;
7748
7749         if ((ffe_ctl->loop == LOOP_CACHING_NOWAIT) &&
7750             ffe_ctl->have_caching_bg && !ffe_ctl->orig_have_caching_bg)
7751                 ffe_ctl->orig_have_caching_bg = true;
7752
7753         if (!ins->objectid && ffe_ctl->loop >= LOOP_CACHING_WAIT &&
7754             ffe_ctl->have_caching_bg)
7755                 return 1;
7756
7757         if (!ins->objectid && ++(ffe_ctl->index) < BTRFS_NR_RAID_TYPES)
7758                 return 1;
7759
7760         if (ins->objectid) {
7761                 if (!use_cluster && last_ptr) {
7762                         spin_lock(&last_ptr->lock);
7763                         last_ptr->window_start = ins->objectid;
7764                         spin_unlock(&last_ptr->lock);
7765                 }
7766                 return 0;
7767         }
7768
7769         /*
7770          * LOOP_CACHING_NOWAIT, search partially cached block groups, kicking
7771          *                      caching kthreads as we move along
7772          * LOOP_CACHING_WAIT, search everything, and wait if our bg is caching
7773          * LOOP_ALLOC_CHUNK, force a chunk allocation and try again
7774          * LOOP_NO_EMPTY_SIZE, set empty_size and empty_cluster to 0 and try
7775          *                     again
7776          */
7777         if (ffe_ctl->loop < LOOP_NO_EMPTY_SIZE) {
7778                 ffe_ctl->index = 0;
7779                 if (ffe_ctl->loop == LOOP_CACHING_NOWAIT) {
7780                         /*
7781                          * We want to skip the LOOP_CACHING_WAIT step if we
7782                          * don't have any uncached bgs and we've already done a
7783                          * full search through.
7784                          */
7785                         if (ffe_ctl->orig_have_caching_bg || !full_search)
7786                                 ffe_ctl->loop = LOOP_CACHING_WAIT;
7787                         else
7788                                 ffe_ctl->loop = LOOP_ALLOC_CHUNK;
7789                 } else {
7790                         ffe_ctl->loop++;
7791                 }
7792
7793                 if (ffe_ctl->loop == LOOP_ALLOC_CHUNK) {
7794                         struct btrfs_trans_handle *trans;
7795                         int exist = 0;
7796
7797                         trans = current->journal_info;
7798                         if (trans)
7799                                 exist = 1;
7800                         else
7801                                 trans = btrfs_join_transaction(root);
7802
7803                         if (IS_ERR(trans)) {
7804                                 ret = PTR_ERR(trans);
7805                                 return ret;
7806                         }
7807
7808                         ret = do_chunk_alloc(trans, ffe_ctl->flags,
7809                                              CHUNK_ALLOC_FORCE);
7810
7811                         /*
7812                          * If we can't allocate a new chunk we've already looped
7813                          * through at least once, move on to the NO_EMPTY_SIZE
7814                          * case.
7815                          */
7816                         if (ret == -ENOSPC)
7817                                 ffe_ctl->loop = LOOP_NO_EMPTY_SIZE;
7818
7819                         /* Do not bail out on ENOSPC since we can do more. */
7820                         if (ret < 0 && ret != -ENOSPC)
7821                                 btrfs_abort_transaction(trans, ret);
7822                         else
7823                                 ret = 0;
7824                         if (!exist)
7825                                 btrfs_end_transaction(trans);
7826                         if (ret)
7827                                 return ret;
7828                 }
7829
7830                 if (ffe_ctl->loop == LOOP_NO_EMPTY_SIZE) {
7831                         /*
7832                          * Don't loop again if we already have no empty_size and
7833                          * no empty_cluster.
7834                          */
7835                         if (ffe_ctl->empty_size == 0 &&
7836                             ffe_ctl->empty_cluster == 0)
7837                                 return -ENOSPC;
7838                         ffe_ctl->empty_size = 0;
7839                         ffe_ctl->empty_cluster = 0;
7840                 }
7841                 return 1;
7842         }
7843         return -ENOSPC;
7844 }
7845
7846 /*
7847  * walks the btree of allocated extents and find a hole of a given size.
7848  * The key ins is changed to record the hole:
7849  * ins->objectid == start position
7850  * ins->flags = BTRFS_EXTENT_ITEM_KEY
7851  * ins->offset == the size of the hole.
7852  * Any available blocks before search_start are skipped.
7853  *
7854  * If there is no suitable free space, we will record the max size of
7855  * the free space extent currently.
7856  *
7857  * The overall logic and call chain:
7858  *
7859  * find_free_extent()
7860  * |- Iterate through all block groups
7861  * |  |- Get a valid block group
7862  * |  |- Try to do clustered allocation in that block group
7863  * |  |- Try to do unclustered allocation in that block group
7864  * |  |- Check if the result is valid
7865  * |  |  |- If valid, then exit
7866  * |  |- Jump to next block group
7867  * |
7868  * |- Push harder to find free extents
7869  *    |- If not found, re-iterate all block groups
7870  */
7871 static noinline int find_free_extent(struct btrfs_fs_info *fs_info,
7872                                 u64 ram_bytes, u64 num_bytes, u64 empty_size,
7873                                 u64 hint_byte, struct btrfs_key *ins,
7874                                 u64 flags, int delalloc)
7875 {
7876         int ret = 0;
7877         struct btrfs_free_cluster *last_ptr = NULL;
7878         struct btrfs_block_group_cache *block_group = NULL;
7879         struct find_free_extent_ctl ffe_ctl = {0};
7880         struct btrfs_space_info *space_info;
7881         bool use_cluster = true;
7882         bool full_search = false;
7883
7884         WARN_ON(num_bytes < fs_info->sectorsize);
7885
7886         ffe_ctl.ram_bytes = ram_bytes;
7887         ffe_ctl.num_bytes = num_bytes;
7888         ffe_ctl.empty_size = empty_size;
7889         ffe_ctl.flags = flags;
7890         ffe_ctl.search_start = 0;
7891         ffe_ctl.retry_clustered = false;
7892         ffe_ctl.retry_unclustered = false;
7893         ffe_ctl.delalloc = delalloc;
7894         ffe_ctl.index = btrfs_bg_flags_to_raid_index(flags);
7895         ffe_ctl.have_caching_bg = false;
7896         ffe_ctl.orig_have_caching_bg = false;
7897         ffe_ctl.found_offset = 0;
7898
7899         ins->type = BTRFS_EXTENT_ITEM_KEY;
7900         ins->objectid = 0;
7901         ins->offset = 0;
7902
7903         trace_find_free_extent(fs_info, num_bytes, empty_size, flags);
7904
7905         space_info = __find_space_info(fs_info, flags);
7906         if (!space_info) {
7907                 btrfs_err(fs_info, "No space info for %llu", flags);
7908                 return -ENOSPC;
7909         }
7910
7911         /*
7912          * If our free space is heavily fragmented we may not be able to make
7913          * big contiguous allocations, so instead of doing the expensive search
7914          * for free space, simply return ENOSPC with our max_extent_size so we
7915          * can go ahead and search for a more manageable chunk.
7916          *
7917          * If our max_extent_size is large enough for our allocation simply
7918          * disable clustering since we will likely not be able to find enough
7919          * space to create a cluster and induce latency trying.
7920          */
7921         if (unlikely(space_info->max_extent_size)) {
7922                 spin_lock(&space_info->lock);
7923                 if (space_info->max_extent_size &&
7924                     num_bytes > space_info->max_extent_size) {
7925                         ins->offset = space_info->max_extent_size;
7926                         spin_unlock(&space_info->lock);
7927                         return -ENOSPC;
7928                 } else if (space_info->max_extent_size) {
7929                         use_cluster = false;
7930                 }
7931                 spin_unlock(&space_info->lock);
7932         }
7933
7934         last_ptr = fetch_cluster_info(fs_info, space_info,
7935                                       &ffe_ctl.empty_cluster);
7936         if (last_ptr) {
7937                 spin_lock(&last_ptr->lock);
7938                 if (last_ptr->block_group)
7939                         hint_byte = last_ptr->window_start;
7940                 if (last_ptr->fragmented) {
7941                         /*
7942                          * We still set window_start so we can keep track of the
7943                          * last place we found an allocation to try and save
7944                          * some time.
7945                          */
7946                         hint_byte = last_ptr->window_start;
7947                         use_cluster = false;
7948                 }
7949                 spin_unlock(&last_ptr->lock);
7950         }
7951
7952         ffe_ctl.search_start = max(ffe_ctl.search_start,
7953                                    first_logical_byte(fs_info, 0));
7954         ffe_ctl.search_start = max(ffe_ctl.search_start, hint_byte);
7955         if (ffe_ctl.search_start == hint_byte) {
7956                 block_group = btrfs_lookup_block_group(fs_info,
7957                                                        ffe_ctl.search_start);
7958                 /*
7959                  * we don't want to use the block group if it doesn't match our
7960                  * allocation bits, or if its not cached.
7961                  *
7962                  * However if we are re-searching with an ideal block group
7963                  * picked out then we don't care that the block group is cached.
7964                  */
7965                 if (block_group && block_group_bits(block_group, flags) &&
7966                     block_group->cached != BTRFS_CACHE_NO) {
7967                         down_read(&space_info->groups_sem);
7968                         if (list_empty(&block_group->list) ||
7969                             block_group->ro) {
7970                                 /*
7971                                  * someone is removing this block group,
7972                                  * we can't jump into the have_block_group
7973                                  * target because our list pointers are not
7974                                  * valid
7975                                  */
7976                                 btrfs_put_block_group(block_group);
7977                                 up_read(&space_info->groups_sem);
7978                         } else {
7979                                 ffe_ctl.index = btrfs_bg_flags_to_raid_index(
7980                                                 block_group->flags);
7981                                 btrfs_lock_block_group(block_group, delalloc);
7982                                 goto have_block_group;
7983                         }
7984                 } else if (block_group) {
7985                         btrfs_put_block_group(block_group);
7986                 }
7987         }
7988 search:
7989         ffe_ctl.have_caching_bg = false;
7990         if (ffe_ctl.index == btrfs_bg_flags_to_raid_index(flags) ||
7991             ffe_ctl.index == 0)
7992                 full_search = true;
7993         down_read(&space_info->groups_sem);
7994         list_for_each_entry(block_group,
7995                             &space_info->block_groups[ffe_ctl.index], list) {
7996                 /* If the block group is read-only, we can skip it entirely. */
7997                 if (unlikely(block_group->ro))
7998                         continue;
7999
8000                 btrfs_grab_block_group(block_group, delalloc);
8001                 ffe_ctl.search_start = block_group->key.objectid;
8002
8003                 /*
8004                  * this can happen if we end up cycling through all the
8005                  * raid types, but we want to make sure we only allocate
8006                  * for the proper type.
8007                  */
8008                 if (!block_group_bits(block_group, flags)) {
8009                         u64 extra = BTRFS_BLOCK_GROUP_DUP |
8010                                 BTRFS_BLOCK_GROUP_RAID1 |
8011                                 BTRFS_BLOCK_GROUP_RAID5 |
8012                                 BTRFS_BLOCK_GROUP_RAID6 |
8013                                 BTRFS_BLOCK_GROUP_RAID10;
8014
8015                         /*
8016                          * if they asked for extra copies and this block group
8017                          * doesn't provide them, bail.  This does allow us to
8018                          * fill raid0 from raid1.
8019                          */
8020                         if ((flags & extra) && !(block_group->flags & extra))
8021                                 goto loop;
8022                 }
8023
8024 have_block_group:
8025                 ffe_ctl.cached = block_group_cache_done(block_group);
8026                 if (unlikely(!ffe_ctl.cached)) {
8027                         ffe_ctl.have_caching_bg = true;
8028                         ret = cache_block_group(block_group, 0);
8029                         BUG_ON(ret < 0);
8030                         ret = 0;
8031                 }
8032
8033                 if (unlikely(block_group->cached == BTRFS_CACHE_ERROR))
8034                         goto loop;
8035
8036                 /*
8037                  * Ok we want to try and use the cluster allocator, so
8038                  * lets look there
8039                  */
8040                 if (last_ptr && use_cluster) {
8041                         struct btrfs_block_group_cache *cluster_bg = NULL;
8042
8043                         ret = find_free_extent_clustered(block_group, last_ptr,
8044                                                          &ffe_ctl, &cluster_bg);
8045
8046                         if (ret == 0) {
8047                                 if (cluster_bg && cluster_bg != block_group) {
8048                                         btrfs_release_block_group(block_group,
8049                                                                   delalloc);
8050                                         block_group = cluster_bg;
8051                                 }
8052                                 goto checks;
8053                         } else if (ret == -EAGAIN) {
8054                                 goto have_block_group;
8055                         } else if (ret > 0) {
8056                                 goto loop;
8057                         }
8058                         /* ret == -ENOENT case falls through */
8059                 }
8060
8061                 ret = find_free_extent_unclustered(block_group, last_ptr,
8062                                                    &ffe_ctl);
8063                 if (ret == -EAGAIN)
8064                         goto have_block_group;
8065                 else if (ret > 0)
8066                         goto loop;
8067                 /* ret == 0 case falls through */
8068 checks:
8069                 ffe_ctl.search_start = round_up(ffe_ctl.found_offset,
8070                                              fs_info->stripesize);
8071
8072                 /* move on to the next group */
8073                 if (ffe_ctl.search_start + num_bytes >
8074                     block_group->key.objectid + block_group->key.offset) {
8075                         btrfs_add_free_space(block_group, ffe_ctl.found_offset,
8076                                              num_bytes);
8077                         goto loop;
8078                 }
8079
8080                 if (ffe_ctl.found_offset < ffe_ctl.search_start)
8081                         btrfs_add_free_space(block_group, ffe_ctl.found_offset,
8082                                 ffe_ctl.search_start - ffe_ctl.found_offset);
8083
8084                 ret = btrfs_add_reserved_bytes(block_group, ram_bytes,
8085                                 num_bytes, delalloc);
8086                 if (ret == -EAGAIN) {
8087                         btrfs_add_free_space(block_group, ffe_ctl.found_offset,
8088                                              num_bytes);
8089                         goto loop;
8090                 }
8091                 btrfs_inc_block_group_reservations(block_group);
8092
8093                 /* we are all good, lets return */
8094                 ins->objectid = ffe_ctl.search_start;
8095                 ins->offset = num_bytes;
8096
8097                 trace_btrfs_reserve_extent(block_group, ffe_ctl.search_start,
8098                                            num_bytes);
8099                 btrfs_release_block_group(block_group, delalloc);
8100                 break;
8101 loop:
8102                 ffe_ctl.retry_clustered = false;
8103                 ffe_ctl.retry_unclustered = false;
8104                 BUG_ON(btrfs_bg_flags_to_raid_index(block_group->flags) !=
8105                        ffe_ctl.index);
8106                 btrfs_release_block_group(block_group, delalloc);
8107                 cond_resched();
8108         }
8109         up_read(&space_info->groups_sem);
8110
8111         ret = find_free_extent_update_loop(fs_info, last_ptr, ins, &ffe_ctl,
8112                                            full_search, use_cluster);
8113         if (ret > 0)
8114                 goto search;
8115
8116         if (ret == -ENOSPC) {
8117                 /*
8118                  * Use ffe_ctl->total_free_space as fallback if we can't find
8119                  * any contiguous hole.
8120                  */
8121                 if (!ffe_ctl.max_extent_size)
8122                         ffe_ctl.max_extent_size = ffe_ctl.total_free_space;
8123                 spin_lock(&space_info->lock);
8124                 space_info->max_extent_size = ffe_ctl.max_extent_size;
8125                 spin_unlock(&space_info->lock);
8126                 ins->offset = ffe_ctl.max_extent_size;
8127         }
8128         return ret;
8129 }
8130
8131 #define DUMP_BLOCK_RSV(fs_info, rsv_name)                               \
8132 do {                                                                    \
8133         struct btrfs_block_rsv *__rsv = &(fs_info)->rsv_name;           \
8134         spin_lock(&__rsv->lock);                                        \
8135         btrfs_info(fs_info, #rsv_name ": size %llu reserved %llu",      \
8136                    __rsv->size, __rsv->reserved);                       \
8137         spin_unlock(&__rsv->lock);                                      \
8138 } while (0)
8139
8140 static void dump_space_info(struct btrfs_fs_info *fs_info,
8141                             struct btrfs_space_info *info, u64 bytes,
8142                             int dump_block_groups)
8143 {
8144         struct btrfs_block_group_cache *cache;
8145         int index = 0;
8146
8147         spin_lock(&info->lock);
8148         btrfs_info(fs_info, "space_info %llu has %llu free, is %sfull",
8149                    info->flags,
8150                    info->total_bytes - btrfs_space_info_used(info, true),
8151                    info->full ? "" : "not ");
8152         btrfs_info(fs_info,
8153                 "space_info total=%llu, used=%llu, pinned=%llu, reserved=%llu, may_use=%llu, readonly=%llu",
8154                 info->total_bytes, info->bytes_used, info->bytes_pinned,
8155                 info->bytes_reserved, info->bytes_may_use,
8156                 info->bytes_readonly);
8157         spin_unlock(&info->lock);
8158
8159         DUMP_BLOCK_RSV(fs_info, global_block_rsv);
8160         DUMP_BLOCK_RSV(fs_info, trans_block_rsv);
8161         DUMP_BLOCK_RSV(fs_info, chunk_block_rsv);
8162         DUMP_BLOCK_RSV(fs_info, delayed_block_rsv);
8163         DUMP_BLOCK_RSV(fs_info, delayed_refs_rsv);
8164
8165         if (!dump_block_groups)
8166                 return;
8167
8168         down_read(&info->groups_sem);
8169 again:
8170         list_for_each_entry(cache, &info->block_groups[index], list) {
8171                 spin_lock(&cache->lock);
8172                 btrfs_info(fs_info,
8173                         "block group %llu has %llu bytes, %llu used %llu pinned %llu reserved %s",
8174                         cache->key.objectid, cache->key.offset,
8175                         btrfs_block_group_used(&cache->item), cache->pinned,
8176                         cache->reserved, cache->ro ? "[readonly]" : "");
8177                 btrfs_dump_free_space(cache, bytes);
8178                 spin_unlock(&cache->lock);
8179         }
8180         if (++index < BTRFS_NR_RAID_TYPES)
8181                 goto again;
8182         up_read(&info->groups_sem);
8183 }
8184
8185 /*
8186  * btrfs_reserve_extent - entry point to the extent allocator. Tries to find a
8187  *                        hole that is at least as big as @num_bytes.
