fs/btrfs/extent-tree.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Copyright (C) 2007 Oracle.  All rights reserved.
   4  */
   5
   6 #include <linux/sched.h>
   7 #include <linux/sched/signal.h>
   8 #include <linux/pagemap.h>
   9 #include <linux/writeback.h>
  10 #include <linux/blkdev.h>
  11 #include <linux/sort.h>
  12 #include <linux/rcupdate.h>
  13 #include <linux/kthread.h>
  14 #include <linux/slab.h>
  15 #include <linux/ratelimit.h>
  16 #include <linux/percpu_counter.h>
  17 #include <linux/lockdep.h>
  18 #include <linux/crc32c.h>
  19 #include "tree-log.h"
  20 #include "disk-io.h"
  21 #include "print-tree.h"
  22 #include "volumes.h"
  23 #include "raid56.h"
  24 #include "locking.h"
  25 #include "free-space-cache.h"
  26 #include "free-space-tree.h"
  27 #include "math.h"
  28 #include "sysfs.h"
  29 #include "qgroup.h"
  30 #include "ref-verify.h"
  31
  32 #undef SCRAMBLE_DELAYED_REFS
  33
  34 /*
  35  * control flags for do_chunk_alloc's force field
  36  * CHUNK_ALLOC_NO_FORCE means to only allocate a chunk
  37  * if we really need one.
  38  *
  39  * CHUNK_ALLOC_LIMITED means to only try and allocate one
  40  * if we have very few chunks already allocated.  This is
  41  * used as part of the clustering code to help make sure
  42  * we have a good pool of storage to cluster in, without
  43  * filling the FS with empty chunks
  44  *
  45  * CHUNK_ALLOC_FORCE means it must try to allocate one
  46  *
  47  */
  48 enum {
  49         CHUNK_ALLOC_NO_FORCE = 0,
  50         CHUNK_ALLOC_LIMITED = 1,
  51         CHUNK_ALLOC_FORCE = 2,
  52 };
  53
  54 /*
  55  * Declare a helper function to detect underflow of various space info members
  56  */
  57 #define DECLARE_SPACE_INFO_UPDATE(name)                                 \
  58 static inline void update_##name(struct btrfs_space_info *sinfo,        \
  59                                  s64 bytes)                             \
  60 {                                                                       \
  61         if (bytes < 0 && sinfo->name < -bytes) {                        \
  62                 WARN_ON(1);                                             \
  63                 sinfo->name = 0;                                        \
  64                 return;                                                 \
  65         }                                                               \
  66         sinfo->name += bytes;                                           \
  67 }
  68
  69 DECLARE_SPACE_INFO_UPDATE(bytes_may_use);
  70 DECLARE_SPACE_INFO_UPDATE(bytes_pinned);
  71
  72 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
  73                                struct btrfs_delayed_ref_node *node, u64 parent,
  74                                u64 root_objectid, u64 owner_objectid,
  75                                u64 owner_offset, int refs_to_drop,
  76                                struct btrfs_delayed_extent_op *extra_op);
  77 static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op,
  78                                     struct extent_buffer *leaf,
  79                                     struct btrfs_extent_item *ei);
  80 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
  81                                       u64 parent, u64 root_objectid,
  82                                       u64 flags, u64 owner, u64 offset,
  83                                       struct btrfs_key *ins, int ref_mod);
  84 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
  85                                      struct btrfs_delayed_ref_node *node,
  86                                      struct btrfs_delayed_extent_op *extent_op);
  87 static int do_chunk_alloc(struct btrfs_trans_handle *trans, u64 flags,
  88                           int force);
  89 static int find_next_key(struct btrfs_path *path, int level,
  90                          struct btrfs_key *key);
  91 static void dump_space_info(struct btrfs_fs_info *fs_info,
  92                             struct btrfs_space_info *info, u64 bytes,
  93                             int dump_block_groups);
  94 static int block_rsv_use_bytes(struct btrfs_block_rsv *block_rsv,
  95                                u64 num_bytes);
  96 static void space_info_add_new_bytes(struct btrfs_fs_info *fs_info,
  97                                      struct btrfs_space_info *space_info,
  98                                      u64 num_bytes);
  99 static void space_info_add_old_bytes(struct btrfs_fs_info *fs_info,
 100                                      struct btrfs_space_info *space_info,
 101                                      u64 num_bytes);
 102
 103 static noinline int
 104 block_group_cache_done(struct btrfs_block_group_cache *cache)
 105 {
 106         smp_mb();
 107         return cache->cached == BTRFS_CACHE_FINISHED ||
 108                 cache->cached == BTRFS_CACHE_ERROR;
 109 }
 110
 111 static int block_group_bits(struct btrfs_block_group_cache *cache, u64 bits)
 112 {
 113         return (cache->flags & bits) == bits;
 114 }
 115
 116 void btrfs_get_block_group(struct btrfs_block_group_cache *cache)
 117 {
 118         atomic_inc(&cache->count);
 119 }
 120
 121 void btrfs_put_block_group(struct btrfs_block_group_cache *cache)
 122 {
 123         if (atomic_dec_and_test(&cache->count)) {
 124                 WARN_ON(cache->pinned > 0);
 125                 WARN_ON(cache->reserved > 0);
 126
 127                 /*
 128                  * If not empty, someone is still holding mutex of
 129                  * full_stripe_lock, which can only be released by caller.
 130                  * And it will definitely cause use-after-free when caller
 131                  * tries to release full stripe lock.
 132                  *
 133                  * No better way to resolve, but only to warn.
 134                  */
 135                 WARN_ON(!RB_EMPTY_ROOT(&cache->full_stripe_locks_root.root));
 136                 kfree(cache->free_space_ctl);
 137                 kfree(cache);
 138         }
 139 }
 140
 141 /*
 142  * this adds the block group to the fs_info rb tree for the block group
 143  * cache
 144  */
 145 static int btrfs_add_block_group_cache(struct btrfs_fs_info *info,
 146                                 struct btrfs_block_group_cache *block_group)
 147 {
 148         struct rb_node **p;
 149         struct rb_node *parent = NULL;
 150         struct btrfs_block_group_cache *cache;
 151
 152         spin_lock(&info->block_group_cache_lock);
 153         p = &info->block_group_cache_tree.rb_node;
 154
 155         while (*p) {
 156                 parent = *p;
 157                 cache = rb_entry(parent, struct btrfs_block_group_cache,
 158                                  cache_node);
 159                 if (block_group->key.objectid < cache->key.objectid) {
 160                         p = &(*p)->rb_left;
 161                 } else if (block_group->key.objectid > cache->key.objectid) {
 162                         p = &(*p)->rb_right;
 163                 } else {
 164                         spin_unlock(&info->block_group_cache_lock);
 165                         return -EEXIST;
 166                 }
 167         }
 168
 169         rb_link_node(&block_group->cache_node, parent, p);
 170         rb_insert_color(&block_group->cache_node,
 171                         &info->block_group_cache_tree);
 172
 173         if (info->first_logical_byte > block_group->key.objectid)
 174                 info->first_logical_byte = block_group->key.objectid;
 175
 176         spin_unlock(&info->block_group_cache_lock);
 177
 178         return 0;
 179 }
 180
 181 /*
 182  * This will return the block group at or after bytenr if contains is 0, else
 183  * it will return the block group that contains the bytenr
 184  */
 185 static struct btrfs_block_group_cache *
 186 block_group_cache_tree_search(struct btrfs_fs_info *info, u64 bytenr,
 187                               int contains)
 188 {
 189         struct btrfs_block_group_cache *cache, *ret = NULL;
 190         struct rb_node *n;
 191         u64 end, start;
 192
 193         spin_lock(&info->block_group_cache_lock);
 194         n = info->block_group_cache_tree.rb_node;
 195
 196         while (n) {
 197                 cache = rb_entry(n, struct btrfs_block_group_cache,
 198                                  cache_node);
 199                 end = cache->key.objectid + cache->key.offset - 1;
 200                 start = cache->key.objectid;
 201
 202                 if (bytenr < start) {
 203                         if (!contains && (!ret || start < ret->key.objectid))
 204                                 ret = cache;
 205                         n = n->rb_left;
 206                 } else if (bytenr > start) {
 207                         if (contains && bytenr <= end) {
 208                                 ret = cache;
 209                                 break;
 210                         }
 211                         n = n->rb_right;
 212                 } else {
 213                         ret = cache;
 214                         break;
 215                 }
 216         }
 217         if (ret) {
 218                 btrfs_get_block_group(ret);
 219                 if (bytenr == 0 && info->first_logical_byte > ret->key.objectid)
 220                         info->first_logical_byte = ret->key.objectid;
 221         }
 222         spin_unlock(&info->block_group_cache_lock);
 223
 224         return ret;
 225 }
 226
 227 static int add_excluded_extent(struct btrfs_fs_info *fs_info,
 228                                u64 start, u64 num_bytes)
 229 {
 230         u64 end = start + num_bytes - 1;
 231         set_extent_bits(&fs_info->freed_extents[0],
 232                         start, end, EXTENT_UPTODATE);
 233         set_extent_bits(&fs_info->freed_extents[1],
 234                         start, end, EXTENT_UPTODATE);
 235         return 0;
 236 }
 237
 238 static void free_excluded_extents(struct btrfs_block_group_cache *cache)
 239 {
 240         struct btrfs_fs_info *fs_info = cache->fs_info;
 241         u64 start, end;
 242
 243         start = cache->key.objectid;
 244         end = start + cache->key.offset - 1;
 245
 246         clear_extent_bits(&fs_info->freed_extents[0],
 247                           start, end, EXTENT_UPTODATE);
 248         clear_extent_bits(&fs_info->freed_extents[1],
 249                           start, end, EXTENT_UPTODATE);
 250 }
 251
 252 static int exclude_super_stripes(struct btrfs_block_group_cache *cache)
 253 {
 254         struct btrfs_fs_info *fs_info = cache->fs_info;
 255         u64 bytenr;
 256         u64 *logical;
 257         int stripe_len;
 258         int i, nr, ret;
 259
 260         if (cache->key.objectid < BTRFS_SUPER_INFO_OFFSET) {
 261                 stripe_len = BTRFS_SUPER_INFO_OFFSET - cache->key.objectid;
 262                 cache->bytes_super += stripe_len;
 263                 ret = add_excluded_extent(fs_info, cache->key.objectid,
 264                                           stripe_len);
 265                 if (ret)
 266                         return ret;
 267         }
 268
 269         for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
 270                 bytenr = btrfs_sb_offset(i);
 271                 ret = btrfs_rmap_block(fs_info, cache->key.objectid,
 272                                        bytenr, &logical, &nr, &stripe_len);
 273                 if (ret)
 274                         return ret;
 275
 276                 while (nr--) {
 277                         u64 start, len;
 278
 279                         if (logical[nr] > cache->key.objectid +
 280                             cache->key.offset)
 281                                 continue;
 282
 283                         if (logical[nr] + stripe_len <= cache->key.objectid)
 284                                 continue;
 285
 286                         start = logical[nr];
 287                         if (start < cache->key.objectid) {
 288                                 start = cache->key.objectid;
 289                                 len = (logical[nr] + stripe_len) - start;
 290                         } else {
 291                                 len = min_t(u64, stripe_len,
 292                                             cache->key.objectid +
 293                                             cache->key.offset - start);
 294                         }
 295
 296                         cache->bytes_super += len;
 297                         ret = add_excluded_extent(fs_info, start, len);
 298                         if (ret) {
 299                                 kfree(logical);
 300                                 return ret;
 301                         }
 302                 }
 303
 304                 kfree(logical);
 305         }
 306         return 0;
 307 }
 308
 309 static struct btrfs_caching_control *
 310 get_caching_control(struct btrfs_block_group_cache *cache)
 311 {
 312         struct btrfs_caching_control *ctl;
 313
 314         spin_lock(&cache->lock);
 315         if (!cache->caching_ctl) {
 316                 spin_unlock(&cache->lock);
 317                 return NULL;
 318         }
 319
 320         ctl = cache->caching_ctl;
 321         refcount_inc(&ctl->count);
 322         spin_unlock(&cache->lock);
 323         return ctl;
 324 }
 325
 326 static void put_caching_control(struct btrfs_caching_control *ctl)
 327 {
 328         if (refcount_dec_and_test(&ctl->count))
 329                 kfree(ctl);
 330 }
 331
 332 #ifdef CONFIG_BTRFS_DEBUG
 333 static void fragment_free_space(struct btrfs_block_group_cache *block_group)
 334 {
 335         struct btrfs_fs_info *fs_info = block_group->fs_info;
 336         u64 start = block_group->key.objectid;
 337         u64 len = block_group->key.offset;
 338         u64 chunk = block_group->flags & BTRFS_BLOCK_GROUP_METADATA ?
 339                 fs_info->nodesize : fs_info->sectorsize;
 340         u64 step = chunk << 1;
 341
 342         while (len > chunk) {
 343                 btrfs_remove_free_space(block_group, start, chunk);
 344                 start += step;
 345                 if (len < step)
 346                         len = 0;
 347                 else
 348                         len -= step;
 349         }
 350 }
 351 #endif
 352
 353 /*
 354  * this is only called by cache_block_group, since we could have freed extents
 355  * we need to check the pinned_extents for any extents that can't be used yet
 356  * since their free space will be released as soon as the transaction commits.
 357  */
 358 u64 add_new_free_space(struct btrfs_block_group_cache *block_group,
 359                        u64 start, u64 end)
 360 {
 361         struct btrfs_fs_info *info = block_group->fs_info;
 362         u64 extent_start, extent_end, size, total_added = 0;
 363         int ret;
 364
 365         while (start < end) {
 366                 ret = find_first_extent_bit(info->pinned_extents, start,
 367                                             &extent_start, &extent_end,
 368                                             EXTENT_DIRTY | EXTENT_UPTODATE,
 369                                             NULL);
 370                 if (ret)
 371                         break;
 372
 373                 if (extent_start <= start) {
 374                         start = extent_end + 1;
 375                 } else if (extent_start > start && extent_start < end) {
 376                         size = extent_start - start;
 377                         total_added += size;
 378                         ret = btrfs_add_free_space(block_group, start,
 379                                                    size);
 380                         BUG_ON(ret); /* -ENOMEM or logic error */
 381                         start = extent_end + 1;
 382                 } else {
 383                         break;
 384                 }
 385         }
 386
 387         if (start < end) {
 388                 size = end - start;
 389                 total_added += size;
 390                 ret = btrfs_add_free_space(block_group, start, size);
 391                 BUG_ON(ret); /* -ENOMEM or logic error */
 392         }
 393
 394         return total_added;
 395 }
 396
 397 static int load_extent_tree_free(struct btrfs_caching_control *caching_ctl)
 398 {
 399         struct btrfs_block_group_cache *block_group = caching_ctl->block_group;
 400         struct btrfs_fs_info *fs_info = block_group->fs_info;
 401         struct btrfs_root *extent_root = fs_info->extent_root;
 402         struct btrfs_path *path;
 403         struct extent_buffer *leaf;
 404         struct btrfs_key key;
 405         u64 total_found = 0;
 406         u64 last = 0;
 407         u32 nritems;
 408         int ret;
 409         bool wakeup = true;
 410
 411         path = btrfs_alloc_path();
 412         if (!path)
 413                 return -ENOMEM;
 414
 415         last = max_t(u64, block_group->key.objectid, BTRFS_SUPER_INFO_OFFSET);
 416
 417 #ifdef CONFIG_BTRFS_DEBUG
 418         /*
 419          * If we're fragmenting we don't want to make anybody think we can
 420          * allocate from this block group until we've had a chance to fragment
 421          * the free space.
 422          */
 423         if (btrfs_should_fragment_free_space(block_group))
 424                 wakeup = false;
 425 #endif
 426         /*
 427          * We don't want to deadlock with somebody trying to allocate a new
 428          * extent for the extent root while also trying to search the extent
 429          * root to add free space.  So we skip locking and search the commit
 430          * root, since its read-only
 431          */
 432         path->skip_locking = 1;
 433         path->search_commit_root = 1;
 434         path->reada = READA_FORWARD;
 435
 436         key.objectid = last;
 437         key.offset = 0;
 438         key.type = BTRFS_EXTENT_ITEM_KEY;
 439
 440 next:
 441         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
 442         if (ret < 0)
 443                 goto out;
 444
 445         leaf = path->nodes[0];
 446         nritems = btrfs_header_nritems(leaf);
 447
 448         while (1) {
 449                 if (btrfs_fs_closing(fs_info) > 1) {
 450                         last = (u64)-1;
 451                         break;
 452                 }
 453
 454                 if (path->slots[0] < nritems) {
 455                         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
 456                 } else {
 457                         ret = find_next_key(path, 0, &key);
 458                         if (ret)
 459                                 break;
 460
 461                         if (need_resched() ||
 462                             rwsem_is_contended(&fs_info->commit_root_sem)) {
 463                                 if (wakeup)
 464                                         caching_ctl->progress = last;
 465                                 btrfs_release_path(path);
 466                                 up_read(&fs_info->commit_root_sem);
 467                                 mutex_unlock(&caching_ctl->mutex);
 468                                 cond_resched();
 469                                 mutex_lock(&caching_ctl->mutex);
 470                                 down_read(&fs_info->commit_root_sem);
 471                                 goto next;
 472                         }
 473
 474                         ret = btrfs_next_leaf(extent_root, path);
 475                         if (ret < 0)
 476                                 goto out;
 477                         if (ret)
 478                                 break;
 479                         leaf = path->nodes[0];
 480                         nritems = btrfs_header_nritems(leaf);
 481                         continue;
 482                 }
 483
 484                 if (key.objectid < last) {
 485                         key.objectid = last;
 486                         key.offset = 0;
 487                         key.type = BTRFS_EXTENT_ITEM_KEY;
 488
 489                         if (wakeup)
 490                                 caching_ctl->progress = last;
 491                         btrfs_release_path(path);
 492                         goto next;
 493                 }
 494
 495                 if (key.objectid < block_group->key.objectid) {
 496                         path->slots[0]++;
 497                         continue;
 498                 }
 499
 500                 if (key.objectid >= block_group->key.objectid +
 501                     block_group->key.offset)
 502                         break;
 503
 504                 if (key.type == BTRFS_EXTENT_ITEM_KEY ||
 505                     key.type == BTRFS_METADATA_ITEM_KEY) {
 506                         total_found += add_new_free_space(block_group, last,
 507                                                           key.objectid);
 508                         if (key.type == BTRFS_METADATA_ITEM_KEY)
 509                                 last = key.objectid +
 510                                         fs_info->nodesize;
 511                         else
 512                                 last = key.objectid + key.offset;
 513
 514                         if (total_found > CACHING_CTL_WAKE_UP) {
 515                                 total_found = 0;
 516                                 if (wakeup)
 517                                         wake_up(&caching_ctl->wait);
 518                         }
 519                 }
 520                 path->slots[0]++;
 521         }
 522         ret = 0;
 523
 524         total_found += add_new_free_space(block_group, last,
 525                                           block_group->key.objectid +
 526                                           block_group->key.offset);
 527         caching_ctl->progress = (u64)-1;
 528
 529 out:
 530         btrfs_free_path(path);
 531         return ret;
 532 }
 533
 534 static noinline void caching_thread(struct btrfs_work *work)
 535 {
 536         struct btrfs_block_group_cache *block_group;
 537         struct btrfs_fs_info *fs_info;
 538         struct btrfs_caching_control *caching_ctl;
 539         int ret;
 540
 541         caching_ctl = container_of(work, struct btrfs_caching_control, work);
 542         block_group = caching_ctl->block_group;
 543         fs_info = block_group->fs_info;
 544
 545         mutex_lock(&caching_ctl->mutex);
 546         down_read(&fs_info->commit_root_sem);
 547
 548         if (btrfs_fs_compat_ro(fs_info, FREE_SPACE_TREE))
 549                 ret = load_free_space_tree(caching_ctl);
 550         else
 551                 ret = load_extent_tree_free(caching_ctl);
 552
 553         spin_lock(&block_group->lock);
 554         block_group->caching_ctl = NULL;
 555         block_group->cached = ret ? BTRFS_CACHE_ERROR : BTRFS_CACHE_FINISHED;
 556         spin_unlock(&block_group->lock);
 557
 558 #ifdef CONFIG_BTRFS_DEBUG
 559         if (btrfs_should_fragment_free_space(block_group)) {
 560                 u64 bytes_used;
 561
 562                 spin_lock(&block_group->space_info->lock);
 563                 spin_lock(&block_group->lock);
 564                 bytes_used = block_group->key.offset -
 565                         btrfs_block_group_used(&block_group->item);
 566                 block_group->space_info->bytes_used += bytes_used >> 1;
 567                 spin_unlock(&block_group->lock);
 568                 spin_unlock(&block_group->space_info->lock);
 569                 fragment_free_space(block_group);
 570         }
 571 #endif
 572
 573         caching_ctl->progress = (u64)-1;
 574
 575         up_read(&fs_info->commit_root_sem);
 576         free_excluded_extents(block_group);
 577         mutex_unlock(&caching_ctl->mutex);
 578
 579         wake_up(&caching_ctl->wait);
 580
 581         put_caching_control(caching_ctl);
 582         btrfs_put_block_group(block_group);
 583 }
 584
 585 static int cache_block_group(struct btrfs_block_group_cache *cache,
 586                              int load_cache_only)
 587 {
 588         DEFINE_WAIT(wait);
 589         struct btrfs_fs_info *fs_info = cache->fs_info;
 590         struct btrfs_caching_control *caching_ctl;
 591         int ret = 0;
 592
 593         caching_ctl = kzalloc(sizeof(*caching_ctl), GFP_NOFS);
 594         if (!caching_ctl)
 595                 return -ENOMEM;
 596
 597         INIT_LIST_HEAD(&caching_ctl->list);
 598         mutex_init(&caching_ctl->mutex);
 599         init_waitqueue_head(&caching_ctl->wait);
 600         caching_ctl->block_group = cache;
 601         caching_ctl->progress = cache->key.objectid;
 602         refcount_set(&caching_ctl->count, 1);
 603         btrfs_init_work(&caching_ctl->work, btrfs_cache_helper,
 604                         caching_thread, NULL, NULL);
 605
 606         spin_lock(&cache->lock);
 607         /*
 608          * This should be a rare occasion, but this could happen I think in the
 609          * case where one thread starts to load the space cache info, and then
 610          * some other thread starts a transaction commit which tries to do an
 611          * allocation while the other thread is still loading the space cache
 612          * info.  The previous loop should have kept us from choosing this block
 613          * group, but if we've moved to the state where we will wait on caching
 614          * block groups we need to first check if we're doing a fast load here,
 615          * so we can wait for it to finish, otherwise we could end up allocating
 616          * from a block group who's cache gets evicted for one reason or
 617          * another.
