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 static int btrfs_issue_discard(struct block_device *bdev, u64 start, u64 len,
1909                                u64 *discarded_bytes)
1910 {
1911         int j, ret = 0;
1912         u64 bytes_left, end;
1913         u64 aligned_start = ALIGN(start, 1 << 9);
1914
1915         if (WARN_ON(start != aligned_start)) {
1916                 len -= aligned_start - start;
1917                 len = round_down(len, 1 << 9);
1918                 start = aligned_start;
1919         }
1920
1921         *discarded_bytes = 0;
1922
1923         if (!len)
1924                 return 0;
1925
1926         end = start + len;
1927         bytes_left = len;
1928
1929         /* Skip any superblocks on this device. */
1930         for (j = 0; j < BTRFS_SUPER_MIRROR_MAX; j++) {
1931                 u64 sb_start = btrfs_sb_offset(j);
1932                 u64 sb_end = sb_start + BTRFS_SUPER_INFO_SIZE;
1933                 u64 size = sb_start - start;
1934
1935                 if (!in_range(sb_start, start, bytes_left) &&
1936                     !in_range(sb_end, start, bytes_left) &&
1937                     !in_range(start, sb_start, BTRFS_SUPER_INFO_SIZE))
1938                         continue;
1939
1940                 /*
1941                  * Superblock spans beginning of range.  Adjust start and
1942                  * try again.
1943                  */
1944                 if (sb_start <= start) {
1945                         start += sb_end - start;
1946                         if (start > end) {
1947                                 bytes_left = 0;
1948                                 break;
1949                         }
1950                         bytes_left = end - start;
1951                         continue;
1952                 }
1953
1954                 if (size) {
1955                         ret = blkdev_issue_discard(bdev, start >> 9, size >> 9,
1956                                                    GFP_NOFS, 0);
1957                         if (!ret)
1958                                 *discarded_bytes += size;
1959                         else if (ret != -EOPNOTSUPP)
1960                                 return ret;
1961                 }
1962
1963                 start = sb_end;
1964                 if (start > end) {
1965                         bytes_left = 0;
1966                         break;
1967                 }
1968                 bytes_left = end - start;
1969         }
1970
1971         if (bytes_left) {
1972                 ret = blkdev_issue_discard(bdev, start >> 9, bytes_left >> 9,
1973                                            GFP_NOFS, 0);
1974                 if (!ret)
1975                         *discarded_bytes += bytes_left;
1976         }
1977         return ret;
1978 }
1979
1980 int btrfs_discard_extent(struct btrfs_fs_info *fs_info, u64 bytenr,
1981                          u64 num_bytes, u64 *actual_bytes)
1982 {
1983         int ret;
1984         u64 discarded_bytes = 0;
1985         struct btrfs_bio *bbio = NULL;
1986
1987
1988         /*
1989          * Avoid races with device replace and make sure our bbio has devices
1990          * associated to its stripes that don't go away while we are discarding.
1991          */
1992         btrfs_bio_counter_inc_blocked(fs_info);
1993         /* Tell the block device(s) that the sectors can be discarded */
1994         ret = btrfs_map_block(fs_info, BTRFS_MAP_DISCARD, bytenr, &num_bytes,
1995                               &bbio, 0);
1996         /* Error condition is -ENOMEM */
1997         if (!ret) {
1998                 struct btrfs_bio_stripe *stripe = bbio->stripes;
1999                 int i;
2000
2001
2002                 for (i = 0; i < bbio->num_stripes; i++, stripe++) {
2003                         u64 bytes;
2004                         struct request_queue *req_q;
2005
2006                         if (!stripe->dev->bdev) {
2007                                 ASSERT(btrfs_test_opt(fs_info, DEGRADED));
2008                                 continue;
2009                         }
2010                         req_q = bdev_get_queue(stripe->dev->bdev);
2011                         if (!blk_queue_discard(req_q))
2012                                 continue;
2013
2014                         ret = btrfs_issue_discard(stripe->dev->bdev,
2015                                                   stripe->physical,
2016                                                   stripe->length,
2017                                                   &bytes);
2018                         if (!ret)
2019                                 discarded_bytes += bytes;
2020                         else if (ret != -EOPNOTSUPP)
2021                                 break; /* Logic errors or -ENOMEM, or -EIO but I don't know how that could happen JDM */
2022
2023                         /*
2024                          * Just in case we get back EOPNOTSUPP for some reason,
2025                          * just ignore the return value so we don't screw up
2026                          * people calling discard_extent.
2027                          */
2028                         ret = 0;
2029                 }
2030                 btrfs_put_bbio(bbio);
2031         }
2032         btrfs_bio_counter_dec(fs_info);
2033
2034         if (actual_bytes)
2035                 *actual_bytes = discarded_bytes;
2036
2037
2038         if (ret == -EOPNOTSUPP)
2039                 ret = 0;
2040         return ret;
2041 }
2042
2043 /* Can return -ENOMEM */
2044 int btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
2045                          struct btrfs_root *root,
2046                          u64 bytenr, u64 num_bytes, u64 parent,
2047                          u64 root_objectid, u64 owner, u64 offset)
2048 {
2049         struct btrfs_fs_info *fs_info = root->fs_info;
2050         int old_ref_mod, new_ref_mod;
2051         int ret;
2052
2053         BUG_ON(owner < BTRFS_FIRST_FREE_OBJECTID &&
2054                root_objectid == BTRFS_TREE_LOG_OBJECTID);
2055
2056         btrfs_ref_tree_mod(root, bytenr, num_bytes, parent, root_objectid,
2057                            owner, offset, BTRFS_ADD_DELAYED_REF);
2058
2059         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
2060                 ret = btrfs_add_delayed_tree_ref(trans, bytenr,
2061                                                  num_bytes, parent,
2062                                                  root_objectid, (int)owner,
2063                                                  BTRFS_ADD_DELAYED_REF, NULL,
2064                                                  &old_ref_mod, &new_ref_mod);
2065         } else {
2066                 ret = btrfs_add_delayed_data_ref(trans, bytenr,
2067                                                  num_bytes, parent,
2068                                                  root_objectid, owner, offset,
2069                                                  0, BTRFS_ADD_DELAYED_REF,
2070                                                  &old_ref_mod, &new_ref_mod);
2071         }
2072
2073         if (ret == 0 && old_ref_mod < 0 && new_ref_mod >= 0) {
2074                 bool metadata = owner < BTRFS_FIRST_FREE_OBJECTID;
2075
2076                 add_pinned_bytes(fs_info, -num_bytes, metadata, root_objectid);
2077         }
2078
2079         return ret;
2080 }
2081
2082 /*
2083  * __btrfs_inc_extent_ref - insert backreference for a given extent
2084  *
2085  * @trans:          Handle of transaction
2086  *
2087  * @node:           The delayed ref node used to get the bytenr/length for
2088  *                  extent whose references are incremented.
2089  *
2090  * @parent:         If this is a shared extent (BTRFS_SHARED_DATA_REF_KEY/
2091  *                  BTRFS_SHARED_BLOCK_REF_KEY) then it holds the logical
2092  *                  bytenr of the parent block. Since new extents are always
2093  *                  created with indirect references, this will only be the case
2094  *                  when relocating a shared extent. In that case, root_objectid
2095  *                  will be BTRFS_TREE_RELOC_OBJECTID. Otheriwse, parent must
2096  *                  be 0
2097  *
2098  * @root_objectid:  The id of the root where this modification has originated,
2099  *                  this can be either one of the well-known metadata trees or
2100  *                  the subvolume id which references this extent.
2101  *
2102  * @owner:          For data extents it is the inode number of the owning file.
2103  *                  For metadata extents this parameter holds the level in the
2104  *                  tree of the extent.
2105  *
2106  * @offset:         For metadata extents the offset is ignored and is currently
2107  *                  always passed as 0. For data extents it is the fileoffset
2108  *                  this extent belongs to.
2109  *
2110  * @refs_to_add     Number of references to add
2111  *
2112  * @extent_op       Pointer to a structure, holding information necessary when
2113  *                  updating a tree block's flags
2114  *
2115  */
2116 static int __btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
2117                                   struct btrfs_delayed_ref_node *node,
2118                                   u64 parent, u64 root_objectid,
2119                                   u64 owner, u64 offset, int refs_to_add,
2120                                   struct btrfs_delayed_extent_op *extent_op)
2121 {
2122         struct btrfs_path *path;
2123         struct extent_buffer *leaf;
2124         struct btrfs_extent_item *item;
2125         struct btrfs_key key;
2126         u64 bytenr = node->bytenr;
2127         u64 num_bytes = node->num_bytes;
2128         u64 refs;
2129         int ret;
2130
2131         path = btrfs_alloc_path();
2132         if (!path)
2133                 return -ENOMEM;
2134
2135         path->reada = READA_FORWARD;
2136         path->leave_spinning = 1;
2137         /* this will setup the path even if it fails to insert the back ref */
2138         ret = insert_inline_extent_backref(trans, path, bytenr, num_bytes,
2139                                            parent, root_objectid, owner,
2140                                            offset, refs_to_add, extent_op);
2141         if ((ret < 0 && ret != -EAGAIN) || !ret)
2142                 goto out;
2143
2144         /*
2145          * Ok we had -EAGAIN which means we didn't have space to insert and
2146          * inline extent ref, so just update the reference count and add a
2147          * normal backref.
2148          */
2149         leaf = path->nodes[0];
2150         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
2151         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2152         refs = btrfs_extent_refs(leaf, item);
2153         btrfs_set_extent_refs(leaf, item, refs + refs_to_add);
2154         if (extent_op)
2155                 __run_delayed_extent_op(extent_op, leaf, item);
2156
2157         btrfs_mark_buffer_dirty(leaf);
2158         btrfs_release_path(path);
2159
2160         path->reada = READA_FORWARD;
2161         path->leave_spinning = 1;
2162         /* now insert the actual backref */
2163         ret = insert_extent_backref(trans, path, bytenr, parent, root_objectid,
2164                                     owner, offset, refs_to_add);
2165         if (ret)
2166                 btrfs_abort_transaction(trans, ret);
2167 out:
2168         btrfs_free_path(path);
2169         return ret;
2170 }
2171
2172 static int run_delayed_data_ref(struct btrfs_trans_handle *trans,
2173                                 struct btrfs_delayed_ref_node *node,
2174                                 struct btrfs_delayed_extent_op *extent_op,
2175                                 int insert_reserved)
2176 {
2177         int ret = 0;
2178         struct btrfs_delayed_data_ref *ref;
2179         struct btrfs_key ins;
2180         u64 parent = 0;
2181         u64 ref_root = 0;
2182         u64 flags = 0;
2183
2184         ins.objectid = node->bytenr;
2185         ins.offset = node->num_bytes;
2186         ins.type = BTRFS_EXTENT_ITEM_KEY;
2187
2188         ref = btrfs_delayed_node_to_data_ref(node);
2189         trace_run_delayed_data_ref(trans->fs_info, node, ref, node->action);
2190
2191         if (node->type == BTRFS_SHARED_DATA_REF_KEY)
2192                 parent = ref->parent;
2193         ref_root = ref->root;
2194
2195         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
2196                 if (extent_op)
2197                         flags |= extent_op->flags_to_set;
2198                 ret = alloc_reserved_file_extent(trans, parent, ref_root,
2199                                                  flags, ref->objectid,
2200                                                  ref->offset, &ins,
2201                                                  node->ref_mod);
2202         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
2203                 ret = __btrfs_inc_extent_ref(trans, node, parent, ref_root,
2204                                              ref->objectid, ref->offset,
2205                                              node->ref_mod, extent_op);
2206         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
2207                 ret = __btrfs_free_extent(trans, node, parent,
2208                                           ref_root, ref->objectid,
2209                                           ref->offset, node->ref_mod,
2210                                           extent_op);
2211         } else {
2212                 BUG();
2213         }
2214         return ret;
2215 }
2216
2217 static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op,
2218                                     struct extent_buffer *leaf,
2219                                     struct btrfs_extent_item *ei)
2220 {
2221         u64 flags = btrfs_extent_flags(leaf, ei);
2222         if (extent_op->update_flags) {
2223                 flags |= extent_op->flags_to_set;
2224                 btrfs_set_extent_flags(leaf, ei, flags);
2225         }
2226
2227         if (extent_op->update_key) {
2228                 struct btrfs_tree_block_info *bi;
2229                 BUG_ON(!(flags & BTRFS_EXTENT_FLAG_TREE_BLOCK));
2230                 bi = (struct btrfs_tree_block_info *)(ei + 1);
2231                 btrfs_set_tree_block_key(leaf, bi, &extent_op->key);
2232         }
2233 }
2234
2235 static int run_delayed_extent_op(struct btrfs_trans_handle *trans,
2236                                  struct btrfs_delayed_ref_head *head,
2237                                  struct btrfs_delayed_extent_op *extent_op)
2238 {
2239         struct btrfs_fs_info *fs_info = trans->fs_info;
2240         struct btrfs_key key;
2241         struct btrfs_path *path;
2242         struct btrfs_extent_item *ei;
2243         struct extent_buffer *leaf;
2244         u32 item_size;
2245         int ret;
2246         int err = 0;
2247         int metadata = !extent_op->is_data;
2248
2249         if (trans->aborted)
2250                 return 0;
2251
2252         if (metadata && !btrfs_fs_incompat(fs_info, SKINNY_METADATA))
2253                 metadata = 0;
2254
2255         path = btrfs_alloc_path();
2256         if (!path)
2257                 return -ENOMEM;
2258
2259         key.objectid = head->bytenr;
2260
2261         if (metadata) {
2262                 key.type = BTRFS_METADATA_ITEM_KEY;
2263                 key.offset = extent_op->level;
2264         } else {
2265                 key.type = BTRFS_EXTENT_ITEM_KEY;
2266                 key.offset = head->num_bytes;
2267         }
2268
2269 again:
2270         path->reada = READA_FORWARD;
2271         path->leave_spinning = 1;
2272         ret = btrfs_search_slot(trans, fs_info->extent_root, &key, path, 0, 1);
2273         if (ret < 0) {
2274                 err = ret;
2275                 goto out;
2276         }
2277         if (ret > 0) {
2278                 if (metadata) {
2279                         if (path->slots[0] > 0) {
2280                                 path->slots[0]--;
2281                                 btrfs_item_key_to_cpu(path->nodes[0], &key,
2282                                                       path->slots[0]);
2283                                 if (key.objectid == head->bytenr &&
2284                                     key.type == BTRFS_EXTENT_ITEM_KEY &&
2285                                     key.offset == head->num_bytes)
2286                                         ret = 0;
2287                         }
2288                         if (ret > 0) {
2289                                 btrfs_release_path(path);
2290                                 metadata = 0;
2291
2292                                 key.objectid = head->bytenr;
2293                                 key.offset = head->num_bytes;
2294                                 key.type = BTRFS_EXTENT_ITEM_KEY;
2295                                 goto again;
2296                         }
2297                 } else {
2298                         err = -EIO;
2299                         goto out;
2300                 }
2301         }
2302
2303         leaf = path->nodes[0];
2304         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2305
2306         if (unlikely(item_size < sizeof(*ei))) {
2307                 err = -EINVAL;
2308                 btrfs_print_v0_err(fs_info);
2309                 btrfs_abort_transaction(trans, err);
2310                 goto out;
2311         }
2312
2313         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2314         __run_delayed_extent_op(extent_op, leaf, ei);
2315
2316         btrfs_mark_buffer_dirty(leaf);
2317 out:
2318         btrfs_free_path(path);
2319         return err;
2320 }
2321
2322 static int run_delayed_tree_ref(struct btrfs_trans_handle *trans,
2323                                 struct btrfs_delayed_ref_node *node,
2324                                 struct btrfs_delayed_extent_op *extent_op,
2325                                 int insert_reserved)
2326 {
2327         int ret = 0;
2328         struct btrfs_delayed_tree_ref *ref;
2329         u64 parent = 0;
2330         u64 ref_root = 0;
2331
2332         ref = btrfs_delayed_node_to_tree_ref(node);
2333         trace_run_delayed_tree_ref(trans->fs_info, node, ref, node->action);
2334
2335         if (node->type == BTRFS_SHARED_BLOCK_REF_KEY)
2336                 parent = ref->parent;
2337         ref_root = ref->root;
2338
2339         if (node->ref_mod != 1) {
2340                 btrfs_err(trans->fs_info,
2341         "btree block(%llu) has %d references rather than 1: action %d ref_root %llu parent %llu",
2342                           node->bytenr, node->ref_mod, node->action, ref_root,
2343                           parent);
2344                 return -EIO;
2345         }
2346         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
2347                 BUG_ON(!extent_op || !extent_op->update_flags);
2348                 ret = alloc_reserved_tree_block(trans, node, extent_op);
2349         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
2350                 ret = __btrfs_inc_extent_ref(trans, node, parent, ref_root,
2351                                              ref->level, 0, 1, extent_op);
2352         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
2353                 ret = __btrfs_free_extent(trans, node, parent, ref_root,
2354                                           ref->level, 0, 1, extent_op);
2355         } else {
2356                 BUG();
2357         }
2358         return ret;
2359 }
2360
2361 /* helper function to actually process a single delayed ref entry */
2362 static int run_one_delayed_ref(struct btrfs_trans_handle *trans,
2363                                struct btrfs_delayed_ref_node *node,
2364                                struct btrfs_delayed_extent_op *extent_op,
2365                                int insert_reserved)
2366 {
2367         int ret = 0;
2368
2369         if (trans->aborted) {
2370                 if (insert_reserved)
2371                         btrfs_pin_extent(trans->fs_info, node->bytenr,
2372                                          node->num_bytes, 1);
2373                 return 0;
2374         }
2375
2376         if (node->type == BTRFS_TREE_BLOCK_REF_KEY ||
2377             node->type == BTRFS_SHARED_BLOCK_REF_KEY)
2378                 ret = run_delayed_tree_ref(trans, node, extent_op,
2379                                            insert_reserved);
2380         else if (node->type == BTRFS_EXTENT_DATA_REF_KEY ||
2381                  node->type == BTRFS_SHARED_DATA_REF_KEY)
2382                 ret = run_delayed_data_ref(trans, node, extent_op,
2383                                            insert_reserved);
2384         else
2385                 BUG();
2386         if (ret && insert_reserved)
2387                 btrfs_pin_extent(trans->fs_info, node->bytenr,
2388                                  node->num_bytes, 1);
2389         return ret;
2390 }
2391
2392 static inline struct btrfs_delayed_ref_node *
2393 select_delayed_ref(struct btrfs_delayed_ref_head *head)
2394 {
2395         struct btrfs_delayed_ref_node *ref;
2396
2397         if (RB_EMPTY_ROOT(&head->ref_tree.rb_root))
2398                 return NULL;
2399
2400         /*
2401          * Select a delayed ref of type BTRFS_ADD_DELAYED_REF first.
2402          * This is to prevent a ref count from going down to zero, which deletes
2403          * the extent item from the extent tree, when there still are references
2404          * to add, which would fail because they would not find the extent item.
2405          */
2406         if (!list_empty(&head->ref_add_list))
2407                 return list_first_entry(&head->ref_add_list,
2408                                 struct btrfs_delayed_ref_node, add_list);
2409
2410         ref = rb_entry(rb_first_cached(&head->ref_tree),
2411                        struct btrfs_delayed_ref_node, ref_node);
2412         ASSERT(list_empty(&ref->add_list));
2413         return ref;
2414 }
2415
2416 static void unselect_delayed_ref_head(struct btrfs_delayed_ref_root *delayed_refs,
2417                                       struct btrfs_delayed_ref_head *head)
2418 {
2419         spin_lock(&delayed_refs->lock);
2420         head->processing = 0;
2421         delayed_refs->num_heads_ready++;
2422         spin_unlock(&delayed_refs->lock);
2423         btrfs_delayed_ref_unlock(head);
2424 }
2425
2426 static struct btrfs_delayed_extent_op *cleanup_extent_op(
2427                                 struct btrfs_delayed_ref_head *head)
2428 {
2429         struct btrfs_delayed_extent_op *extent_op = head->extent_op;
2430
2431         if (!extent_op)
2432                 return NULL;
2433
2434         if (head->must_insert_reserved) {
2435                 head->extent_op = NULL;
2436                 btrfs_free_delayed_extent_op(extent_op);
2437                 return NULL;
2438         }
2439         return extent_op;
2440 }
2441
2442 static int run_and_cleanup_extent_op(struct btrfs_trans_handle *trans,
2443                                      struct btrfs_delayed_ref_head *head)
2444 {
2445         struct btrfs_delayed_extent_op *extent_op;
2446         int ret;
2447
2448         extent_op = cleanup_extent_op(head);
2449         if (!extent_op)
2450                 return 0;
2451         head->extent_op = NULL;
2452         spin_unlock(&head->lock);
2453         ret = run_delayed_extent_op(trans, head, extent_op);
2454         btrfs_free_delayed_extent_op(extent_op);
2455         return ret ? ret : 1;
2456 }
2457
2458 void btrfs_cleanup_ref_head_accounting(struct btrfs_fs_info *fs_info,
2459                                   struct btrfs_delayed_ref_root *delayed_refs,
2460                                   struct btrfs_delayed_ref_head *head)
2461 {
2462         int nr_items = 1;       /* Dropping this ref head update. */
2463
2464         if (head->total_ref_mod < 0) {
2465                 struct btrfs_space_info *space_info;
2466                 u64 flags;
2467
2468                 if (head->is_data)
2469                         flags = BTRFS_BLOCK_GROUP_DATA;
2470                 else if (head->is_system)
2471                         flags = BTRFS_BLOCK_GROUP_SYSTEM;
2472                 else
2473                         flags = BTRFS_BLOCK_GROUP_METADATA;
2474                 space_info = __find_space_info(fs_info, flags);
2475                 ASSERT(space_info);
2476                 percpu_counter_add_batch(&space_info->total_bytes_pinned,
2477                                    -head->num_bytes,
2478                                    BTRFS_TOTAL_BYTES_PINNED_BATCH);
2479
2480                 /*
2481                  * We had csum deletions accounted for in our delayed refs rsv,
2482                  * we need to drop the csum leaves for this update from our
2483                  * delayed_refs_rsv.
2484                  */
2485                 if (head->is_data) {
2486                         spin_lock(&delayed_refs->lock);
2487                         delayed_refs->pending_csums -= head->num_bytes;
2488                         spin_unlock(&delayed_refs->lock);
2489                         nr_items += btrfs_csum_bytes_to_leaves(fs_info,
2490                                 head->num_bytes);
2491                 }
2492         }
2493
2494         btrfs_delayed_refs_rsv_release(fs_info, nr_items);
2495 }
2496
2497 static int cleanup_ref_head(struct btrfs_trans_handle *trans,
2498                             struct btrfs_delayed_ref_head *head)
2499 {
2500
2501         struct btrfs_fs_info *fs_info = trans->fs_info;
2502         struct btrfs_delayed_ref_root *delayed_refs;
2503         int ret;
2504
2505         delayed_refs = &trans->transaction->delayed_refs;
2506
2507         ret = run_and_cleanup_extent_op(trans, head);
2508         if (ret < 0) {
2509                 unselect_delayed_ref_head(delayed_refs, head);
2510                 btrfs_debug(fs_info, "run_delayed_extent_op returned %d", ret);
2511                 return ret;
2512         } else if (ret) {
2513                 return ret;
2514         }
2515
2516         /*
2517          * Need to drop our head ref lock and re-acquire the delayed ref lock
2518          * and then re-check to make sure nobody got added.
2519          */
2520         spin_unlock(&head->lock);
2521         spin_lock(&delayed_refs->lock);
2522         spin_lock(&head->lock);
2523         if (!RB_EMPTY_ROOT(&head->ref_tree.rb_root) || head->extent_op) {
2524                 spin_unlock(&head->lock);
2525                 spin_unlock(&delayed_refs->lock);
2526                 return 1;
2527         }
2528         btrfs_delete_ref_head(delayed_refs, head);
2529         spin_unlock(&head->lock);
2530         spin_unlock(&delayed_refs->lock);
2531
2532         if (head->must_insert_reserved) {
2533                 btrfs_pin_extent(fs_info, head->bytenr,
2534                                  head->num_bytes, 1);
2535                 if (head->is_data) {
2536                         ret = btrfs_del_csums(trans, fs_info, head->bytenr,
2537                                               head->num_bytes);
2538                 }
2539         }
2540
2541         btrfs_cleanup_ref_head_accounting(fs_info, delayed_refs, head);
2542
2543         trace_run_delayed_ref_head(fs_info, head, 0);
2544         btrfs_delayed_ref_unlock(head);
2545         btrfs_put_delayed_ref_head(head);
2546         return 0;
2547 }
2548
2549 static struct btrfs_delayed_ref_head *btrfs_obtain_ref_head(
2550                                         struct btrfs_trans_handle *trans)
2551 {
2552         struct btrfs_delayed_ref_root *delayed_refs =
2553                 &trans->transaction->delayed_refs;
2554         struct btrfs_delayed_ref_head *head = NULL;
2555         int ret;
2556
2557         spin_lock(&delayed_refs->lock);
2558         head = btrfs_select_ref_head(delayed_refs);
2559         if (!head) {
2560                 spin_unlock(&delayed_refs->lock);
2561                 return head;
2562         }
2563
2564         /*
2565          * Grab the lock that says we are going to process all the refs for
2566          * this head
2567          */
2568         ret = btrfs_delayed_ref_lock(delayed_refs, head);
2569         spin_unlock(&delayed_refs->lock);
2570
2571         /*
2572          * We may have dropped the spin lock to get the head mutex lock, and
2573          * that might have given someone else time to free the head.  If that's
2574          * true, it has been removed from our list and we can move on.
2575          */
2576         if (ret == -EAGAIN)
2577                 head = ERR_PTR(-EAGAIN);
2578
2579         return head;
2580 }
2581
2582 static int btrfs_run_delayed_refs_for_head(struct btrfs_trans_handle *trans,
2583                                     struct btrfs_delayed_ref_head *locked_ref,
2584                                     unsigned long *run_refs)
2585 {
2586         struct btrfs_fs_info *fs_info = trans->fs_info;
2587         struct btrfs_delayed_ref_root *delayed_refs;
2588         struct btrfs_delayed_extent_op *extent_op;
2589         struct btrfs_delayed_ref_node *ref;
2590         int must_insert_reserved = 0;
2591         int ret;
2592
2593         delayed_refs = &trans->transaction->delayed_refs;
2594
2595         lockdep_assert_held(&locked_ref->mutex);
2596         lockdep_assert_held(&locked_ref->lock);
2597
2598         while ((ref = select_delayed_ref(locked_ref))) {
2599                 if (ref->seq &&
2600                     btrfs_check_delayed_seq(fs_info, ref->seq)) {
2601                         spin_unlock(&locked_ref->lock);
2602                         unselect_delayed_ref_head(delayed_refs, locked_ref);
2603                         return -EAGAIN;
2604                 }
2605
2606                 (*run_refs)++;
2607                 ref->in_tree = 0;
2608                 rb_erase_cached(&ref->ref_node, &locked_ref->ref_tree);
2609                 RB_CLEAR_NODE(&ref->ref_node);
2610                 if (!list_empty(&ref->add_list))
2611                         list_del(&ref->add_list);
2612                 /*
2613                  * When we play the delayed ref, also correct the ref_mod on
2614                  * head
2615                  */
2616                 switch (ref->action) {
2617                 case BTRFS_ADD_DELAYED_REF:
2618                 case BTRFS_ADD_DELAYED_EXTENT:
2619                         locked_ref->ref_mod -= ref->ref_mod;
2620                         break;
2621                 case BTRFS_DROP_DELAYED_REF:
2622                         locked_ref->ref_mod += ref->ref_mod;
2623                         break;
2624                 default:
2625                         WARN_ON(1);
2626                 }
2627                 atomic_dec(&delayed_refs->num_entries);
2628
2629                 /*
2630                  * Record the must_insert_reserved flag before we drop the
2631                  * spin lock.
2632                  */
2633                 must_insert_reserved = locked_ref->must_insert_reserved;
2634                 locked_ref->must_insert_reserved = 0;
2635
2636                 extent_op = locked_ref->extent_op;
2637                 locked_ref->extent_op = NULL;
2638                 spin_unlock(&locked_ref->lock);
2639
2640                 ret = run_one_delayed_ref(trans, ref, extent_op,
2641                                           must_insert_reserved);
2642
2643                 btrfs_free_delayed_extent_op(extent_op);
2644                 if (ret) {
2645                         unselect_delayed_ref_head(delayed_refs, locked_ref);
2646                         btrfs_put_delayed_ref(ref);
2647                         btrfs_debug(fs_info, "run_one_delayed_ref returned %d",
2648                                     ret);
2649                         return ret;
2650                 }
2651
2652                 btrfs_put_delayed_ref(ref);
2653                 cond_resched();
2654
2655                 spin_lock(&locked_ref->lock);
2656                 btrfs_merge_delayed_refs(trans, delayed_refs, locked_ref);
2657         }
2658
2659         return 0;
2660 }
2661
2662 /*
2663  * Returns 0 on success or if called with an already aborted transaction.
2664  * Returns -ENOMEM or -EIO on failure and will abort the transaction.
2665  */
2666 static noinline int __btrfs_run_delayed_refs(struct btrfs_trans_handle *trans,
2667                                              unsigned long nr)
2668 {
2669         struct btrfs_fs_info *fs_info = trans->fs_info;
2670         struct btrfs_delayed_ref_root *delayed_refs;
2671         struct btrfs_delayed_ref_head *locked_ref = NULL;
2672         ktime_t start = ktime_get();
2673         int ret;
2674         unsigned long count = 0;
2675         unsigned long actual_count = 0;
2676
2677         delayed_refs = &trans->transaction->delayed_refs;
2678         do {
2679                 if (!locked_ref) {
2680                         locked_ref = btrfs_obtain_ref_head(trans);
2681                         if (IS_ERR_OR_NULL(locked_ref)) {
2682                                 if (PTR_ERR(locked_ref) == -EAGAIN) {
2683                                         continue;
2684                                 } else {
2685                                         break;
2686                                 }
2687                         }
2688                         count++;
2689                 }
2690                 /*
2691                  * We need to try and merge add/drops of the same ref since we
2692                  * can run into issues with relocate dropping the implicit ref
2693                  * and then it being added back again before the drop can
2694                  * finish.  If we merged anything we need to re-loop so we can
2695                  * get a good ref.
2696                  * Or we can get node references of the same type that weren't
2697                  * merged when created due to bumps in the tree mod seq, and
2698                  * we need to merge them to prevent adding an inline extent
2699                  * backref before dropping it (triggering a BUG_ON at
2700                  * insert_inline_extent_backref()).
2701                  */
2702                 spin_lock(&locked_ref->lock);
2703                 btrfs_merge_delayed_refs(trans, delayed_refs, locked_ref);
2704
2705                 ret = btrfs_run_delayed_refs_for_head(trans, locked_ref,
2706                                                       &actual_count);
2707                 if (ret < 0 && ret != -EAGAIN) {
2708                         /*
2709                          * Error, btrfs_run_delayed_refs_for_head already
2710                          * unlocked everything so just bail out
2711                          */
2712                         return ret;
2713                 } else if (!ret) {
2714                         /*
2715                          * Success, perform the usual cleanup of a processed
2716                          * head
2717                          */
2718                         ret = cleanup_ref_head(trans, locked_ref);
2719                         if (ret > 0 ) {
2720                                 /* We dropped our lock, we need to loop. */
2721                                 ret = 0;
2722                                 continue;
2723                         } else if (ret) {
2724                                 return ret;
2725                         }
2726                 }
2727
2728                 /*
2729                  * Either success case or btrfs_run_delayed_refs_for_head
2730                  * returned -EAGAIN, meaning we need to select another head
2731                  */
2732
2733                 locked_ref = NULL;
2734                 cond_resched();
2735         } while ((nr != -1 && count < nr) || locked_ref);
2736
2737         /*
2738          * We don't want to include ref heads since we can have empty ref heads
2739          * and those will drastically skew our runtime down since we just do
2740          * accounting, no actual extent tree updates.
2741          */
2742         if (actual_count > 0) {
2743                 u64 runtime = ktime_to_ns(ktime_sub(ktime_get(), start));
2744                 u64 avg;
2745
2746                 /*
2747                  * We weigh the current average higher than our current runtime
2748                  * to avoid large swings in the average.
2749                  */
2750                 spin_lock(&delayed_refs->lock);
2751                 avg = fs_info->avg_delayed_ref_runtime * 3 + runtime;
2752                 fs_info->avg_delayed_ref_runtime = avg >> 2;    /* div by 4 */
2753                 spin_unlock(&delayed_refs->lock);
2754         }
2755         return 0;
2756 }
2757
2758 #ifdef SCRAMBLE_DELAYED_REFS
2759 /*
2760  * Normally delayed refs get processed in ascending bytenr order. This
2761  * correlates in most cases to the order added. To expose dependencies on this
2762  * order, we start to process the tree in the middle instead of the beginning
2763  */
2764 static u64 find_middle(struct rb_root *root)
2765 {
2766         struct rb_node *n = root->rb_node;
2767         struct btrfs_delayed_ref_node *entry;
2768         int alt = 1;
2769         u64 middle;
2770         u64 first = 0, last = 0;
2771
2772         n = rb_first(root);
2773         if (n) {
2774                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2775                 first = entry->bytenr;
2776         }
2777         n = rb_last(root);
2778         if (n) {
2779                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2780                 last = entry->bytenr;
2781         }
2782         n = root->rb_node;
2783
2784         while (n) {
2785                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2786                 WARN_ON(!entry->in_tree);
2787
2788                 middle = entry->bytenr;
2789
2790                 if (alt)
2791                         n = n->rb_left;
2792                 else
2793                         n = n->rb_right;
2794
2795                 alt = 1 - alt;
2796         }
2797         return middle;
2798 }
2799 #endif
2800
2801 static inline u64 heads_to_leaves(struct btrfs_fs_info *fs_info, u64 heads)
2802 {
2803         u64 num_bytes;
2804
2805         num_bytes = heads * (sizeof(struct btrfs_extent_item) +
2806                              sizeof(struct btrfs_extent_inline_ref));
2807         if (!btrfs_fs_incompat(fs_info, SKINNY_METADATA))
2808                 num_bytes += heads * sizeof(struct btrfs_tree_block_info);
2809
2810         /*
2811          * We don't ever fill up leaves all the way so multiply by 2 just to be
2812          * closer to what we're really going to want to use.
2813          */
2814         return div_u64(num_bytes, BTRFS_LEAF_DATA_SIZE(fs_info));
2815 }
2816
2817 /*
2818  * Takes the number of bytes to be csumm'ed and figures out how many leaves it
2819  * would require to store the csums for that many bytes.
2820  */
2821 u64 btrfs_csum_bytes_to_leaves(struct btrfs_fs_info *fs_info, u64 csum_bytes)
2822 {
2823         u64 csum_size;
2824         u64 num_csums_per_leaf;
2825         u64 num_csums;
2826
2827         csum_size = BTRFS_MAX_ITEM_SIZE(fs_info);
2828         num_csums_per_leaf = div64_u64(csum_size,
2829                         (u64)btrfs_super_csum_size(fs_info->super_copy));
2830         num_csums = div64_u64(csum_bytes, fs_info->sectorsize);
2831         num_csums += num_csums_per_leaf - 1;
2832         num_csums = div64_u64(num_csums, num_csums_per_leaf);
2833         return num_csums;
2834 }
2835
2836 bool btrfs_check_space_for_delayed_refs(struct btrfs_fs_info *fs_info)
2837 {
2838         struct btrfs_block_rsv *delayed_refs_rsv = &fs_info->delayed_refs_rsv;
2839         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
2840         bool ret = false;
2841         u64 reserved;
2842
2843         spin_lock(&global_rsv->lock);
2844         reserved = global_rsv->reserved;
2845         spin_unlock(&global_rsv->lock);
2846
2847         /*
2848          * Since the global reserve is just kind of magic we don't really want
2849          * to rely on it to save our bacon, so if our size is more than the
2850          * delayed_refs_rsv and the global rsv then it's time to think about
2851          * bailing.
2852          */
2853         spin_lock(&delayed_refs_rsv->lock);
2854         reserved += delayed_refs_rsv->reserved;
2855         if (delayed_refs_rsv->size >= reserved)
2856                 ret = true;
2857         spin_unlock(&delayed_refs_rsv->lock);
2858         return ret;
2859 }
2860
2861 int btrfs_should_throttle_delayed_refs(struct btrfs_trans_handle *trans)
2862 {
2863         u64 num_entries =
2864                 atomic_read(&trans->transaction->delayed_refs.num_entries);
2865         u64 avg_runtime;
2866         u64 val;
2867
2868         smp_mb();
2869         avg_runtime = trans->fs_info->avg_delayed_ref_runtime;
2870         val = num_entries * avg_runtime;
2871         if (val >= NSEC_PER_SEC)
2872                 return 1;
2873         if (val >= NSEC_PER_SEC / 2)
2874                 return 2;
2875
2876         return btrfs_check_space_for_delayed_refs(trans->fs_info);
2877 }
2878
2879 struct async_delayed_refs {
2880         struct btrfs_root *root;
2881         u64 transid;
2882         int count;
2883         int error;
2884         int sync;
2885         struct completion wait;
2886         struct btrfs_work work;
2887 };
2888
2889 static inline struct async_delayed_refs *
2890 to_async_delayed_refs(struct btrfs_work *work)
2891 {
2892         return container_of(work, struct async_delayed_refs, work);
2893 }
2894
2895 static void delayed_ref_async_start(struct btrfs_work *work)
2896 {
2897         struct async_delayed_refs *async = to_async_delayed_refs(work);
2898         struct btrfs_trans_handle *trans;
2899         struct btrfs_fs_info *fs_info = async->root->fs_info;
2900         int ret;
2901
2902         /* if the commit is already started, we don't need to wait here */
2903         if (btrfs_transaction_blocked(fs_info))
2904                 goto done;
2905
2906         trans = btrfs_join_transaction(async->root);
2907         if (IS_ERR(trans)) {
2908                 async->error = PTR_ERR(trans);
2909                 goto done;
2910         }
2911
2912         /* Don't bother flushing if we got into a different transaction */
2913         if (trans->transid > async->transid)
2914                 goto end;
2915
2916         ret = btrfs_run_delayed_refs(trans, async->count);
2917         if (ret)
2918                 async->error = ret;
2919 end:
2920         ret = btrfs_end_transaction(trans);
2921         if (ret && !async->error)
2922                 async->error = ret;
2923 done:
2924         if (async->sync)
2925                 complete(&async->wait);
2926         else
2927                 kfree(async);
2928 }
2929
2930 int btrfs_async_run_delayed_refs(struct btrfs_fs_info *fs_info,
2931                                  unsigned long count, u64 transid, int wait)
2932 {
2933         struct async_delayed_refs *async;
2934         int ret;
2935
2936         async = kmalloc(sizeof(*async), GFP_NOFS);
2937         if (!async)
2938                 return -ENOMEM;
2939
2940         async->root = fs_info->tree_root;
2941         async->count = count;
2942         async->error = 0;
2943         async->transid = transid;
2944         if (wait)
2945                 async->sync = 1;
2946         else
2947                 async->sync = 0;
2948         init_completion(&async->wait);
2949
2950         btrfs_init_work(&async->work, btrfs_extent_refs_helper,
2951                         delayed_ref_async_start, NULL, NULL);
2952
2953         btrfs_queue_work(fs_info->extent_workers, &async->work);
2954
2955         if (wait) {
2956                 wait_for_completion(&async->wait);
2957                 ret = async->error;
2958                 kfree(async);
2959                 return ret;
2960         }
2961         return 0;
2962 }
2963
2964 /*
2965  * this starts processing the delayed reference count updates and
2966  * extent insertions we have queued up so far.  count can be
2967  * 0, which means to process everything in the tree at the start
2968  * of the run (but not newly added entries), or it can be some target
2969  * number you'd like to process.
2970  *
2971  * Returns 0 on success or if called with an aborted transaction
2972  * Returns <0 on error and aborts the transaction
2973  */
2974 int btrfs_run_delayed_refs(struct btrfs_trans_handle *trans,
2975                            unsigned long count)
2976 {
2977         struct btrfs_fs_info *fs_info = trans->fs_info;
2978         struct rb_node *node;
2979         struct btrfs_delayed_ref_root *delayed_refs;
2980         struct btrfs_delayed_ref_head *head;
2981         int ret;
2982         int run_all = count == (unsigned long)-1;
2983
2984         /* We'll clean this up in btrfs_cleanup_transaction */
2985         if (trans->aborted)
2986                 return 0;
2987
2988         if (test_bit(BTRFS_FS_CREATING_FREE_SPACE_TREE, &fs_info->flags))
2989                 return 0;
2990
2991         delayed_refs = &trans->transaction->delayed_refs;
2992         if (count == 0)
2993                 count = atomic_read(&delayed_refs->num_entries) * 2;
2994
2995 again:
2996 #ifdef SCRAMBLE_DELAYED_REFS
2997         delayed_refs->run_delayed_start = find_middle(&delayed_refs->root);
2998 #endif
2999         ret = __btrfs_run_delayed_refs(trans, count);
3000         if (ret < 0) {
3001                 btrfs_abort_transaction(trans, ret);
3002                 return ret;
3003         }
3004
3005         if (run_all) {
3006                 btrfs_create_pending_block_groups(trans);
3007
3008                 spin_lock(&delayed_refs->lock);
3009                 node = rb_first_cached(&delayed_refs->href_root);
3010                 if (!node) {
3011                         spin_unlock(&delayed_refs->lock);
3012                         goto out;
3013                 }
3014                 head = rb_entry(node, struct btrfs_delayed_ref_head,
3015                                 href_node);
3016                 refcount_inc(&head->refs);
3017                 spin_unlock(&delayed_refs->lock);
3018
3019                 /* Mutex was contended, block until it's released and retry. */
3020                 mutex_lock(&head->mutex);
3021                 mutex_unlock(&head->mutex);
3022
3023                 btrfs_put_delayed_ref_head(head);
3024                 cond_resched();
3025                 goto again;
3026         }
3027 out:
3028         return 0;
3029 }
3030
3031 int btrfs_set_disk_extent_flags(struct btrfs_trans_handle *trans,
3032                                 struct btrfs_fs_info *fs_info,
3033                                 u64 bytenr, u64 num_bytes, u64 flags,
3034                                 int level, int is_data)
3035 {
3036         struct btrfs_delayed_extent_op *extent_op;
3037         int ret;
3038
3039         extent_op = btrfs_alloc_delayed_extent_op();
3040         if (!extent_op)
3041                 return -ENOMEM;
3042
3043         extent_op->flags_to_set = flags;
3044         extent_op->update_flags = true;
3045         extent_op->update_key = false;
3046         extent_op->is_data = is_data ? true : false;
3047         extent_op->level = level;
3048
3049         ret = btrfs_add_delayed_extent_op(fs_info, trans, bytenr,
3050                                           num_bytes, extent_op);
3051         if (ret)
3052                 btrfs_free_delayed_extent_op(extent_op);
3053         return ret;
3054 }
3055
3056 static noinline int check_delayed_ref(struct btrfs_root *root,
3057                                       struct btrfs_path *path,
3058                                       u64 objectid, u64 offset, u64 bytenr)
3059 {
3060         struct btrfs_delayed_ref_head *head;
3061         struct btrfs_delayed_ref_node *ref;
3062         struct btrfs_delayed_data_ref *data_ref;
3063         struct btrfs_delayed_ref_root *delayed_refs;
3064         struct btrfs_transaction *cur_trans;
3065         struct rb_node *node;
3066         int ret = 0;
3067
3068         spin_lock(&root->fs_info->trans_lock);
3069         cur_trans = root->fs_info->running_transaction;
3070         if (cur_trans)
3071                 refcount_inc(&cur_trans->use_count);
3072         spin_unlock(&root->fs_info->trans_lock);
3073         if (!cur_trans)
3074                 return 0;
3075
3076         delayed_refs = &cur_trans->delayed_refs;
3077         spin_lock(&delayed_refs->lock);
3078         head = btrfs_find_delayed_ref_head(delayed_refs, bytenr);
3079         if (!head) {
3080                 spin_unlock(&delayed_refs->lock);
3081                 btrfs_put_transaction(cur_trans);
3082                 return 0;
3083         }
3084
3085         if (!mutex_trylock(&head->mutex)) {
3086                 refcount_inc(&head->refs);
3087                 spin_unlock(&delayed_refs->lock);
3088
3089                 btrfs_release_path(path);
3090
3091                 /*
3092                  * Mutex was contended, block until it's released and let
3093                  * caller try again
3094                  */
3095                 mutex_lock(&head->mutex);
3096                 mutex_unlock(&head->mutex);
3097                 btrfs_put_delayed_ref_head(head);
3098                 btrfs_put_transaction(cur_trans);
3099                 return -EAGAIN;
3100         }
3101         spin_unlock(&delayed_refs->lock);
3102
3103         spin_lock(&head->lock);
3104         /*
3105          * XXX: We should replace this with a proper search function in the
3106          * future.
3107          */
3108         for (node = rb_first_cached(&head->ref_tree); node;
3109              node = rb_next(node)) {
3110                 ref = rb_entry(node, struct btrfs_delayed_ref_node, ref_node);
3111                 /* If it's a shared ref we know a cross reference exists */
3112                 if (ref->type != BTRFS_EXTENT_DATA_REF_KEY) {
3113                         ret = 1;
3114                         break;
3115                 }
3116
3117                 data_ref = btrfs_delayed_node_to_data_ref(ref);
3118
3119                 /*
3120                  * If our ref doesn't match the one we're currently looking at
3121                  * then we have a cross reference.