8188  *
8189  * @root           -    The root that will contain this extent
8190  *
8191  * @ram_bytes      -    The amount of space in ram that @num_bytes take. This
8192  *                      is used for accounting purposes. This value differs
8193  *                      from @num_bytes only in the case of compressed extents.
8194  *
8195  * @num_bytes      -    Number of bytes to allocate on-disk.
8196  *
8197  * @min_alloc_size -    Indicates the minimum amount of space that the
8198  *                      allocator should try to satisfy. In some cases
8199  *                      @num_bytes may be larger than what is required and if
8200  *                      the filesystem is fragmented then allocation fails.
8201  *                      However, the presence of @min_alloc_size gives a
8202  *                      chance to try and satisfy the smaller allocation.
8203  *
8204  * @empty_size     -    A hint that you plan on doing more COW. This is the
8205  *                      size in bytes the allocator should try to find free
8206  *                      next to the block it returns.  This is just a hint and
8207  *                      may be ignored by the allocator.
8208  *
8209  * @hint_byte      -    Hint to the allocator to start searching above the byte
8210  *                      address passed. It might be ignored.
8211  *
8212  * @ins            -    This key is modified to record the found hole. It will
8213  *                      have the following values:
8214  *                      ins->objectid == start position
8215  *                      ins->flags = BTRFS_EXTENT_ITEM_KEY
8216  *                      ins->offset == the size of the hole.
8217  *
8218  * @is_data        -    Boolean flag indicating whether an extent is
8219  *                      allocated for data (true) or metadata (false)
8220  *
8221  * @delalloc       -    Boolean flag indicating whether this allocation is for
8222  *                      delalloc or not. If 'true' data_rwsem of block groups
8223  *                      is going to be acquired.
8224  *
8225  *
8226  * Returns 0 when an allocation succeeded or < 0 when an error occurred. In
8227  * case -ENOSPC is returned then @ins->offset will contain the size of the
8228  * largest available hole the allocator managed to find.
8229  */
8230 int btrfs_reserve_extent(struct btrfs_root *root, u64 ram_bytes,
8231                          u64 num_bytes, u64 min_alloc_size,
8232                          u64 empty_size, u64 hint_byte,
8233                          struct btrfs_key *ins, int is_data, int delalloc)
8234 {
8235         struct btrfs_fs_info *fs_info = root->fs_info;
8236         bool final_tried = num_bytes == min_alloc_size;
8237         u64 flags;
8238         int ret;
8239
8240         flags = get_alloc_profile_by_root(root, is_data);
8241 again:
8242         WARN_ON(num_bytes < fs_info->sectorsize);
8243         ret = find_free_extent(fs_info, ram_bytes, num_bytes, empty_size,
8244                                hint_byte, ins, flags, delalloc);
8245         if (!ret && !is_data) {
8246                 btrfs_dec_block_group_reservations(fs_info, ins->objectid);
8247         } else if (ret == -ENOSPC) {
8248                 if (!final_tried && ins->offset) {
8249                         num_bytes = min(num_bytes >> 1, ins->offset);
8250                         num_bytes = round_down(num_bytes,
8251                                                fs_info->sectorsize);
8252                         num_bytes = max(num_bytes, min_alloc_size);
8253                         ram_bytes = num_bytes;
8254                         if (num_bytes == min_alloc_size)
8255                                 final_tried = true;
8256                         goto again;
8257                 } else if (btrfs_test_opt(fs_info, ENOSPC_DEBUG)) {
8258                         struct btrfs_space_info *sinfo;
8259
8260                         sinfo = __find_space_info(fs_info, flags);
8261                         btrfs_err(fs_info,
8262                                   "allocation failed flags %llu, wanted %llu",
8263                                   flags, num_bytes);
8264                         if (sinfo)
8265                                 dump_space_info(fs_info, sinfo, num_bytes, 1);
8266                 }
8267         }
8268
8269         return ret;
8270 }
8271
8272 static int __btrfs_free_reserved_extent(struct btrfs_fs_info *fs_info,
8273                                         u64 start, u64 len,
8274                                         int pin, int delalloc)
8275 {
8276         struct btrfs_block_group_cache *cache;
8277         int ret = 0;
8278
8279         cache = btrfs_lookup_block_group(fs_info, start);
8280         if (!cache) {
8281                 btrfs_err(fs_info, "Unable to find block group for %llu",
8282                           start);
8283                 return -ENOSPC;
8284         }
8285
8286         if (pin)
8287                 pin_down_extent(fs_info, cache, start, len, 1);
8288         else {
8289                 if (btrfs_test_opt(fs_info, DISCARD))
8290                         ret = btrfs_discard_extent(fs_info, start, len, NULL);
8291                 btrfs_add_free_space(cache, start, len);
8292                 btrfs_free_reserved_bytes(cache, len, delalloc);
8293                 trace_btrfs_reserved_extent_free(fs_info, start, len);
8294         }
8295
8296         btrfs_put_block_group(cache);
8297         return ret;
8298 }
8299
8300 int btrfs_free_reserved_extent(struct btrfs_fs_info *fs_info,
8301                                u64 start, u64 len, int delalloc)
8302 {
8303         return __btrfs_free_reserved_extent(fs_info, start, len, 0, delalloc);
8304 }
8305
8306 int btrfs_free_and_pin_reserved_extent(struct btrfs_fs_info *fs_info,
8307                                        u64 start, u64 len)
8308 {
8309         return __btrfs_free_reserved_extent(fs_info, start, len, 1, 0);
8310 }
8311
8312 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
8313                                       u64 parent, u64 root_objectid,
8314                                       u64 flags, u64 owner, u64 offset,
8315                                       struct btrfs_key *ins, int ref_mod)
8316 {
8317         struct btrfs_fs_info *fs_info = trans->fs_info;
8318         int ret;
8319         struct btrfs_extent_item *extent_item;
8320         struct btrfs_extent_inline_ref *iref;
8321         struct btrfs_path *path;
8322         struct extent_buffer *leaf;
8323         int type;
8324         u32 size;
8325
8326         if (parent > 0)
8327                 type = BTRFS_SHARED_DATA_REF_KEY;
8328         else
8329                 type = BTRFS_EXTENT_DATA_REF_KEY;
8330
8331         size = sizeof(*extent_item) + btrfs_extent_inline_ref_size(type);
8332
8333         path = btrfs_alloc_path();
8334         if (!path)
8335                 return -ENOMEM;
8336
8337         path->leave_spinning = 1;
8338         ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path,
8339                                       ins, size);
8340         if (ret) {
8341                 btrfs_free_path(path);
8342                 return ret;
8343         }
8344
8345         leaf = path->nodes[0];
8346         extent_item = btrfs_item_ptr(leaf, path->slots[0],
8347                                      struct btrfs_extent_item);
8348         btrfs_set_extent_refs(leaf, extent_item, ref_mod);
8349         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
8350         btrfs_set_extent_flags(leaf, extent_item,
8351                                flags | BTRFS_EXTENT_FLAG_DATA);
8352
8353         iref = (struct btrfs_extent_inline_ref *)(extent_item + 1);
8354         btrfs_set_extent_inline_ref_type(leaf, iref, type);
8355         if (parent > 0) {
8356                 struct btrfs_shared_data_ref *ref;
8357                 ref = (struct btrfs_shared_data_ref *)(iref + 1);
8358                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
8359                 btrfs_set_shared_data_ref_count(leaf, ref, ref_mod);
8360         } else {
8361                 struct btrfs_extent_data_ref *ref;
8362                 ref = (struct btrfs_extent_data_ref *)(&iref->offset);
8363                 btrfs_set_extent_data_ref_root(leaf, ref, root_objectid);
8364                 btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
8365                 btrfs_set_extent_data_ref_offset(leaf, ref, offset);
8366                 btrfs_set_extent_data_ref_count(leaf, ref, ref_mod);
8367         }
8368
8369         btrfs_mark_buffer_dirty(path->nodes[0]);
8370         btrfs_free_path(path);
8371
8372         ret = remove_from_free_space_tree(trans, ins->objectid, ins->offset);
8373         if (ret)
8374                 return ret;
8375
8376         ret = update_block_group(trans, fs_info, ins->objectid, ins->offset, 1);
8377         if (ret) { /* -ENOENT, logic error */
8378                 btrfs_err(fs_info, "update block group failed for %llu %llu",
8379                         ins->objectid, ins->offset);
8380                 BUG();
8381         }
8382         trace_btrfs_reserved_extent_alloc(fs_info, ins->objectid, ins->offset);
8383         return ret;
8384 }
8385
8386 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
8387                                      struct btrfs_delayed_ref_node *node,
8388                                      struct btrfs_delayed_extent_op *extent_op)
8389 {
8390         struct btrfs_fs_info *fs_info = trans->fs_info;
8391         int ret;
8392         struct btrfs_extent_item *extent_item;
8393         struct btrfs_key extent_key;
8394         struct btrfs_tree_block_info *block_info;
8395         struct btrfs_extent_inline_ref *iref;
8396         struct btrfs_path *path;
8397         struct extent_buffer *leaf;
8398         struct btrfs_delayed_tree_ref *ref;
8399         u32 size = sizeof(*extent_item) + sizeof(*iref);
8400         u64 num_bytes;
8401         u64 flags = extent_op->flags_to_set;
8402         bool skinny_metadata = btrfs_fs_incompat(fs_info, SKINNY_METADATA);
8403
8404         ref = btrfs_delayed_node_to_tree_ref(node);
8405
8406         extent_key.objectid = node->bytenr;
8407         if (skinny_metadata) {
8408                 extent_key.offset = ref->level;
8409                 extent_key.type = BTRFS_METADATA_ITEM_KEY;
8410                 num_bytes = fs_info->nodesize;
8411         } else {
8412                 extent_key.offset = node->num_bytes;
8413                 extent_key.type = BTRFS_EXTENT_ITEM_KEY;
8414                 size += sizeof(*block_info);
8415                 num_bytes = node->num_bytes;
8416         }
8417
8418         path = btrfs_alloc_path();
8419         if (!path)
8420                 return -ENOMEM;
8421
8422         path->leave_spinning = 1;
8423         ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path,
8424                                       &extent_key, size);
8425         if (ret) {
8426                 btrfs_free_path(path);
8427                 return ret;
8428         }
8429
8430         leaf = path->nodes[0];
8431         extent_item = btrfs_item_ptr(leaf, path->slots[0],
8432                                      struct btrfs_extent_item);
8433         btrfs_set_extent_refs(leaf, extent_item, 1);
8434         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
8435         btrfs_set_extent_flags(leaf, extent_item,
8436                                flags | BTRFS_EXTENT_FLAG_TREE_BLOCK);
8437
8438         if (skinny_metadata) {
8439                 iref = (struct btrfs_extent_inline_ref *)(extent_item + 1);
8440         } else {
8441                 block_info = (struct btrfs_tree_block_info *)(extent_item + 1);
8442                 btrfs_set_tree_block_key(leaf, block_info, &extent_op->key);
8443                 btrfs_set_tree_block_level(leaf, block_info, ref->level);
8444                 iref = (struct btrfs_extent_inline_ref *)(block_info + 1);
8445         }
8446
8447         if (node->type == BTRFS_SHARED_BLOCK_REF_KEY) {
8448                 BUG_ON(!(flags & BTRFS_BLOCK_FLAG_FULL_BACKREF));
8449                 btrfs_set_extent_inline_ref_type(leaf, iref,
8450                                                  BTRFS_SHARED_BLOCK_REF_KEY);
8451                 btrfs_set_extent_inline_ref_offset(leaf, iref, ref->parent);
8452         } else {
8453                 btrfs_set_extent_inline_ref_type(leaf, iref,
8454                                                  BTRFS_TREE_BLOCK_REF_KEY);
8455                 btrfs_set_extent_inline_ref_offset(leaf, iref, ref->root);
8456         }
8457
8458         btrfs_mark_buffer_dirty(leaf);
8459         btrfs_free_path(path);
8460
8461         ret = remove_from_free_space_tree(trans, extent_key.objectid,
8462                                           num_bytes);
8463         if (ret)
8464                 return ret;
8465
8466         ret = update_block_group(trans, fs_info, extent_key.objectid,
8467                                  fs_info->nodesize, 1);
8468         if (ret) { /* -ENOENT, logic error */
8469                 btrfs_err(fs_info, "update block group failed for %llu %llu",
8470                         extent_key.objectid, extent_key.offset);
8471                 BUG();
8472         }
8473
8474         trace_btrfs_reserved_extent_alloc(fs_info, extent_key.objectid,
8475                                           fs_info->nodesize);
8476         return ret;
8477 }
8478
8479 int btrfs_alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
8480                                      struct btrfs_root *root, u64 owner,
8481                                      u64 offset, u64 ram_bytes,
8482                                      struct btrfs_key *ins)
8483 {
8484         int ret;
8485
8486         BUG_ON(root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID);
8487
8488         btrfs_ref_tree_mod(root, ins->objectid, ins->offset, 0,
8489                            root->root_key.objectid, owner, offset,
8490                            BTRFS_ADD_DELAYED_EXTENT);
8491
8492         ret = btrfs_add_delayed_data_ref(trans, ins->objectid,
8493                                          ins->offset, 0,
8494                                          root->root_key.objectid, owner,
8495                                          offset, ram_bytes,
8496                                          BTRFS_ADD_DELAYED_EXTENT, NULL, NULL);
8497         return ret;
8498 }
8499
8500 /*
8501  * this is used by the tree logging recovery code.  It records that
8502  * an extent has been allocated and makes sure to clear the free
8503  * space cache bits as well
8504  */
8505 int btrfs_alloc_logged_file_extent(struct btrfs_trans_handle *trans,
8506                                    u64 root_objectid, u64 owner, u64 offset,
8507                                    struct btrfs_key *ins)
8508 {
8509         struct btrfs_fs_info *fs_info = trans->fs_info;
8510         int ret;
8511         struct btrfs_block_group_cache *block_group;
8512         struct btrfs_space_info *space_info;
8513
8514         /*
8515          * Mixed block groups will exclude before processing the log so we only
8516          * need to do the exclude dance if this fs isn't mixed.
8517          */
8518         if (!btrfs_fs_incompat(fs_info, MIXED_GROUPS)) {
8519                 ret = __exclude_logged_extent(fs_info, ins->objectid,
8520                                               ins->offset);
8521                 if (ret)
8522                         return ret;
8523         }
8524
8525         block_group = btrfs_lookup_block_group(fs_info, ins->objectid);
8526         if (!block_group)
8527                 return -EINVAL;
8528
8529         space_info = block_group->space_info;
8530         spin_lock(&space_info->lock);
8531         spin_lock(&block_group->lock);
8532         space_info->bytes_reserved += ins->offset;
8533         block_group->reserved += ins->offset;
8534         spin_unlock(&block_group->lock);
8535         spin_unlock(&space_info->lock);
8536
8537         ret = alloc_reserved_file_extent(trans, 0, root_objectid, 0, owner,
8538                                          offset, ins, 1);
8539         btrfs_put_block_group(block_group);
8540         return ret;
8541 }
8542
8543 static struct extent_buffer *
8544 btrfs_init_new_buffer(struct btrfs_trans_handle *trans, struct btrfs_root *root,
8545                       u64 bytenr, int level, u64 owner)
8546 {
8547         struct btrfs_fs_info *fs_info = root->fs_info;
8548         struct extent_buffer *buf;
8549
8550         buf = btrfs_find_create_tree_block(fs_info, bytenr);
8551         if (IS_ERR(buf))
8552                 return buf;
8553
8554         /*
8555          * Extra safety check in case the extent tree is corrupted and extent
8556          * allocator chooses to use a tree block which is already used and
8557          * locked.
8558          */
8559         if (buf->lock_owner == current->pid) {
8560                 btrfs_err_rl(fs_info,
8561 "tree block %llu owner %llu already locked by pid=%d, extent tree corruption detected",
8562                         buf->start, btrfs_header_owner(buf), current->pid);
8563                 free_extent_buffer(buf);
8564                 return ERR_PTR(-EUCLEAN);
8565         }
8566
8567         btrfs_set_buffer_lockdep_class(root->root_key.objectid, buf, level);
8568         btrfs_tree_lock(buf);
8569         clean_tree_block(fs_info, buf);
8570         clear_bit(EXTENT_BUFFER_STALE, &buf->bflags);
8571
8572         btrfs_set_lock_blocking_write(buf);
8573         set_extent_buffer_uptodate(buf);
8574
8575         memzero_extent_buffer(buf, 0, sizeof(struct btrfs_header));
8576         btrfs_set_header_level(buf, level);
8577         btrfs_set_header_bytenr(buf, buf->start);
8578         btrfs_set_header_generation(buf, trans->transid);
8579         btrfs_set_header_backref_rev(buf, BTRFS_MIXED_BACKREF_REV);
8580         btrfs_set_header_owner(buf, owner);
8581         write_extent_buffer_fsid(buf, fs_info->fs_devices->metadata_uuid);
8582         write_extent_buffer_chunk_tree_uuid(buf, fs_info->chunk_tree_uuid);
8583         if (root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID) {
8584                 buf->log_index = root->log_transid % 2;
8585                 /*
8586                  * we allow two log transactions at a time, use different
8587                  * EXTENT bit to differentiate dirty pages.
8588                  */
8589                 if (buf->log_index == 0)
8590                         set_extent_dirty(&root->dirty_log_pages, buf->start,
8591                                         buf->start + buf->len - 1, GFP_NOFS);
8592                 else
8593                         set_extent_new(&root->dirty_log_pages, buf->start,
8594                                         buf->start + buf->len - 1);
8595         } else {
8596                 buf->log_index = -1;
8597                 set_extent_dirty(&trans->transaction->dirty_pages, buf->start,
8598                          buf->start + buf->len - 1, GFP_NOFS);
8599         }
8600         trans->dirty = true;
8601         /* this returns a buffer locked for blocking */
8602         return buf;
8603 }
8604
8605 static struct btrfs_block_rsv *
8606 use_block_rsv(struct btrfs_trans_handle *trans,
8607               struct btrfs_root *root, u32 blocksize)
8608 {
8609         struct btrfs_fs_info *fs_info = root->fs_info;
8610         struct btrfs_block_rsv *block_rsv;
8611         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
8612         int ret;
8613         bool global_updated = false;
8614
8615         block_rsv = get_block_rsv(trans, root);
8616
8617         if (unlikely(block_rsv->size == 0))
8618                 goto try_reserve;
8619 again:
8620         ret = block_rsv_use_bytes(block_rsv, blocksize);
8621         if (!ret)
8622                 return block_rsv;
8623
8624         if (block_rsv->failfast)
8625                 return ERR_PTR(ret);
8626
8627         if (block_rsv->type == BTRFS_BLOCK_RSV_GLOBAL && !global_updated) {
8628                 global_updated = true;
8629                 update_global_block_rsv(fs_info);
8630                 goto again;
8631         }
8632
8633         /*
8634          * The global reserve still exists to save us from ourselves, so don't
8635          * warn_on if we are short on our delayed refs reserve.