 618          */
 619         while (cache->cached == BTRFS_CACHE_FAST) {
 620                 struct btrfs_caching_control *ctl;
 621
 622                 ctl = cache->caching_ctl;
 623                 refcount_inc(&ctl->count);
 624                 prepare_to_wait(&ctl->wait, &wait, TASK_UNINTERRUPTIBLE);
 625                 spin_unlock(&cache->lock);
 626
 627                 schedule();
 628
 629                 finish_wait(&ctl->wait, &wait);
 630                 put_caching_control(ctl);
 631                 spin_lock(&cache->lock);
 632         }
 633
 634         if (cache->cached != BTRFS_CACHE_NO) {
 635                 spin_unlock(&cache->lock);
 636                 kfree(caching_ctl);
 637                 return 0;
 638         }
 639         WARN_ON(cache->caching_ctl);
 640         cache->caching_ctl = caching_ctl;
 641         cache->cached = BTRFS_CACHE_FAST;
 642         spin_unlock(&cache->lock);
 643
 644         if (btrfs_test_opt(fs_info, SPACE_CACHE)) {
 645                 mutex_lock(&caching_ctl->mutex);
 646                 ret = load_free_space_cache(cache);
 647
 648                 spin_lock(&cache->lock);
 649                 if (ret == 1) {
 650                         cache->caching_ctl = NULL;
 651                         cache->cached = BTRFS_CACHE_FINISHED;
 652                         cache->last_byte_to_unpin = (u64)-1;
 653                         caching_ctl->progress = (u64)-1;
 654                 } else {
 655                         if (load_cache_only) {
 656                                 cache->caching_ctl = NULL;
 657                                 cache->cached = BTRFS_CACHE_NO;
 658                         } else {
 659                                 cache->cached = BTRFS_CACHE_STARTED;
 660                                 cache->has_caching_ctl = 1;
 661                         }
 662                 }
 663                 spin_unlock(&cache->lock);
 664 #ifdef CONFIG_BTRFS_DEBUG
 665                 if (ret == 1 &&
 666                     btrfs_should_fragment_free_space(cache)) {
 667                         u64 bytes_used;
 668
 669                         spin_lock(&cache->space_info->lock);
 670                         spin_lock(&cache->lock);
 671                         bytes_used = cache->key.offset -
 672                                 btrfs_block_group_used(&cache->item);
 673                         cache->space_info->bytes_used += bytes_used >> 1;
 674                         spin_unlock(&cache->lock);
 675                         spin_unlock(&cache->space_info->lock);
 676                         fragment_free_space(cache);
 677                 }
 678 #endif
 679                 mutex_unlock(&caching_ctl->mutex);
 680
 681                 wake_up(&caching_ctl->wait);
 682                 if (ret == 1) {
 683                         put_caching_control(caching_ctl);
 684                         free_excluded_extents(cache);
 685                         return 0;
 686                 }
 687         } else {
 688                 /*
 689                  * We're either using the free space tree or no caching at all.
 690                  * Set cached to the appropriate value and wakeup any waiters.
 691                  */
 692                 spin_lock(&cache->lock);
 693                 if (load_cache_only) {
 694                         cache->caching_ctl = NULL;
 695                         cache->cached = BTRFS_CACHE_NO;
 696                 } else {
 697                         cache->cached = BTRFS_CACHE_STARTED;
 698                         cache->has_caching_ctl = 1;
 699                 }
 700                 spin_unlock(&cache->lock);
 701                 wake_up(&caching_ctl->wait);
 702         }
 703
 704         if (load_cache_only) {
 705                 put_caching_control(caching_ctl);
 706                 return 0;
 707         }
 708
 709         down_write(&fs_info->commit_root_sem);
 710         refcount_inc(&caching_ctl->count);
 711         list_add_tail(&caching_ctl->list, &fs_info->caching_block_groups);
 712         up_write(&fs_info->commit_root_sem);
 713
 714         btrfs_get_block_group(cache);
 715
 716         btrfs_queue_work(fs_info->caching_workers, &caching_ctl->work);
 717
 718         return ret;
 719 }
 720
 721 /*
 722  * return the block group that starts at or after bytenr
 723  */
 724 static struct btrfs_block_group_cache *
 725 btrfs_lookup_first_block_group(struct btrfs_fs_info *info, u64 bytenr)
 726 {
 727         return block_group_cache_tree_search(info, bytenr, 0);
 728 }
 729
 730 /*
 731  * return the block group that contains the given bytenr
 732  */
 733 struct btrfs_block_group_cache *btrfs_lookup_block_group(
 734                                                  struct btrfs_fs_info *info,
 735                                                  u64 bytenr)
 736 {
 737         return block_group_cache_tree_search(info, bytenr, 1);
 738 }
 739
 740 static struct btrfs_space_info *__find_space_info(struct btrfs_fs_info *info,
 741                                                   u64 flags)
 742 {
 743         struct list_head *head = &info->space_info;
 744         struct btrfs_space_info *found;
 745
 746         flags &= BTRFS_BLOCK_GROUP_TYPE_MASK;
 747
 748         rcu_read_lock();
 749         list_for_each_entry_rcu(found, head, list) {
 750                 if (found->flags & flags) {
 751                         rcu_read_unlock();
 752                         return found;
 753                 }
 754         }
 755         rcu_read_unlock();
 756         return NULL;
 757 }
 758
 759 static void add_pinned_bytes(struct btrfs_fs_info *fs_info,
 760                              struct btrfs_ref *ref, int sign)
 761 {
 762         struct btrfs_space_info *space_info;
 763         s64 num_bytes;
 764         u64 flags;
 765
 766         ASSERT(sign == 1 || sign == -1);
 767         num_bytes = sign * ref->len;
 768         if (ref->type == BTRFS_REF_METADATA) {
 769                 if (ref->tree_ref.root == BTRFS_CHUNK_TREE_OBJECTID)
 770                         flags = BTRFS_BLOCK_GROUP_SYSTEM;
 771                 else
 772                         flags = BTRFS_BLOCK_GROUP_METADATA;
 773         } else {
 774                 flags = BTRFS_BLOCK_GROUP_DATA;
 775         }
 776
 777         space_info = __find_space_info(fs_info, flags);
 778         ASSERT(space_info);
 779         percpu_counter_add_batch(&space_info->total_bytes_pinned, num_bytes,
 780                     BTRFS_TOTAL_BYTES_PINNED_BATCH);
 781 }
 782
 783 /*
 784  * after adding space to the filesystem, we need to clear the full flags
 785  * on all the space infos.
 786  */
 787 void btrfs_clear_space_info_full(struct btrfs_fs_info *info)
 788 {
 789         struct list_head *head = &info->space_info;
 790         struct btrfs_space_info *found;
 791
 792         rcu_read_lock();
 793         list_for_each_entry_rcu(found, head, list)
 794                 found->full = 0;
 795         rcu_read_unlock();
 796 }
 797
 798 /* simple helper to search for an existing data extent at a given offset */
 799 int btrfs_lookup_data_extent(struct btrfs_fs_info *fs_info, u64 start, u64 len)
 800 {
 801         int ret;
 802         struct btrfs_key key;
 803         struct btrfs_path *path;
 804
 805         path = btrfs_alloc_path();
 806         if (!path)
 807                 return -ENOMEM;
 808
 809         key.objectid = start;
 810         key.offset = len;
 811         key.type = BTRFS_EXTENT_ITEM_KEY;
 812         ret = btrfs_search_slot(NULL, fs_info->extent_root, &key, path, 0, 0);
 813         btrfs_free_path(path);
 814         return ret;
 815 }
 816
 817 /*
 818  * helper function to lookup reference count and flags of a tree block.
 819  *
 820  * the head node for delayed ref is used to store the sum of all the
 821  * reference count modifications queued up in the rbtree. the head
 822  * node may also store the extent flags to set. This way you can check
 823  * to see what the reference count and extent flags would be if all of
 824  * the delayed refs are not processed.
 825  */
 826 int btrfs_lookup_extent_info(struct btrfs_trans_handle *trans,
 827                              struct btrfs_fs_info *fs_info, u64 bytenr,
 828                              u64 offset, int metadata, u64 *refs, u64 *flags)
 829 {
 830         struct btrfs_delayed_ref_head *head;
 831         struct btrfs_delayed_ref_root *delayed_refs;
 832         struct btrfs_path *path;
 833         struct btrfs_extent_item *ei;
 834         struct extent_buffer *leaf;
 835         struct btrfs_key key;
 836         u32 item_size;
 837         u64 num_refs;
 838         u64 extent_flags;
 839         int ret;
 840
 841         /*
 842          * If we don't have skinny metadata, don't bother doing anything
 843          * different
 844          */
 845         if (metadata && !btrfs_fs_incompat(fs_info, SKINNY_METADATA)) {
 846                 offset = fs_info->nodesize;
 847                 metadata = 0;
 848         }
 849
 850         path = btrfs_alloc_path();
 851         if (!path)
 852                 return -ENOMEM;
 853
 854         if (!trans) {
 855                 path->skip_locking = 1;
 856                 path->search_commit_root = 1;
 857         }
 858
 859 search_again:
 860         key.objectid = bytenr;
 861         key.offset = offset;
 862         if (metadata)
 863                 key.type = BTRFS_METADATA_ITEM_KEY;
 864         else
 865                 key.type = BTRFS_EXTENT_ITEM_KEY;
 866
 867         ret = btrfs_search_slot(trans, fs_info->extent_root, &key, path, 0, 0);
 868         if (ret < 0)
 869                 goto out_free;
 870
 871         if (ret > 0 && metadata && key.type == BTRFS_METADATA_ITEM_KEY) {
 872                 if (path->slots[0]) {
 873                         path->slots[0]--;
 874                         btrfs_item_key_to_cpu(path->nodes[0], &key,
 875                                               path->slots[0]);
 876                         if (key.objectid == bytenr &&
 877                             key.type == BTRFS_EXTENT_ITEM_KEY &&
 878                             key.offset == fs_info->nodesize)
 879                                 ret = 0;
 880                 }
 881         }
 882
 883         if (ret == 0) {
 884                 leaf = path->nodes[0];
 885                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
 886                 if (item_size >= sizeof(*ei)) {
 887                         ei = btrfs_item_ptr(leaf, path->slots[0],
 888                                             struct btrfs_extent_item);
 889                         num_refs = btrfs_extent_refs(leaf, ei);
 890                         extent_flags = btrfs_extent_flags(leaf, ei);
 891                 } else {
 892                         ret = -EINVAL;
 893                         btrfs_print_v0_err(fs_info);
 894                         if (trans)
 895                                 btrfs_abort_transaction(trans, ret);
 896                         else
 897                                 btrfs_handle_fs_error(fs_info, ret, NULL);
 898
 899                         goto out_free;
 900                 }
 901
 902                 BUG_ON(num_refs == 0);
 903         } else {
 904                 num_refs = 0;
 905                 extent_flags = 0;
 906                 ret = 0;
 907         }
 908
 909         if (!trans)
 910                 goto out;
 911
 912         delayed_refs = &trans->transaction->delayed_refs;
 913         spin_lock(&delayed_refs->lock);
 914         head = btrfs_find_delayed_ref_head(delayed_refs, bytenr);
 915         if (head) {
 916                 if (!mutex_trylock(&head->mutex)) {
 917                         refcount_inc(&head->refs);
 918                         spin_unlock(&delayed_refs->lock);
 919
 920                         btrfs_release_path(path);
 921
 922                         /*
 923                          * Mutex was contended, block until it's released and try
 924                          * again
 925                          */
 926                         mutex_lock(&head->mutex);
 927                         mutex_unlock(&head->mutex);
 928                         btrfs_put_delayed_ref_head(head);
 929                         goto search_again;
 930                 }
 931                 spin_lock(&head->lock);
 932                 if (head->extent_op && head->extent_op->update_flags)
 933                         extent_flags |= head->extent_op->flags_to_set;
 934                 else
 935                         BUG_ON(num_refs == 0);
 936
 937                 num_refs += head->ref_mod;
 938                 spin_unlock(&head->lock);
 939                 mutex_unlock(&head->mutex);
 940         }
 941         spin_unlock(&delayed_refs->lock);
 942 out:
 943         WARN_ON(num_refs == 0);
 944         if (refs)
 945                 *refs = num_refs;
 946         if (flags)
 947                 *flags = extent_flags;
 948 out_free:
 949         btrfs_free_path(path);
 950         return ret;
 951 }
 952
 953 /*
 954  * Back reference rules.  Back refs have three main goals:
 955  *
 956  * 1) differentiate between all holders of references to an extent so that
 957  *    when a reference is dropped we can make sure it was a valid reference
 958  *    before freeing the extent.
 959  *
 960  * 2) Provide enough information to quickly find the holders of an extent
 961  *    if we notice a given block is corrupted or bad.
 962  *
 963  * 3) Make it easy to migrate blocks for FS shrinking or storage pool
 964  *    maintenance.  This is actually the same as #2, but with a slightly
 965  *    different use case.
 966  *
 967  * There are two kinds of back refs. The implicit back refs is optimized
 968  * for pointers in non-shared tree blocks. For a given pointer in a block,
 969  * back refs of this kind provide information about the block's owner tree
 970  * and the pointer's key. These information allow us to find the block by
 971  * b-tree searching. The full back refs is for pointers in tree blocks not
 972  * referenced by their owner trees. The location of tree block is recorded
 973  * in the back refs. Actually the full back refs is generic, and can be
 974  * used in all cases the implicit back refs is used. The major shortcoming
 975  * of the full back refs is its overhead. Every time a tree block gets
 976  * COWed, we have to update back refs entry for all pointers in it.
 977  *
 978  * For a newly allocated tree block, we use implicit back refs for
 979  * pointers in it. This means most tree related operations only involve
 980  * implicit back refs. For a tree block created in old transaction, the
 981  * only way to drop a reference to it is COW it. So we can detect the
 982  * event that tree block loses its owner tree's reference and do the
 983  * back refs conversion.
 984  *
 985  * When a tree block is COWed through a tree, there are four cases:
 986  *
 987  * The reference count of the block is one and the tree is the block's
 988  * owner tree. Nothing to do in this case.
 989  *
 990  * The reference count of the block is one and the tree is not the
 991  * block's owner tree. In this case, full back refs is used for pointers
 992  * in the block. Remove these full back refs, add implicit back refs for
 993  * every pointers in the new block.
 994  *
 995  * The reference count of the block is greater than one and the tree is
 996  * the block's owner tree. In this case, implicit back refs is used for
 997  * pointers in the block. Add full back refs for every pointers in the
 998  * block, increase lower level extents' reference counts. The original
 999  * implicit back refs are entailed to the new block.
1000  *
1001  * The reference count of the block is greater than one and the tree is
1002  * not the block's owner tree. Add implicit back refs for every pointer in
1003  * the new block, increase lower level extents' reference count.
1004  *
1005  * Back Reference Key composing:
1006  *
1007  * The key objectid corresponds to the first byte in the extent,
1008  * The key type is used to differentiate between types of back refs.
1009  * There are different meanings of the key offset for different types
1010  * of back refs.
1011  *
1012  * File extents can be referenced by:
1013  *
1014  * - multiple snapshots, subvolumes, or different generations in one subvol
1015  * - different files inside a single subvolume
1016  * - different offsets inside a file (bookend extents in file.c)
1017  *
1018  * The extent ref structure for the implicit back refs has fields for:
1019  *
1020  * - Objectid of the subvolume root
1021  * - objectid of the file holding the reference
1022  * - original offset in the file
1023  * - how many bookend extents
1024  *
1025  * The key offset for the implicit back refs is hash of the first
1026  * three fields.
1027  *
1028  * The extent ref structure for the full back refs has field for:
1029  *
1030  * - number of pointers in the tree leaf
1031  *
1032  * The key offset for the implicit back refs is the first byte of
1033  * the tree leaf
1034  *
1035  * When a file extent is allocated, The implicit back refs is used.