3122                  */
3123                 if (data_ref->root != root->root_key.objectid ||
3124                     data_ref->objectid != objectid ||
3125                     data_ref->offset != offset) {
3126                         ret = 1;
3127                         break;
3128                 }
3129         }
3130         spin_unlock(&head->lock);
3131         mutex_unlock(&head->mutex);
3132         btrfs_put_transaction(cur_trans);
3133         return ret;
3134 }
3135
3136 static noinline int check_committed_ref(struct btrfs_root *root,
3137                                         struct btrfs_path *path,
3138                                         u64 objectid, u64 offset, u64 bytenr)
3139 {
3140         struct btrfs_fs_info *fs_info = root->fs_info;
3141         struct btrfs_root *extent_root = fs_info->extent_root;
3142         struct extent_buffer *leaf;
3143         struct btrfs_extent_data_ref *ref;
3144         struct btrfs_extent_inline_ref *iref;
3145         struct btrfs_extent_item *ei;
3146         struct btrfs_key key;
3147         u32 item_size;
3148         int type;
3149         int ret;
3150
3151         key.objectid = bytenr;
3152         key.offset = (u64)-1;
3153         key.type = BTRFS_EXTENT_ITEM_KEY;
3154
3155         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
3156         if (ret < 0)
3157                 goto out;
3158         BUG_ON(ret == 0); /* Corruption */
3159
3160         ret = -ENOENT;
3161         if (path->slots[0] == 0)
3162                 goto out;
3163
3164         path->slots[0]--;
3165         leaf = path->nodes[0];
3166         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
3167
3168         if (key.objectid != bytenr || key.type != BTRFS_EXTENT_ITEM_KEY)
3169                 goto out;
3170
3171         ret = 1;
3172         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
3173         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
3174
3175         if (item_size != sizeof(*ei) +
3176             btrfs_extent_inline_ref_size(BTRFS_EXTENT_DATA_REF_KEY))
3177                 goto out;
3178
3179         if (btrfs_extent_generation(leaf, ei) <=
3180             btrfs_root_last_snapshot(&root->root_item))
3181                 goto out;
3182
3183         iref = (struct btrfs_extent_inline_ref *)(ei + 1);
3184
3185         type = btrfs_get_extent_inline_ref_type(leaf, iref, BTRFS_REF_TYPE_DATA);
3186         if (type != BTRFS_EXTENT_DATA_REF_KEY)
3187                 goto out;
3188
3189         ref = (struct btrfs_extent_data_ref *)(&iref->offset);
3190         if (btrfs_extent_refs(leaf, ei) !=
3191             btrfs_extent_data_ref_count(leaf, ref) ||
3192             btrfs_extent_data_ref_root(leaf, ref) !=
3193             root->root_key.objectid ||
3194             btrfs_extent_data_ref_objectid(leaf, ref) != objectid ||
3195             btrfs_extent_data_ref_offset(leaf, ref) != offset)
3196                 goto out;
3197
3198         ret = 0;
3199 out:
3200         return ret;
3201 }
3202
3203 int btrfs_cross_ref_exist(struct btrfs_root *root, u64 objectid, u64 offset,
3204                           u64 bytenr)
3205 {
3206         struct btrfs_path *path;
3207         int ret;
3208
3209         path = btrfs_alloc_path();
3210         if (!path)
3211                 return -ENOMEM;
3212
3213         do {
3214                 ret = check_committed_ref(root, path, objectid,
3215                                           offset, bytenr);
3216                 if (ret && ret != -ENOENT)
3217                         goto out;
3218
3219                 ret = check_delayed_ref(root, path, objectid, offset, bytenr);
3220         } while (ret == -EAGAIN);
3221
3222 out:
3223         btrfs_free_path(path);
3224         if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID)
3225                 WARN_ON(ret > 0);
3226         return ret;
3227 }
3228
3229 static int __btrfs_mod_ref(struct btrfs_trans_handle *trans,
3230                            struct btrfs_root *root,
3231                            struct extent_buffer *buf,
3232                            int full_backref, int inc)
3233 {
3234         struct btrfs_fs_info *fs_info = root->fs_info;
3235         u64 bytenr;
3236         u64 num_bytes;
3237         u64 parent;
3238         u64 ref_root;
3239         u32 nritems;
3240         struct btrfs_key key;
3241         struct btrfs_file_extent_item *fi;
3242         int i;
3243         int level;
3244         int ret = 0;
3245         int (*process_func)(struct btrfs_trans_handle *,
3246                             struct btrfs_root *,
3247                             u64, u64, u64, u64, u64, u64);
3248
3249
3250         if (btrfs_is_testing(fs_info))
3251                 return 0;
3252
3253         ref_root = btrfs_header_owner(buf);
3254         nritems = btrfs_header_nritems(buf);
3255         level = btrfs_header_level(buf);
3256
3257         if (!test_bit(BTRFS_ROOT_REF_COWS, &root->state) && level == 0)
3258                 return 0;
3259
3260         if (inc)
3261                 process_func = btrfs_inc_extent_ref;
3262         else
3263                 process_func = btrfs_free_extent;
3264
3265         if (full_backref)
3266                 parent = buf->start;
3267         else
3268                 parent = 0;
3269
3270         for (i = 0; i < nritems; i++) {
3271                 if (level == 0) {
3272                         btrfs_item_key_to_cpu(buf, &key, i);
3273                         if (key.type != BTRFS_EXTENT_DATA_KEY)
3274                                 continue;
3275                         fi = btrfs_item_ptr(buf, i,
3276                                             struct btrfs_file_extent_item);
3277                         if (btrfs_file_extent_type(buf, fi) ==
3278                             BTRFS_FILE_EXTENT_INLINE)
3279                                 continue;
3280                         bytenr = btrfs_file_extent_disk_bytenr(buf, fi);
3281                         if (bytenr == 0)
3282                                 continue;
3283
3284                         num_bytes = btrfs_file_extent_disk_num_bytes(buf, fi);
3285                         key.offset -= btrfs_file_extent_offset(buf, fi);
3286                         ret = process_func(trans, root, bytenr, num_bytes,
3287                                            parent, ref_root, key.objectid,
3288                                            key.offset);
3289                         if (ret)
3290                                 goto fail;
3291                 } else {
3292                         bytenr = btrfs_node_blockptr(buf, i);
3293                         num_bytes = fs_info->nodesize;
3294                         ret = process_func(trans, root, bytenr, num_bytes,
3295                                            parent, ref_root, level - 1, 0);
3296                         if (ret)
3297                                 goto fail;
3298                 }
3299         }
3300         return 0;
3301 fail:
3302         return ret;
3303 }
3304
3305 int btrfs_inc_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
3306                   struct extent_buffer *buf, int full_backref)
3307 {
3308         return __btrfs_mod_ref(trans, root, buf, full_backref, 1);
3309 }
3310
3311 int btrfs_dec_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
3312                   struct extent_buffer *buf, int full_backref)
3313 {
3314         return __btrfs_mod_ref(trans, root, buf, full_backref, 0);
3315 }
3316
3317 static int write_one_cache_group(struct btrfs_trans_handle *trans,
3318                                  struct btrfs_fs_info *fs_info,
3319                                  struct btrfs_path *path,
3320                                  struct btrfs_block_group_cache *cache)
3321 {
3322         int ret;
3323         struct btrfs_root *extent_root = fs_info->extent_root;
3324         unsigned long bi;
3325         struct extent_buffer *leaf;
3326
3327         ret = btrfs_search_slot(trans, extent_root, &cache->key, path, 0, 1);
3328         if (ret) {
3329                 if (ret > 0)
3330                         ret = -ENOENT;
3331                 goto fail;
3332         }
3333
3334         leaf = path->nodes[0];
3335         bi = btrfs_item_ptr_offset(leaf, path->slots[0]);
3336         write_extent_buffer(leaf, &cache->item, bi, sizeof(cache->item));
3337         btrfs_mark_buffer_dirty(leaf);
3338 fail:
3339         btrfs_release_path(path);
3340         return ret;
3341
3342 }
3343
3344 static struct btrfs_block_group_cache *
3345 next_block_group(struct btrfs_fs_info *fs_info,
3346                  struct btrfs_block_group_cache *cache)
3347 {
3348         struct rb_node *node;
3349
3350         spin_lock(&fs_info->block_group_cache_lock);
3351
3352         /* If our block group was removed, we need a full search. */
3353         if (RB_EMPTY_NODE(&cache->cache_node)) {
3354                 const u64 next_bytenr = cache->key.objectid + cache->key.offset;
3355
3356                 spin_unlock(&fs_info->block_group_cache_lock);
3357                 btrfs_put_block_group(cache);
3358                 cache = btrfs_lookup_first_block_group(fs_info, next_bytenr); return cache;
3359         }
3360         node = rb_next(&cache->cache_node);
3361         btrfs_put_block_group(cache);
3362         if (node) {
3363                 cache = rb_entry(node, struct btrfs_block_group_cache,
3364                                  cache_node);
3365                 btrfs_get_block_group(cache);
3366         } else
3367                 cache = NULL;
3368         spin_unlock(&fs_info->block_group_cache_lock);
3369         return cache;
3370 }
3371
3372 static int cache_save_setup(struct btrfs_block_group_cache *block_group,
3373                             struct btrfs_trans_handle *trans,
3374                             struct btrfs_path *path)
3375 {
3376         struct btrfs_fs_info *fs_info = block_group->fs_info;
3377         struct btrfs_root *root = fs_info->tree_root;
3378         struct inode *inode = NULL;
3379         struct extent_changeset *data_reserved = NULL;
3380         u64 alloc_hint = 0;
3381         int dcs = BTRFS_DC_ERROR;
3382         u64 num_pages = 0;
3383         int retries = 0;
3384         int ret = 0;
3385
3386         /*
3387          * If this block group is smaller than 100 megs don't bother caching the
3388          * block group.
3389          */
3390         if (block_group->key.offset < (100 * SZ_1M)) {
3391                 spin_lock(&block_group->lock);
3392                 block_group->disk_cache_state = BTRFS_DC_WRITTEN;
3393                 spin_unlock(&block_group->lock);
3394                 return 0;
3395         }
3396
3397         if (trans->aborted)
3398                 return 0;
3399 again:
3400         inode = lookup_free_space_inode(fs_info, block_group, path);
3401         if (IS_ERR(inode) && PTR_ERR(inode) != -ENOENT) {
3402                 ret = PTR_ERR(inode);
3403                 btrfs_release_path(path);
3404                 goto out;
3405         }
3406
3407         if (IS_ERR(inode)) {
3408                 BUG_ON(retries);
3409                 retries++;
3410
3411                 if (block_group->ro)
3412                         goto out_free;
3413
3414                 ret = create_free_space_inode(fs_info, trans, block_group,
3415                                               path);
3416                 if (ret)
3417                         goto out_free;
3418                 goto again;
3419         }
3420
3421         /*
3422          * We want to set the generation to 0, that way if anything goes wrong
3423          * from here on out we know not to trust this cache when we load up next
3424          * time.
3425          */
3426         BTRFS_I(inode)->generation = 0;
3427         ret = btrfs_update_inode(trans, root, inode);
3428         if (ret) {
3429                 /*
3430                  * So theoretically we could recover from this, simply set the
3431                  * super cache generation to 0 so we know to invalidate the
3432                  * cache, but then we'd have to keep track of the block groups
3433                  * that fail this way so we know we _have_ to reset this cache
3434                  * before the next commit or risk reading stale cache.  So to
3435                  * limit our exposure to horrible edge cases lets just abort the
3436                  * transaction, this only happens in really bad situations
3437                  * anyway.
3438                  */
3439                 btrfs_abort_transaction(trans, ret);
3440                 goto out_put;
3441         }
3442         WARN_ON(ret);
3443
3444         /* We've already setup this transaction, go ahead and exit */
3445         if (block_group->cache_generation == trans->transid &&
3446             i_size_read(inode)) {
3447                 dcs = BTRFS_DC_SETUP;
3448                 goto out_put;
3449         }
3450
3451         if (i_size_read(inode) > 0) {
3452                 ret = btrfs_check_trunc_cache_free_space(fs_info,
3453                                         &fs_info->global_block_rsv);
3454                 if (ret)
3455                         goto out_put;
3456
3457                 ret = btrfs_truncate_free_space_cache(trans, NULL, inode);
3458                 if (ret)
3459                         goto out_put;
3460         }
3461
3462         spin_lock(&block_group->lock);
3463         if (block_group->cached != BTRFS_CACHE_FINISHED ||
3464             !btrfs_test_opt(fs_info, SPACE_CACHE)) {
3465                 /*
3466                  * don't bother trying to write stuff out _if_
3467                  * a) we're not cached,
3468                  * b) we're with nospace_cache mount option,
3469                  * c) we're with v2 space_cache (FREE_SPACE_TREE).
3470                  */
3471                 dcs = BTRFS_DC_WRITTEN;
3472                 spin_unlock(&block_group->lock);
3473                 goto out_put;
3474         }
3475         spin_unlock(&block_group->lock);
3476
3477         /*
3478          * We hit an ENOSPC when setting up the cache in this transaction, just
3479          * skip doing the setup, we've already cleared the cache so we're safe.
3480          */
3481         if (test_bit(BTRFS_TRANS_CACHE_ENOSPC, &trans->transaction->flags)) {
3482                 ret = -ENOSPC;
3483                 goto out_put;
3484         }
3485
3486         /*
3487          * Try to preallocate enough space based on how big the block group is.
3488          * Keep in mind this has to include any pinned space which could end up
3489          * taking up quite a bit since it's not folded into the other space
3490          * cache.
3491          */
3492         num_pages = div_u64(block_group->key.offset, SZ_256M);
3493         if (!num_pages)
3494                 num_pages = 1;
3495
3496         num_pages *= 16;
3497         num_pages *= PAGE_SIZE;
3498
3499         ret = btrfs_check_data_free_space(inode, &data_reserved, 0, num_pages);
3500         if (ret)
3501                 goto out_put;
3502
3503         ret = btrfs_prealloc_file_range_trans(inode, trans, 0, 0, num_pages,
3504                                               num_pages, num_pages,
3505                                               &alloc_hint);
3506         /*
3507          * Our cache requires contiguous chunks so that we don't modify a bunch
3508          * of metadata or split extents when writing the cache out, which means
3509          * we can enospc if we are heavily fragmented in addition to just normal
3510          * out of space conditions.  So if we hit this just skip setting up any
3511          * other block groups for this transaction, maybe we'll unpin enough
3512          * space the next time around.
3513          */
3514         if (!ret)
3515                 dcs = BTRFS_DC_SETUP;
3516         else if (ret == -ENOSPC)
3517                 set_bit(BTRFS_TRANS_CACHE_ENOSPC, &trans->transaction->flags);
3518
3519 out_put:
3520         iput(inode);
3521 out_free:
3522         btrfs_release_path(path);
3523 out:
3524         spin_lock(&block_group->lock);
3525         if (!ret && dcs == BTRFS_DC_SETUP)
3526                 block_group->cache_generation = trans->transid;
3527         block_group->disk_cache_state = dcs;
3528         spin_unlock(&block_group->lock);
3529
3530         extent_changeset_free(data_reserved);
3531         return ret;
3532 }
3533
3534 int btrfs_setup_space_cache(struct btrfs_trans_handle *trans,
3535                             struct btrfs_fs_info *fs_info)
3536 {
3537         struct btrfs_block_group_cache *cache, *tmp;
3538         struct btrfs_transaction *cur_trans = trans->transaction;
3539         struct btrfs_path *path;
3540
3541         if (list_empty(&cur_trans->dirty_bgs) ||
3542             !btrfs_test_opt(fs_info, SPACE_CACHE))
3543                 return 0;
3544
3545         path = btrfs_alloc_path();
3546         if (!path)
3547                 return -ENOMEM;
3548
3549         /* Could add new block groups, use _safe just in case */
3550         list_for_each_entry_safe(cache, tmp, &cur_trans->dirty_bgs,
3551                                  dirty_list) {
3552                 if (cache->disk_cache_state == BTRFS_DC_CLEAR)
3553                         cache_save_setup(cache, trans, path);
3554         }
3555
3556         btrfs_free_path(path);
3557         return 0;
3558 }
3559
3560 /*
3561  * transaction commit does final block group cache writeback during a
3562  * critical section where nothing is allowed to change the FS.  This is
3563  * required in order for the cache to actually match the block group,
3564  * but can introduce a lot of latency into the commit.
3565  *
3566  * So, btrfs_start_dirty_block_groups is here to kick off block group
3567  * cache IO.  There's a chance we'll have to redo some of it if the
3568  * block group changes again during the commit, but it greatly reduces
3569  * the commit latency by getting rid of the easy block groups while
3570  * we're still allowing others to join the commit.
3571  */
3572 int btrfs_start_dirty_block_groups(struct btrfs_trans_handle *trans)
3573 {
3574         struct btrfs_fs_info *fs_info = trans->fs_info;
3575         struct btrfs_block_group_cache *cache;
3576         struct btrfs_transaction *cur_trans = trans->transaction;
3577         int ret = 0;
3578         int should_put;
3579         struct btrfs_path *path = NULL;
3580         LIST_HEAD(dirty);
3581         struct list_head *io = &cur_trans->io_bgs;
3582         int num_started = 0;
3583         int loops = 0;
3584
3585         spin_lock(&cur_trans->dirty_bgs_lock);
3586         if (list_empty(&cur_trans->dirty_bgs)) {
3587                 spin_unlock(&cur_trans->dirty_bgs_lock);
3588                 return 0;
3589         }
3590         list_splice_init(&cur_trans->dirty_bgs, &dirty);
3591         spin_unlock(&cur_trans->dirty_bgs_lock);
3592
3593 again:
3594         /*
3595          * make sure all the block groups on our dirty list actually
3596          * exist
3597          */
3598         btrfs_create_pending_block_groups(trans);
3599
3600         if (!path) {
3601                 path = btrfs_alloc_path();
3602                 if (!path)
3603                         return -ENOMEM;
3604         }
3605
3606         /*
3607          * cache_write_mutex is here only to save us from balance or automatic
3608          * removal of empty block groups deleting this block group while we are
3609          * writing out the cache
3610          */
3611         mutex_lock(&trans->transaction->cache_write_mutex);
3612         while (!list_empty(&dirty)) {
3613                 bool drop_reserve = true;
3614
3615                 cache = list_first_entry(&dirty,
3616                                          struct btrfs_block_group_cache,
3617                                          dirty_list);
3618                 /*
3619                  * this can happen if something re-dirties a block
3620                  * group that is already under IO.  Just wait for it to
3621                  * finish and then do it all again
3622                  */
3623                 if (!list_empty(&cache->io_list)) {
3624                         list_del_init(&cache->io_list);
3625                         btrfs_wait_cache_io(trans, cache, path);
3626                         btrfs_put_block_group(cache);
3627                 }
3628
3629
3630                 /*
3631                  * btrfs_wait_cache_io uses the cache->dirty_list to decide
3632                  * if it should update the cache_state.  Don't delete
3633                  * until after we wait.
3634                  *
3635                  * Since we're not running in the commit critical section
3636                  * we need the dirty_bgs_lock to protect from update_block_group
3637                  */
3638                 spin_lock(&cur_trans->dirty_bgs_lock);
3639                 list_del_init(&cache->dirty_list);
3640                 spin_unlock(&cur_trans->dirty_bgs_lock);
3641
3642                 should_put = 1;
3643
3644                 cache_save_setup(cache, trans, path);
3645
3646                 if (cache->disk_cache_state == BTRFS_DC_SETUP) {
3647                         cache->io_ctl.inode = NULL;
3648                         ret = btrfs_write_out_cache(fs_info, trans,
3649                                                     cache, path);
3650                         if (ret == 0 && cache->io_ctl.inode) {
3651                                 num_started++;
3652                                 should_put = 0;
3653
3654                                 /*
3655                                  * The cache_write_mutex is protecting the
3656                                  * io_list, also refer to the definition of
3657                                  * btrfs_transaction::io_bgs for more details
3658                                  */
3659                                 list_add_tail(&cache->io_list, io);
3660                         } else {
3661                                 /*
3662                                  * if we failed to write the cache, the
3663                                  * generation will be bad and life goes on
3664                                  */
3665                                 ret = 0;
3666                         }
3667                 }
3668                 if (!ret) {
3669                         ret = write_one_cache_group(trans, fs_info,
3670                                                     path, cache);
3671                         /*
3672                          * Our block group might still be attached to the list
3673                          * of new block groups in the transaction handle of some
3674                          * other task (struct btrfs_trans_handle->new_bgs). This
3675                          * means its block group item isn't yet in the extent
3676                          * tree. If this happens ignore the error, as we will
3677                          * try again later in the critical section of the
3678                          * transaction commit.
3679                          */
3680                         if (ret == -ENOENT) {
3681                                 ret = 0;
3682                                 spin_lock(&cur_trans->dirty_bgs_lock);
3683                                 if (list_empty(&cache->dirty_list)) {
3684                                         list_add_tail(&cache->dirty_list,
3685                                                       &cur_trans->dirty_bgs);
3686                                         btrfs_get_block_group(cache);
3687                                         drop_reserve = false;
3688                                 }
3689                                 spin_unlock(&cur_trans->dirty_bgs_lock);
3690                         } else if (ret) {
3691                                 btrfs_abort_transaction(trans, ret);
3692                         }
3693                 }
3694
3695                 /* if it's not on the io list, we need to put the block group */
3696                 if (should_put)
3697                         btrfs_put_block_group(cache);
3698                 if (drop_reserve)
3699                         btrfs_delayed_refs_rsv_release(fs_info, 1);
3700
3701                 if (ret)
3702                         break;
3703
3704                 /*
3705                  * Avoid blocking other tasks for too long. It might even save
3706                  * us from writing caches for block groups that are going to be
3707                  * removed.
3708                  */
3709                 mutex_unlock(&trans->transaction->cache_write_mutex);
3710                 mutex_lock(&trans->transaction->cache_write_mutex);
3711         }
3712         mutex_unlock(&trans->transaction->cache_write_mutex);
3713
3714         /*
3715          * go through delayed refs for all the stuff we've just kicked off
3716          * and then loop back (just once)
3717          */
3718         ret = btrfs_run_delayed_refs(trans, 0);
3719         if (!ret && loops == 0) {
3720                 loops++;
3721                 spin_lock(&cur_trans->dirty_bgs_lock);
3722                 list_splice_init(&cur_trans->dirty_bgs, &dirty);
3723                 /*
3724                  * dirty_bgs_lock protects us from concurrent block group
3725                  * deletes too (not just cache_write_mutex).
3726                  */
3727                 if (!list_empty(&dirty)) {
3728                         spin_unlock(&cur_trans->dirty_bgs_lock);
3729                         goto again;
3730                 }
3731                 spin_unlock(&cur_trans->dirty_bgs_lock);
3732         } else if (ret < 0) {
3733                 btrfs_cleanup_dirty_bgs(cur_trans, fs_info);
3734         }
3735
3736         btrfs_free_path(path);
3737         return ret;
3738 }
3739
3740 int btrfs_write_dirty_block_groups(struct btrfs_trans_handle *trans,
3741                                    struct btrfs_fs_info *fs_info)
3742 {
3743         struct btrfs_block_group_cache *cache;
3744         struct btrfs_transaction *cur_trans = trans->transaction;
3745         int ret = 0;
3746         int should_put;
3747         struct btrfs_path *path;
3748         struct list_head *io = &cur_trans->io_bgs;
3749         int num_started = 0;
3750
3751         path = btrfs_alloc_path();
3752         if (!path)
3753                 return -ENOMEM;
3754
3755         /*
3756          * Even though we are in the critical section of the transaction commit,
3757          * we can still have concurrent tasks adding elements to this
3758          * transaction's list of dirty block groups. These tasks correspond to
3759          * endio free space workers started when writeback finishes for a
3760          * space cache, which run inode.c:btrfs_finish_ordered_io(), and can
3761          * allocate new block groups as a result of COWing nodes of the root
3762          * tree when updating the free space inode. The writeback for the space
3763          * caches is triggered by an earlier call to
3764          * btrfs_start_dirty_block_groups() and iterations of the following
3765          * loop.
3766          * Also we want to do the cache_save_setup first and then run the
3767          * delayed refs to make sure we have the best chance at doing this all
3768          * in one shot.
3769          */
3770         spin_lock(&cur_trans->dirty_bgs_lock);
3771         while (!list_empty(&cur_trans->dirty_bgs)) {
3772                 cache = list_first_entry(&cur_trans->dirty_bgs,
3773                                          struct btrfs_block_group_cache,
3774                                          dirty_list);
3775
3776                 /*
3777                  * this can happen if cache_save_setup re-dirties a block
3778                  * group that is already under IO.  Just wait for it to
3779                  * finish and then do it all again
3780                  */
3781                 if (!list_empty(&cache->io_list)) {
3782                         spin_unlock(&cur_trans->dirty_bgs_lock);
3783                         list_del_init(&cache->io_list);
3784                         btrfs_wait_cache_io(trans, cache, path);
3785                         btrfs_put_block_group(cache);
3786                         spin_lock(&cur_trans->dirty_bgs_lock);
3787                 }
3788
3789                 /*
3790                  * don't remove from the dirty list until after we've waited
3791                  * on any pending IO
3792                  */
3793                 list_del_init(&cache->dirty_list);
3794                 spin_unlock(&cur_trans->dirty_bgs_lock);
3795                 should_put = 1;
3796
3797                 cache_save_setup(cache, trans, path);
3798
3799                 if (!ret)
3800                         ret = btrfs_run_delayed_refs(trans,
3801                                                      (unsigned long) -1);
3802
3803                 if (!ret && cache->disk_cache_state == BTRFS_DC_SETUP) {
3804                         cache->io_ctl.inode = NULL;
3805                         ret = btrfs_write_out_cache(fs_info, trans,
3806                                                     cache, path);
3807                         if (ret == 0 && cache->io_ctl.inode) {
3808                                 num_started++;
3809                                 should_put = 0;
3810                                 list_add_tail(&cache->io_list, io);
3811                         } else {
3812                                 /*
3813                                  * if we failed to write the cache, the
3814                                  * generation will be bad and life goes on
3815                                  */
3816                                 ret = 0;
3817                         }
3818                 }
3819                 if (!ret) {
3820                         ret = write_one_cache_group(trans, fs_info,
3821                                                     path, cache);
3822                         /*
3823                          * One of the free space endio workers might have
3824                          * created a new block group while updating a free space
3825                          * cache's inode (at inode.c:btrfs_finish_ordered_io())
3826                          * and hasn't released its transaction handle yet, in
3827                          * which case the new block group is still attached to
3828                          * its transaction handle and its creation has not
3829                          * finished yet (no block group item in the extent tree
3830                          * yet, etc). If this is the case, wait for all free
3831                          * space endio workers to finish and retry. This is a
3832                          * a very rare case so no need for a more efficient and
3833                          * complex approach.
3834                          */
3835                         if (ret == -ENOENT) {
3836                                 wait_event(cur_trans->writer_wait,
3837                                    atomic_read(&cur_trans->num_writers) == 1);
3838                                 ret = write_one_cache_group(trans, fs_info,
3839                                                             path, cache);
3840                         }
3841                         if (ret)
3842                                 btrfs_abort_transaction(trans, ret);
3843                 }
3844
3845                 /* if its not on the io list, we need to put the block group */
3846                 if (should_put)
3847                         btrfs_put_block_group(cache);
3848                 btrfs_delayed_refs_rsv_release(fs_info, 1);
3849                 spin_lock(&cur_trans->dirty_bgs_lock);
3850         }
3851         spin_unlock(&cur_trans->dirty_bgs_lock);
3852
3853         /*
3854          * Refer to the definition of io_bgs member for details why it's safe
3855          * to use it without any locking
3856          */
3857         while (!list_empty(io)) {
3858                 cache = list_first_entry(io, struct btrfs_block_group_cache,
3859                                          io_list);
3860                 list_del_init(&cache->io_list);
3861                 btrfs_wait_cache_io(trans, cache, path);
3862                 btrfs_put_block_group(cache);
3863         }
3864
3865         btrfs_free_path(path);
3866         return ret;
3867 }
3868
3869 int btrfs_extent_readonly(struct btrfs_fs_info *fs_info, u64 bytenr)
3870 {
3871         struct btrfs_block_group_cache *block_group;
3872         int readonly = 0;
3873
3874         block_group = btrfs_lookup_block_group(fs_info, bytenr);
3875         if (!block_group || block_group->ro)
3876                 readonly = 1;
3877         if (block_group)
3878                 btrfs_put_block_group(block_group);
3879         return readonly;
3880 }
3881
3882 bool btrfs_inc_nocow_writers(struct btrfs_fs_info *fs_info, u64 bytenr)
3883 {
3884         struct btrfs_block_group_cache *bg;
3885         bool ret = true;
3886
3887         bg = btrfs_lookup_block_group(fs_info, bytenr);
3888         if (!bg)
3889                 return false;
3890
3891         spin_lock(&bg->lock);
3892         if (bg->ro)
3893                 ret = false;
3894         else
3895                 atomic_inc(&bg->nocow_writers);
3896         spin_unlock(&bg->lock);
3897
3898         /* no put on block group, done by btrfs_dec_nocow_writers */
3899         if (!ret)
3900                 btrfs_put_block_group(bg);
3901
3902         return ret;
3903
3904 }
3905
3906 void btrfs_dec_nocow_writers(struct btrfs_fs_info *fs_info, u64 bytenr)
3907 {
3908         struct btrfs_block_group_cache *bg;
3909
3910         bg = btrfs_lookup_block_group(fs_info, bytenr);
3911         ASSERT(bg);
3912         if (atomic_dec_and_test(&bg->nocow_writers))
3913                 wake_up_var(&bg->nocow_writers);
3914         /*
3915          * Once for our lookup and once for the lookup done by a previous call
3916          * to btrfs_inc_nocow_writers()
3917          */
3918         btrfs_put_block_group(bg);
3919         btrfs_put_block_group(bg);
3920 }
3921
3922 void btrfs_wait_nocow_writers(struct btrfs_block_group_cache *bg)
3923 {
3924         wait_var_event(&bg->nocow_writers, !atomic_read(&bg->nocow_writers));
3925 }
3926
3927 static const char *alloc_name(u64 flags)
3928 {
3929         switch (flags) {
3930         case BTRFS_BLOCK_GROUP_METADATA|BTRFS_BLOCK_GROUP_DATA:
3931                 return "mixed";
3932         case BTRFS_BLOCK_GROUP_METADATA:
3933                 return "metadata";
3934         case BTRFS_BLOCK_GROUP_DATA:
3935                 return "data";
3936         case BTRFS_BLOCK_GROUP_SYSTEM:
3937                 return "system";
3938         default:
3939                 WARN_ON(1);
3940                 return "invalid-combination";
3941         };
3942 }
3943
3944 static int create_space_info(struct btrfs_fs_info *info, u64 flags)
3945 {
3946
3947         struct btrfs_space_info *space_info;
3948         int i;
3949         int ret;
3950
3951         space_info = kzalloc(sizeof(*space_info), GFP_NOFS);
3952         if (!space_info)
3953                 return -ENOMEM;
3954
3955         ret = percpu_counter_init(&space_info->total_bytes_pinned, 0,
3956                                  GFP_KERNEL);
3957         if (ret) {
3958                 kfree(space_info);
3959                 return ret;
3960         }
3961
3962         for (i = 0; i < BTRFS_NR_RAID_TYPES; i++)
3963                 INIT_LIST_HEAD(&space_info->block_groups[i]);
3964         init_rwsem(&space_info->groups_sem);
3965         spin_lock_init(&space_info->lock);
3966         space_info->flags = flags & BTRFS_BLOCK_GROUP_TYPE_MASK;
3967         space_info->force_alloc = CHUNK_ALLOC_NO_FORCE;
3968         init_waitqueue_head(&space_info->wait);
3969         INIT_LIST_HEAD(&space_info->ro_bgs);
3970         INIT_LIST_HEAD(&space_info->tickets);
3971         INIT_LIST_HEAD(&space_info->priority_tickets);
3972
3973         ret = kobject_init_and_add(&space_info->kobj, &space_info_ktype,
3974                                     info->space_info_kobj, "%s",
3975                                     alloc_name(space_info->flags));
3976         if (ret) {
3977                 percpu_counter_destroy(&space_info->total_bytes_pinned);
3978                 kfree(space_info);
3979                 return ret;
3980         }
3981
3982         list_add_rcu(&space_info->list, &info->space_info);
3983         if (flags & BTRFS_BLOCK_GROUP_DATA)
3984                 info->data_sinfo = space_info;
3985
3986         return ret;
3987 }
3988
3989 static void update_space_info(struct btrfs_fs_info *info, u64 flags,
3990                              u64 total_bytes, u64 bytes_used,
3991                              u64 bytes_readonly,
3992                              struct btrfs_space_info **space_info)
3993 {
3994         struct btrfs_space_info *found;
3995         int factor;
3996
3997         factor = btrfs_bg_type_to_factor(flags);
3998
3999         found = __find_space_info(info, flags);
4000         ASSERT(found);
4001         spin_lock(&found->lock);
4002         found->total_bytes += total_bytes;
4003         found->disk_total += total_bytes * factor;
4004         found->bytes_used += bytes_used;
4005         found->disk_used += bytes_used * factor;
4006         found->bytes_readonly += bytes_readonly;
4007         if (total_bytes > 0)
4008                 found->full = 0;
4009         space_info_add_new_bytes(info, found, total_bytes -
4010                                  bytes_used - bytes_readonly);
4011         spin_unlock(&found->lock);
4012         *space_info = found;
4013 }
4014
4015 static void set_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
4016 {
4017         u64 extra_flags = chunk_to_extended(flags) &
4018                                 BTRFS_EXTENDED_PROFILE_MASK;
4019
4020         write_seqlock(&fs_info->profiles_lock);
4021         if (flags & BTRFS_BLOCK_GROUP_DATA)
4022                 fs_info->avail_data_alloc_bits |= extra_flags;
4023         if (flags & BTRFS_BLOCK_GROUP_METADATA)
4024                 fs_info->avail_metadata_alloc_bits |= extra_flags;
4025         if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
4026                 fs_info->avail_system_alloc_bits |= extra_flags;
4027         write_sequnlock(&fs_info->profiles_lock);
4028 }
4029
4030 /*
4031  * returns target flags in extended format or 0 if restripe for this
4032  * chunk_type is not in progress
4033  *
4034  * should be called with balance_lock held
4035  */
4036 static u64 get_restripe_target(struct btrfs_fs_info *fs_info, u64 flags)
4037 {
4038         struct btrfs_balance_control *bctl = fs_info->balance_ctl;
4039         u64 target = 0;
4040
4041         if (!bctl)
4042                 return 0;
4043
4044         if (flags & BTRFS_BLOCK_GROUP_DATA &&
4045             bctl->data.flags & BTRFS_BALANCE_ARGS_CONVERT) {
4046                 target = BTRFS_BLOCK_GROUP_DATA | bctl->data.target;
4047         } else if (flags & BTRFS_BLOCK_GROUP_SYSTEM &&
4048                    bctl->sys.flags & BTRFS_BALANCE_ARGS_CONVERT) {
4049                 target = BTRFS_BLOCK_GROUP_SYSTEM | bctl->sys.target;
4050         } else if (flags & BTRFS_BLOCK_GROUP_METADATA &&
4051                    bctl->meta.flags & BTRFS_BALANCE_ARGS_CONVERT) {
4052                 target = BTRFS_BLOCK_GROUP_METADATA | bctl->meta.target;
4053         }
4054
4055         return target;
4056 }
4057
4058 /*
4059  * @flags: available profiles in extended format (see ctree.h)
4060  *
4061  * Returns reduced profile in chunk format.  If profile changing is in
4062  * progress (either running or paused) picks the target profile (if it's
4063  * already available), otherwise falls back to plain reducing.
4064  */
4065 static u64 btrfs_reduce_alloc_profile(struct btrfs_fs_info *fs_info, u64 flags)
4066 {
4067         u64 num_devices = fs_info->fs_devices->rw_devices;
4068         u64 target;
4069         u64 raid_type;
4070         u64 allowed = 0;
4071
4072         /*
4073          * see if restripe for this chunk_type is in progress, if so
4074          * try to reduce to the target profile
4075          */
4076         spin_lock(&fs_info->balance_lock);
4077         target = get_restripe_target(fs_info, flags);
4078         if (target) {
4079                 /* pick target profile only if it's already available */
4080                 if ((flags & target) & BTRFS_EXTENDED_PROFILE_MASK) {
4081                         spin_unlock(&fs_info->balance_lock);
4082                         return extended_to_chunk(target);
4083                 }
4084         }
4085         spin_unlock(&fs_info->balance_lock);
4086
4087         /* First, mask out the RAID levels which aren't possible */
4088         for (raid_type = 0; raid_type < BTRFS_NR_RAID_TYPES; raid_type++) {
4089                 if (num_devices >= btrfs_raid_array[raid_type].devs_min)
4090                         allowed |= btrfs_raid_array[raid_type].bg_flag;
4091         }
4092         allowed &= flags;
4093
4094         if (allowed & BTRFS_BLOCK_GROUP_RAID6)
4095                 allowed = BTRFS_BLOCK_GROUP_RAID6;
4096         else if (allowed & BTRFS_BLOCK_GROUP_RAID5)
4097                 allowed = BTRFS_BLOCK_GROUP_RAID5;
4098         else if (allowed & BTRFS_BLOCK_GROUP_RAID10)
4099                 allowed = BTRFS_BLOCK_GROUP_RAID10;
4100         else if (allowed & BTRFS_BLOCK_GROUP_RAID1)
4101                 allowed = BTRFS_BLOCK_GROUP_RAID1;
4102         else if (allowed & BTRFS_BLOCK_GROUP_RAID0)
4103                 allowed = BTRFS_BLOCK_GROUP_RAID0;
4104
4105         flags &= ~BTRFS_BLOCK_GROUP_PROFILE_MASK;
4106
4107         return extended_to_chunk(flags | allowed);
4108 }
4109
4110 static u64 get_alloc_profile(struct btrfs_fs_info *fs_info, u64 orig_flags)
4111 {
4112         unsigned seq;
4113         u64 flags;
4114
4115         do {
4116                 flags = orig_flags;
4117                 seq = read_seqbegin(&fs_info->profiles_lock);
4118
4119                 if (flags & BTRFS_BLOCK_GROUP_DATA)
4120                         flags |= fs_info->avail_data_alloc_bits;
4121                 else if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
4122                         flags |= fs_info->avail_system_alloc_bits;
4123                 else if (flags & BTRFS_BLOCK_GROUP_METADATA)
4124                         flags |= fs_info->avail_metadata_alloc_bits;
4125         } while (read_seqretry(&fs_info->profiles_lock, seq));
4126
4127         return btrfs_reduce_alloc_profile(fs_info, flags);
4128 }
4129
4130 static u64 get_alloc_profile_by_root(struct btrfs_root *root, int data)
4131 {
4132         struct btrfs_fs_info *fs_info = root->fs_info;
4133         u64 flags;
4134         u64 ret;
4135
4136         if (data)
4137                 flags = BTRFS_BLOCK_GROUP_DATA;
4138         else if (root == fs_info->chunk_root)
4139                 flags = BTRFS_BLOCK_GROUP_SYSTEM;
4140         else
4141                 flags = BTRFS_BLOCK_GROUP_METADATA;
4142
4143         ret = get_alloc_profile(fs_info, flags);
4144         return ret;
4145 }
4146
4147 u64 btrfs_data_alloc_profile(struct btrfs_fs_info *fs_info)
4148 {
4149         return get_alloc_profile(fs_info, BTRFS_BLOCK_GROUP_DATA);
4150 }
4151
4152 u64 btrfs_metadata_alloc_profile(struct btrfs_fs_info *fs_info)
4153 {
4154         return get_alloc_profile(fs_info, BTRFS_BLOCK_GROUP_METADATA);
4155 }
4156
4157 u64 btrfs_system_alloc_profile(struct btrfs_fs_info *fs_info)
4158 {
4159         return get_alloc_profile(fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
4160 }
4161
4162 static u64 btrfs_space_info_used(struct btrfs_space_info *s_info,
4163                                  bool may_use_included)
4164 {
4165         ASSERT(s_info);
4166         return s_info->bytes_used + s_info->bytes_reserved +
4167                 s_info->bytes_pinned + s_info->bytes_readonly +
4168                 (may_use_included ? s_info->bytes_may_use : 0);
4169 }
4170
4171 int btrfs_alloc_data_chunk_ondemand(struct btrfs_inode *inode, u64 bytes)
4172 {
4173         struct btrfs_root *root = inode->root;
4174         struct btrfs_fs_info *fs_info = root->fs_info;
4175         struct btrfs_space_info *data_sinfo = fs_info->data_sinfo;
4176         u64 used;
4177         int ret = 0;
4178         int need_commit = 2;
4179         int have_pinned_space;
4180
4181         /* make sure bytes are sectorsize aligned */
4182         bytes = ALIGN(bytes, fs_info->sectorsize);
4183
4184         if (btrfs_is_free_space_inode(inode)) {
4185                 need_commit = 0;
4186                 ASSERT(current->journal_info);
4187         }
4188
4189 again:
4190         /* make sure we have enough space to handle the data first */
4191         spin_lock(&data_sinfo->lock);
4192         used = btrfs_space_info_used(data_sinfo, true);
4193
4194         if (used + bytes > data_sinfo->total_bytes) {
4195                 struct btrfs_trans_handle *trans;
4196
4197                 /*
4198                  * if we don't have enough free bytes in this space then we need
4199                  * to alloc a new chunk.
4200                  */
4201                 if (!data_sinfo->full) {
4202                         u64 alloc_target;
4203
4204                         data_sinfo->force_alloc = CHUNK_ALLOC_FORCE;
4205                         spin_unlock(&data_sinfo->lock);
4206
4207                         alloc_target = btrfs_data_alloc_profile(fs_info);
4208                         /*
4209                          * It is ugly that we don't call nolock join
4210                          * transaction for the free space inode case here.
4211                          * But it is safe because we only do the data space
4212                          * reservation for the free space cache in the
4213                          * transaction context, the common join transaction
4214                          * just increase the counter of the current transaction
4215                          * handler, doesn't try to acquire the trans_lock of
4216                          * the fs.
4217                          */
4218                         trans = btrfs_join_transaction(root);
4219                         if (IS_ERR(trans))
4220                                 return PTR_ERR(trans);
4221
4222                         ret = do_chunk_alloc(trans, alloc_target,
4223                                              CHUNK_ALLOC_NO_FORCE);
4224                         btrfs_end_transaction(trans);
4225                         if (ret < 0) {
4226                                 if (ret != -ENOSPC)
4227                                         return ret;
4228                                 else {
4229                                         have_pinned_space = 1;
4230                                         goto commit_trans;
4231                                 }
4232                         }
4233
4234                         goto again;
4235                 }
4236
4237                 /*
4238                  * If we don't have enough pinned space to deal with this
4239                  * allocation, and no removed chunk in current transaction,
4240                  * don't bother committing the transaction.
4241                  */
4242                 have_pinned_space = __percpu_counter_compare(
4243                         &data_sinfo->total_bytes_pinned,
4244                         used + bytes - data_sinfo->total_bytes,
4245                         BTRFS_TOTAL_BYTES_PINNED_BATCH);
4246                 spin_unlock(&data_sinfo->lock);
4247
4248                 /* commit the current transaction and try again */
4249 commit_trans:
4250                 if (need_commit) {
4251                         need_commit--;
4252
4253                         if (need_commit > 0) {
4254                                 btrfs_start_delalloc_roots(fs_info, -1);
4255                                 btrfs_wait_ordered_roots(fs_info, U64_MAX, 0,
4256                                                          (u64)-1);
4257                         }
4258
4259                         trans = btrfs_join_transaction(root);
4260                         if (IS_ERR(trans))
4261                                 return PTR_ERR(trans);
4262                         if (have_pinned_space >= 0 ||
4263                             test_bit(BTRFS_TRANS_HAVE_FREE_BGS,
4264                                      &trans->transaction->flags) ||
4265                             need_commit > 0) {
4266                                 ret = btrfs_commit_transaction(trans);
4267                                 if (ret)
4268                                         return ret;
4269                                 /*
4270                                  * The cleaner kthread might still be doing iput
4271                                  * operations. Wait for it to finish so that
4272                                  * more space is released.  We don't need to
4273                                  * explicitly run the delayed iputs here because
4274                                  * the commit_transaction would have woken up
4275                                  * the cleaner.
4276                                  */
4277                                 ret = btrfs_wait_on_delayed_iputs(fs_info);
4278                                 if (ret)
4279                                         return ret;
4280                                 goto again;
4281                         } else {
4282                                 btrfs_end_transaction(trans);
4283                         }
4284                 }
4285
4286                 trace_btrfs_space_reservation(fs_info,
4287                                               "space_info:enospc",
4288                                               data_sinfo->flags, bytes, 1);
4289                 return -ENOSPC;
4290         }
4291         update_bytes_may_use(data_sinfo, bytes);
4292         trace_btrfs_space_reservation(fs_info, "space_info",
4293                                       data_sinfo->flags, bytes, 1);
4294         spin_unlock(&data_sinfo->lock);
4295
4296         return 0;
4297 }
4298
4299 int btrfs_check_data_free_space(struct inode *inode,
4300                         struct extent_changeset **reserved, u64 start, u64 len)
4301 {
4302         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
4303         int ret;
4304
4305         /* align the range */
4306         len = round_up(start + len, fs_info->sectorsize) -
4307               round_down(start, fs_info->sectorsize);
4308         start = round_down(start, fs_info->sectorsize);
4309
4310         ret = btrfs_alloc_data_chunk_ondemand(BTRFS_I(inode), len);
4311         if (ret < 0)
4312                 return ret;
4313
4314         /* Use new btrfs_qgroup_reserve_data to reserve precious data space. */
4315         ret = btrfs_qgroup_reserve_data(inode, reserved, start, len);
4316         if (ret < 0)
4317                 btrfs_free_reserved_data_space_noquota(inode, start, len);
4318         else
4319                 ret = 0;
4320         return ret;
4321 }
4322
4323 /*
4324  * Called if we need to clear a data reservation for this inode
4325  * Normally in a error case.
4326  *
4327  * This one will *NOT* use accurate qgroup reserved space API, just for case
4328  * which we can't sleep and is sure it won't affect qgroup reserved space.
4329  * Like clear_bit_hook().
4330  */
4331 void btrfs_free_reserved_data_space_noquota(struct inode *inode, u64 start,
4332                                             u64 len)
4333 {
4334         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
4335         struct btrfs_space_info *data_sinfo;
4336
4337         /* Make sure the range is aligned to sectorsize */
4338         len = round_up(start + len, fs_info->sectorsize) -
4339               round_down(start, fs_info->sectorsize);
4340         start = round_down(start, fs_info->sectorsize);
4341
4342         data_sinfo = fs_info->data_sinfo;
4343         spin_lock(&data_sinfo->lock);
4344         update_bytes_may_use(data_sinfo, -len);
4345         trace_btrfs_space_reservation(fs_info, "space_info",
4346                                       data_sinfo->flags, len, 0);
4347         spin_unlock(&data_sinfo->lock);
4348 }
4349
4350 /*
4351  * Called if we need to clear a data reservation for this inode
4352  * Normally in a error case.
4353  *
4354  * This one will handle the per-inode data rsv map for accurate reserved
4355  * space framework.