8636          */
8637         if (block_rsv->type != BTRFS_BLOCK_RSV_DELREFS &&
8638             btrfs_test_opt(fs_info, ENOSPC_DEBUG)) {
8639                 static DEFINE_RATELIMIT_STATE(_rs,
8640                                 DEFAULT_RATELIMIT_INTERVAL * 10,
8641                                 /*DEFAULT_RATELIMIT_BURST*/ 1);
8642                 if (__ratelimit(&_rs))
8643                         WARN(1, KERN_DEBUG
8644                                 "BTRFS: block rsv returned %d\n", ret);
8645         }
8646 try_reserve:
8647         ret = reserve_metadata_bytes(root, block_rsv, blocksize,
8648                                      BTRFS_RESERVE_NO_FLUSH);
8649         if (!ret)
8650                 return block_rsv;
8651         /*
8652          * If we couldn't reserve metadata bytes try and use some from
8653          * the global reserve if its space type is the same as the global
8654          * reservation.
8655          */
8656         if (block_rsv->type != BTRFS_BLOCK_RSV_GLOBAL &&
8657             block_rsv->space_info == global_rsv->space_info) {
8658                 ret = block_rsv_use_bytes(global_rsv, blocksize);
8659                 if (!ret)
8660                         return global_rsv;
8661         }
8662         return ERR_PTR(ret);
8663 }
8664
8665 static void unuse_block_rsv(struct btrfs_fs_info *fs_info,
8666                             struct btrfs_block_rsv *block_rsv, u32 blocksize)
8667 {
8668         block_rsv_add_bytes(block_rsv, blocksize, false);
8669         block_rsv_release_bytes(fs_info, block_rsv, NULL, 0, NULL);
8670 }
8671
8672 /*
8673  * finds a free extent and does all the dirty work required for allocation
8674  * returns the tree buffer or an ERR_PTR on error.
8675  */
8676 struct extent_buffer *btrfs_alloc_tree_block(struct btrfs_trans_handle *trans,
8677                                              struct btrfs_root *root,
8678                                              u64 parent, u64 root_objectid,
8679                                              const struct btrfs_disk_key *key,
8680                                              int level, u64 hint,
8681                                              u64 empty_size)
8682 {
8683         struct btrfs_fs_info *fs_info = root->fs_info;
8684         struct btrfs_key ins;
8685         struct btrfs_block_rsv *block_rsv;
8686         struct extent_buffer *buf;
8687         struct btrfs_delayed_extent_op *extent_op;
8688         u64 flags = 0;
8689         int ret;
8690         u32 blocksize = fs_info->nodesize;
8691         bool skinny_metadata = btrfs_fs_incompat(fs_info, SKINNY_METADATA);
8692
8693 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
8694         if (btrfs_is_testing(fs_info)) {
8695                 buf = btrfs_init_new_buffer(trans, root, root->alloc_bytenr,
8696                                             level, root_objectid);
8697                 if (!IS_ERR(buf))
8698                         root->alloc_bytenr += blocksize;
8699                 return buf;
8700         }
8701 #endif
8702
8703         block_rsv = use_block_rsv(trans, root, blocksize);
8704         if (IS_ERR(block_rsv))
8705                 return ERR_CAST(block_rsv);
8706
8707         ret = btrfs_reserve_extent(root, blocksize, blocksize, blocksize,
8708                                    empty_size, hint, &ins, 0, 0);
8709         if (ret)
8710                 goto out_unuse;
8711
8712         buf = btrfs_init_new_buffer(trans, root, ins.objectid, level,
8713                                     root_objectid);
8714         if (IS_ERR(buf)) {
8715                 ret = PTR_ERR(buf);
8716                 goto out_free_reserved;
8717         }
8718
8719         if (root_objectid == BTRFS_TREE_RELOC_OBJECTID) {
8720                 if (parent == 0)
8721                         parent = ins.objectid;
8722                 flags |= BTRFS_BLOCK_FLAG_FULL_BACKREF;
8723         } else
8724                 BUG_ON(parent > 0);
8725
8726         if (root_objectid != BTRFS_TREE_LOG_OBJECTID) {
8727                 extent_op = btrfs_alloc_delayed_extent_op();
8728                 if (!extent_op) {
8729                         ret = -ENOMEM;
8730                         goto out_free_buf;
8731                 }
8732                 if (key)
8733                         memcpy(&extent_op->key, key, sizeof(extent_op->key));
8734                 else
8735                         memset(&extent_op->key, 0, sizeof(extent_op->key));
8736                 extent_op->flags_to_set = flags;
8737                 extent_op->update_key = skinny_metadata ? false : true;
8738                 extent_op->update_flags = true;
8739                 extent_op->is_data = false;
8740                 extent_op->level = level;
8741
8742                 btrfs_ref_tree_mod(root, ins.objectid, ins.offset, parent,
8743                                    root_objectid, level, 0,
8744                                    BTRFS_ADD_DELAYED_EXTENT);
8745                 ret = btrfs_add_delayed_tree_ref(trans, ins.objectid,
8746                                                  ins.offset, parent,
8747                                                  root_objectid, level,
8748                                                  BTRFS_ADD_DELAYED_EXTENT,
8749                                                  extent_op, NULL, NULL);
8750                 if (ret)
8751                         goto out_free_delayed;
8752         }
8753         return buf;
8754
8755 out_free_delayed:
8756         btrfs_free_delayed_extent_op(extent_op);
8757 out_free_buf:
8758         free_extent_buffer(buf);
8759 out_free_reserved:
8760         btrfs_free_reserved_extent(fs_info, ins.objectid, ins.offset, 0);
8761 out_unuse:
8762         unuse_block_rsv(fs_info, block_rsv, blocksize);
8763         return ERR_PTR(ret);
8764 }
8765
8766 struct walk_control {
8767         u64 refs[BTRFS_MAX_LEVEL];
8768         u64 flags[BTRFS_MAX_LEVEL];
8769         struct btrfs_key update_progress;
8770         int stage;
8771         int level;
8772         int shared_level;
8773         int update_ref;
8774         int keep_locks;
8775         int reada_slot;
8776         int reada_count;
8777 };
8778
8779 #define DROP_REFERENCE  1
8780 #define UPDATE_BACKREF  2
8781
8782 static noinline void reada_walk_down(struct btrfs_trans_handle *trans,
8783                                      struct btrfs_root *root,
8784                                      struct walk_control *wc,
8785                                      struct btrfs_path *path)
8786 {
8787         struct btrfs_fs_info *fs_info = root->fs_info;
8788         u64 bytenr;
8789         u64 generation;
8790         u64 refs;
8791         u64 flags;
8792         u32 nritems;
8793         struct btrfs_key key;
8794         struct extent_buffer *eb;
8795         int ret;
8796         int slot;
8797         int nread = 0;
8798
8799         if (path->slots[wc->level] < wc->reada_slot) {
8800                 wc->reada_count = wc->reada_count * 2 / 3;
8801                 wc->reada_count = max(wc->reada_count, 2);
8802         } else {
8803                 wc->reada_count = wc->reada_count * 3 / 2;
8804                 wc->reada_count = min_t(int, wc->reada_count,
8805                                         BTRFS_NODEPTRS_PER_BLOCK(fs_info));
8806         }
8807
8808         eb = path->nodes[wc->level];
8809         nritems = btrfs_header_nritems(eb);
8810
8811         for (slot = path->slots[wc->level]; slot < nritems; slot++) {
8812                 if (nread >= wc->reada_count)
8813                         break;
8814
8815                 cond_resched();
8816                 bytenr = btrfs_node_blockptr(eb, slot);
8817                 generation = btrfs_node_ptr_generation(eb, slot);
8818
8819                 if (slot == path->slots[wc->level])
8820                         goto reada;
8821
8822                 if (wc->stage == UPDATE_BACKREF &&
8823                     generation <= root->root_key.offset)
8824                         continue;
8825
8826                 /* We don't lock the tree block, it's OK to be racy here */
8827                 ret = btrfs_lookup_extent_info(trans, fs_info, bytenr,
8828                                                wc->level - 1, 1, &refs,
8829                                                &flags);
8830                 /* We don't care about errors in readahead. */
8831                 if (ret < 0)
8832                         continue;
8833                 BUG_ON(refs == 0);
8834
8835                 if (wc->stage == DROP_REFERENCE) {
8836                         if (refs == 1)
8837                                 goto reada;
8838
8839                         if (wc->level == 1 &&
8840                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
8841                                 continue;
8842                         if (!wc->update_ref ||
8843                             generation <= root->root_key.offset)
8844                                 continue;
8845                         btrfs_node_key_to_cpu(eb, &key, slot);
8846                         ret = btrfs_comp_cpu_keys(&key,
8847                                                   &wc->update_progress);
8848                         if (ret < 0)
8849                                 continue;
8850                 } else {
8851                         if (wc->level == 1 &&
8852                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
8853                                 continue;
8854                 }
8855 reada:
8856                 readahead_tree_block(fs_info, bytenr);
8857                 nread++;
8858         }
8859         wc->reada_slot = slot;
8860 }
8861
8862 /*
8863  * helper to process tree block while walking down the tree.
8864  *
8865  * when wc->stage == UPDATE_BACKREF, this function updates
8866  * back refs for pointers in the block.
8867  *
8868  * NOTE: return value 1 means we should stop walking down.
8869  */
8870 static noinline int walk_down_proc(struct btrfs_trans_handle *trans,
8871                                    struct btrfs_root *root,
8872                                    struct btrfs_path *path,
8873                                    struct walk_control *wc, int lookup_info)
8874 {
8875         struct btrfs_fs_info *fs_info = root->fs_info;
8876         int level = wc->level;
8877         struct extent_buffer *eb = path->nodes[level];
8878         u64 flag = BTRFS_BLOCK_FLAG_FULL_BACKREF;
8879         int ret;
8880
8881         if (wc->stage == UPDATE_BACKREF &&
8882             btrfs_header_owner(eb) != root->root_key.objectid)
8883                 return 1;
8884
8885         /*
8886          * when reference count of tree block is 1, it won't increase
8887          * again. once full backref flag is set, we never clear it.
8888          */
8889         if (lookup_info &&
8890             ((wc->stage == DROP_REFERENCE && wc->refs[level] != 1) ||
8891              (wc->stage == UPDATE_BACKREF && !(wc->flags[level] & flag)))) {
8892                 BUG_ON(!path->locks[level]);
8893                 ret = btrfs_lookup_extent_info(trans, fs_info,
8894                                                eb->start, level, 1,
8895                                                &wc->refs[level],
8896                                                &wc->flags[level]);
8897                 BUG_ON(ret == -ENOMEM);
8898                 if (ret)
8899                         return ret;
8900                 BUG_ON(wc->refs[level] == 0);
8901         }
8902
8903         if (wc->stage == DROP_REFERENCE) {
8904                 if (wc->refs[level] > 1)
8905                         return 1;
8906
8907                 if (path->locks[level] && !wc->keep_locks) {
8908                         btrfs_tree_unlock_rw(eb, path->locks[level]);
8909                         path->locks[level] = 0;
8910                 }
8911                 return 0;
8912         }
8913
8914         /* wc->stage == UPDATE_BACKREF */
8915         if (!(wc->flags[level] & flag)) {
8916                 BUG_ON(!path->locks[level]);
8917                 ret = btrfs_inc_ref(trans, root, eb, 1);
8918                 BUG_ON(ret); /* -ENOMEM */
8919                 ret = btrfs_dec_ref(trans, root, eb, 0);
8920                 BUG_ON(ret); /* -ENOMEM */
8921                 ret = btrfs_set_disk_extent_flags(trans, fs_info, eb->start,
8922                                                   eb->len, flag,
8923                                                   btrfs_header_level(eb), 0);
8924                 BUG_ON(ret); /* -ENOMEM */
8925                 wc->flags[level] |= flag;
8926         }
8927
8928         /*
8929          * the block is shared by multiple trees, so it's not good to
8930          * keep the tree lock
8931          */
8932         if (path->locks[level] && level > 0) {
8933                 btrfs_tree_unlock_rw(eb, path->locks[level]);
8934                 path->locks[level] = 0;
8935         }
8936         return 0;
8937 }
8938
8939 /*
8940  * helper to process tree block pointer.
8941  *
8942  * when wc->stage == DROP_REFERENCE, this function checks
8943  * reference count of the block pointed to. if the block
8944  * is shared and we need update back refs for the subtree
8945  * rooted at the block, this function changes wc->stage to
8946  * UPDATE_BACKREF. if the block is shared and there is no
8947  * need to update back, this function drops the reference
8948  * to the block.
8949  *
8950  * NOTE: return value 1 means we should stop walking down.
8951  */
8952 static noinline int do_walk_down(struct btrfs_trans_handle *trans,
8953                                  struct btrfs_root *root,
8954                                  struct btrfs_path *path,
8955                                  struct walk_control *wc, int *lookup_info)
8956 {
8957         struct btrfs_fs_info *fs_info = root->fs_info;
8958         u64 bytenr;
8959         u64 generation;
8960         u64 parent;
8961         struct btrfs_key key;
8962         struct btrfs_key first_key;
8963         struct extent_buffer *next;
8964         int level = wc->level;
8965         int reada = 0;
8966         int ret = 0;
8967         bool need_account = false;
8968
8969         generation = btrfs_node_ptr_generation(path->nodes[level],
8970                                                path->slots[level]);
8971         /*
8972          * if the lower level block was created before the snapshot
8973          * was created, we know there is no need to update back refs
8974          * for the subtree
8975          */
8976         if (wc->stage == UPDATE_BACKREF &&
8977             generation <= root->root_key.offset) {
8978                 *lookup_info = 1;
8979                 return 1;
8980         }
8981
8982         bytenr = btrfs_node_blockptr(path->nodes[level], path->slots[level]);
8983         btrfs_node_key_to_cpu(path->nodes[level], &first_key,
8984                               path->slots[level]);
8985
8986         next = find_extent_buffer(fs_info, bytenr);
8987         if (!next) {
8988                 next = btrfs_find_create_tree_block(fs_info, bytenr);
8989                 if (IS_ERR(next))
8990                         return PTR_ERR(next);
8991
8992                 btrfs_set_buffer_lockdep_class(root->root_key.objectid, next,
8993                                                level - 1);
8994                 reada = 1;
8995         }
8996         btrfs_tree_lock(next);
8997         btrfs_set_lock_blocking_write(next);
8998
8999         ret = btrfs_lookup_extent_info(trans, fs_info, bytenr, level - 1, 1,
9000                                        &wc->refs[level - 1],
9001                                        &wc->flags[level - 1]);
9002         if (ret < 0)
9003                 goto out_unlock;
9004
9005         if (unlikely(wc->refs[level - 1] == 0)) {
9006                 btrfs_err(fs_info, "Missing references.");
9007                 ret = -EIO;
9008                 goto out_unlock;
9009         }
9010         *lookup_info = 0;
9011
9012         if (wc->stage == DROP_REFERENCE) {
9013                 if (wc->refs[level - 1] > 1) {
9014                         need_account = true;
9015                         if (level == 1 &&
9016                             (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
9017                                 goto skip;
9018
9019                         if (!wc->update_ref ||
9020                             generation <= root->root_key.offset)
9021                                 goto skip;
9022
9023                         btrfs_node_key_to_cpu(path->nodes[level], &key,
9024                                               path->slots[level]);
9025                         ret = btrfs_comp_cpu_keys(&key, &wc->update_progress);
9026                         if (ret < 0)
9027                                 goto skip;
9028
9029                         wc->stage = UPDATE_BACKREF;
9030                         wc->shared_level = level - 1;
9031                 }
9032         } else {
9033                 if (level == 1 &&
9034                     (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
9035                         goto skip;
9036         }
9037
9038         if (!btrfs_buffer_uptodate(next, generation, 0)) {
9039                 btrfs_tree_unlock(next);
9040                 free_extent_buffer(next);
9041                 next = NULL;
9042                 *lookup_info = 1;
9043         }
9044
9045         if (!next) {
9046                 if (reada && level == 1)
9047                         reada_walk_down(trans, root, wc, path);
9048                 next = read_tree_block(fs_info, bytenr, generation, level - 1,
9049                                        &first_key);
9050                 if (IS_ERR(next)) {
9051                         return PTR_ERR(next);
9052                 } else if (!extent_buffer_uptodate(next)) {
9053                         free_extent_buffer(next);
9054                         return -EIO;
9055                 }
9056                 btrfs_tree_lock(next);
9057                 btrfs_set_lock_blocking_write(next);
9058         }
9059
9060         level--;
9061         ASSERT(level == btrfs_header_level(next));
9062         if (level != btrfs_header_level(next)) {
9063                 btrfs_err(root->fs_info, "mismatched level");
9064                 ret = -EIO;
9065                 goto out_unlock;
9066         }
9067         path->nodes[level] = next;
9068         path->slots[level] = 0;
9069         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
9070         wc->level = level;
9071         if (wc->level == 1)
9072                 wc->reada_slot = 0;
9073         return 0;
9074 skip:
9075         wc->refs[level - 1] = 0;
9076         wc->flags[level - 1] = 0;
9077         if (wc->stage == DROP_REFERENCE) {
9078                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF) {
9079                         parent = path->nodes[level]->start;
9080                 } else {
9081                         ASSERT(root->root_key.objectid ==
9082                                btrfs_header_owner(path->nodes[level]));
9083                         if (root->root_key.objectid !=
9084                             btrfs_header_owner(path->nodes[level])) {
9085                                 btrfs_err(root->fs_info,
9086                                                 "mismatched block owner");
9087                                 ret = -EIO;
9088                                 goto out_unlock;
9089                         }
9090                         parent = 0;
9091                 }
9092
9093                 /*
9094                  * Reloc tree doesn't contribute to qgroup numbers, and we have
9095                  * already accounted them at merge time (replace_path),
9096                  * thus we could skip expensive subtree trace here.