1036  * the fields are filled in:
1037  *
1038  *     (root_key.objectid, inode objectid, offset in file, 1)
1039  *
1040  * When a file extent is removed file truncation, we find the
1041  * corresponding implicit back refs and check the following fields:
1042  *
1043  *     (btrfs_header_owner(leaf), inode objectid, offset in file)
1044  *
1045  * Btree extents can be referenced by:
1046  *
1047  * - Different subvolumes
1048  *
1049  * Both the implicit back refs and the full back refs for tree blocks
1050  * only consist of key. The key offset for the implicit back refs is
1051  * objectid of block's owner tree. The key offset for the full back refs
1052  * is the first byte of parent block.
1053  *
1054  * When implicit back refs is used, information about the lowest key and
1055  * level of the tree block are required. These information are stored in
1056  * tree block info structure.
1057  */
1058
1059 /*
1060  * is_data == BTRFS_REF_TYPE_BLOCK, tree block type is required,
1061  * is_data == BTRFS_REF_TYPE_DATA, data type is requiried,
1062  * is_data == BTRFS_REF_TYPE_ANY, either type is OK.
1063  */
1064 int btrfs_get_extent_inline_ref_type(const struct extent_buffer *eb,
1065                                      struct btrfs_extent_inline_ref *iref,
1066                                      enum btrfs_inline_ref_type is_data)
1067 {
1068         int type = btrfs_extent_inline_ref_type(eb, iref);
1069         u64 offset = btrfs_extent_inline_ref_offset(eb, iref);
1070
1071         if (type == BTRFS_TREE_BLOCK_REF_KEY ||
1072             type == BTRFS_SHARED_BLOCK_REF_KEY ||
1073             type == BTRFS_SHARED_DATA_REF_KEY ||
1074             type == BTRFS_EXTENT_DATA_REF_KEY) {
1075                 if (is_data == BTRFS_REF_TYPE_BLOCK) {
1076                         if (type == BTRFS_TREE_BLOCK_REF_KEY)
1077                                 return type;
1078                         if (type == BTRFS_SHARED_BLOCK_REF_KEY) {
1079                                 ASSERT(eb->fs_info);
1080                                 /*
1081                                  * Every shared one has parent tree
1082                                  * block, which must be aligned to
1083                                  * nodesize.
1084                                  */
1085                                 if (offset &&
1086                                     IS_ALIGNED(offset, eb->fs_info->nodesize))
1087                                         return type;
1088                         }
1089                 } else if (is_data == BTRFS_REF_TYPE_DATA) {
1090                         if (type == BTRFS_EXTENT_DATA_REF_KEY)
1091                                 return type;
1092                         if (type == BTRFS_SHARED_DATA_REF_KEY) {
1093                                 ASSERT(eb->fs_info);
1094                                 /*
1095                                  * Every shared one has parent tree
1096                                  * block, which must be aligned to
1097                                  * nodesize.
1098                                  */
1099                                 if (offset &&
1100                                     IS_ALIGNED(offset, eb->fs_info->nodesize))
1101                                         return type;
1102                         }
1103                 } else {
1104                         ASSERT(is_data == BTRFS_REF_TYPE_ANY);
1105                         return type;
1106                 }
1107         }
1108
1109         btrfs_print_leaf((struct extent_buffer *)eb);
1110         btrfs_err(eb->fs_info, "eb %llu invalid extent inline ref type %d",
1111                   eb->start, type);
1112         WARN_ON(1);
1113
1114         return BTRFS_REF_TYPE_INVALID;
1115 }
1116
1117 static u64 hash_extent_data_ref(u64 root_objectid, u64 owner, u64 offset)
1118 {
1119         u32 high_crc = ~(u32)0;
1120         u32 low_crc = ~(u32)0;
1121         __le64 lenum;
1122
1123         lenum = cpu_to_le64(root_objectid);
1124         high_crc = crc32c(high_crc, &lenum, sizeof(lenum));
1125         lenum = cpu_to_le64(owner);
1126         low_crc = crc32c(low_crc, &lenum, sizeof(lenum));
1127         lenum = cpu_to_le64(offset);
1128         low_crc = crc32c(low_crc, &lenum, sizeof(lenum));
1129
1130         return ((u64)high_crc << 31) ^ (u64)low_crc;
1131 }
1132
1133 static u64 hash_extent_data_ref_item(struct extent_buffer *leaf,
1134                                      struct btrfs_extent_data_ref *ref)
1135 {
1136         return hash_extent_data_ref(btrfs_extent_data_ref_root(leaf, ref),
1137                                     btrfs_extent_data_ref_objectid(leaf, ref),
1138                                     btrfs_extent_data_ref_offset(leaf, ref));
1139 }
1140
1141 static int match_extent_data_ref(struct extent_buffer *leaf,
1142                                  struct btrfs_extent_data_ref *ref,
1143                                  u64 root_objectid, u64 owner, u64 offset)
1144 {
1145         if (btrfs_extent_data_ref_root(leaf, ref) != root_objectid ||
1146             btrfs_extent_data_ref_objectid(leaf, ref) != owner ||
1147             btrfs_extent_data_ref_offset(leaf, ref) != offset)
1148                 return 0;
1149         return 1;
1150 }
1151
1152 static noinline int lookup_extent_data_ref(struct btrfs_trans_handle *trans,
1153                                            struct btrfs_path *path,
1154                                            u64 bytenr, u64 parent,
1155                                            u64 root_objectid,
1156                                            u64 owner, u64 offset)
1157 {
1158         struct btrfs_root *root = trans->fs_info->extent_root;
1159         struct btrfs_key key;
1160         struct btrfs_extent_data_ref *ref;
1161         struct extent_buffer *leaf;
1162         u32 nritems;
1163         int ret;
1164         int recow;
1165         int err = -ENOENT;
1166
1167         key.objectid = bytenr;
1168         if (parent) {
1169                 key.type = BTRFS_SHARED_DATA_REF_KEY;
1170                 key.offset = parent;
1171         } else {
1172                 key.type = BTRFS_EXTENT_DATA_REF_KEY;
1173                 key.offset = hash_extent_data_ref(root_objectid,
1174                                                   owner, offset);
1175         }
1176 again:
1177         recow = 0;
1178         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1179         if (ret < 0) {
1180                 err = ret;
1181                 goto fail;
1182         }
1183
1184         if (parent) {
1185                 if (!ret)
1186                         return 0;
1187                 goto fail;
1188         }
1189
1190         leaf = path->nodes[0];
1191         nritems = btrfs_header_nritems(leaf);
1192         while (1) {
1193                 if (path->slots[0] >= nritems) {
1194                         ret = btrfs_next_leaf(root, path);
1195                         if (ret < 0)
1196                                 err = ret;
1197                         if (ret)
1198                                 goto fail;
1199
1200                         leaf = path->nodes[0];
1201                         nritems = btrfs_header_nritems(leaf);
1202                         recow = 1;
1203                 }
1204
1205                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1206                 if (key.objectid != bytenr ||
1207                     key.type != BTRFS_EXTENT_DATA_REF_KEY)
1208                         goto fail;
1209
1210                 ref = btrfs_item_ptr(leaf, path->slots[0],
1211                                      struct btrfs_extent_data_ref);
1212
1213                 if (match_extent_data_ref(leaf, ref, root_objectid,
1214                                           owner, offset)) {
1215                         if (recow) {
1216                                 btrfs_release_path(path);
1217                                 goto again;
1218                         }
1219                         err = 0;
1220                         break;
1221                 }
1222                 path->slots[0]++;
1223         }
1224 fail:
1225         return err;
1226 }
1227
1228 static noinline int insert_extent_data_ref(struct btrfs_trans_handle *trans,
1229                                            struct btrfs_path *path,
1230                                            u64 bytenr, u64 parent,
1231                                            u64 root_objectid, u64 owner,
1232                                            u64 offset, int refs_to_add)
1233 {
1234         struct btrfs_root *root = trans->fs_info->extent_root;
1235         struct btrfs_key key;
1236         struct extent_buffer *leaf;
1237         u32 size;
1238         u32 num_refs;
1239         int ret;
1240
1241         key.objectid = bytenr;
1242         if (parent) {
1243                 key.type = BTRFS_SHARED_DATA_REF_KEY;
1244                 key.offset = parent;
1245                 size = sizeof(struct btrfs_shared_data_ref);
1246         } else {
1247                 key.type = BTRFS_EXTENT_DATA_REF_KEY;
1248                 key.offset = hash_extent_data_ref(root_objectid,
1249                                                   owner, offset);
1250                 size = sizeof(struct btrfs_extent_data_ref);
1251         }
1252
1253         ret = btrfs_insert_empty_item(trans, root, path, &key, size);
1254         if (ret && ret != -EEXIST)
1255                 goto fail;
1256
1257         leaf = path->nodes[0];
1258         if (parent) {
1259                 struct btrfs_shared_data_ref *ref;
1260                 ref = btrfs_item_ptr(leaf, path->slots[0],
1261                                      struct btrfs_shared_data_ref);
1262                 if (ret == 0) {
1263                         btrfs_set_shared_data_ref_count(leaf, ref, refs_to_add);
1264                 } else {
1265                         num_refs = btrfs_shared_data_ref_count(leaf, ref);
1266                         num_refs += refs_to_add;
1267                         btrfs_set_shared_data_ref_count(leaf, ref, num_refs);
1268                 }
1269         } else {
1270                 struct btrfs_extent_data_ref *ref;
1271                 while (ret == -EEXIST) {
1272                         ref = btrfs_item_ptr(leaf, path->slots[0],
1273                                              struct btrfs_extent_data_ref);
1274                         if (match_extent_data_ref(leaf, ref, root_objectid,
1275                                                   owner, offset))
1276                                 break;
1277                         btrfs_release_path(path);
1278                         key.offset++;
1279                         ret = btrfs_insert_empty_item(trans, root, path, &key,
1280                                                       size);
1281                         if (ret && ret != -EEXIST)
1282                                 goto fail;
1283
1284                         leaf = path->nodes[0];
1285                 }
1286                 ref = btrfs_item_ptr(leaf, path->slots[0],
1287                                      struct btrfs_extent_data_ref);
1288                 if (ret == 0) {
1289                         btrfs_set_extent_data_ref_root(leaf, ref,
1290                                                        root_objectid);
1291                         btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
1292                         btrfs_set_extent_data_ref_offset(leaf, ref, offset);
1293                         btrfs_set_extent_data_ref_count(leaf, ref, refs_to_add);
1294                 } else {
1295                         num_refs = btrfs_extent_data_ref_count(leaf, ref);
1296                         num_refs += refs_to_add;
1297                         btrfs_set_extent_data_ref_count(leaf, ref, num_refs);
1298                 }
1299         }
1300         btrfs_mark_buffer_dirty(leaf);
1301         ret = 0;
1302 fail:
1303         btrfs_release_path(path);
1304         return ret;
1305 }
1306
1307 static noinline int remove_extent_data_ref(struct btrfs_trans_handle *trans,
1308                                            struct btrfs_path *path,
1309                                            int refs_to_drop, int *last_ref)
1310 {
1311         struct btrfs_key key;
1312         struct btrfs_extent_data_ref *ref1 = NULL;
1313         struct btrfs_shared_data_ref *ref2 = NULL;
1314         struct extent_buffer *leaf;
1315         u32 num_refs = 0;
1316         int ret = 0;
1317
1318         leaf = path->nodes[0];
1319         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1320
1321         if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
1322                 ref1 = btrfs_item_ptr(leaf, path->slots[0],
1323                                       struct btrfs_extent_data_ref);
1324                 num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1325         } else if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
1326                 ref2 = btrfs_item_ptr(leaf, path->slots[0],
1327                                       struct btrfs_shared_data_ref);
1328                 num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1329         } else if (unlikely(key.type == BTRFS_EXTENT_REF_V0_KEY)) {
1330                 btrfs_print_v0_err(trans->fs_info);
1331                 btrfs_abort_transaction(trans, -EINVAL);
1332                 return -EINVAL;
1333         } else {
1334                 BUG();
1335         }
1336
1337         BUG_ON(num_refs < refs_to_drop);
1338         num_refs -= refs_to_drop;
1339
1340         if (num_refs == 0) {
1341                 ret = btrfs_del_item(trans, trans->fs_info->extent_root, path);
1342                 *last_ref = 1;
1343         } else {
1344                 if (key.type == BTRFS_EXTENT_DATA_REF_KEY)
1345                         btrfs_set_extent_data_ref_count(leaf, ref1, num_refs);
1346                 else if (key.type == BTRFS_SHARED_DATA_REF_KEY)
1347                         btrfs_set_shared_data_ref_count(leaf, ref2, num_refs);
1348                 btrfs_mark_buffer_dirty(leaf);
1349         }
1350         return ret;
1351 }
1352
1353 static noinline u32 extent_data_ref_count(struct btrfs_path *path,
1354                                           struct btrfs_extent_inline_ref *iref)
1355 {
1356         struct btrfs_key key;
1357         struct extent_buffer *leaf;
1358         struct btrfs_extent_data_ref *ref1;
1359         struct btrfs_shared_data_ref *ref2;
1360         u32 num_refs = 0;
1361         int type;
1362
1363         leaf = path->nodes[0];
1364         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1365
1366         BUG_ON(key.type == BTRFS_EXTENT_REF_V0_KEY);
1367         if (iref) {
1368                 /*
1369                  * If type is invalid, we should have bailed out earlier than
1370                  * this call.
1371                  */
1372                 type = btrfs_get_extent_inline_ref_type(leaf, iref, BTRFS_REF_TYPE_DATA);
1373                 ASSERT(type != BTRFS_REF_TYPE_INVALID);
1374                 if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1375                         ref1 = (struct btrfs_extent_data_ref *)(&iref->offset);
1376                         num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1377                 } else {
1378                         ref2 = (struct btrfs_shared_data_ref *)(iref + 1);
1379                         num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1380                 }
1381         } else if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
1382                 ref1 = btrfs_item_ptr(leaf, path->slots[0],
1383                                       struct btrfs_extent_data_ref);
1384                 num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1385         } else if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
1386                 ref2 = btrfs_item_ptr(leaf, path->slots[0],
1387                                       struct btrfs_shared_data_ref);
1388                 num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1389         } else {
1390                 WARN_ON(1);
1391         }
1392         return num_refs;
1393 }
1394
1395 static noinline int lookup_tree_block_ref(struct btrfs_trans_handle *trans,
1396                                           struct btrfs_path *path,
1397                                           u64 bytenr, u64 parent,
1398                                           u64 root_objectid)
1399 {
1400         struct btrfs_root *root = trans->fs_info->extent_root;
1401         struct btrfs_key key;
1402         int ret;
1403
1404         key.objectid = bytenr;
1405         if (parent) {
1406                 key.type = BTRFS_SHARED_BLOCK_REF_KEY;
1407                 key.offset = parent;
1408         } else {
1409                 key.type = BTRFS_TREE_BLOCK_REF_KEY;
1410                 key.offset = root_objectid;
1411         }
1412
1413         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1414         if (ret > 0)
1415                 ret = -ENOENT;
1416         return ret;
1417 }
1418
1419 static noinline int insert_tree_block_ref(struct btrfs_trans_handle *trans,
1420                                           struct btrfs_path *path,
1421                                           u64 bytenr, u64 parent,
1422                                           u64 root_objectid)
1423 {
1424         struct btrfs_key key;
1425         int ret;
1426
1427         key.objectid = bytenr;
1428         if (parent) {
1429                 key.type = BTRFS_SHARED_BLOCK_REF_KEY;
1430                 key.offset = parent;
1431         } else {
1432                 key.type = BTRFS_TREE_BLOCK_REF_KEY;
1433                 key.offset = root_objectid;
1434         }
1435
1436         ret = btrfs_insert_empty_item(trans, trans->fs_info->extent_root,
1437                                       path, &key, 0);
1438         btrfs_release_path(path);
1439         return ret;
1440 }
1441
1442 static inline int extent_ref_type(u64 parent, u64 owner)
1443 {
1444         int type;
1445         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1446                 if (parent > 0)
1447                         type = BTRFS_SHARED_BLOCK_REF_KEY;
1448                 else
1449                         type = BTRFS_TREE_BLOCK_REF_KEY;
1450         } else {
1451                 if (parent > 0)
1452                         type = BTRFS_SHARED_DATA_REF_KEY;
1453                 else
1454                         type = BTRFS_EXTENT_DATA_REF_KEY;
1455         }
1456         return type;
1457 }
1458
1459 static int find_next_key(struct btrfs_path *path, int level,
1460                          struct btrfs_key *key)
1461
1462 {
1463         for (; level < BTRFS_MAX_LEVEL; level++) {
1464                 if (!path->nodes[level])
1465                         break;
1466                 if (path->slots[level] + 1 >=
1467                     btrfs_header_nritems(path->nodes[level]))
1468                         continue;
1469                 if (level == 0)
1470                         btrfs_item_key_to_cpu(path->nodes[level], key,
1471                                               path->slots[level] + 1);
1472                 else
1473                         btrfs_node_key_to_cpu(path->nodes[level], key,
1474                                               path->slots[level] + 1);
1475                 return 0;
1476         }
1477         return 1;
1478 }
1479
1480 /*
1481  * look for inline back ref. if back ref is found, *ref_ret is set
1482  * to the address of inline back ref, and 0 is returned.
1483  *
1484  * if back ref isn't found, *ref_ret is set to the address where it
1485  * should be inserted, and -ENOENT is returned.
1486  *
1487  * if insert is true and there are too many inline back refs, the path
1488  * points to the extent item, and -EAGAIN is returned.
1489  *
1490  * NOTE: inline back refs are ordered in the same way that back ref
1491  *       items in the tree are ordered.
1492  */
1493 static noinline_for_stack
1494 int lookup_inline_extent_backref(struct btrfs_trans_handle *trans,
1495                                  struct btrfs_path *path,
1496                                  struct btrfs_extent_inline_ref **ref_ret,
1497                                  u64 bytenr, u64 num_bytes,
1498                                  u64 parent, u64 root_objectid,
1499                                  u64 owner, u64 offset, int insert)
1500 {
1501         struct btrfs_fs_info *fs_info = trans->fs_info;
1502         struct btrfs_root *root = fs_info->extent_root;
1503         struct btrfs_key key;
1504         struct extent_buffer *leaf;
1505         struct btrfs_extent_item *ei;
1506         struct btrfs_extent_inline_ref *iref;
1507         u64 flags;
1508         u64 item_size;
1509         unsigned long ptr;
1510         unsigned long end;
1511         int extra_size;
1512         int type;
1513         int want;
1514         int ret;
1515         int err = 0;
1516         bool skinny_metadata = btrfs_fs_incompat(fs_info, SKINNY_METADATA);
1517         int needed;
1518
1519         key.objectid = bytenr;
1520         key.type = BTRFS_EXTENT_ITEM_KEY;
1521         key.offset = num_bytes;
1522
1523         want = extent_ref_type(parent, owner);
1524         if (insert) {
1525                 extra_size = btrfs_extent_inline_ref_size(want);
1526                 path->keep_locks = 1;
1527         } else
1528                 extra_size = -1;
1529
1530         /*
1531          * Owner is our level, so we can just add one to get the level for the
1532          * block we are interested in.