4356  */
4357 void btrfs_free_reserved_data_space(struct inode *inode,
4358                         struct extent_changeset *reserved, u64 start, u64 len)
4359 {
4360         struct btrfs_root *root = BTRFS_I(inode)->root;
4361
4362         /* Make sure the range is aligned to sectorsize */
4363         len = round_up(start + len, root->fs_info->sectorsize) -
4364               round_down(start, root->fs_info->sectorsize);
4365         start = round_down(start, root->fs_info->sectorsize);
4366
4367         btrfs_free_reserved_data_space_noquota(inode, start, len);
4368         btrfs_qgroup_free_data(inode, reserved, start, len);
4369 }
4370
4371 static void force_metadata_allocation(struct btrfs_fs_info *info)
4372 {
4373         struct list_head *head = &info->space_info;
4374         struct btrfs_space_info *found;
4375
4376         rcu_read_lock();
4377         list_for_each_entry_rcu(found, head, list) {
4378                 if (found->flags & BTRFS_BLOCK_GROUP_METADATA)
4379                         found->force_alloc = CHUNK_ALLOC_FORCE;
4380         }
4381         rcu_read_unlock();
4382 }
4383
4384 static inline u64 calc_global_rsv_need_space(struct btrfs_block_rsv *global)
4385 {
4386         return (global->size << 1);
4387 }
4388
4389 static int should_alloc_chunk(struct btrfs_fs_info *fs_info,
4390                               struct btrfs_space_info *sinfo, int force)
4391 {
4392         u64 bytes_used = btrfs_space_info_used(sinfo, false);
4393         u64 thresh;
4394
4395         if (force == CHUNK_ALLOC_FORCE)
4396                 return 1;
4397
4398         /*
4399          * in limited mode, we want to have some free space up to
4400          * about 1% of the FS size.
4401          */
4402         if (force == CHUNK_ALLOC_LIMITED) {
4403                 thresh = btrfs_super_total_bytes(fs_info->super_copy);
4404                 thresh = max_t(u64, SZ_64M, div_factor_fine(thresh, 1));
4405
4406                 if (sinfo->total_bytes - bytes_used < thresh)
4407                         return 1;
4408         }
4409
4410         if (bytes_used + SZ_2M < div_factor(sinfo->total_bytes, 8))
4411                 return 0;
4412         return 1;
4413 }
4414
4415 static u64 get_profile_num_devs(struct btrfs_fs_info *fs_info, u64 type)
4416 {
4417         u64 num_dev;
4418
4419         if (type & (BTRFS_BLOCK_GROUP_RAID10 |
4420                     BTRFS_BLOCK_GROUP_RAID0 |
4421                     BTRFS_BLOCK_GROUP_RAID5 |
4422                     BTRFS_BLOCK_GROUP_RAID6))
4423                 num_dev = fs_info->fs_devices->rw_devices;
4424         else if (type & BTRFS_BLOCK_GROUP_RAID1)
4425                 num_dev = 2;
4426         else
4427                 num_dev = 1;    /* DUP or single */
4428
4429         return num_dev;
4430 }
4431
4432 /*
4433  * If @is_allocation is true, reserve space in the system space info necessary
4434  * for allocating a chunk, otherwise if it's false, reserve space necessary for
4435  * removing a chunk.
4436  */
4437 void check_system_chunk(struct btrfs_trans_handle *trans, u64 type)
4438 {
4439         struct btrfs_fs_info *fs_info = trans->fs_info;
4440         struct btrfs_space_info *info;
4441         u64 left;
4442         u64 thresh;
4443         int ret = 0;
4444         u64 num_devs;
4445
4446         /*
4447          * Needed because we can end up allocating a system chunk and for an
4448          * atomic and race free space reservation in the chunk block reserve.
4449          */
4450         lockdep_assert_held(&fs_info->chunk_mutex);
4451
4452         info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
4453         spin_lock(&info->lock);
4454         left = info->total_bytes - btrfs_space_info_used(info, true);
4455         spin_unlock(&info->lock);
4456
4457         num_devs = get_profile_num_devs(fs_info, type);
4458
4459         /* num_devs device items to update and 1 chunk item to add or remove */
4460         thresh = btrfs_calc_trunc_metadata_size(fs_info, num_devs) +
4461                 btrfs_calc_trans_metadata_size(fs_info, 1);
4462
4463         if (left < thresh && btrfs_test_opt(fs_info, ENOSPC_DEBUG)) {
4464                 btrfs_info(fs_info, "left=%llu, need=%llu, flags=%llu",
4465                            left, thresh, type);
4466                 dump_space_info(fs_info, info, 0, 0);
4467         }
4468
4469         if (left < thresh) {
4470                 u64 flags = btrfs_system_alloc_profile(fs_info);
4471
4472                 /*
4473                  * Ignore failure to create system chunk. We might end up not
4474                  * needing it, as we might not need to COW all nodes/leafs from
4475                  * the paths we visit in the chunk tree (they were already COWed
4476                  * or created in the current transaction for example).
4477                  */
4478                 ret = btrfs_alloc_chunk(trans, flags);
4479         }
4480
4481         if (!ret) {
4482                 ret = btrfs_block_rsv_add(fs_info->chunk_root,
4483                                           &fs_info->chunk_block_rsv,
4484                                           thresh, BTRFS_RESERVE_NO_FLUSH);
4485                 if (!ret)
4486                         trans->chunk_bytes_reserved += thresh;
4487         }
4488 }
4489
4490 /*
4491  * If force is CHUNK_ALLOC_FORCE:
4492  *    - return 1 if it successfully allocates a chunk,
4493  *    - return errors including -ENOSPC otherwise.
4494  * If force is NOT CHUNK_ALLOC_FORCE:
4495  *    - return 0 if it doesn't need to allocate a new chunk,
4496  *    - return 1 if it successfully allocates a chunk,
4497  *    - return errors including -ENOSPC otherwise.
4498  */
4499 static int do_chunk_alloc(struct btrfs_trans_handle *trans, u64 flags,
4500                           int force)
4501 {
4502         struct btrfs_fs_info *fs_info = trans->fs_info;
4503         struct btrfs_space_info *space_info;
4504         bool wait_for_alloc = false;
4505         bool should_alloc = false;
4506         int ret = 0;
4507
4508         /* Don't re-enter if we're already allocating a chunk */
4509         if (trans->allocating_chunk)
4510                 return -ENOSPC;
4511
4512         space_info = __find_space_info(fs_info, flags);
4513         ASSERT(space_info);
4514
4515         do {
4516                 spin_lock(&space_info->lock);
4517                 if (force < space_info->force_alloc)
4518                         force = space_info->force_alloc;
4519                 should_alloc = should_alloc_chunk(fs_info, space_info, force);
4520                 if (space_info->full) {
4521                         /* No more free physical space */
4522                         if (should_alloc)
4523                                 ret = -ENOSPC;
4524                         else
4525                                 ret = 0;
4526                         spin_unlock(&space_info->lock);
4527                         return ret;
4528                 } else if (!should_alloc) {
4529                         spin_unlock(&space_info->lock);
4530                         return 0;
4531                 } else if (space_info->chunk_alloc) {
4532                         /*
4533                          * Someone is already allocating, so we need to block
4534                          * until this someone is finished and then loop to
4535                          * recheck if we should continue with our allocation
4536                          * attempt.
4537                          */
4538                         wait_for_alloc = true;
4539                         spin_unlock(&space_info->lock);
4540                         mutex_lock(&fs_info->chunk_mutex);
4541                         mutex_unlock(&fs_info->chunk_mutex);
4542                 } else {
4543                         /* Proceed with allocation */
4544                         space_info->chunk_alloc = 1;
4545                         wait_for_alloc = false;
4546                         spin_unlock(&space_info->lock);
4547                 }
4548
4549                 cond_resched();
4550         } while (wait_for_alloc);
4551
4552         mutex_lock(&fs_info->chunk_mutex);
4553         trans->allocating_chunk = true;
4554
4555         /*
4556          * If we have mixed data/metadata chunks we want to make sure we keep
4557          * allocating mixed chunks instead of individual chunks.
4558          */
4559         if (btrfs_mixed_space_info(space_info))
4560                 flags |= (BTRFS_BLOCK_GROUP_DATA | BTRFS_BLOCK_GROUP_METADATA);
4561
4562         /*
4563          * if we're doing a data chunk, go ahead and make sure that
4564          * we keep a reasonable number of metadata chunks allocated in the
4565          * FS as well.
4566          */
4567         if (flags & BTRFS_BLOCK_GROUP_DATA && fs_info->metadata_ratio) {
4568                 fs_info->data_chunk_allocations++;
4569                 if (!(fs_info->data_chunk_allocations %
4570                       fs_info->metadata_ratio))
4571                         force_metadata_allocation(fs_info);
4572         }
4573
4574         /*
4575          * Check if we have enough space in SYSTEM chunk because we may need
4576          * to update devices.
4577          */
4578         check_system_chunk(trans, flags);
4579
4580         ret = btrfs_alloc_chunk(trans, flags);
4581         trans->allocating_chunk = false;
4582
4583         spin_lock(&space_info->lock);
4584         if (ret < 0) {
4585                 if (ret == -ENOSPC)
4586                         space_info->full = 1;
4587                 else
4588                         goto out;
4589         } else {
4590                 ret = 1;
4591                 space_info->max_extent_size = 0;
4592         }
4593
4594         space_info->force_alloc = CHUNK_ALLOC_NO_FORCE;
4595 out:
4596         space_info->chunk_alloc = 0;
4597         spin_unlock(&space_info->lock);
4598         mutex_unlock(&fs_info->chunk_mutex);
4599         /*
4600          * When we allocate a new chunk we reserve space in the chunk block
4601          * reserve to make sure we can COW nodes/leafs in the chunk tree or
4602          * add new nodes/leafs to it if we end up needing to do it when
4603          * inserting the chunk item and updating device items as part of the
4604          * second phase of chunk allocation, performed by
4605          * btrfs_finish_chunk_alloc(). So make sure we don't accumulate a
4606          * large number of new block groups to create in our transaction
4607          * handle's new_bgs list to avoid exhausting the chunk block reserve
4608          * in extreme cases - like having a single transaction create many new
4609          * block groups when starting to write out the free space caches of all
4610          * the block groups that were made dirty during the lifetime of the
4611          * transaction.
4612          */
4613         if (trans->chunk_bytes_reserved >= (u64)SZ_2M)
4614                 btrfs_create_pending_block_groups(trans);
4615
4616         return ret;
4617 }
4618
4619 static int can_overcommit(struct btrfs_fs_info *fs_info,
4620                           struct btrfs_space_info *space_info, u64 bytes,
4621                           enum btrfs_reserve_flush_enum flush,
4622                           bool system_chunk)
4623 {
4624         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
4625         u64 profile;
4626         u64 space_size;
4627         u64 avail;
4628         u64 used;
4629         int factor;
4630
4631         /* Don't overcommit when in mixed mode. */
4632         if (space_info->flags & BTRFS_BLOCK_GROUP_DATA)
4633                 return 0;
4634
4635         if (system_chunk)
4636                 profile = btrfs_system_alloc_profile(fs_info);
4637         else
4638                 profile = btrfs_metadata_alloc_profile(fs_info);
4639
4640         used = btrfs_space_info_used(space_info, false);
4641
4642         /*
4643          * We only want to allow over committing if we have lots of actual space
4644          * free, but if we don't have enough space to handle the global reserve
4645          * space then we could end up having a real enospc problem when trying
4646          * to allocate a chunk or some other such important allocation.
4647          */
4648         spin_lock(&global_rsv->lock);
4649         space_size = calc_global_rsv_need_space(global_rsv);
4650         spin_unlock(&global_rsv->lock);
4651         if (used + space_size >= space_info->total_bytes)
4652                 return 0;
4653
4654         used += space_info->bytes_may_use;
4655
4656         avail = atomic64_read(&fs_info->free_chunk_space);
4657
4658         /*
4659          * If we have dup, raid1 or raid10 then only half of the free
4660          * space is actually usable.  For raid56, the space info used
4661          * doesn't include the parity drive, so we don't have to
4662          * change the math
4663          */
4664         factor = btrfs_bg_type_to_factor(profile);
4665         avail = div_u64(avail, factor);
4666
4667         /*
4668          * If we aren't flushing all things, let us overcommit up to
4669          * 1/2th of the space. If we can flush, don't let us overcommit
4670          * too much, let it overcommit up to 1/8 of the space.
4671          */
4672         if (flush == BTRFS_RESERVE_FLUSH_ALL)
4673                 avail >>= 3;
4674         else
4675                 avail >>= 1;
4676
4677         if (used + bytes < space_info->total_bytes + avail)
4678                 return 1;
4679         return 0;
4680 }
4681
4682 static void btrfs_writeback_inodes_sb_nr(struct btrfs_fs_info *fs_info,
4683                                          unsigned long nr_pages, int nr_items)
4684 {
4685         struct super_block *sb = fs_info->sb;
4686
4687         if (down_read_trylock(&sb->s_umount)) {
4688                 writeback_inodes_sb_nr(sb, nr_pages, WB_REASON_FS_FREE_SPACE);
4689                 up_read(&sb->s_umount);
4690         } else {
4691                 /*
4692                  * We needn't worry the filesystem going from r/w to r/o though
4693                  * we don't acquire ->s_umount mutex, because the filesystem
4694                  * should guarantee the delalloc inodes list be empty after
4695                  * the filesystem is readonly(all dirty pages are written to
4696                  * the disk).
4697                  */
4698                 btrfs_start_delalloc_roots(fs_info, nr_items);
4699                 if (!current->journal_info)
4700                         btrfs_wait_ordered_roots(fs_info, nr_items, 0, (u64)-1);
4701         }
4702 }
4703
4704 static inline u64 calc_reclaim_items_nr(struct btrfs_fs_info *fs_info,
4705                                         u64 to_reclaim)
4706 {
4707         u64 bytes;
4708         u64 nr;
4709
4710         bytes = btrfs_calc_trans_metadata_size(fs_info, 1);
4711         nr = div64_u64(to_reclaim, bytes);
4712         if (!nr)
4713                 nr = 1;
4714         return nr;
4715 }
4716
4717 #define EXTENT_SIZE_PER_ITEM    SZ_256K
4718
4719 /*
4720  * shrink metadata reservation for delalloc
4721  */
4722 static void shrink_delalloc(struct btrfs_fs_info *fs_info, u64 to_reclaim,
4723                             u64 orig, bool wait_ordered)
4724 {
4725         struct btrfs_space_info *space_info;
4726         struct btrfs_trans_handle *trans;
4727         u64 delalloc_bytes;
4728         u64 async_pages;
4729         u64 items;
4730         long time_left;
4731         unsigned long nr_pages;
4732         int loops;
4733
4734         /* Calc the number of the pages we need flush for space reservation */
4735         items = calc_reclaim_items_nr(fs_info, to_reclaim);
4736         to_reclaim = items * EXTENT_SIZE_PER_ITEM;
4737
4738         trans = (struct btrfs_trans_handle *)current->journal_info;
4739         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
4740
4741         delalloc_bytes = percpu_counter_sum_positive(
4742                                                 &fs_info->delalloc_bytes);
4743         if (delalloc_bytes == 0) {
4744                 if (trans)
4745                         return;
4746                 if (wait_ordered)
4747                         btrfs_wait_ordered_roots(fs_info, items, 0, (u64)-1);
4748                 return;
4749         }
4750
4751         loops = 0;
4752         while (delalloc_bytes && loops < 3) {
4753                 nr_pages = min(delalloc_bytes, to_reclaim) >> PAGE_SHIFT;
4754
4755                 /*
4756                  * Triggers inode writeback for up to nr_pages. This will invoke
4757                  * ->writepages callback and trigger delalloc filling
4758                  *  (btrfs_run_delalloc_range()).
4759                  */
4760                 btrfs_writeback_inodes_sb_nr(fs_info, nr_pages, items);
4761
4762                 /*
4763                  * We need to wait for the compressed pages to start before
4764                  * we continue.
4765                  */
4766                 async_pages = atomic_read(&fs_info->async_delalloc_pages);
4767                 if (!async_pages)
4768                         goto skip_async;
4769
4770                 /*
4771                  * Calculate how many compressed pages we want to be written
4772                  * before we continue. I.e if there are more async pages than we
4773                  * require wait_event will wait until nr_pages are written.
4774                  */
4775                 if (async_pages <= nr_pages)
4776                         async_pages = 0;
4777                 else
4778                         async_pages -= nr_pages;
4779
4780                 wait_event(fs_info->async_submit_wait,
4781                            atomic_read(&fs_info->async_delalloc_pages) <=
4782                            (int)async_pages);
4783 skip_async:
4784                 spin_lock(&space_info->lock);
4785                 if (list_empty(&space_info->tickets) &&
4786                     list_empty(&space_info->priority_tickets)) {
4787                         spin_unlock(&space_info->lock);
4788                         break;
4789                 }
4790                 spin_unlock(&space_info->lock);
4791
4792                 loops++;
4793                 if (wait_ordered && !trans) {
4794                         btrfs_wait_ordered_roots(fs_info, items, 0, (u64)-1);
4795                 } else {
4796                         time_left = schedule_timeout_killable(1);
4797                         if (time_left)
4798                                 break;
4799                 }
4800                 delalloc_bytes = percpu_counter_sum_positive(
4801                                                 &fs_info->delalloc_bytes);
4802         }
4803 }
4804
4805 struct reserve_ticket {
4806         u64 orig_bytes;
4807         u64 bytes;
4808         int error;
4809         struct list_head list;
4810         wait_queue_head_t wait;
4811 };
4812
4813 /**
4814  * maybe_commit_transaction - possibly commit the transaction if its ok to
4815  * @root - the root we're allocating for
4816  * @bytes - the number of bytes we want to reserve
4817  * @force - force the commit
4818  *
4819  * This will check to make sure that committing the transaction will actually
4820  * get us somewhere and then commit the transaction if it does.  Otherwise it
4821  * will return -ENOSPC.
4822  */
4823 static int may_commit_transaction(struct btrfs_fs_info *fs_info,
4824                                   struct btrfs_space_info *space_info)
4825 {
4826         struct reserve_ticket *ticket = NULL;
4827         struct btrfs_block_rsv *delayed_rsv = &fs_info->delayed_block_rsv;
4828         struct btrfs_block_rsv *delayed_refs_rsv = &fs_info->delayed_refs_rsv;
4829         struct btrfs_trans_handle *trans;
4830         u64 bytes_needed;
4831         u64 reclaim_bytes = 0;
4832
4833         trans = (struct btrfs_trans_handle *)current->journal_info;
4834         if (trans)
4835                 return -EAGAIN;
4836
4837         spin_lock(&space_info->lock);
4838         if (!list_empty(&space_info->priority_tickets))
4839                 ticket = list_first_entry(&space_info->priority_tickets,
4840                                           struct reserve_ticket, list);
4841         else if (!list_empty(&space_info->tickets))
4842                 ticket = list_first_entry(&space_info->tickets,
4843                                           struct reserve_ticket, list);
4844         bytes_needed = (ticket) ? ticket->bytes : 0;
4845         spin_unlock(&space_info->lock);
4846
4847         if (!bytes_needed)
4848                 return 0;
4849
4850         trans = btrfs_join_transaction(fs_info->extent_root);
4851         if (IS_ERR(trans))
4852                 return PTR_ERR(trans);
4853
4854         /*
4855          * See if there is enough pinned space to make this reservation, or if
4856          * we have block groups that are going to be freed, allowing us to
4857          * possibly do a chunk allocation the next loop through.
4858          */
4859         if (test_bit(BTRFS_TRANS_HAVE_FREE_BGS, &trans->transaction->flags) ||
4860             __percpu_counter_compare(&space_info->total_bytes_pinned,
4861                                      bytes_needed,
4862                                      BTRFS_TOTAL_BYTES_PINNED_BATCH) >= 0)
4863                 goto commit;
4864
4865         /*
4866          * See if there is some space in the delayed insertion reservation for
4867          * this reservation.
4868          */
4869         if (space_info != delayed_rsv->space_info)
4870                 goto enospc;
4871
4872         spin_lock(&delayed_rsv->lock);
4873         reclaim_bytes += delayed_rsv->reserved;
4874         spin_unlock(&delayed_rsv->lock);
4875
4876         spin_lock(&delayed_refs_rsv->lock);
4877         reclaim_bytes += delayed_refs_rsv->reserved;
4878         spin_unlock(&delayed_refs_rsv->lock);
4879         if (reclaim_bytes >= bytes_needed)
4880                 goto commit;
4881         bytes_needed -= reclaim_bytes;
4882
4883         if (__percpu_counter_compare(&space_info->total_bytes_pinned,
4884                                    bytes_needed,
4885                                    BTRFS_TOTAL_BYTES_PINNED_BATCH) < 0)
4886                 goto enospc;
4887
4888 commit:
4889         return btrfs_commit_transaction(trans);
4890 enospc:
4891         btrfs_end_transaction(trans);
4892         return -ENOSPC;
4893 }
4894
4895 /*
4896  * Try to flush some data based on policy set by @state. This is only advisory
4897  * and may fail for various reasons. The caller is supposed to examine the
4898  * state of @space_info to detect the outcome.
4899  */
4900 static void flush_space(struct btrfs_fs_info *fs_info,
4901                        struct btrfs_space_info *space_info, u64 num_bytes,
4902                        int state)
4903 {
4904         struct btrfs_root *root = fs_info->extent_root;
4905         struct btrfs_trans_handle *trans;
4906         int nr;
4907         int ret = 0;
4908
4909         switch (state) {
4910         case FLUSH_DELAYED_ITEMS_NR:
4911         case FLUSH_DELAYED_ITEMS:
4912                 if (state == FLUSH_DELAYED_ITEMS_NR)
4913                         nr = calc_reclaim_items_nr(fs_info, num_bytes) * 2;
4914                 else
4915                         nr = -1;
4916
4917                 trans = btrfs_join_transaction(root);
4918                 if (IS_ERR(trans)) {
4919                         ret = PTR_ERR(trans);
4920                         break;
4921                 }
4922                 ret = btrfs_run_delayed_items_nr(trans, nr);
4923                 btrfs_end_transaction(trans);
4924                 break;
4925         case FLUSH_DELALLOC:
4926         case FLUSH_DELALLOC_WAIT:
4927                 shrink_delalloc(fs_info, num_bytes * 2, num_bytes,
4928                                 state == FLUSH_DELALLOC_WAIT);
4929                 break;
4930         case FLUSH_DELAYED_REFS_NR:
4931         case FLUSH_DELAYED_REFS:
4932                 trans = btrfs_join_transaction(root);
4933                 if (IS_ERR(trans)) {
4934                         ret = PTR_ERR(trans);
4935                         break;
4936                 }
4937                 if (state == FLUSH_DELAYED_REFS_NR)
4938                         nr = calc_reclaim_items_nr(fs_info, num_bytes);
4939                 else
4940                         nr = 0;
4941                 btrfs_run_delayed_refs(trans, nr);
4942                 btrfs_end_transaction(trans);
4943                 break;
4944         case ALLOC_CHUNK:
4945         case ALLOC_CHUNK_FORCE:
4946                 trans = btrfs_join_transaction(root);
4947                 if (IS_ERR(trans)) {
4948                         ret = PTR_ERR(trans);
4949                         break;
4950                 }
4951                 ret = do_chunk_alloc(trans,
4952                                      btrfs_metadata_alloc_profile(fs_info),
4953                                      (state == ALLOC_CHUNK) ?
4954                                       CHUNK_ALLOC_NO_FORCE : CHUNK_ALLOC_FORCE);
4955                 btrfs_end_transaction(trans);
4956                 if (ret > 0 || ret == -ENOSPC)
4957                         ret = 0;
4958                 break;
4959         case COMMIT_TRANS:
4960                 /*
4961                  * If we have pending delayed iputs then we could free up a
4962                  * bunch of pinned space, so make sure we run the iputs before
4963                  * we do our pinned bytes check below.
4964                  */
4965                 btrfs_run_delayed_iputs(fs_info);
4966                 btrfs_wait_on_delayed_iputs(fs_info);
4967
4968                 ret = may_commit_transaction(fs_info, space_info);
4969                 break;
4970         default:
4971                 ret = -ENOSPC;
4972                 break;
4973         }
4974
4975         trace_btrfs_flush_space(fs_info, space_info->flags, num_bytes, state,
4976                                 ret);
4977         return;
4978 }
4979
4980 static inline u64
4981 btrfs_calc_reclaim_metadata_size(struct btrfs_fs_info *fs_info,
4982                                  struct btrfs_space_info *space_info,
4983                                  bool system_chunk)
4984 {
4985         struct reserve_ticket *ticket;
4986         u64 used;
4987         u64 expected;
4988         u64 to_reclaim = 0;
4989
4990         list_for_each_entry(ticket, &space_info->tickets, list)
4991                 to_reclaim += ticket->bytes;
4992         list_for_each_entry(ticket, &space_info->priority_tickets, list)
4993                 to_reclaim += ticket->bytes;
4994         if (to_reclaim)
4995                 return to_reclaim;
4996
4997         to_reclaim = min_t(u64, num_online_cpus() * SZ_1M, SZ_16M);
4998         if (can_overcommit(fs_info, space_info, to_reclaim,
4999                            BTRFS_RESERVE_FLUSH_ALL, system_chunk))
5000                 return 0;
5001
5002         used = btrfs_space_info_used(space_info, true);
5003
5004         if (can_overcommit(fs_info, space_info, SZ_1M,
5005                            BTRFS_RESERVE_FLUSH_ALL, system_chunk))
5006                 expected = div_factor_fine(space_info->total_bytes, 95);
5007         else
5008                 expected = div_factor_fine(space_info->total_bytes, 90);
5009
5010         if (used > expected)
5011                 to_reclaim = used - expected;
5012         else
5013                 to_reclaim = 0;
5014         to_reclaim = min(to_reclaim, space_info->bytes_may_use +
5015                                      space_info->bytes_reserved);
5016         return to_reclaim;
5017 }
5018
5019 static inline int need_do_async_reclaim(struct btrfs_fs_info *fs_info,
5020                                         struct btrfs_space_info *space_info,
5021                                         u64 used, bool system_chunk)
5022 {
5023         u64 thresh = div_factor_fine(space_info->total_bytes, 98);
5024
5025         /* If we're just plain full then async reclaim just slows us down. */
5026         if ((space_info->bytes_used + space_info->bytes_reserved) >= thresh)
5027                 return 0;
5028
5029         if (!btrfs_calc_reclaim_metadata_size(fs_info, space_info,
5030                                               system_chunk))
5031                 return 0;
5032
5033         return (used >= thresh && !btrfs_fs_closing(fs_info) &&
5034                 !test_bit(BTRFS_FS_STATE_REMOUNTING, &fs_info->fs_state));
5035 }
5036
5037 static bool wake_all_tickets(struct list_head *head)
5038 {
5039         struct reserve_ticket *ticket;
5040
5041         while (!list_empty(head)) {
5042                 ticket = list_first_entry(head, struct reserve_ticket, list);
5043                 list_del_init(&ticket->list);
5044                 ticket->error = -ENOSPC;
5045                 wake_up(&ticket->wait);
5046                 if (ticket->bytes != ticket->orig_bytes)
5047                         return true;
5048         }
5049         return false;
5050 }
5051
5052 /*
5053  * This is for normal flushers, we can wait all goddamned day if we want to.  We
5054  * will loop and continuously try to flush as long as we are making progress.
5055  * We count progress as clearing off tickets each time we have to loop.
5056  */
5057 static void btrfs_async_reclaim_metadata_space(struct work_struct *work)
5058 {
5059         struct btrfs_fs_info *fs_info;
5060         struct btrfs_space_info *space_info;
5061         u64 to_reclaim;
5062         int flush_state;
5063         int commit_cycles = 0;
5064         u64 last_tickets_id;
5065
5066         fs_info = container_of(work, struct btrfs_fs_info, async_reclaim_work);
5067         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
5068
5069         spin_lock(&space_info->lock);
5070         to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info, space_info,
5071                                                       false);
5072         if (!to_reclaim) {
5073                 space_info->flush = 0;
5074                 spin_unlock(&space_info->lock);
5075                 return;
5076         }
5077         last_tickets_id = space_info->tickets_id;
5078         spin_unlock(&space_info->lock);
5079
5080         flush_state = FLUSH_DELAYED_ITEMS_NR;
5081         do {
5082                 flush_space(fs_info, space_info, to_reclaim, flush_state);
5083                 spin_lock(&space_info->lock);
5084                 if (list_empty(&space_info->tickets)) {
5085                         space_info->flush = 0;
5086                         spin_unlock(&space_info->lock);
5087                         return;
5088                 }
5089                 to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info,
5090                                                               space_info,
5091                                                               false);
5092                 if (last_tickets_id == space_info->tickets_id) {
5093                         flush_state++;
5094                 } else {
5095                         last_tickets_id = space_info->tickets_id;
5096                         flush_state = FLUSH_DELAYED_ITEMS_NR;
5097                         if (commit_cycles)
5098                                 commit_cycles--;
5099                 }
5100
5101                 /*
5102                  * We don't want to force a chunk allocation until we've tried
5103                  * pretty hard to reclaim space.  Think of the case where we
5104                  * freed up a bunch of space and so have a lot of pinned space
5105                  * to reclaim.  We would rather use that than possibly create a
5106                  * underutilized metadata chunk.  So if this is our first run
5107                  * through the flushing state machine skip ALLOC_CHUNK_FORCE and
5108                  * commit the transaction.  If nothing has changed the next go
5109                  * around then we can force a chunk allocation.
5110                  */
5111                 if (flush_state == ALLOC_CHUNK_FORCE && !commit_cycles)
5112                         flush_state++;
5113
5114                 if (flush_state > COMMIT_TRANS) {
5115                         commit_cycles++;
5116                         if (commit_cycles > 2) {
5117                                 if (wake_all_tickets(&space_info->tickets)) {
5118                                         flush_state = FLUSH_DELAYED_ITEMS_NR;
5119                                         commit_cycles--;
5120                                 } else {
5121                                         space_info->flush = 0;
5122                                 }
5123                         } else {
5124                                 flush_state = FLUSH_DELAYED_ITEMS_NR;
5125                         }
5126                 }
5127                 spin_unlock(&space_info->lock);
5128         } while (flush_state <= COMMIT_TRANS);
5129 }
5130
5131 void btrfs_init_async_reclaim_work(struct work_struct *work)
5132 {
5133         INIT_WORK(work, btrfs_async_reclaim_metadata_space);
5134 }
5135
5136 static const enum btrfs_flush_state priority_flush_states[] = {
5137         FLUSH_DELAYED_ITEMS_NR,
5138         FLUSH_DELAYED_ITEMS,
5139         ALLOC_CHUNK,
5140 };
5141
5142 static void priority_reclaim_metadata_space(struct btrfs_fs_info *fs_info,
5143                                             struct btrfs_space_info *space_info,
5144                                             struct reserve_ticket *ticket)
5145 {
5146         u64 to_reclaim;
5147         int flush_state;
5148
5149         spin_lock(&space_info->lock);
5150         to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info, space_info,
5151                                                       false);
5152         if (!to_reclaim) {
5153                 spin_unlock(&space_info->lock);
5154                 return;
5155         }
5156         spin_unlock(&space_info->lock);
5157
5158         flush_state = 0;
5159         do {
5160                 flush_space(fs_info, space_info, to_reclaim,
5161                             priority_flush_states[flush_state]);
5162                 flush_state++;
5163                 spin_lock(&space_info->lock);
5164                 if (ticket->bytes == 0) {
5165                         spin_unlock(&space_info->lock);
5166                         return;
5167                 }
5168                 spin_unlock(&space_info->lock);
5169         } while (flush_state < ARRAY_SIZE(priority_flush_states));
5170 }
5171
5172 static int wait_reserve_ticket(struct btrfs_fs_info *fs_info,
5173                                struct btrfs_space_info *space_info,
5174                                struct reserve_ticket *ticket)
5175
5176 {
5177         DEFINE_WAIT(wait);
5178         u64 reclaim_bytes = 0;
5179         int ret = 0;
5180
5181         spin_lock(&space_info->lock);
5182         while (ticket->bytes > 0 && ticket->error == 0) {
5183                 ret = prepare_to_wait_event(&ticket->wait, &wait, TASK_KILLABLE);
5184                 if (ret) {
5185                         ret = -EINTR;
5186                         break;
5187                 }
5188                 spin_unlock(&space_info->lock);
5189
5190                 schedule();
5191
5192                 finish_wait(&ticket->wait, &wait);
5193                 spin_lock(&space_info->lock);
5194         }
5195         if (!ret)
5196                 ret = ticket->error;
5197         if (!list_empty(&ticket->list))
5198                 list_del_init(&ticket->list);
5199         if (ticket->bytes && ticket->bytes < ticket->orig_bytes)
5200                 reclaim_bytes = ticket->orig_bytes - ticket->bytes;
5201         spin_unlock(&space_info->lock);
5202
5203         if (reclaim_bytes)
5204                 space_info_add_old_bytes(fs_info, space_info, reclaim_bytes);
5205         return ret;
5206 }
5207
5208 /**
5209  * reserve_metadata_bytes - try to reserve bytes from the block_rsv's space
5210  * @root - the root we're allocating for
5211  * @space_info - the space info we want to allocate from
5212  * @orig_bytes - the number of bytes we want
5213  * @flush - whether or not we can flush to make our reservation
5214  *
5215  * This will reserve orig_bytes number of bytes from the space info associated
5216  * with the block_rsv.  If there is not enough space it will make an attempt to
5217  * flush out space to make room.  It will do this by flushing delalloc if
5218  * possible or committing the transaction.  If flush is 0 then no attempts to
5219  * regain reservations will be made and this will fail if there is not enough
5220  * space already.
5221  */
5222 static int __reserve_metadata_bytes(struct btrfs_fs_info *fs_info,
5223                                     struct btrfs_space_info *space_info,
5224                                     u64 orig_bytes,
5225                                     enum btrfs_reserve_flush_enum flush,
5226                                     bool system_chunk)
5227 {
5228         struct reserve_ticket ticket;
5229         u64 used;
5230         u64 reclaim_bytes = 0;
5231         int ret = 0;
5232
5233         ASSERT(orig_bytes);
5234         ASSERT(!current->journal_info || flush != BTRFS_RESERVE_FLUSH_ALL);
5235
5236         spin_lock(&space_info->lock);
5237         ret = -ENOSPC;
5238         used = btrfs_space_info_used(space_info, true);
5239
5240         /*
5241          * If we have enough space then hooray, make our reservation and carry
5242          * on.  If not see if we can overcommit, and if we can, hooray carry on.
5243          * If not things get more complicated.
5244          */
5245         if (used + orig_bytes <= space_info->total_bytes) {
5246                 update_bytes_may_use(space_info, orig_bytes);
5247                 trace_btrfs_space_reservation(fs_info, "space_info",
5248                                               space_info->flags, orig_bytes, 1);
5249                 ret = 0;
5250         } else if (can_overcommit(fs_info, space_info, orig_bytes, flush,
5251                                   system_chunk)) {
5252                 update_bytes_may_use(space_info, orig_bytes);
5253                 trace_btrfs_space_reservation(fs_info, "space_info",
5254                                               space_info->flags, orig_bytes, 1);
5255                 ret = 0;
5256         }
5257
5258         /*
5259          * If we couldn't make a reservation then setup our reservation ticket
5260          * and kick the async worker if it's not already running.
5261          *
5262          * If we are a priority flusher then we just need to add our ticket to
5263          * the list and we will do our own flushing further down.
5264          */
5265         if (ret && flush != BTRFS_RESERVE_NO_FLUSH) {
5266                 ticket.orig_bytes = orig_bytes;
5267                 ticket.bytes = orig_bytes;
5268                 ticket.error = 0;
5269                 init_waitqueue_head(&ticket.wait);
5270                 if (flush == BTRFS_RESERVE_FLUSH_ALL) {
5271                         list_add_tail(&ticket.list, &space_info->tickets);
5272                         if (!space_info->flush) {
5273                                 space_info->flush = 1;
5274                                 trace_btrfs_trigger_flush(fs_info,
5275                                                           space_info->flags,
5276                                                           orig_bytes, flush,
5277                                                           "enospc");
5278                                 queue_work(system_unbound_wq,
5279                                            &fs_info->async_reclaim_work);
5280                         }
5281                 } else {
5282                         list_add_tail(&ticket.list,
5283                                       &space_info->priority_tickets);
5284                 }
5285         } else if (!ret && space_info->flags & BTRFS_BLOCK_GROUP_METADATA) {
5286                 used += orig_bytes;
5287                 /*
5288                  * We will do the space reservation dance during log replay,
5289                  * which means we won't have fs_info->fs_root set, so don't do
5290                  * the async reclaim as we will panic.
5291                  */
5292                 if (!test_bit(BTRFS_FS_LOG_RECOVERING, &fs_info->flags) &&
5293                     need_do_async_reclaim(fs_info, space_info,
5294                                           used, system_chunk) &&
5295                     !work_busy(&fs_info->async_reclaim_work)) {
5296                         trace_btrfs_trigger_flush(fs_info, space_info->flags,
5297                                                   orig_bytes, flush, "preempt");
5298                         queue_work(system_unbound_wq,
5299                                    &fs_info->async_reclaim_work);
5300                 }
5301         }
5302         spin_unlock(&space_info->lock);
5303         if (!ret || flush == BTRFS_RESERVE_NO_FLUSH)
5304                 return ret;
5305
5306         if (flush == BTRFS_RESERVE_FLUSH_ALL)
5307                 return wait_reserve_ticket(fs_info, space_info, &ticket);
5308
5309         ret = 0;
5310         priority_reclaim_metadata_space(fs_info, space_info, &ticket);
5311         spin_lock(&space_info->lock);
5312         if (ticket.bytes) {
5313                 if (ticket.bytes < orig_bytes)
5314                         reclaim_bytes = orig_bytes - ticket.bytes;
5315                 list_del_init(&ticket.list);
5316                 ret = -ENOSPC;
5317         }
5318         spin_unlock(&space_info->lock);
5319
5320         if (reclaim_bytes)
5321                 space_info_add_old_bytes(fs_info, space_info, reclaim_bytes);
5322         ASSERT(list_empty(&ticket.list));
5323         return ret;
5324 }
5325
5326 /**
5327  * reserve_metadata_bytes - try to reserve bytes from the block_rsv's space
5328  * @root - the root we're allocating for
5329  * @block_rsv - the block_rsv we're allocating for
5330  * @orig_bytes - the number of bytes we want
5331  * @flush - whether or not we can flush to make our reservation
5332  *
5333  * This will reserve orig_bytes number of bytes from the space info associated
5334  * with the block_rsv.  If there is not enough space it will make an attempt to
5335  * flush out space to make room.  It will do this by flushing delalloc if
5336  * possible or committing the transaction.  If flush is 0 then no attempts to
5337  * regain reservations will be made and this will fail if there is not enough
5338  * space already.
5339  */
5340 static int reserve_metadata_bytes(struct btrfs_root *root,
5341                                   struct btrfs_block_rsv *block_rsv,
5342                                   u64 orig_bytes,
5343                                   enum btrfs_reserve_flush_enum flush)
5344 {
5345         struct btrfs_fs_info *fs_info = root->fs_info;
5346         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
5347         int ret;
5348         bool system_chunk = (root == fs_info->chunk_root);
5349
5350         ret = __reserve_metadata_bytes(fs_info, block_rsv->space_info,
5351                                        orig_bytes, flush, system_chunk);
5352         if (ret == -ENOSPC &&
5353             unlikely(root->orphan_cleanup_state == ORPHAN_CLEANUP_STARTED)) {
5354                 if (block_rsv != global_rsv &&
5355                     !block_rsv_use_bytes(global_rsv, orig_bytes))
5356                         ret = 0;
5357         }
5358         if (ret == -ENOSPC) {
5359                 trace_btrfs_space_reservation(fs_info, "space_info:enospc",
5360                                               block_rsv->space_info->flags,
5361                                               orig_bytes, 1);
5362
5363                 if (btrfs_test_opt(fs_info, ENOSPC_DEBUG))
5364                         dump_space_info(fs_info, block_rsv->space_info,
5365                                         orig_bytes, 0);
5366         }
5367         return ret;
5368 }
5369
5370 static struct btrfs_block_rsv *get_block_rsv(
5371                                         const struct btrfs_trans_handle *trans,
5372                                         const struct btrfs_root *root)
5373 {
5374         struct btrfs_fs_info *fs_info = root->fs_info;
5375         struct btrfs_block_rsv *block_rsv = NULL;
5376
5377         if (test_bit(BTRFS_ROOT_REF_COWS, &root->state) ||
5378             (root == fs_info->csum_root && trans->adding_csums) ||
5379             (root == fs_info->uuid_root))
5380                 block_rsv = trans->block_rsv;
5381
5382         if (!block_rsv)
5383                 block_rsv = root->block_rsv;
5384
5385         if (!block_rsv)
5386                 block_rsv = &fs_info->empty_block_rsv;
5387
5388         return block_rsv;
5389 }
5390
5391 static int block_rsv_use_bytes(struct btrfs_block_rsv *block_rsv,
5392                                u64 num_bytes)
5393 {
5394         int ret = -ENOSPC;
5395         spin_lock(&block_rsv->lock);
5396         if (block_rsv->reserved >= num_bytes) {
5397                 block_rsv->reserved -= num_bytes;
5398                 if (block_rsv->reserved < block_rsv->size)
5399                         block_rsv->full = 0;
5400                 ret = 0;
5401         }
5402         spin_unlock(&block_rsv->lock);
5403         return ret;
5404 }
5405
5406 static void block_rsv_add_bytes(struct btrfs_block_rsv *block_rsv,
5407                                 u64 num_bytes, bool update_size)
5408 {
5409         spin_lock(&block_rsv->lock);
5410         block_rsv->reserved += num_bytes;
5411         if (update_size)
5412                 block_rsv->size += num_bytes;
5413         else if (block_rsv->reserved >= block_rsv->size)
5414                 block_rsv->full = 1;
5415         spin_unlock(&block_rsv->lock);
5416 }
5417
5418 int btrfs_cond_migrate_bytes(struct btrfs_fs_info *fs_info,
5419                              struct btrfs_block_rsv *dest, u64 num_bytes,
5420                              int min_factor)
5421 {
5422         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
5423         u64 min_bytes;
5424
5425         if (global_rsv->space_info != dest->space_info)
5426                 return -ENOSPC;
5427
5428         spin_lock(&global_rsv->lock);
5429         min_bytes = div_factor(global_rsv->size, min_factor);
5430         if (global_rsv->reserved < min_bytes + num_bytes) {
5431                 spin_unlock(&global_rsv->lock);
5432                 return -ENOSPC;
5433         }
5434         global_rsv->reserved -= num_bytes;
5435         if (global_rsv->reserved < global_rsv->size)
5436                 global_rsv->full = 0;
5437         spin_unlock(&global_rsv->lock);
5438
5439         block_rsv_add_bytes(dest, num_bytes, true);
5440         return 0;
5441 }
5442
5443 /**
5444  * btrfs_migrate_to_delayed_refs_rsv - transfer bytes to our delayed refs rsv.
5445  * @fs_info - the fs info for our fs.
5446  * @src - the source block rsv to transfer from.
5447  * @num_bytes - the number of bytes to transfer.
5448  *
5449  * This transfers up to the num_bytes amount from the src rsv to the
5450  * delayed_refs_rsv.  Any extra bytes are returned to the space info.
5451  */
5452 void btrfs_migrate_to_delayed_refs_rsv(struct btrfs_fs_info *fs_info,
5453                                        struct btrfs_block_rsv *src,
5454                                        u64 num_bytes)
5455 {
5456         struct btrfs_block_rsv *delayed_refs_rsv = &fs_info->delayed_refs_rsv;
5457         u64 to_free = 0;
5458
5459         spin_lock(&src->lock);
5460         src->reserved -= num_bytes;
5461         src->size -= num_bytes;
5462         spin_unlock(&src->lock);
5463
5464         spin_lock(&delayed_refs_rsv->lock);
5465         if (delayed_refs_rsv->size > delayed_refs_rsv->reserved) {
5466                 u64 delta = delayed_refs_rsv->size -
5467                         delayed_refs_rsv->reserved;
5468                 if (num_bytes > delta) {
5469                         to_free = num_bytes - delta;
5470                         num_bytes = delta;
5471                 }
5472         } else {
5473                 to_free = num_bytes;
5474                 num_bytes = 0;
5475         }
5476
5477         if (num_bytes)
5478                 delayed_refs_rsv->reserved += num_bytes;
5479         if (delayed_refs_rsv->reserved >= delayed_refs_rsv->size)
5480                 delayed_refs_rsv->full = 1;
5481         spin_unlock(&delayed_refs_rsv->lock);
5482
5483         if (num_bytes)
5484                 trace_btrfs_space_reservation(fs_info, "delayed_refs_rsv",
5485                                               0, num_bytes, 1);
5486         if (to_free)
5487                 space_info_add_old_bytes(fs_info, delayed_refs_rsv->space_info,
5488                                          to_free);
5489 }
5490
5491 /**
5492  * btrfs_delayed_refs_rsv_refill - refill based on our delayed refs usage.
5493  * @fs_info - the fs_info for our fs.
5494  * @flush - control how we can flush for this reservation.
5495  *
5496  * This will refill the delayed block_rsv up to 1 items size worth of space and
5497  * will return -ENOSPC if we can't make the reservation.
5498  */
5499 int btrfs_delayed_refs_rsv_refill(struct btrfs_fs_info *fs_info,
5500                                   enum btrfs_reserve_flush_enum flush)
5501 {
5502         struct btrfs_block_rsv *block_rsv = &fs_info->delayed_refs_rsv;
5503         u64 limit = btrfs_calc_trans_metadata_size(fs_info, 1);
5504         u64 num_bytes = 0;
5505         int ret = -ENOSPC;
5506
5507         spin_lock(&block_rsv->lock);
5508         if (block_rsv->reserved < block_rsv->size) {
5509                 num_bytes = block_rsv->size - block_rsv->reserved;
5510                 num_bytes = min(num_bytes, limit);
5511         }
5512         spin_unlock(&block_rsv->lock);
5513
5514         if (!num_bytes)
5515                 return 0;
5516
5517         ret = reserve_metadata_bytes(fs_info->extent_root, block_rsv,
5518                                      num_bytes, flush);
5519         if (ret)
5520                 return ret;
5521         block_rsv_add_bytes(block_rsv, num_bytes, 0);
5522         trace_btrfs_space_reservation(fs_info, "delayed_refs_rsv",
5523                                       0, num_bytes, 1);
5524         return 0;
5525 }
5526
5527 /*
5528  * This is for space we already have accounted in space_info->bytes_may_use, so
5529  * basically when we're returning space from block_rsv's.
5530  */
5531 static void space_info_add_old_bytes(struct btrfs_fs_info *fs_info,
5532                                      struct btrfs_space_info *space_info,
5533                                      u64 num_bytes)
5534 {
5535         struct reserve_ticket *ticket;
5536         struct list_head *head;
5537         u64 used;
5538         enum btrfs_reserve_flush_enum flush = BTRFS_RESERVE_NO_FLUSH;
5539         bool check_overcommit = false;
5540
5541         spin_lock(&space_info->lock);
5542         head = &space_info->priority_tickets;
5543
5544         /*
5545          * If we are over our limit then we need to check and see if we can
5546          * overcommit, and if we can't then we just need to free up our space
5547          * and not satisfy any requests.