9097                  */
9098                 if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID &&
9099                     need_account) {
9100                         ret = btrfs_qgroup_trace_subtree(trans, next,
9101                                                          generation, level - 1);
9102                         if (ret) {
9103                                 btrfs_err_rl(fs_info,
9104                                              "Error %d accounting shared subtree. Quota is out of sync, rescan required.",
9105                                              ret);
9106                         }
9107                 }
9108                 ret = btrfs_free_extent(trans, root, bytenr, fs_info->nodesize,
9109                                         parent, root->root_key.objectid,
9110                                         level - 1, 0);
9111                 if (ret)
9112                         goto out_unlock;
9113         }
9114
9115         *lookup_info = 1;
9116         ret = 1;
9117
9118 out_unlock:
9119         btrfs_tree_unlock(next);
9120         free_extent_buffer(next);
9121
9122         return ret;
9123 }
9124
9125 /*
9126  * helper to process tree block while walking up the tree.
9127  *
9128  * when wc->stage == DROP_REFERENCE, this function drops
9129  * reference count on the block.
9130  *
9131  * when wc->stage == UPDATE_BACKREF, this function changes
9132  * wc->stage back to DROP_REFERENCE if we changed wc->stage
9133  * to UPDATE_BACKREF previously while processing the block.
9134  *
9135  * NOTE: return value 1 means we should stop walking up.
9136  */
9137 static noinline int walk_up_proc(struct btrfs_trans_handle *trans,
9138                                  struct btrfs_root *root,
9139                                  struct btrfs_path *path,
9140                                  struct walk_control *wc)
9141 {
9142         struct btrfs_fs_info *fs_info = root->fs_info;
9143         int ret;
9144         int level = wc->level;
9145         struct extent_buffer *eb = path->nodes[level];
9146         u64 parent = 0;
9147
9148         if (wc->stage == UPDATE_BACKREF) {
9149                 BUG_ON(wc->shared_level < level);
9150                 if (level < wc->shared_level)
9151                         goto out;
9152
9153                 ret = find_next_key(path, level + 1, &wc->update_progress);
9154                 if (ret > 0)
9155                         wc->update_ref = 0;
9156
9157                 wc->stage = DROP_REFERENCE;
9158                 wc->shared_level = -1;
9159                 path->slots[level] = 0;
9160
9161                 /*
9162                  * check reference count again if the block isn't locked.
9163                  * we should start walking down the tree again if reference
9164                  * count is one.
9165                  */
9166                 if (!path->locks[level]) {
9167                         BUG_ON(level == 0);
9168                         btrfs_tree_lock(eb);
9169                         btrfs_set_lock_blocking_write(eb);
9170                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
9171
9172                         ret = btrfs_lookup_extent_info(trans, fs_info,
9173                                                        eb->start, level, 1,
9174                                                        &wc->refs[level],
9175                                                        &wc->flags[level]);
9176                         if (ret < 0) {
9177                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
9178                                 path->locks[level] = 0;
9179                                 return ret;
9180                         }
9181                         BUG_ON(wc->refs[level] == 0);
9182                         if (wc->refs[level] == 1) {
9183                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
9184                                 path->locks[level] = 0;
9185                                 return 1;
9186                         }
9187                 }
9188         }
9189
9190         /* wc->stage == DROP_REFERENCE */
9191         BUG_ON(wc->refs[level] > 1 && !path->locks[level]);
9192
9193         if (wc->refs[level] == 1) {
9194                 if (level == 0) {
9195                         if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
9196                                 ret = btrfs_dec_ref(trans, root, eb, 1);
9197                         else
9198                                 ret = btrfs_dec_ref(trans, root, eb, 0);
9199                         BUG_ON(ret); /* -ENOMEM */
9200                         ret = btrfs_qgroup_trace_leaf_items(trans, eb);
9201                         if (ret) {
9202                                 btrfs_err_rl(fs_info,
9203                                              "error %d accounting leaf items. Quota is out of sync, rescan required.",
9204                                              ret);
9205                         }
9206                 }
9207                 /* make block locked assertion in clean_tree_block happy */
9208                 if (!path->locks[level] &&
9209                     btrfs_header_generation(eb) == trans->transid) {
9210                         btrfs_tree_lock(eb);
9211                         btrfs_set_lock_blocking_write(eb);
9212                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
9213                 }
9214                 clean_tree_block(fs_info, eb);
9215         }
9216
9217         if (eb == root->node) {
9218                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
9219                         parent = eb->start;
9220                 else if (root->root_key.objectid != btrfs_header_owner(eb))
9221                         goto owner_mismatch;
9222         } else {
9223                 if (wc->flags[level + 1] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
9224                         parent = path->nodes[level + 1]->start;
9225                 else if (root->root_key.objectid !=
9226                          btrfs_header_owner(path->nodes[level + 1]))
9227                         goto owner_mismatch;
9228         }
9229
9230         btrfs_free_tree_block(trans, root, eb, parent, wc->refs[level] == 1);
9231 out:
9232         wc->refs[level] = 0;
9233         wc->flags[level] = 0;
9234         return 0;
9235
9236 owner_mismatch:
9237         btrfs_err_rl(fs_info, "unexpected tree owner, have %llu expect %llu",
9238                      btrfs_header_owner(eb), root->root_key.objectid);
9239         return -EUCLEAN;
9240 }
9241
9242 static noinline int walk_down_tree(struct btrfs_trans_handle *trans,
9243                                    struct btrfs_root *root,
9244                                    struct btrfs_path *path,
9245                                    struct walk_control *wc)
9246 {
9247         int level = wc->level;
9248         int lookup_info = 1;
9249         int ret;
9250
9251         while (level >= 0) {
9252                 ret = walk_down_proc(trans, root, path, wc, lookup_info);
9253                 if (ret > 0)
9254                         break;
9255
9256                 if (level == 0)
9257                         break;
9258
9259                 if (path->slots[level] >=
9260                     btrfs_header_nritems(path->nodes[level]))
9261                         break;
9262
9263                 ret = do_walk_down(trans, root, path, wc, &lookup_info);
9264                 if (ret > 0) {
9265                         path->slots[level]++;
9266                         continue;
9267                 } else if (ret < 0)
9268                         return ret;
9269                 level = wc->level;
9270         }
9271         return 0;
9272 }
9273
9274 static noinline int walk_up_tree(struct btrfs_trans_handle *trans,
9275                                  struct btrfs_root *root,
9276                                  struct btrfs_path *path,
9277                                  struct walk_control *wc, int max_level)
9278 {
9279         int level = wc->level;
9280         int ret;
9281
9282         path->slots[level] = btrfs_header_nritems(path->nodes[level]);
9283         while (level < max_level && path->nodes[level]) {
9284                 wc->level = level;
9285                 if (path->slots[level] + 1 <
9286                     btrfs_header_nritems(path->nodes[level])) {
9287                         path->slots[level]++;
9288                         return 0;
9289                 } else {
9290                         ret = walk_up_proc(trans, root, path, wc);
9291                         if (ret > 0)
9292                                 return 0;
9293                         if (ret < 0)
9294                                 return ret;
9295
9296                         if (path->locks[level]) {
9297                                 btrfs_tree_unlock_rw(path->nodes[level],
9298                                                      path->locks[level]);
9299                                 path->locks[level] = 0;
9300                         }
9301                         free_extent_buffer(path->nodes[level]);
9302                         path->nodes[level] = NULL;
9303                         level++;
9304                 }
9305         }
9306         return 1;
9307 }
9308
9309 /*
9310  * drop a subvolume tree.
9311  *
9312  * this function traverses the tree freeing any blocks that only
9313  * referenced by the tree.
9314  *
9315  * when a shared tree block is found. this function decreases its
9316  * reference count by one. if update_ref is true, this function
9317  * also make sure backrefs for the shared block and all lower level
9318  * blocks are properly updated.
9319  *
9320  * If called with for_reloc == 0, may exit early with -EAGAIN
9321  */
9322 int btrfs_drop_snapshot(struct btrfs_root *root,
9323                          struct btrfs_block_rsv *block_rsv, int update_ref,
9324                          int for_reloc)
9325 {
9326         struct btrfs_fs_info *fs_info = root->fs_info;
9327         struct btrfs_path *path;
9328         struct btrfs_trans_handle *trans;
9329         struct btrfs_root *tree_root = fs_info->tree_root;
9330         struct btrfs_root_item *root_item = &root->root_item;
9331         struct walk_control *wc;
9332         struct btrfs_key key;
9333         int err = 0;
9334         int ret;
9335         int level;
9336         bool root_dropped = false;
9337
9338         btrfs_debug(fs_info, "Drop subvolume %llu", root->root_key.objectid);
9339
9340         path = btrfs_alloc_path();
9341         if (!path) {
9342                 err = -ENOMEM;
9343                 goto out;
9344         }
9345
9346         wc = kzalloc(sizeof(*wc), GFP_NOFS);
9347         if (!wc) {
9348                 btrfs_free_path(path);
9349                 err = -ENOMEM;
9350                 goto out;
9351         }
9352
9353         trans = btrfs_start_transaction(tree_root, 0);
9354         if (IS_ERR(trans)) {
9355                 err = PTR_ERR(trans);
9356                 goto out_free;
9357         }
9358
9359         err = btrfs_run_delayed_items(trans);
9360         if (err)
9361                 goto out_end_trans;
9362
9363         if (block_rsv)
9364                 trans->block_rsv = block_rsv;
9365
9366         /*
9367          * This will help us catch people modifying the fs tree while we're
9368          * dropping it.  It is unsafe to mess with the fs tree while it's being
9369          * dropped as we unlock the root node and parent nodes as we walk down
9370          * the tree, assuming nothing will change.  If something does change
9371          * then we'll have stale information and drop references to blocks we've
9372          * already dropped.
9373          */
9374         set_bit(BTRFS_ROOT_DELETING, &root->state);
9375         if (btrfs_disk_key_objectid(&root_item->drop_progress) == 0) {
9376                 level = btrfs_header_level(root->node);
9377                 path->nodes[level] = btrfs_lock_root_node(root);
9378                 btrfs_set_lock_blocking_write(path->nodes[level]);
9379                 path->slots[level] = 0;
9380                 path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
9381                 memset(&wc->update_progress, 0,
9382                        sizeof(wc->update_progress));
9383         } else {
9384                 btrfs_disk_key_to_cpu(&key, &root_item->drop_progress);
9385                 memcpy(&wc->update_progress, &key,
9386                        sizeof(wc->update_progress));
9387
9388                 level = root_item->drop_level;
9389                 BUG_ON(level == 0);
9390                 path->lowest_level = level;
9391                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
9392                 path->lowest_level = 0;
9393                 if (ret < 0) {
9394                         err = ret;
9395                         goto out_end_trans;
9396                 }
9397                 WARN_ON(ret > 0);
9398
9399                 /*
9400                  * unlock our path, this is safe because only this
9401                  * function is allowed to delete this snapshot
9402                  */
9403                 btrfs_unlock_up_safe(path, 0);
9404
9405                 level = btrfs_header_level(root->node);
9406                 while (1) {
9407                         btrfs_tree_lock(path->nodes[level]);
9408                         btrfs_set_lock_blocking_write(path->nodes[level]);
9409                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
9410
9411                         ret = btrfs_lookup_extent_info(trans, fs_info,
9412                                                 path->nodes[level]->start,
9413                                                 level, 1, &wc->refs[level],
9414                                                 &wc->flags[level]);
9415                         if (ret < 0) {
9416                                 err = ret;
9417                                 goto out_end_trans;
9418                         }
9419                         BUG_ON(wc->refs[level] == 0);
9420
9421                         if (level == root_item->drop_level)
9422                                 break;
9423
9424                         btrfs_tree_unlock(path->nodes[level]);
9425                         path->locks[level] = 0;
9426                         WARN_ON(wc->refs[level] != 1);
9427                         level--;
9428                 }
9429         }
9430
9431         wc->level = level;
9432         wc->shared_level = -1;
9433         wc->stage = DROP_REFERENCE;
9434         wc->update_ref = update_ref;
9435         wc->keep_locks = 0;
9436         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(fs_info);
9437
9438         while (1) {
9439
9440                 ret = walk_down_tree(trans, root, path, wc);
9441                 if (ret < 0) {
9442                         err = ret;
9443                         break;
9444                 }
9445
9446                 ret = walk_up_tree(trans, root, path, wc, BTRFS_MAX_LEVEL);
9447                 if (ret < 0) {
9448                         err = ret;
9449                         break;
9450                 }
9451
9452                 if (ret > 0) {
9453                         BUG_ON(wc->stage != DROP_REFERENCE);
9454                         break;
9455                 }
9456
9457                 if (wc->stage == DROP_REFERENCE) {
9458                         level = wc->level;
9459                         btrfs_node_key(path->nodes[level],
9460                                        &root_item->drop_progress,
9461                                        path->slots[level]);
9462                         root_item->drop_level = level;
9463                 }
9464
9465                 BUG_ON(wc->level == 0);
9466                 if (btrfs_should_end_transaction(trans) ||
9467                     (!for_reloc && btrfs_need_cleaner_sleep(fs_info))) {
9468                         ret = btrfs_update_root(trans, tree_root,
9469                                                 &root->root_key,
9470                                                 root_item);
9471                         if (ret) {
9472                                 btrfs_abort_transaction(trans, ret);
9473                                 err = ret;
9474                                 goto out_end_trans;
9475                         }
9476
9477                         btrfs_end_transaction_throttle(trans);
9478                         if (!for_reloc && btrfs_need_cleaner_sleep(fs_info)) {
9479                                 btrfs_debug(fs_info,
9480                                             "drop snapshot early exit");
9481                                 err = -EAGAIN;
9482                                 goto out_free;
9483                         }
9484
9485                         trans = btrfs_start_transaction(tree_root, 0);
9486                         if (IS_ERR(trans)) {
9487                                 err = PTR_ERR(trans);
9488                                 goto out_free;
9489                         }
9490                         if (block_rsv)
9491                                 trans->block_rsv = block_rsv;
9492                 }
9493         }
9494         btrfs_release_path(path);
9495         if (err)
9496                 goto out_end_trans;
9497
9498         ret = btrfs_del_root(trans, &root->root_key);
9499         if (ret) {
9500                 btrfs_abort_transaction(trans, ret);
9501                 err = ret;
9502                 goto out_end_trans;
9503         }
9504
9505         if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) {
9506                 ret = btrfs_find_root(tree_root, &root->root_key, path,
9507                                       NULL, NULL);
9508                 if (ret < 0) {
9509                         btrfs_abort_transaction(trans, ret);
9510                         err = ret;
9511                         goto out_end_trans;
9512                 } else if (ret > 0) {
9513                         /* if we fail to delete the orphan item this time
9514                          * around, it'll get picked up the next time.
9515                          *
9516                          * The most common failure here is just -ENOENT.
9517                          */
9518                         btrfs_del_orphan_item(trans, tree_root,
9519                                               root->root_key.objectid);
9520                 }
9521         }
9522
9523         if (test_bit(BTRFS_ROOT_IN_RADIX, &root->state)) {
9524                 btrfs_add_dropped_root(trans, root);
9525         } else {
9526                 free_extent_buffer(root->node);
9527                 free_extent_buffer(root->commit_root);
9528                 btrfs_put_fs_root(root);
9529         }
9530         root_dropped = true;
9531 out_end_trans:
9532         btrfs_end_transaction_throttle(trans);
9533 out_free:
9534         kfree(wc);
9535         btrfs_free_path(path);
9536 out:
9537         /*
9538          * So if we need to stop dropping the snapshot for whatever reason we
9539          * need to make sure to add it back to the dead root list so that we
9540          * keep trying to do the work later.  This also cleans up roots if we
9541          * don't have it in the radix (like when we recover after a power fail
9542          * or unmount) so we don't leak memory.
9543          */
9544         if (!for_reloc && !root_dropped)
9545                 btrfs_add_dead_root(root);
9546         if (err && err != -EAGAIN)
9547                 btrfs_handle_fs_error(fs_info, err, NULL);
9548         return err;
9549 }
9550
9551 /*
9552  * drop subtree rooted at tree block 'node'.