1533          */
1534         if (skinny_metadata && owner < BTRFS_FIRST_FREE_OBJECTID) {
1535                 key.type = BTRFS_METADATA_ITEM_KEY;
1536                 key.offset = owner;
1537         }
1538
1539 again:
1540         ret = btrfs_search_slot(trans, root, &key, path, extra_size, 1);
1541         if (ret < 0) {
1542                 err = ret;
1543                 goto out;
1544         }
1545
1546         /*
1547          * We may be a newly converted file system which still has the old fat
1548          * extent entries for metadata, so try and see if we have one of those.
1549          */
1550         if (ret > 0 && skinny_metadata) {
1551                 skinny_metadata = false;
1552                 if (path->slots[0]) {
1553                         path->slots[0]--;
1554                         btrfs_item_key_to_cpu(path->nodes[0], &key,
1555                                               path->slots[0]);
1556                         if (key.objectid == bytenr &&
1557                             key.type == BTRFS_EXTENT_ITEM_KEY &&
1558                             key.offset == num_bytes)
1559                                 ret = 0;
1560                 }
1561                 if (ret) {
1562                         key.objectid = bytenr;
1563                         key.type = BTRFS_EXTENT_ITEM_KEY;
1564                         key.offset = num_bytes;
1565                         btrfs_release_path(path);
1566                         goto again;
1567                 }
1568         }
1569
1570         if (ret && !insert) {
1571                 err = -ENOENT;
1572                 goto out;
1573         } else if (WARN_ON(ret)) {
1574                 err = -EIO;
1575                 goto out;
1576         }
1577
1578         leaf = path->nodes[0];
1579         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1580         if (unlikely(item_size < sizeof(*ei))) {
1581                 err = -EINVAL;
1582                 btrfs_print_v0_err(fs_info);
1583                 btrfs_abort_transaction(trans, err);
1584                 goto out;
1585         }
1586
1587         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1588         flags = btrfs_extent_flags(leaf, ei);
1589
1590         ptr = (unsigned long)(ei + 1);
1591         end = (unsigned long)ei + item_size;
1592
1593         if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK && !skinny_metadata) {
1594                 ptr += sizeof(struct btrfs_tree_block_info);
1595                 BUG_ON(ptr > end);
1596         }
1597
1598         if (owner >= BTRFS_FIRST_FREE_OBJECTID)
1599                 needed = BTRFS_REF_TYPE_DATA;
1600         else
1601                 needed = BTRFS_REF_TYPE_BLOCK;
1602
1603         err = -ENOENT;
1604         while (1) {
1605                 if (ptr >= end) {
1606                         WARN_ON(ptr > end);
1607                         break;
1608                 }
1609                 iref = (struct btrfs_extent_inline_ref *)ptr;
1610                 type = btrfs_get_extent_inline_ref_type(leaf, iref, needed);
1611                 if (type == BTRFS_REF_TYPE_INVALID) {
1612                         err = -EUCLEAN;
1613                         goto out;
1614                 }
1615
1616                 if (want < type)
1617                         break;
1618                 if (want > type) {
1619                         ptr += btrfs_extent_inline_ref_size(type);
1620                         continue;
1621                 }
1622
1623                 if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1624                         struct btrfs_extent_data_ref *dref;
1625                         dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1626                         if (match_extent_data_ref(leaf, dref, root_objectid,
1627                                                   owner, offset)) {
1628                                 err = 0;
1629                                 break;
1630                         }
1631                         if (hash_extent_data_ref_item(leaf, dref) <
1632                             hash_extent_data_ref(root_objectid, owner, offset))
1633                                 break;
1634                 } else {
1635                         u64 ref_offset;
1636                         ref_offset = btrfs_extent_inline_ref_offset(leaf, iref);
1637                         if (parent > 0) {
1638                                 if (parent == ref_offset) {
1639                                         err = 0;
1640                                         break;
1641                                 }
1642                                 if (ref_offset < parent)
1643                                         break;
1644                         } else {
1645                                 if (root_objectid == ref_offset) {
1646                                         err = 0;
1647                                         break;
1648                                 }
1649                                 if (ref_offset < root_objectid)
1650                                         break;
1651                         }
1652                 }
1653                 ptr += btrfs_extent_inline_ref_size(type);
1654         }
1655         if (err == -ENOENT && insert) {
1656                 if (item_size + extra_size >=
1657                     BTRFS_MAX_EXTENT_ITEM_SIZE(root)) {
1658                         err = -EAGAIN;
1659                         goto out;
1660                 }
1661                 /*
1662                  * To add new inline back ref, we have to make sure
1663                  * there is no corresponding back ref item.
1664                  * For simplicity, we just do not add new inline back
1665                  * ref if there is any kind of item for this block
1666                  */
1667                 if (find_next_key(path, 0, &key) == 0 &&
1668                     key.objectid == bytenr &&
1669                     key.type < BTRFS_BLOCK_GROUP_ITEM_KEY) {
1670                         err = -EAGAIN;
1671                         goto out;
1672                 }
1673         }
1674         *ref_ret = (struct btrfs_extent_inline_ref *)ptr;
1675 out:
1676         if (insert) {
1677                 path->keep_locks = 0;
1678                 btrfs_unlock_up_safe(path, 1);
1679         }
1680         return err;
1681 }
1682
1683 /*
1684  * helper to add new inline back ref
1685  */
1686 static noinline_for_stack
1687 void setup_inline_extent_backref(struct btrfs_fs_info *fs_info,
1688                                  struct btrfs_path *path,
1689                                  struct btrfs_extent_inline_ref *iref,
1690                                  u64 parent, u64 root_objectid,
1691                                  u64 owner, u64 offset, int refs_to_add,
1692                                  struct btrfs_delayed_extent_op *extent_op)
1693 {
1694         struct extent_buffer *leaf;
1695         struct btrfs_extent_item *ei;
1696         unsigned long ptr;
1697         unsigned long end;
1698         unsigned long item_offset;
1699         u64 refs;
1700         int size;
1701         int type;
1702
1703         leaf = path->nodes[0];
1704         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1705         item_offset = (unsigned long)iref - (unsigned long)ei;
1706
1707         type = extent_ref_type(parent, owner);
1708         size = btrfs_extent_inline_ref_size(type);
1709
1710         btrfs_extend_item(path, size);
1711
1712         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1713         refs = btrfs_extent_refs(leaf, ei);
1714         refs += refs_to_add;
1715         btrfs_set_extent_refs(leaf, ei, refs);
1716         if (extent_op)
1717                 __run_delayed_extent_op(extent_op, leaf, ei);
1718
1719         ptr = (unsigned long)ei + item_offset;
1720         end = (unsigned long)ei + btrfs_item_size_nr(leaf, path->slots[0]);
1721         if (ptr < end - size)
1722                 memmove_extent_buffer(leaf, ptr + size, ptr,
1723                                       end - size - ptr);
1724
1725         iref = (struct btrfs_extent_inline_ref *)ptr;
1726         btrfs_set_extent_inline_ref_type(leaf, iref, type);
1727         if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1728                 struct btrfs_extent_data_ref *dref;
1729                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1730                 btrfs_set_extent_data_ref_root(leaf, dref, root_objectid);
1731                 btrfs_set_extent_data_ref_objectid(leaf, dref, owner);
1732                 btrfs_set_extent_data_ref_offset(leaf, dref, offset);
1733                 btrfs_set_extent_data_ref_count(leaf, dref, refs_to_add);
1734         } else if (type == BTRFS_SHARED_DATA_REF_KEY) {
1735                 struct btrfs_shared_data_ref *sref;
1736                 sref = (struct btrfs_shared_data_ref *)(iref + 1);
1737                 btrfs_set_shared_data_ref_count(leaf, sref, refs_to_add);
1738                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
1739         } else if (type == BTRFS_SHARED_BLOCK_REF_KEY) {
1740                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
1741         } else {
1742                 btrfs_set_extent_inline_ref_offset(leaf, iref, root_objectid);
1743         }
1744         btrfs_mark_buffer_dirty(leaf);
1745 }
1746
1747 static int lookup_extent_backref(struct btrfs_trans_handle *trans,
1748                                  struct btrfs_path *path,
1749                                  struct btrfs_extent_inline_ref **ref_ret,
1750                                  u64 bytenr, u64 num_bytes, u64 parent,
1751                                  u64 root_objectid, u64 owner, u64 offset)
1752 {
1753         int ret;
1754
1755         ret = lookup_inline_extent_backref(trans, path, ref_ret, bytenr,
1756                                            num_bytes, parent, root_objectid,
1757                                            owner, offset, 0);
1758         if (ret != -ENOENT)
1759                 return ret;
1760
1761         btrfs_release_path(path);
1762         *ref_ret = NULL;
1763
1764         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1765                 ret = lookup_tree_block_ref(trans, path, bytenr, parent,
1766                                             root_objectid);
1767         } else {
1768                 ret = lookup_extent_data_ref(trans, path, bytenr, parent,
1769                                              root_objectid, owner, offset);
1770         }
1771         return ret;
1772 }
1773
1774 /*
1775  * helper to update/remove inline back ref
1776  */
1777 static noinline_for_stack
1778 void update_inline_extent_backref(struct btrfs_path *path,
1779                                   struct btrfs_extent_inline_ref *iref,
1780                                   int refs_to_mod,
1781                                   struct btrfs_delayed_extent_op *extent_op,
1782                                   int *last_ref)
1783 {
1784         struct extent_buffer *leaf = path->nodes[0];
1785         struct btrfs_extent_item *ei;
1786         struct btrfs_extent_data_ref *dref = NULL;
1787         struct btrfs_shared_data_ref *sref = NULL;
1788         unsigned long ptr;
1789         unsigned long end;
1790         u32 item_size;
1791         int size;
1792         int type;
1793         u64 refs;
1794
1795         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1796         refs = btrfs_extent_refs(leaf, ei);
1797         WARN_ON(refs_to_mod < 0 && refs + refs_to_mod <= 0);
1798         refs += refs_to_mod;
1799         btrfs_set_extent_refs(leaf, ei, refs);
1800         if (extent_op)
1801                 __run_delayed_extent_op(extent_op, leaf, ei);
1802
1803         /*
1804          * If type is invalid, we should have bailed out after
1805          * lookup_inline_extent_backref().
1806          */
1807         type = btrfs_get_extent_inline_ref_type(leaf, iref, BTRFS_REF_TYPE_ANY);
1808         ASSERT(type != BTRFS_REF_TYPE_INVALID);
1809
1810         if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1811                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1812                 refs = btrfs_extent_data_ref_count(leaf, dref);
1813         } else if (type == BTRFS_SHARED_DATA_REF_KEY) {
1814                 sref = (struct btrfs_shared_data_ref *)(iref + 1);
1815                 refs = btrfs_shared_data_ref_count(leaf, sref);
1816         } else {
1817                 refs = 1;
1818                 BUG_ON(refs_to_mod != -1);
1819         }
1820
1821         BUG_ON(refs_to_mod < 0 && refs < -refs_to_mod);
1822         refs += refs_to_mod;
1823
1824         if (refs > 0) {
1825                 if (type == BTRFS_EXTENT_DATA_REF_KEY)
1826                         btrfs_set_extent_data_ref_count(leaf, dref, refs);
1827                 else
1828                         btrfs_set_shared_data_ref_count(leaf, sref, refs);
1829         } else {
1830                 *last_ref = 1;
1831                 size =  btrfs_extent_inline_ref_size(type);
1832                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1833                 ptr = (unsigned long)iref;
1834                 end = (unsigned long)ei + item_size;
1835                 if (ptr + size < end)
1836                         memmove_extent_buffer(leaf, ptr, ptr + size,
1837                                               end - ptr - size);
1838                 item_size -= size;
1839                 btrfs_truncate_item(path, item_size, 1);
1840         }
1841         btrfs_mark_buffer_dirty(leaf);
1842 }
1843
1844 static noinline_for_stack
1845 int insert_inline_extent_backref(struct btrfs_trans_handle *trans,
1846                                  struct btrfs_path *path,
1847                                  u64 bytenr, u64 num_bytes, u64 parent,
1848                                  u64 root_objectid, u64 owner,
1849                                  u64 offset, int refs_to_add,
1850                                  struct btrfs_delayed_extent_op *extent_op)
1851 {
1852         struct btrfs_extent_inline_ref *iref;
1853         int ret;
1854
1855         ret = lookup_inline_extent_backref(trans, path, &iref, bytenr,
1856                                            num_bytes, parent, root_objectid,
1857                                            owner, offset, 1);
1858         if (ret == 0) {
1859                 BUG_ON(owner < BTRFS_FIRST_FREE_OBJECTID);
1860                 update_inline_extent_backref(path, iref, refs_to_add,
1861                                              extent_op, NULL);
1862         } else if (ret == -ENOENT) {
1863                 setup_inline_extent_backref(trans->fs_info, path, iref, parent,
1864                                             root_objectid, owner, offset,
1865                                             refs_to_add, extent_op);
1866                 ret = 0;
1867         }
1868         return ret;
1869 }
1870
1871 static int insert_extent_backref(struct btrfs_trans_handle *trans,
1872                                  struct btrfs_path *path,
1873                                  u64 bytenr, u64 parent, u64 root_objectid,
1874                                  u64 owner, u64 offset, int refs_to_add)
1875 {
1876         int ret;
1877         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1878                 BUG_ON(refs_to_add != 1);
1879                 ret = insert_tree_block_ref(trans, path, bytenr, parent,
1880                                             root_objectid);
1881         } else {
1882                 ret = insert_extent_data_ref(trans, path, bytenr, parent,
1883                                              root_objectid, owner, offset,
1884                                              refs_to_add);
1885         }
1886         return ret;
1887 }
1888
1889 static int remove_extent_backref(struct btrfs_trans_handle *trans,
1890                                  struct btrfs_path *path,
1891                                  struct btrfs_extent_inline_ref *iref,
1892                                  int refs_to_drop, int is_data, int *last_ref)
1893 {
1894         int ret = 0;
1895
1896         BUG_ON(!is_data && refs_to_drop != 1);
1897         if (iref) {
1898                 update_inline_extent_backref(path, iref, -refs_to_drop, NULL,
1899                                              last_ref);
1900         } else if (is_data) {
1901                 ret = remove_extent_data_ref(trans, path, refs_to_drop,
1902                                              last_ref);
1903         } else {
1904                 *last_ref = 1;
1905                 ret = btrfs_del_item(trans, trans->fs_info->extent_root, path);
1906         }
1907         return ret;
1908 }
1909
1910 static int btrfs_issue_discard(struct block_device *bdev, u64 start, u64 len,
1911                                u64 *discarded_bytes)
1912 {
1913         int j, ret = 0;
1914         u64 bytes_left, end;
1915         u64 aligned_start = ALIGN(start, 1 << 9);
1916
1917         if (WARN_ON(start != aligned_start)) {
1918                 len -= aligned_start - start;
1919                 len = round_down(len, 1 << 9);
1920                 start = aligned_start;
1921         }
1922
1923         *discarded_bytes = 0;
1924
1925         if (!len)
1926                 return 0;
1927
1928         end = start + len;
1929         bytes_left = len;
1930
1931         /* Skip any superblocks on this device. */
1932         for (j = 0; j < BTRFS_SUPER_MIRROR_MAX; j++) {
1933                 u64 sb_start = btrfs_sb_offset(j);
1934                 u64 sb_end = sb_start + BTRFS_SUPER_INFO_SIZE;
1935                 u64 size = sb_start - start;
1936
1937                 if (!in_range(sb_start, start, bytes_left) &&
1938                     !in_range(sb_end, start, bytes_left) &&
1939                     !in_range(start, sb_start, BTRFS_SUPER_INFO_SIZE))
1940                         continue;
1941
1942                 /*
1943                  * Superblock spans beginning of range.  Adjust start and
1944                  * try again.
1945                  */
1946                 if (sb_start <= start) {
1947                         start += sb_end - start;
1948                         if (start > end) {
1949                                 bytes_left = 0;
1950                                 break;
1951                         }
1952                         bytes_left = end - start;
1953                         continue;
1954                 }
1955
1956                 if (size) {
1957                         ret = blkdev_issue_discard(bdev, start >> 9, size >> 9,
1958                                                    GFP_NOFS, 0);
1959                         if (!ret)
1960                                 *discarded_bytes += size;
1961                         else if (ret != -EOPNOTSUPP)
1962                                 return ret;
1963                 }
1964
1965                 start = sb_end;
1966                 if (start > end) {
1967                         bytes_left = 0;
1968                         break;
1969                 }
1970                 bytes_left = end - start;
1971         }
1972
1973         if (bytes_left) {
1974                 ret = blkdev_issue_discard(bdev, start >> 9, bytes_left >> 9,
1975                                            GFP_NOFS, 0);
1976                 if (!ret)
1977                         *discarded_bytes += bytes_left;
1978         }
1979         return ret;
1980 }
1981
1982 int btrfs_discard_extent(struct btrfs_fs_info *fs_info, u64 bytenr,
1983                          u64 num_bytes, u64 *actual_bytes)
1984 {
1985         int ret;
1986         u64 discarded_bytes = 0;
1987         struct btrfs_bio *bbio = NULL;
1988
1989
1990         /*
1991          * Avoid races with device replace and make sure our bbio has devices
1992          * associated to its stripes that don't go away while we are discarding.
1993          */
1994         btrfs_bio_counter_inc_blocked(fs_info);
1995         /* Tell the block device(s) that the sectors can be discarded */
1996         ret = btrfs_map_block(fs_info, BTRFS_MAP_DISCARD, bytenr, &num_bytes,
1997                               &bbio, 0);
1998         /* Error condition is -ENOMEM */
1999         if (!ret) {
2000                 struct btrfs_bio_stripe *stripe = bbio->stripes;
2001                 int i;
2002
2003
2004                 for (i = 0; i < bbio->num_stripes; i++, stripe++) {
2005                         u64 bytes;
2006                         struct request_queue *req_q;
2007
2008                         if (!stripe->dev->bdev) {
2009                                 ASSERT(btrfs_test_opt(fs_info, DEGRADED));
2010                                 continue;
2011                         }
2012                         req_q = bdev_get_queue(stripe->dev->bdev);
2013                         if (!blk_queue_discard(req_q))
2014                                 continue;
2015
2016                         ret = btrfs_issue_discard(stripe->dev->bdev,
2017                                                   stripe->physical,
2018                                                   stripe->length,
2019                                                   &bytes);
2020                         if (!ret)
2021                                 discarded_bytes += bytes;
2022                         else if (ret != -EOPNOTSUPP)
2023                                 break; /* Logic errors or -ENOMEM, or -EIO but I don't know how that could happen JDM */
2024
2025                         /*
2026                          * Just in case we get back EOPNOTSUPP for some reason,
2027                          * just ignore the return value so we don't screw up
2028                          * people calling discard_extent.