5548          */
5549         used = btrfs_space_info_used(space_info, true);
5550         if (used - num_bytes >= space_info->total_bytes)
5551                 check_overcommit = true;
5552 again:
5553         while (!list_empty(head) && num_bytes) {
5554                 ticket = list_first_entry(head, struct reserve_ticket,
5555                                           list);
5556                 /*
5557                  * We use 0 bytes because this space is already reserved, so
5558                  * adding the ticket space would be a double count.
5559                  */
5560                 if (check_overcommit &&
5561                     !can_overcommit(fs_info, space_info, 0, flush, false))
5562                         break;
5563                 if (num_bytes >= ticket->bytes) {
5564                         list_del_init(&ticket->list);
5565                         num_bytes -= ticket->bytes;
5566                         ticket->bytes = 0;
5567                         space_info->tickets_id++;
5568                         wake_up(&ticket->wait);
5569                 } else {
5570                         ticket->bytes -= num_bytes;
5571                         num_bytes = 0;
5572                 }
5573         }
5574
5575         if (num_bytes && head == &space_info->priority_tickets) {
5576                 head = &space_info->tickets;
5577                 flush = BTRFS_RESERVE_FLUSH_ALL;
5578                 goto again;
5579         }
5580         update_bytes_may_use(space_info, -num_bytes);
5581         trace_btrfs_space_reservation(fs_info, "space_info",
5582                                       space_info->flags, num_bytes, 0);
5583         spin_unlock(&space_info->lock);
5584 }
5585
5586 /*
5587  * This is for newly allocated space that isn't accounted in
5588  * space_info->bytes_may_use yet.  So if we allocate a chunk or unpin an extent
5589  * we use this helper.
5590  */
5591 static void space_info_add_new_bytes(struct btrfs_fs_info *fs_info,
5592                                      struct btrfs_space_info *space_info,
5593                                      u64 num_bytes)
5594 {
5595         struct reserve_ticket *ticket;
5596         struct list_head *head = &space_info->priority_tickets;
5597
5598 again:
5599         while (!list_empty(head) && num_bytes) {
5600                 ticket = list_first_entry(head, struct reserve_ticket,
5601                                           list);
5602                 if (num_bytes >= ticket->bytes) {
5603                         trace_btrfs_space_reservation(fs_info, "space_info",
5604                                                       space_info->flags,
5605                                                       ticket->bytes, 1);
5606                         list_del_init(&ticket->list);
5607                         num_bytes -= ticket->bytes;
5608                         update_bytes_may_use(space_info, ticket->bytes);
5609                         ticket->bytes = 0;
5610                         space_info->tickets_id++;
5611                         wake_up(&ticket->wait);
5612                 } else {
5613                         trace_btrfs_space_reservation(fs_info, "space_info",
5614                                                       space_info->flags,
5615                                                       num_bytes, 1);
5616                         update_bytes_may_use(space_info, num_bytes);
5617                         ticket->bytes -= num_bytes;
5618                         num_bytes = 0;
5619                 }
5620         }
5621
5622         if (num_bytes && head == &space_info->priority_tickets) {
5623                 head = &space_info->tickets;
5624                 goto again;
5625         }
5626 }
5627
5628 static u64 block_rsv_release_bytes(struct btrfs_fs_info *fs_info,
5629                                     struct btrfs_block_rsv *block_rsv,
5630                                     struct btrfs_block_rsv *dest, u64 num_bytes,
5631                                     u64 *qgroup_to_release_ret)
5632 {
5633         struct btrfs_space_info *space_info = block_rsv->space_info;
5634         u64 qgroup_to_release = 0;
5635         u64 ret;
5636
5637         spin_lock(&block_rsv->lock);
5638         if (num_bytes == (u64)-1) {
5639                 num_bytes = block_rsv->size;
5640                 qgroup_to_release = block_rsv->qgroup_rsv_size;
5641         }
5642         block_rsv->size -= num_bytes;
5643         if (block_rsv->reserved >= block_rsv->size) {
5644                 num_bytes = block_rsv->reserved - block_rsv->size;
5645                 block_rsv->reserved = block_rsv->size;
5646                 block_rsv->full = 1;
5647         } else {
5648                 num_bytes = 0;
5649         }
5650         if (block_rsv->qgroup_rsv_reserved >= block_rsv->qgroup_rsv_size) {
5651                 qgroup_to_release = block_rsv->qgroup_rsv_reserved -
5652                                     block_rsv->qgroup_rsv_size;
5653                 block_rsv->qgroup_rsv_reserved = block_rsv->qgroup_rsv_size;
5654         } else {
5655                 qgroup_to_release = 0;
5656         }
5657         spin_unlock(&block_rsv->lock);
5658
5659         ret = num_bytes;
5660         if (num_bytes > 0) {
5661                 if (dest) {
5662                         spin_lock(&dest->lock);
5663                         if (!dest->full) {
5664                                 u64 bytes_to_add;
5665
5666                                 bytes_to_add = dest->size - dest->reserved;
5667                                 bytes_to_add = min(num_bytes, bytes_to_add);
5668                                 dest->reserved += bytes_to_add;
5669                                 if (dest->reserved >= dest->size)
5670                                         dest->full = 1;
5671                                 num_bytes -= bytes_to_add;
5672                         }
5673                         spin_unlock(&dest->lock);
5674                 }
5675                 if (num_bytes)
5676                         space_info_add_old_bytes(fs_info, space_info,
5677                                                  num_bytes);
5678         }
5679         if (qgroup_to_release_ret)
5680                 *qgroup_to_release_ret = qgroup_to_release;
5681         return ret;
5682 }
5683
5684 int btrfs_block_rsv_migrate(struct btrfs_block_rsv *src,
5685                             struct btrfs_block_rsv *dst, u64 num_bytes,
5686                             bool update_size)
5687 {
5688         int ret;
5689
5690         ret = block_rsv_use_bytes(src, num_bytes);
5691         if (ret)
5692                 return ret;
5693
5694         block_rsv_add_bytes(dst, num_bytes, update_size);
5695         return 0;
5696 }
5697
5698 void btrfs_init_block_rsv(struct btrfs_block_rsv *rsv, unsigned short type)
5699 {
5700         memset(rsv, 0, sizeof(*rsv));
5701         spin_lock_init(&rsv->lock);
5702         rsv->type = type;
5703 }
5704
5705 void btrfs_init_metadata_block_rsv(struct btrfs_fs_info *fs_info,
5706                                    struct btrfs_block_rsv *rsv,
5707                                    unsigned short type)
5708 {
5709         btrfs_init_block_rsv(rsv, type);
5710         rsv->space_info = __find_space_info(fs_info,
5711                                             BTRFS_BLOCK_GROUP_METADATA);
5712 }
5713
5714 struct btrfs_block_rsv *btrfs_alloc_block_rsv(struct btrfs_fs_info *fs_info,
5715                                               unsigned short type)
5716 {
5717         struct btrfs_block_rsv *block_rsv;
5718
5719         block_rsv = kmalloc(sizeof(*block_rsv), GFP_NOFS);
5720         if (!block_rsv)
5721                 return NULL;
5722
5723         btrfs_init_metadata_block_rsv(fs_info, block_rsv, type);
5724         return block_rsv;
5725 }
5726
5727 void btrfs_free_block_rsv(struct btrfs_fs_info *fs_info,
5728                           struct btrfs_block_rsv *rsv)
5729 {
5730         if (!rsv)
5731                 return;
5732         btrfs_block_rsv_release(fs_info, rsv, (u64)-1);
5733         kfree(rsv);
5734 }
5735
5736 int btrfs_block_rsv_add(struct btrfs_root *root,
5737                         struct btrfs_block_rsv *block_rsv, u64 num_bytes,
5738                         enum btrfs_reserve_flush_enum flush)
5739 {
5740         int ret;
5741
5742         if (num_bytes == 0)
5743                 return 0;
5744
5745         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
5746         if (!ret)
5747                 block_rsv_add_bytes(block_rsv, num_bytes, true);
5748
5749         return ret;
5750 }
5751
5752 int btrfs_block_rsv_check(struct btrfs_block_rsv *block_rsv, int min_factor)
5753 {
5754         u64 num_bytes = 0;
5755         int ret = -ENOSPC;
5756
5757         if (!block_rsv)
5758                 return 0;
5759
5760         spin_lock(&block_rsv->lock);
5761         num_bytes = div_factor(block_rsv->size, min_factor);
5762         if (block_rsv->reserved >= num_bytes)
5763                 ret = 0;
5764         spin_unlock(&block_rsv->lock);
5765
5766         return ret;
5767 }
5768
5769 int btrfs_block_rsv_refill(struct btrfs_root *root,
5770                            struct btrfs_block_rsv *block_rsv, u64 min_reserved,
5771                            enum btrfs_reserve_flush_enum flush)
5772 {
5773         u64 num_bytes = 0;
5774         int ret = -ENOSPC;
5775
5776         if (!block_rsv)
5777                 return 0;
5778
5779         spin_lock(&block_rsv->lock);
5780         num_bytes = min_reserved;
5781         if (block_rsv->reserved >= num_bytes)
5782                 ret = 0;
5783         else
5784                 num_bytes -= block_rsv->reserved;
5785         spin_unlock(&block_rsv->lock);
5786
5787         if (!ret)
5788                 return 0;
5789
5790         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
5791         if (!ret) {
5792                 block_rsv_add_bytes(block_rsv, num_bytes, false);
5793                 return 0;
5794         }
5795
5796         return ret;
5797 }
5798
5799 static void calc_refill_bytes(struct btrfs_block_rsv *block_rsv,
5800                                 u64 *metadata_bytes, u64 *qgroup_bytes)
5801 {
5802         *metadata_bytes = 0;
5803         *qgroup_bytes = 0;
5804
5805         spin_lock(&block_rsv->lock);
5806         if (block_rsv->reserved < block_rsv->size)
5807                 *metadata_bytes = block_rsv->size - block_rsv->reserved;
5808         if (block_rsv->qgroup_rsv_reserved < block_rsv->qgroup_rsv_size)
5809                 *qgroup_bytes = block_rsv->qgroup_rsv_size -
5810                         block_rsv->qgroup_rsv_reserved;
5811         spin_unlock(&block_rsv->lock);
5812 }
5813
5814 /**
5815  * btrfs_inode_rsv_refill - refill the inode block rsv.
5816  * @inode - the inode we are refilling.
5817  * @flush - the flushing restriction.
5818  *
5819  * Essentially the same as btrfs_block_rsv_refill, except it uses the
5820  * block_rsv->size as the minimum size.  We'll either refill the missing amount
5821  * or return if we already have enough space.  This will also handle the reserve
5822  * tracepoint for the reserved amount.
5823  */
5824 static int btrfs_inode_rsv_refill(struct btrfs_inode *inode,
5825                                   enum btrfs_reserve_flush_enum flush)
5826 {
5827         struct btrfs_root *root = inode->root;
5828         struct btrfs_block_rsv *block_rsv = &inode->block_rsv;
5829         u64 num_bytes, last = 0;
5830         u64 qgroup_num_bytes;
5831         int ret = -ENOSPC;
5832
5833         calc_refill_bytes(block_rsv, &num_bytes, &qgroup_num_bytes);
5834         if (num_bytes == 0)
5835                 return 0;
5836
5837         do {
5838                 ret = btrfs_qgroup_reserve_meta_prealloc(root, qgroup_num_bytes,
5839                                                          true);
5840                 if (ret)
5841                         return ret;
5842                 ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
5843                 if (ret) {
5844                         btrfs_qgroup_free_meta_prealloc(root, qgroup_num_bytes);
5845                         last = num_bytes;
5846                         /*
5847                          * If we are fragmented we can end up with a lot of
5848                          * outstanding extents which will make our size be much
5849                          * larger than our reserved amount.
5850                          *
5851                          * If the reservation happens here, it might be very
5852                          * big though not needed in the end, if the delalloc
5853                          * flushing happens.
5854                          *
5855                          * If this is the case try and do the reserve again.
5856                          */
5857                         if (flush == BTRFS_RESERVE_FLUSH_ALL)
5858                                 calc_refill_bytes(block_rsv, &num_bytes,
5859                                                    &qgroup_num_bytes);
5860                         if (num_bytes == 0)
5861                                 return 0;
5862                 }
5863         } while (ret && last != num_bytes);
5864
5865         if (!ret) {
5866                 block_rsv_add_bytes(block_rsv, num_bytes, false);
5867                 trace_btrfs_space_reservation(root->fs_info, "delalloc",
5868                                               btrfs_ino(inode), num_bytes, 1);
5869
5870                 /* Don't forget to increase qgroup_rsv_reserved */
5871                 spin_lock(&block_rsv->lock);
5872                 block_rsv->qgroup_rsv_reserved += qgroup_num_bytes;
5873                 spin_unlock(&block_rsv->lock);
5874         }
5875         return ret;
5876 }
5877
5878 static u64 __btrfs_block_rsv_release(struct btrfs_fs_info *fs_info,
5879                                      struct btrfs_block_rsv *block_rsv,
5880                                      u64 num_bytes, u64 *qgroup_to_release)
5881 {
5882         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
5883         struct btrfs_block_rsv *delayed_rsv = &fs_info->delayed_refs_rsv;
5884         struct btrfs_block_rsv *target = delayed_rsv;
5885
5886         if (target->full || target == block_rsv)
5887                 target = global_rsv;
5888
5889         if (block_rsv->space_info != target->space_info)
5890                 target = NULL;
5891
5892         return block_rsv_release_bytes(fs_info, block_rsv, target, num_bytes,
5893                                        qgroup_to_release);
5894 }
5895
5896 void btrfs_block_rsv_release(struct btrfs_fs_info *fs_info,
5897                              struct btrfs_block_rsv *block_rsv,
5898                              u64 num_bytes)
5899 {
5900         __btrfs_block_rsv_release(fs_info, block_rsv, num_bytes, NULL);
5901 }
5902
5903 /**
5904  * btrfs_inode_rsv_release - release any excessive reservation.
5905  * @inode - the inode we need to release from.
5906  * @qgroup_free - free or convert qgroup meta.
5907  *   Unlike normal operation, qgroup meta reservation needs to know if we are
5908  *   freeing qgroup reservation or just converting it into per-trans.  Normally
5909  *   @qgroup_free is true for error handling, and false for normal release.
5910  *
5911  * This is the same as btrfs_block_rsv_release, except that it handles the
5912  * tracepoint for the reservation.
5913  */
5914 static void btrfs_inode_rsv_release(struct btrfs_inode *inode, bool qgroup_free)
5915 {
5916         struct btrfs_fs_info *fs_info = inode->root->fs_info;
5917         struct btrfs_block_rsv *block_rsv = &inode->block_rsv;
5918         u64 released = 0;
5919         u64 qgroup_to_release = 0;
5920
5921         /*
5922          * Since we statically set the block_rsv->size we just want to say we
5923          * are releasing 0 bytes, and then we'll just get the reservation over
5924          * the size free'd.
5925          */
5926         released = __btrfs_block_rsv_release(fs_info, block_rsv, 0,
5927                                              &qgroup_to_release);
5928         if (released > 0)
5929                 trace_btrfs_space_reservation(fs_info, "delalloc",
5930                                               btrfs_ino(inode), released, 0);
5931         if (qgroup_free)
5932                 btrfs_qgroup_free_meta_prealloc(inode->root, qgroup_to_release);
5933         else
5934                 btrfs_qgroup_convert_reserved_meta(inode->root,
5935                                                    qgroup_to_release);
5936 }
5937
5938 /**
5939  * btrfs_delayed_refs_rsv_release - release a ref head's reservation.
5940  * @fs_info - the fs_info for our fs.
5941  * @nr - the number of items to drop.
5942  *
5943  * This drops the delayed ref head's count from the delayed refs rsv and frees
5944  * any excess reservation we had.
5945  */
5946 void btrfs_delayed_refs_rsv_release(struct btrfs_fs_info *fs_info, int nr)
5947 {
5948         struct btrfs_block_rsv *block_rsv = &fs_info->delayed_refs_rsv;
5949         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
5950         u64 num_bytes = btrfs_calc_trans_metadata_size(fs_info, nr);
5951         u64 released = 0;
5952
5953         released = block_rsv_release_bytes(fs_info, block_rsv, global_rsv,
5954                                            num_bytes, NULL);
5955         if (released)
5956                 trace_btrfs_space_reservation(fs_info, "delayed_refs_rsv",
5957                                               0, released, 0);
5958 }
5959
5960 static void update_global_block_rsv(struct btrfs_fs_info *fs_info)
5961 {
5962         struct btrfs_block_rsv *block_rsv = &fs_info->global_block_rsv;
5963         struct btrfs_space_info *sinfo = block_rsv->space_info;
5964         u64 num_bytes;
5965
5966         /*
5967          * The global block rsv is based on the size of the extent tree, the
5968          * checksum tree and the root tree.  If the fs is empty we want to set
5969          * it to a minimal amount for safety.
5970          */
5971         num_bytes = btrfs_root_used(&fs_info->extent_root->root_item) +
5972                 btrfs_root_used(&fs_info->csum_root->root_item) +
5973                 btrfs_root_used(&fs_info->tree_root->root_item);
5974         num_bytes = max_t(u64, num_bytes, SZ_16M);
5975
5976         spin_lock(&sinfo->lock);
5977         spin_lock(&block_rsv->lock);
5978
5979         block_rsv->size = min_t(u64, num_bytes, SZ_512M);
5980
5981         if (block_rsv->reserved < block_rsv->size) {
5982                 num_bytes = btrfs_space_info_used(sinfo, true);
5983                 if (sinfo->total_bytes > num_bytes) {
5984                         num_bytes = sinfo->total_bytes - num_bytes;
5985                         num_bytes = min(num_bytes,
5986                                         block_rsv->size - block_rsv->reserved);
5987                         block_rsv->reserved += num_bytes;
5988                         update_bytes_may_use(sinfo, num_bytes);
5989                         trace_btrfs_space_reservation(fs_info, "space_info",
5990                                                       sinfo->flags, num_bytes,
5991                                                       1);
5992                 }
5993         } else if (block_rsv->reserved > block_rsv->size) {
5994                 num_bytes = block_rsv->reserved - block_rsv->size;
5995                 update_bytes_may_use(sinfo, -num_bytes);
5996                 trace_btrfs_space_reservation(fs_info, "space_info",
5997                                       sinfo->flags, num_bytes, 0);
5998                 block_rsv->reserved = block_rsv->size;
5999         }
6000
6001         if (block_rsv->reserved == block_rsv->size)
6002                 block_rsv->full = 1;
6003         else
6004                 block_rsv->full = 0;
6005
6006         spin_unlock(&block_rsv->lock);
6007         spin_unlock(&sinfo->lock);
6008 }
6009
6010 static void init_global_block_rsv(struct btrfs_fs_info *fs_info)
6011 {
6012         struct btrfs_space_info *space_info;
6013
6014         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
6015         fs_info->chunk_block_rsv.space_info = space_info;
6016
6017         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
6018         fs_info->global_block_rsv.space_info = space_info;
6019         fs_info->trans_block_rsv.space_info = space_info;
6020         fs_info->empty_block_rsv.space_info = space_info;
6021         fs_info->delayed_block_rsv.space_info = space_info;
6022         fs_info->delayed_refs_rsv.space_info = space_info;
6023
6024         fs_info->extent_root->block_rsv = &fs_info->delayed_refs_rsv;
6025         fs_info->csum_root->block_rsv = &fs_info->delayed_refs_rsv;
6026         fs_info->dev_root->block_rsv = &fs_info->global_block_rsv;
6027         fs_info->tree_root->block_rsv = &fs_info->global_block_rsv;
6028         if (fs_info->quota_root)
6029                 fs_info->quota_root->block_rsv = &fs_info->global_block_rsv;
6030         fs_info->chunk_root->block_rsv = &fs_info->chunk_block_rsv;
6031
6032         update_global_block_rsv(fs_info);
6033 }
6034
6035 static void release_global_block_rsv(struct btrfs_fs_info *fs_info)
6036 {
6037         block_rsv_release_bytes(fs_info, &fs_info->global_block_rsv, NULL,
6038                                 (u64)-1, NULL);
6039         WARN_ON(fs_info->trans_block_rsv.size > 0);
6040         WARN_ON(fs_info->trans_block_rsv.reserved > 0);
6041         WARN_ON(fs_info->chunk_block_rsv.size > 0);
6042         WARN_ON(fs_info->chunk_block_rsv.reserved > 0);
6043         WARN_ON(fs_info->delayed_block_rsv.size > 0);
6044         WARN_ON(fs_info->delayed_block_rsv.reserved > 0);
6045         WARN_ON(fs_info->delayed_refs_rsv.reserved > 0);
6046         WARN_ON(fs_info->delayed_refs_rsv.size > 0);
6047 }
6048
6049 /*
6050  * btrfs_update_delayed_refs_rsv - adjust the size of the delayed refs rsv
6051  * @trans - the trans that may have generated delayed refs
6052  *
6053  * This is to be called anytime we may have adjusted trans->delayed_ref_updates,
6054  * it'll calculate the additional size and add it to the delayed_refs_rsv.
6055  */
6056 void btrfs_update_delayed_refs_rsv(struct btrfs_trans_handle *trans)
6057 {
6058         struct btrfs_fs_info *fs_info = trans->fs_info;
6059         struct btrfs_block_rsv *delayed_rsv = &fs_info->delayed_refs_rsv;
6060         u64 num_bytes;
6061
6062         if (!trans->delayed_ref_updates)
6063                 return;
6064
6065         num_bytes = btrfs_calc_trans_metadata_size(fs_info,
6066                                                    trans->delayed_ref_updates);
6067         spin_lock(&delayed_rsv->lock);
6068         delayed_rsv->size += num_bytes;
6069         delayed_rsv->full = 0;
6070         spin_unlock(&delayed_rsv->lock);
6071         trans->delayed_ref_updates = 0;
6072 }
6073
6074 /*
6075  * To be called after all the new block groups attached to the transaction
6076  * handle have been created (btrfs_create_pending_block_groups()).
6077  */
6078 void btrfs_trans_release_chunk_metadata(struct btrfs_trans_handle *trans)
6079 {
6080         struct btrfs_fs_info *fs_info = trans->fs_info;
6081
6082         if (!trans->chunk_bytes_reserved)
6083                 return;
6084
6085         WARN_ON_ONCE(!list_empty(&trans->new_bgs));
6086
6087         block_rsv_release_bytes(fs_info, &fs_info->chunk_block_rsv, NULL,
6088                                 trans->chunk_bytes_reserved, NULL);
6089         trans->chunk_bytes_reserved = 0;
6090 }
6091
6092 /*
6093  * btrfs_subvolume_reserve_metadata() - reserve space for subvolume operation
6094  * root: the root of the parent directory
6095  * rsv: block reservation
6096  * items: the number of items that we need do reservation
6097  * use_global_rsv: allow fallback to the global block reservation
6098  *
6099  * This function is used to reserve the space for snapshot/subvolume
6100  * creation and deletion. Those operations are different with the
6101  * common file/directory operations, they change two fs/file trees
6102  * and root tree, the number of items that the qgroup reserves is
6103  * different with the free space reservation. So we can not use
6104  * the space reservation mechanism in start_transaction().
6105  */
6106 int btrfs_subvolume_reserve_metadata(struct btrfs_root *root,
6107                                      struct btrfs_block_rsv *rsv, int items,
6108                                      bool use_global_rsv)
6109 {
6110         u64 qgroup_num_bytes = 0;
6111         u64 num_bytes;
6112         int ret;
6113         struct btrfs_fs_info *fs_info = root->fs_info;
6114         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
6115
6116         if (test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags)) {
6117                 /* One for parent inode, two for dir entries */
6118                 qgroup_num_bytes = 3 * fs_info->nodesize;
6119                 ret = btrfs_qgroup_reserve_meta_prealloc(root,
6120                                 qgroup_num_bytes, true);
6121                 if (ret)
6122                         return ret;
6123         }
6124
6125         num_bytes = btrfs_calc_trans_metadata_size(fs_info, items);
6126         rsv->space_info = __find_space_info(fs_info,
6127                                             BTRFS_BLOCK_GROUP_METADATA);
6128         ret = btrfs_block_rsv_add(root, rsv, num_bytes,
6129                                   BTRFS_RESERVE_FLUSH_ALL);
6130
6131         if (ret == -ENOSPC && use_global_rsv)
6132                 ret = btrfs_block_rsv_migrate(global_rsv, rsv, num_bytes, true);
6133
6134         if (ret && qgroup_num_bytes)
6135                 btrfs_qgroup_free_meta_prealloc(root, qgroup_num_bytes);
6136
6137         return ret;
6138 }
6139
6140 void btrfs_subvolume_release_metadata(struct btrfs_fs_info *fs_info,
6141                                       struct btrfs_block_rsv *rsv)
6142 {
6143         btrfs_block_rsv_release(fs_info, rsv, (u64)-1);
6144 }
6145
6146 static void btrfs_calculate_inode_block_rsv_size(struct btrfs_fs_info *fs_info,
6147                                                  struct btrfs_inode *inode)
6148 {
6149         struct btrfs_block_rsv *block_rsv = &inode->block_rsv;
6150         u64 reserve_size = 0;
6151         u64 qgroup_rsv_size = 0;
6152         u64 csum_leaves;
6153         unsigned outstanding_extents;
6154
6155         lockdep_assert_held(&inode->lock);
6156         outstanding_extents = inode->outstanding_extents;
6157         if (outstanding_extents)
6158                 reserve_size = btrfs_calc_trans_metadata_size(fs_info,
6159                                                 outstanding_extents + 1);
6160         csum_leaves = btrfs_csum_bytes_to_leaves(fs_info,
6161                                                  inode->csum_bytes);
6162         reserve_size += btrfs_calc_trans_metadata_size(fs_info,
6163                                                        csum_leaves);
6164         /*
6165          * For qgroup rsv, the calculation is very simple:
6166          * account one nodesize for each outstanding extent
6167          *
6168          * This is overestimating in most cases.
6169          */
6170         qgroup_rsv_size = (u64)outstanding_extents * fs_info->nodesize;
6171
6172         spin_lock(&block_rsv->lock);
6173         block_rsv->size = reserve_size;
6174         block_rsv->qgroup_rsv_size = qgroup_rsv_size;
6175         spin_unlock(&block_rsv->lock);
6176 }
6177
6178 int btrfs_delalloc_reserve_metadata(struct btrfs_inode *inode, u64 num_bytes)
6179 {
6180         struct btrfs_fs_info *fs_info = inode->root->fs_info;
6181         unsigned nr_extents;
6182         enum btrfs_reserve_flush_enum flush = BTRFS_RESERVE_FLUSH_ALL;
6183         int ret = 0;
6184         bool delalloc_lock = true;
6185
6186         /* If we are a free space inode we need to not flush since we will be in
6187          * the middle of a transaction commit.  We also don't need the delalloc
6188          * mutex since we won't race with anybody.  We need this mostly to make
6189          * lockdep shut its filthy mouth.
6190          *
6191          * If we have a transaction open (can happen if we call truncate_block
6192          * from truncate), then we need FLUSH_LIMIT so we don't deadlock.
6193          */
6194         if (btrfs_is_free_space_inode(inode)) {
6195                 flush = BTRFS_RESERVE_NO_FLUSH;
6196                 delalloc_lock = false;
6197         } else {
6198                 if (current->journal_info)
6199                         flush = BTRFS_RESERVE_FLUSH_LIMIT;
6200
6201                 if (btrfs_transaction_in_commit(fs_info))
6202                         schedule_timeout(1);
6203         }
6204
6205         if (delalloc_lock)
6206                 mutex_lock(&inode->delalloc_mutex);
6207
6208         num_bytes = ALIGN(num_bytes, fs_info->sectorsize);
6209
6210         /* Add our new extents and calculate the new rsv size. */
6211         spin_lock(&inode->lock);
6212         nr_extents = count_max_extents(num_bytes);
6213         btrfs_mod_outstanding_extents(inode, nr_extents);
6214         inode->csum_bytes += num_bytes;
6215         btrfs_calculate_inode_block_rsv_size(fs_info, inode);
6216         spin_unlock(&inode->lock);
6217
6218         ret = btrfs_inode_rsv_refill(inode, flush);
6219         if (unlikely(ret))
6220                 goto out_fail;
6221
6222         if (delalloc_lock)
6223                 mutex_unlock(&inode->delalloc_mutex);
6224         return 0;
6225
6226 out_fail:
6227         spin_lock(&inode->lock);
6228         nr_extents = count_max_extents(num_bytes);
6229         btrfs_mod_outstanding_extents(inode, -nr_extents);
6230         inode->csum_bytes -= num_bytes;
6231         btrfs_calculate_inode_block_rsv_size(fs_info, inode);
6232         spin_unlock(&inode->lock);
6233
6234         btrfs_inode_rsv_release(inode, true);
6235         if (delalloc_lock)
6236                 mutex_unlock(&inode->delalloc_mutex);
6237         return ret;
6238 }
6239
6240 /**
6241  * btrfs_delalloc_release_metadata - release a metadata reservation for an inode
6242  * @inode: the inode to release the reservation for.
6243  * @num_bytes: the number of bytes we are releasing.
6244  * @qgroup_free: free qgroup reservation or convert it to per-trans reservation
6245  *
6246  * This will release the metadata reservation for an inode.  This can be called
6247  * once we complete IO for a given set of bytes to release their metadata
6248  * reservations, or on error for the same reason.
6249  */
6250 void btrfs_delalloc_release_metadata(struct btrfs_inode *inode, u64 num_bytes,
6251                                      bool qgroup_free)
6252 {
6253         struct btrfs_fs_info *fs_info = inode->root->fs_info;
6254
6255         num_bytes = ALIGN(num_bytes, fs_info->sectorsize);
6256         spin_lock(&inode->lock);
6257         inode->csum_bytes -= num_bytes;
6258         btrfs_calculate_inode_block_rsv_size(fs_info, inode);
6259         spin_unlock(&inode->lock);
6260
6261         if (btrfs_is_testing(fs_info))
6262                 return;
6263
6264         btrfs_inode_rsv_release(inode, qgroup_free);
6265 }
6266
6267 /**
6268  * btrfs_delalloc_release_extents - release our outstanding_extents
6269  * @inode: the inode to balance the reservation for.
6270  * @num_bytes: the number of bytes we originally reserved with
6271  * @qgroup_free: do we need to free qgroup meta reservation or convert them.
6272  *
6273  * When we reserve space we increase outstanding_extents for the extents we may
6274  * add.  Once we've set the range as delalloc or created our ordered extents we
6275  * have outstanding_extents to track the real usage, so we use this to free our
6276  * temporarily tracked outstanding_extents.  This _must_ be used in conjunction
6277  * with btrfs_delalloc_reserve_metadata.
6278  */
6279 void btrfs_delalloc_release_extents(struct btrfs_inode *inode, u64 num_bytes,
6280                                     bool qgroup_free)
6281 {
6282         struct btrfs_fs_info *fs_info = inode->root->fs_info;
6283         unsigned num_extents;
6284
6285         spin_lock(&inode->lock);
6286         num_extents = count_max_extents(num_bytes);
6287         btrfs_mod_outstanding_extents(inode, -num_extents);
6288         btrfs_calculate_inode_block_rsv_size(fs_info, inode);
6289         spin_unlock(&inode->lock);
6290
6291         if (btrfs_is_testing(fs_info))
6292                 return;
6293
6294         btrfs_inode_rsv_release(inode, qgroup_free);
6295 }
6296
6297 /**
6298  * btrfs_delalloc_reserve_space - reserve data and metadata space for
6299  * delalloc
6300  * @inode: inode we're writing to
6301  * @start: start range we are writing to
6302  * @len: how long the range we are writing to
6303  * @reserved: mandatory parameter, record actually reserved qgroup ranges of
6304  *            current reservation.
6305  *
6306  * This will do the following things
6307  *
6308  * o reserve space in data space info for num bytes
6309  *   and reserve precious corresponding qgroup space
6310  *   (Done in check_data_free_space)
6311  *
6312  * o reserve space for metadata space, based on the number of outstanding
6313  *   extents and how much csums will be needed
6314  *   also reserve metadata space in a per root over-reserve method.
6315  * o add to the inodes->delalloc_bytes
6316  * o add it to the fs_info's delalloc inodes list.
6317  *   (Above 3 all done in delalloc_reserve_metadata)
6318  *
6319  * Return 0 for success
6320  * Return <0 for error(-ENOSPC or -EQUOT)
6321  */
6322 int btrfs_delalloc_reserve_space(struct inode *inode,
6323                         struct extent_changeset **reserved, u64 start, u64 len)
6324 {
6325         int ret;
6326
6327         ret = btrfs_check_data_free_space(inode, reserved, start, len);
6328         if (ret < 0)
6329                 return ret;
6330         ret = btrfs_delalloc_reserve_metadata(BTRFS_I(inode), len);
6331         if (ret < 0)
6332                 btrfs_free_reserved_data_space(inode, *reserved, start, len);
6333         return ret;
6334 }
6335
6336 /**
6337  * btrfs_delalloc_release_space - release data and metadata space for delalloc
6338  * @inode: inode we're releasing space for
6339  * @start: start position of the space already reserved
6340  * @len: the len of the space already reserved
6341  * @release_bytes: the len of the space we consumed or didn't use
6342  *
6343  * This function will release the metadata space that was not used and will
6344  * decrement ->delalloc_bytes and remove it from the fs_info delalloc_inodes
6345  * list if there are no delalloc bytes left.
6346  * Also it will handle the qgroup reserved space.
6347  */
6348 void btrfs_delalloc_release_space(struct inode *inode,
6349                                   struct extent_changeset *reserved,
6350                                   u64 start, u64 len, bool qgroup_free)
6351 {
6352         btrfs_delalloc_release_metadata(BTRFS_I(inode), len, qgroup_free);
6353         btrfs_free_reserved_data_space(inode, reserved, start, len);
6354 }
6355
6356 static int update_block_group(struct btrfs_trans_handle *trans,
6357                               struct btrfs_fs_info *info, u64 bytenr,
6358                               u64 num_bytes, int alloc)
6359 {
6360         struct btrfs_block_group_cache *cache = NULL;
6361         u64 total = num_bytes;
6362         u64 old_val;
6363         u64 byte_in_group;
6364         int factor;
6365         int ret = 0;
6366
6367         /* block accounting for super block */
6368         spin_lock(&info->delalloc_root_lock);
6369         old_val = btrfs_super_bytes_used(info->super_copy);
6370         if (alloc)
6371                 old_val += num_bytes;
6372         else
6373                 old_val -= num_bytes;
6374         btrfs_set_super_bytes_used(info->super_copy, old_val);
6375         spin_unlock(&info->delalloc_root_lock);
6376
6377         while (total) {
6378                 cache = btrfs_lookup_block_group(info, bytenr);
6379                 if (!cache) {
6380                         ret = -ENOENT;
6381                         break;
6382                 }
6383                 factor = btrfs_bg_type_to_factor(cache->flags);
6384
6385                 /*
6386                  * If this block group has free space cache written out, we
6387                  * need to make sure to load it if we are removing space.  This
6388                  * is because we need the unpinning stage to actually add the
6389                  * space back to the block group, otherwise we will leak space.
6390                  */
6391                 if (!alloc && cache->cached == BTRFS_CACHE_NO)
6392                         cache_block_group(cache, 1);
6393
6394                 byte_in_group = bytenr - cache->key.objectid;
6395                 WARN_ON(byte_in_group > cache->key.offset);
6396
6397                 spin_lock(&cache->space_info->lock);
6398                 spin_lock(&cache->lock);
6399
6400                 if (btrfs_test_opt(info, SPACE_CACHE) &&
6401                     cache->disk_cache_state < BTRFS_DC_CLEAR)
6402                         cache->disk_cache_state = BTRFS_DC_CLEAR;
6403
6404                 old_val = btrfs_block_group_used(&cache->item);
6405                 num_bytes = min(total, cache->key.offset - byte_in_group);
6406                 if (alloc) {
6407                         old_val += num_bytes;
6408                         btrfs_set_block_group_used(&cache->item, old_val);
6409                         cache->reserved -= num_bytes;
6410                         cache->space_info->bytes_reserved -= num_bytes;
6411                         cache->space_info->bytes_used += num_bytes;
6412                         cache->space_info->disk_used += num_bytes * factor;
6413                         spin_unlock(&cache->lock);
6414                         spin_unlock(&cache->space_info->lock);
6415                 } else {
6416                         old_val -= num_bytes;
6417                         btrfs_set_block_group_used(&cache->item, old_val);
6418                         cache->pinned += num_bytes;
6419                         update_bytes_pinned(cache->space_info, num_bytes);
6420                         cache->space_info->bytes_used -= num_bytes;
6421                         cache->space_info->disk_used -= num_bytes * factor;
6422                         spin_unlock(&cache->lock);
6423                         spin_unlock(&cache->space_info->lock);
6424
6425                         trace_btrfs_space_reservation(info, "pinned",
6426                                                       cache->space_info->flags,
6427                                                       num_bytes, 1);
6428                         percpu_counter_add_batch(&cache->space_info->total_bytes_pinned,
6429                                            num_bytes,
6430                                            BTRFS_TOTAL_BYTES_PINNED_BATCH);
6431                         set_extent_dirty(info->pinned_extents,
6432                                          bytenr, bytenr + num_bytes - 1,
6433                                          GFP_NOFS | __GFP_NOFAIL);
6434                 }
6435
6436                 spin_lock(&trans->transaction->dirty_bgs_lock);
6437                 if (list_empty(&cache->dirty_list)) {
6438                         list_add_tail(&cache->dirty_list,
6439                                       &trans->transaction->dirty_bgs);
6440                         trans->transaction->num_dirty_bgs++;
6441                         trans->delayed_ref_updates++;
6442                         btrfs_get_block_group(cache);
6443                 }
6444                 spin_unlock(&trans->transaction->dirty_bgs_lock);
6445
6446                 /*
6447                  * No longer have used bytes in this block group, queue it for
6448                  * deletion. We do this after adding the block group to the
6449                  * dirty list to avoid races between cleaner kthread and space
6450                  * cache writeout.
6451                  */
6452                 if (!alloc && old_val == 0)
6453                         btrfs_mark_bg_unused(cache);
6454
6455                 btrfs_put_block_group(cache);
6456                 total -= num_bytes;
6457                 bytenr += num_bytes;
6458         }
6459
6460         /* Modified block groups are accounted for in the delayed_refs_rsv. */
6461         btrfs_update_delayed_refs_rsv(trans);
6462         return ret;
6463 }
6464
6465 static u64 first_logical_byte(struct btrfs_fs_info *fs_info, u64 search_start)
6466 {
6467         struct btrfs_block_group_cache *cache;
6468         u64 bytenr;
6469
6470         spin_lock(&fs_info->block_group_cache_lock);
6471         bytenr = fs_info->first_logical_byte;
6472         spin_unlock(&fs_info->block_group_cache_lock);
6473
6474         if (bytenr < (u64)-1)
6475                 return bytenr;
6476
6477         cache = btrfs_lookup_first_block_group(fs_info, search_start);
6478         if (!cache)
6479                 return 0;
6480
6481         bytenr = cache->key.objectid;
6482         btrfs_put_block_group(cache);
6483
6484         return bytenr;
6485 }
6486
6487 static int pin_down_extent(struct btrfs_fs_info *fs_info,
6488                            struct btrfs_block_group_cache *cache,
6489                            u64 bytenr, u64 num_bytes, int reserved)
6490 {
6491         spin_lock(&cache->space_info->lock);
6492         spin_lock(&cache->lock);
6493         cache->pinned += num_bytes;
6494         update_bytes_pinned(cache->space_info, num_bytes);
6495         if (reserved) {
6496                 cache->reserved -= num_bytes;
6497                 cache->space_info->bytes_reserved -= num_bytes;
6498         }
6499         spin_unlock(&cache->lock);
6500         spin_unlock(&cache->space_info->lock);
6501
6502         trace_btrfs_space_reservation(fs_info, "pinned",
6503                                       cache->space_info->flags, num_bytes, 1);
6504         percpu_counter_add_batch(&cache->space_info->total_bytes_pinned,
6505                     num_bytes, BTRFS_TOTAL_BYTES_PINNED_BATCH);
6506         set_extent_dirty(fs_info->pinned_extents, bytenr,
6507                          bytenr + num_bytes - 1, GFP_NOFS | __GFP_NOFAIL);
6508         return 0;
6509 }
6510
6511 /*
6512  * this function must be called within transaction
6513  */
6514 int btrfs_pin_extent(struct btrfs_fs_info *fs_info,
6515                      u64 bytenr, u64 num_bytes, int reserved)
6516 {
6517         struct btrfs_block_group_cache *cache;
6518
6519         cache = btrfs_lookup_block_group(fs_info, bytenr);
6520         BUG_ON(!cache); /* Logic error */
6521
6522         pin_down_extent(fs_info, cache, bytenr, num_bytes, reserved);
6523
6524         btrfs_put_block_group(cache);
6525         return 0;
6526 }
6527
6528 /*
6529  * this function must be called within transaction
6530  */
6531 int btrfs_pin_extent_for_log_replay(struct btrfs_fs_info *fs_info,
6532                                     u64 bytenr, u64 num_bytes)
6533 {
6534         struct btrfs_block_group_cache *cache;
6535         int ret;
6536
6537         cache = btrfs_lookup_block_group(fs_info, bytenr);
6538         if (!cache)
6539                 return -EINVAL;
6540
6541         /*
6542          * pull in the free space cache (if any) so that our pin
6543          * removes the free space from the cache.  We have load_only set
6544          * to one because the slow code to read in the free extents does check
6545          * the pinned extents.
6546          */
6547         cache_block_group(cache, 1);
6548
6549         pin_down_extent(fs_info, cache, bytenr, num_bytes, 0);
6550
6551         /* remove us from the free space cache (if we're there at all) */
6552         ret = btrfs_remove_free_space(cache, bytenr, num_bytes);
6553         btrfs_put_block_group(cache);
6554         return ret;
6555 }
6556
6557 static int __exclude_logged_extent(struct btrfs_fs_info *fs_info,
6558                                    u64 start, u64 num_bytes)
6559 {
6560         int ret;
6561         struct btrfs_block_group_cache *block_group;
6562         struct btrfs_caching_control *caching_ctl;
6563
6564         block_group = btrfs_lookup_block_group(fs_info, start);
6565         if (!block_group)
6566                 return -EINVAL;
6567
6568         cache_block_group(block_group, 0);
6569         caching_ctl = get_caching_control(block_group);
6570
6571         if (!caching_ctl) {
6572                 /* Logic error */
6573                 BUG_ON(!block_group_cache_done(block_group));
6574                 ret = btrfs_remove_free_space(block_group, start, num_bytes);
6575         } else {
6576                 mutex_lock(&caching_ctl->mutex);
6577
6578                 if (start >= caching_ctl->progress) {
6579                         ret = add_excluded_extent(fs_info, start, num_bytes);
6580                 } else if (start + num_bytes <= caching_ctl->progress) {
6581                         ret = btrfs_remove_free_space(block_group,
6582                                                       start, num_bytes);
6583                 } else {
6584                         num_bytes = caching_ctl->progress - start;
6585                         ret = btrfs_remove_free_space(block_group,
6586                                                       start, num_bytes);
6587                         if (ret)
6588                                 goto out_lock;
6589
6590                         num_bytes = (start + num_bytes) -
6591                                 caching_ctl->progress;
6592                         start = caching_ctl->progress;
6593                         ret = add_excluded_extent(fs_info, start, num_bytes);
6594                 }
6595 out_lock:
6596                 mutex_unlock(&caching_ctl->mutex);
6597                 put_caching_control(caching_ctl);
6598         }
6599         btrfs_put_block_group(block_group);
6600         return ret;
6601 }
6602
6603 int btrfs_exclude_logged_extents(struct extent_buffer *eb)
6604 {
6605         struct btrfs_fs_info *fs_info = eb->fs_info;
6606         struct btrfs_file_extent_item *item;
6607         struct btrfs_key key;
6608         int found_type;
6609         int i;
6610         int ret = 0;
6611
6612         if (!btrfs_fs_incompat(fs_info, MIXED_GROUPS))
6613                 return 0;
6614
6615         for (i = 0; i < btrfs_header_nritems(eb); i++) {
6616                 btrfs_item_key_to_cpu(eb, &key, i);
6617                 if (key.type != BTRFS_EXTENT_DATA_KEY)
6618                         continue;
6619                 item = btrfs_item_ptr(eb, i, struct btrfs_file_extent_item);
6620                 found_type = btrfs_file_extent_type(eb, item);
6621                 if (found_type == BTRFS_FILE_EXTENT_INLINE)
6622                         continue;
6623                 if (btrfs_file_extent_disk_bytenr(eb, item) == 0)
6624                         continue;
6625                 key.objectid = btrfs_file_extent_disk_bytenr(eb, item);
6626                 key.offset = btrfs_file_extent_disk_num_bytes(eb, item);
6627                 ret = __exclude_logged_extent(fs_info, key.objectid, key.offset);
6628                 if (ret)
6629                         break;
6630         }
6631
6632         return ret;
6633 }
6634
6635 static void
6636 btrfs_inc_block_group_reservations(struct btrfs_block_group_cache *bg)
6637 {
6638         atomic_inc(&bg->reservations);
6639 }
6640
6641 void btrfs_dec_block_group_reservations(struct btrfs_fs_info *fs_info,
6642                                         const u64 start)
6643 {
6644         struct btrfs_block_group_cache *bg;
6645
6646         bg = btrfs_lookup_block_group(fs_info, start);
6647         ASSERT(bg);
6648         if (atomic_dec_and_test(&bg->reservations))
6649                 wake_up_var(&bg->reservations);
6650         btrfs_put_block_group(bg);
6651 }
6652
6653 void btrfs_wait_block_group_reservations(struct btrfs_block_group_cache *bg)
6654 {
6655         struct btrfs_space_info *space_info = bg->space_info;
6656
6657         ASSERT(bg->ro);
6658
6659         if (!(bg->flags & BTRFS_BLOCK_GROUP_DATA))
6660                 return;
6661
6662         /*
6663          * Our block group is read only but before we set it to read only,
6664          * some task might have had allocated an extent from it already, but it
6665          * has not yet created a respective ordered extent (and added it to a
6666          * root's list of ordered extents).
6667          * Therefore wait for any task currently allocating extents, since the
6668          * block group's reservations counter is incremented while a read lock
6669          * on the groups' semaphore is held and decremented after releasing
6670          * the read access on that semaphore and creating the ordered extent.
6671          */
6672         down_write(&space_info->groups_sem);
6673         up_write(&space_info->groups_sem);
6674
6675         wait_var_event(&bg->reservations, !atomic_read(&bg->reservations));
6676 }
6677
6678 /**
6679  * btrfs_add_reserved_bytes - update the block_group and space info counters
6680  * @cache:      The cache we are manipulating
6681  * @ram_bytes:  The number of bytes of file content, and will be same to
6682  *              @num_bytes except for the compress path.
6683  * @num_bytes:  The number of bytes in question
6684  * @delalloc:   The blocks are allocated for the delalloc write
6685  *
6686  * This is called by the allocator when it reserves space. If this is a
6687  * reservation and the block group has become read only we cannot make the
6688  * reservation and return -EAGAIN, otherwise this function always succeeds.