9553  *
9554  * NOTE: this function will unlock and release tree block 'node'
9555  * only used by relocation code
9556  */
9557 int btrfs_drop_subtree(struct btrfs_trans_handle *trans,
9558                         struct btrfs_root *root,
9559                         struct extent_buffer *node,
9560                         struct extent_buffer *parent)
9561 {
9562         struct btrfs_fs_info *fs_info = root->fs_info;
9563         struct btrfs_path *path;
9564         struct walk_control *wc;
9565         int level;
9566         int parent_level;
9567         int ret = 0;
9568         int wret;
9569
9570         BUG_ON(root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID);
9571
9572         path = btrfs_alloc_path();
9573         if (!path)
9574                 return -ENOMEM;
9575
9576         wc = kzalloc(sizeof(*wc), GFP_NOFS);
9577         if (!wc) {
9578                 btrfs_free_path(path);
9579                 return -ENOMEM;
9580         }
9581
9582         btrfs_assert_tree_locked(parent);
9583         parent_level = btrfs_header_level(parent);
9584         extent_buffer_get(parent);
9585         path->nodes[parent_level] = parent;
9586         path->slots[parent_level] = btrfs_header_nritems(parent);
9587
9588         btrfs_assert_tree_locked(node);
9589         level = btrfs_header_level(node);
9590         path->nodes[level] = node;
9591         path->slots[level] = 0;
9592         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
9593
9594         wc->refs[parent_level] = 1;
9595         wc->flags[parent_level] = BTRFS_BLOCK_FLAG_FULL_BACKREF;
9596         wc->level = level;
9597         wc->shared_level = -1;
9598         wc->stage = DROP_REFERENCE;
9599         wc->update_ref = 0;
9600         wc->keep_locks = 1;
9601         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(fs_info);
9602
9603         while (1) {
9604                 wret = walk_down_tree(trans, root, path, wc);
9605                 if (wret < 0) {
9606                         ret = wret;
9607                         break;
9608                 }
9609
9610                 wret = walk_up_tree(trans, root, path, wc, parent_level);
9611                 if (wret < 0)
9612                         ret = wret;
9613                 if (wret != 0)
9614                         break;
9615         }
9616
9617         kfree(wc);
9618         btrfs_free_path(path);
9619         return ret;
9620 }
9621
9622 static u64 update_block_group_flags(struct btrfs_fs_info *fs_info, u64 flags)
9623 {
9624         u64 num_devices;
9625         u64 stripped;
9626
9627         /*
9628          * if restripe for this chunk_type is on pick target profile and
9629          * return, otherwise do the usual balance
9630          */
9631         stripped = get_restripe_target(fs_info, flags);
9632         if (stripped)
9633                 return extended_to_chunk(stripped);
9634
9635         num_devices = fs_info->fs_devices->rw_devices;
9636
9637         stripped = BTRFS_BLOCK_GROUP_RAID0 |
9638                 BTRFS_BLOCK_GROUP_RAID5 | BTRFS_BLOCK_GROUP_RAID6 |
9639                 BTRFS_BLOCK_GROUP_RAID1 | BTRFS_BLOCK_GROUP_RAID10;
9640
9641         if (num_devices == 1) {
9642                 stripped |= BTRFS_BLOCK_GROUP_DUP;
9643                 stripped = flags & ~stripped;
9644
9645                 /* turn raid0 into single device chunks */
9646                 if (flags & BTRFS_BLOCK_GROUP_RAID0)
9647                         return stripped;
9648
9649                 /* turn mirroring into duplication */
9650                 if (flags & (BTRFS_BLOCK_GROUP_RAID1 |
9651                              BTRFS_BLOCK_GROUP_RAID10))
9652                         return stripped | BTRFS_BLOCK_GROUP_DUP;
9653         } else {
9654                 /* they already had raid on here, just return */
9655                 if (flags & stripped)
9656                         return flags;
9657
9658                 stripped |= BTRFS_BLOCK_GROUP_DUP;
9659                 stripped = flags & ~stripped;
9660
9661                 /* switch duplicated blocks with raid1 */
9662                 if (flags & BTRFS_BLOCK_GROUP_DUP)
9663                         return stripped | BTRFS_BLOCK_GROUP_RAID1;
9664
9665                 /* this is drive concat, leave it alone */
9666         }
9667
9668         return flags;
9669 }
9670
9671 static int inc_block_group_ro(struct btrfs_block_group_cache *cache, int force)
9672 {
9673         struct btrfs_space_info *sinfo = cache->space_info;
9674         u64 num_bytes;
9675         u64 sinfo_used;
9676         u64 min_allocable_bytes;
9677         int ret = -ENOSPC;
9678
9679         /*
9680          * We need some metadata space and system metadata space for
9681          * allocating chunks in some corner cases until we force to set
9682          * it to be readonly.
9683          */
9684         if ((sinfo->flags &
9685              (BTRFS_BLOCK_GROUP_SYSTEM | BTRFS_BLOCK_GROUP_METADATA)) &&
9686             !force)
9687                 min_allocable_bytes = SZ_1M;
9688         else
9689                 min_allocable_bytes = 0;
9690
9691         spin_lock(&sinfo->lock);
9692         spin_lock(&cache->lock);
9693
9694         if (cache->ro) {
9695                 cache->ro++;
9696                 ret = 0;
9697                 goto out;
9698         }
9699
9700         num_bytes = cache->key.offset - cache->reserved - cache->pinned -
9701                     cache->bytes_super - btrfs_block_group_used(&cache->item);
9702         sinfo_used = btrfs_space_info_used(sinfo, true);
9703
9704         if (sinfo_used + num_bytes + min_allocable_bytes <=
9705             sinfo->total_bytes) {
9706                 sinfo->bytes_readonly += num_bytes;
9707                 cache->ro++;
9708                 list_add_tail(&cache->ro_list, &sinfo->ro_bgs);
9709                 ret = 0;
9710         }
9711 out:
9712         spin_unlock(&cache->lock);
9713         spin_unlock(&sinfo->lock);
9714         if (ret == -ENOSPC && btrfs_test_opt(cache->fs_info, ENOSPC_DEBUG)) {
9715                 btrfs_info(cache->fs_info,
9716                         "unable to make block group %llu ro",
9717                         cache->key.objectid);
9718                 btrfs_info(cache->fs_info,
9719                         "sinfo_used=%llu bg_num_bytes=%llu min_allocable=%llu",
9720                         sinfo_used, num_bytes, min_allocable_bytes);
9721                 dump_space_info(cache->fs_info, cache->space_info, 0, 0);
9722         }
9723         return ret;
9724 }
9725
9726 int btrfs_inc_block_group_ro(struct btrfs_block_group_cache *cache)
9727
9728 {
9729         struct btrfs_fs_info *fs_info = cache->fs_info;
9730         struct btrfs_trans_handle *trans;
9731         u64 alloc_flags;
9732         int ret;
9733
9734 again:
9735         trans = btrfs_join_transaction(fs_info->extent_root);
9736         if (IS_ERR(trans))
9737                 return PTR_ERR(trans);
9738
9739         /*
9740          * we're not allowed to set block groups readonly after the dirty
9741          * block groups cache has started writing.  If it already started,
9742          * back off and let this transaction commit
9743          */
9744         mutex_lock(&fs_info->ro_block_group_mutex);
9745         if (test_bit(BTRFS_TRANS_DIRTY_BG_RUN, &trans->transaction->flags)) {
9746                 u64 transid = trans->transid;
9747
9748                 mutex_unlock(&fs_info->ro_block_group_mutex);
9749                 btrfs_end_transaction(trans);
9750
9751                 ret = btrfs_wait_for_commit(fs_info, transid);
9752                 if (ret)
9753                         return ret;
9754                 goto again;
9755         }
9756
9757         /*
9758          * if we are changing raid levels, try to allocate a corresponding
9759          * block group with the new raid level.
9760          */
9761         alloc_flags = update_block_group_flags(fs_info, cache->flags);
9762         if (alloc_flags != cache->flags) {
9763                 ret = do_chunk_alloc(trans, alloc_flags,
9764                                      CHUNK_ALLOC_FORCE);
9765                 /*
9766                  * ENOSPC is allowed here, we may have enough space
9767                  * already allocated at the new raid level to
9768                  * carry on
9769                  */
9770                 if (ret == -ENOSPC)
9771                         ret = 0;
9772                 if (ret < 0)
9773                         goto out;
9774         }
9775
9776         ret = inc_block_group_ro(cache, 0);
9777         if (!ret)
9778                 goto out;
9779         alloc_flags = get_alloc_profile(fs_info, cache->space_info->flags);
9780         ret = do_chunk_alloc(trans, alloc_flags, CHUNK_ALLOC_FORCE);
9781         if (ret < 0)
9782                 goto out;
9783         ret = inc_block_group_ro(cache, 0);
9784 out:
9785         if (cache->flags & BTRFS_BLOCK_GROUP_SYSTEM) {
9786                 alloc_flags = update_block_group_flags(fs_info, cache->flags);
9787                 mutex_lock(&fs_info->chunk_mutex);
9788                 check_system_chunk(trans, alloc_flags);
9789                 mutex_unlock(&fs_info->chunk_mutex);
9790         }
9791         mutex_unlock(&fs_info->ro_block_group_mutex);
9792
9793         btrfs_end_transaction(trans);
9794         return ret;
9795 }
9796
9797 int btrfs_force_chunk_alloc(struct btrfs_trans_handle *trans, u64 type)
9798 {
9799         u64 alloc_flags = get_alloc_profile(trans->fs_info, type);
9800
9801         return do_chunk_alloc(trans, alloc_flags, CHUNK_ALLOC_FORCE);
9802 }
9803
9804 /*
9805  * helper to account the unused space of all the readonly block group in the
9806  * space_info. takes mirrors into account.
9807  */
9808 u64 btrfs_account_ro_block_groups_free_space(struct btrfs_space_info *sinfo)
9809 {
9810         struct btrfs_block_group_cache *block_group;
9811         u64 free_bytes = 0;
9812         int factor;
9813
9814         /* It's df, we don't care if it's racy */
9815         if (list_empty(&sinfo->ro_bgs))
9816                 return 0;
9817
9818         spin_lock(&sinfo->lock);
9819         list_for_each_entry(block_group, &sinfo->ro_bgs, ro_list) {
9820                 spin_lock(&block_group->lock);
9821
9822                 if (!block_group->ro) {
9823                         spin_unlock(&block_group->lock);
9824                         continue;
9825                 }
9826
9827                 factor = btrfs_bg_type_to_factor(block_group->flags);
9828                 free_bytes += (block_group->key.offset -
9829                                btrfs_block_group_used(&block_group->item)) *
9830                                factor;
9831
9832                 spin_unlock(&block_group->lock);
9833         }
9834         spin_unlock(&sinfo->lock);
9835
9836         return free_bytes;
9837 }
9838
9839 void btrfs_dec_block_group_ro(struct btrfs_block_group_cache *cache)
9840 {
9841         struct btrfs_space_info *sinfo = cache->space_info;
9842         u64 num_bytes;
9843
9844         BUG_ON(!cache->ro);
9845
9846         spin_lock(&sinfo->lock);
9847         spin_lock(&cache->lock);
9848         if (!--cache->ro) {
9849                 num_bytes = cache->key.offset - cache->reserved -
9850                             cache->pinned - cache->bytes_super -
9851                             btrfs_block_group_used(&cache->item);
9852                 sinfo->bytes_readonly -= num_bytes;
9853                 list_del_init(&cache->ro_list);
9854         }
9855         spin_unlock(&cache->lock);
9856         spin_unlock(&sinfo->lock);
9857 }
9858
9859 /*
9860  * Checks to see if it's even possible to relocate this block group.
9861  *
9862  * @return - -1 if it's not a good idea to relocate this block group, 0 if its
9863  * ok to go ahead and try.
9864  */
9865 int btrfs_can_relocate(struct btrfs_fs_info *fs_info, u64 bytenr)
9866 {
9867         struct btrfs_root *root = fs_info->extent_root;
9868         struct btrfs_block_group_cache *block_group;
9869         struct btrfs_space_info *space_info;
9870         struct btrfs_fs_devices *fs_devices = fs_info->fs_devices;
9871         struct btrfs_device *device;
9872         struct btrfs_trans_handle *trans;
9873         u64 min_free;
9874         u64 dev_min = 1;
9875         u64 dev_nr = 0;
9876         u64 target;
9877         int debug;
9878         int index;
9879         int full = 0;
9880         int ret = 0;
9881
9882         debug = btrfs_test_opt(fs_info, ENOSPC_DEBUG);
9883
9884         block_group = btrfs_lookup_block_group(fs_info, bytenr);
9885
9886         /* odd, couldn't find the block group, leave it alone */
9887         if (!block_group) {
9888                 if (debug)
9889                         btrfs_warn(fs_info,
9890                                    "can't find block group for bytenr %llu",
9891                                    bytenr);
9892                 return -1;
9893         }
9894
9895         min_free = btrfs_block_group_used(&block_group->item);
9896
9897         /* no bytes used, we're good */
9898         if (!min_free)
9899                 goto out;
9900
9901         space_info = block_group->space_info;
9902         spin_lock(&space_info->lock);
9903
9904         full = space_info->full;
9905
9906         /*
9907          * if this is the last block group we have in this space, we can't
9908          * relocate it unless we're able to allocate a new chunk below.
9909          *
9910          * Otherwise, we need to make sure we have room in the space to handle
9911          * all of the extents from this block group.  If we can, we're good
9912          */
9913         if ((space_info->total_bytes != block_group->key.offset) &&
9914             (btrfs_space_info_used(space_info, false) + min_free <
9915              space_info->total_bytes)) {
9916                 spin_unlock(&space_info->lock);
9917                 goto out;
9918         }
9919         spin_unlock(&space_info->lock);
9920
9921         /*
9922          * ok we don't have enough space, but maybe we have free space on our
9923          * devices to allocate new chunks for relocation, so loop through our
9924          * alloc devices and guess if we have enough space.  if this block
9925          * group is going to be restriped, run checks against the target
9926          * profile instead of the current one.
9927          */
9928         ret = -1;
9929
9930         /*
9931          * index:
9932          *      0: raid10
9933          *      1: raid1
9934          *      2: dup
9935          *      3: raid0
9936          *      4: single
9937          */
9938         target = get_restripe_target(fs_info, block_group->flags);
9939         if (target) {
9940                 index = btrfs_bg_flags_to_raid_index(extended_to_chunk(target));
9941         } else {
9942                 /*
9943                  * this is just a balance, so if we were marked as full
9944                  * we know there is no space for a new chunk
9945                  */
9946                 if (full) {
9947                         if (debug)
9948                                 btrfs_warn(fs_info,
9949                                            "no space to alloc new chunk for block group %llu",
9950                                            block_group->key.objectid);
9951                         goto out;
9952                 }
9953
9954                 index = btrfs_bg_flags_to_raid_index(block_group->flags);
9955         }
9956
9957         if (index == BTRFS_RAID_RAID10) {
9958                 dev_min = 4;
9959                 /* Divide by 2 */
9960                 min_free >>= 1;
9961         } else if (index == BTRFS_RAID_RAID1) {
9962                 dev_min = 2;
9963         } else if (index == BTRFS_RAID_DUP) {
9964                 /* Multiply by 2 */
9965                 min_free <<= 1;
9966         } else if (index == BTRFS_RAID_RAID0) {
9967                 dev_min = fs_devices->rw_devices;
9968                 min_free = div64_u64(min_free, dev_min);
9969         }
9970
9971         /* We need to do this so that we can look at pending chunks */
9972         trans = btrfs_join_transaction(root);
9973         if (IS_ERR(trans)) {
9974                 ret = PTR_ERR(trans);
9975                 goto out;
9976         }
9977
9978         mutex_lock(&fs_info->chunk_mutex);
9979         list_for_each_entry(device, &fs_devices->alloc_list, dev_alloc_list) {
9980                 u64 dev_offset;
9981
9982                 /*
9983                  * check to make sure we can actually find a chunk with enough
9984                  * space to fit our block group in.
9985                  */
9986                 if (device->total_bytes > device->bytes_used + min_free &&
9987                     !test_bit(BTRFS_DEV_STATE_REPLACE_TGT, &device->dev_state)) {
9988                         ret = find_free_dev_extent(trans, device, min_free,
9989                                                    &dev_offset, NULL);
9990                         if (!ret)
9991                                 dev_nr++;
9992
9993                         if (dev_nr >= dev_min)
9994                                 break;
9995
9996                         ret = -1;
9997                 }
9998         }
9999         if (debug && ret == -1)
10000                 btrfs_warn(fs_info,
10001                            "no space to allocate a new chunk for block group %llu",
10002                            block_group->key.objectid);
10003         mutex_unlock(&fs_info->chunk_mutex);
10004         btrfs_end_transaction(trans);
10005 out:
10006         btrfs_put_block_group(block_group);
10007         return ret;
10008 }
10009
10010 static int find_first_block_group(struct btrfs_fs_info *fs_info,
10011                                   struct btrfs_path *path,
10012                                   struct btrfs_key *key)
10013 {
10014         struct btrfs_root *root = fs_info->extent_root;
10015         int ret = 0;
10016         struct btrfs_key found_key;
10017         struct extent_buffer *leaf;
10018         struct btrfs_block_group_item bg;
10019         u64 flags;
10020         int slot;
10021
10022         ret = btrfs_search_slot(NULL, root, key, path, 0, 0);
10023         if (ret < 0)
10024                 goto out;
10025
10026         while (1) {
10027                 slot = path->slots[0];
10028                 leaf = path->nodes[0];
10029                 if (slot >= btrfs_header_nritems(leaf)) {
10030                         ret = btrfs_next_leaf(root, path);
10031                         if (ret == 0)
10032                                 continue;
10033                         if (ret < 0)
10034                                 goto out;
10035                         break;
10036                 }
10037                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
10038
10039                 if (found_key.objectid >= key->objectid &&
10040                     found_key.type == BTRFS_BLOCK_GROUP_ITEM_KEY) {
10041                         struct extent_map_tree *em_tree;
10042                         struct extent_map *em;
10043
10044                         em_tree = &root->fs_info->mapping_tree.map_tree;
10045                         read_lock(&em_tree->lock);
10046                         em = lookup_extent_mapping(em_tree, found_key.objectid,
10047                                                    found_key.offset);
10048                         read_unlock(&em_tree->lock);
10049                         if (!em) {
10050                                 btrfs_err(fs_info,
10051                         "logical %llu len %llu found bg but no related chunk",
10052                                           found_key.objectid, found_key.offset);
10053                                 ret = -ENOENT;
10054                         } else if (em->start != found_key.objectid ||
10055                                    em->len != found_key.offset) {
10056                                 btrfs_err(fs_info,
10057                 "block group %llu len %llu mismatch with chunk %llu len %llu",
10058                                           found_key.objectid, found_key.offset,
10059                                           em->start, em->len);
10060                                 ret = -EUCLEAN;
10061                         } else {
10062                                 read_extent_buffer(leaf, &bg,
10063                                         btrfs_item_ptr_offset(leaf, slot),
10064                                         sizeof(bg));
10065                                 flags = btrfs_block_group_flags(&bg) &
10066                                         BTRFS_BLOCK_GROUP_TYPE_MASK;
10067
10068                                 if (flags != (em->map_lookup->type &
10069                                               BTRFS_BLOCK_GROUP_TYPE_MASK)) {
10070                                         btrfs_err(fs_info,
10071 "block group %llu len %llu type flags 0x%llx mismatch with chunk type flags 0x%llx",
10072                                                 found_key.objectid,
10073                                                 found_key.offset, flags,
10074                                                 (BTRFS_BLOCK_GROUP_TYPE_MASK &
10075                                                  em->map_lookup->type));
10076                                         ret = -EUCLEAN;
10077                                 } else {
10078                                         ret = 0;
10079                                 }
10080                         }
10081                         free_extent_map(em);
10082                         goto out;
10083                 }
10084                 path->slots[0]++;
10085         }
10086 out:
10087         return ret;
10088 }
10089
10090 void btrfs_put_block_group_cache(struct btrfs_fs_info *info)
10091 {
10092         struct btrfs_block_group_cache *block_group;
10093         u64 last = 0;
10094
10095         while (1) {
10096                 struct inode *inode;
10097
10098                 block_group = btrfs_lookup_first_block_group(info, last);
10099                 while (block_group) {
10100                         wait_block_group_cache_done(block_group);
10101                         spin_lock(&block_group->lock);
10102                         if (block_group->iref)
10103                                 break;
10104                         spin_unlock(&block_group->lock);
10105                         block_group = next_block_group(info, block_group);
10106                 }
10107                 if (!block_group) {
10108                         if (last == 0)
10109                                 break;
10110                         last = 0;
10111                         continue;
10112                 }
10113
10114                 inode = block_group->inode;
10115                 block_group->iref = 0;
10116                 block_group->inode = NULL;
10117                 spin_unlock(&block_group->lock);
10118                 ASSERT(block_group->io_ctl.inode == NULL);
10119                 iput(inode);
10120                 last = block_group->key.objectid + block_group->key.offset;
10121                 btrfs_put_block_group(block_group);
10122         }
10123 }
10124
10125 /*
10126  * Must be called only after stopping all workers, since we could have block
10127  * group caching kthreads running, and therefore they could race with us if we
10128  * freed the block groups before stopping them.