2029                          */
2030                         ret = 0;
2031                 }
2032                 btrfs_put_bbio(bbio);
2033         }
2034         btrfs_bio_counter_dec(fs_info);
2035
2036         if (actual_bytes)
2037                 *actual_bytes = discarded_bytes;
2038
2039
2040         if (ret == -EOPNOTSUPP)
2041                 ret = 0;
2042         return ret;
2043 }
2044
2045 /* Can return -ENOMEM */
2046 int btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
2047                          struct btrfs_ref *generic_ref)
2048 {
2049         struct btrfs_fs_info *fs_info = trans->fs_info;
2050         int old_ref_mod, new_ref_mod;
2051         int ret;
2052
2053         ASSERT(generic_ref->type != BTRFS_REF_NOT_SET &&
2054                generic_ref->action);
2055         BUG_ON(generic_ref->type == BTRFS_REF_METADATA &&
2056                generic_ref->tree_ref.root == BTRFS_TREE_LOG_OBJECTID);
2057
2058         if (generic_ref->type == BTRFS_REF_METADATA)
2059                 ret = btrfs_add_delayed_tree_ref(trans, generic_ref,
2060                                 NULL, &old_ref_mod, &new_ref_mod);
2061         else
2062                 ret = btrfs_add_delayed_data_ref(trans, generic_ref, 0,
2063                                                  &old_ref_mod, &new_ref_mod);
2064
2065         btrfs_ref_tree_mod(fs_info, generic_ref);
2066
2067         if (ret == 0 && old_ref_mod < 0 && new_ref_mod >= 0)
2068                 add_pinned_bytes(fs_info, generic_ref, -1);
2069
2070         return ret;
2071 }
2072
2073 /*
2074  * __btrfs_inc_extent_ref - insert backreference for a given extent
2075  *
2076  * @trans:          Handle of transaction
2077  *
2078  * @node:           The delayed ref node used to get the bytenr/length for
2079  *                  extent whose references are incremented.
2080  *
2081  * @parent:         If this is a shared extent (BTRFS_SHARED_DATA_REF_KEY/
2082  *                  BTRFS_SHARED_BLOCK_REF_KEY) then it holds the logical
2083  *                  bytenr of the parent block. Since new extents are always
2084  *                  created with indirect references, this will only be the case
2085  *                  when relocating a shared extent. In that case, root_objectid
2086  *                  will be BTRFS_TREE_RELOC_OBJECTID. Otheriwse, parent must
2087  *                  be 0
2088  *
2089  * @root_objectid:  The id of the root where this modification has originated,
2090  *                  this can be either one of the well-known metadata trees or
2091  *                  the subvolume id which references this extent.
2092  *
2093  * @owner:          For data extents it is the inode number of the owning file.
2094  *                  For metadata extents this parameter holds the level in the
2095  *                  tree of the extent.
2096  *
2097  * @offset:         For metadata extents the offset is ignored and is currently
2098  *                  always passed as 0. For data extents it is the fileoffset
2099  *                  this extent belongs to.
2100  *
2101  * @refs_to_add     Number of references to add
2102  *
2103  * @extent_op       Pointer to a structure, holding information necessary when
2104  *                  updating a tree block's flags
2105  *
2106  */
2107 static int __btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
2108                                   struct btrfs_delayed_ref_node *node,
2109                                   u64 parent, u64 root_objectid,
2110                                   u64 owner, u64 offset, int refs_to_add,
2111                                   struct btrfs_delayed_extent_op *extent_op)
2112 {
2113         struct btrfs_path *path;
2114         struct extent_buffer *leaf;
2115         struct btrfs_extent_item *item;
2116         struct btrfs_key key;
2117         u64 bytenr = node->bytenr;
2118         u64 num_bytes = node->num_bytes;
2119         u64 refs;
2120         int ret;
2121
2122         path = btrfs_alloc_path();
2123         if (!path)
2124                 return -ENOMEM;
2125
2126         path->reada = READA_FORWARD;
2127         path->leave_spinning = 1;
2128         /* this will setup the path even if it fails to insert the back ref */
2129         ret = insert_inline_extent_backref(trans, path, bytenr, num_bytes,
2130                                            parent, root_objectid, owner,
2131                                            offset, refs_to_add, extent_op);
2132         if ((ret < 0 && ret != -EAGAIN) || !ret)
2133                 goto out;
2134
2135         /*
2136          * Ok we had -EAGAIN which means we didn't have space to insert and
2137          * inline extent ref, so just update the reference count and add a
2138          * normal backref.
2139          */
2140         leaf = path->nodes[0];
2141         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
2142         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2143         refs = btrfs_extent_refs(leaf, item);
2144         btrfs_set_extent_refs(leaf, item, refs + refs_to_add);
2145         if (extent_op)
2146                 __run_delayed_extent_op(extent_op, leaf, item);
2147
2148         btrfs_mark_buffer_dirty(leaf);
2149         btrfs_release_path(path);
2150
2151         path->reada = READA_FORWARD;
2152         path->leave_spinning = 1;
2153         /* now insert the actual backref */
2154         ret = insert_extent_backref(trans, path, bytenr, parent, root_objectid,
2155                                     owner, offset, refs_to_add);
2156         if (ret)
2157                 btrfs_abort_transaction(trans, ret);
2158 out:
2159         btrfs_free_path(path);
2160         return ret;
2161 }
2162
2163 static int run_delayed_data_ref(struct btrfs_trans_handle *trans,
2164                                 struct btrfs_delayed_ref_node *node,
2165                                 struct btrfs_delayed_extent_op *extent_op,
2166                                 int insert_reserved)
2167 {
2168         int ret = 0;
2169         struct btrfs_delayed_data_ref *ref;
2170         struct btrfs_key ins;
2171         u64 parent = 0;
2172         u64 ref_root = 0;
2173         u64 flags = 0;
2174
2175         ins.objectid = node->bytenr;
2176         ins.offset = node->num_bytes;
2177         ins.type = BTRFS_EXTENT_ITEM_KEY;
2178
2179         ref = btrfs_delayed_node_to_data_ref(node);
2180         trace_run_delayed_data_ref(trans->fs_info, node, ref, node->action);
2181
2182         if (node->type == BTRFS_SHARED_DATA_REF_KEY)
2183                 parent = ref->parent;
2184         ref_root = ref->root;
2185
2186         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
2187                 if (extent_op)
2188                         flags |= extent_op->flags_to_set;
2189                 ret = alloc_reserved_file_extent(trans, parent, ref_root,
2190                                                  flags, ref->objectid,
2191                                                  ref->offset, &ins,
2192                                                  node->ref_mod);
2193         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
2194                 ret = __btrfs_inc_extent_ref(trans, node, parent, ref_root,
2195                                              ref->objectid, ref->offset,
2196                                              node->ref_mod, extent_op);
2197         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
2198                 ret = __btrfs_free_extent(trans, node, parent,
2199                                           ref_root, ref->objectid,
2200                                           ref->offset, node->ref_mod,
2201                                           extent_op);
2202         } else {
2203                 BUG();
2204         }
2205         return ret;
2206 }
2207
2208 static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op,
2209                                     struct extent_buffer *leaf,
2210                                     struct btrfs_extent_item *ei)
2211 {
2212         u64 flags = btrfs_extent_flags(leaf, ei);
2213         if (extent_op->update_flags) {
2214                 flags |= extent_op->flags_to_set;
2215                 btrfs_set_extent_flags(leaf, ei, flags);
2216         }
2217
2218         if (extent_op->update_key) {
2219                 struct btrfs_tree_block_info *bi;
2220                 BUG_ON(!(flags & BTRFS_EXTENT_FLAG_TREE_BLOCK));
2221                 bi = (struct btrfs_tree_block_info *)(ei + 1);
2222                 btrfs_set_tree_block_key(leaf, bi, &extent_op->key);
2223         }
2224 }
2225
2226 static int run_delayed_extent_op(struct btrfs_trans_handle *trans,
2227                                  struct btrfs_delayed_ref_head *head,
2228                                  struct btrfs_delayed_extent_op *extent_op)
2229 {
2230         struct btrfs_fs_info *fs_info = trans->fs_info;
2231         struct btrfs_key key;
2232         struct btrfs_path *path;
2233         struct btrfs_extent_item *ei;
2234         struct extent_buffer *leaf;
2235         u32 item_size;
2236         int ret;
2237         int err = 0;
2238         int metadata = !extent_op->is_data;
2239
2240         if (trans->aborted)
2241                 return 0;
2242
2243         if (metadata && !btrfs_fs_incompat(fs_info, SKINNY_METADATA))
2244                 metadata = 0;
2245
2246         path = btrfs_alloc_path();
2247         if (!path)
2248                 return -ENOMEM;
2249
2250         key.objectid = head->bytenr;
2251
2252         if (metadata) {
2253                 key.type = BTRFS_METADATA_ITEM_KEY;
2254                 key.offset = extent_op->level;
2255         } else {
2256                 key.type = BTRFS_EXTENT_ITEM_KEY;
2257                 key.offset = head->num_bytes;
2258         }
2259
2260 again:
2261         path->reada = READA_FORWARD;
2262         path->leave_spinning = 1;
2263         ret = btrfs_search_slot(trans, fs_info->extent_root, &key, path, 0, 1);
2264         if (ret < 0) {
2265                 err = ret;
2266                 goto out;
2267         }
2268         if (ret > 0) {
2269                 if (metadata) {
2270                         if (path->slots[0] > 0) {
2271                                 path->slots[0]--;
2272                                 btrfs_item_key_to_cpu(path->nodes[0], &key,
2273                                                       path->slots[0]);
2274                                 if (key.objectid == head->bytenr &&
2275                                     key.type == BTRFS_EXTENT_ITEM_KEY &&
2276                                     key.offset == head->num_bytes)
2277                                         ret = 0;
2278                         }
2279                         if (ret > 0) {
2280                                 btrfs_release_path(path);
2281                                 metadata = 0;
2282
2283                                 key.objectid = head->bytenr;
2284                                 key.offset = head->num_bytes;
2285                                 key.type = BTRFS_EXTENT_ITEM_KEY;
2286                                 goto again;
2287                         }
2288                 } else {
2289                         err = -EIO;
2290                         goto out;
2291                 }
2292         }
2293
2294         leaf = path->nodes[0];
2295         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2296
2297         if (unlikely(item_size < sizeof(*ei))) {
2298                 err = -EINVAL;
2299                 btrfs_print_v0_err(fs_info);
2300                 btrfs_abort_transaction(trans, err);
2301                 goto out;
2302         }
2303
2304         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2305         __run_delayed_extent_op(extent_op, leaf, ei);
2306
2307         btrfs_mark_buffer_dirty(leaf);
2308 out:
2309         btrfs_free_path(path);
2310         return err;
2311 }
2312
2313 static int run_delayed_tree_ref(struct btrfs_trans_handle *trans,
2314                                 struct btrfs_delayed_ref_node *node,
2315                                 struct btrfs_delayed_extent_op *extent_op,
2316                                 int insert_reserved)
2317 {
2318         int ret = 0;
2319         struct btrfs_delayed_tree_ref *ref;
2320         u64 parent = 0;
2321         u64 ref_root = 0;
2322
2323         ref = btrfs_delayed_node_to_tree_ref(node);
2324         trace_run_delayed_tree_ref(trans->fs_info, node, ref, node->action);
2325
2326         if (node->type == BTRFS_SHARED_BLOCK_REF_KEY)
2327                 parent = ref->parent;
2328         ref_root = ref->root;
2329
2330         if (node->ref_mod != 1) {
2331                 btrfs_err(trans->fs_info,
2332         "btree block(%llu) has %d references rather than 1: action %d ref_root %llu parent %llu",
2333                           node->bytenr, node->ref_mod, node->action, ref_root,
2334                           parent);
2335                 return -EIO;
2336         }
2337         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
2338                 BUG_ON(!extent_op || !extent_op->update_flags);
2339                 ret = alloc_reserved_tree_block(trans, node, extent_op);
2340         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
2341                 ret = __btrfs_inc_extent_ref(trans, node, parent, ref_root,
2342                                              ref->level, 0, 1, extent_op);
2343         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
2344                 ret = __btrfs_free_extent(trans, node, parent, ref_root,
2345                                           ref->level, 0, 1, extent_op);
2346         } else {
2347                 BUG();
2348         }
2349         return ret;
2350 }
2351
2352 /* helper function to actually process a single delayed ref entry */
2353 static int run_one_delayed_ref(struct btrfs_trans_handle *trans,
2354                                struct btrfs_delayed_ref_node *node,
2355                                struct btrfs_delayed_extent_op *extent_op,
2356                                int insert_reserved)
2357 {
2358         int ret = 0;
2359
2360         if (trans->aborted) {
2361                 if (insert_reserved)
2362                         btrfs_pin_extent(trans->fs_info, node->bytenr,
2363                                          node->num_bytes, 1);
2364                 return 0;
2365         }
2366
2367         if (node->type == BTRFS_TREE_BLOCK_REF_KEY ||
2368             node->type == BTRFS_SHARED_BLOCK_REF_KEY)
2369                 ret = run_delayed_tree_ref(trans, node, extent_op,
2370                                            insert_reserved);
2371         else if (node->type == BTRFS_EXTENT_DATA_REF_KEY ||
2372                  node->type == BTRFS_SHARED_DATA_REF_KEY)
2373                 ret = run_delayed_data_ref(trans, node, extent_op,
2374                                            insert_reserved);
2375         else
2376                 BUG();
2377         if (ret && insert_reserved)
2378                 btrfs_pin_extent(trans->fs_info, node->bytenr,
2379                                  node->num_bytes, 1);
2380         return ret;
2381 }
2382
2383 static inline struct btrfs_delayed_ref_node *
2384 select_delayed_ref(struct btrfs_delayed_ref_head *head)
2385 {
2386         struct btrfs_delayed_ref_node *ref;
2387
2388         if (RB_EMPTY_ROOT(&head->ref_tree.rb_root))
2389                 return NULL;
2390
2391         /*
2392          * Select a delayed ref of type BTRFS_ADD_DELAYED_REF first.
2393          * This is to prevent a ref count from going down to zero, which deletes
2394          * the extent item from the extent tree, when there still are references
2395          * to add, which would fail because they would not find the extent item.
2396          */
2397         if (!list_empty(&head->ref_add_list))
2398                 return list_first_entry(&head->ref_add_list,
2399                                 struct btrfs_delayed_ref_node, add_list);
2400
2401         ref = rb_entry(rb_first_cached(&head->ref_tree),
2402                        struct btrfs_delayed_ref_node, ref_node);
2403         ASSERT(list_empty(&ref->add_list));
2404         return ref;
2405 }
2406
2407 static void unselect_delayed_ref_head(struct btrfs_delayed_ref_root *delayed_refs,
2408                                       struct btrfs_delayed_ref_head *head)
2409 {
2410         spin_lock(&delayed_refs->lock);
2411         head->processing = 0;
2412         delayed_refs->num_heads_ready++;
2413         spin_unlock(&delayed_refs->lock);
2414         btrfs_delayed_ref_unlock(head);
2415 }
2416
2417 static struct btrfs_delayed_extent_op *cleanup_extent_op(
2418                                 struct btrfs_delayed_ref_head *head)
2419 {
2420         struct btrfs_delayed_extent_op *extent_op = head->extent_op;
2421
2422         if (!extent_op)
2423                 return NULL;
2424
2425         if (head->must_insert_reserved) {
2426                 head->extent_op = NULL;
2427                 btrfs_free_delayed_extent_op(extent_op);
2428                 return NULL;
2429         }
2430         return extent_op;
2431 }
2432
2433 static int run_and_cleanup_extent_op(struct btrfs_trans_handle *trans,
2434                                      struct btrfs_delayed_ref_head *head)
2435 {
2436         struct btrfs_delayed_extent_op *extent_op;
2437         int ret;
2438
2439         extent_op = cleanup_extent_op(head);
2440         if (!extent_op)
2441                 return 0;
2442         head->extent_op = NULL;
2443         spin_unlock(&head->lock);
2444         ret = run_delayed_extent_op(trans, head, extent_op);
2445         btrfs_free_delayed_extent_op(extent_op);
2446         return ret ? ret : 1;
2447 }
2448
2449 void btrfs_cleanup_ref_head_accounting(struct btrfs_fs_info *fs_info,
2450                                   struct btrfs_delayed_ref_root *delayed_refs,
2451                                   struct btrfs_delayed_ref_head *head)
2452 {
2453         int nr_items = 1;       /* Dropping this ref head update. */
2454
2455         if (head->total_ref_mod < 0) {
2456                 struct btrfs_space_info *space_info;
2457                 u64 flags;
2458
2459                 if (head->is_data)
2460                         flags = BTRFS_BLOCK_GROUP_DATA;
2461                 else if (head->is_system)
2462                         flags = BTRFS_BLOCK_GROUP_SYSTEM;
2463                 else
2464                         flags = BTRFS_BLOCK_GROUP_METADATA;
2465                 space_info = __find_space_info(fs_info, flags);
2466                 ASSERT(space_info);
2467                 percpu_counter_add_batch(&space_info->total_bytes_pinned,
2468                                    -head->num_bytes,
2469                                    BTRFS_TOTAL_BYTES_PINNED_BATCH);
2470
2471                 /*
2472                  * We had csum deletions accounted for in our delayed refs rsv,
2473                  * we need to drop the csum leaves for this update from our
2474                  * delayed_refs_rsv.
2475                  */
2476                 if (head->is_data) {
2477                         spin_lock(&delayed_refs->lock);
2478                         delayed_refs->pending_csums -= head->num_bytes;
2479                         spin_unlock(&delayed_refs->lock);
2480                         nr_items += btrfs_csum_bytes_to_leaves(fs_info,
2481                                 head->num_bytes);
2482                 }
2483         }
2484
2485         btrfs_delayed_refs_rsv_release(fs_info, nr_items);
2486 }
2487
2488 static int cleanup_ref_head(struct btrfs_trans_handle *trans,
2489                             struct btrfs_delayed_ref_head *head)
2490 {
2491
2492         struct btrfs_fs_info *fs_info = trans->fs_info;
2493         struct btrfs_delayed_ref_root *delayed_refs;
2494         int ret;
2495
2496         delayed_refs = &trans->transaction->delayed_refs;
2497
2498         ret = run_and_cleanup_extent_op(trans, head);
2499         if (ret < 0) {
2500                 unselect_delayed_ref_head(delayed_refs, head);
2501                 btrfs_debug(fs_info, "run_delayed_extent_op returned %d", ret);
2502                 return ret;
2503         } else if (ret) {
2504                 return ret;
2505         }
2506
2507         /*
2508          * Need to drop our head ref lock and re-acquire the delayed ref lock
2509          * and then re-check to make sure nobody got added.
2510          */
2511         spin_unlock(&head->lock);
2512         spin_lock(&delayed_refs->lock);
2513         spin_lock(&head->lock);
2514         if (!RB_EMPTY_ROOT(&head->ref_tree.rb_root) || head->extent_op) {
2515                 spin_unlock(&head->lock);
2516                 spin_unlock(&delayed_refs->lock);
2517                 return 1;
2518         }
2519         btrfs_delete_ref_head(delayed_refs, head);
2520         spin_unlock(&head->lock);
2521         spin_unlock(&delayed_refs->lock);
2522
2523         if (head->must_insert_reserved) {
2524                 btrfs_pin_extent(fs_info, head->bytenr,
2525                                  head->num_bytes, 1);
2526                 if (head->is_data) {
2527                         ret = btrfs_del_csums(trans, fs_info, head->bytenr,
2528                                               head->num_bytes);
2529                 }
2530         }
2531
2532         btrfs_cleanup_ref_head_accounting(fs_info, delayed_refs, head);
2533
2534         trace_run_delayed_ref_head(fs_info, head, 0);
2535         btrfs_delayed_ref_unlock(head);
2536         btrfs_put_delayed_ref_head(head);
2537         return 0;
2538 }
2539
2540 static struct btrfs_delayed_ref_head *btrfs_obtain_ref_head(
2541                                         struct btrfs_trans_handle *trans)
2542 {
2543         struct btrfs_delayed_ref_root *delayed_refs =
2544                 &trans->transaction->delayed_refs;
2545         struct btrfs_delayed_ref_head *head = NULL;
2546         int ret;
2547
2548         spin_lock(&delayed_refs->lock);
2549         head = btrfs_select_ref_head(delayed_refs);
2550         if (!head) {
2551                 spin_unlock(&delayed_refs->lock);
2552                 return head;
2553         }
2554
2555         /*
2556          * Grab the lock that says we are going to process all the refs for
2557          * this head
2558          */
2559         ret = btrfs_delayed_ref_lock(delayed_refs, head);
2560         spin_unlock(&delayed_refs->lock);
2561
2562         /*
2563          * We may have dropped the spin lock to get the head mutex lock, and
2564          * that might have given someone else time to free the head.  If that's
2565          * true, it has been removed from our list and we can move on.