6689  */
6690 static int btrfs_add_reserved_bytes(struct btrfs_block_group_cache *cache,
6691                                     u64 ram_bytes, u64 num_bytes, int delalloc)
6692 {
6693         struct btrfs_space_info *space_info = cache->space_info;
6694         int ret = 0;
6695
6696         spin_lock(&space_info->lock);
6697         spin_lock(&cache->lock);
6698         if (cache->ro) {
6699                 ret = -EAGAIN;
6700         } else {
6701                 cache->reserved += num_bytes;
6702                 space_info->bytes_reserved += num_bytes;
6703                 update_bytes_may_use(space_info, -ram_bytes);
6704                 if (delalloc)
6705                         cache->delalloc_bytes += num_bytes;
6706         }
6707         spin_unlock(&cache->lock);
6708         spin_unlock(&space_info->lock);
6709         return ret;
6710 }
6711
6712 /**
6713  * btrfs_free_reserved_bytes - update the block_group and space info counters
6714  * @cache:      The cache we are manipulating
6715  * @num_bytes:  The number of bytes in question
6716  * @delalloc:   The blocks are allocated for the delalloc write
6717  *
6718  * This is called by somebody who is freeing space that was never actually used
6719  * on disk.  For example if you reserve some space for a new leaf in transaction
6720  * A and before transaction A commits you free that leaf, you call this with
6721  * reserve set to 0 in order to clear the reservation.
6722  */
6723
6724 static void btrfs_free_reserved_bytes(struct btrfs_block_group_cache *cache,
6725                                       u64 num_bytes, int delalloc)
6726 {
6727         struct btrfs_space_info *space_info = cache->space_info;
6728
6729         spin_lock(&space_info->lock);
6730         spin_lock(&cache->lock);
6731         if (cache->ro)
6732                 space_info->bytes_readonly += num_bytes;
6733         cache->reserved -= num_bytes;
6734         space_info->bytes_reserved -= num_bytes;
6735         space_info->max_extent_size = 0;
6736
6737         if (delalloc)
6738                 cache->delalloc_bytes -= num_bytes;
6739         spin_unlock(&cache->lock);
6740         spin_unlock(&space_info->lock);
6741 }
6742 void btrfs_prepare_extent_commit(struct btrfs_fs_info *fs_info)
6743 {
6744         struct btrfs_caching_control *next;
6745         struct btrfs_caching_control *caching_ctl;
6746         struct btrfs_block_group_cache *cache;
6747
6748         down_write(&fs_info->commit_root_sem);
6749
6750         list_for_each_entry_safe(caching_ctl, next,
6751                                  &fs_info->caching_block_groups, list) {
6752                 cache = caching_ctl->block_group;
6753                 if (block_group_cache_done(cache)) {
6754                         cache->last_byte_to_unpin = (u64)-1;
6755                         list_del_init(&caching_ctl->list);
6756                         put_caching_control(caching_ctl);
6757                 } else {
6758                         cache->last_byte_to_unpin = caching_ctl->progress;
6759                 }
6760         }
6761
6762         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
6763                 fs_info->pinned_extents = &fs_info->freed_extents[1];
6764         else
6765                 fs_info->pinned_extents = &fs_info->freed_extents[0];
6766
6767         up_write(&fs_info->commit_root_sem);
6768
6769         update_global_block_rsv(fs_info);
6770 }
6771
6772 /*
6773  * Returns the free cluster for the given space info and sets empty_cluster to
6774  * what it should be based on the mount options.
6775  */
6776 static struct btrfs_free_cluster *
6777 fetch_cluster_info(struct btrfs_fs_info *fs_info,
6778                    struct btrfs_space_info *space_info, u64 *empty_cluster)
6779 {
6780         struct btrfs_free_cluster *ret = NULL;
6781
6782         *empty_cluster = 0;
6783         if (btrfs_mixed_space_info(space_info))
6784                 return ret;
6785
6786         if (space_info->flags & BTRFS_BLOCK_GROUP_METADATA) {
6787                 ret = &fs_info->meta_alloc_cluster;
6788                 if (btrfs_test_opt(fs_info, SSD))
6789                         *empty_cluster = SZ_2M;
6790                 else
6791                         *empty_cluster = SZ_64K;
6792         } else if ((space_info->flags & BTRFS_BLOCK_GROUP_DATA) &&
6793                    btrfs_test_opt(fs_info, SSD_SPREAD)) {
6794                 *empty_cluster = SZ_2M;
6795                 ret = &fs_info->data_alloc_cluster;
6796         }
6797
6798         return ret;
6799 }
6800
6801 static int unpin_extent_range(struct btrfs_fs_info *fs_info,
6802                               u64 start, u64 end,
6803                               const bool return_free_space)
6804 {
6805         struct btrfs_block_group_cache *cache = NULL;
6806         struct btrfs_space_info *space_info;
6807         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
6808         struct btrfs_free_cluster *cluster = NULL;
6809         u64 len;
6810         u64 total_unpinned = 0;
6811         u64 empty_cluster = 0;
6812         bool readonly;
6813
6814         while (start <= end) {
6815                 readonly = false;
6816                 if (!cache ||
6817                     start >= cache->key.objectid + cache->key.offset) {
6818                         if (cache)
6819                                 btrfs_put_block_group(cache);
6820                         total_unpinned = 0;
6821                         cache = btrfs_lookup_block_group(fs_info, start);
6822                         BUG_ON(!cache); /* Logic error */
6823
6824                         cluster = fetch_cluster_info(fs_info,
6825                                                      cache->space_info,
6826                                                      &empty_cluster);
6827                         empty_cluster <<= 1;
6828                 }
6829
6830                 len = cache->key.objectid + cache->key.offset - start;
6831                 len = min(len, end + 1 - start);
6832
6833                 if (start < cache->last_byte_to_unpin) {
6834                         len = min(len, cache->last_byte_to_unpin - start);
6835                         if (return_free_space)
6836                                 btrfs_add_free_space(cache, start, len);
6837                 }
6838
6839                 start += len;
6840                 total_unpinned += len;
6841                 space_info = cache->space_info;
6842
6843                 /*
6844                  * If this space cluster has been marked as fragmented and we've
6845                  * unpinned enough in this block group to potentially allow a
6846                  * cluster to be created inside of it go ahead and clear the
6847                  * fragmented check.
6848                  */
6849                 if (cluster && cluster->fragmented &&
6850                     total_unpinned > empty_cluster) {
6851                         spin_lock(&cluster->lock);
6852                         cluster->fragmented = 0;
6853                         spin_unlock(&cluster->lock);
6854                 }
6855
6856                 spin_lock(&space_info->lock);
6857                 spin_lock(&cache->lock);
6858                 cache->pinned -= len;
6859                 update_bytes_pinned(space_info, -len);
6860
6861                 trace_btrfs_space_reservation(fs_info, "pinned",
6862                                               space_info->flags, len, 0);
6863                 space_info->max_extent_size = 0;
6864                 percpu_counter_add_batch(&space_info->total_bytes_pinned,
6865                             -len, BTRFS_TOTAL_BYTES_PINNED_BATCH);
6866                 if (cache->ro) {
6867                         space_info->bytes_readonly += len;
6868                         readonly = true;
6869                 }
6870                 spin_unlock(&cache->lock);
6871                 if (!readonly && return_free_space &&
6872                     global_rsv->space_info == space_info) {
6873                         u64 to_add = len;
6874
6875                         spin_lock(&global_rsv->lock);
6876                         if (!global_rsv->full) {
6877                                 to_add = min(len, global_rsv->size -
6878                                              global_rsv->reserved);
6879                                 global_rsv->reserved += to_add;
6880                                 update_bytes_may_use(space_info, to_add);
6881                                 if (global_rsv->reserved >= global_rsv->size)
6882                                         global_rsv->full = 1;
6883                                 trace_btrfs_space_reservation(fs_info,
6884                                                               "space_info",
6885                                                               space_info->flags,
6886                                                               to_add, 1);
6887                                 len -= to_add;
6888                         }
6889                         spin_unlock(&global_rsv->lock);
6890                         /* Add to any tickets we may have */
6891                         if (len)
6892                                 space_info_add_new_bytes(fs_info, space_info,
6893                                                          len);
6894                 }
6895                 spin_unlock(&space_info->lock);
6896         }
6897
6898         if (cache)
6899                 btrfs_put_block_group(cache);
6900         return 0;
6901 }
6902
6903 int btrfs_finish_extent_commit(struct btrfs_trans_handle *trans)
6904 {
6905         struct btrfs_fs_info *fs_info = trans->fs_info;
6906         struct btrfs_block_group_cache *block_group, *tmp;
6907         struct list_head *deleted_bgs;
6908         struct extent_io_tree *unpin;
6909         u64 start;
6910         u64 end;
6911         int ret;
6912
6913         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
6914                 unpin = &fs_info->freed_extents[1];
6915         else
6916                 unpin = &fs_info->freed_extents[0];
6917
6918         while (!trans->aborted) {
6919                 struct extent_state *cached_state = NULL;
6920
6921                 mutex_lock(&fs_info->unused_bg_unpin_mutex);
6922                 ret = find_first_extent_bit(unpin, 0, &start, &end,
6923                                             EXTENT_DIRTY, &cached_state);
6924                 if (ret) {
6925                         mutex_unlock(&fs_info->unused_bg_unpin_mutex);
6926                         break;
6927                 }
6928
6929                 if (btrfs_test_opt(fs_info, DISCARD))
6930                         ret = btrfs_discard_extent(fs_info, start,
6931                                                    end + 1 - start, NULL);
6932
6933                 clear_extent_dirty(unpin, start, end, &cached_state);
6934                 unpin_extent_range(fs_info, start, end, true);
6935                 mutex_unlock(&fs_info->unused_bg_unpin_mutex);
6936                 free_extent_state(cached_state);
6937                 cond_resched();
6938         }
6939
6940         /*
6941          * Transaction is finished.  We don't need the lock anymore.  We
6942          * do need to clean up the block groups in case of a transaction
6943          * abort.
6944          */
6945         deleted_bgs = &trans->transaction->deleted_bgs;
6946         list_for_each_entry_safe(block_group, tmp, deleted_bgs, bg_list) {
6947                 u64 trimmed = 0;
6948
6949                 ret = -EROFS;
6950                 if (!trans->aborted)
6951                         ret = btrfs_discard_extent(fs_info,
6952                                                    block_group->key.objectid,
6953                                                    block_group->key.offset,
6954                                                    &trimmed);
6955
6956                 list_del_init(&block_group->bg_list);
6957                 btrfs_put_block_group_trimming(block_group);
6958                 btrfs_put_block_group(block_group);
6959
6960                 if (ret) {
6961                         const char *errstr = btrfs_decode_error(ret);
6962                         btrfs_warn(fs_info,
6963                            "discard failed while removing blockgroup: errno=%d %s",
6964                                    ret, errstr);
6965                 }
6966         }
6967
6968         return 0;
6969 }
6970
6971 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
6972                                struct btrfs_delayed_ref_node *node, u64 parent,
6973                                u64 root_objectid, u64 owner_objectid,
6974                                u64 owner_offset, int refs_to_drop,
6975                                struct btrfs_delayed_extent_op *extent_op)
6976 {
6977         struct btrfs_fs_info *info = trans->fs_info;
6978         struct btrfs_key key;
6979         struct btrfs_path *path;
6980         struct btrfs_root *extent_root = info->extent_root;
6981         struct extent_buffer *leaf;
6982         struct btrfs_extent_item *ei;
6983         struct btrfs_extent_inline_ref *iref;
6984         int ret;
6985         int is_data;
6986         int extent_slot = 0;
6987         int found_extent = 0;
6988         int num_to_del = 1;
6989         u32 item_size;
6990         u64 refs;
6991         u64 bytenr = node->bytenr;
6992         u64 num_bytes = node->num_bytes;
6993         int last_ref = 0;
6994         bool skinny_metadata = btrfs_fs_incompat(info, SKINNY_METADATA);
6995
6996         path = btrfs_alloc_path();
6997         if (!path)
6998                 return -ENOMEM;
6999
7000         path->reada = READA_FORWARD;
7001         path->leave_spinning = 1;
7002
7003         is_data = owner_objectid >= BTRFS_FIRST_FREE_OBJECTID;
7004         BUG_ON(!is_data && refs_to_drop != 1);
7005
7006         if (is_data)
7007                 skinny_metadata = false;
7008
7009         ret = lookup_extent_backref(trans, path, &iref, bytenr, num_bytes,
7010                                     parent, root_objectid, owner_objectid,
7011                                     owner_offset);
7012         if (ret == 0) {
7013                 extent_slot = path->slots[0];
7014                 while (extent_slot >= 0) {
7015                         btrfs_item_key_to_cpu(path->nodes[0], &key,
7016                                               extent_slot);
7017                         if (key.objectid != bytenr)
7018                                 break;
7019                         if (key.type == BTRFS_EXTENT_ITEM_KEY &&
7020                             key.offset == num_bytes) {
7021                                 found_extent = 1;
7022                                 break;
7023                         }
7024                         if (key.type == BTRFS_METADATA_ITEM_KEY &&
7025                             key.offset == owner_objectid) {
7026                                 found_extent = 1;
7027                                 break;
7028                         }
7029                         if (path->slots[0] - extent_slot > 5)
7030                                 break;
7031                         extent_slot--;
7032                 }
7033
7034                 if (!found_extent) {
7035                         BUG_ON(iref);
7036                         ret = remove_extent_backref(trans, path, NULL,
7037                                                     refs_to_drop,
7038                                                     is_data, &last_ref);
7039                         if (ret) {
7040                                 btrfs_abort_transaction(trans, ret);
7041                                 goto out;
7042                         }
7043                         btrfs_release_path(path);
7044                         path->leave_spinning = 1;
7045
7046                         key.objectid = bytenr;
7047                         key.type = BTRFS_EXTENT_ITEM_KEY;
7048                         key.offset = num_bytes;
7049
7050                         if (!is_data && skinny_metadata) {
7051                                 key.type = BTRFS_METADATA_ITEM_KEY;
7052                                 key.offset = owner_objectid;
7053                         }
7054
7055                         ret = btrfs_search_slot(trans, extent_root,
7056                                                 &key, path, -1, 1);
7057                         if (ret > 0 && skinny_metadata && path->slots[0]) {
7058                                 /*
7059                                  * Couldn't find our skinny metadata item,
7060                                  * see if we have ye olde extent item.
7061                                  */
7062                                 path->slots[0]--;
7063                                 btrfs_item_key_to_cpu(path->nodes[0], &key,
7064                                                       path->slots[0]);
7065                                 if (key.objectid == bytenr &&
7066                                     key.type == BTRFS_EXTENT_ITEM_KEY &&
7067                                     key.offset == num_bytes)
7068                                         ret = 0;
7069                         }
7070
7071                         if (ret > 0 && skinny_metadata) {
7072                                 skinny_metadata = false;
7073                                 key.objectid = bytenr;
7074                                 key.type = BTRFS_EXTENT_ITEM_KEY;
7075                                 key.offset = num_bytes;
7076                                 btrfs_release_path(path);
7077                                 ret = btrfs_search_slot(trans, extent_root,
7078                                                         &key, path, -1, 1);
7079                         }
7080
7081                         if (ret) {
7082                                 btrfs_err(info,
7083                                           "umm, got %d back from search, was looking for %llu",
7084                                           ret, bytenr);
7085                                 if (ret > 0)
7086                                         btrfs_print_leaf(path->nodes[0]);
7087                         }
7088                         if (ret < 0) {
7089                                 btrfs_abort_transaction(trans, ret);
7090                                 goto out;
7091                         }
7092                         extent_slot = path->slots[0];
7093                 }
7094         } else if (WARN_ON(ret == -ENOENT)) {
7095                 btrfs_print_leaf(path->nodes[0]);
7096                 btrfs_err(info,
7097                         "unable to find ref byte nr %llu parent %llu root %llu  owner %llu offset %llu",
7098                         bytenr, parent, root_objectid, owner_objectid,
7099                         owner_offset);
7100                 btrfs_abort_transaction(trans, ret);
7101                 goto out;
7102         } else {
7103                 btrfs_abort_transaction(trans, ret);
7104                 goto out;
7105         }
7106
7107         leaf = path->nodes[0];
7108         item_size = btrfs_item_size_nr(leaf, extent_slot);
7109         if (unlikely(item_size < sizeof(*ei))) {
7110                 ret = -EINVAL;
7111                 btrfs_print_v0_err(info);
7112                 btrfs_abort_transaction(trans, ret);
7113                 goto out;
7114         }
7115         ei = btrfs_item_ptr(leaf, extent_slot,
7116                             struct btrfs_extent_item);
7117         if (owner_objectid < BTRFS_FIRST_FREE_OBJECTID &&
7118             key.type == BTRFS_EXTENT_ITEM_KEY) {
7119                 struct btrfs_tree_block_info *bi;
7120                 BUG_ON(item_size < sizeof(*ei) + sizeof(*bi));
7121                 bi = (struct btrfs_tree_block_info *)(ei + 1);
7122                 WARN_ON(owner_objectid != btrfs_tree_block_level(leaf, bi));
7123         }
7124
7125         refs = btrfs_extent_refs(leaf, ei);
7126         if (refs < refs_to_drop) {
7127                 btrfs_err(info,
7128                           "trying to drop %d refs but we only have %Lu for bytenr %Lu",
7129                           refs_to_drop, refs, bytenr);
7130                 ret = -EINVAL;
7131                 btrfs_abort_transaction(trans, ret);
7132                 goto out;
7133         }
7134         refs -= refs_to_drop;
7135
7136         if (refs > 0) {
7137                 if (extent_op)
7138                         __run_delayed_extent_op(extent_op, leaf, ei);
7139                 /*
7140                  * In the case of inline back ref, reference count will
7141                  * be updated by remove_extent_backref
7142                  */
7143                 if (iref) {
7144                         BUG_ON(!found_extent);
7145                 } else {
7146                         btrfs_set_extent_refs(leaf, ei, refs);
7147                         btrfs_mark_buffer_dirty(leaf);
7148                 }
7149                 if (found_extent) {
7150                         ret = remove_extent_backref(trans, path, iref,
7151                                                     refs_to_drop, is_data,
7152                                                     &last_ref);
7153                         if (ret) {
7154                                 btrfs_abort_transaction(trans, ret);
7155                                 goto out;
7156                         }
7157                 }
7158         } else {
7159                 if (found_extent) {
7160                         BUG_ON(is_data && refs_to_drop !=
7161                                extent_data_ref_count(path, iref));
7162                         if (iref) {
7163                                 BUG_ON(path->slots[0] != extent_slot);
7164                         } else {
7165                                 BUG_ON(path->slots[0] != extent_slot + 1);
7166                                 path->slots[0] = extent_slot;
7167                                 num_to_del = 2;
7168                         }
7169                 }
7170
7171                 last_ref = 1;
7172                 ret = btrfs_del_items(trans, extent_root, path, path->slots[0],
7173                                       num_to_del);
7174                 if (ret) {
7175                         btrfs_abort_transaction(trans, ret);
7176                         goto out;
7177                 }
7178                 btrfs_release_path(path);
7179
7180                 if (is_data) {
7181                         ret = btrfs_del_csums(trans, info, bytenr, num_bytes);
7182                         if (ret) {
7183                                 btrfs_abort_transaction(trans, ret);
7184                                 goto out;
7185                         }
7186                 }
7187
7188                 ret = add_to_free_space_tree(trans, bytenr, num_bytes);
7189                 if (ret) {
7190                         btrfs_abort_transaction(trans, ret);
7191                         goto out;
7192                 }
7193
7194                 ret = update_block_group(trans, info, bytenr, num_bytes, 0);
7195                 if (ret) {
7196                         btrfs_abort_transaction(trans, ret);
7197                         goto out;
7198                 }
7199         }
7200         btrfs_release_path(path);
7201
7202 out:
7203         btrfs_free_path(path);
7204         return ret;
7205 }
7206
7207 /*
7208  * when we free an block, it is possible (and likely) that we free the last
7209  * delayed ref for that extent as well.  This searches the delayed ref tree for
7210  * a given extent, and if there are no other delayed refs to be processed, it
7211  * removes it from the tree.
7212  */
7213 static noinline int check_ref_cleanup(struct btrfs_trans_handle *trans,
7214                                       u64 bytenr)
7215 {
7216         struct btrfs_delayed_ref_head *head;
7217         struct btrfs_delayed_ref_root *delayed_refs;
7218         int ret = 0;
7219
7220         delayed_refs = &trans->transaction->delayed_refs;
7221         spin_lock(&delayed_refs->lock);
7222         head = btrfs_find_delayed_ref_head(delayed_refs, bytenr);
7223         if (!head)
7224                 goto out_delayed_unlock;
7225
7226         spin_lock(&head->lock);
7227         if (!RB_EMPTY_ROOT(&head->ref_tree.rb_root))
7228                 goto out;
7229
7230         if (cleanup_extent_op(head) != NULL)
7231                 goto out;
7232
7233         /*
7234          * waiting for the lock here would deadlock.  If someone else has it
7235          * locked they are already in the process of dropping it anyway
7236          */
7237         if (!mutex_trylock(&head->mutex))
7238                 goto out;
7239
7240         btrfs_delete_ref_head(delayed_refs, head);
7241         head->processing = 0;
7242
7243         spin_unlock(&head->lock);
7244         spin_unlock(&delayed_refs->lock);
7245
7246         BUG_ON(head->extent_op);
7247         if (head->must_insert_reserved)
7248                 ret = 1;
7249
7250         btrfs_cleanup_ref_head_accounting(trans->fs_info, delayed_refs, head);
7251         mutex_unlock(&head->mutex);
7252         btrfs_put_delayed_ref_head(head);
7253         return ret;
7254 out:
7255         spin_unlock(&head->lock);
7256
7257 out_delayed_unlock:
7258         spin_unlock(&delayed_refs->lock);
7259         return 0;
7260 }
7261
7262 void btrfs_free_tree_block(struct btrfs_trans_handle *trans,
7263                            struct btrfs_root *root,
7264                            struct extent_buffer *buf,
7265                            u64 parent, int last_ref)
7266 {
7267         struct btrfs_fs_info *fs_info = root->fs_info;
7268         int pin = 1;
7269         int ret;
7270
7271         if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
7272                 int old_ref_mod, new_ref_mod;
7273
7274                 btrfs_ref_tree_mod(root, buf->start, buf->len, parent,
7275                                    root->root_key.objectid,
7276                                    btrfs_header_level(buf), 0,
7277                                    BTRFS_DROP_DELAYED_REF);
7278                 ret = btrfs_add_delayed_tree_ref(trans, buf->start,
7279                                                  buf->len, parent,
7280                                                  root->root_key.objectid,
7281                                                  btrfs_header_level(buf),
7282                                                  BTRFS_DROP_DELAYED_REF, NULL,
7283                                                  &old_ref_mod, &new_ref_mod);
7284                 BUG_ON(ret); /* -ENOMEM */
7285                 pin = old_ref_mod >= 0 && new_ref_mod < 0;
7286         }
7287
7288         if (last_ref && btrfs_header_generation(buf) == trans->transid) {
7289                 struct btrfs_block_group_cache *cache;
7290
7291                 if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
7292                         ret = check_ref_cleanup(trans, buf->start);
7293                         if (!ret)
7294                                 goto out;
7295                 }
7296
7297                 pin = 0;
7298                 cache = btrfs_lookup_block_group(fs_info, buf->start);
7299
7300                 if (btrfs_header_flag(buf, BTRFS_HEADER_FLAG_WRITTEN)) {
7301                         pin_down_extent(fs_info, cache, buf->start,
7302                                         buf->len, 1);
7303                         btrfs_put_block_group(cache);
7304                         goto out;
7305                 }
7306
7307                 WARN_ON(test_bit(EXTENT_BUFFER_DIRTY, &buf->bflags));
7308
7309                 btrfs_add_free_space(cache, buf->start, buf->len);
7310                 btrfs_free_reserved_bytes(cache, buf->len, 0);
7311                 btrfs_put_block_group(cache);
7312                 trace_btrfs_reserved_extent_free(fs_info, buf->start, buf->len);
7313         }
7314 out:
7315         if (pin)
7316                 add_pinned_bytes(fs_info, buf->len, true,
7317                                  root->root_key.objectid);
7318
7319         if (last_ref) {
7320                 /*
7321                  * Deleting the buffer, clear the corrupt flag since it doesn't
7322                  * matter anymore.
7323                  */
7324                 clear_bit(EXTENT_BUFFER_CORRUPT, &buf->bflags);
7325         }
7326 }
7327
7328 /* Can return -ENOMEM */
7329 int btrfs_free_extent(struct btrfs_trans_handle *trans,
7330                       struct btrfs_root *root,
7331                       u64 bytenr, u64 num_bytes, u64 parent, u64 root_objectid,
7332                       u64 owner, u64 offset)
7333 {
7334         struct btrfs_fs_info *fs_info = root->fs_info;
7335         int old_ref_mod, new_ref_mod;
7336         int ret;
7337
7338         if (btrfs_is_testing(fs_info))
7339                 return 0;
7340
7341         if (root_objectid != BTRFS_TREE_LOG_OBJECTID)
7342                 btrfs_ref_tree_mod(root, bytenr, num_bytes, parent,
7343                                    root_objectid, owner, offset,
7344                                    BTRFS_DROP_DELAYED_REF);
7345
7346         /*
7347          * tree log blocks never actually go into the extent allocation
7348          * tree, just update pinning info and exit early.
7349          */
7350         if (root_objectid == BTRFS_TREE_LOG_OBJECTID) {
7351                 WARN_ON(owner >= BTRFS_FIRST_FREE_OBJECTID);
7352                 /* unlocks the pinned mutex */
7353                 btrfs_pin_extent(fs_info, bytenr, num_bytes, 1);
7354                 old_ref_mod = new_ref_mod = 0;
7355                 ret = 0;
7356         } else if (owner < BTRFS_FIRST_FREE_OBJECTID) {
7357                 ret = btrfs_add_delayed_tree_ref(trans, bytenr,
7358                                                  num_bytes, parent,
7359                                                  root_objectid, (int)owner,
7360                                                  BTRFS_DROP_DELAYED_REF, NULL,
7361                                                  &old_ref_mod, &new_ref_mod);
7362         } else {
7363                 ret = btrfs_add_delayed_data_ref(trans, bytenr,
7364                                                  num_bytes, parent,
7365                                                  root_objectid, owner, offset,
7366                                                  0, BTRFS_DROP_DELAYED_REF,
7367                                                  &old_ref_mod, &new_ref_mod);
7368         }
7369
7370         if (ret == 0 && old_ref_mod >= 0 && new_ref_mod < 0) {
7371                 bool metadata = owner < BTRFS_FIRST_FREE_OBJECTID;
7372
7373                 add_pinned_bytes(fs_info, num_bytes, metadata, root_objectid);
7374         }
7375
7376         return ret;
7377 }
7378
7379 /*
7380  * when we wait for progress in the block group caching, its because
7381  * our allocation attempt failed at least once.  So, we must sleep
7382  * and let some progress happen before we try again.
7383  *
7384  * This function will sleep at least once waiting for new free space to
7385  * show up, and then it will check the block group free space numbers
7386  * for our min num_bytes.  Another option is to have it go ahead
7387  * and look in the rbtree for a free extent of a given size, but this
7388  * is a good start.
7389  *
7390  * Callers of this must check if cache->cached == BTRFS_CACHE_ERROR before using
7391  * any of the information in this block group.
7392  */
7393 static noinline void
7394 wait_block_group_cache_progress(struct btrfs_block_group_cache *cache,
7395                                 u64 num_bytes)
7396 {
7397         struct btrfs_caching_control *caching_ctl;
7398
7399         caching_ctl = get_caching_control(cache);
7400         if (!caching_ctl)
7401                 return;
7402
7403         wait_event(caching_ctl->wait, block_group_cache_done(cache) ||
7404                    (cache->free_space_ctl->free_space >= num_bytes));
7405
7406         put_caching_control(caching_ctl);
7407 }
7408
7409 static noinline int
7410 wait_block_group_cache_done(struct btrfs_block_group_cache *cache)
7411 {
7412         struct btrfs_caching_control *caching_ctl;
7413         int ret = 0;
7414
7415         caching_ctl = get_caching_control(cache);
7416         if (!caching_ctl)
7417                 return (cache->cached == BTRFS_CACHE_ERROR) ? -EIO : 0;
7418
7419         wait_event(caching_ctl->wait, block_group_cache_done(cache));
7420         if (cache->cached == BTRFS_CACHE_ERROR)
7421                 ret = -EIO;
7422         put_caching_control(caching_ctl);
7423         return ret;
7424 }
7425
7426 enum btrfs_loop_type {
7427         LOOP_CACHING_NOWAIT = 0,
7428         LOOP_CACHING_WAIT = 1,
7429         LOOP_ALLOC_CHUNK = 2,
7430         LOOP_NO_EMPTY_SIZE = 3,
7431 };
7432
7433 static inline void
7434 btrfs_lock_block_group(struct btrfs_block_group_cache *cache,
7435                        int delalloc)
7436 {
7437         if (delalloc)
7438                 down_read(&cache->data_rwsem);
7439 }
7440
7441 static inline void
7442 btrfs_grab_block_group(struct btrfs_block_group_cache *cache,
7443                        int delalloc)
7444 {
7445         btrfs_get_block_group(cache);
7446         if (delalloc)
7447                 down_read(&cache->data_rwsem);
7448 }
7449
7450 static struct btrfs_block_group_cache *
7451 btrfs_lock_cluster(struct btrfs_block_group_cache *block_group,
7452                    struct btrfs_free_cluster *cluster,
7453                    int delalloc)
7454 {
7455         struct btrfs_block_group_cache *used_bg = NULL;
7456
7457         spin_lock(&cluster->refill_lock);
7458         while (1) {
7459                 used_bg = cluster->block_group;
7460                 if (!used_bg)
7461                         return NULL;
7462
7463                 if (used_bg == block_group)
7464                         return used_bg;
7465
7466                 btrfs_get_block_group(used_bg);
7467
7468                 if (!delalloc)
7469                         return used_bg;
7470
7471                 if (down_read_trylock(&used_bg->data_rwsem))
7472                         return used_bg;
7473
7474                 spin_unlock(&cluster->refill_lock);
7475
7476                 /* We should only have one-level nested. */
7477                 down_read_nested(&used_bg->data_rwsem, SINGLE_DEPTH_NESTING);
7478
7479                 spin_lock(&cluster->refill_lock);
7480                 if (used_bg == cluster->block_group)
7481                         return used_bg;
7482
7483                 up_read(&used_bg->data_rwsem);
7484                 btrfs_put_block_group(used_bg);
7485         }
7486 }
7487
7488 static inline void
7489 btrfs_release_block_group(struct btrfs_block_group_cache *cache,
7490                          int delalloc)
7491 {
7492         if (delalloc)
7493                 up_read(&cache->data_rwsem);
7494         btrfs_put_block_group(cache);
7495 }
7496
7497 /*
7498  * Structure used internally for find_free_extent() function.  Wraps needed
7499  * parameters.
7500  */
7501 struct find_free_extent_ctl {
7502         /* Basic allocation info */
7503         u64 ram_bytes;
7504         u64 num_bytes;
7505         u64 empty_size;
7506         u64 flags;
7507         int delalloc;
7508
7509         /* Where to start the search inside the bg */
7510         u64 search_start;
7511
7512         /* For clustered allocation */
7513         u64 empty_cluster;
7514
7515         bool have_caching_bg;
7516         bool orig_have_caching_bg;
7517
7518         /* RAID index, converted from flags */
7519         int index;
7520
7521         /*
7522          * Current loop number, check find_free_extent_update_loop() for details
7523          */
7524         int loop;
7525
7526         /*
7527          * Whether we're refilling a cluster, if true we need to re-search
7528          * current block group but don't try to refill the cluster again.
7529          */
7530         bool retry_clustered;
7531
7532         /*
7533          * Whether we're updating free space cache, if true we need to re-search
7534          * current block group but don't try updating free space cache again.
7535          */
7536         bool retry_unclustered;
7537
7538         /* If current block group is cached */
7539         int cached;
7540
7541         /* Max contiguous hole found */
7542         u64 max_extent_size;
7543
7544         /* Total free space from free space cache, not always contiguous */
7545         u64 total_free_space;
7546
7547         /* Found result */
7548         u64 found_offset;
7549 };
7550
7551
7552 /*
7553  * Helper function for find_free_extent().
7554  *
7555  * Return -ENOENT to inform caller that we need fallback to unclustered mode.
7556  * Return -EAGAIN to inform caller that we need to re-search this block group
7557  * Return >0 to inform caller that we find nothing
7558  * Return 0 means we have found a location and set ffe_ctl->found_offset.
7559  */
7560 static int find_free_extent_clustered(struct btrfs_block_group_cache *bg,
7561                 struct btrfs_free_cluster *last_ptr,
7562                 struct find_free_extent_ctl *ffe_ctl,
7563                 struct btrfs_block_group_cache **cluster_bg_ret)
7564 {
7565         struct btrfs_fs_info *fs_info = bg->fs_info;
7566         struct btrfs_block_group_cache *cluster_bg;
7567         u64 aligned_cluster;
7568         u64 offset;
7569         int ret;
7570
7571         cluster_bg = btrfs_lock_cluster(bg, last_ptr, ffe_ctl->delalloc);
7572         if (!cluster_bg)
7573                 goto refill_cluster;
7574         if (cluster_bg != bg && (cluster_bg->ro ||
7575             !block_group_bits(cluster_bg, ffe_ctl->flags)))
7576                 goto release_cluster;
7577
7578         offset = btrfs_alloc_from_cluster(cluster_bg, last_ptr,
7579                         ffe_ctl->num_bytes, cluster_bg->key.objectid,
7580                         &ffe_ctl->max_extent_size);
7581         if (offset) {
7582                 /* We have a block, we're done */
7583                 spin_unlock(&last_ptr->refill_lock);
7584                 trace_btrfs_reserve_extent_cluster(cluster_bg,
7585                                 ffe_ctl->search_start, ffe_ctl->num_bytes);
7586                 *cluster_bg_ret = cluster_bg;
7587                 ffe_ctl->found_offset = offset;
7588                 return 0;
7589         }
7590         WARN_ON(last_ptr->block_group != cluster_bg);
7591
7592 release_cluster:
7593         /*
7594          * If we are on LOOP_NO_EMPTY_SIZE, we can't set up a new clusters, so
7595          * lets just skip it and let the allocator find whatever block it can
7596          * find. If we reach this point, we will have tried the cluster
7597          * allocator plenty of times and not have found anything, so we are
7598          * likely way too fragmented for the clustering stuff to find anything.
7599          *
7600          * However, if the cluster is taken from the current block group,
7601          * release the cluster first, so that we stand a better chance of
7602          * succeeding in the unclustered allocation.
7603          */
7604         if (ffe_ctl->loop >= LOOP_NO_EMPTY_SIZE && cluster_bg != bg) {
7605                 spin_unlock(&last_ptr->refill_lock);
7606                 btrfs_release_block_group(cluster_bg, ffe_ctl->delalloc);
7607                 return -ENOENT;
7608         }
7609
7610         /* This cluster didn't work out, free it and start over */
7611         btrfs_return_cluster_to_free_space(NULL, last_ptr);
7612
7613         if (cluster_bg != bg)
7614                 btrfs_release_block_group(cluster_bg, ffe_ctl->delalloc);
7615
7616 refill_cluster:
7617         if (ffe_ctl->loop >= LOOP_NO_EMPTY_SIZE) {
7618                 spin_unlock(&last_ptr->refill_lock);
7619                 return -ENOENT;
7620         }
7621
7622         aligned_cluster = max_t(u64,
7623                         ffe_ctl->empty_cluster + ffe_ctl->empty_size,
7624                         bg->full_stripe_len);
7625         ret = btrfs_find_space_cluster(fs_info, bg, last_ptr,
7626                         ffe_ctl->search_start, ffe_ctl->num_bytes,
7627                         aligned_cluster);
7628         if (ret == 0) {
7629                 /* Now pull our allocation out of this cluster */
7630                 offset = btrfs_alloc_from_cluster(bg, last_ptr,
7631                                 ffe_ctl->num_bytes, ffe_ctl->search_start,
7632                                 &ffe_ctl->max_extent_size);
7633                 if (offset) {
7634                         /* We found one, proceed */
7635                         spin_unlock(&last_ptr->refill_lock);
7636                         trace_btrfs_reserve_extent_cluster(bg,
7637                                         ffe_ctl->search_start,
7638                                         ffe_ctl->num_bytes);
7639                         ffe_ctl->found_offset = offset;
7640                         return 0;
7641                 }
7642         } else if (!ffe_ctl->cached && ffe_ctl->loop > LOOP_CACHING_NOWAIT &&
7643                    !ffe_ctl->retry_clustered) {
7644                 spin_unlock(&last_ptr->refill_lock);
7645
7646                 ffe_ctl->retry_clustered = true;
7647                 wait_block_group_cache_progress(bg, ffe_ctl->num_bytes +
7648                                 ffe_ctl->empty_cluster + ffe_ctl->empty_size);
7649                 return -EAGAIN;
7650         }
7651         /*
7652          * At this point we either didn't find a cluster or we weren't able to
7653          * allocate a block from our cluster.  Free the cluster we've been
7654          * trying to use, and go to the next block group.
7655          */
7656         btrfs_return_cluster_to_free_space(NULL, last_ptr);
7657         spin_unlock(&last_ptr->refill_lock);
7658         return 1;
7659 }
7660
7661 /*
7662  * Return >0 to inform caller that we find nothing
7663  * Return 0 when we found an free extent and set ffe_ctrl->found_offset
7664  * Return -EAGAIN to inform caller that we need to re-search this block group
7665  */
7666 static int find_free_extent_unclustered(struct btrfs_block_group_cache *bg,
7667                 struct btrfs_free_cluster *last_ptr,
7668                 struct find_free_extent_ctl *ffe_ctl)
7669 {
7670         u64 offset;
7671
7672         /*
7673          * We are doing an unclustered allocation, set the fragmented flag so
7674          * we don't bother trying to setup a cluster again until we get more
7675          * space.
7676          */
7677         if (unlikely(last_ptr)) {
7678                 spin_lock(&last_ptr->lock);
7679                 last_ptr->fragmented = 1;
7680                 spin_unlock(&last_ptr->lock);
7681         }
7682         if (ffe_ctl->cached) {
7683                 struct btrfs_free_space_ctl *free_space_ctl;
7684
7685                 free_space_ctl = bg->free_space_ctl;
7686                 spin_lock(&free_space_ctl->tree_lock);
7687                 if (free_space_ctl->free_space <
7688                     ffe_ctl->num_bytes + ffe_ctl->empty_cluster +
7689                     ffe_ctl->empty_size) {
7690                         ffe_ctl->total_free_space = max_t(u64,
7691                                         ffe_ctl->total_free_space,
7692                                         free_space_ctl->free_space);
7693                         spin_unlock(&free_space_ctl->tree_lock);
7694                         return 1;
7695                 }
7696                 spin_unlock(&free_space_ctl->tree_lock);
7697         }
7698
7699         offset = btrfs_find_space_for_alloc(bg, ffe_ctl->search_start,
7700                         ffe_ctl->num_bytes, ffe_ctl->empty_size,
7701                         &ffe_ctl->max_extent_size);
7702
7703         /*
7704          * If we didn't find a chunk, and we haven't failed on this block group
7705          * before, and this block group is in the middle of caching and we are
7706          * ok with waiting, then go ahead and wait for progress to be made, and
7707          * set @retry_unclustered to true.
7708          *
7709          * If @retry_unclustered is true then we've already waited on this
7710          * block group once and should move on to the next block group.
7711          */
7712         if (!offset && !ffe_ctl->retry_unclustered && !ffe_ctl->cached &&
7713             ffe_ctl->loop > LOOP_CACHING_NOWAIT) {
7714                 wait_block_group_cache_progress(bg, ffe_ctl->num_bytes +
7715                                                 ffe_ctl->empty_size);
7716                 ffe_ctl->retry_unclustered = true;
7717                 return -EAGAIN;
7718         } else if (!offset) {
7719                 return 1;
7720         }
7721         ffe_ctl->found_offset = offset;
7722         return 0;
7723 }
7724
7725 /*
7726  * Return >0 means caller needs to re-search for free extent
7727  * Return 0 means we have the needed free extent.
7728  * Return <0 means we failed to locate any free extent.
7729  */
7730 static int find_free_extent_update_loop(struct btrfs_fs_info *fs_info,
7731                                         struct btrfs_free_cluster *last_ptr,
7732                                         struct btrfs_key *ins,
7733                                         struct find_free_extent_ctl *ffe_ctl,
7734                                         int full_search, bool use_cluster)
7735 {
7736         struct btrfs_root *root = fs_info->extent_root;
7737         int ret;
7738
7739         if ((ffe_ctl->loop == LOOP_CACHING_NOWAIT) &&
7740             ffe_ctl->have_caching_bg && !ffe_ctl->orig_have_caching_bg)
7741                 ffe_ctl->orig_have_caching_bg = true;
7742
7743         if (!ins->objectid && ffe_ctl->loop >= LOOP_CACHING_WAIT &&
7744             ffe_ctl->have_caching_bg)
7745                 return 1;
7746
7747         if (!ins->objectid && ++(ffe_ctl->index) < BTRFS_NR_RAID_TYPES)
7748                 return 1;
7749
7750         if (ins->objectid) {
7751                 if (!use_cluster && last_ptr) {
7752                         spin_lock(&last_ptr->lock);
7753                         last_ptr->window_start = ins->objectid;
7754                         spin_unlock(&last_ptr->lock);
7755                 }
7756                 return 0;
7757         }
7758
7759         /*
7760          * LOOP_CACHING_NOWAIT, search partially cached block groups, kicking
7761          *                      caching kthreads as we move along
7762          * LOOP_CACHING_WAIT, search everything, and wait if our bg is caching
7763          * LOOP_ALLOC_CHUNK, force a chunk allocation and try again
7764          * LOOP_NO_EMPTY_SIZE, set empty_size and empty_cluster to 0 and try
7765          *                     again
7766          */
7767         if (ffe_ctl->loop < LOOP_NO_EMPTY_SIZE) {
7768                 ffe_ctl->index = 0;
7769                 if (ffe_ctl->loop == LOOP_CACHING_NOWAIT) {
7770                         /*
7771                          * We want to skip the LOOP_CACHING_WAIT step if we
7772                          * don't have any uncached bgs and we've already done a
7773                          * full search through.
7774                          */
7775                         if (ffe_ctl->orig_have_caching_bg || !full_search)
7776                                 ffe_ctl->loop = LOOP_CACHING_WAIT;
7777                         else
7778                                 ffe_ctl->loop = LOOP_ALLOC_CHUNK;
7779                 } else {
7780                         ffe_ctl->loop++;
7781                 }
7782
7783                 if (ffe_ctl->loop == LOOP_ALLOC_CHUNK) {
7784                         struct btrfs_trans_handle *trans;
7785                         int exist = 0;
7786
7787                         trans = current->journal_info;
7788                         if (trans)
7789                                 exist = 1;
7790                         else
7791                                 trans = btrfs_join_transaction(root);
7792
7793                         if (IS_ERR(trans)) {
7794                                 ret = PTR_ERR(trans);
7795                                 return ret;
7796                         }
7797
7798                         ret = do_chunk_alloc(trans, ffe_ctl->flags,
7799                                              CHUNK_ALLOC_FORCE);
7800
7801                         /*
7802                          * If we can't allocate a new chunk we've already looped
7803                          * through at least once, move on to the NO_EMPTY_SIZE
7804                          * case.
7805                          */
7806                         if (ret == -ENOSPC)
7807                                 ffe_ctl->loop = LOOP_NO_EMPTY_SIZE;
7808
7809                         /* Do not bail out on ENOSPC since we can do more. */
7810                         if (ret < 0 && ret != -ENOSPC)
7811                                 btrfs_abort_transaction(trans, ret);
7812                         else
7813                                 ret = 0;
7814                         if (!exist)
7815                                 btrfs_end_transaction(trans);
7816                         if (ret)
7817                                 return ret;
7818                 }
7819
7820                 if (ffe_ctl->loop == LOOP_NO_EMPTY_SIZE) {
7821                         /*
7822                          * Don't loop again if we already have no empty_size and
7823                          * no empty_cluster.
7824                          */
7825                         if (ffe_ctl->empty_size == 0 &&
7826                             ffe_ctl->empty_cluster == 0)
7827                                 return -ENOSPC;
7828                         ffe_ctl->empty_size = 0;
7829                         ffe_ctl->empty_cluster = 0;
7830                 }
7831                 return 1;
7832         }
7833         return -ENOSPC;
7834 }
7835
7836 /*
7837  * walks the btree of allocated extents and find a hole of a given size.
7838  * The key ins is changed to record the hole:
7839  * ins->objectid == start position
7840  * ins->flags = BTRFS_EXTENT_ITEM_KEY
7841  * ins->offset == the size of the hole.
7842  * Any available blocks before search_start are skipped.
7843  *
7844  * If there is no suitable free space, we will record the max size of
7845  * the free space extent currently.
7846  *
7847  * The overall logic and call chain:
7848  *
7849  * find_free_extent()
7850  * |- Iterate through all block groups
7851  * |  |- Get a valid block group
7852  * |  |- Try to do clustered allocation in that block group
7853  * |  |- Try to do unclustered allocation in that block group
7854  * |  |- Check if the result is valid
7855  * |  |  |- If valid, then exit
7856  * |  |- Jump to next block group
7857  * |
7858  * |- Push harder to find free extents
7859  *    |- If not found, re-iterate all block groups
7860  */
7861 static noinline int find_free_extent(struct btrfs_fs_info *fs_info,
7862                                 u64 ram_bytes, u64 num_bytes, u64 empty_size,
7863                                 u64 hint_byte, struct btrfs_key *ins,
7864                                 u64 flags, int delalloc)
7865 {
7866         int ret = 0;
7867         struct btrfs_free_cluster *last_ptr = NULL;
7868         struct btrfs_block_group_cache *block_group = NULL;
7869         struct find_free_extent_ctl ffe_ctl = {0};
7870         struct btrfs_space_info *space_info;
7871         bool use_cluster = true;
7872         bool full_search = false;
7873
7874         WARN_ON(num_bytes < fs_info->sectorsize);
7875
7876         ffe_ctl.ram_bytes = ram_bytes;
7877         ffe_ctl.num_bytes = num_bytes;
7878         ffe_ctl.empty_size = empty_size;
7879         ffe_ctl.flags = flags;
7880         ffe_ctl.search_start = 0;
7881         ffe_ctl.retry_clustered = false;
7882         ffe_ctl.retry_unclustered = false;
7883         ffe_ctl.delalloc = delalloc;
7884         ffe_ctl.index = btrfs_bg_flags_to_raid_index(flags);
7885         ffe_ctl.have_caching_bg = false;
7886         ffe_ctl.orig_have_caching_bg = false;
7887         ffe_ctl.found_offset = 0;
7888
7889         ins->type = BTRFS_EXTENT_ITEM_KEY;
7890         ins->objectid = 0;
7891         ins->offset = 0;
7892
7893         trace_find_free_extent(fs_info, num_bytes, empty_size, flags);
7894
7895         space_info = __find_space_info(fs_info, flags);
7896         if (!space_info) {
7897                 btrfs_err(fs_info, "No space info for %llu", flags);
7898                 return -ENOSPC;
7899         }
7900
7901         /*
7902          * If our free space is heavily fragmented we may not be able to make
7903          * big contiguous allocations, so instead of doing the expensive search
7904          * for free space, simply return ENOSPC with our max_extent_size so we
7905          * can go ahead and search for a more manageable chunk.