10129  */
10130 int btrfs_free_block_groups(struct btrfs_fs_info *info)
10131 {
10132         struct btrfs_block_group_cache *block_group;
10133         struct btrfs_space_info *space_info;
10134         struct btrfs_caching_control *caching_ctl;
10135         struct rb_node *n;
10136
10137         down_write(&info->commit_root_sem);
10138         while (!list_empty(&info->caching_block_groups)) {
10139                 caching_ctl = list_entry(info->caching_block_groups.next,
10140                                          struct btrfs_caching_control, list);
10141                 list_del(&caching_ctl->list);
10142                 put_caching_control(caching_ctl);
10143         }
10144         up_write(&info->commit_root_sem);
10145
10146         spin_lock(&info->unused_bgs_lock);
10147         while (!list_empty(&info->unused_bgs)) {
10148                 block_group = list_first_entry(&info->unused_bgs,
10149                                                struct btrfs_block_group_cache,
10150                                                bg_list);
10151                 list_del_init(&block_group->bg_list);
10152                 btrfs_put_block_group(block_group);
10153         }
10154         spin_unlock(&info->unused_bgs_lock);
10155
10156         spin_lock(&info->block_group_cache_lock);
10157         while ((n = rb_last(&info->block_group_cache_tree)) != NULL) {
10158                 block_group = rb_entry(n, struct btrfs_block_group_cache,
10159                                        cache_node);
10160                 rb_erase(&block_group->cache_node,
10161                          &info->block_group_cache_tree);
10162                 RB_CLEAR_NODE(&block_group->cache_node);
10163                 spin_unlock(&info->block_group_cache_lock);
10164
10165                 down_write(&block_group->space_info->groups_sem);
10166                 list_del(&block_group->list);
10167                 up_write(&block_group->space_info->groups_sem);
10168
10169                 /*
10170                  * We haven't cached this block group, which means we could
10171                  * possibly have excluded extents on this block group.
10172                  */
10173                 if (block_group->cached == BTRFS_CACHE_NO ||
10174                     block_group->cached == BTRFS_CACHE_ERROR)
10175                         free_excluded_extents(block_group);
10176
10177                 btrfs_remove_free_space_cache(block_group);
10178                 ASSERT(block_group->cached != BTRFS_CACHE_STARTED);
10179                 ASSERT(list_empty(&block_group->dirty_list));
10180                 ASSERT(list_empty(&block_group->io_list));
10181                 ASSERT(list_empty(&block_group->bg_list));
10182                 ASSERT(atomic_read(&block_group->count) == 1);
10183                 btrfs_put_block_group(block_group);
10184
10185                 spin_lock(&info->block_group_cache_lock);
10186         }
10187         spin_unlock(&info->block_group_cache_lock);
10188
10189         /* now that all the block groups are freed, go through and
10190          * free all the space_info structs.  This is only called during
10191          * the final stages of unmount, and so we know nobody is
10192          * using them.  We call synchronize_rcu() once before we start,
10193          * just to be on the safe side.
10194          */
10195         synchronize_rcu();
10196
10197         release_global_block_rsv(info);
10198
10199         while (!list_empty(&info->space_info)) {
10200                 int i;
10201
10202                 space_info = list_entry(info->space_info.next,
10203                                         struct btrfs_space_info,
10204                                         list);
10205
10206                 /*
10207                  * Do not hide this behind enospc_debug, this is actually
10208                  * important and indicates a real bug if this happens.
10209                  */
10210                 if (WARN_ON(space_info->bytes_pinned > 0 ||
10211                             space_info->bytes_reserved > 0 ||
10212                             space_info->bytes_may_use > 0))
10213                         dump_space_info(info, space_info, 0, 0);
10214                 list_del(&space_info->list);
10215                 for (i = 0; i < BTRFS_NR_RAID_TYPES; i++) {
10216                         struct kobject *kobj;
10217                         kobj = space_info->block_group_kobjs[i];
10218                         space_info->block_group_kobjs[i] = NULL;
10219                         if (kobj) {
10220                                 kobject_del(kobj);
10221                                 kobject_put(kobj);
10222                         }
10223                 }
10224                 kobject_del(&space_info->kobj);
10225                 kobject_put(&space_info->kobj);
10226         }
10227         return 0;
10228 }
10229
10230 /* link_block_group will queue up kobjects to add when we're reclaim-safe */
10231 void btrfs_add_raid_kobjects(struct btrfs_fs_info *fs_info)
10232 {
10233         struct btrfs_space_info *space_info;
10234         struct raid_kobject *rkobj;
10235         LIST_HEAD(list);
10236         int index;
10237         int ret = 0;
10238
10239         spin_lock(&fs_info->pending_raid_kobjs_lock);
10240         list_splice_init(&fs_info->pending_raid_kobjs, &list);
10241         spin_unlock(&fs_info->pending_raid_kobjs_lock);
10242
10243         list_for_each_entry(rkobj, &list, list) {
10244                 space_info = __find_space_info(fs_info, rkobj->flags);
10245                 index = btrfs_bg_flags_to_raid_index(rkobj->flags);
10246
10247                 ret = kobject_add(&rkobj->kobj, &space_info->kobj,
10248                                   "%s", get_raid_name(index));
10249                 if (ret) {
10250                         kobject_put(&rkobj->kobj);
10251                         break;
10252                 }
10253         }
10254         if (ret)
10255                 btrfs_warn(fs_info,
10256                            "failed to add kobject for block cache, ignoring");
10257 }
10258
10259 static void link_block_group(struct btrfs_block_group_cache *cache)
10260 {
10261         struct btrfs_space_info *space_info = cache->space_info;
10262         struct btrfs_fs_info *fs_info = cache->fs_info;
10263         int index = btrfs_bg_flags_to_raid_index(cache->flags);
10264         bool first = false;
10265
10266         down_write(&space_info->groups_sem);
10267         if (list_empty(&space_info->block_groups[index]))
10268                 first = true;
10269         list_add_tail(&cache->list, &space_info->block_groups[index]);
10270         up_write(&space_info->groups_sem);
10271
10272         if (first) {
10273                 struct raid_kobject *rkobj = kzalloc(sizeof(*rkobj), GFP_NOFS);
10274                 if (!rkobj) {
10275                         btrfs_warn(cache->fs_info,
10276                                 "couldn't alloc memory for raid level kobject");
10277                         return;
10278                 }
10279                 rkobj->flags = cache->flags;
10280                 kobject_init(&rkobj->kobj, &btrfs_raid_ktype);
10281
10282                 spin_lock(&fs_info->pending_raid_kobjs_lock);
10283                 list_add_tail(&rkobj->list, &fs_info->pending_raid_kobjs);
10284                 spin_unlock(&fs_info->pending_raid_kobjs_lock);
10285                 space_info->block_group_kobjs[index] = &rkobj->kobj;
10286         }
10287 }
10288
10289 static struct btrfs_block_group_cache *
10290 btrfs_create_block_group_cache(struct btrfs_fs_info *fs_info,
10291                                u64 start, u64 size)
10292 {
10293         struct btrfs_block_group_cache *cache;
10294
10295         cache = kzalloc(sizeof(*cache), GFP_NOFS);
10296         if (!cache)
10297                 return NULL;
10298
10299         cache->free_space_ctl = kzalloc(sizeof(*cache->free_space_ctl),
10300                                         GFP_NOFS);
10301         if (!cache->free_space_ctl) {
10302                 kfree(cache);
10303                 return NULL;
10304         }
10305
10306         cache->key.objectid = start;
10307         cache->key.offset = size;
10308         cache->key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
10309
10310         cache->fs_info = fs_info;
10311         cache->full_stripe_len = btrfs_full_stripe_len(fs_info, start);
10312         set_free_space_tree_thresholds(cache);
10313
10314         atomic_set(&cache->count, 1);
10315         spin_lock_init(&cache->lock);
10316         init_rwsem(&cache->data_rwsem);
10317         INIT_LIST_HEAD(&cache->list);
10318         INIT_LIST_HEAD(&cache->cluster_list);
10319         INIT_LIST_HEAD(&cache->bg_list);
10320         INIT_LIST_HEAD(&cache->ro_list);
10321         INIT_LIST_HEAD(&cache->dirty_list);
10322         INIT_LIST_HEAD(&cache->io_list);
10323         btrfs_init_free_space_ctl(cache);
10324         atomic_set(&cache->trimming, 0);
10325         mutex_init(&cache->free_space_lock);
10326         btrfs_init_full_stripe_locks_tree(&cache->full_stripe_locks_root);
10327
10328         return cache;
10329 }
10330
10331
10332 /*
10333  * Iterate all chunks and verify that each of them has the corresponding block
10334  * group
10335  */
10336 static int check_chunk_block_group_mappings(struct btrfs_fs_info *fs_info)
10337 {
10338         struct btrfs_mapping_tree *map_tree = &fs_info->mapping_tree;
10339         struct extent_map *em;
10340         struct btrfs_block_group_cache *bg;
10341         u64 start = 0;
10342         int ret = 0;
10343
10344         while (1) {
10345                 read_lock(&map_tree->map_tree.lock);
10346                 /*
10347                  * lookup_extent_mapping will return the first extent map
10348                  * intersecting the range, so setting @len to 1 is enough to
10349                  * get the first chunk.
10350                  */
10351                 em = lookup_extent_mapping(&map_tree->map_tree, start, 1);
10352                 read_unlock(&map_tree->map_tree.lock);
10353                 if (!em)
10354                         break;
10355
10356                 bg = btrfs_lookup_block_group(fs_info, em->start);
10357                 if (!bg) {
10358                         btrfs_err(fs_info,
10359         "chunk start=%llu len=%llu doesn't have corresponding block group",
10360                                      em->start, em->len);
10361                         ret = -EUCLEAN;
10362                         free_extent_map(em);
10363                         break;
10364                 }
10365                 if (bg->key.objectid != em->start ||
10366                     bg->key.offset != em->len ||
10367                     (bg->flags & BTRFS_BLOCK_GROUP_TYPE_MASK) !=
10368                     (em->map_lookup->type & BTRFS_BLOCK_GROUP_TYPE_MASK)) {
10369                         btrfs_err(fs_info,
10370 "chunk start=%llu len=%llu flags=0x%llx doesn't match block group start=%llu len=%llu flags=0x%llx",
10371                                 em->start, em->len,
10372                                 em->map_lookup->type & BTRFS_BLOCK_GROUP_TYPE_MASK,
10373                                 bg->key.objectid, bg->key.offset,
10374                                 bg->flags & BTRFS_BLOCK_GROUP_TYPE_MASK);
10375                         ret = -EUCLEAN;
10376                         free_extent_map(em);
10377                         btrfs_put_block_group(bg);
10378                         break;
10379                 }
10380                 start = em->start + em->len;
10381                 free_extent_map(em);
10382                 btrfs_put_block_group(bg);
10383         }
10384         return ret;
10385 }
10386
10387 int btrfs_read_block_groups(struct btrfs_fs_info *info)
10388 {
10389         struct btrfs_path *path;
10390         int ret;
10391         struct btrfs_block_group_cache *cache;
10392         struct btrfs_space_info *space_info;
10393         struct btrfs_key key;
10394         struct btrfs_key found_key;
10395         struct extent_buffer *leaf;
10396         int need_clear = 0;
10397         u64 cache_gen;
10398         u64 feature;
10399         int mixed;
10400
10401         feature = btrfs_super_incompat_flags(info->super_copy);
10402         mixed = !!(feature & BTRFS_FEATURE_INCOMPAT_MIXED_GROUPS);
10403
10404         key.objectid = 0;
10405         key.offset = 0;
10406         key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
10407         path = btrfs_alloc_path();
10408         if (!path)
10409                 return -ENOMEM;
10410         path->reada = READA_FORWARD;
10411
10412         cache_gen = btrfs_super_cache_generation(info->super_copy);
10413         if (btrfs_test_opt(info, SPACE_CACHE) &&
10414             btrfs_super_generation(info->super_copy) != cache_gen)
10415                 need_clear = 1;
10416         if (btrfs_test_opt(info, CLEAR_CACHE))
10417                 need_clear = 1;
10418
10419         while (1) {
10420                 ret = find_first_block_group(info, path, &key);
10421                 if (ret > 0)
10422                         break;
10423                 if (ret != 0)
10424                         goto error;
10425
10426                 leaf = path->nodes[0];
10427                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
10428
10429                 cache = btrfs_create_block_group_cache(info, found_key.objectid,
10430                                                        found_key.offset);
10431                 if (!cache) {
10432                         ret = -ENOMEM;
10433                         goto error;
10434                 }
10435
10436                 if (need_clear) {
10437                         /*
10438                          * When we mount with old space cache, we need to
10439                          * set BTRFS_DC_CLEAR and set dirty flag.
10440                          *
10441                          * a) Setting 'BTRFS_DC_CLEAR' makes sure that we
10442                          *    truncate the old free space cache inode and
10443                          *    setup a new one.
10444                          * b) Setting 'dirty flag' makes sure that we flush
10445                          *    the new space cache info onto disk.
10446                          */
10447                         if (btrfs_test_opt(info, SPACE_CACHE))
10448                                 cache->disk_cache_state = BTRFS_DC_CLEAR;
10449                 }
10450
10451                 read_extent_buffer(leaf, &cache->item,
10452                                    btrfs_item_ptr_offset(leaf, path->slots[0]),
10453                                    sizeof(cache->item));
10454                 cache->flags = btrfs_block_group_flags(&cache->item);
10455                 if (!mixed &&
10456                     ((cache->flags & BTRFS_BLOCK_GROUP_METADATA) &&
10457                     (cache->flags & BTRFS_BLOCK_GROUP_DATA))) {
10458                         btrfs_err(info,
10459 "bg %llu is a mixed block group but filesystem hasn't enabled mixed block groups",
10460                                   cache->key.objectid);
10461                         ret = -EINVAL;
10462                         goto error;
10463                 }
10464
10465                 key.objectid = found_key.objectid + found_key.offset;
10466                 btrfs_release_path(path);
10467
10468                 /*
10469                  * We need to exclude the super stripes now so that the space
10470                  * info has super bytes accounted for, otherwise we'll think
10471                  * we have more space than we actually do.
10472                  */
10473                 ret = exclude_super_stripes(cache);
10474                 if (ret) {
10475                         /*
10476                          * We may have excluded something, so call this just in
10477                          * case.
10478                          */
10479                         free_excluded_extents(cache);
10480                         btrfs_put_block_group(cache);
10481                         goto error;
10482                 }
10483
10484                 /*
10485                  * check for two cases, either we are full, and therefore
10486                  * don't need to bother with the caching work since we won't
10487                  * find any space, or we are empty, and we can just add all
10488                  * the space in and be done with it.  This saves us _a_lot_ of
10489                  * time, particularly in the full case.
10490                  */
10491                 if (found_key.offset == btrfs_block_group_used(&cache->item)) {
10492                         cache->last_byte_to_unpin = (u64)-1;
10493                         cache->cached = BTRFS_CACHE_FINISHED;
10494                         free_excluded_extents(cache);
10495                 } else if (btrfs_block_group_used(&cache->item) == 0) {
10496                         cache->last_byte_to_unpin = (u64)-1;
10497                         cache->cached = BTRFS_CACHE_FINISHED;
10498                         add_new_free_space(cache, found_key.objectid,
10499                                            found_key.objectid +
10500                                            found_key.offset);
10501                         free_excluded_extents(cache);
10502                 }
10503
10504                 ret = btrfs_add_block_group_cache(info, cache);
10505                 if (ret) {
10506                         btrfs_remove_free_space_cache(cache);
10507                         btrfs_put_block_group(cache);
10508                         goto error;
10509                 }
10510
10511                 trace_btrfs_add_block_group(info, cache, 0);
10512                 update_space_info(info, cache->flags, found_key.offset,
10513                                   btrfs_block_group_used(&cache->item),
10514                                   cache->bytes_super, &space_info);
10515
10516                 cache->space_info = space_info;
10517
10518                 link_block_group(cache);
10519
10520                 set_avail_alloc_bits(info, cache->flags);
10521                 if (btrfs_chunk_readonly(info, cache->key.objectid)) {
10522                         inc_block_group_ro(cache, 1);
10523                 } else if (btrfs_block_group_used(&cache->item) == 0) {
10524                         ASSERT(list_empty(&cache->bg_list));
10525                         btrfs_mark_bg_unused(cache);
10526                 }
10527         }
10528
10529         list_for_each_entry_rcu(space_info, &info->space_info, list) {
10530                 if (!(get_alloc_profile(info, space_info->flags) &
10531                       (BTRFS_BLOCK_GROUP_RAID10 |
10532                        BTRFS_BLOCK_GROUP_RAID1 |
10533                        BTRFS_BLOCK_GROUP_RAID5 |
10534                        BTRFS_BLOCK_GROUP_RAID6 |
10535                        BTRFS_BLOCK_GROUP_DUP)))
10536                         continue;
10537                 /*
10538                  * avoid allocating from un-mirrored block group if there are
10539                  * mirrored block groups.