2566          */
2567         if (ret == -EAGAIN)
2568                 head = ERR_PTR(-EAGAIN);
2569
2570         return head;
2571 }
2572
2573 static int btrfs_run_delayed_refs_for_head(struct btrfs_trans_handle *trans,
2574                                     struct btrfs_delayed_ref_head *locked_ref,
2575                                     unsigned long *run_refs)
2576 {
2577         struct btrfs_fs_info *fs_info = trans->fs_info;
2578         struct btrfs_delayed_ref_root *delayed_refs;
2579         struct btrfs_delayed_extent_op *extent_op;
2580         struct btrfs_delayed_ref_node *ref;
2581         int must_insert_reserved = 0;
2582         int ret;
2583
2584         delayed_refs = &trans->transaction->delayed_refs;
2585
2586         lockdep_assert_held(&locked_ref->mutex);
2587         lockdep_assert_held(&locked_ref->lock);
2588
2589         while ((ref = select_delayed_ref(locked_ref))) {
2590                 if (ref->seq &&
2591                     btrfs_check_delayed_seq(fs_info, ref->seq)) {
2592                         spin_unlock(&locked_ref->lock);
2593                         unselect_delayed_ref_head(delayed_refs, locked_ref);
2594                         return -EAGAIN;
2595                 }
2596
2597                 (*run_refs)++;
2598                 ref->in_tree = 0;
2599                 rb_erase_cached(&ref->ref_node, &locked_ref->ref_tree);
2600                 RB_CLEAR_NODE(&ref->ref_node);
2601                 if (!list_empty(&ref->add_list))
2602                         list_del(&ref->add_list);
2603                 /*
2604                  * When we play the delayed ref, also correct the ref_mod on
2605                  * head
2606                  */
2607                 switch (ref->action) {
2608                 case BTRFS_ADD_DELAYED_REF:
2609                 case BTRFS_ADD_DELAYED_EXTENT:
2610                         locked_ref->ref_mod -= ref->ref_mod;
2611                         break;
2612                 case BTRFS_DROP_DELAYED_REF:
2613                         locked_ref->ref_mod += ref->ref_mod;
2614                         break;
2615                 default:
2616                         WARN_ON(1);
2617                 }
2618                 atomic_dec(&delayed_refs->num_entries);
2619
2620                 /*
2621                  * Record the must_insert_reserved flag before we drop the
2622                  * spin lock.
2623                  */
2624                 must_insert_reserved = locked_ref->must_insert_reserved;
2625                 locked_ref->must_insert_reserved = 0;
2626
2627                 extent_op = locked_ref->extent_op;
2628                 locked_ref->extent_op = NULL;
2629                 spin_unlock(&locked_ref->lock);
2630
2631                 ret = run_one_delayed_ref(trans, ref, extent_op,
2632                                           must_insert_reserved);
2633
2634                 btrfs_free_delayed_extent_op(extent_op);
2635                 if (ret) {
2636                         unselect_delayed_ref_head(delayed_refs, locked_ref);
2637                         btrfs_put_delayed_ref(ref);
2638                         btrfs_debug(fs_info, "run_one_delayed_ref returned %d",
2639                                     ret);
2640                         return ret;
2641                 }
2642
2643                 btrfs_put_delayed_ref(ref);
2644                 cond_resched();
2645
2646                 spin_lock(&locked_ref->lock);
2647                 btrfs_merge_delayed_refs(trans, delayed_refs, locked_ref);
2648         }
2649
2650         return 0;
2651 }
2652
2653 /*
2654  * Returns 0 on success or if called with an already aborted transaction.
2655  * Returns -ENOMEM or -EIO on failure and will abort the transaction.
2656  */
2657 static noinline int __btrfs_run_delayed_refs(struct btrfs_trans_handle *trans,
2658                                              unsigned long nr)
2659 {
2660         struct btrfs_fs_info *fs_info = trans->fs_info;
2661         struct btrfs_delayed_ref_root *delayed_refs;
2662         struct btrfs_delayed_ref_head *locked_ref = NULL;
2663         ktime_t start = ktime_get();
2664         int ret;
2665         unsigned long count = 0;
2666         unsigned long actual_count = 0;
2667
2668         delayed_refs = &trans->transaction->delayed_refs;
2669         do {
2670                 if (!locked_ref) {
2671                         locked_ref = btrfs_obtain_ref_head(trans);
2672                         if (IS_ERR_OR_NULL(locked_ref)) {
2673                                 if (PTR_ERR(locked_ref) == -EAGAIN) {
2674                                         continue;
2675                                 } else {
2676                                         break;
2677                                 }
2678                         }
2679                         count++;
2680                 }
2681                 /*
2682                  * We need to try and merge add/drops of the same ref since we
2683                  * can run into issues with relocate dropping the implicit ref
2684                  * and then it being added back again before the drop can
2685                  * finish.  If we merged anything we need to re-loop so we can
2686                  * get a good ref.
2687                  * Or we can get node references of the same type that weren't
2688                  * merged when created due to bumps in the tree mod seq, and
2689                  * we need to merge them to prevent adding an inline extent
2690                  * backref before dropping it (triggering a BUG_ON at
2691                  * insert_inline_extent_backref()).
2692                  */
2693                 spin_lock(&locked_ref->lock);
2694                 btrfs_merge_delayed_refs(trans, delayed_refs, locked_ref);
2695
2696                 ret = btrfs_run_delayed_refs_for_head(trans, locked_ref,
2697                                                       &actual_count);
2698                 if (ret < 0 && ret != -EAGAIN) {
2699                         /*
2700                          * Error, btrfs_run_delayed_refs_for_head already
2701                          * unlocked everything so just bail out
2702                          */
2703                         return ret;
2704                 } else if (!ret) {
2705                         /*
2706                          * Success, perform the usual cleanup of a processed
2707                          * head
2708                          */
2709                         ret = cleanup_ref_head(trans, locked_ref);
2710                         if (ret > 0 ) {
2711                                 /* We dropped our lock, we need to loop. */
2712                                 ret = 0;
2713                                 continue;
2714                         } else if (ret) {
2715                                 return ret;
2716                         }
2717                 }
2718
2719                 /*
2720                  * Either success case or btrfs_run_delayed_refs_for_head
2721                  * returned -EAGAIN, meaning we need to select another head
2722                  */
2723
2724                 locked_ref = NULL;
2725                 cond_resched();
2726         } while ((nr != -1 && count < nr) || locked_ref);
2727
2728         /*
2729          * We don't want to include ref heads since we can have empty ref heads
2730          * and those will drastically skew our runtime down since we just do
2731          * accounting, no actual extent tree updates.
2732          */
2733         if (actual_count > 0) {
2734                 u64 runtime = ktime_to_ns(ktime_sub(ktime_get(), start));
2735                 u64 avg;
2736
2737                 /*
2738                  * We weigh the current average higher than our current runtime
2739                  * to avoid large swings in the average.
2740                  */
2741                 spin_lock(&delayed_refs->lock);
2742                 avg = fs_info->avg_delayed_ref_runtime * 3 + runtime;
2743                 fs_info->avg_delayed_ref_runtime = avg >> 2;    /* div by 4 */
2744                 spin_unlock(&delayed_refs->lock);
2745         }
2746         return 0;
2747 }
2748
2749 #ifdef SCRAMBLE_DELAYED_REFS
2750 /*
2751  * Normally delayed refs get processed in ascending bytenr order. This
2752  * correlates in most cases to the order added. To expose dependencies on this
2753  * order, we start to process the tree in the middle instead of the beginning
2754  */
2755 static u64 find_middle(struct rb_root *root)
2756 {
2757         struct rb_node *n = root->rb_node;
2758         struct btrfs_delayed_ref_node *entry;
2759         int alt = 1;
2760         u64 middle;
2761         u64 first = 0, last = 0;
2762
2763         n = rb_first(root);
2764         if (n) {
2765                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2766                 first = entry->bytenr;
2767         }
2768         n = rb_last(root);
2769         if (n) {
2770                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2771                 last = entry->bytenr;
2772         }
2773         n = root->rb_node;
2774
2775         while (n) {
2776                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2777                 WARN_ON(!entry->in_tree);
2778
2779                 middle = entry->bytenr;
2780
2781                 if (alt)
2782                         n = n->rb_left;
2783                 else
2784                         n = n->rb_right;
2785
2786                 alt = 1 - alt;
2787         }
2788         return middle;
2789 }
2790 #endif
2791
2792 static inline u64 heads_to_leaves(struct btrfs_fs_info *fs_info, u64 heads)
2793 {
2794         u64 num_bytes;
2795
2796         num_bytes = heads * (sizeof(struct btrfs_extent_item) +
2797                              sizeof(struct btrfs_extent_inline_ref));
2798         if (!btrfs_fs_incompat(fs_info, SKINNY_METADATA))
2799                 num_bytes += heads * sizeof(struct btrfs_tree_block_info);
2800
2801         /*
2802          * We don't ever fill up leaves all the way so multiply by 2 just to be
2803          * closer to what we're really going to want to use.
2804          */
2805         return div_u64(num_bytes, BTRFS_LEAF_DATA_SIZE(fs_info));
2806 }
2807
2808 /*
2809  * Takes the number of bytes to be csumm'ed and figures out how many leaves it
2810  * would require to store the csums for that many bytes.
2811  */
2812 u64 btrfs_csum_bytes_to_leaves(struct btrfs_fs_info *fs_info, u64 csum_bytes)
2813 {
2814         u64 csum_size;
2815         u64 num_csums_per_leaf;
2816         u64 num_csums;
2817
2818         csum_size = BTRFS_MAX_ITEM_SIZE(fs_info);
2819         num_csums_per_leaf = div64_u64(csum_size,
2820                         (u64)btrfs_super_csum_size(fs_info->super_copy));
2821         num_csums = div64_u64(csum_bytes, fs_info->sectorsize);
2822         num_csums += num_csums_per_leaf - 1;
2823         num_csums = div64_u64(num_csums, num_csums_per_leaf);
2824         return num_csums;
2825 }
2826
2827 bool btrfs_check_space_for_delayed_refs(struct btrfs_fs_info *fs_info)
2828 {
2829         struct btrfs_block_rsv *delayed_refs_rsv = &fs_info->delayed_refs_rsv;
2830         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
2831         bool ret = false;
2832         u64 reserved;
2833
2834         spin_lock(&global_rsv->lock);
2835         reserved = global_rsv->reserved;
2836         spin_unlock(&global_rsv->lock);
2837
2838         /*
2839          * Since the global reserve is just kind of magic we don't really want
2840          * to rely on it to save our bacon, so if our size is more than the
2841          * delayed_refs_rsv and the global rsv then it's time to think about
2842          * bailing.
2843          */
2844         spin_lock(&delayed_refs_rsv->lock);
2845         reserved += delayed_refs_rsv->reserved;
2846         if (delayed_refs_rsv->size >= reserved)
2847                 ret = true;
2848         spin_unlock(&delayed_refs_rsv->lock);
2849         return ret;
2850 }
2851
2852 int btrfs_should_throttle_delayed_refs(struct btrfs_trans_handle *trans)
2853 {
2854         u64 num_entries =
2855                 atomic_read(&trans->transaction->delayed_refs.num_entries);
2856         u64 avg_runtime;
2857         u64 val;
2858
2859         smp_mb();
2860         avg_runtime = trans->fs_info->avg_delayed_ref_runtime;
2861         val = num_entries * avg_runtime;
2862         if (val >= NSEC_PER_SEC)
2863                 return 1;
2864         if (val >= NSEC_PER_SEC / 2)
2865                 return 2;
2866
2867         return btrfs_check_space_for_delayed_refs(trans->fs_info);
2868 }
2869
2870 /*
2871  * this starts processing the delayed reference count updates and
2872  * extent insertions we have queued up so far.  count can be
2873  * 0, which means to process everything in the tree at the start
2874  * of the run (but not newly added entries), or it can be some target
2875  * number you'd like to process.
2876  *
2877  * Returns 0 on success or if called with an aborted transaction
2878  * Returns <0 on error and aborts the transaction
2879  */
2880 int btrfs_run_delayed_refs(struct btrfs_trans_handle *trans,
2881                            unsigned long count)
2882 {
2883         struct btrfs_fs_info *fs_info = trans->fs_info;
2884         struct rb_node *node;
2885         struct btrfs_delayed_ref_root *delayed_refs;
2886         struct btrfs_delayed_ref_head *head;
2887         int ret;
2888         int run_all = count == (unsigned long)-1;
2889
2890         /* We'll clean this up in btrfs_cleanup_transaction */
2891         if (trans->aborted)
2892                 return 0;
2893
2894         if (test_bit(BTRFS_FS_CREATING_FREE_SPACE_TREE, &fs_info->flags))
2895                 return 0;
2896
2897         delayed_refs = &trans->transaction->delayed_refs;
2898         if (count == 0)
2899                 count = atomic_read(&delayed_refs->num_entries) * 2;
2900
2901 again:
2902 #ifdef SCRAMBLE_DELAYED_REFS
2903         delayed_refs->run_delayed_start = find_middle(&delayed_refs->root);
2904 #endif
2905         ret = __btrfs_run_delayed_refs(trans, count);
2906         if (ret < 0) {
2907                 btrfs_abort_transaction(trans, ret);
2908                 return ret;
2909         }
2910
2911         if (run_all) {
2912                 btrfs_create_pending_block_groups(trans);
2913
2914                 spin_lock(&delayed_refs->lock);
2915                 node = rb_first_cached(&delayed_refs->href_root);
2916                 if (!node) {
2917                         spin_unlock(&delayed_refs->lock);
2918                         goto out;
2919                 }
2920                 head = rb_entry(node, struct btrfs_delayed_ref_head,
2921                                 href_node);
2922                 refcount_inc(&head->refs);
2923                 spin_unlock(&delayed_refs->lock);
2924
2925                 /* Mutex was contended, block until it's released and retry. */
2926                 mutex_lock(&head->mutex);
2927                 mutex_unlock(&head->mutex);
2928
2929                 btrfs_put_delayed_ref_head(head);
2930                 cond_resched();
2931                 goto again;
2932         }
2933 out:
2934         return 0;
2935 }
2936
2937 int btrfs_set_disk_extent_flags(struct btrfs_trans_handle *trans,
2938                                 u64 bytenr, u64 num_bytes, u64 flags,
2939                                 int level, int is_data)
2940 {
2941         struct btrfs_delayed_extent_op *extent_op;
2942         int ret;
2943
2944         extent_op = btrfs_alloc_delayed_extent_op();
2945         if (!extent_op)
2946                 return -ENOMEM;
2947
2948         extent_op->flags_to_set = flags;
2949         extent_op->update_flags = true;
2950         extent_op->update_key = false;
2951         extent_op->is_data = is_data ? true : false;
2952         extent_op->level = level;
2953
2954         ret = btrfs_add_delayed_extent_op(trans, bytenr, num_bytes, extent_op);
2955         if (ret)
2956                 btrfs_free_delayed_extent_op(extent_op);
2957         return ret;
2958 }
2959
2960 static noinline int check_delayed_ref(struct btrfs_root *root,
2961                                       struct btrfs_path *path,
2962                                       u64 objectid, u64 offset, u64 bytenr)
2963 {
2964         struct btrfs_delayed_ref_head *head;
2965         struct btrfs_delayed_ref_node *ref;
2966         struct btrfs_delayed_data_ref *data_ref;
2967         struct btrfs_delayed_ref_root *delayed_refs;
2968         struct btrfs_transaction *cur_trans;
2969         struct rb_node *node;
2970         int ret = 0;
2971
2972         spin_lock(&root->fs_info->trans_lock);
2973         cur_trans = root->fs_info->running_transaction;
2974         if (cur_trans)
2975                 refcount_inc(&cur_trans->use_count);
2976         spin_unlock(&root->fs_info->trans_lock);
2977         if (!cur_trans)
2978                 return 0;
2979
2980         delayed_refs = &cur_trans->delayed_refs;
2981         spin_lock(&delayed_refs->lock);
2982         head = btrfs_find_delayed_ref_head(delayed_refs, bytenr);
2983         if (!head) {
2984                 spin_unlock(&delayed_refs->lock);
2985                 btrfs_put_transaction(cur_trans);
2986                 return 0;
2987         }
2988
2989         if (!mutex_trylock(&head->mutex)) {
2990                 refcount_inc(&head->refs);
2991                 spin_unlock(&delayed_refs->lock);
2992
2993                 btrfs_release_path(path);
2994
2995                 /*
2996                  * Mutex was contended, block until it's released and let
2997                  * caller try again
2998                  */
2999                 mutex_lock(&head->mutex);
3000                 mutex_unlock(&head->mutex);
3001                 btrfs_put_delayed_ref_head(head);
3002                 btrfs_put_transaction(cur_trans);
3003                 return -EAGAIN;
3004         }
3005         spin_unlock(&delayed_refs->lock);
3006
3007         spin_lock(&head->lock);
3008         /*
3009          * XXX: We should replace this with a proper search function in the
3010          * future.
3011          */
3012         for (node = rb_first_cached(&head->ref_tree); node;
3013              node = rb_next(node)) {
3014                 ref = rb_entry(node, struct btrfs_delayed_ref_node, ref_node);
3015                 /* If it's a shared ref we know a cross reference exists */
3016                 if (ref->type != BTRFS_EXTENT_DATA_REF_KEY) {
3017                         ret = 1;
3018                         break;
3019                 }
3020
3021                 data_ref = btrfs_delayed_node_to_data_ref(ref);
3022
3023                 /*
3024                  * If our ref doesn't match the one we're currently looking at
3025                  * then we have a cross reference.