7906          *
7907          * If our max_extent_size is large enough for our allocation simply
7908          * disable clustering since we will likely not be able to find enough
7909          * space to create a cluster and induce latency trying.
7910          */
7911         if (unlikely(space_info->max_extent_size)) {
7912                 spin_lock(&space_info->lock);
7913                 if (space_info->max_extent_size &&
7914                     num_bytes > space_info->max_extent_size) {
7915                         ins->offset = space_info->max_extent_size;
7916                         spin_unlock(&space_info->lock);
7917                         return -ENOSPC;
7918                 } else if (space_info->max_extent_size) {
7919                         use_cluster = false;
7920                 }
7921                 spin_unlock(&space_info->lock);
7922         }
7923
7924         last_ptr = fetch_cluster_info(fs_info, space_info,
7925                                       &ffe_ctl.empty_cluster);
7926         if (last_ptr) {
7927                 spin_lock(&last_ptr->lock);
7928                 if (last_ptr->block_group)
7929                         hint_byte = last_ptr->window_start;
7930                 if (last_ptr->fragmented) {
7931                         /*
7932                          * We still set window_start so we can keep track of the
7933                          * last place we found an allocation to try and save
7934                          * some time.
7935                          */
7936                         hint_byte = last_ptr->window_start;
7937                         use_cluster = false;
7938                 }
7939                 spin_unlock(&last_ptr->lock);
7940         }
7941
7942         ffe_ctl.search_start = max(ffe_ctl.search_start,
7943                                    first_logical_byte(fs_info, 0));
7944         ffe_ctl.search_start = max(ffe_ctl.search_start, hint_byte);
7945         if (ffe_ctl.search_start == hint_byte) {
7946                 block_group = btrfs_lookup_block_group(fs_info,
7947                                                        ffe_ctl.search_start);
7948                 /*
7949                  * we don't want to use the block group if it doesn't match our
7950                  * allocation bits, or if its not cached.
7951                  *
7952                  * However if we are re-searching with an ideal block group
7953                  * picked out then we don't care that the block group is cached.
7954                  */
7955                 if (block_group && block_group_bits(block_group, flags) &&
7956                     block_group->cached != BTRFS_CACHE_NO) {
7957                         down_read(&space_info->groups_sem);
7958                         if (list_empty(&block_group->list) ||
7959                             block_group->ro) {
7960                                 /*
7961                                  * someone is removing this block group,
7962                                  * we can't jump into the have_block_group
7963                                  * target because our list pointers are not
7964                                  * valid
7965                                  */
7966                                 btrfs_put_block_group(block_group);
7967                                 up_read(&space_info->groups_sem);
7968                         } else {
7969                                 ffe_ctl.index = btrfs_bg_flags_to_raid_index(
7970                                                 block_group->flags);
7971                                 btrfs_lock_block_group(block_group, delalloc);
7972                                 goto have_block_group;
7973                         }
7974                 } else if (block_group) {
7975                         btrfs_put_block_group(block_group);
7976                 }
7977         }
7978 search:
7979         ffe_ctl.have_caching_bg = false;
7980         if (ffe_ctl.index == btrfs_bg_flags_to_raid_index(flags) ||
7981             ffe_ctl.index == 0)
7982                 full_search = true;
7983         down_read(&space_info->groups_sem);
7984         list_for_each_entry(block_group,
7985                             &space_info->block_groups[ffe_ctl.index], list) {
7986                 /* If the block group is read-only, we can skip it entirely. */
7987                 if (unlikely(block_group->ro))
7988                         continue;
7989
7990                 btrfs_grab_block_group(block_group, delalloc);
7991                 ffe_ctl.search_start = block_group->key.objectid;
7992
7993                 /*
7994                  * this can happen if we end up cycling through all the
7995                  * raid types, but we want to make sure we only allocate
7996                  * for the proper type.
7997                  */
7998                 if (!block_group_bits(block_group, flags)) {
7999                         u64 extra = BTRFS_BLOCK_GROUP_DUP |
8000                                 BTRFS_BLOCK_GROUP_RAID1 |
8001                                 BTRFS_BLOCK_GROUP_RAID5 |
8002                                 BTRFS_BLOCK_GROUP_RAID6 |
8003                                 BTRFS_BLOCK_GROUP_RAID10;
8004
8005                         /*
8006                          * if they asked for extra copies and this block group
8007                          * doesn't provide them, bail.  This does allow us to
8008                          * fill raid0 from raid1.
8009                          */
8010                         if ((flags & extra) && !(block_group->flags & extra))
8011                                 goto loop;
8012                 }
8013
8014 have_block_group:
8015                 ffe_ctl.cached = block_group_cache_done(block_group);
8016                 if (unlikely(!ffe_ctl.cached)) {
8017                         ffe_ctl.have_caching_bg = true;
8018                         ret = cache_block_group(block_group, 0);
8019                         BUG_ON(ret < 0);
8020                         ret = 0;
8021                 }
8022
8023                 if (unlikely(block_group->cached == BTRFS_CACHE_ERROR))
8024                         goto loop;
8025
8026                 /*
8027                  * Ok we want to try and use the cluster allocator, so
8028                  * lets look there
8029                  */
8030                 if (last_ptr && use_cluster) {
8031                         struct btrfs_block_group_cache *cluster_bg = NULL;
8032
8033                         ret = find_free_extent_clustered(block_group, last_ptr,
8034                                                          &ffe_ctl, &cluster_bg);
8035
8036                         if (ret == 0) {
8037                                 if (cluster_bg && cluster_bg != block_group) {
8038                                         btrfs_release_block_group(block_group,
8039                                                                   delalloc);
8040                                         block_group = cluster_bg;
8041                                 }
8042                                 goto checks;
8043                         } else if (ret == -EAGAIN) {
8044                                 goto have_block_group;
8045                         } else if (ret > 0) {
8046                                 goto loop;
8047                         }
8048                         /* ret == -ENOENT case falls through */
8049                 }
8050
8051                 ret = find_free_extent_unclustered(block_group, last_ptr,
8052                                                    &ffe_ctl);
8053                 if (ret == -EAGAIN)
8054                         goto have_block_group;
8055                 else if (ret > 0)
8056                         goto loop;
8057                 /* ret == 0 case falls through */
8058 checks:
8059                 ffe_ctl.search_start = round_up(ffe_ctl.found_offset,
8060                                              fs_info->stripesize);
8061
8062                 /* move on to the next group */
8063                 if (ffe_ctl.search_start + num_bytes >
8064                     block_group->key.objectid + block_group->key.offset) {
8065                         btrfs_add_free_space(block_group, ffe_ctl.found_offset,
8066                                              num_bytes);
8067                         goto loop;
8068                 }
8069
8070                 if (ffe_ctl.found_offset < ffe_ctl.search_start)
8071                         btrfs_add_free_space(block_group, ffe_ctl.found_offset,
8072                                 ffe_ctl.search_start - ffe_ctl.found_offset);
8073
8074                 ret = btrfs_add_reserved_bytes(block_group, ram_bytes,
8075                                 num_bytes, delalloc);
8076                 if (ret == -EAGAIN) {
8077                         btrfs_add_free_space(block_group, ffe_ctl.found_offset,
8078                                              num_bytes);
8079                         goto loop;
8080                 }
8081                 btrfs_inc_block_group_reservations(block_group);
8082
8083                 /* we are all good, lets return */
8084                 ins->objectid = ffe_ctl.search_start;
8085                 ins->offset = num_bytes;
8086
8087                 trace_btrfs_reserve_extent(block_group, ffe_ctl.search_start,
8088                                            num_bytes);
8089                 btrfs_release_block_group(block_group, delalloc);
8090                 break;
8091 loop:
8092                 ffe_ctl.retry_clustered = false;
8093                 ffe_ctl.retry_unclustered = false;
8094                 BUG_ON(btrfs_bg_flags_to_raid_index(block_group->flags) !=
8095                        ffe_ctl.index);
8096                 btrfs_release_block_group(block_group, delalloc);
8097                 cond_resched();
8098         }
8099         up_read(&space_info->groups_sem);
8100
8101         ret = find_free_extent_update_loop(fs_info, last_ptr, ins, &ffe_ctl,
8102                                            full_search, use_cluster);
8103         if (ret > 0)
8104                 goto search;
8105
8106         if (ret == -ENOSPC) {
8107                 /*
8108                  * Use ffe_ctl->total_free_space as fallback if we can't find
8109                  * any contiguous hole.
8110                  */
8111                 if (!ffe_ctl.max_extent_size)
8112                         ffe_ctl.max_extent_size = ffe_ctl.total_free_space;
8113                 spin_lock(&space_info->lock);
8114                 space_info->max_extent_size = ffe_ctl.max_extent_size;
8115                 spin_unlock(&space_info->lock);
8116                 ins->offset = ffe_ctl.max_extent_size;
8117         }
8118         return ret;
8119 }
8120
8121 #define DUMP_BLOCK_RSV(fs_info, rsv_name)                               \
8122 do {                                                                    \
8123         struct btrfs_block_rsv *__rsv = &(fs_info)->rsv_name;           \
8124         spin_lock(&__rsv->lock);                                        \
8125         btrfs_info(fs_info, #rsv_name ": size %llu reserved %llu",      \
8126                    __rsv->size, __rsv->reserved);                       \
8127         spin_unlock(&__rsv->lock);                                      \
8128 } while (0)
8129
8130 static void dump_space_info(struct btrfs_fs_info *fs_info,
8131                             struct btrfs_space_info *info, u64 bytes,
8132                             int dump_block_groups)
8133 {
8134         struct btrfs_block_group_cache *cache;
8135         int index = 0;
8136
8137         spin_lock(&info->lock);
8138         btrfs_info(fs_info, "space_info %llu has %llu free, is %sfull",
8139                    info->flags,
8140                    info->total_bytes - btrfs_space_info_used(info, true),
8141                    info->full ? "" : "not ");
8142         btrfs_info(fs_info,
8143                 "space_info total=%llu, used=%llu, pinned=%llu, reserved=%llu, may_use=%llu, readonly=%llu",
8144                 info->total_bytes, info->bytes_used, info->bytes_pinned,
8145                 info->bytes_reserved, info->bytes_may_use,
8146                 info->bytes_readonly);
8147         spin_unlock(&info->lock);
8148
8149         DUMP_BLOCK_RSV(fs_info, global_block_rsv);
8150         DUMP_BLOCK_RSV(fs_info, trans_block_rsv);
8151         DUMP_BLOCK_RSV(fs_info, chunk_block_rsv);
8152         DUMP_BLOCK_RSV(fs_info, delayed_block_rsv);
8153         DUMP_BLOCK_RSV(fs_info, delayed_refs_rsv);
8154
8155         if (!dump_block_groups)
8156                 return;
8157
8158         down_read(&info->groups_sem);
8159 again:
8160         list_for_each_entry(cache, &info->block_groups[index], list) {
8161                 spin_lock(&cache->lock);
8162                 btrfs_info(fs_info,
8163                         "block group %llu has %llu bytes, %llu used %llu pinned %llu reserved %s",
8164                         cache->key.objectid, cache->key.offset,
8165                         btrfs_block_group_used(&cache->item), cache->pinned,
8166                         cache->reserved, cache->ro ? "[readonly]" : "");
8167                 btrfs_dump_free_space(cache, bytes);
8168                 spin_unlock(&cache->lock);
8169         }
8170         if (++index < BTRFS_NR_RAID_TYPES)
8171                 goto again;
8172         up_read(&info->groups_sem);
8173 }
8174
8175 /*
8176  * btrfs_reserve_extent - entry point to the extent allocator. Tries to find a
8177  *                        hole that is at least as big as @num_bytes.
8178  *
8179  * @root           -    The root that will contain this extent
8180  *
8181  * @ram_bytes      -    The amount of space in ram that @num_bytes take. This
8182  *                      is used for accounting purposes. This value differs
8183  *                      from @num_bytes only in the case of compressed extents.
8184  *
8185  * @num_bytes      -    Number of bytes to allocate on-disk.
8186  *
8187  * @min_alloc_size -    Indicates the minimum amount of space that the
8188  *                      allocator should try to satisfy. In some cases
8189  *                      @num_bytes may be larger than what is required and if
8190  *                      the filesystem is fragmented then allocation fails.
8191  *                      However, the presence of @min_alloc_size gives a
8192  *                      chance to try and satisfy the smaller allocation.
8193  *
8194  * @empty_size     -    A hint that you plan on doing more COW. This is the
8195  *                      size in bytes the allocator should try to find free
8196  *                      next to the block it returns.  This is just a hint and
8197  *                      may be ignored by the allocator.
8198  *
8199  * @hint_byte      -    Hint to the allocator to start searching above the byte
8200  *                      address passed. It might be ignored.
8201  *
8202  * @ins            -    This key is modified to record the found hole. It will
8203  *                      have the following values:
8204  *                      ins->objectid == start position
8205  *                      ins->flags = BTRFS_EXTENT_ITEM_KEY
8206  *                      ins->offset == the size of the hole.
8207  *
8208  * @is_data        -    Boolean flag indicating whether an extent is
8209  *                      allocated for data (true) or metadata (false)
8210  *
8211  * @delalloc       -    Boolean flag indicating whether this allocation is for
8212  *                      delalloc or not. If 'true' data_rwsem of block groups
8213  *                      is going to be acquired.
8214  *
8215  *
8216  * Returns 0 when an allocation succeeded or < 0 when an error occurred. In
8217  * case -ENOSPC is returned then @ins->offset will contain the size of the
8218  * largest available hole the allocator managed to find.
8219  */
8220 int btrfs_reserve_extent(struct btrfs_root *root, u64 ram_bytes,
8221                          u64 num_bytes, u64 min_alloc_size,
8222                          u64 empty_size, u64 hint_byte,
8223                          struct btrfs_key *ins, int is_data, int delalloc)
8224 {
8225         struct btrfs_fs_info *fs_info = root->fs_info;
8226         bool final_tried = num_bytes == min_alloc_size;
8227         u64 flags;
8228         int ret;
8229
8230         flags = get_alloc_profile_by_root(root, is_data);
8231 again:
8232         WARN_ON(num_bytes < fs_info->sectorsize);
8233         ret = find_free_extent(fs_info, ram_bytes, num_bytes, empty_size,
8234                                hint_byte, ins, flags, delalloc);
8235         if (!ret && !is_data) {
8236                 btrfs_dec_block_group_reservations(fs_info, ins->objectid);
8237         } else if (ret == -ENOSPC) {
8238                 if (!final_tried && ins->offset) {
8239                         num_bytes = min(num_bytes >> 1, ins->offset);
8240                         num_bytes = round_down(num_bytes,
8241                                                fs_info->sectorsize);
8242                         num_bytes = max(num_bytes, min_alloc_size);
8243                         ram_bytes = num_bytes;
8244                         if (num_bytes == min_alloc_size)
8245                                 final_tried = true;
8246                         goto again;
8247                 } else if (btrfs_test_opt(fs_info, ENOSPC_DEBUG)) {
8248                         struct btrfs_space_info *sinfo;
8249
8250                         sinfo = __find_space_info(fs_info, flags);
8251                         btrfs_err(fs_info,
8252                                   "allocation failed flags %llu, wanted %llu",
8253                                   flags, num_bytes);
8254                         if (sinfo)
8255                                 dump_space_info(fs_info, sinfo, num_bytes, 1);
8256                 }
8257         }
8258
8259         return ret;
8260 }
8261
8262 static int __btrfs_free_reserved_extent(struct btrfs_fs_info *fs_info,
8263                                         u64 start, u64 len,
8264                                         int pin, int delalloc)
8265 {
8266         struct btrfs_block_group_cache *cache;
8267         int ret = 0;
8268
8269         cache = btrfs_lookup_block_group(fs_info, start);
8270         if (!cache) {
8271                 btrfs_err(fs_info, "Unable to find block group for %llu",
8272                           start);
8273                 return -ENOSPC;
8274         }
8275
8276         if (pin)
8277                 pin_down_extent(fs_info, cache, start, len, 1);
8278         else {
8279                 if (btrfs_test_opt(fs_info, DISCARD))
8280                         ret = btrfs_discard_extent(fs_info, start, len, NULL);
8281                 btrfs_add_free_space(cache, start, len);
8282                 btrfs_free_reserved_bytes(cache, len, delalloc);
8283                 trace_btrfs_reserved_extent_free(fs_info, start, len);
8284         }
8285
8286         btrfs_put_block_group(cache);
8287         return ret;
8288 }
8289
8290 int btrfs_free_reserved_extent(struct btrfs_fs_info *fs_info,
8291                                u64 start, u64 len, int delalloc)
8292 {
8293         return __btrfs_free_reserved_extent(fs_info, start, len, 0, delalloc);
8294 }
8295
8296 int btrfs_free_and_pin_reserved_extent(struct btrfs_fs_info *fs_info,
8297                                        u64 start, u64 len)
8298 {
8299         return __btrfs_free_reserved_extent(fs_info, start, len, 1, 0);
8300 }
8301
8302 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
8303                                       u64 parent, u64 root_objectid,
8304                                       u64 flags, u64 owner, u64 offset,
8305                                       struct btrfs_key *ins, int ref_mod)
8306 {
8307         struct btrfs_fs_info *fs_info = trans->fs_info;
8308         int ret;
8309         struct btrfs_extent_item *extent_item;
8310         struct btrfs_extent_inline_ref *iref;
8311         struct btrfs_path *path;
8312         struct extent_buffer *leaf;
8313         int type;
8314         u32 size;
8315
8316         if (parent > 0)
8317                 type = BTRFS_SHARED_DATA_REF_KEY;
8318         else
8319                 type = BTRFS_EXTENT_DATA_REF_KEY;
8320
8321         size = sizeof(*extent_item) + btrfs_extent_inline_ref_size(type);
8322
8323         path = btrfs_alloc_path();
8324         if (!path)
8325                 return -ENOMEM;
8326
8327         path->leave_spinning = 1;
8328         ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path,
8329                                       ins, size);
8330         if (ret) {
8331                 btrfs_free_path(path);
8332                 return ret;
8333         }
8334
8335         leaf = path->nodes[0];
8336         extent_item = btrfs_item_ptr(leaf, path->slots[0],
8337                                      struct btrfs_extent_item);
8338         btrfs_set_extent_refs(leaf, extent_item, ref_mod);
8339         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
8340         btrfs_set_extent_flags(leaf, extent_item,
8341                                flags | BTRFS_EXTENT_FLAG_DATA);
8342
8343         iref = (struct btrfs_extent_inline_ref *)(extent_item + 1);
8344         btrfs_set_extent_inline_ref_type(leaf, iref, type);
8345         if (parent > 0) {
8346                 struct btrfs_shared_data_ref *ref;
8347                 ref = (struct btrfs_shared_data_ref *)(iref + 1);
8348                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
8349                 btrfs_set_shared_data_ref_count(leaf, ref, ref_mod);
8350         } else {
8351                 struct btrfs_extent_data_ref *ref;
8352                 ref = (struct btrfs_extent_data_ref *)(&iref->offset);
8353                 btrfs_set_extent_data_ref_root(leaf, ref, root_objectid);
8354                 btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
8355                 btrfs_set_extent_data_ref_offset(leaf, ref, offset);
8356                 btrfs_set_extent_data_ref_count(leaf, ref, ref_mod);
8357         }
8358
8359         btrfs_mark_buffer_dirty(path->nodes[0]);
8360         btrfs_free_path(path);
8361
8362         ret = remove_from_free_space_tree(trans, ins->objectid, ins->offset);
8363         if (ret)
8364                 return ret;
8365
8366         ret = update_block_group(trans, fs_info, ins->objectid, ins->offset, 1);
8367         if (ret) { /* -ENOENT, logic error */
8368                 btrfs_err(fs_info, "update block group failed for %llu %llu",
8369                         ins->objectid, ins->offset);
8370                 BUG();
8371         }
8372         trace_btrfs_reserved_extent_alloc(fs_info, ins->objectid, ins->offset);
8373         return ret;
8374 }
8375
8376 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
8377                                      struct btrfs_delayed_ref_node *node,
8378                                      struct btrfs_delayed_extent_op *extent_op)
8379 {
8380         struct btrfs_fs_info *fs_info = trans->fs_info;
8381         int ret;
8382         struct btrfs_extent_item *extent_item;
8383         struct btrfs_key extent_key;
8384         struct btrfs_tree_block_info *block_info;
8385         struct btrfs_extent_inline_ref *iref;
8386         struct btrfs_path *path;
8387         struct extent_buffer *leaf;
8388         struct btrfs_delayed_tree_ref *ref;
8389         u32 size = sizeof(*extent_item) + sizeof(*iref);
8390         u64 num_bytes;
8391         u64 flags = extent_op->flags_to_set;
8392         bool skinny_metadata = btrfs_fs_incompat(fs_info, SKINNY_METADATA);
8393
8394         ref = btrfs_delayed_node_to_tree_ref(node);
8395
8396         extent_key.objectid = node->bytenr;
8397         if (skinny_metadata) {
8398                 extent_key.offset = ref->level;
8399                 extent_key.type = BTRFS_METADATA_ITEM_KEY;
8400                 num_bytes = fs_info->nodesize;
8401         } else {
8402                 extent_key.offset = node->num_bytes;
8403                 extent_key.type = BTRFS_EXTENT_ITEM_KEY;
8404                 size += sizeof(*block_info);
8405                 num_bytes = node->num_bytes;
8406         }
8407
8408         path = btrfs_alloc_path();
8409         if (!path)
8410                 return -ENOMEM;
8411
8412         path->leave_spinning = 1;
8413         ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path,
8414                                       &extent_key, size);
8415         if (ret) {
8416                 btrfs_free_path(path);
8417                 return ret;
8418         }
8419
8420         leaf = path->nodes[0];
8421         extent_item = btrfs_item_ptr(leaf, path->slots[0],
8422                                      struct btrfs_extent_item);
8423         btrfs_set_extent_refs(leaf, extent_item, 1);
8424         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
8425         btrfs_set_extent_flags(leaf, extent_item,
8426                                flags | BTRFS_EXTENT_FLAG_TREE_BLOCK);
8427
8428         if (skinny_metadata) {
8429                 iref = (struct btrfs_extent_inline_ref *)(extent_item + 1);
8430         } else {
8431                 block_info = (struct btrfs_tree_block_info *)(extent_item + 1);
8432                 btrfs_set_tree_block_key(leaf, block_info, &extent_op->key);
8433                 btrfs_set_tree_block_level(leaf, block_info, ref->level);
8434                 iref = (struct btrfs_extent_inline_ref *)(block_info + 1);
8435         }
8436
8437         if (node->type == BTRFS_SHARED_BLOCK_REF_KEY) {
8438                 BUG_ON(!(flags & BTRFS_BLOCK_FLAG_FULL_BACKREF));
8439                 btrfs_set_extent_inline_ref_type(leaf, iref,
8440                                                  BTRFS_SHARED_BLOCK_REF_KEY);
8441                 btrfs_set_extent_inline_ref_offset(leaf, iref, ref->parent);
8442         } else {
8443                 btrfs_set_extent_inline_ref_type(leaf, iref,
8444                                                  BTRFS_TREE_BLOCK_REF_KEY);
8445                 btrfs_set_extent_inline_ref_offset(leaf, iref, ref->root);
8446         }
8447
8448         btrfs_mark_buffer_dirty(leaf);
8449         btrfs_free_path(path);
8450
8451         ret = remove_from_free_space_tree(trans, extent_key.objectid,
8452                                           num_bytes);
8453         if (ret)
8454                 return ret;
8455
8456         ret = update_block_group(trans, fs_info, extent_key.objectid,
8457                                  fs_info->nodesize, 1);
8458         if (ret) { /* -ENOENT, logic error */
8459                 btrfs_err(fs_info, "update block group failed for %llu %llu",
8460                         extent_key.objectid, extent_key.offset);
8461                 BUG();
8462         }
8463
8464         trace_btrfs_reserved_extent_alloc(fs_info, extent_key.objectid,
8465                                           fs_info->nodesize);
8466         return ret;
8467 }
8468
8469 int btrfs_alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
8470                                      struct btrfs_root *root, u64 owner,
8471                                      u64 offset, u64 ram_bytes,
8472                                      struct btrfs_key *ins)
8473 {
8474         int ret;
8475
8476         BUG_ON(root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID);
8477
8478         btrfs_ref_tree_mod(root, ins->objectid, ins->offset, 0,
8479                            root->root_key.objectid, owner, offset,
8480                            BTRFS_ADD_DELAYED_EXTENT);
8481
8482         ret = btrfs_add_delayed_data_ref(trans, ins->objectid,
8483                                          ins->offset, 0,
8484                                          root->root_key.objectid, owner,
8485                                          offset, ram_bytes,
8486                                          BTRFS_ADD_DELAYED_EXTENT, NULL, NULL);
8487         return ret;
8488 }
8489
8490 /*
8491  * this is used by the tree logging recovery code.  It records that
8492  * an extent has been allocated and makes sure to clear the free
8493  * space cache bits as well
8494  */
8495 int btrfs_alloc_logged_file_extent(struct btrfs_trans_handle *trans,
8496                                    u64 root_objectid, u64 owner, u64 offset,
8497                                    struct btrfs_key *ins)
8498 {
8499         struct btrfs_fs_info *fs_info = trans->fs_info;
8500         int ret;
8501         struct btrfs_block_group_cache *block_group;
8502         struct btrfs_space_info *space_info;
8503
8504         /*
8505          * Mixed block groups will exclude before processing the log so we only
8506          * need to do the exclude dance if this fs isn't mixed.
8507          */
8508         if (!btrfs_fs_incompat(fs_info, MIXED_GROUPS)) {
8509                 ret = __exclude_logged_extent(fs_info, ins->objectid,
8510                                               ins->offset);
8511                 if (ret)
8512                         return ret;
8513         }
8514
8515         block_group = btrfs_lookup_block_group(fs_info, ins->objectid);
8516         if (!block_group)
8517                 return -EINVAL;
8518
8519         space_info = block_group->space_info;
8520         spin_lock(&space_info->lock);
8521         spin_lock(&block_group->lock);
8522         space_info->bytes_reserved += ins->offset;
8523         block_group->reserved += ins->offset;
8524         spin_unlock(&block_group->lock);
8525         spin_unlock(&space_info->lock);
8526
8527         ret = alloc_reserved_file_extent(trans, 0, root_objectid, 0, owner,
8528                                          offset, ins, 1);
8529         btrfs_put_block_group(block_group);
8530         return ret;
8531 }
8532
8533 static struct extent_buffer *
8534 btrfs_init_new_buffer(struct btrfs_trans_handle *trans, struct btrfs_root *root,
8535                       u64 bytenr, int level, u64 owner)
8536 {
8537         struct btrfs_fs_info *fs_info = root->fs_info;
8538         struct extent_buffer *buf;
8539
8540         buf = btrfs_find_create_tree_block(fs_info, bytenr);
8541         if (IS_ERR(buf))
8542                 return buf;
8543
8544         /*
8545          * Extra safety check in case the extent tree is corrupted and extent
8546          * allocator chooses to use a tree block which is already used and
8547          * locked.
8548          */
8549         if (buf->lock_owner == current->pid) {
8550                 btrfs_err_rl(fs_info,
8551 "tree block %llu owner %llu already locked by pid=%d, extent tree corruption detected",
8552                         buf->start, btrfs_header_owner(buf), current->pid);
8553                 free_extent_buffer(buf);
8554                 return ERR_PTR(-EUCLEAN);
8555         }
8556
8557         btrfs_set_buffer_lockdep_class(root->root_key.objectid, buf, level);
8558         btrfs_tree_lock(buf);
8559         btrfs_clean_tree_block(buf);
8560         clear_bit(EXTENT_BUFFER_STALE, &buf->bflags);
8561
8562         btrfs_set_lock_blocking_write(buf);
8563         set_extent_buffer_uptodate(buf);
8564
8565         memzero_extent_buffer(buf, 0, sizeof(struct btrfs_header));
8566         btrfs_set_header_level(buf, level);
8567         btrfs_set_header_bytenr(buf, buf->start);
8568         btrfs_set_header_generation(buf, trans->transid);
8569         btrfs_set_header_backref_rev(buf, BTRFS_MIXED_BACKREF_REV);
8570         btrfs_set_header_owner(buf, owner);
8571         write_extent_buffer_fsid(buf, fs_info->fs_devices->metadata_uuid);
8572         write_extent_buffer_chunk_tree_uuid(buf, fs_info->chunk_tree_uuid);
8573         if (root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID) {
8574                 buf->log_index = root->log_transid % 2;
8575                 /*
8576                  * we allow two log transactions at a time, use different
8577                  * EXTENT bit to differentiate dirty pages.
8578                  */
8579                 if (buf->log_index == 0)
8580                         set_extent_dirty(&root->dirty_log_pages, buf->start,
8581                                         buf->start + buf->len - 1, GFP_NOFS);
8582                 else
8583                         set_extent_new(&root->dirty_log_pages, buf->start,
8584                                         buf->start + buf->len - 1);
8585         } else {
8586                 buf->log_index = -1;
8587                 set_extent_dirty(&trans->transaction->dirty_pages, buf->start,
8588                          buf->start + buf->len - 1, GFP_NOFS);
8589         }
8590         trans->dirty = true;
8591         /* this returns a buffer locked for blocking */
8592         return buf;
8593 }
8594
8595 static struct btrfs_block_rsv *
8596 use_block_rsv(struct btrfs_trans_handle *trans,
8597               struct btrfs_root *root, u32 blocksize)
8598 {
8599         struct btrfs_fs_info *fs_info = root->fs_info;
8600         struct btrfs_block_rsv *block_rsv;
8601         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
8602         int ret;
8603         bool global_updated = false;
8604
8605         block_rsv = get_block_rsv(trans, root);
8606
8607         if (unlikely(block_rsv->size == 0))
8608                 goto try_reserve;
8609 again:
8610         ret = block_rsv_use_bytes(block_rsv, blocksize);
8611         if (!ret)
8612                 return block_rsv;
8613
8614         if (block_rsv->failfast)
8615                 return ERR_PTR(ret);
8616
8617         if (block_rsv->type == BTRFS_BLOCK_RSV_GLOBAL && !global_updated) {
8618                 global_updated = true;
8619                 update_global_block_rsv(fs_info);
8620                 goto again;
8621         }
8622
8623         /*
8624          * The global reserve still exists to save us from ourselves, so don't
8625          * warn_on if we are short on our delayed refs reserve.
8626          */
8627         if (block_rsv->type != BTRFS_BLOCK_RSV_DELREFS &&
8628             btrfs_test_opt(fs_info, ENOSPC_DEBUG)) {
8629                 static DEFINE_RATELIMIT_STATE(_rs,
8630                                 DEFAULT_RATELIMIT_INTERVAL * 10,
8631                                 /*DEFAULT_RATELIMIT_BURST*/ 1);
8632                 if (__ratelimit(&_rs))
8633                         WARN(1, KERN_DEBUG
8634                                 "BTRFS: block rsv returned %d\n", ret);
8635         }
8636 try_reserve:
8637         ret = reserve_metadata_bytes(root, block_rsv, blocksize,
8638                                      BTRFS_RESERVE_NO_FLUSH);
8639         if (!ret)
8640                 return block_rsv;
8641         /*
8642          * If we couldn't reserve metadata bytes try and use some from
8643          * the global reserve if its space type is the same as the global
8644          * reservation.
8645          */
8646         if (block_rsv->type != BTRFS_BLOCK_RSV_GLOBAL &&
8647             block_rsv->space_info == global_rsv->space_info) {
8648                 ret = block_rsv_use_bytes(global_rsv, blocksize);
8649                 if (!ret)
8650                         return global_rsv;
8651         }
8652         return ERR_PTR(ret);
8653 }
8654
8655 static void unuse_block_rsv(struct btrfs_fs_info *fs_info,
8656                             struct btrfs_block_rsv *block_rsv, u32 blocksize)
8657 {
8658         block_rsv_add_bytes(block_rsv, blocksize, false);
8659         block_rsv_release_bytes(fs_info, block_rsv, NULL, 0, NULL);
8660 }
8661
8662 /*
8663  * finds a free extent and does all the dirty work required for allocation
8664  * returns the tree buffer or an ERR_PTR on error.
8665  */
8666 struct extent_buffer *btrfs_alloc_tree_block(struct btrfs_trans_handle *trans,
8667                                              struct btrfs_root *root,
8668                                              u64 parent, u64 root_objectid,
8669                                              const struct btrfs_disk_key *key,
8670                                              int level, u64 hint,
8671                                              u64 empty_size)
8672 {
8673         struct btrfs_fs_info *fs_info = root->fs_info;
8674         struct btrfs_key ins;
8675         struct btrfs_block_rsv *block_rsv;
8676         struct extent_buffer *buf;
8677         struct btrfs_delayed_extent_op *extent_op;
8678         u64 flags = 0;
8679         int ret;
8680         u32 blocksize = fs_info->nodesize;
8681         bool skinny_metadata = btrfs_fs_incompat(fs_info, SKINNY_METADATA);
8682
8683 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
8684         if (btrfs_is_testing(fs_info)) {
8685                 buf = btrfs_init_new_buffer(trans, root, root->alloc_bytenr,
8686                                             level, root_objectid);
8687                 if (!IS_ERR(buf))
8688                         root->alloc_bytenr += blocksize;
8689                 return buf;
8690         }
8691 #endif
8692
8693         block_rsv = use_block_rsv(trans, root, blocksize);
8694         if (IS_ERR(block_rsv))
8695                 return ERR_CAST(block_rsv);
8696
8697         ret = btrfs_reserve_extent(root, blocksize, blocksize, blocksize,
8698                                    empty_size, hint, &ins, 0, 0);
8699         if (ret)
8700                 goto out_unuse;
8701
8702         buf = btrfs_init_new_buffer(trans, root, ins.objectid, level,
8703                                     root_objectid);
8704         if (IS_ERR(buf)) {
8705                 ret = PTR_ERR(buf);
8706                 goto out_free_reserved;
8707         }
8708
8709         if (root_objectid == BTRFS_TREE_RELOC_OBJECTID) {
8710                 if (parent == 0)
8711                         parent = ins.objectid;
8712                 flags |= BTRFS_BLOCK_FLAG_FULL_BACKREF;
8713         } else
8714                 BUG_ON(parent > 0);
8715
8716         if (root_objectid != BTRFS_TREE_LOG_OBJECTID) {
8717                 extent_op = btrfs_alloc_delayed_extent_op();
8718                 if (!extent_op) {
8719                         ret = -ENOMEM;
8720                         goto out_free_buf;
8721                 }
8722                 if (key)
8723                         memcpy(&extent_op->key, key, sizeof(extent_op->key));
8724                 else
8725                         memset(&extent_op->key, 0, sizeof(extent_op->key));
8726                 extent_op->flags_to_set = flags;
8727                 extent_op->update_key = skinny_metadata ? false : true;
8728                 extent_op->update_flags = true;
8729                 extent_op->is_data = false;
8730                 extent_op->level = level;
8731
8732                 btrfs_ref_tree_mod(root, ins.objectid, ins.offset, parent,
8733                                    root_objectid, level, 0,
8734                                    BTRFS_ADD_DELAYED_EXTENT);
8735                 ret = btrfs_add_delayed_tree_ref(trans, ins.objectid,
8736                                                  ins.offset, parent,
8737                                                  root_objectid, level,
8738                                                  BTRFS_ADD_DELAYED_EXTENT,
8739                                                  extent_op, NULL, NULL);
8740                 if (ret)
8741                         goto out_free_delayed;
8742         }
8743         return buf;
8744
8745 out_free_delayed:
8746         btrfs_free_delayed_extent_op(extent_op);
8747 out_free_buf:
8748         free_extent_buffer(buf);
8749 out_free_reserved:
8750         btrfs_free_reserved_extent(fs_info, ins.objectid, ins.offset, 0);
8751 out_unuse:
8752         unuse_block_rsv(fs_info, block_rsv, blocksize);
8753         return ERR_PTR(ret);
8754 }
8755
8756 struct walk_control {
8757         u64 refs[BTRFS_MAX_LEVEL];
8758         u64 flags[BTRFS_MAX_LEVEL];
8759         struct btrfs_key update_progress;
8760         struct btrfs_key drop_progress;
8761         int drop_level;
8762         int stage;
8763         int level;
8764         int shared_level;
8765         int update_ref;
8766         int keep_locks;
8767         int reada_slot;
8768         int reada_count;
8769         int restarted;
8770 };
8771
8772 #define DROP_REFERENCE  1
8773 #define UPDATE_BACKREF  2
8774
8775 static noinline void reada_walk_down(struct btrfs_trans_handle *trans,
8776                                      struct btrfs_root *root,
8777                                      struct walk_control *wc,
8778                                      struct btrfs_path *path)
8779 {
8780         struct btrfs_fs_info *fs_info = root->fs_info;
8781         u64 bytenr;
8782         u64 generation;
8783         u64 refs;
8784         u64 flags;
8785         u32 nritems;
8786         struct btrfs_key key;
8787         struct extent_buffer *eb;
8788         int ret;
8789         int slot;
8790         int nread = 0;
8791
8792         if (path->slots[wc->level] < wc->reada_slot) {
8793                 wc->reada_count = wc->reada_count * 2 / 3;
8794                 wc->reada_count = max(wc->reada_count, 2);
8795         } else {
8796                 wc->reada_count = wc->reada_count * 3 / 2;
8797                 wc->reada_count = min_t(int, wc->reada_count,
8798                                         BTRFS_NODEPTRS_PER_BLOCK(fs_info));
8799         }
8800
8801         eb = path->nodes[wc->level];
8802         nritems = btrfs_header_nritems(eb);
8803
8804         for (slot = path->slots[wc->level]; slot < nritems; slot++) {
8805                 if (nread >= wc->reada_count)
8806                         break;
8807
8808                 cond_resched();
8809                 bytenr = btrfs_node_blockptr(eb, slot);
8810                 generation = btrfs_node_ptr_generation(eb, slot);
8811
8812                 if (slot == path->slots[wc->level])
8813                         goto reada;
8814
8815                 if (wc->stage == UPDATE_BACKREF &&
8816                     generation <= root->root_key.offset)
8817                         continue;
8818
8819                 /* We don't lock the tree block, it's OK to be racy here */
8820                 ret = btrfs_lookup_extent_info(trans, fs_info, bytenr,
8821                                                wc->level - 1, 1, &refs,
8822                                                &flags);
8823                 /* We don't care about errors in readahead. */
8824                 if (ret < 0)
8825                         continue;
8826                 BUG_ON(refs == 0);
8827
8828                 if (wc->stage == DROP_REFERENCE) {
8829                         if (refs == 1)
8830                                 goto reada;
8831
8832                         if (wc->level == 1 &&
8833                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
8834                                 continue;
8835                         if (!wc->update_ref ||
8836                             generation <= root->root_key.offset)
8837                                 continue;
8838                         btrfs_node_key_to_cpu(eb, &key, slot);
8839                         ret = btrfs_comp_cpu_keys(&key,
8840                                                   &wc->update_progress);
8841                         if (ret < 0)
8842                                 continue;
8843                 } else {
8844                         if (wc->level == 1 &&
8845                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
8846                                 continue;
8847                 }
8848 reada:
8849                 readahead_tree_block(fs_info, bytenr);
8850                 nread++;
8851         }
8852         wc->reada_slot = slot;
8853 }
8854
8855 /*
8856  * helper to process tree block while walking down the tree.
8857  *
8858  * when wc->stage == UPDATE_BACKREF, this function updates
8859  * back refs for pointers in the block.
8860  *
8861  * NOTE: return value 1 means we should stop walking down.
8862  */
8863 static noinline int walk_down_proc(struct btrfs_trans_handle *trans,
8864                                    struct btrfs_root *root,
8865                                    struct btrfs_path *path,
8866                                    struct walk_control *wc, int lookup_info)
8867 {
8868         struct btrfs_fs_info *fs_info = root->fs_info;
8869         int level = wc->level;
8870         struct extent_buffer *eb = path->nodes[level];
8871         u64 flag = BTRFS_BLOCK_FLAG_FULL_BACKREF;
8872         int ret;
8873
8874         if (wc->stage == UPDATE_BACKREF &&
8875             btrfs_header_owner(eb) != root->root_key.objectid)
8876                 return 1;
8877
8878         /*
8879          * when reference count of tree block is 1, it won't increase
8880          * again. once full backref flag is set, we never clear it.
8881          */
8882         if (lookup_info &&
8883             ((wc->stage == DROP_REFERENCE && wc->refs[level] != 1) ||
8884              (wc->stage == UPDATE_BACKREF && !(wc->flags[level] & flag)))) {
8885                 BUG_ON(!path->locks[level]);
8886                 ret = btrfs_lookup_extent_info(trans, fs_info,
8887                                                eb->start, level, 1,
8888                                                &wc->refs[level],
8889                                                &wc->flags[level]);
8890                 BUG_ON(ret == -ENOMEM);
8891                 if (ret)
8892                         return ret;
8893                 BUG_ON(wc->refs[level] == 0);
8894         }
8895
8896         if (wc->stage == DROP_REFERENCE) {
8897                 if (wc->refs[level] > 1)
8898                         return 1;
8899
8900                 if (path->locks[level] && !wc->keep_locks) {
8901                         btrfs_tree_unlock_rw(eb, path->locks[level]);
8902                         path->locks[level] = 0;
8903                 }
8904                 return 0;
8905         }
8906
8907         /* wc->stage == UPDATE_BACKREF */
8908         if (!(wc->flags[level] & flag)) {
8909                 BUG_ON(!path->locks[level]);
8910                 ret = btrfs_inc_ref(trans, root, eb, 1);
8911                 BUG_ON(ret); /* -ENOMEM */
8912                 ret = btrfs_dec_ref(trans, root, eb, 0);
8913                 BUG_ON(ret); /* -ENOMEM */
8914                 ret = btrfs_set_disk_extent_flags(trans, fs_info, eb->start,
8915                                                   eb->len, flag,
8916                                                   btrfs_header_level(eb), 0);
8917                 BUG_ON(ret); /* -ENOMEM */
8918                 wc->flags[level] |= flag;
8919         }
8920
8921         /*
8922          * the block is shared by multiple trees, so it's not good to
8923          * keep the tree lock
8924          */
8925         if (path->locks[level] && level > 0) {
8926                 btrfs_tree_unlock_rw(eb, path->locks[level]);
8927                 path->locks[level] = 0;
8928         }
8929         return 0;
8930 }
8931
8932 /*
8933  * This is used to verify a ref exists for this root to deal with a bug where we
8934  * would have a drop_progress key that hadn't been updated properly.
8935  */
8936 static int check_ref_exists(struct btrfs_trans_handle *trans,
8937                             struct btrfs_root *root, u64 bytenr, u64 parent,
8938                             int level)
8939 {
8940         struct btrfs_path *path;
8941         struct btrfs_extent_inline_ref *iref;
8942         int ret;
8943
8944         path = btrfs_alloc_path();
8945         if (!path)
8946                 return -ENOMEM;
8947
8948         ret = lookup_extent_backref(trans, path, &iref, bytenr,
8949                                     root->fs_info->nodesize, parent,
8950                                     root->root_key.objectid, level, 0);
8951         btrfs_free_path(path);
8952         if (ret == -ENOENT)
8953                 return 0;
8954         if (ret < 0)
8955                 return ret;
8956         return 1;
8957 }
8958
8959 /*
8960  * helper to process tree block pointer.
8961  *
8962  * when wc->stage == DROP_REFERENCE, this function checks
8963  * reference count of the block pointed to. if the block
8964  * is shared and we need update back refs for the subtree
8965  * rooted at the block, this function changes wc->stage to
8966  * UPDATE_BACKREF. if the block is shared and there is no
8967  * need to update back, this function drops the reference
8968  * to the block.
8969  *
8970  * NOTE: return value 1 means we should stop walking down.