10540                  */
10541                 list_for_each_entry(cache,
10542                                 &space_info->block_groups[BTRFS_RAID_RAID0],
10543                                 list)
10544                         inc_block_group_ro(cache, 1);
10545                 list_for_each_entry(cache,
10546                                 &space_info->block_groups[BTRFS_RAID_SINGLE],
10547                                 list)
10548                         inc_block_group_ro(cache, 1);
10549         }
10550
10551         btrfs_add_raid_kobjects(info);
10552         init_global_block_rsv(info);
10553         ret = check_chunk_block_group_mappings(info);
10554 error:
10555         btrfs_free_path(path);
10556         return ret;
10557 }
10558
10559 void btrfs_create_pending_block_groups(struct btrfs_trans_handle *trans)
10560 {
10561         struct btrfs_fs_info *fs_info = trans->fs_info;
10562         struct btrfs_block_group_cache *block_group;
10563         struct btrfs_root *extent_root = fs_info->extent_root;
10564         struct btrfs_block_group_item item;
10565         struct btrfs_key key;
10566         int ret = 0;
10567
10568         if (!trans->can_flush_pending_bgs)
10569                 return;
10570
10571         while (!list_empty(&trans->new_bgs)) {
10572                 block_group = list_first_entry(&trans->new_bgs,
10573                                                struct btrfs_block_group_cache,
10574                                                bg_list);
10575                 if (ret)
10576                         goto next;
10577
10578                 spin_lock(&block_group->lock);
10579                 memcpy(&item, &block_group->item, sizeof(item));
10580                 memcpy(&key, &block_group->key, sizeof(key));
10581                 spin_unlock(&block_group->lock);
10582
10583                 ret = btrfs_insert_item(trans, extent_root, &key, &item,
10584                                         sizeof(item));
10585                 if (ret)
10586                         btrfs_abort_transaction(trans, ret);
10587                 ret = btrfs_finish_chunk_alloc(trans, key.objectid, key.offset);
10588                 if (ret)
10589                         btrfs_abort_transaction(trans, ret);
10590                 add_block_group_free_space(trans, block_group);
10591                 /* already aborted the transaction if it failed. */
10592 next:
10593                 btrfs_delayed_refs_rsv_release(fs_info, 1);
10594                 list_del_init(&block_group->bg_list);
10595         }
10596         btrfs_trans_release_chunk_metadata(trans);
10597 }
10598
10599 int btrfs_make_block_group(struct btrfs_trans_handle *trans, u64 bytes_used,
10600                            u64 type, u64 chunk_offset, u64 size)
10601 {
10602         struct btrfs_fs_info *fs_info = trans->fs_info;
10603         struct btrfs_block_group_cache *cache;
10604         int ret;
10605
10606         btrfs_set_log_full_commit(fs_info, trans);
10607
10608         cache = btrfs_create_block_group_cache(fs_info, chunk_offset, size);
10609         if (!cache)
10610                 return -ENOMEM;
10611
10612         btrfs_set_block_group_used(&cache->item, bytes_used);
10613         btrfs_set_block_group_chunk_objectid(&cache->item,
10614                                              BTRFS_FIRST_CHUNK_TREE_OBJECTID);
10615         btrfs_set_block_group_flags(&cache->item, type);
10616
10617         cache->flags = type;
10618         cache->last_byte_to_unpin = (u64)-1;
10619         cache->cached = BTRFS_CACHE_FINISHED;
10620         cache->needs_free_space = 1;
10621         ret = exclude_super_stripes(cache);
10622         if (ret) {
10623                 /*
10624                  * We may have excluded something, so call this just in
10625                  * case.
10626                  */
10627                 free_excluded_extents(cache);
10628                 btrfs_put_block_group(cache);
10629                 return ret;
10630         }
10631
10632         add_new_free_space(cache, chunk_offset, chunk_offset + size);
10633
10634         free_excluded_extents(cache);
10635
10636 #ifdef CONFIG_BTRFS_DEBUG
10637         if (btrfs_should_fragment_free_space(cache)) {
10638                 u64 new_bytes_used = size - bytes_used;
10639
10640                 bytes_used += new_bytes_used >> 1;
10641                 fragment_free_space(cache);
10642         }
10643 #endif
10644         /*
10645          * Ensure the corresponding space_info object is created and
10646          * assigned to our block group. We want our bg to be added to the rbtree
10647          * with its ->space_info set.
10648          */
10649         cache->space_info = __find_space_info(fs_info, cache->flags);
10650         ASSERT(cache->space_info);
10651
10652         ret = btrfs_add_block_group_cache(fs_info, cache);
10653         if (ret) {
10654                 btrfs_remove_free_space_cache(cache);
10655                 btrfs_put_block_group(cache);
10656                 return ret;
10657         }
10658
10659         /*
10660          * Now that our block group has its ->space_info set and is inserted in
10661          * the rbtree, update the space info's counters.
10662          */
10663         trace_btrfs_add_block_group(fs_info, cache, 1);
10664         update_space_info(fs_info, cache->flags, size, bytes_used,
10665                                 cache->bytes_super, &cache->space_info);
10666         update_global_block_rsv(fs_info);
10667
10668         link_block_group(cache);
10669
10670         list_add_tail(&cache->bg_list, &trans->new_bgs);
10671         trans->delayed_ref_updates++;
10672         btrfs_update_delayed_refs_rsv(trans);
10673
10674         set_avail_alloc_bits(fs_info, type);
10675         return 0;
10676 }
10677
10678 static void clear_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
10679 {
10680         u64 extra_flags = chunk_to_extended(flags) &
10681                                 BTRFS_EXTENDED_PROFILE_MASK;
10682
10683         write_seqlock(&fs_info->profiles_lock);
10684         if (flags & BTRFS_BLOCK_GROUP_DATA)
10685                 fs_info->avail_data_alloc_bits &= ~extra_flags;
10686         if (flags & BTRFS_BLOCK_GROUP_METADATA)
10687                 fs_info->avail_metadata_alloc_bits &= ~extra_flags;
10688         if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
10689                 fs_info->avail_system_alloc_bits &= ~extra_flags;
10690         write_sequnlock(&fs_info->profiles_lock);
10691 }
10692
10693 int btrfs_remove_block_group(struct btrfs_trans_handle *trans,
10694                              u64 group_start, struct extent_map *em)
10695 {
10696         struct btrfs_fs_info *fs_info = trans->fs_info;
10697         struct btrfs_root *root = fs_info->extent_root;
10698         struct btrfs_path *path;
10699         struct btrfs_block_group_cache *block_group;
10700         struct btrfs_free_cluster *cluster;
10701         struct btrfs_root *tree_root = fs_info->tree_root;
10702         struct btrfs_key key;
10703         struct inode *inode;
10704         struct kobject *kobj = NULL;
10705         int ret;
10706         int index;
10707         int factor;
10708         struct btrfs_caching_control *caching_ctl = NULL;
10709         bool remove_em;
10710         bool remove_rsv = false;
10711
10712         block_group = btrfs_lookup_block_group(fs_info, group_start);
10713         BUG_ON(!block_group);
10714         BUG_ON(!block_group->ro);
10715
10716         trace_btrfs_remove_block_group(block_group);
10717         /*
10718          * Free the reserved super bytes from this block group before
10719          * remove it.
10720          */
10721         free_excluded_extents(block_group);
10722         btrfs_free_ref_tree_range(fs_info, block_group->key.objectid,
10723                                   block_group->key.offset);
10724
10725         memcpy(&key, &block_group->key, sizeof(key));
10726         index = btrfs_bg_flags_to_raid_index(block_group->flags);
10727         factor = btrfs_bg_type_to_factor(block_group->flags);
10728
10729         /* make sure this block group isn't part of an allocation cluster */
10730         cluster = &fs_info->data_alloc_cluster;
10731         spin_lock(&cluster->refill_lock);
10732         btrfs_return_cluster_to_free_space(block_group, cluster);
10733         spin_unlock(&cluster->refill_lock);
10734
10735         /*
10736          * make sure this block group isn't part of a metadata
10737          * allocation cluster
10738          */
10739         cluster = &fs_info->meta_alloc_cluster;
10740         spin_lock(&cluster->refill_lock);
10741         btrfs_return_cluster_to_free_space(block_group, cluster);
10742         spin_unlock(&cluster->refill_lock);
10743
10744         path = btrfs_alloc_path();
10745         if (!path) {
10746                 ret = -ENOMEM;
10747                 goto out;
10748         }
10749
10750         /*
10751          * get the inode first so any iput calls done for the io_list
10752          * aren't the final iput (no unlinks allowed now)
10753          */
10754         inode = lookup_free_space_inode(fs_info, block_group, path);
10755
10756         mutex_lock(&trans->transaction->cache_write_mutex);
10757         /*
10758          * Make sure our free space cache IO is done before removing the
10759          * free space inode
10760          */
10761         spin_lock(&trans->transaction->dirty_bgs_lock);
10762         if (!list_empty(&block_group->io_list)) {
10763                 list_del_init(&block_group->io_list);
10764
10765                 WARN_ON(!IS_ERR(inode) && inode != block_group->io_ctl.inode);
10766
10767                 spin_unlock(&trans->transaction->dirty_bgs_lock);
10768                 btrfs_wait_cache_io(trans, block_group, path);
10769                 btrfs_put_block_group(block_group);
10770                 spin_lock(&trans->transaction->dirty_bgs_lock);
10771         }
10772
10773         if (!list_empty(&block_group->dirty_list)) {
10774                 list_del_init(&block_group->dirty_list);
10775                 remove_rsv = true;
10776                 btrfs_put_block_group(block_group);
10777         }
10778         spin_unlock(&trans->transaction->dirty_bgs_lock);
10779         mutex_unlock(&trans->transaction->cache_write_mutex);
10780
10781         if (!IS_ERR(inode)) {
10782                 ret = btrfs_orphan_add(trans, BTRFS_I(inode));
10783                 if (ret) {
10784                         btrfs_add_delayed_iput(inode);
10785                         goto out;
10786                 }
10787                 clear_nlink(inode);
10788                 /* One for the block groups ref */
10789                 spin_lock(&block_group->lock);
10790                 if (block_group->iref) {
10791                         block_group->iref = 0;
10792                         block_group->inode = NULL;
10793                         spin_unlock(&block_group->lock);
10794                         iput(inode);
10795                 } else {
10796                         spin_unlock(&block_group->lock);
10797                 }
10798                 /* One for our lookup ref */
10799                 btrfs_add_delayed_iput(inode);
10800         }
10801
10802         key.objectid = BTRFS_FREE_SPACE_OBJECTID;
10803         key.offset = block_group->key.objectid;
10804         key.type = 0;
10805
10806         ret = btrfs_search_slot(trans, tree_root, &key, path, -1, 1);
10807         if (ret < 0)
10808                 goto out;
10809         if (ret > 0)
10810                 btrfs_release_path(path);
10811         if (ret == 0) {
10812                 ret = btrfs_del_item(trans, tree_root, path);
10813                 if (ret)
10814                         goto out;
10815                 btrfs_release_path(path);
10816         }
10817
10818         spin_lock(&fs_info->block_group_cache_lock);
10819         rb_erase(&block_group->cache_node,
10820                  &fs_info->block_group_cache_tree);
10821         RB_CLEAR_NODE(&block_group->cache_node);
10822
10823         if (fs_info->first_logical_byte == block_group->key.objectid)
10824                 fs_info->first_logical_byte = (u64)-1;
10825         spin_unlock(&fs_info->block_group_cache_lock);
10826
10827         down_write(&block_group->space_info->groups_sem);
10828         /*
10829          * we must use list_del_init so people can check to see if they
10830          * are still on the list after taking the semaphore
10831          */
10832         list_del_init(&block_group->list);
10833         if (list_empty(&block_group->space_info->block_groups[index])) {
10834                 kobj = block_group->space_info->block_group_kobjs[index];
10835                 block_group->space_info->block_group_kobjs[index] = NULL;
10836                 clear_avail_alloc_bits(fs_info, block_group->flags);
10837         }
10838         up_write(&block_group->space_info->groups_sem);
10839         if (kobj) {
10840                 kobject_del(kobj);
10841                 kobject_put(kobj);
10842         }
10843
10844         if (block_group->has_caching_ctl)
10845                 caching_ctl = get_caching_control(block_group);
10846         if (block_group->cached == BTRFS_CACHE_STARTED)
10847                 wait_block_group_cache_done(block_group);
10848         if (block_group->has_caching_ctl) {
10849                 down_write(&fs_info->commit_root_sem);
10850                 if (!caching_ctl) {
10851                         struct btrfs_caching_control *ctl;
10852
10853                         list_for_each_entry(ctl,
10854                                     &fs_info->caching_block_groups, list)
10855                                 if (ctl->block_group == block_group) {
10856                                         caching_ctl = ctl;
10857                                         refcount_inc(&caching_ctl->count);
10858                                         break;
10859                                 }
10860                 }
10861                 if (caching_ctl)
10862                         list_del_init(&caching_ctl->list);
10863                 up_write(&fs_info->commit_root_sem);
10864                 if (caching_ctl) {
10865                         /* Once for the caching bgs list and once for us. */
10866                         put_caching_control(caching_ctl);
10867                         put_caching_control(caching_ctl);
10868                 }
10869         }
10870
10871         spin_lock(&trans->transaction->dirty_bgs_lock);
10872         if (!list_empty(&block_group->dirty_list)) {
10873                 WARN_ON(1);
10874         }
10875         if (!list_empty(&block_group->io_list)) {
10876                 WARN_ON(1);
10877         }
10878         spin_unlock(&trans->transaction->dirty_bgs_lock);
10879         btrfs_remove_free_space_cache(block_group);
10880
10881         spin_lock(&block_group->space_info->lock);
10882         list_del_init(&block_group->ro_list);
10883
10884         if (btrfs_test_opt(fs_info, ENOSPC_DEBUG)) {
10885                 WARN_ON(block_group->space_info->total_bytes
10886                         < block_group->key.offset);
10887                 WARN_ON(block_group->space_info->bytes_readonly
10888                         < block_group->key.offset);
10889                 WARN_ON(block_group->space_info->disk_total
10890                         < block_group->key.offset * factor);
10891         }
10892         block_group->space_info->total_bytes -= block_group->key.offset;
10893         block_group->space_info->bytes_readonly -= block_group->key.offset;
10894         block_group->space_info->disk_total -= block_group->key.offset * factor;
10895
10896         spin_unlock(&block_group->space_info->lock);
10897
10898         memcpy(&key, &block_group->key, sizeof(key));
10899
10900         mutex_lock(&fs_info->chunk_mutex);
10901         if (!list_empty(&em->list)) {
10902                 /* We're in the transaction->pending_chunks list. */
10903                 free_extent_map(em);
10904         }
10905         spin_lock(&block_group->lock);
10906         block_group->removed = 1;
10907         /*
10908          * At this point trimming can't start on this block group, because we
10909          * removed the block group from the tree fs_info->block_group_cache_tree
10910          * so no one can't find it anymore and even if someone already got this
10911          * block group before we removed it from the rbtree, they have already
10912          * incremented block_group->trimming - if they didn't, they won't find
10913          * any free space entries because we already removed them all when we
10914          * called btrfs_remove_free_space_cache().
10915          *
10916          * And we must not remove the extent map from the fs_info->mapping_tree
10917          * to prevent the same logical address range and physical device space
10918          * ranges from being reused for a new block group. This is because our
10919          * fs trim operation (btrfs_trim_fs() / btrfs_ioctl_fitrim()) is
10920          * completely transactionless, so while it is trimming a range the
10921          * currently running transaction might finish and a new one start,
10922          * allowing for new block groups to be created that can reuse the same
10923          * physical device locations unless we take this special care.
10924          *
10925          * There may also be an implicit trim operation if the file system
10926          * is mounted with -odiscard. The same protections must remain
10927          * in place until the extents have been discarded completely when
10928          * the transaction commit has completed.
10929          */
10930         remove_em = (atomic_read(&block_group->trimming) == 0);
10931         /*
10932          * Make sure a trimmer task always sees the em in the pinned_chunks list
10933          * if it sees block_group->removed == 1 (needs to lock block_group->lock
10934          * before checking block_group->removed).
10935          */
10936         if (!remove_em) {
10937                 /*
10938                  * Our em might be in trans->transaction->pending_chunks which
10939                  * is protected by fs_info->chunk_mutex ([lock|unlock]_chunks),
10940                  * and so is the fs_info->pinned_chunks list.
10941                  *
10942                  * So at this point we must be holding the chunk_mutex to avoid
10943                  * any races with chunk allocation (more specifically at
10944                  * volumes.c:contains_pending_extent()), to ensure it always
10945                  * sees the em, either in the pending_chunks list or in the
10946                  * pinned_chunks list.
10947                  */
10948                 list_move_tail(&em->list, &fs_info->pinned_chunks);
10949         }
10950         spin_unlock(&block_group->lock);
10951
10952         if (remove_em) {
10953                 struct extent_map_tree *em_tree;
10954
10955                 em_tree = &fs_info->mapping_tree.map_tree;
10956                 write_lock(&em_tree->lock);
10957                 /*
10958                  * The em might be in the pending_chunks list, so make sure the
10959                  * chunk mutex is locked, since remove_extent_mapping() will
10960                  * delete us from that list.
10961                  */
10962                 remove_extent_mapping(em_tree, em);
10963                 write_unlock(&em_tree->lock);
10964                 /* once for the tree */
10965                 free_extent_map(em);
10966         }
10967
10968         mutex_unlock(&fs_info->chunk_mutex);
10969
10970         ret = remove_block_group_free_space(trans, block_group);
10971         if (ret)
10972                 goto out;
10973
10974         btrfs_put_block_group(block_group);
10975         btrfs_put_block_group(block_group);
10976
10977         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
10978         if (ret > 0)
10979                 ret = -EIO;
10980         if (ret < 0)
10981                 goto out;
10982
10983         ret = btrfs_del_item(trans, root, path);
10984 out:
10985         if (remove_rsv)
10986                 btrfs_delayed_refs_rsv_release(fs_info, 1);
10987         btrfs_free_path(path);
10988         return ret;
10989 }
10990
10991 struct btrfs_trans_handle *
10992 btrfs_start_trans_remove_block_group(struct btrfs_fs_info *fs_info,
10993                                      const u64 chunk_offset)
10994 {
10995         struct extent_map_tree *em_tree = &fs_info->mapping_tree.map_tree;
10996         struct extent_map *em;
10997         struct map_lookup *map;
10998         unsigned int num_items;
10999
11000         read_lock(&em_tree->lock);
11001         em = lookup_extent_mapping(em_tree, chunk_offset, 1);
11002         read_unlock(&em_tree->lock);
11003         ASSERT(em && em->start == chunk_offset);
11004
11005         /*
11006          * We need to reserve 3 + N units from the metadata space info in order
11007          * to remove a block group (done at btrfs_remove_chunk() and at
11008          * btrfs_remove_block_group()), which are used for:
11009          *
11010          * 1 unit for adding the free space inode's orphan (located in the tree
11011          * of tree roots).