3026                  */
3027                 if (data_ref->root != root->root_key.objectid ||
3028                     data_ref->objectid != objectid ||
3029                     data_ref->offset != offset) {
3030                         ret = 1;
3031                         break;
3032                 }
3033         }
3034         spin_unlock(&head->lock);
3035         mutex_unlock(&head->mutex);
3036         btrfs_put_transaction(cur_trans);
3037         return ret;
3038 }
3039
3040 static noinline int check_committed_ref(struct btrfs_root *root,
3041                                         struct btrfs_path *path,
3042                                         u64 objectid, u64 offset, u64 bytenr)
3043 {
3044         struct btrfs_fs_info *fs_info = root->fs_info;
3045         struct btrfs_root *extent_root = fs_info->extent_root;
3046         struct extent_buffer *leaf;
3047         struct btrfs_extent_data_ref *ref;
3048         struct btrfs_extent_inline_ref *iref;
3049         struct btrfs_extent_item *ei;
3050         struct btrfs_key key;
3051         u32 item_size;
3052         int type;
3053         int ret;
3054
3055         key.objectid = bytenr;
3056         key.offset = (u64)-1;
3057         key.type = BTRFS_EXTENT_ITEM_KEY;
3058
3059         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
3060         if (ret < 0)
3061                 goto out;
3062         BUG_ON(ret == 0); /* Corruption */
3063
3064         ret = -ENOENT;
3065         if (path->slots[0] == 0)
3066                 goto out;
3067
3068         path->slots[0]--;
3069         leaf = path->nodes[0];
3070         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
3071
3072         if (key.objectid != bytenr || key.type != BTRFS_EXTENT_ITEM_KEY)
3073                 goto out;
3074
3075         ret = 1;
3076         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
3077         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
3078
3079         if (item_size != sizeof(*ei) +
3080             btrfs_extent_inline_ref_size(BTRFS_EXTENT_DATA_REF_KEY))
3081                 goto out;
3082
3083         if (btrfs_extent_generation(leaf, ei) <=
3084             btrfs_root_last_snapshot(&root->root_item))
3085                 goto out;
3086
3087         iref = (struct btrfs_extent_inline_ref *)(ei + 1);
3088
3089         type = btrfs_get_extent_inline_ref_type(leaf, iref, BTRFS_REF_TYPE_DATA);
3090         if (type != BTRFS_EXTENT_DATA_REF_KEY)
3091                 goto out;
3092
3093         ref = (struct btrfs_extent_data_ref *)(&iref->offset);
3094         if (btrfs_extent_refs(leaf, ei) !=
3095             btrfs_extent_data_ref_count(leaf, ref) ||
3096             btrfs_extent_data_ref_root(leaf, ref) !=
3097             root->root_key.objectid ||
3098             btrfs_extent_data_ref_objectid(leaf, ref) != objectid ||
3099             btrfs_extent_data_ref_offset(leaf, ref) != offset)
3100                 goto out;
3101
3102         ret = 0;
3103 out:
3104         return ret;
3105 }
3106
3107 int btrfs_cross_ref_exist(struct btrfs_root *root, u64 objectid, u64 offset,
3108                           u64 bytenr)
3109 {
3110         struct btrfs_path *path;
3111         int ret;
3112
3113         path = btrfs_alloc_path();
3114         if (!path)
3115                 return -ENOMEM;
3116
3117         do {
3118                 ret = check_committed_ref(root, path, objectid,
3119                                           offset, bytenr);
3120                 if (ret && ret != -ENOENT)
3121                         goto out;
3122
3123                 ret = check_delayed_ref(root, path, objectid, offset, bytenr);
3124         } while (ret == -EAGAIN);
3125
3126 out:
3127         btrfs_free_path(path);
3128         if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID)
3129                 WARN_ON(ret > 0);
3130         return ret;
3131 }
3132
3133 static int __btrfs_mod_ref(struct btrfs_trans_handle *trans,
3134                            struct btrfs_root *root,
3135                            struct extent_buffer *buf,
3136                            int full_backref, int inc)
3137 {
3138         struct btrfs_fs_info *fs_info = root->fs_info;
3139         u64 bytenr;
3140         u64 num_bytes;
3141         u64 parent;
3142         u64 ref_root;
3143         u32 nritems;
3144         struct btrfs_key key;
3145         struct btrfs_file_extent_item *fi;
3146         struct btrfs_ref generic_ref = { 0 };
3147         bool for_reloc = btrfs_header_flag(buf, BTRFS_HEADER_FLAG_RELOC);
3148         int i;
3149         int action;
3150         int level;
3151         int ret = 0;
3152
3153         if (btrfs_is_testing(fs_info))
3154                 return 0;
3155
3156         ref_root = btrfs_header_owner(buf);
3157         nritems = btrfs_header_nritems(buf);
3158         level = btrfs_header_level(buf);
3159
3160         if (!test_bit(BTRFS_ROOT_REF_COWS, &root->state) && level == 0)
3161                 return 0;
3162
3163         if (full_backref)
3164                 parent = buf->start;
3165         else
3166                 parent = 0;
3167         if (inc)
3168                 action = BTRFS_ADD_DELAYED_REF;
3169         else
3170                 action = BTRFS_DROP_DELAYED_REF;
3171
3172         for (i = 0; i < nritems; i++) {
3173                 if (level == 0) {
3174                         btrfs_item_key_to_cpu(buf, &key, i);
3175                         if (key.type != BTRFS_EXTENT_DATA_KEY)
3176                                 continue;
3177                         fi = btrfs_item_ptr(buf, i,
3178                                             struct btrfs_file_extent_item);
3179                         if (btrfs_file_extent_type(buf, fi) ==
3180                             BTRFS_FILE_EXTENT_INLINE)
3181                                 continue;
3182                         bytenr = btrfs_file_extent_disk_bytenr(buf, fi);
3183                         if (bytenr == 0)
3184                                 continue;
3185
3186                         num_bytes = btrfs_file_extent_disk_num_bytes(buf, fi);
3187                         key.offset -= btrfs_file_extent_offset(buf, fi);
3188                         btrfs_init_generic_ref(&generic_ref, action, bytenr,
3189                                                num_bytes, parent);
3190                         generic_ref.real_root = root->root_key.objectid;
3191                         btrfs_init_data_ref(&generic_ref, ref_root, key.objectid,
3192                                             key.offset);
3193                         generic_ref.skip_qgroup = for_reloc;
3194                         if (inc)
3195                                 ret = btrfs_inc_extent_ref(trans, &generic_ref);
3196                         else
3197                                 ret = btrfs_free_extent(trans, &generic_ref);
3198                         if (ret)
3199                                 goto fail;
3200                 } else {
3201                         bytenr = btrfs_node_blockptr(buf, i);
3202                         num_bytes = fs_info->nodesize;
3203                         btrfs_init_generic_ref(&generic_ref, action, bytenr,
3204                                                num_bytes, parent);
3205                         generic_ref.real_root = root->root_key.objectid;
3206                         btrfs_init_tree_ref(&generic_ref, level - 1, ref_root);
3207                         generic_ref.skip_qgroup = for_reloc;
3208                         if (inc)
3209                                 ret = btrfs_inc_extent_ref(trans, &generic_ref);
3210                         else
3211                                 ret = btrfs_free_extent(trans, &generic_ref);
3212                         if (ret)
3213                                 goto fail;
3214                 }
3215         }
3216         return 0;
3217 fail:
3218         return ret;
3219 }
3220
3221 int btrfs_inc_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
3222                   struct extent_buffer *buf, int full_backref)
3223 {
3224         return __btrfs_mod_ref(trans, root, buf, full_backref, 1);
3225 }
3226
3227 int btrfs_dec_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
3228                   struct extent_buffer *buf, int full_backref)
3229 {
3230         return __btrfs_mod_ref(trans, root, buf, full_backref, 0);
3231 }
3232
3233 static int write_one_cache_group(struct btrfs_trans_handle *trans,
3234                                  struct btrfs_path *path,
3235                                  struct btrfs_block_group_cache *cache)
3236 {
3237         struct btrfs_fs_info *fs_info = trans->fs_info;
3238         int ret;
3239         struct btrfs_root *extent_root = fs_info->extent_root;
3240         unsigned long bi;
3241         struct extent_buffer *leaf;
3242
3243         ret = btrfs_search_slot(trans, extent_root, &cache->key, path, 0, 1);
3244         if (ret) {
3245                 if (ret > 0)
3246                         ret = -ENOENT;
3247                 goto fail;
3248         }
3249
3250         leaf = path->nodes[0];
3251         bi = btrfs_item_ptr_offset(leaf, path->slots[0]);
3252         write_extent_buffer(leaf, &cache->item, bi, sizeof(cache->item));
3253         btrfs_mark_buffer_dirty(leaf);
3254 fail:
3255         btrfs_release_path(path);
3256         return ret;
3257
3258 }
3259
3260 static struct btrfs_block_group_cache *next_block_group(
3261                 struct btrfs_block_group_cache *cache)
3262 {
3263         struct btrfs_fs_info *fs_info = cache->fs_info;
3264         struct rb_node *node;
3265
3266         spin_lock(&fs_info->block_group_cache_lock);
3267
3268         /* If our block group was removed, we need a full search. */
3269         if (RB_EMPTY_NODE(&cache->cache_node)) {
3270                 const u64 next_bytenr = cache->key.objectid + cache->key.offset;
3271
3272                 spin_unlock(&fs_info->block_group_cache_lock);
3273                 btrfs_put_block_group(cache);
3274                 cache = btrfs_lookup_first_block_group(fs_info, next_bytenr); return cache;
3275         }
3276         node = rb_next(&cache->cache_node);
3277         btrfs_put_block_group(cache);
3278         if (node) {
3279                 cache = rb_entry(node, struct btrfs_block_group_cache,
3280                                  cache_node);
3281                 btrfs_get_block_group(cache);
3282         } else
3283                 cache = NULL;
3284         spin_unlock(&fs_info->block_group_cache_lock);
3285         return cache;
3286 }
3287
3288 static int cache_save_setup(struct btrfs_block_group_cache *block_group,
3289                             struct btrfs_trans_handle *trans,
3290                             struct btrfs_path *path)
3291 {
3292         struct btrfs_fs_info *fs_info = block_group->fs_info;
3293         struct btrfs_root *root = fs_info->tree_root;
3294         struct inode *inode = NULL;
3295         struct extent_changeset *data_reserved = NULL;
3296         u64 alloc_hint = 0;
3297         int dcs = BTRFS_DC_ERROR;
3298         u64 num_pages = 0;
3299         int retries = 0;
3300         int ret = 0;
3301
3302         /*
3303          * If this block group is smaller than 100 megs don't bother caching the
3304          * block group.
3305          */
3306         if (block_group->key.offset < (100 * SZ_1M)) {
3307                 spin_lock(&block_group->lock);
3308                 block_group->disk_cache_state = BTRFS_DC_WRITTEN;
3309                 spin_unlock(&block_group->lock);
3310                 return 0;
3311         }
3312
3313         if (trans->aborted)
3314                 return 0;
3315 again:
3316         inode = lookup_free_space_inode(block_group, path);
3317         if (IS_ERR(inode) && PTR_ERR(inode) != -ENOENT) {
3318                 ret = PTR_ERR(inode);
3319                 btrfs_release_path(path);
3320                 goto out;
3321         }
3322
3323         if (IS_ERR(inode)) {
3324                 BUG_ON(retries);
3325                 retries++;
3326
3327                 if (block_group->ro)
3328                         goto out_free;
3329
3330                 ret = create_free_space_inode(trans, block_group, path);
3331                 if (ret)
3332                         goto out_free;
3333                 goto again;
3334         }
3335
3336         /*
3337          * We want to set the generation to 0, that way if anything goes wrong
3338          * from here on out we know not to trust this cache when we load up next
3339          * time.
3340          */
3341         BTRFS_I(inode)->generation = 0;
3342         ret = btrfs_update_inode(trans, root, inode);
3343         if (ret) {
3344                 /*
3345                  * So theoretically we could recover from this, simply set the
3346                  * super cache generation to 0 so we know to invalidate the
3347                  * cache, but then we'd have to keep track of the block groups
3348                  * that fail this way so we know we _have_ to reset this cache
3349                  * before the next commit or risk reading stale cache.  So to
3350                  * limit our exposure to horrible edge cases lets just abort the
3351                  * transaction, this only happens in really bad situations
3352                  * anyway.
3353                  */
3354                 btrfs_abort_transaction(trans, ret);
3355                 goto out_put;
3356         }
3357         WARN_ON(ret);
3358
3359         /* We've already setup this transaction, go ahead and exit */
3360         if (block_group->cache_generation == trans->transid &&
3361             i_size_read(inode)) {
3362                 dcs = BTRFS_DC_SETUP;
3363                 goto out_put;
3364         }
3365
3366         if (i_size_read(inode) > 0) {
3367                 ret = btrfs_check_trunc_cache_free_space(fs_info,
3368                                         &fs_info->global_block_rsv);
3369                 if (ret)
3370                         goto out_put;
3371
3372                 ret = btrfs_truncate_free_space_cache(trans, NULL, inode);
3373                 if (ret)
3374                         goto out_put;
3375         }
3376
3377         spin_lock(&block_group->lock);
3378         if (block_group->cached != BTRFS_CACHE_FINISHED ||
3379             !btrfs_test_opt(fs_info, SPACE_CACHE)) {
3380                 /*
3381                  * don't bother trying to write stuff out _if_
3382                  * a) we're not cached,
3383                  * b) we're with nospace_cache mount option,
3384                  * c) we're with v2 space_cache (FREE_SPACE_TREE).
3385                  */
3386                 dcs = BTRFS_DC_WRITTEN;
3387                 spin_unlock(&block_group->lock);
3388                 goto out_put;
3389         }
3390         spin_unlock(&block_group->lock);
3391
3392         /*
3393          * We hit an ENOSPC when setting up the cache in this transaction, just
3394          * skip doing the setup, we've already cleared the cache so we're safe.
3395          */
3396         if (test_bit(BTRFS_TRANS_CACHE_ENOSPC, &trans->transaction->flags)) {
3397                 ret = -ENOSPC;
3398                 goto out_put;
3399         }
3400
3401         /*
3402          * Try to preallocate enough space based on how big the block group is.
3403          * Keep in mind this has to include any pinned space which could end up
3404          * taking up quite a bit since it's not folded into the other space
3405          * cache.
3406          */
3407         num_pages = div_u64(block_group->key.offset, SZ_256M);
3408         if (!num_pages)
3409                 num_pages = 1;
3410
3411         num_pages *= 16;
3412         num_pages *= PAGE_SIZE;
3413
3414         ret = btrfs_check_data_free_space(inode, &data_reserved, 0, num_pages);
3415         if (ret)
3416                 goto out_put;
3417
3418         ret = btrfs_prealloc_file_range_trans(inode, trans, 0, 0, num_pages,
3419                                               num_pages, num_pages,
3420                                               &alloc_hint);
3421         /*
3422          * Our cache requires contiguous chunks so that we don't modify a bunch
3423          * of metadata or split extents when writing the cache out, which means
3424          * we can enospc if we are heavily fragmented in addition to just normal
3425          * out of space conditions.  So if we hit this just skip setting up any
3426          * other block groups for this transaction, maybe we'll unpin enough
3427          * space the next time around.
3428          */
3429         if (!ret)
3430                 dcs = BTRFS_DC_SETUP;
3431         else if (ret == -ENOSPC)
3432                 set_bit(BTRFS_TRANS_CACHE_ENOSPC, &trans->transaction->flags);
3433
3434 out_put:
3435         iput(inode);
3436 out_free:
3437         btrfs_release_path(path);
3438 out:
3439         spin_lock(&block_group->lock);
3440         if (!ret && dcs == BTRFS_DC_SETUP)
3441                 block_group->cache_generation = trans->transid;
3442         block_group->disk_cache_state = dcs;
3443         spin_unlock(&block_group->lock);
3444
3445         extent_changeset_free(data_reserved);
3446         return ret;
3447 }
3448
3449 int btrfs_setup_space_cache(struct btrfs_trans_handle *trans)
3450 {
3451         struct btrfs_fs_info *fs_info = trans->fs_info;
3452         struct btrfs_block_group_cache *cache, *tmp;
3453         struct btrfs_transaction *cur_trans = trans->transaction;
3454         struct btrfs_path *path;
3455
3456         if (list_empty(&cur_trans->dirty_bgs) ||
3457             !btrfs_test_opt(fs_info, SPACE_CACHE))
3458                 return 0;
3459
3460         path = btrfs_alloc_path();
3461         if (!path)
3462                 return -ENOMEM;
3463
3464         /* Could add new block groups, use _safe just in case */
3465         list_for_each_entry_safe(cache, tmp, &cur_trans->dirty_bgs,
3466                                  dirty_list) {
3467                 if (cache->disk_cache_state == BTRFS_DC_CLEAR)
3468                         cache_save_setup(cache, trans, path);
3469         }
3470
3471         btrfs_free_path(path);
3472         return 0;
3473 }
3474
3475 /*
3476  * transaction commit does final block group cache writeback during a
3477  * critical section where nothing is allowed to change the FS.  This is
3478  * required in order for the cache to actually match the block group,
3479  * but can introduce a lot of latency into the commit.
3480  *
3481  * So, btrfs_start_dirty_block_groups is here to kick off block group
3482  * cache IO.  There's a chance we'll have to redo some of it if the
3483  * block group changes again during the commit, but it greatly reduces
3484  * the commit latency by getting rid of the easy block groups while
3485  * we're still allowing others to join the commit.
3486  */
3487 int btrfs_start_dirty_block_groups(struct btrfs_trans_handle *trans)
3488 {
3489         struct btrfs_fs_info *fs_info = trans->fs_info;
3490         struct btrfs_block_group_cache *cache;
3491         struct btrfs_transaction *cur_trans = trans->transaction;
3492         int ret = 0;
3493         int should_put;
3494         struct btrfs_path *path = NULL;
3495         LIST_HEAD(dirty);
3496         struct list_head *io = &cur_trans->io_bgs;
3497         int num_started = 0;
3498         int loops = 0;
3499
3500         spin_lock(&cur_trans->dirty_bgs_lock);
3501         if (list_empty(&cur_trans->dirty_bgs)) {
3502                 spin_unlock(&cur_trans->dirty_bgs_lock);
3503                 return 0;
3504         }
3505         list_splice_init(&cur_trans->dirty_bgs, &dirty);
3506         spin_unlock(&cur_trans->dirty_bgs_lock);
3507
3508 again:
3509         /*
3510          * make sure all the block groups on our dirty list actually
3511          * exist
3512          */
3513         btrfs_create_pending_block_groups(trans);
3514
3515         if (!path) {
3516                 path = btrfs_alloc_path();
3517                 if (!path)
3518                         return -ENOMEM;
3519         }
3520
3521         /*
3522          * cache_write_mutex is here only to save us from balance or automatic
3523          * removal of empty block groups deleting this block group while we are
3524          * writing out the cache
3525          */
3526         mutex_lock(&trans->transaction->cache_write_mutex);
3527         while (!list_empty(&dirty)) {
3528                 bool drop_reserve = true;
3529
3530                 cache = list_first_entry(&dirty,
3531                                          struct btrfs_block_group_cache,
3532                                          dirty_list);
3533                 /*
3534                  * this can happen if something re-dirties a block
3535                  * group that is already under IO.  Just wait for it to
3536                  * finish and then do it all again
3537                  */
3538                 if (!list_empty(&cache->io_list)) {
3539                         list_del_init(&cache->io_list);
3540                         btrfs_wait_cache_io(trans, cache, path);
3541                         btrfs_put_block_group(cache);
3542                 }
3543
3544
3545                 /*
3546                  * btrfs_wait_cache_io uses the cache->dirty_list to decide
3547                  * if it should update the cache_state.  Don't delete
3548                  * until after we wait.