8971  */
8972 static noinline int do_walk_down(struct btrfs_trans_handle *trans,
8973                                  struct btrfs_root *root,
8974                                  struct btrfs_path *path,
8975                                  struct walk_control *wc, int *lookup_info)
8976 {
8977         struct btrfs_fs_info *fs_info = root->fs_info;
8978         u64 bytenr;
8979         u64 generation;
8980         u64 parent;
8981         struct btrfs_key key;
8982         struct btrfs_key first_key;
8983         struct extent_buffer *next;
8984         int level = wc->level;
8985         int reada = 0;
8986         int ret = 0;
8987         bool need_account = false;
8988
8989         generation = btrfs_node_ptr_generation(path->nodes[level],
8990                                                path->slots[level]);
8991         /*
8992          * if the lower level block was created before the snapshot
8993          * was created, we know there is no need to update back refs
8994          * for the subtree
8995          */
8996         if (wc->stage == UPDATE_BACKREF &&
8997             generation <= root->root_key.offset) {
8998                 *lookup_info = 1;
8999                 return 1;
9000         }
9001
9002         bytenr = btrfs_node_blockptr(path->nodes[level], path->slots[level]);
9003         btrfs_node_key_to_cpu(path->nodes[level], &first_key,
9004                               path->slots[level]);
9005
9006         next = find_extent_buffer(fs_info, bytenr);
9007         if (!next) {
9008                 next = btrfs_find_create_tree_block(fs_info, bytenr);
9009                 if (IS_ERR(next))
9010                         return PTR_ERR(next);
9011
9012                 btrfs_set_buffer_lockdep_class(root->root_key.objectid, next,
9013                                                level - 1);
9014                 reada = 1;
9015         }
9016         btrfs_tree_lock(next);
9017         btrfs_set_lock_blocking_write(next);
9018
9019         ret = btrfs_lookup_extent_info(trans, fs_info, bytenr, level - 1, 1,
9020                                        &wc->refs[level - 1],
9021                                        &wc->flags[level - 1]);
9022         if (ret < 0)
9023                 goto out_unlock;
9024
9025         if (unlikely(wc->refs[level - 1] == 0)) {
9026                 btrfs_err(fs_info, "Missing references.");
9027                 ret = -EIO;
9028                 goto out_unlock;
9029         }
9030         *lookup_info = 0;
9031
9032         if (wc->stage == DROP_REFERENCE) {
9033                 if (wc->refs[level - 1] > 1) {
9034                         need_account = true;
9035                         if (level == 1 &&
9036                             (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
9037                                 goto skip;
9038
9039                         if (!wc->update_ref ||
9040                             generation <= root->root_key.offset)
9041                                 goto skip;
9042
9043                         btrfs_node_key_to_cpu(path->nodes[level], &key,
9044                                               path->slots[level]);
9045                         ret = btrfs_comp_cpu_keys(&key, &wc->update_progress);
9046                         if (ret < 0)
9047                                 goto skip;
9048
9049                         wc->stage = UPDATE_BACKREF;
9050                         wc->shared_level = level - 1;
9051                 }
9052         } else {
9053                 if (level == 1 &&
9054                     (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
9055                         goto skip;
9056         }
9057
9058         if (!btrfs_buffer_uptodate(next, generation, 0)) {
9059                 btrfs_tree_unlock(next);
9060                 free_extent_buffer(next);
9061                 next = NULL;
9062                 *lookup_info = 1;
9063         }
9064
9065         if (!next) {
9066                 if (reada && level == 1)
9067                         reada_walk_down(trans, root, wc, path);
9068                 next = read_tree_block(fs_info, bytenr, generation, level - 1,
9069                                        &first_key);
9070                 if (IS_ERR(next)) {
9071                         return PTR_ERR(next);
9072                 } else if (!extent_buffer_uptodate(next)) {
9073                         free_extent_buffer(next);
9074                         return -EIO;
9075                 }
9076                 btrfs_tree_lock(next);
9077                 btrfs_set_lock_blocking_write(next);
9078         }
9079
9080         level--;
9081         ASSERT(level == btrfs_header_level(next));
9082         if (level != btrfs_header_level(next)) {
9083                 btrfs_err(root->fs_info, "mismatched level");
9084                 ret = -EIO;
9085                 goto out_unlock;
9086         }
9087         path->nodes[level] = next;
9088         path->slots[level] = 0;
9089         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
9090         wc->level = level;
9091         if (wc->level == 1)
9092                 wc->reada_slot = 0;
9093         return 0;
9094 skip:
9095         wc->refs[level - 1] = 0;
9096         wc->flags[level - 1] = 0;
9097         if (wc->stage == DROP_REFERENCE) {
9098                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF) {
9099                         parent = path->nodes[level]->start;
9100                 } else {
9101                         ASSERT(root->root_key.objectid ==
9102                                btrfs_header_owner(path->nodes[level]));
9103                         if (root->root_key.objectid !=
9104                             btrfs_header_owner(path->nodes[level])) {
9105                                 btrfs_err(root->fs_info,
9106                                                 "mismatched block owner");
9107                                 ret = -EIO;
9108                                 goto out_unlock;
9109                         }
9110                         parent = 0;
9111                 }
9112
9113                 /*
9114                  * If we had a drop_progress we need to verify the refs are set
9115                  * as expected.  If we find our ref then we know that from here
9116                  * on out everything should be correct, and we can clear the
9117                  * ->restarted flag.
9118                  */
9119                 if (wc->restarted) {
9120                         ret = check_ref_exists(trans, root, bytenr, parent,
9121                                                level - 1);
9122                         if (ret < 0)
9123                                 goto out_unlock;
9124                         if (ret == 0)
9125                                 goto no_delete;
9126                         ret = 0;
9127                         wc->restarted = 0;
9128                 }
9129
9130                 /*
9131                  * Reloc tree doesn't contribute to qgroup numbers, and we have
9132                  * already accounted them at merge time (replace_path),
9133                  * thus we could skip expensive subtree trace here.
9134                  */
9135                 if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID &&
9136                     need_account) {
9137                         ret = btrfs_qgroup_trace_subtree(trans, next,
9138                                                          generation, level - 1);
9139                         if (ret) {
9140                                 btrfs_err_rl(fs_info,
9141                                              "Error %d accounting shared subtree. Quota is out of sync, rescan required.",
9142                                              ret);
9143                         }
9144                 }
9145
9146                 /*
9147                  * We need to update the next key in our walk control so we can
9148                  * update the drop_progress key accordingly.  We don't care if
9149                  * find_next_key doesn't find a key because that means we're at
9150                  * the end and are going to clean up now.
9151                  */
9152                 wc->drop_level = level;
9153                 find_next_key(path, level, &wc->drop_progress);
9154
9155                 ret = btrfs_free_extent(trans, root, bytenr, fs_info->nodesize,
9156                                         parent, root->root_key.objectid,
9157                                         level - 1, 0);
9158                 if (ret)
9159                         goto out_unlock;
9160         }
9161 no_delete:
9162         *lookup_info = 1;
9163         ret = 1;
9164
9165 out_unlock:
9166         btrfs_tree_unlock(next);
9167         free_extent_buffer(next);
9168
9169         return ret;
9170 }
9171
9172 /*
9173  * helper to process tree block while walking up the tree.
9174  *
9175  * when wc->stage == DROP_REFERENCE, this function drops
9176  * reference count on the block.
9177  *
9178  * when wc->stage == UPDATE_BACKREF, this function changes
9179  * wc->stage back to DROP_REFERENCE if we changed wc->stage
9180  * to UPDATE_BACKREF previously while processing the block.
9181  *
9182  * NOTE: return value 1 means we should stop walking up.
9183  */
9184 static noinline int walk_up_proc(struct btrfs_trans_handle *trans,
9185                                  struct btrfs_root *root,
9186                                  struct btrfs_path *path,
9187                                  struct walk_control *wc)
9188 {
9189         struct btrfs_fs_info *fs_info = root->fs_info;
9190         int ret;
9191         int level = wc->level;
9192         struct extent_buffer *eb = path->nodes[level];
9193         u64 parent = 0;
9194
9195         if (wc->stage == UPDATE_BACKREF) {
9196                 BUG_ON(wc->shared_level < level);
9197                 if (level < wc->shared_level)
9198                         goto out;
9199
9200                 ret = find_next_key(path, level + 1, &wc->update_progress);
9201                 if (ret > 0)
9202                         wc->update_ref = 0;
9203
9204                 wc->stage = DROP_REFERENCE;
9205                 wc->shared_level = -1;
9206                 path->slots[level] = 0;
9207
9208                 /*
9209                  * check reference count again if the block isn't locked.
9210                  * we should start walking down the tree again if reference
9211                  * count is one.
9212                  */
9213                 if (!path->locks[level]) {
9214                         BUG_ON(level == 0);
9215                         btrfs_tree_lock(eb);
9216                         btrfs_set_lock_blocking_write(eb);
9217                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
9218
9219                         ret = btrfs_lookup_extent_info(trans, fs_info,
9220                                                        eb->start, level, 1,
9221                                                        &wc->refs[level],
9222                                                        &wc->flags[level]);
9223                         if (ret < 0) {
9224                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
9225                                 path->locks[level] = 0;
9226                                 return ret;
9227                         }
9228                         BUG_ON(wc->refs[level] == 0);
9229                         if (wc->refs[level] == 1) {
9230                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
9231                                 path->locks[level] = 0;
9232                                 return 1;
9233                         }
9234                 }
9235         }
9236
9237         /* wc->stage == DROP_REFERENCE */
9238         BUG_ON(wc->refs[level] > 1 && !path->locks[level]);
9239
9240         if (wc->refs[level] == 1) {
9241                 if (level == 0) {
9242                         if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
9243                                 ret = btrfs_dec_ref(trans, root, eb, 1);
9244                         else
9245                                 ret = btrfs_dec_ref(trans, root, eb, 0);
9246                         BUG_ON(ret); /* -ENOMEM */
9247                         ret = btrfs_qgroup_trace_leaf_items(trans, eb);
9248                         if (ret) {
9249                                 btrfs_err_rl(fs_info,
9250                                              "error %d accounting leaf items. Quota is out of sync, rescan required.",
9251                                              ret);
9252                         }
9253                 }
9254                 /* make block locked assertion in btrfs_clean_tree_block happy */
9255                 if (!path->locks[level] &&
9256                     btrfs_header_generation(eb) == trans->transid) {
9257                         btrfs_tree_lock(eb);
9258                         btrfs_set_lock_blocking_write(eb);
9259                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
9260                 }
9261                 btrfs_clean_tree_block(eb);
9262         }
9263
9264         if (eb == root->node) {
9265                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
9266                         parent = eb->start;
9267                 else if (root->root_key.objectid != btrfs_header_owner(eb))
9268                         goto owner_mismatch;
9269         } else {
9270                 if (wc->flags[level + 1] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
9271                         parent = path->nodes[level + 1]->start;
9272                 else if (root->root_key.objectid !=
9273                          btrfs_header_owner(path->nodes[level + 1]))
9274                         goto owner_mismatch;
9275         }
9276
9277         btrfs_free_tree_block(trans, root, eb, parent, wc->refs[level] == 1);
9278 out:
9279         wc->refs[level] = 0;
9280         wc->flags[level] = 0;
9281         return 0;
9282
9283 owner_mismatch:
9284         btrfs_err_rl(fs_info, "unexpected tree owner, have %llu expect %llu",
9285                      btrfs_header_owner(eb), root->root_key.objectid);
9286         return -EUCLEAN;
9287 }
9288
9289 static noinline int walk_down_tree(struct btrfs_trans_handle *trans,
9290                                    struct btrfs_root *root,
9291                                    struct btrfs_path *path,
9292                                    struct walk_control *wc)
9293 {
9294         int level = wc->level;
9295         int lookup_info = 1;
9296         int ret;
9297
9298         while (level >= 0) {
9299                 ret = walk_down_proc(trans, root, path, wc, lookup_info);
9300                 if (ret > 0)
9301                         break;
9302
9303                 if (level == 0)
9304                         break;
9305
9306                 if (path->slots[level] >=
9307                     btrfs_header_nritems(path->nodes[level]))
9308                         break;
9309
9310                 ret = do_walk_down(trans, root, path, wc, &lookup_info);
9311                 if (ret > 0) {
9312                         path->slots[level]++;
9313                         continue;
9314                 } else if (ret < 0)
9315                         return ret;
9316                 level = wc->level;
9317         }
9318         return 0;
9319 }
9320
9321 static noinline int walk_up_tree(struct btrfs_trans_handle *trans,
9322                                  struct btrfs_root *root,
9323                                  struct btrfs_path *path,
9324                                  struct walk_control *wc, int max_level)
9325 {
9326         int level = wc->level;
9327         int ret;
9328
9329         path->slots[level] = btrfs_header_nritems(path->nodes[level]);
9330         while (level < max_level && path->nodes[level]) {
9331                 wc->level = level;
9332                 if (path->slots[level] + 1 <
9333                     btrfs_header_nritems(path->nodes[level])) {
9334                         path->slots[level]++;
9335                         return 0;
9336                 } else {
9337                         ret = walk_up_proc(trans, root, path, wc);
9338                         if (ret > 0)
9339                                 return 0;
9340                         if (ret < 0)
9341                                 return ret;
9342
9343                         if (path->locks[level]) {
9344                                 btrfs_tree_unlock_rw(path->nodes[level],
9345                                                      path->locks[level]);
9346                                 path->locks[level] = 0;
9347                         }
9348                         free_extent_buffer(path->nodes[level]);
9349                         path->nodes[level] = NULL;
9350                         level++;
9351                 }
9352         }
9353         return 1;
9354 }
9355
9356 /*
9357  * drop a subvolume tree.
9358  *
9359  * this function traverses the tree freeing any blocks that only
9360  * referenced by the tree.
9361  *
9362  * when a shared tree block is found. this function decreases its
9363  * reference count by one. if update_ref is true, this function
9364  * also make sure backrefs for the shared block and all lower level
9365  * blocks are properly updated.
9366  *
9367  * If called with for_reloc == 0, may exit early with -EAGAIN
9368  */
9369 int btrfs_drop_snapshot(struct btrfs_root *root,
9370                          struct btrfs_block_rsv *block_rsv, int update_ref,
9371                          int for_reloc)
9372 {
9373         struct btrfs_fs_info *fs_info = root->fs_info;
9374         struct btrfs_path *path;
9375         struct btrfs_trans_handle *trans;
9376         struct btrfs_root *tree_root = fs_info->tree_root;
9377         struct btrfs_root_item *root_item = &root->root_item;
9378         struct walk_control *wc;
9379         struct btrfs_key key;
9380         int err = 0;
9381         int ret;
9382         int level;
9383         bool root_dropped = false;
9384
9385         btrfs_debug(fs_info, "Drop subvolume %llu", root->root_key.objectid);
9386
9387         path = btrfs_alloc_path();
9388         if (!path) {
9389                 err = -ENOMEM;
9390                 goto out;
9391         }
9392
9393         wc = kzalloc(sizeof(*wc), GFP_NOFS);
9394         if (!wc) {
9395                 btrfs_free_path(path);
9396                 err = -ENOMEM;
9397                 goto out;
9398         }
9399
9400         trans = btrfs_start_transaction(tree_root, 0);
9401         if (IS_ERR(trans)) {
9402                 err = PTR_ERR(trans);
9403                 goto out_free;
9404         }
9405
9406         err = btrfs_run_delayed_items(trans);
9407         if (err)
9408                 goto out_end_trans;
9409
9410         if (block_rsv)
9411                 trans->block_rsv = block_rsv;
9412
9413         /*
9414          * This will help us catch people modifying the fs tree while we're
9415          * dropping it.  It is unsafe to mess with the fs tree while it's being
9416          * dropped as we unlock the root node and parent nodes as we walk down
9417          * the tree, assuming nothing will change.  If something does change
9418          * then we'll have stale information and drop references to blocks we've
9419          * already dropped.
9420          */
9421         set_bit(BTRFS_ROOT_DELETING, &root->state);
9422         if (btrfs_disk_key_objectid(&root_item->drop_progress) == 0) {
9423                 level = btrfs_header_level(root->node);
9424                 path->nodes[level] = btrfs_lock_root_node(root);
9425                 btrfs_set_lock_blocking_write(path->nodes[level]);
9426                 path->slots[level] = 0;
9427                 path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
9428                 memset(&wc->update_progress, 0,
9429                        sizeof(wc->update_progress));
9430         } else {
9431                 btrfs_disk_key_to_cpu(&key, &root_item->drop_progress);
9432                 memcpy(&wc->update_progress, &key,
9433                        sizeof(wc->update_progress));
9434
9435                 level = root_item->drop_level;
9436                 BUG_ON(level == 0);
9437                 path->lowest_level = level;
9438                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
9439                 path->lowest_level = 0;
9440                 if (ret < 0) {
9441                         err = ret;
9442                         goto out_end_trans;
9443                 }
9444                 WARN_ON(ret > 0);
9445
9446                 /*
9447                  * unlock our path, this is safe because only this
9448                  * function is allowed to delete this snapshot
9449                  */
9450                 btrfs_unlock_up_safe(path, 0);
9451
9452                 level = btrfs_header_level(root->node);
9453                 while (1) {
9454                         btrfs_tree_lock(path->nodes[level]);
9455                         btrfs_set_lock_blocking_write(path->nodes[level]);
9456                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
9457
9458                         ret = btrfs_lookup_extent_info(trans, fs_info,
9459                                                 path->nodes[level]->start,
9460                                                 level, 1, &wc->refs[level],
9461                                                 &wc->flags[level]);
9462                         if (ret < 0) {
9463                                 err = ret;
9464                                 goto out_end_trans;
9465                         }
9466                         BUG_ON(wc->refs[level] == 0);
9467
9468                         if (level == root_item->drop_level)
9469                                 break;
9470
9471                         btrfs_tree_unlock(path->nodes[level]);
9472                         path->locks[level] = 0;
9473                         WARN_ON(wc->refs[level] != 1);
9474                         level--;
9475                 }
9476         }
9477
9478         wc->restarted = test_bit(BTRFS_ROOT_DEAD_TREE, &root->state);
9479         wc->level = level;
9480         wc->shared_level = -1;
9481         wc->stage = DROP_REFERENCE;
9482         wc->update_ref = update_ref;
9483         wc->keep_locks = 0;
9484         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(fs_info);
9485
9486         while (1) {
9487
9488                 ret = walk_down_tree(trans, root, path, wc);
9489                 if (ret < 0) {
9490                         err = ret;
9491                         break;
9492                 }
9493
9494                 ret = walk_up_tree(trans, root, path, wc, BTRFS_MAX_LEVEL);
9495                 if (ret < 0) {
9496                         err = ret;
9497                         break;
9498                 }
9499
9500                 if (ret > 0) {
9501                         BUG_ON(wc->stage != DROP_REFERENCE);
9502                         break;
9503                 }
9504
9505                 if (wc->stage == DROP_REFERENCE) {
9506                         wc->drop_level = wc->level;
9507                         btrfs_node_key_to_cpu(path->nodes[wc->drop_level],
9508                                               &wc->drop_progress,
9509                                               path->slots[wc->drop_level]);
9510                 }
9511                 btrfs_cpu_key_to_disk(&root_item->drop_progress,
9512                                       &wc->drop_progress);
9513                 root_item->drop_level = wc->drop_level;
9514
9515                 BUG_ON(wc->level == 0);
9516                 if (btrfs_should_end_transaction(trans) ||
9517                     (!for_reloc && btrfs_need_cleaner_sleep(fs_info))) {
9518                         ret = btrfs_update_root(trans, tree_root,
9519                                                 &root->root_key,
9520                                                 root_item);
9521                         if (ret) {
9522                                 btrfs_abort_transaction(trans, ret);
9523                                 err = ret;
9524                                 goto out_end_trans;
9525                         }
9526
9527                         btrfs_end_transaction_throttle(trans);
9528                         if (!for_reloc && btrfs_need_cleaner_sleep(fs_info)) {
9529                                 btrfs_debug(fs_info,
9530                                             "drop snapshot early exit");
9531                                 err = -EAGAIN;
9532                                 goto out_free;
9533                         }
9534
9535                         trans = btrfs_start_transaction(tree_root, 0);
9536                         if (IS_ERR(trans)) {
9537                                 err = PTR_ERR(trans);
9538                                 goto out_free;
9539                         }
9540                         if (block_rsv)
9541                                 trans->block_rsv = block_rsv;
9542                 }
9543         }
9544         btrfs_release_path(path);
9545         if (err)
9546                 goto out_end_trans;
9547
9548         ret = btrfs_del_root(trans, &root->root_key);
9549         if (ret) {
9550                 btrfs_abort_transaction(trans, ret);
9551                 err = ret;
9552                 goto out_end_trans;
9553         }
9554
9555         if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) {
9556                 ret = btrfs_find_root(tree_root, &root->root_key, path,
9557                                       NULL, NULL);
9558                 if (ret < 0) {
9559                         btrfs_abort_transaction(trans, ret);
9560                         err = ret;
9561                         goto out_end_trans;
9562                 } else if (ret > 0) {
9563                         /* if we fail to delete the orphan item this time
9564                          * around, it'll get picked up the next time.
9565                          *
9566                          * The most common failure here is just -ENOENT.
9567                          */
9568                         btrfs_del_orphan_item(trans, tree_root,
9569                                               root->root_key.objectid);
9570                 }
9571         }
9572
9573         if (test_bit(BTRFS_ROOT_IN_RADIX, &root->state)) {
9574                 btrfs_add_dropped_root(trans, root);
9575         } else {
9576                 free_extent_buffer(root->node);
9577                 free_extent_buffer(root->commit_root);
9578                 btrfs_put_fs_root(root);
9579         }
9580         root_dropped = true;
9581 out_end_trans:
9582         btrfs_end_transaction_throttle(trans);
9583 out_free:
9584         kfree(wc);
9585         btrfs_free_path(path);
9586 out:
9587         /*
9588          * So if we need to stop dropping the snapshot for whatever reason we
9589          * need to make sure to add it back to the dead root list so that we
9590          * keep trying to do the work later.  This also cleans up roots if we
9591          * don't have it in the radix (like when we recover after a power fail
9592          * or unmount) so we don't leak memory.
9593          */
9594         if (!for_reloc && !root_dropped)
9595                 btrfs_add_dead_root(root);
9596         if (err && err != -EAGAIN)
9597                 btrfs_handle_fs_error(fs_info, err, NULL);
9598         return err;
9599 }
9600
9601 /*
9602  * drop subtree rooted at tree block 'node'.
9603  *
9604  * NOTE: this function will unlock and release tree block 'node'
9605  * only used by relocation code
9606  */
9607 int btrfs_drop_subtree(struct btrfs_trans_handle *trans,
9608                         struct btrfs_root *root,
9609                         struct extent_buffer *node,
9610                         struct extent_buffer *parent)
9611 {
9612         struct btrfs_fs_info *fs_info = root->fs_info;
9613         struct btrfs_path *path;
9614         struct walk_control *wc;
9615         int level;
9616         int parent_level;
9617         int ret = 0;
9618         int wret;
9619
9620         BUG_ON(root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID);
9621
9622         path = btrfs_alloc_path();
9623         if (!path)
9624                 return -ENOMEM;
9625
9626         wc = kzalloc(sizeof(*wc), GFP_NOFS);
9627         if (!wc) {
9628                 btrfs_free_path(path);
9629                 return -ENOMEM;
9630         }
9631
9632         btrfs_assert_tree_locked(parent);
9633         parent_level = btrfs_header_level(parent);
9634         extent_buffer_get(parent);
9635         path->nodes[parent_level] = parent;
9636         path->slots[parent_level] = btrfs_header_nritems(parent);
9637
9638         btrfs_assert_tree_locked(node);
9639         level = btrfs_header_level(node);
9640         path->nodes[level] = node;
9641         path->slots[level] = 0;
9642         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
9643
9644         wc->refs[parent_level] = 1;
9645         wc->flags[parent_level] = BTRFS_BLOCK_FLAG_FULL_BACKREF;
9646         wc->level = level;
9647         wc->shared_level = -1;
9648         wc->stage = DROP_REFERENCE;
9649         wc->update_ref = 0;
9650         wc->keep_locks = 1;
9651         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(fs_info);
9652
9653         while (1) {
9654                 wret = walk_down_tree(trans, root, path, wc);
9655                 if (wret < 0) {
9656                         ret = wret;
9657                         break;
9658                 }
9659
9660                 wret = walk_up_tree(trans, root, path, wc, parent_level);
9661                 if (wret < 0)
9662                         ret = wret;
9663                 if (wret != 0)
9664                         break;
9665         }
9666
9667         kfree(wc);
9668         btrfs_free_path(path);
9669         return ret;
9670 }
9671
9672 static u64 update_block_group_flags(struct btrfs_fs_info *fs_info, u64 flags)
9673 {
9674         u64 num_devices;
9675         u64 stripped;
9676
9677         /*
9678          * if restripe for this chunk_type is on pick target profile and
9679          * return, otherwise do the usual balance
9680          */
9681         stripped = get_restripe_target(fs_info, flags);
9682         if (stripped)
9683                 return extended_to_chunk(stripped);
9684
9685         num_devices = fs_info->fs_devices->rw_devices;
9686
9687         stripped = BTRFS_BLOCK_GROUP_RAID0 |
9688                 BTRFS_BLOCK_GROUP_RAID5 | BTRFS_BLOCK_GROUP_RAID6 |
9689                 BTRFS_BLOCK_GROUP_RAID1 | BTRFS_BLOCK_GROUP_RAID10;
9690
9691         if (num_devices == 1) {
9692                 stripped |= BTRFS_BLOCK_GROUP_DUP;
9693                 stripped = flags & ~stripped;
9694
9695                 /* turn raid0 into single device chunks */
9696                 if (flags & BTRFS_BLOCK_GROUP_RAID0)
9697                         return stripped;
9698
9699                 /* turn mirroring into duplication */
9700                 if (flags & (BTRFS_BLOCK_GROUP_RAID1 |
9701                              BTRFS_BLOCK_GROUP_RAID10))
9702                         return stripped | BTRFS_BLOCK_GROUP_DUP;
9703         } else {
9704                 /* they already had raid on here, just return */
9705                 if (flags & stripped)
9706                         return flags;
9707
9708                 stripped |= BTRFS_BLOCK_GROUP_DUP;
9709                 stripped = flags & ~stripped;
9710
9711                 /* switch duplicated blocks with raid1 */
9712                 if (flags & BTRFS_BLOCK_GROUP_DUP)
9713                         return stripped | BTRFS_BLOCK_GROUP_RAID1;
9714
9715                 /* this is drive concat, leave it alone */
9716         }
9717
9718         return flags;
9719 }
9720
9721 static int inc_block_group_ro(struct btrfs_block_group_cache *cache, int force)
9722 {
9723         struct btrfs_space_info *sinfo = cache->space_info;
9724         u64 num_bytes;
9725         u64 sinfo_used;
9726         u64 min_allocable_bytes;
9727         int ret = -ENOSPC;
9728
9729         /*
9730          * We need some metadata space and system metadata space for
9731          * allocating chunks in some corner cases until we force to set
9732          * it to be readonly.
9733          */
9734         if ((sinfo->flags &
9735              (BTRFS_BLOCK_GROUP_SYSTEM | BTRFS_BLOCK_GROUP_METADATA)) &&
9736             !force)
9737                 min_allocable_bytes = SZ_1M;
9738         else
9739                 min_allocable_bytes = 0;
9740
9741         spin_lock(&sinfo->lock);
9742         spin_lock(&cache->lock);
9743
9744         if (cache->ro) {
9745                 cache->ro++;
9746                 ret = 0;
9747                 goto out;
9748         }
9749
9750         num_bytes = cache->key.offset - cache->reserved - cache->pinned -
9751                     cache->bytes_super - btrfs_block_group_used(&cache->item);
9752         sinfo_used = btrfs_space_info_used(sinfo, true);
9753
9754         if (sinfo_used + num_bytes + min_allocable_bytes <=
9755             sinfo->total_bytes) {
9756                 sinfo->bytes_readonly += num_bytes;
9757                 cache->ro++;
9758                 list_add_tail(&cache->ro_list, &sinfo->ro_bgs);
9759                 ret = 0;
9760         }
9761 out:
9762         spin_unlock(&cache->lock);
9763         spin_unlock(&sinfo->lock);
9764         if (ret == -ENOSPC && btrfs_test_opt(cache->fs_info, ENOSPC_DEBUG)) {
9765                 btrfs_info(cache->fs_info,
9766                         "unable to make block group %llu ro",
9767                         cache->key.objectid);
9768                 btrfs_info(cache->fs_info,
9769                         "sinfo_used=%llu bg_num_bytes=%llu min_allocable=%llu",
9770                         sinfo_used, num_bytes, min_allocable_bytes);
9771                 dump_space_info(cache->fs_info, cache->space_info, 0, 0);
9772         }
9773         return ret;
9774 }
9775
9776 int btrfs_inc_block_group_ro(struct btrfs_block_group_cache *cache)
9777
9778 {
9779         struct btrfs_fs_info *fs_info = cache->fs_info;
9780         struct btrfs_trans_handle *trans;
9781         u64 alloc_flags;
9782         int ret;
9783
9784 again:
9785         trans = btrfs_join_transaction(fs_info->extent_root);
9786         if (IS_ERR(trans))
9787                 return PTR_ERR(trans);
9788
9789         /*
9790          * we're not allowed to set block groups readonly after the dirty
9791          * block groups cache has started writing.  If it already started,
9792          * back off and let this transaction commit
9793          */
9794         mutex_lock(&fs_info->ro_block_group_mutex);
9795         if (test_bit(BTRFS_TRANS_DIRTY_BG_RUN, &trans->transaction->flags)) {
9796                 u64 transid = trans->transid;
9797
9798                 mutex_unlock(&fs_info->ro_block_group_mutex);
9799                 btrfs_end_transaction(trans);
9800
9801                 ret = btrfs_wait_for_commit(fs_info, transid);
9802                 if (ret)
9803                         return ret;
9804                 goto again;
9805         }
9806
9807         /*
9808          * if we are changing raid levels, try to allocate a corresponding
9809          * block group with the new raid level.
9810          */
9811         alloc_flags = update_block_group_flags(fs_info, cache->flags);
9812         if (alloc_flags != cache->flags) {
9813                 ret = do_chunk_alloc(trans, alloc_flags,
9814                                      CHUNK_ALLOC_FORCE);
9815                 /*
9816                  * ENOSPC is allowed here, we may have enough space
9817                  * already allocated at the new raid level to
9818                  * carry on
9819                  */
9820                 if (ret == -ENOSPC)
9821                         ret = 0;
9822                 if (ret < 0)
9823                         goto out;
9824         }
9825
9826         ret = inc_block_group_ro(cache, 0);
9827         if (!ret)
9828                 goto out;
9829         alloc_flags = get_alloc_profile(fs_info, cache->space_info->flags);
9830         ret = do_chunk_alloc(trans, alloc_flags, CHUNK_ALLOC_FORCE);
9831         if (ret < 0)
9832                 goto out;
9833         ret = inc_block_group_ro(cache, 0);
9834 out:
9835         if (cache->flags & BTRFS_BLOCK_GROUP_SYSTEM) {
9836                 alloc_flags = update_block_group_flags(fs_info, cache->flags);
9837                 mutex_lock(&fs_info->chunk_mutex);
9838                 check_system_chunk(trans, alloc_flags);
9839                 mutex_unlock(&fs_info->chunk_mutex);
9840         }
9841         mutex_unlock(&fs_info->ro_block_group_mutex);
9842
9843         btrfs_end_transaction(trans);
9844         return ret;
9845 }
9846
9847 int btrfs_force_chunk_alloc(struct btrfs_trans_handle *trans, u64 type)
9848 {
9849         u64 alloc_flags = get_alloc_profile(trans->fs_info, type);
9850
9851         return do_chunk_alloc(trans, alloc_flags, CHUNK_ALLOC_FORCE);
9852 }
9853
9854 /*
9855  * helper to account the unused space of all the readonly block group in the
9856  * space_info. takes mirrors into account.
9857  */
9858 u64 btrfs_account_ro_block_groups_free_space(struct btrfs_space_info *sinfo)
9859 {
9860         struct btrfs_block_group_cache *block_group;
9861         u64 free_bytes = 0;
9862         int factor;
9863
9864         /* It's df, we don't care if it's racy */
9865         if (list_empty(&sinfo->ro_bgs))
9866                 return 0;
9867
9868         spin_lock(&sinfo->lock);
9869         list_for_each_entry(block_group, &sinfo->ro_bgs, ro_list) {
9870                 spin_lock(&block_group->lock);
9871
9872                 if (!block_group->ro) {
9873                         spin_unlock(&block_group->lock);
9874                         continue;
9875                 }
9876
9877                 factor = btrfs_bg_type_to_factor(block_group->flags);
9878                 free_bytes += (block_group->key.offset -
9879                                btrfs_block_group_used(&block_group->item)) *
9880                                factor;
9881
9882                 spin_unlock(&block_group->lock);
9883         }
9884         spin_unlock(&sinfo->lock);
9885
9886         return free_bytes;
9887 }
9888
9889 void btrfs_dec_block_group_ro(struct btrfs_block_group_cache *cache)
9890 {
9891         struct btrfs_space_info *sinfo = cache->space_info;
9892         u64 num_bytes;
9893
9894         BUG_ON(!cache->ro);
9895
9896         spin_lock(&sinfo->lock);
9897         spin_lock(&cache->lock);
9898         if (!--cache->ro) {
9899                 num_bytes = cache->key.offset - cache->reserved -
9900                             cache->pinned - cache->bytes_super -
9901                             btrfs_block_group_used(&cache->item);
9902                 sinfo->bytes_readonly -= num_bytes;
9903                 list_del_init(&cache->ro_list);
9904         }
9905         spin_unlock(&cache->lock);
9906         spin_unlock(&sinfo->lock);
9907 }
9908
9909 /*
9910  * Checks to see if it's even possible to relocate this block group.
9911  *
9912  * @return - -1 if it's not a good idea to relocate this block group, 0 if its
9913  * ok to go ahead and try.
9914  */
9915 int btrfs_can_relocate(struct btrfs_fs_info *fs_info, u64 bytenr)
9916 {
9917         struct btrfs_block_group_cache *block_group;
9918         struct btrfs_space_info *space_info;
9919         struct btrfs_fs_devices *fs_devices = fs_info->fs_devices;
9920         struct btrfs_device *device;
9921         u64 min_free;
9922         u64 dev_min = 1;
9923         u64 dev_nr = 0;
9924         u64 target;
9925         int debug;
9926         int index;
9927         int full = 0;
9928         int ret = 0;
9929
9930         debug = btrfs_test_opt(fs_info, ENOSPC_DEBUG);
9931
9932         block_group = btrfs_lookup_block_group(fs_info, bytenr);
9933
9934         /* odd, couldn't find the block group, leave it alone */
9935         if (!block_group) {
9936                 if (debug)
9937                         btrfs_warn(fs_info,
9938                                    "can't find block group for bytenr %llu",
9939                                    bytenr);
9940                 return -1;
9941         }
9942
9943         min_free = btrfs_block_group_used(&block_group->item);
9944
9945         /* no bytes used, we're good */
9946         if (!min_free)
9947                 goto out;
9948
9949         space_info = block_group->space_info;
9950         spin_lock(&space_info->lock);
9951
9952         full = space_info->full;
9953
9954         /*
9955          * if this is the last block group we have in this space, we can't
9956          * relocate it unless we're able to allocate a new chunk below.
9957          *
9958          * Otherwise, we need to make sure we have room in the space to handle
9959          * all of the extents from this block group.  If we can, we're good
9960          */
9961         if ((space_info->total_bytes != block_group->key.offset) &&
9962             (btrfs_space_info_used(space_info, false) + min_free <
9963              space_info->total_bytes)) {
9964                 spin_unlock(&space_info->lock);
9965                 goto out;
9966         }
9967         spin_unlock(&space_info->lock);
9968
9969         /*
9970          * ok we don't have enough space, but maybe we have free space on our
9971          * devices to allocate new chunks for relocation, so loop through our
9972          * alloc devices and guess if we have enough space.  if this block
9973          * group is going to be restriped, run checks against the target
9974          * profile instead of the current one.
9975          */
9976         ret = -1;
9977
9978         /*
9979          * index:
9980          *      0: raid10
9981          *      1: raid1
9982          *      2: dup
9983          *      3: raid0
9984          *      4: single
9985          */
9986         target = get_restripe_target(fs_info, block_group->flags);
9987         if (target) {
9988                 index = btrfs_bg_flags_to_raid_index(extended_to_chunk(target));
9989         } else {
9990                 /*
9991                  * this is just a balance, so if we were marked as full
9992                  * we know there is no space for a new chunk
9993                  */
9994                 if (full) {
9995                         if (debug)
9996                                 btrfs_warn(fs_info,
9997                                            "no space to alloc new chunk for block group %llu",
9998                                            block_group->key.objectid);
9999                         goto out;
10000                 }
10001
10002                 index = btrfs_bg_flags_to_raid_index(block_group->flags);
10003         }
10004
10005         if (index == BTRFS_RAID_RAID10) {
10006                 dev_min = 4;
10007                 /* Divide by 2 */
10008                 min_free >>= 1;
10009         } else if (index == BTRFS_RAID_RAID1) {
10010                 dev_min = 2;
10011         } else if (index == BTRFS_RAID_DUP) {
10012                 /* Multiply by 2 */
10013                 min_free <<= 1;
10014         } else if (index == BTRFS_RAID_RAID0) {
10015                 dev_min = fs_devices->rw_devices;
10016                 min_free = div64_u64(min_free, dev_min);
10017         }
10018
10019         mutex_lock(&fs_info->chunk_mutex);
10020         list_for_each_entry(device, &fs_devices->alloc_list, dev_alloc_list) {
10021                 u64 dev_offset;
10022
10023                 /*
10024                  * check to make sure we can actually find a chunk with enough
10025                  * space to fit our block group in.
10026                  */
10027                 if (device->total_bytes > device->bytes_used + min_free &&
10028                     !test_bit(BTRFS_DEV_STATE_REPLACE_TGT, &device->dev_state)) {
10029                         ret = find_free_dev_extent(device, min_free,
10030                                                    &dev_offset, NULL);
10031                         if (!ret)
10032                                 dev_nr++;
10033
10034                         if (dev_nr >= dev_min)
10035                                 break;
10036
10037                         ret = -1;
10038                 }
10039         }
10040         if (debug && ret == -1)
10041                 btrfs_warn(fs_info,
10042                            "no space to allocate a new chunk for block group %llu",
10043                            block_group->key.objectid);
10044         mutex_unlock(&fs_info->chunk_mutex);
10045 out:
10046         btrfs_put_block_group(block_group);
10047         return ret;
10048 }
10049
10050 static int find_first_block_group(struct btrfs_fs_info *fs_info,
10051                                   struct btrfs_path *path,
10052                                   struct btrfs_key *key)
10053 {
10054         struct btrfs_root *root = fs_info->extent_root;
10055         int ret = 0;
10056         struct btrfs_key found_key;
10057         struct extent_buffer *leaf;
10058         struct btrfs_block_group_item bg;
10059         u64 flags;
10060         int slot;
10061
10062         ret = btrfs_search_slot(NULL, root, key, path, 0, 0);
10063         if (ret < 0)
10064                 goto out;
10065
10066         while (1) {
10067                 slot = path->slots[0];
10068                 leaf = path->nodes[0];
10069                 if (slot >= btrfs_header_nritems(leaf)) {
10070                         ret = btrfs_next_leaf(root, path);
10071                         if (ret == 0)
10072                                 continue;
10073                         if (ret < 0)
10074                                 goto out;
10075                         break;
10076                 }
10077                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
10078
10079                 if (found_key.objectid >= key->objectid &&
10080                     found_key.type == BTRFS_BLOCK_GROUP_ITEM_KEY) {
10081                         struct extent_map_tree *em_tree;
10082                         struct extent_map *em;
10083
10084                         em_tree = &root->fs_info->mapping_tree.map_tree;
10085                         read_lock(&em_tree->lock);
10086                         em = lookup_extent_mapping(em_tree, found_key.objectid,
10087                                                    found_key.offset);
10088                         read_unlock(&em_tree->lock);
10089                         if (!em) {
10090                                 btrfs_err(fs_info,
10091                         "logical %llu len %llu found bg but no related chunk",
10092                                           found_key.objectid, found_key.offset);
10093                                 ret = -ENOENT;
10094                         } else if (em->start != found_key.objectid ||
10095                                    em->len != found_key.offset) {
10096                                 btrfs_err(fs_info,
10097                 "block group %llu len %llu mismatch with chunk %llu len %llu",
10098                                           found_key.objectid, found_key.offset,
10099                                           em->start, em->len);
10100                                 ret = -EUCLEAN;
10101                         } else {
10102                                 read_extent_buffer(leaf, &bg,
10103                                         btrfs_item_ptr_offset(leaf, slot),
10104                                         sizeof(bg));
10105                                 flags = btrfs_block_group_flags(&bg) &
10106                                         BTRFS_BLOCK_GROUP_TYPE_MASK;
10107
10108                                 if (flags != (em->map_lookup->type &
10109                                               BTRFS_BLOCK_GROUP_TYPE_MASK)) {
10110                                         btrfs_err(fs_info,
10111 "block group %llu len %llu type flags 0x%llx mismatch with chunk type flags 0x%llx",
10112                                                 found_key.objectid,
10113                                                 found_key.offset, flags,
10114                                                 (BTRFS_BLOCK_GROUP_TYPE_MASK &
10115                                                  em->map_lookup->type));
10116                                         ret = -EUCLEAN;
10117                                 } else {
10118                                         ret = 0;
10119                                 }
10120                         }
10121                         free_extent_map(em);
10122                         goto out;
10123                 }
10124                 path->slots[0]++;
10125         }
10126 out:
10127         return ret;
10128 }
10129
10130 void btrfs_put_block_group_cache(struct btrfs_fs_info *info)
10131 {
10132         struct btrfs_block_group_cache *block_group;
10133         u64 last = 0;
10134
10135         while (1) {
10136                 struct inode *inode;
10137
10138                 block_group = btrfs_lookup_first_block_group(info, last);
10139                 while (block_group) {
10140                         wait_block_group_cache_done(block_group);
10141                         spin_lock(&block_group->lock);
10142                         if (block_group->iref)
10143                                 break;
10144                         spin_unlock(&block_group->lock);
10145                         block_group = next_block_group(info, block_group);
10146                 }
10147                 if (!block_group) {
10148                         if (last == 0)
10149                                 break;
10150                         last = 0;
10151                         continue;
10152                 }
10153
10154                 inode = block_group->inode;
10155                 block_group->iref = 0;
10156                 block_group->inode = NULL;
10157                 spin_unlock(&block_group->lock);
10158                 ASSERT(block_group->io_ctl.inode == NULL);
10159                 iput(inode);
10160                 last = block_group->key.objectid + block_group->key.offset;
10161                 btrfs_put_block_group(block_group);
10162         }
10163 }
10164
10165 /*
10166  * Must be called only after stopping all workers, since we could have block
10167  * group caching kthreads running, and therefore they could race with us if we
10168  * freed the block groups before stopping them.
10169  */
10170 int btrfs_free_block_groups(struct btrfs_fs_info *info)
10171 {
10172         struct btrfs_block_group_cache *block_group;
10173         struct btrfs_space_info *space_info;
10174         struct btrfs_caching_control *caching_ctl;
10175         struct rb_node *n;
10176
10177         down_write(&info->commit_root_sem);
10178         while (!list_empty(&info->caching_block_groups)) {
10179                 caching_ctl = list_entry(info->caching_block_groups.next,
10180                                          struct btrfs_caching_control, list);
10181                 list_del(&caching_ctl->list);
10182                 put_caching_control(caching_ctl);
10183         }
10184         up_write(&info->commit_root_sem);
10185
10186         spin_lock(&info->unused_bgs_lock);
10187         while (!list_empty(&info->unused_bgs)) {
10188                 block_group = list_first_entry(&info->unused_bgs,
10189                                                struct btrfs_block_group_cache,
10190                                                bg_list);
10191                 list_del_init(&block_group->bg_list);
10192                 btrfs_put_block_group(block_group);
10193         }
10194         spin_unlock(&info->unused_bgs_lock);
10195
10196         spin_lock(&info->block_group_cache_lock);
10197         while ((n = rb_last(&info->block_group_cache_tree)) != NULL) {
10198                 block_group = rb_entry(n, struct btrfs_block_group_cache,
10199                                        cache_node);
10200                 rb_erase(&block_group->cache_node,
10201                          &info->block_group_cache_tree);
10202                 RB_CLEAR_NODE(&block_group->cache_node);
10203                 spin_unlock(&info->block_group_cache_lock);
10204
10205                 down_write(&block_group->space_info->groups_sem);
10206                 list_del(&block_group->list);
10207                 up_write(&block_group->space_info->groups_sem);
10208
10209                 /*
10210                  * We haven't cached this block group, which means we could
10211                  * possibly have excluded extents on this block group.
10212                  */
10213                 if (block_group->cached == BTRFS_CACHE_NO ||
10214                     block_group->cached == BTRFS_CACHE_ERROR)
10215                         free_excluded_extents(block_group);
10216
10217                 btrfs_remove_free_space_cache(block_group);
10218                 ASSERT(block_group->cached != BTRFS_CACHE_STARTED);
10219                 ASSERT(list_empty(&block_group->dirty_list));
10220                 ASSERT(list_empty(&block_group->io_list));
10221                 ASSERT(list_empty(&block_group->bg_list));
10222                 ASSERT(atomic_read(&block_group->count) == 1);
10223                 btrfs_put_block_group(block_group);
10224
10225                 spin_lock(&info->block_group_cache_lock);
10226         }
10227         spin_unlock(&info->block_group_cache_lock);
10228
10229         /* now that all the block groups are freed, go through and
10230          * free all the space_info structs.  This is only called during
10231          * the final stages of unmount, and so we know nobody is
10232          * using them.  We call synchronize_rcu() once before we start,
10233          * just to be on the safe side.
10234          */
10235         synchronize_rcu();
10236
10237         release_global_block_rsv(info);
10238
10239         while (!list_empty(&info->space_info)) {
10240                 int i;
10241
10242                 space_info = list_entry(info->space_info.next,
10243                                         struct btrfs_space_info,
10244                                         list);
10245
10246                 /*
10247                  * Do not hide this behind enospc_debug, this is actually
10248                  * important and indicates a real bug if this happens.