11012          * 1 unit for deleting the block group item (located in the extent
11013          * tree).
11014          * 1 unit for deleting the free space item (located in tree of tree
11015          * roots).
11016          * N units for deleting N device extent items corresponding to each
11017          * stripe (located in the device tree).
11018          *
11019          * In order to remove a block group we also need to reserve units in the
11020          * system space info in order to update the chunk tree (update one or
11021          * more device items and remove one chunk item), but this is done at
11022          * btrfs_remove_chunk() through a call to check_system_chunk().
11023          */
11024         map = em->map_lookup;
11025         num_items = 3 + map->num_stripes;
11026         free_extent_map(em);
11027
11028         return btrfs_start_transaction_fallback_global_rsv(fs_info->extent_root,
11029                                                            num_items, 1);
11030 }
11031
11032 /*
11033  * Process the unused_bgs list and remove any that don't have any allocated
11034  * space inside of them.
11035  */
11036 void btrfs_delete_unused_bgs(struct btrfs_fs_info *fs_info)
11037 {
11038         struct btrfs_block_group_cache *block_group;
11039         struct btrfs_space_info *space_info;
11040         struct btrfs_trans_handle *trans;
11041         int ret = 0;
11042
11043         if (!test_bit(BTRFS_FS_OPEN, &fs_info->flags))
11044                 return;
11045
11046         spin_lock(&fs_info->unused_bgs_lock);
11047         while (!list_empty(&fs_info->unused_bgs)) {
11048                 u64 start, end;
11049                 int trimming;
11050
11051                 block_group = list_first_entry(&fs_info->unused_bgs,
11052                                                struct btrfs_block_group_cache,
11053                                                bg_list);
11054                 list_del_init(&block_group->bg_list);
11055
11056                 space_info = block_group->space_info;
11057
11058                 if (ret || btrfs_mixed_space_info(space_info)) {
11059                         btrfs_put_block_group(block_group);
11060                         continue;
11061                 }
11062                 spin_unlock(&fs_info->unused_bgs_lock);
11063
11064                 mutex_lock(&fs_info->delete_unused_bgs_mutex);
11065
11066                 /* Don't want to race with allocators so take the groups_sem */
11067                 down_write(&space_info->groups_sem);
11068                 spin_lock(&block_group->lock);
11069                 if (block_group->reserved || block_group->pinned ||
11070                     btrfs_block_group_used(&block_group->item) ||
11071                     block_group->ro ||
11072                     list_is_singular(&block_group->list)) {
11073                         /*
11074                          * We want to bail if we made new allocations or have
11075                          * outstanding allocations in this block group.  We do
11076                          * the ro check in case balance is currently acting on
11077                          * this block group.
11078                          */
11079                         trace_btrfs_skip_unused_block_group(block_group);
11080                         spin_unlock(&block_group->lock);
11081                         up_write(&space_info->groups_sem);
11082                         goto next;
11083                 }
11084                 spin_unlock(&block_group->lock);
11085
11086                 /* We don't want to force the issue, only flip if it's ok. */
11087                 ret = inc_block_group_ro(block_group, 0);
11088                 up_write(&space_info->groups_sem);
11089                 if (ret < 0) {
11090                         ret = 0;
11091                         goto next;
11092                 }
11093
11094                 /*
11095                  * Want to do this before we do anything else so we can recover
11096                  * properly if we fail to join the transaction.
11097                  */
11098                 trans = btrfs_start_trans_remove_block_group(fs_info,
11099                                                      block_group->key.objectid);
11100                 if (IS_ERR(trans)) {
11101                         btrfs_dec_block_group_ro(block_group);
11102                         ret = PTR_ERR(trans);
11103                         goto next;
11104                 }
11105
11106                 /*
11107                  * We could have pending pinned extents for this block group,
11108                  * just delete them, we don't care about them anymore.
11109                  */
11110                 start = block_group->key.objectid;
11111                 end = start + block_group->key.offset - 1;
11112                 /*
11113                  * Hold the unused_bg_unpin_mutex lock to avoid racing with
11114                  * btrfs_finish_extent_commit(). If we are at transaction N,
11115                  * another task might be running finish_extent_commit() for the
11116                  * previous transaction N - 1, and have seen a range belonging
11117                  * to the block group in freed_extents[] before we were able to
11118                  * clear the whole block group range from freed_extents[]. This
11119                  * means that task can lookup for the block group after we
11120                  * unpinned it from freed_extents[] and removed it, leading to
11121                  * a BUG_ON() at btrfs_unpin_extent_range().
11122                  */
11123                 mutex_lock(&fs_info->unused_bg_unpin_mutex);
11124                 ret = clear_extent_bits(&fs_info->freed_extents[0], start, end,
11125                                   EXTENT_DIRTY);
11126                 if (ret) {
11127                         mutex_unlock(&fs_info->unused_bg_unpin_mutex);
11128                         btrfs_dec_block_group_ro(block_group);
11129                         goto end_trans;
11130                 }
11131                 ret = clear_extent_bits(&fs_info->freed_extents[1], start, end,
11132                                   EXTENT_DIRTY);
11133                 if (ret) {
11134                         mutex_unlock(&fs_info->unused_bg_unpin_mutex);
11135                         btrfs_dec_block_group_ro(block_group);
11136                         goto end_trans;
11137                 }
11138                 mutex_unlock(&fs_info->unused_bg_unpin_mutex);
11139
11140                 /* Reset pinned so btrfs_put_block_group doesn't complain */
11141                 spin_lock(&space_info->lock);
11142                 spin_lock(&block_group->lock);
11143
11144                 update_bytes_pinned(space_info, -block_group->pinned);
11145                 space_info->bytes_readonly += block_group->pinned;
11146                 percpu_counter_add_batch(&space_info->total_bytes_pinned,
11147                                    -block_group->pinned,
11148                                    BTRFS_TOTAL_BYTES_PINNED_BATCH);
11149                 block_group->pinned = 0;
11150
11151                 spin_unlock(&block_group->lock);
11152                 spin_unlock(&space_info->lock);
11153
11154                 /* DISCARD can flip during remount */
11155                 trimming = btrfs_test_opt(fs_info, DISCARD);
11156
11157                 /* Implicit trim during transaction commit. */
11158                 if (trimming)
11159                         btrfs_get_block_group_trimming(block_group);
11160
11161                 /*
11162                  * Btrfs_remove_chunk will abort the transaction if things go
11163                  * horribly wrong.
11164                  */
11165                 ret = btrfs_remove_chunk(trans, block_group->key.objectid);
11166
11167                 if (ret) {
11168                         if (trimming)
11169                                 btrfs_put_block_group_trimming(block_group);
11170                         goto end_trans;
11171                 }
11172
11173                 /*
11174                  * If we're not mounted with -odiscard, we can just forget
11175                  * about this block group. Otherwise we'll need to wait
11176                  * until transaction commit to do the actual discard.
11177                  */
11178                 if (trimming) {
11179                         spin_lock(&fs_info->unused_bgs_lock);
11180                         /*
11181                          * A concurrent scrub might have added us to the list
11182                          * fs_info->unused_bgs, so use a list_move operation
11183                          * to add the block group to the deleted_bgs list.
11184                          */
11185                         list_move(&block_group->bg_list,
11186                                   &trans->transaction->deleted_bgs);
11187                         spin_unlock(&fs_info->unused_bgs_lock);
11188                         btrfs_get_block_group(block_group);
11189                 }
11190 end_trans:
11191                 btrfs_end_transaction(trans);
11192 next:
11193                 mutex_unlock(&fs_info->delete_unused_bgs_mutex);
11194                 btrfs_put_block_group(block_group);
11195                 spin_lock(&fs_info->unused_bgs_lock);
11196         }
11197         spin_unlock(&fs_info->unused_bgs_lock);
11198 }
11199
11200 int btrfs_init_space_info(struct btrfs_fs_info *fs_info)
11201 {
11202         struct btrfs_super_block *disk_super;
11203         u64 features;
11204         u64 flags;
11205         int mixed = 0;
11206         int ret;
11207
11208         disk_super = fs_info->super_copy;
11209         if (!btrfs_super_root(disk_super))
11210                 return -EINVAL;
11211
11212         features = btrfs_super_incompat_flags(disk_super);
11213         if (features & BTRFS_FEATURE_INCOMPAT_MIXED_GROUPS)
11214                 mixed = 1;
11215
11216         flags = BTRFS_BLOCK_GROUP_SYSTEM;
11217         ret = create_space_info(fs_info, flags);
11218         if (ret)
11219                 goto out;
11220
11221         if (mixed) {
11222                 flags = BTRFS_BLOCK_GROUP_METADATA | BTRFS_BLOCK_GROUP_DATA;
11223                 ret = create_space_info(fs_info, flags);
11224         } else {
11225                 flags = BTRFS_BLOCK_GROUP_METADATA;
11226                 ret = create_space_info(fs_info, flags);
11227                 if (ret)
11228                         goto out;
11229
11230                 flags = BTRFS_BLOCK_GROUP_DATA;
11231                 ret = create_space_info(fs_info, flags);
11232         }
11233 out:
11234         return ret;
11235 }
11236
11237 int btrfs_error_unpin_extent_range(struct btrfs_fs_info *fs_info,
11238                                    u64 start, u64 end)
11239 {
11240         return unpin_extent_range(fs_info, start, end, false);
11241 }
11242
11243 /*
11244  * It used to be that old block groups would be left around forever.
11245  * Iterating over them would be enough to trim unused space.  Since we
11246  * now automatically remove them, we also need to iterate over unallocated
11247  * space.
11248  *
11249  * We don't want a transaction for this since the discard may take a
11250  * substantial amount of time.  We don't require that a transaction be
11251  * running, but we do need to take a running transaction into account
11252  * to ensure that we're not discarding chunks that were released or
11253  * allocated in the current transaction.
11254  *
11255  * Holding the chunks lock will prevent other threads from allocating
11256  * or releasing chunks, but it won't prevent a running transaction
11257  * from committing and releasing the memory that the pending chunks
11258  * list head uses.  For that, we need to take a reference to the
11259  * transaction and hold the commit root sem.  We only need to hold
11260  * it while performing the free space search since we have already
11261  * held back allocations.
11262  */
11263 static int btrfs_trim_free_extents(struct btrfs_device *device,
11264                                    u64 minlen, u64 *trimmed)
11265 {
11266         u64 start = 0, len = 0;
11267         int ret;
11268
11269         *trimmed = 0;
11270
11271         /* Discard not supported = nothing to do. */
11272         if (!blk_queue_discard(bdev_get_queue(device->bdev)))
11273                 return 0;
11274
11275         /* Not writable = nothing to do. */
11276         if (!test_bit(BTRFS_DEV_STATE_WRITEABLE, &device->dev_state))
11277                 return 0;
11278
11279         /* No free space = nothing to do. */
11280         if (device->total_bytes <= device->bytes_used)
11281                 return 0;
11282
11283         ret = 0;
11284
11285         while (1) {
11286                 struct btrfs_fs_info *fs_info = device->fs_info;
11287                 struct btrfs_transaction *trans;
11288                 u64 bytes;
11289
11290                 ret = mutex_lock_interruptible(&fs_info->chunk_mutex);
11291                 if (ret)
11292                         break;
11293
11294                 ret = down_read_killable(&fs_info->commit_root_sem);
11295                 if (ret) {
11296                         mutex_unlock(&fs_info->chunk_mutex);
11297                         break;
11298                 }
11299
11300                 spin_lock(&fs_info->trans_lock);
11301                 trans = fs_info->running_transaction;
11302                 if (trans)
11303                         refcount_inc(&trans->use_count);
11304                 spin_unlock(&fs_info->trans_lock);
11305
11306                 if (!trans)
11307                         up_read(&fs_info->commit_root_sem);
11308
11309                 ret = find_free_dev_extent_start(trans, device, minlen, start,
11310                                                  &start, &len);
11311                 if (trans) {
11312                         up_read(&fs_info->commit_root_sem);
11313                         btrfs_put_transaction(trans);
11314                 }
11315
11316                 if (ret) {
11317                         mutex_unlock(&fs_info->chunk_mutex);
11318                         if (ret == -ENOSPC)
11319                                 ret = 0;
11320                         break;
11321                 }
11322
11323                 ret = btrfs_issue_discard(device->bdev, start, len, &bytes);
11324                 mutex_unlock(&fs_info->chunk_mutex);
11325
11326                 if (ret)
11327                         break;
11328
11329                 start += len;
11330                 *trimmed += bytes;
11331
11332                 if (fatal_signal_pending(current)) {
11333                         ret = -ERESTARTSYS;
11334                         break;
11335                 }
11336
11337                 cond_resched();
11338         }
11339
11340         return ret;
11341 }
11342
11343 /*
11344  * Trim the whole filesystem by:
11345  * 1) trimming the free space in each block group
11346  * 2) trimming the unallocated space on each device
11347  *
11348  * This will also continue trimming even if a block group or device encounters
11349  * an error.  The return value will be the last error, or 0 if nothing bad
11350  * happens.
11351  */
11352 int btrfs_trim_fs(struct btrfs_fs_info *fs_info, struct fstrim_range *range)
11353 {
11354         struct btrfs_block_group_cache *cache = NULL;
11355         struct btrfs_device *device;
11356         struct list_head *devices;
11357         u64 group_trimmed;
11358         u64 start;
11359         u64 end;
11360         u64 trimmed = 0;
11361         u64 bg_failed = 0;
11362         u64 dev_failed = 0;
11363         int bg_ret = 0;
11364         int dev_ret = 0;
11365         int ret = 0;
11366
11367         cache = btrfs_lookup_first_block_group(fs_info, range->start);
11368         for (; cache; cache = next_block_group(fs_info, cache)) {
11369                 if (cache->key.objectid >= (range->start + range->len)) {
11370                         btrfs_put_block_group(cache);
11371                         break;
11372                 }
11373
11374                 start = max(range->start, cache->key.objectid);
11375                 end = min(range->start + range->len,
11376                                 cache->key.objectid + cache->key.offset);
11377
11378                 if (end - start >= range->minlen) {
11379                         if (!block_group_cache_done(cache)) {
11380                                 ret = cache_block_group(cache, 0);
11381                                 if (ret) {
11382                                         bg_failed++;
11383                                         bg_ret = ret;
11384                                         continue;
11385                                 }
11386                                 ret = wait_block_group_cache_done(cache);
11387                                 if (ret) {
11388                                         bg_failed++;
11389                                         bg_ret = ret;
11390                                         continue;
11391                                 }
11392                         }
11393                         ret = btrfs_trim_block_group(cache,
11394                                                      &group_trimmed,
11395                                                      start,
11396                                                      end,
11397                                                      range->minlen);
11398
11399                         trimmed += group_trimmed;
11400                         if (ret) {
11401                                 bg_failed++;
11402                                 bg_ret = ret;
11403                                 continue;
11404                         }
11405                 }
11406         }
11407
11408         if (bg_failed)
11409                 btrfs_warn(fs_info,
11410                         "failed to trim %llu block group(s), last error %d",
11411                         bg_failed, bg_ret);
11412         mutex_lock(&fs_info->fs_devices->device_list_mutex);
11413         devices = &fs_info->fs_devices->devices;
11414         list_for_each_entry(device, devices, dev_list) {
11415                 ret = btrfs_trim_free_extents(device, range->minlen,
11416                                               &group_trimmed);
11417                 if (ret) {
11418                         dev_failed++;
11419                         dev_ret = ret;
11420                         break;
11421                 }
11422
11423                 trimmed += group_trimmed;
11424         }
11425         mutex_unlock(&fs_info->fs_devices->device_list_mutex);
11426
11427         if (dev_failed)
11428                 btrfs_warn(fs_info,
11429                         "failed to trim %llu device(s), last error %d",
11430                         dev_failed, dev_ret);
11431         range->len = trimmed;
11432         if (bg_ret)
11433                 return bg_ret;
11434         return dev_ret;
11435 }
11436
11437 /*
11438  * btrfs_{start,end}_write_no_snapshotting() are similar to
11439  * mnt_{want,drop}_write(), they are used to prevent some tasks from writing
11440  * data into the page cache through nocow before the subvolume is snapshoted,
11441  * but flush the data into disk after the snapshot creation, or to prevent
11442  * operations while snapshotting is ongoing and that cause the snapshot to be
11443  * inconsistent (writes followed by expanding truncates for example).
11444  */
11445 void btrfs_end_write_no_snapshotting(struct btrfs_root *root)
11446 {
11447         percpu_counter_dec(&root->subv_writers->counter);
11448         cond_wake_up(&root->subv_writers->wait);
11449 }
11450
11451 int btrfs_start_write_no_snapshotting(struct btrfs_root *root)
11452 {
11453         if (atomic_read(&root->will_be_snapshotted))
11454                 return 0;
11455
11456         percpu_counter_inc(&root->subv_writers->counter);
11457         /*
11458          * Make sure counter is updated before we check for snapshot creation.
11459          */
11460         smp_mb();
11461         if (atomic_read(&root->will_be_snapshotted)) {
11462                 btrfs_end_write_no_snapshotting(root);
11463                 return 0;
11464         }
11465         return 1;
11466 }
11467
11468 void btrfs_wait_for_snapshot_creation(struct btrfs_root *root)
11469 {
11470         while (true) {
11471                 int ret;
11472
11473                 ret = btrfs_start_write_no_snapshotting(root);
11474                 if (ret)
11475                         break;
11476                 wait_var_event(&root->will_be_snapshotted,
11477                                !atomic_read(&root->will_be_snapshotted));
11478         }
11479 }
11480
11481 void btrfs_mark_bg_unused(struct btrfs_block_group_cache *bg)
11482 {
11483         struct btrfs_fs_info *fs_info = bg->fs_info;
11484
11485         spin_lock(&fs_info->unused_bgs_lock);
11486         if (list_empty(&bg->bg_list)) {
11487                 btrfs_get_block_group(bg);
11488                 trace_btrfs_add_unused_block_group(bg);
11489                 list_add_tail(&bg->bg_list, &fs_info->unused_bgs);
11490         }
11491         spin_unlock(&fs_info->unused_bgs_lock);
11492 }