10249                  */
10250                 if (WARN_ON(space_info->bytes_pinned > 0 ||
10251                             space_info->bytes_reserved > 0 ||
10252                             space_info->bytes_may_use > 0))
10253                         dump_space_info(info, space_info, 0, 0);
10254                 list_del(&space_info->list);
10255                 for (i = 0; i < BTRFS_NR_RAID_TYPES; i++) {
10256                         struct kobject *kobj;
10257                         kobj = space_info->block_group_kobjs[i];
10258                         space_info->block_group_kobjs[i] = NULL;
10259                         if (kobj) {
10260                                 kobject_del(kobj);
10261                                 kobject_put(kobj);
10262                         }
10263                 }
10264                 kobject_del(&space_info->kobj);
10265                 kobject_put(&space_info->kobj);
10266         }
10267         return 0;
10268 }
10269
10270 /* link_block_group will queue up kobjects to add when we're reclaim-safe */
10271 void btrfs_add_raid_kobjects(struct btrfs_fs_info *fs_info)
10272 {
10273         struct btrfs_space_info *space_info;
10274         struct raid_kobject *rkobj;
10275         LIST_HEAD(list);
10276         int index;
10277         int ret = 0;
10278
10279         spin_lock(&fs_info->pending_raid_kobjs_lock);
10280         list_splice_init(&fs_info->pending_raid_kobjs, &list);
10281         spin_unlock(&fs_info->pending_raid_kobjs_lock);
10282
10283         list_for_each_entry(rkobj, &list, list) {
10284                 space_info = __find_space_info(fs_info, rkobj->flags);
10285                 index = btrfs_bg_flags_to_raid_index(rkobj->flags);
10286
10287                 ret = kobject_add(&rkobj->kobj, &space_info->kobj,
10288                                   "%s", get_raid_name(index));
10289                 if (ret) {
10290                         kobject_put(&rkobj->kobj);
10291                         break;
10292                 }
10293         }
10294         if (ret)
10295                 btrfs_warn(fs_info,
10296                            "failed to add kobject for block cache, ignoring");
10297 }
10298
10299 static void link_block_group(struct btrfs_block_group_cache *cache)
10300 {
10301         struct btrfs_space_info *space_info = cache->space_info;
10302         struct btrfs_fs_info *fs_info = cache->fs_info;
10303         int index = btrfs_bg_flags_to_raid_index(cache->flags);
10304         bool first = false;
10305
10306         down_write(&space_info->groups_sem);
10307         if (list_empty(&space_info->block_groups[index]))
10308                 first = true;
10309         list_add_tail(&cache->list, &space_info->block_groups[index]);
10310         up_write(&space_info->groups_sem);
10311
10312         if (first) {
10313                 struct raid_kobject *rkobj = kzalloc(sizeof(*rkobj), GFP_NOFS);
10314                 if (!rkobj) {
10315                         btrfs_warn(cache->fs_info,
10316                                 "couldn't alloc memory for raid level kobject");
10317                         return;
10318                 }
10319                 rkobj->flags = cache->flags;
10320                 kobject_init(&rkobj->kobj, &btrfs_raid_ktype);
10321
10322                 spin_lock(&fs_info->pending_raid_kobjs_lock);
10323                 list_add_tail(&rkobj->list, &fs_info->pending_raid_kobjs);
10324                 spin_unlock(&fs_info->pending_raid_kobjs_lock);
10325                 space_info->block_group_kobjs[index] = &rkobj->kobj;
10326         }
10327 }
10328
10329 static struct btrfs_block_group_cache *
10330 btrfs_create_block_group_cache(struct btrfs_fs_info *fs_info,
10331                                u64 start, u64 size)
10332 {
10333         struct btrfs_block_group_cache *cache;
10334
10335         cache = kzalloc(sizeof(*cache), GFP_NOFS);
10336         if (!cache)
10337                 return NULL;
10338
10339         cache->free_space_ctl = kzalloc(sizeof(*cache->free_space_ctl),
10340                                         GFP_NOFS);
10341         if (!cache->free_space_ctl) {
10342                 kfree(cache);
10343                 return NULL;
10344         }
10345
10346         cache->key.objectid = start;
10347         cache->key.offset = size;
10348         cache->key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
10349
10350         cache->fs_info = fs_info;
10351         cache->full_stripe_len = btrfs_full_stripe_len(fs_info, start);
10352         set_free_space_tree_thresholds(cache);
10353
10354         atomic_set(&cache->count, 1);
10355         spin_lock_init(&cache->lock);
10356         init_rwsem(&cache->data_rwsem);
10357         INIT_LIST_HEAD(&cache->list);
10358         INIT_LIST_HEAD(&cache->cluster_list);
10359         INIT_LIST_HEAD(&cache->bg_list);
10360         INIT_LIST_HEAD(&cache->ro_list);
10361         INIT_LIST_HEAD(&cache->dirty_list);
10362         INIT_LIST_HEAD(&cache->io_list);
10363         btrfs_init_free_space_ctl(cache);
10364         atomic_set(&cache->trimming, 0);
10365         mutex_init(&cache->free_space_lock);
10366         btrfs_init_full_stripe_locks_tree(&cache->full_stripe_locks_root);
10367
10368         return cache;
10369 }
10370
10371
10372 /*
10373  * Iterate all chunks and verify that each of them has the corresponding block
10374  * group
10375  */
10376 static int check_chunk_block_group_mappings(struct btrfs_fs_info *fs_info)
10377 {
10378         struct btrfs_mapping_tree *map_tree = &fs_info->mapping_tree;
10379         struct extent_map *em;
10380         struct btrfs_block_group_cache *bg;
10381         u64 start = 0;
10382         int ret = 0;
10383
10384         while (1) {
10385                 read_lock(&map_tree->map_tree.lock);
10386                 /*
10387                  * lookup_extent_mapping will return the first extent map
10388                  * intersecting the range, so setting @len to 1 is enough to
10389                  * get the first chunk.
10390                  */
10391                 em = lookup_extent_mapping(&map_tree->map_tree, start, 1);
10392                 read_unlock(&map_tree->map_tree.lock);
10393                 if (!em)
10394                         break;
10395
10396                 bg = btrfs_lookup_block_group(fs_info, em->start);
10397                 if (!bg) {
10398                         btrfs_err(fs_info,
10399         "chunk start=%llu len=%llu doesn't have corresponding block group",
10400                                      em->start, em->len);
10401                         ret = -EUCLEAN;
10402                         free_extent_map(em);
10403                         break;
10404                 }
10405                 if (bg->key.objectid != em->start ||
10406                     bg->key.offset != em->len ||
10407                     (bg->flags & BTRFS_BLOCK_GROUP_TYPE_MASK) !=
10408                     (em->map_lookup->type & BTRFS_BLOCK_GROUP_TYPE_MASK)) {
10409                         btrfs_err(fs_info,
10410 "chunk start=%llu len=%llu flags=0x%llx doesn't match block group start=%llu len=%llu flags=0x%llx",
10411                                 em->start, em->len,
10412                                 em->map_lookup->type & BTRFS_BLOCK_GROUP_TYPE_MASK,
10413                                 bg->key.objectid, bg->key.offset,
10414                                 bg->flags & BTRFS_BLOCK_GROUP_TYPE_MASK);
10415                         ret = -EUCLEAN;
10416                         free_extent_map(em);
10417                         btrfs_put_block_group(bg);
10418                         break;
10419                 }
10420                 start = em->start + em->len;
10421                 free_extent_map(em);
10422                 btrfs_put_block_group(bg);
10423         }
10424         return ret;
10425 }
10426
10427 int btrfs_read_block_groups(struct btrfs_fs_info *info)
10428 {
10429         struct btrfs_path *path;
10430         int ret;
10431         struct btrfs_block_group_cache *cache;
10432         struct btrfs_space_info *space_info;
10433         struct btrfs_key key;
10434         struct btrfs_key found_key;
10435         struct extent_buffer *leaf;
10436         int need_clear = 0;
10437         u64 cache_gen;
10438         u64 feature;
10439         int mixed;
10440
10441         feature = btrfs_super_incompat_flags(info->super_copy);
10442         mixed = !!(feature & BTRFS_FEATURE_INCOMPAT_MIXED_GROUPS);
10443
10444         key.objectid = 0;
10445         key.offset = 0;
10446         key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
10447         path = btrfs_alloc_path();
10448         if (!path)
10449                 return -ENOMEM;
10450         path->reada = READA_FORWARD;
10451
10452         cache_gen = btrfs_super_cache_generation(info->super_copy);
10453         if (btrfs_test_opt(info, SPACE_CACHE) &&
10454             btrfs_super_generation(info->super_copy) != cache_gen)
10455                 need_clear = 1;
10456         if (btrfs_test_opt(info, CLEAR_CACHE))
10457                 need_clear = 1;
10458
10459         while (1) {
10460                 ret = find_first_block_group(info, path, &key);
10461                 if (ret > 0)
10462                         break;
10463                 if (ret != 0)
10464                         goto error;
10465
10466                 leaf = path->nodes[0];
10467                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
10468
10469                 cache = btrfs_create_block_group_cache(info, found_key.objectid,
10470                                                        found_key.offset);
10471                 if (!cache) {
10472                         ret = -ENOMEM;
10473                         goto error;
10474                 }
10475
10476                 if (need_clear) {
10477                         /*
10478                          * When we mount with old space cache, we need to
10479                          * set BTRFS_DC_CLEAR and set dirty flag.
10480                          *
10481                          * a) Setting 'BTRFS_DC_CLEAR' makes sure that we
10482                          *    truncate the old free space cache inode and
10483                          *    setup a new one.
10484                          * b) Setting 'dirty flag' makes sure that we flush
10485                          *    the new space cache info onto disk.
10486                          */
10487                         if (btrfs_test_opt(info, SPACE_CACHE))
10488                                 cache->disk_cache_state = BTRFS_DC_CLEAR;
10489                 }
10490
10491                 read_extent_buffer(leaf, &cache->item,
10492                                    btrfs_item_ptr_offset(leaf, path->slots[0]),
10493                                    sizeof(cache->item));
10494                 cache->flags = btrfs_block_group_flags(&cache->item);
10495                 if (!mixed &&
10496                     ((cache->flags & BTRFS_BLOCK_GROUP_METADATA) &&
10497                     (cache->flags & BTRFS_BLOCK_GROUP_DATA))) {
10498                         btrfs_err(info,
10499 "bg %llu is a mixed block group but filesystem hasn't enabled mixed block groups",
10500                                   cache->key.objectid);
10501                         ret = -EINVAL;
10502                         goto error;
10503                 }
10504
10505                 key.objectid = found_key.objectid + found_key.offset;
10506                 btrfs_release_path(path);
10507
10508                 /*
10509                  * We need to exclude the super stripes now so that the space
10510                  * info has super bytes accounted for, otherwise we'll think
10511                  * we have more space than we actually do.
10512                  */
10513                 ret = exclude_super_stripes(cache);
10514                 if (ret) {
10515                         /*
10516                          * We may have excluded something, so call this just in
10517                          * case.
10518                          */
10519                         free_excluded_extents(cache);
10520                         btrfs_put_block_group(cache);
10521                         goto error;
10522                 }
10523
10524                 /*
10525                  * check for two cases, either we are full, and therefore
10526                  * don't need to bother with the caching work since we won't
10527                  * find any space, or we are empty, and we can just add all
10528                  * the space in and be done with it.  This saves us _a_lot_ of
10529                  * time, particularly in the full case.
10530                  */
10531                 if (found_key.offset == btrfs_block_group_used(&cache->item)) {
10532                         cache->last_byte_to_unpin = (u64)-1;
10533                         cache->cached = BTRFS_CACHE_FINISHED;
10534                         free_excluded_extents(cache);
10535                 } else if (btrfs_block_group_used(&cache->item) == 0) {
10536                         cache->last_byte_to_unpin = (u64)-1;
10537                         cache->cached = BTRFS_CACHE_FINISHED;
10538                         add_new_free_space(cache, found_key.objectid,
10539                                            found_key.objectid +
10540                                            found_key.offset);
10541                         free_excluded_extents(cache);
10542                 }
10543
10544                 ret = btrfs_add_block_group_cache(info, cache);
10545                 if (ret) {
10546                         btrfs_remove_free_space_cache(cache);
10547                         btrfs_put_block_group(cache);
10548                         goto error;
10549                 }
10550
10551                 trace_btrfs_add_block_group(info, cache, 0);
10552                 update_space_info(info, cache->flags, found_key.offset,
10553                                   btrfs_block_group_used(&cache->item),
10554                                   cache->bytes_super, &space_info);
10555
10556                 cache->space_info = space_info;
10557
10558                 link_block_group(cache);
10559
10560                 set_avail_alloc_bits(info, cache->flags);
10561                 if (btrfs_chunk_readonly(info, cache->key.objectid)) {
10562                         inc_block_group_ro(cache, 1);
10563                 } else if (btrfs_block_group_used(&cache->item) == 0) {
10564                         ASSERT(list_empty(&cache->bg_list));
10565                         btrfs_mark_bg_unused(cache);
10566                 }
10567         }
10568
10569         list_for_each_entry_rcu(space_info, &info->space_info, list) {
10570                 if (!(get_alloc_profile(info, space_info->flags) &
10571                       (BTRFS_BLOCK_GROUP_RAID10 |
10572                        BTRFS_BLOCK_GROUP_RAID1 |
10573                        BTRFS_BLOCK_GROUP_RAID5 |
10574                        BTRFS_BLOCK_GROUP_RAID6 |
10575                        BTRFS_BLOCK_GROUP_DUP)))
10576                         continue;
10577                 /*
10578                  * avoid allocating from un-mirrored block group if there are
10579                  * mirrored block groups.
10580                  */
10581                 list_for_each_entry(cache,
10582                                 &space_info->block_groups[BTRFS_RAID_RAID0],
10583                                 list)
10584                         inc_block_group_ro(cache, 1);
10585                 list_for_each_entry(cache,
10586                                 &space_info->block_groups[BTRFS_RAID_SINGLE],
10587                                 list)
10588                         inc_block_group_ro(cache, 1);
10589         }
10590
10591         btrfs_add_raid_kobjects(info);
10592         init_global_block_rsv(info);
10593         ret = check_chunk_block_group_mappings(info);
10594 error:
10595         btrfs_free_path(path);
10596         return ret;
10597 }
10598
10599 void btrfs_create_pending_block_groups(struct btrfs_trans_handle *trans)
10600 {
10601         struct btrfs_fs_info *fs_info = trans->fs_info;
10602         struct btrfs_block_group_cache *block_group;
10603         struct btrfs_root *extent_root = fs_info->extent_root;
10604         struct btrfs_block_group_item item;
10605         struct btrfs_key key;
10606         int ret = 0;
10607
10608         if (!trans->can_flush_pending_bgs)
10609                 return;
10610
10611         while (!list_empty(&trans->new_bgs)) {
10612                 block_group = list_first_entry(&trans->new_bgs,
10613                                                struct btrfs_block_group_cache,
10614                                                bg_list);
10615                 if (ret)
10616                         goto next;
10617
10618                 spin_lock(&block_group->lock);
10619                 memcpy(&item, &block_group->item, sizeof(item));
10620                 memcpy(&key, &block_group->key, sizeof(key));
10621                 spin_unlock(&block_group->lock);
10622
10623                 ret = btrfs_insert_item(trans, extent_root, &key, &item,
10624                                         sizeof(item));
10625                 if (ret)
10626                         btrfs_abort_transaction(trans, ret);
10627                 ret = btrfs_finish_chunk_alloc(trans, key.objectid, key.offset);
10628                 if (ret)
10629                         btrfs_abort_transaction(trans, ret);
10630                 add_block_group_free_space(trans, block_group);
10631                 /* already aborted the transaction if it failed. */
10632 next:
10633                 btrfs_delayed_refs_rsv_release(fs_info, 1);
10634                 list_del_init(&block_group->bg_list);
10635         }
10636         btrfs_trans_release_chunk_metadata(trans);
10637 }
10638
10639 int btrfs_make_block_group(struct btrfs_trans_handle *trans, u64 bytes_used,
10640                            u64 type, u64 chunk_offset, u64 size)
10641 {
10642         struct btrfs_fs_info *fs_info = trans->fs_info;
10643         struct btrfs_block_group_cache *cache;
10644         int ret;
10645
10646         btrfs_set_log_full_commit(fs_info, trans);
10647
10648         cache = btrfs_create_block_group_cache(fs_info, chunk_offset, size);
10649         if (!cache)
10650                 return -ENOMEM;
10651
10652         btrfs_set_block_group_used(&cache->item, bytes_used);
10653         btrfs_set_block_group_chunk_objectid(&cache->item,
10654                                              BTRFS_FIRST_CHUNK_TREE_OBJECTID);
10655         btrfs_set_block_group_flags(&cache->item, type);
10656
10657         cache->flags = type;
10658         cache->last_byte_to_unpin = (u64)-1;
10659         cache->cached = BTRFS_CACHE_FINISHED;
10660         cache->needs_free_space = 1;
10661         ret = exclude_super_stripes(cache);
10662         if (ret) {
10663                 /*
10664                  * We may have excluded something, so call this just in
10665                  * case.
10666                  */
10667                 free_excluded_extents(cache);
10668                 btrfs_put_block_group(cache);
10669                 return ret;
10670         }
10671
10672         add_new_free_space(cache, chunk_offset, chunk_offset + size);
10673
10674         free_excluded_extents(cache);
10675
10676 #ifdef CONFIG_BTRFS_DEBUG
10677         if (btrfs_should_fragment_free_space(cache)) {
10678                 u64 new_bytes_used = size - bytes_used;
10679
10680                 bytes_used += new_bytes_used >> 1;
10681                 fragment_free_space(cache);
10682         }
10683 #endif
10684         /*
10685          * Ensure the corresponding space_info object is created and
10686          * assigned to our block group. We want our bg to be added to the rbtree
10687          * with its ->space_info set.
10688          */
10689         cache->space_info = __find_space_info(fs_info, cache->flags);
10690         ASSERT(cache->space_info);
10691
10692         ret = btrfs_add_block_group_cache(fs_info, cache);
10693         if (ret) {
10694                 btrfs_remove_free_space_cache(cache);
10695                 btrfs_put_block_group(cache);
10696                 return ret;
10697         }
10698
10699         /*
10700          * Now that our block group has its ->space_info set and is inserted in
10701          * the rbtree, update the space info's counters.
10702          */
10703         trace_btrfs_add_block_group(fs_info, cache, 1);
10704         update_space_info(fs_info, cache->flags, size, bytes_used,
10705                                 cache->bytes_super, &cache->space_info);
10706         update_global_block_rsv(fs_info);
10707
10708         link_block_group(cache);
10709
10710         list_add_tail(&cache->bg_list, &trans->new_bgs);
10711         trans->delayed_ref_updates++;
10712         btrfs_update_delayed_refs_rsv(trans);
10713
10714         set_avail_alloc_bits(fs_info, type);
10715         return 0;
10716 }
10717
10718 static void clear_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
10719 {
10720         u64 extra_flags = chunk_to_extended(flags) &
10721                                 BTRFS_EXTENDED_PROFILE_MASK;
10722
10723         write_seqlock(&fs_info->profiles_lock);
10724         if (flags & BTRFS_BLOCK_GROUP_DATA)
10725                 fs_info->avail_data_alloc_bits &= ~extra_flags;
10726         if (flags & BTRFS_BLOCK_GROUP_METADATA)
10727                 fs_info->avail_metadata_alloc_bits &= ~extra_flags;
10728         if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
10729                 fs_info->avail_system_alloc_bits &= ~extra_flags;
10730         write_sequnlock(&fs_info->profiles_lock);
10731 }
10732
10733 int btrfs_remove_block_group(struct btrfs_trans_handle *trans,
10734                              u64 group_start, struct extent_map *em)
10735 {
10736         struct btrfs_fs_info *fs_info = trans->fs_info;
10737         struct btrfs_root *root = fs_info->extent_root;
10738         struct btrfs_path *path;
10739         struct btrfs_block_group_cache *block_group;
10740         struct btrfs_free_cluster *cluster;
10741         struct btrfs_root *tree_root = fs_info->tree_root;
10742         struct btrfs_key key;
10743         struct inode *inode;
10744         struct kobject *kobj = NULL;
10745         int ret;
10746         int index;
10747         int factor;
10748         struct btrfs_caching_control *caching_ctl = NULL;
10749         bool remove_em;
10750         bool remove_rsv = false;
10751
10752         block_group = btrfs_lookup_block_group(fs_info, group_start);
10753         BUG_ON(!block_group);
10754         BUG_ON(!block_group->ro);
10755
10756         trace_btrfs_remove_block_group(block_group);
10757         /*
10758          * Free the reserved super bytes from this block group before
10759          * remove it.
10760          */
10761         free_excluded_extents(block_group);
10762         btrfs_free_ref_tree_range(fs_info, block_group->key.objectid,
10763                                   block_group->key.offset);
10764
10765         memcpy(&key, &block_group->key, sizeof(key));
10766         index = btrfs_bg_flags_to_raid_index(block_group->flags);
10767         factor = btrfs_bg_type_to_factor(block_group->flags);
10768
10769         /* make sure this block group isn't part of an allocation cluster */
10770         cluster = &fs_info->data_alloc_cluster;
10771         spin_lock(&cluster->refill_lock);
10772         btrfs_return_cluster_to_free_space(block_group, cluster);
10773         spin_unlock(&cluster->refill_lock);
10774
10775         /*
10776          * make sure this block group isn't part of a metadata
10777          * allocation cluster
10778          */
10779         cluster = &fs_info->meta_alloc_cluster;
10780         spin_lock(&cluster->refill_lock);
10781         btrfs_return_cluster_to_free_space(block_group, cluster);
10782         spin_unlock(&cluster->refill_lock);
10783
10784         path = btrfs_alloc_path();
10785         if (!path) {
10786                 ret = -ENOMEM;
10787                 goto out;
10788         }
10789
10790         /*
10791          * get the inode first so any iput calls done for the io_list
10792          * aren't the final iput (no unlinks allowed now)
10793          */
10794         inode = lookup_free_space_inode(fs_info, block_group, path);
10795
10796         mutex_lock(&trans->transaction->cache_write_mutex);
10797         /*
10798          * Make sure our free space cache IO is done before removing the
10799          * free space inode
10800          */
10801         spin_lock(&trans->transaction->dirty_bgs_lock);
10802         if (!list_empty(&block_group->io_list)) {
10803                 list_del_init(&block_group->io_list);
10804
10805                 WARN_ON(!IS_ERR(inode) && inode != block_group->io_ctl.inode);
10806
10807                 spin_unlock(&trans->transaction->dirty_bgs_lock);
10808                 btrfs_wait_cache_io(trans, block_group, path);
10809                 btrfs_put_block_group(block_group);
10810                 spin_lock(&trans->transaction->dirty_bgs_lock);
10811         }
10812
10813         if (!list_empty(&block_group->dirty_list)) {
10814                 list_del_init(&block_group->dirty_list);
10815                 remove_rsv = true;
10816                 btrfs_put_block_group(block_group);
10817         }
10818         spin_unlock(&trans->transaction->dirty_bgs_lock);
10819         mutex_unlock(&trans->transaction->cache_write_mutex);
10820
10821         if (!IS_ERR(inode)) {
10822                 ret = btrfs_orphan_add(trans, BTRFS_I(inode));
10823                 if (ret) {
10824                         btrfs_add_delayed_iput(inode);
10825                         goto out;
10826                 }
10827                 clear_nlink(inode);
10828                 /* One for the block groups ref */
10829                 spin_lock(&block_group->lock);
10830                 if (block_group->iref) {
10831                         block_group->iref = 0;
10832                         block_group->inode = NULL;
10833                         spin_unlock(&block_group->lock);
10834                         iput(inode);
10835                 } else {
10836                         spin_unlock(&block_group->lock);
10837                 }
10838                 /* One for our lookup ref */
10839                 btrfs_add_delayed_iput(inode);
10840         }
10841
10842         key.objectid = BTRFS_FREE_SPACE_OBJECTID;
10843         key.offset = block_group->key.objectid;
10844         key.type = 0;
10845
10846         ret = btrfs_search_slot(trans, tree_root, &key, path, -1, 1);
10847         if (ret < 0)
10848                 goto out;
10849         if (ret > 0)
10850                 btrfs_release_path(path);
10851         if (ret == 0) {
10852                 ret = btrfs_del_item(trans, tree_root, path);
10853                 if (ret)
10854                         goto out;
10855                 btrfs_release_path(path);
10856         }
10857
10858         spin_lock(&fs_info->block_group_cache_lock);
10859         rb_erase(&block_group->cache_node,
10860                  &fs_info->block_group_cache_tree);
10861         RB_CLEAR_NODE(&block_group->cache_node);
10862
10863         if (fs_info->first_logical_byte == block_group->key.objectid)
10864                 fs_info->first_logical_byte = (u64)-1;
10865         spin_unlock(&fs_info->block_group_cache_lock);
10866
10867         down_write(&block_group->space_info->groups_sem);
10868         /*
10869          * we must use list_del_init so people can check to see if they
10870          * are still on the list after taking the semaphore
10871          */
10872         list_del_init(&block_group->list);
10873         if (list_empty(&block_group->space_info->block_groups[index])) {
10874                 kobj = block_group->space_info->block_group_kobjs[index];
10875                 block_group->space_info->block_group_kobjs[index] = NULL;
10876                 clear_avail_alloc_bits(fs_info, block_group->flags);
10877         }
10878         up_write(&block_group->space_info->groups_sem);
10879         if (kobj) {
10880                 kobject_del(kobj);
10881                 kobject_put(kobj);
10882         }
10883
10884         if (block_group->has_caching_ctl)
10885                 caching_ctl = get_caching_control(block_group);
10886         if (block_group->cached == BTRFS_CACHE_STARTED)
10887                 wait_block_group_cache_done(block_group);
10888         if (block_group->has_caching_ctl) {
10889                 down_write(&fs_info->commit_root_sem);
10890                 if (!caching_ctl) {
10891                         struct btrfs_caching_control *ctl;
10892
10893                         list_for_each_entry(ctl,
10894                                     &fs_info->caching_block_groups, list)
10895                                 if (ctl->block_group == block_group) {
10896                                         caching_ctl = ctl;
10897                                         refcount_inc(&caching_ctl->count);
10898                                         break;
10899                                 }
10900                 }
10901                 if (caching_ctl)
10902                         list_del_init(&caching_ctl->list);
10903                 up_write(&fs_info->commit_root_sem);
10904                 if (caching_ctl) {
10905                         /* Once for the caching bgs list and once for us. */
10906                         put_caching_control(caching_ctl);
10907                         put_caching_control(caching_ctl);
10908                 }
10909         }
10910
10911         spin_lock(&trans->transaction->dirty_bgs_lock);
10912         WARN_ON(!list_empty(&block_group->dirty_list));
10913         WARN_ON(!list_empty(&block_group->io_list));
10914         spin_unlock(&trans->transaction->dirty_bgs_lock);
10915
10916         btrfs_remove_free_space_cache(block_group);
10917
10918         spin_lock(&block_group->space_info->lock);
10919         list_del_init(&block_group->ro_list);
10920
10921         if (btrfs_test_opt(fs_info, ENOSPC_DEBUG)) {
10922                 WARN_ON(block_group->space_info->total_bytes
10923                         < block_group->key.offset);
10924                 WARN_ON(block_group->space_info->bytes_readonly
10925                         < block_group->key.offset);
10926                 WARN_ON(block_group->space_info->disk_total
10927                         < block_group->key.offset * factor);
10928         }
10929         block_group->space_info->total_bytes -= block_group->key.offset;
10930         block_group->space_info->bytes_readonly -= block_group->key.offset;
10931         block_group->space_info->disk_total -= block_group->key.offset * factor;
10932
10933         spin_unlock(&block_group->space_info->lock);
10934
10935         memcpy(&key, &block_group->key, sizeof(key));
10936
10937         mutex_lock(&fs_info->chunk_mutex);
10938         spin_lock(&block_group->lock);
10939         block_group->removed = 1;
10940         /*
10941          * At this point trimming can't start on this block group, because we
10942          * removed the block group from the tree fs_info->block_group_cache_tree
10943          * so no one can't find it anymore and even if someone already got this
10944          * block group before we removed it from the rbtree, they have already
10945          * incremented block_group->trimming - if they didn't, they won't find
10946          * any free space entries because we already removed them all when we
10947          * called btrfs_remove_free_space_cache().
10948          *
10949          * And we must not remove the extent map from the fs_info->mapping_tree
10950          * to prevent the same logical address range and physical device space
10951          * ranges from being reused for a new block group. This is because our
10952          * fs trim operation (btrfs_trim_fs() / btrfs_ioctl_fitrim()) is
10953          * completely transactionless, so while it is trimming a range the
10954          * currently running transaction might finish and a new one start,
10955          * allowing for new block groups to be created that can reuse the same
10956          * physical device locations unless we take this special care.
10957          *
10958          * There may also be an implicit trim operation if the file system
10959          * is mounted with -odiscard. The same protections must remain
10960          * in place until the extents have been discarded completely when
10961          * the transaction commit has completed.
10962          */
10963         remove_em = (atomic_read(&block_group->trimming) == 0);
10964         spin_unlock(&block_group->lock);
10965
10966         if (remove_em) {
10967                 struct extent_map_tree *em_tree;
10968
10969                 em_tree = &fs_info->mapping_tree.map_tree;
10970                 write_lock(&em_tree->lock);
10971                 remove_extent_mapping(em_tree, em);
10972                 write_unlock(&em_tree->lock);
10973                 /* once for the tree */
10974                 free_extent_map(em);
10975         }
10976
10977         mutex_unlock(&fs_info->chunk_mutex);
10978
10979         ret = remove_block_group_free_space(trans, block_group);
10980         if (ret)
10981                 goto out;
10982
10983         btrfs_put_block_group(block_group);
10984         btrfs_put_block_group(block_group);
10985
10986         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
10987         if (ret > 0)
10988                 ret = -EIO;
10989         if (ret < 0)
10990                 goto out;
10991
10992         ret = btrfs_del_item(trans, root, path);
10993 out:
10994         if (remove_rsv)
10995                 btrfs_delayed_refs_rsv_release(fs_info, 1);
10996         btrfs_free_path(path);
10997         return ret;
10998 }
10999
11000 struct btrfs_trans_handle *
11001 btrfs_start_trans_remove_block_group(struct btrfs_fs_info *fs_info,
11002                                      const u64 chunk_offset)
11003 {
11004         struct extent_map_tree *em_tree = &fs_info->mapping_tree.map_tree;
11005         struct extent_map *em;
11006         struct map_lookup *map;
11007         unsigned int num_items;
11008
11009         read_lock(&em_tree->lock);
11010         em = lookup_extent_mapping(em_tree, chunk_offset, 1);
11011         read_unlock(&em_tree->lock);
11012         ASSERT(em && em->start == chunk_offset);
11013
11014         /*
11015          * We need to reserve 3 + N units from the metadata space info in order
11016          * to remove a block group (done at btrfs_remove_chunk() and at
11017          * btrfs_remove_block_group()), which are used for:
11018          *
11019          * 1 unit for adding the free space inode's orphan (located in the tree
11020          * of tree roots).
11021          * 1 unit for deleting the block group item (located in the extent
11022          * tree).
11023          * 1 unit for deleting the free space item (located in tree of tree
11024          * roots).
11025          * N units for deleting N device extent items corresponding to each
11026          * stripe (located in the device tree).
11027          *
11028          * In order to remove a block group we also need to reserve units in the
11029          * system space info in order to update the chunk tree (update one or
11030          * more device items and remove one chunk item), but this is done at
11031          * btrfs_remove_chunk() through a call to check_system_chunk().
11032          */
11033         map = em->map_lookup;
11034         num_items = 3 + map->num_stripes;
11035         free_extent_map(em);
11036
11037         return btrfs_start_transaction_fallback_global_rsv(fs_info->extent_root,
11038                                                            num_items, 1);
11039 }
11040
11041 /*
11042  * Process the unused_bgs list and remove any that don't have any allocated
11043  * space inside of them.
11044  */
11045 void btrfs_delete_unused_bgs(struct btrfs_fs_info *fs_info)
11046 {
11047         struct btrfs_block_group_cache *block_group;
11048         struct btrfs_space_info *space_info;
11049         struct btrfs_trans_handle *trans;
11050         int ret = 0;
11051
11052         if (!test_bit(BTRFS_FS_OPEN, &fs_info->flags))
11053                 return;
11054
11055         spin_lock(&fs_info->unused_bgs_lock);
11056         while (!list_empty(&fs_info->unused_bgs)) {
11057                 u64 start, end;
11058                 int trimming;
11059
11060                 block_group = list_first_entry(&fs_info->unused_bgs,
11061                                                struct btrfs_block_group_cache,
11062                                                bg_list);
11063                 list_del_init(&block_group->bg_list);
11064
11065                 space_info = block_group->space_info;
11066
11067                 if (ret || btrfs_mixed_space_info(space_info)) {
11068                         btrfs_put_block_group(block_group);
11069                         continue;
11070                 }
11071                 spin_unlock(&fs_info->unused_bgs_lock);
11072
11073                 mutex_lock(&fs_info->delete_unused_bgs_mutex);
11074
11075                 /* Don't want to race with allocators so take the groups_sem */
11076                 down_write(&space_info->groups_sem);
11077                 spin_lock(&block_group->lock);
11078                 if (block_group->reserved || block_group->pinned ||
11079                     btrfs_block_group_used(&block_group->item) ||
11080                     block_group->ro ||
11081                     list_is_singular(&block_group->list)) {
11082                         /*
11083                          * We want to bail if we made new allocations or have
11084                          * outstanding allocations in this block group.  We do
11085                          * the ro check in case balance is currently acting on
11086                          * this block group.
11087                          */
11088                         trace_btrfs_skip_unused_block_group(block_group);
11089                         spin_unlock(&block_group->lock);
11090                         up_write(&space_info->groups_sem);
11091                         goto next;
11092                 }
11093                 spin_unlock(&block_group->lock);
11094
11095                 /* We don't want to force the issue, only flip if it's ok. */
11096                 ret = inc_block_group_ro(block_group, 0);
11097                 up_write(&space_info->groups_sem);
11098                 if (ret < 0) {
11099                         ret = 0;
11100                         goto next;
11101                 }
11102
11103                 /*
11104                  * Want to do this before we do anything else so we can recover
11105                  * properly if we fail to join the transaction.
11106                  */
11107                 trans = btrfs_start_trans_remove_block_group(fs_info,
11108                                                      block_group->key.objectid);
11109                 if (IS_ERR(trans)) {
11110                         btrfs_dec_block_group_ro(block_group);
11111                         ret = PTR_ERR(trans);
11112                         goto next;
11113                 }
11114
11115                 /*
11116                  * We could have pending pinned extents for this block group,
11117                  * just delete them, we don't care about them anymore.
11118                  */
11119                 start = block_group->key.objectid;
11120                 end = start + block_group->key.offset - 1;
11121                 /*
11122                  * Hold the unused_bg_unpin_mutex lock to avoid racing with
11123                  * btrfs_finish_extent_commit(). If we are at transaction N,
11124                  * another task might be running finish_extent_commit() for the
11125                  * previous transaction N - 1, and have seen a range belonging
11126                  * to the block group in freed_extents[] before we were able to
11127                  * clear the whole block group range from freed_extents[]. This
11128                  * means that task can lookup for the block group after we
11129                  * unpinned it from freed_extents[] and removed it, leading to
11130                  * a BUG_ON() at btrfs_unpin_extent_range().
11131                  */
11132                 mutex_lock(&fs_info->unused_bg_unpin_mutex);
11133                 ret = clear_extent_bits(&fs_info->freed_extents[0], start, end,
11134                                   EXTENT_DIRTY);
11135                 if (ret) {
11136                         mutex_unlock(&fs_info->unused_bg_unpin_mutex);
11137                         btrfs_dec_block_group_ro(block_group);
11138                         goto end_trans;
11139                 }
11140                 ret = clear_extent_bits(&fs_info->freed_extents[1], start, end,
11141                                   EXTENT_DIRTY);
11142                 if (ret) {
11143                         mutex_unlock(&fs_info->unused_bg_unpin_mutex);
11144                         btrfs_dec_block_group_ro(block_group);
11145                         goto end_trans;
11146                 }
11147                 mutex_unlock(&fs_info->unused_bg_unpin_mutex);
11148
11149                 /* Reset pinned so btrfs_put_block_group doesn't complain */
11150                 spin_lock(&space_info->lock);
11151                 spin_lock(&block_group->lock);
11152
11153                 update_bytes_pinned(space_info, -block_group->pinned);
11154                 space_info->bytes_readonly += block_group->pinned;
11155                 percpu_counter_add_batch(&space_info->total_bytes_pinned,
11156                                    -block_group->pinned,
11157                                    BTRFS_TOTAL_BYTES_PINNED_BATCH);
11158                 block_group->pinned = 0;
11159
11160                 spin_unlock(&block_group->lock);
11161                 spin_unlock(&space_info->lock);
11162
11163                 /* DISCARD can flip during remount */
11164                 trimming = btrfs_test_opt(fs_info, DISCARD);
11165
11166                 /* Implicit trim during transaction commit. */
11167                 if (trimming)
11168                         btrfs_get_block_group_trimming(block_group);
11169
11170                 /*
11171                  * Btrfs_remove_chunk will abort the transaction if things go
11172                  * horribly wrong.
11173                  */
11174                 ret = btrfs_remove_chunk(trans, block_group->key.objectid);
11175
11176                 if (ret) {
11177                         if (trimming)
11178                                 btrfs_put_block_group_trimming(block_group);
11179                         goto end_trans;
11180                 }
11181
11182                 /*
11183                  * If we're not mounted with -odiscard, we can just forget
11184                  * about this block group. Otherwise we'll need to wait
11185                  * until transaction commit to do the actual discard.
11186                  */
11187                 if (trimming) {
11188                         spin_lock(&fs_info->unused_bgs_lock);
11189                         /*
11190                          * A concurrent scrub might have added us to the list
11191                          * fs_info->unused_bgs, so use a list_move operation
11192                          * to add the block group to the deleted_bgs list.
11193                          */
11194                         list_move(&block_group->bg_list,
11195                                   &trans->transaction->deleted_bgs);
11196                         spin_unlock(&fs_info->unused_bgs_lock);
11197                         btrfs_get_block_group(block_group);
11198                 }
11199 end_trans:
11200                 btrfs_end_transaction(trans);
11201 next:
11202                 mutex_unlock(&fs_info->delete_unused_bgs_mutex);
11203                 btrfs_put_block_group(block_group);
11204                 spin_lock(&fs_info->unused_bgs_lock);
11205         }
11206         spin_unlock(&fs_info->unused_bgs_lock);
11207 }
11208
11209 int btrfs_init_space_info(struct btrfs_fs_info *fs_info)
11210 {
11211         struct btrfs_super_block *disk_super;
11212         u64 features;
11213         u64 flags;
11214         int mixed = 0;
11215         int ret;
11216
11217         disk_super = fs_info->super_copy;
11218         if (!btrfs_super_root(disk_super))
11219                 return -EINVAL;
11220
11221         features = btrfs_super_incompat_flags(disk_super);
11222         if (features & BTRFS_FEATURE_INCOMPAT_MIXED_GROUPS)
11223                 mixed = 1;
11224
11225         flags = BTRFS_BLOCK_GROUP_SYSTEM;
11226         ret = create_space_info(fs_info, flags);
11227         if (ret)
11228                 goto out;
11229
11230         if (mixed) {
11231                 flags = BTRFS_BLOCK_GROUP_METADATA | BTRFS_BLOCK_GROUP_DATA;
11232                 ret = create_space_info(fs_info, flags);
11233         } else {
11234                 flags = BTRFS_BLOCK_GROUP_METADATA;
11235                 ret = create_space_info(fs_info, flags);
11236                 if (ret)
11237                         goto out;
11238
11239                 flags = BTRFS_BLOCK_GROUP_DATA;
11240                 ret = create_space_info(fs_info, flags);
11241         }
11242 out:
11243         return ret;
11244 }
11245
11246 int btrfs_error_unpin_extent_range(struct btrfs_fs_info *fs_info,
11247                                    u64 start, u64 end)
11248 {
11249         return unpin_extent_range(fs_info, start, end, false);
11250 }
11251
11252 /*
11253  * It used to be that old block groups would be left around forever.
11254  * Iterating over them would be enough to trim unused space.  Since we
11255  * now automatically remove them, we also need to iterate over unallocated
11256  * space.
11257  *
11258  * We don't want a transaction for this since the discard may take a
11259  * substantial amount of time.  We don't require that a transaction be
11260  * running, but we do need to take a running transaction into account
11261  * to ensure that we're not discarding chunks that were released or
11262  * allocated in the current transaction.
11263  *
11264  * Holding the chunks lock will prevent other threads from allocating
11265  * or releasing chunks, but it won't prevent a running transaction
11266  * from committing and releasing the memory that the pending chunks
11267  * list head uses.  For that, we need to take a reference to the
11268  * transaction and hold the commit root sem.  We only need to hold
11269  * it while performing the free space search since we have already
11270  * held back allocations.
11271  */
11272 static int btrfs_trim_free_extents(struct btrfs_device *device,
11273                                    struct fstrim_range *range, u64 *trimmed)
11274 {
11275         u64 start, len = 0, end = 0;
11276         int ret;
11277
11278         start = max_t(u64, range->start, SZ_1M);
11279         *trimmed = 0;
11280
11281         /* Discard not supported = nothing to do. */
11282         if (!blk_queue_discard(bdev_get_queue(device->bdev)))
11283                 return 0;
11284
11285         /* Not writable = nothing to do. */
11286         if (!test_bit(BTRFS_DEV_STATE_WRITEABLE, &device->dev_state))
11287                 return 0;
11288
11289         /* No free space = nothing to do. */
11290         if (device->total_bytes <= device->bytes_used)
11291                 return 0;
11292
11293         ret = 0;
11294
11295         while (1) {
11296                 struct btrfs_fs_info *fs_info = device->fs_info;
11297                 u64 bytes;
11298
11299                 ret = mutex_lock_interruptible(&fs_info->chunk_mutex);
11300                 if (ret)
11301                         break;
11302
11303                 find_first_clear_extent_bit(&device->alloc_state, start,
11304                                             &start, &end,
11305                                             CHUNK_TRIMMED | CHUNK_ALLOCATED);
11306                 /*
11307                  * If find_first_clear_extent_bit find a range that spans the
11308                  * end of the device it will set end to -1, in this case it's up
11309                  * to the caller to trim the value to the size of the device.
11310                  */
11311                 end = min(end, device->total_bytes - 1);
11312                 len = end - start + 1;
11313
11314                 /* We didn't find any extents */
11315                 if (!len) {
11316                         mutex_unlock(&fs_info->chunk_mutex);
11317                         ret = 0;
11318                         break;
11319                 }
11320
11321                 /* Keep going until we satisfy minlen or reach end of space */
11322                 if (len < range->minlen) {
11323                         mutex_unlock(&fs_info->chunk_mutex);
11324                         start += len;
11325                         continue;
11326                 }
11327
11328                 /* If we are out of the passed range break */
11329                 if (start > range->start + range->len - 1) {
11330                         mutex_unlock(&fs_info->chunk_mutex);
11331                         break;
11332                 }
11333
11334                 start = max(range->start, start);
11335                 len = min(range->len, len);
11336
11337                 ret = btrfs_issue_discard(device->bdev, start, len,
11338                                           &bytes);
11339                 if (!ret)
11340                         set_extent_bits(&device->alloc_state, start,
11341                                         start + bytes - 1,
11342                                         CHUNK_TRIMMED);
11343                 mutex_unlock(&fs_info->chunk_mutex);
11344
11345                 if (ret)
11346                         break;
11347
11348                 start += len;
11349                 *trimmed += bytes;
11350
11351                 /* We've trimmed enough */
11352                 if (*trimmed >= range->len)
11353                         break;
11354
11355                 if (fatal_signal_pending(current)) {
11356                         ret = -ERESTARTSYS;
11357                         break;
11358                 }
11359
11360                 cond_resched();
11361         }
11362
11363         return ret;
11364 }
11365
11366 /*
11367  * Trim the whole filesystem by:
11368  * 1) trimming the free space in each block group
11369  * 2) trimming the unallocated space on each device
11370  *
11371  * This will also continue trimming even if a block group or device encounters
11372  * an error.  The return value will be the last error, or 0 if nothing bad
11373  * happens.
11374  */
11375 int btrfs_trim_fs(struct btrfs_fs_info *fs_info, struct fstrim_range *range)
11376 {
11377         struct btrfs_block_group_cache *cache = NULL;
11378         struct btrfs_device *device;
11379         struct list_head *devices;
11380         u64 group_trimmed;
11381         u64 start;
11382         u64 end;
11383         u64 trimmed = 0;
11384         u64 bg_failed = 0;
11385         u64 dev_failed = 0;
11386         int bg_ret = 0;
11387         int dev_ret = 0;
11388         int ret = 0;
11389
11390         cache = btrfs_lookup_first_block_group(fs_info, range->start);
11391         for (; cache; cache = next_block_group(fs_info, cache)) {
11392                 if (cache->key.objectid >= (range->start + range->len)) {
11393                         btrfs_put_block_group(cache);
11394                         break;
11395                 }
11396
11397                 start = max(range->start, cache->key.objectid);
11398                 end = min(range->start + range->len,
11399                                 cache->key.objectid + cache->key.offset);
11400
11401                 if (end - start >= range->minlen) {
11402                         if (!block_group_cache_done(cache)) {
11403                                 ret = cache_block_group(cache, 0);
11404                                 if (ret) {
11405                                         bg_failed++;
11406                                         bg_ret = ret;
11407                                         continue;
11408                                 }
11409                                 ret = wait_block_group_cache_done(cache);
11410                                 if (ret) {
11411                                         bg_failed++;
11412                                         bg_ret = ret;
11413                                         continue;
11414                                 }
11415                         }
11416                         ret = btrfs_trim_block_group(cache,
11417                                                      &group_trimmed,
11418                                                      start,
11419                                                      end,
11420                                                      range->minlen);
11421
11422                         trimmed += group_trimmed;
11423                         if (ret) {
11424                                 bg_failed++;
11425                                 bg_ret = ret;
11426                                 continue;
11427                         }
11428                 }
11429         }
11430
11431         if (bg_failed)
11432                 btrfs_warn(fs_info,
11433                         "failed to trim %llu block group(s), last error %d",
11434                         bg_failed, bg_ret);
11435         mutex_lock(&fs_info->fs_devices->device_list_mutex);
11436         devices = &fs_info->fs_devices->devices;
11437         list_for_each_entry(device, devices, dev_list) {
11438                 ret = btrfs_trim_free_extents(device, range, &group_trimmed);
11439                 if (ret) {
11440                         dev_failed++;
11441                         dev_ret = ret;
11442                         break;
11443                 }
11444
11445                 trimmed += group_trimmed;
11446         }
11447         mutex_unlock(&fs_info->fs_devices->device_list_mutex);
11448
11449         if (dev_failed)
11450                 btrfs_warn(fs_info,
11451                         "failed to trim %llu device(s), last error %d",
11452                         dev_failed, dev_ret);
11453         range->len = trimmed;
11454         if (bg_ret)
11455                 return bg_ret;
11456         return dev_ret;
11457 }
11458
11459 /*
11460  * btrfs_{start,end}_write_no_snapshotting() are similar to
11461  * mnt_{want,drop}_write(), they are used to prevent some tasks from writing
11462  * data into the page cache through nocow before the subvolume is snapshoted,
11463  * but flush the data into disk after the snapshot creation, or to prevent
11464  * operations while snapshotting is ongoing and that cause the snapshot to be
11465  * inconsistent (writes followed by expanding truncates for example).
11466  */
11467 void btrfs_end_write_no_snapshotting(struct btrfs_root *root)
11468 {
11469         percpu_counter_dec(&root->subv_writers->counter);
11470         cond_wake_up(&root->subv_writers->wait);
11471 }
11472
11473 int btrfs_start_write_no_snapshotting(struct btrfs_root *root)
11474 {
11475         if (atomic_read(&root->will_be_snapshotted))
11476                 return 0;
11477
11478         percpu_counter_inc(&root->subv_writers->counter);
11479         /*
11480          * Make sure counter is updated before we check for snapshot creation.
11481          */
11482         smp_mb();
11483         if (atomic_read(&root->will_be_snapshotted)) {
11484                 btrfs_end_write_no_snapshotting(root);
11485                 return 0;
11486         }
11487         return 1;
11488 }
11489
11490 void btrfs_wait_for_snapshot_creation(struct btrfs_root *root)
11491 {
11492         while (true) {
11493                 int ret;
11494
11495                 ret = btrfs_start_write_no_snapshotting(root);
11496                 if (ret)
11497                         break;
11498                 wait_var_event(&root->will_be_snapshotted,
11499                                !atomic_read(&root->will_be_snapshotted));
11500         }
11501 }
11502
11503 void btrfs_mark_bg_unused(struct btrfs_block_group_cache *bg)
11504 {
11505         struct btrfs_fs_info *fs_info = bg->fs_info;
11506
11507         spin_lock(&fs_info->unused_bgs_lock);
11508         if (list_empty(&bg->bg_list)) {
11509                 btrfs_get_block_group(bg);
11510                 trace_btrfs_add_unused_block_group(bg);
11511                 list_add_tail(&bg->bg_list, &fs_info->unused_bgs);
11512         }
11513         spin_unlock(&fs_info->unused_bgs_lock);
11514 }