3549                  *
3550                  * Since we're not running in the commit critical section
3551                  * we need the dirty_bgs_lock to protect from update_block_group
3552                  */
3553                 spin_lock(&cur_trans->dirty_bgs_lock);
3554                 list_del_init(&cache->dirty_list);
3555                 spin_unlock(&cur_trans->dirty_bgs_lock);
3556
3557                 should_put = 1;
3558
3559                 cache_save_setup(cache, trans, path);
3560
3561                 if (cache->disk_cache_state == BTRFS_DC_SETUP) {
3562                         cache->io_ctl.inode = NULL;
3563                         ret = btrfs_write_out_cache(trans, cache, path);
3564                         if (ret == 0 && cache->io_ctl.inode) {
3565                                 num_started++;
3566                                 should_put = 0;
3567
3568                                 /*
3569                                  * The cache_write_mutex is protecting the
3570                                  * io_list, also refer to the definition of
3571                                  * btrfs_transaction::io_bgs for more details
3572                                  */
3573                                 list_add_tail(&cache->io_list, io);
3574                         } else {
3575                                 /*
3576                                  * if we failed to write the cache, the
3577                                  * generation will be bad and life goes on
3578                                  */
3579                                 ret = 0;
3580                         }
3581                 }
3582                 if (!ret) {
3583                         ret = write_one_cache_group(trans, path, cache);
3584                         /*
3585                          * Our block group might still be attached to the list
3586                          * of new block groups in the transaction handle of some
3587                          * other task (struct btrfs_trans_handle->new_bgs). This
3588                          * means its block group item isn't yet in the extent
3589                          * tree. If this happens ignore the error, as we will
3590                          * try again later in the critical section of the
3591                          * transaction commit.
3592                          */
3593                         if (ret == -ENOENT) {
3594                                 ret = 0;
3595                                 spin_lock(&cur_trans->dirty_bgs_lock);
3596                                 if (list_empty(&cache->dirty_list)) {
3597                                         list_add_tail(&cache->dirty_list,
3598                                                       &cur_trans->dirty_bgs);
3599                                         btrfs_get_block_group(cache);
3600                                         drop_reserve = false;
3601                                 }
3602                                 spin_unlock(&cur_trans->dirty_bgs_lock);
3603                         } else if (ret) {
3604                                 btrfs_abort_transaction(trans, ret);
3605                         }
3606                 }
3607
3608                 /* if it's not on the io list, we need to put the block group */
3609                 if (should_put)
3610                         btrfs_put_block_group(cache);
3611                 if (drop_reserve)
3612                         btrfs_delayed_refs_rsv_release(fs_info, 1);
3613
3614                 if (ret)
3615                         break;
3616
3617                 /*
3618                  * Avoid blocking other tasks for too long. It might even save
3619                  * us from writing caches for block groups that are going to be
3620                  * removed.
3621                  */
3622                 mutex_unlock(&trans->transaction->cache_write_mutex);
3623                 mutex_lock(&trans->transaction->cache_write_mutex);
3624         }
3625         mutex_unlock(&trans->transaction->cache_write_mutex);
3626
3627         /*
3628          * go through delayed refs for all the stuff we've just kicked off
3629          * and then loop back (just once)
3630          */
3631         ret = btrfs_run_delayed_refs(trans, 0);
3632         if (!ret && loops == 0) {
3633                 loops++;
3634                 spin_lock(&cur_trans->dirty_bgs_lock);
3635                 list_splice_init(&cur_trans->dirty_bgs, &dirty);
3636                 /*
3637                  * dirty_bgs_lock protects us from concurrent block group
3638                  * deletes too (not just cache_write_mutex).
3639                  */
3640                 if (!list_empty(&dirty)) {
3641                         spin_unlock(&cur_trans->dirty_bgs_lock);
3642                         goto again;
3643                 }
3644                 spin_unlock(&cur_trans->dirty_bgs_lock);
3645         } else if (ret < 0) {
3646                 btrfs_cleanup_dirty_bgs(cur_trans, fs_info);
3647         }
3648
3649         btrfs_free_path(path);
3650         return ret;
3651 }
3652
3653 int btrfs_write_dirty_block_groups(struct btrfs_trans_handle *trans)
3654 {
3655         struct btrfs_fs_info *fs_info = trans->fs_info;
3656         struct btrfs_block_group_cache *cache;
3657         struct btrfs_transaction *cur_trans = trans->transaction;
3658         int ret = 0;
3659         int should_put;
3660         struct btrfs_path *path;
3661         struct list_head *io = &cur_trans->io_bgs;
3662         int num_started = 0;
3663
3664         path = btrfs_alloc_path();
3665         if (!path)
3666                 return -ENOMEM;
3667
3668         /*
3669          * Even though we are in the critical section of the transaction commit,
3670          * we can still have concurrent tasks adding elements to this
3671          * transaction's list of dirty block groups. These tasks correspond to
3672          * endio free space workers started when writeback finishes for a
3673          * space cache, which run inode.c:btrfs_finish_ordered_io(), and can
3674          * allocate new block groups as a result of COWing nodes of the root
3675          * tree when updating the free space inode. The writeback for the space
3676          * caches is triggered by an earlier call to
3677          * btrfs_start_dirty_block_groups() and iterations of the following
3678          * loop.
3679          * Also we want to do the cache_save_setup first and then run the
3680          * delayed refs to make sure we have the best chance at doing this all
3681          * in one shot.
3682          */
3683         spin_lock(&cur_trans->dirty_bgs_lock);
3684         while (!list_empty(&cur_trans->dirty_bgs)) {
3685                 cache = list_first_entry(&cur_trans->dirty_bgs,
3686                                          struct btrfs_block_group_cache,
3687                                          dirty_list);
3688
3689                 /*
3690                  * this can happen if cache_save_setup re-dirties a block
3691                  * group that is already under IO.  Just wait for it to
3692                  * finish and then do it all again
3693                  */
3694                 if (!list_empty(&cache->io_list)) {
3695                         spin_unlock(&cur_trans->dirty_bgs_lock);
3696                         list_del_init(&cache->io_list);
3697                         btrfs_wait_cache_io(trans, cache, path);
3698                         btrfs_put_block_group(cache);
3699                         spin_lock(&cur_trans->dirty_bgs_lock);
3700                 }
3701
3702                 /*
3703                  * don't remove from the dirty list until after we've waited
3704                  * on any pending IO
3705                  */
3706                 list_del_init(&cache->dirty_list);
3707                 spin_unlock(&cur_trans->dirty_bgs_lock);
3708                 should_put = 1;
3709
3710                 cache_save_setup(cache, trans, path);
3711
3712                 if (!ret)
3713                         ret = btrfs_run_delayed_refs(trans,
3714                                                      (unsigned long) -1);
3715
3716                 if (!ret && cache->disk_cache_state == BTRFS_DC_SETUP) {
3717                         cache->io_ctl.inode = NULL;
3718                         ret = btrfs_write_out_cache(trans, cache, path);
3719                         if (ret == 0 && cache->io_ctl.inode) {
3720                                 num_started++;
3721                                 should_put = 0;
3722                                 list_add_tail(&cache->io_list, io);
3723                         } else {
3724                                 /*
3725                                  * if we failed to write the cache, the
3726                                  * generation will be bad and life goes on
3727                                  */
3728                                 ret = 0;
3729                         }
3730                 }
3731                 if (!ret) {
3732                         ret = write_one_cache_group(trans, path, cache);
3733                         /*
3734                          * One of the free space endio workers might have
3735                          * created a new block group while updating a free space
3736                          * cache's inode (at inode.c:btrfs_finish_ordered_io())
3737                          * and hasn't released its transaction handle yet, in
3738                          * which case the new block group is still attached to
3739                          * its transaction handle and its creation has not
3740                          * finished yet (no block group item in the extent tree
3741                          * yet, etc). If this is the case, wait for all free
3742                          * space endio workers to finish and retry. This is a
3743                          * a very rare case so no need for a more efficient and
3744                          * complex approach.
3745                          */
3746                         if (ret == -ENOENT) {
3747                                 wait_event(cur_trans->writer_wait,
3748                                    atomic_read(&cur_trans->num_writers) == 1);
3749                                 ret = write_one_cache_group(trans, path, cache);
3750                         }
3751                         if (ret)
3752                                 btrfs_abort_transaction(trans, ret);
3753                 }
3754
3755                 /* if its not on the io list, we need to put the block group */
3756                 if (should_put)
3757                         btrfs_put_block_group(cache);
3758                 btrfs_delayed_refs_rsv_release(fs_info, 1);
3759                 spin_lock(&cur_trans->dirty_bgs_lock);
3760         }
3761         spin_unlock(&cur_trans->dirty_bgs_lock);
3762
3763         /*
3764          * Refer to the definition of io_bgs member for details why it's safe
3765          * to use it without any locking
3766          */
3767         while (!list_empty(io)) {
3768                 cache = list_first_entry(io, struct btrfs_block_group_cache,
3769                                          io_list);
3770                 list_del_init(&cache->io_list);
3771                 btrfs_wait_cache_io(trans, cache, path);
3772                 btrfs_put_block_group(cache);
3773         }
3774
3775         btrfs_free_path(path);
3776         return ret;
3777 }
3778
3779 int btrfs_extent_readonly(struct btrfs_fs_info *fs_info, u64 bytenr)
3780 {
3781         struct btrfs_block_group_cache *block_group;
3782         int readonly = 0;
3783
3784         block_group = btrfs_lookup_block_group(fs_info, bytenr);
3785         if (!block_group || block_group->ro)
3786                 readonly = 1;
3787         if (block_group)
3788                 btrfs_put_block_group(block_group);
3789         return readonly;
3790 }
3791
3792 bool btrfs_inc_nocow_writers(struct btrfs_fs_info *fs_info, u64 bytenr)
3793 {
3794         struct btrfs_block_group_cache *bg;
3795         bool ret = true;
3796
3797         bg = btrfs_lookup_block_group(fs_info, bytenr);
3798         if (!bg)
3799                 return false;
3800
3801         spin_lock(&bg->lock);
3802         if (bg->ro)
3803                 ret = false;
3804         else
3805                 atomic_inc(&bg->nocow_writers);
3806         spin_unlock(&bg->lock);
3807
3808         /* no put on block group, done by btrfs_dec_nocow_writers */
3809         if (!ret)
3810                 btrfs_put_block_group(bg);
3811
3812         return ret;
3813
3814 }
3815
3816 void btrfs_dec_nocow_writers(struct btrfs_fs_info *fs_info, u64 bytenr)
3817 {
3818         struct btrfs_block_group_cache *bg;
3819
3820         bg = btrfs_lookup_block_group(fs_info, bytenr);
3821         ASSERT(bg);
3822         if (atomic_dec_and_test(&bg->nocow_writers))
3823                 wake_up_var(&bg->nocow_writers);
3824         /*
3825          * Once for our lookup and once for the lookup done by a previous call
3826          * to btrfs_inc_nocow_writers()
3827          */
3828         btrfs_put_block_group(bg);
3829         btrfs_put_block_group(bg);
3830 }
3831
3832 void btrfs_wait_nocow_writers(struct btrfs_block_group_cache *bg)
3833 {
3834         wait_var_event(&bg->nocow_writers, !atomic_read(&bg->nocow_writers));
3835 }
3836
3837 static const char *alloc_name(u64 flags)
3838 {
3839         switch (flags) {
3840         case BTRFS_BLOCK_GROUP_METADATA|BTRFS_BLOCK_GROUP_DATA:
3841                 return "mixed";
3842         case BTRFS_BLOCK_GROUP_METADATA:
3843                 return "metadata";
3844         case BTRFS_BLOCK_GROUP_DATA:
3845                 return "data";
3846         case BTRFS_BLOCK_GROUP_SYSTEM:
3847                 return "system";
3848         default:
3849                 WARN_ON(1);
3850                 return "invalid-combination";
3851         };
3852 }
3853
3854 static int create_space_info(struct btrfs_fs_info *info, u64 flags)
3855 {
3856
3857         struct btrfs_space_info *space_info;
3858         int i;
3859         int ret;
3860
3861         space_info = kzalloc(sizeof(*space_info), GFP_NOFS);
3862         if (!space_info)
3863                 return -ENOMEM;
3864
3865         ret = percpu_counter_init(&space_info->total_bytes_pinned, 0,
3866                                  GFP_KERNEL);
3867         if (ret) {
3868                 kfree(space_info);
3869                 return ret;
3870         }
3871
3872         for (i = 0; i < BTRFS_NR_RAID_TYPES; i++)
3873                 INIT_LIST_HEAD(&space_info->block_groups[i]);
3874         init_rwsem(&space_info->groups_sem);
3875         spin_lock_init(&space_info->lock);
3876         space_info->flags = flags & BTRFS_BLOCK_GROUP_TYPE_MASK;
3877         space_info->force_alloc = CHUNK_ALLOC_NO_FORCE;
3878         init_waitqueue_head(&space_info->wait);
3879         INIT_LIST_HEAD(&space_info->ro_bgs);
3880         INIT_LIST_HEAD(&space_info->tickets);
3881         INIT_LIST_HEAD(&space_info->priority_tickets);
3882
3883         ret = kobject_init_and_add(&space_info->kobj, &space_info_ktype,
3884                                     info->space_info_kobj, "%s",
3885                                     alloc_name(space_info->flags));
3886         if (ret) {
3887                 kobject_put(&space_info->kobj);
3888                 return ret;
3889         }
3890
3891         list_add_rcu(&space_info->list, &info->space_info);
3892         if (flags & BTRFS_BLOCK_GROUP_DATA)
3893                 info->data_sinfo = space_info;
3894
3895         return ret;
3896 }
3897
3898 static void update_space_info(struct btrfs_fs_info *info, u64 flags,
3899                              u64 total_bytes, u64 bytes_used,
3900                              u64 bytes_readonly,
3901                              struct btrfs_space_info **space_info)
3902 {
3903         struct btrfs_space_info *found;
3904         int factor;
3905
3906         factor = btrfs_bg_type_to_factor(flags);
3907
3908         found = __find_space_info(info, flags);
3909         ASSERT(found);
3910         spin_lock(&found->lock);
3911         found->total_bytes += total_bytes;
3912         found->disk_total += total_bytes * factor;
3913         found->bytes_used += bytes_used;
3914         found->disk_used += bytes_used * factor;
3915         found->bytes_readonly += bytes_readonly;
3916         if (total_bytes > 0)
3917                 found->full = 0;
3918         space_info_add_new_bytes(info, found, total_bytes -
3919                                  bytes_used - bytes_readonly);
3920         spin_unlock(&found->lock);
3921         *space_info = found;
3922 }
3923
3924 static void set_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
3925 {
3926         u64 extra_flags = chunk_to_extended(flags) &
3927                                 BTRFS_EXTENDED_PROFILE_MASK;
3928
3929         write_seqlock(&fs_info->profiles_lock);
3930         if (flags & BTRFS_BLOCK_GROUP_DATA)
3931                 fs_info->avail_data_alloc_bits |= extra_flags;
3932         if (flags & BTRFS_BLOCK_GROUP_METADATA)
3933                 fs_info->avail_metadata_alloc_bits |= extra_flags;
3934         if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
3935                 fs_info->avail_system_alloc_bits |= extra_flags;
3936         write_sequnlock(&fs_info->profiles_lock);
3937 }
3938
3939 /*
3940  * returns target flags in extended format or 0 if restripe for this
3941  * chunk_type is not in progress
3942  *
3943  * should be called with balance_lock held
3944  */
3945 static u64 get_restripe_target(struct btrfs_fs_info *fs_info, u64 flags)
3946 {
3947         struct btrfs_balance_control *bctl = fs_info->balance_ctl;
3948         u64 target = 0;
3949
3950         if (!bctl)
3951                 return 0;
3952
3953         if (flags & BTRFS_BLOCK_GROUP_DATA &&
3954             bctl->data.flags & BTRFS_BALANCE_ARGS_CONVERT) {
3955                 target = BTRFS_BLOCK_GROUP_DATA | bctl->data.target;
3956         } else if (flags & BTRFS_BLOCK_GROUP_SYSTEM &&
3957                    bctl->sys.flags & BTRFS_BALANCE_ARGS_CONVERT) {
3958                 target = BTRFS_BLOCK_GROUP_SYSTEM | bctl->sys.target;
3959         } else if (flags & BTRFS_BLOCK_GROUP_METADATA &&
3960                    bctl->meta.flags & BTRFS_BALANCE_ARGS_CONVERT) {
3961                 target = BTRFS_BLOCK_GROUP_METADATA | bctl->meta.target;
3962         }
3963
3964         return target;
3965 }
3966
3967 /*
3968  * @flags: available profiles in extended format (see ctree.h)
3969  *
3970  * Returns reduced profile in chunk format.  If profile changing is in
3971  * progress (either running or paused) picks the target profile (if it's
3972  * already available), otherwise falls back to plain reducing.
3973  */
3974 static u64 btrfs_reduce_alloc_profile(struct btrfs_fs_info *fs_info, u64 flags)
3975 {
3976         u64 num_devices = fs_info->fs_devices->rw_devices;
3977         u64 target;
3978         u64 raid_type;
3979         u64 allowed = 0;
3980
3981         /*
3982          * see if restripe for this chunk_type is in progress, if so
3983          * try to reduce to the target profile
3984          */
3985         spin_lock(&fs_info->balance_lock);
3986         target = get_restripe_target(fs_info, flags);
3987         if (target) {
3988                 /* pick target profile only if it's already available */
3989                 if ((flags & target) & BTRFS_EXTENDED_PROFILE_MASK) {
3990                         spin_unlock(&fs_info->balance_lock);
3991                         return extended_to_chunk(target);
3992                 }
3993         }
3994         spin_unlock(&fs_info->balance_lock);
3995
3996         /* First, mask out the RAID levels which aren't possible */
3997         for (raid_type = 0; raid_type < BTRFS_NR_RAID_TYPES; raid_type++) {
3998                 if (num_devices >= btrfs_raid_array[raid_type].devs_min)
3999                         allowed |= btrfs_raid_array[raid_type].bg_flag;
4000         }
4001         allowed &= flags;
4002
4003         if (allowed & BTRFS_BLOCK_GROUP_RAID6)
4004                 allowed = BTRFS_BLOCK_GROUP_RAID6;
4005         else if (allowed & BTRFS_BLOCK_GROUP_RAID5)
4006                 allowed = BTRFS_BLOCK_GROUP_RAID5;
4007         else if (allowed & BTRFS_BLOCK_GROUP_RAID10)
4008                 allowed = BTRFS_BLOCK_GROUP_RAID10;
4009         else if (allowed & BTRFS_BLOCK_GROUP_RAID1)
4010                 allowed = BTRFS_BLOCK_GROUP_RAID1;
4011         else if (allowed & BTRFS_BLOCK_GROUP_RAID0)
4012                 allowed = BTRFS_BLOCK_GROUP_RAID0;
4013
4014         flags &= ~BTRFS_BLOCK_GROUP_PROFILE_MASK;
4015
4016         return extended_to_chunk(flags | allowed);
4017 }
4018
4019 static u64 get_alloc_profile(struct btrfs_fs_info *fs_info, u64 orig_flags)
4020 {
4021         unsigned seq;
4022         u64 flags;
4023
4024         do {
4025                 flags = orig_flags;
4026                 seq = read_seqbegin(&fs_info->profiles_lock);
4027
4028                 if (flags & BTRFS_BLOCK_GROUP_DATA)
4029                         flags |= fs_info->avail_data_alloc_bits;
4030                 else if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
4031                         flags |= fs_info->avail_system_alloc_bits;
4032                 else if (flags & BTRFS_BLOCK_GROUP_METADATA)
4033                         flags |= fs_info->avail_metadata_alloc_bits;
4034         } while (read_seqretry(&fs_info->profiles_lock, seq));
4035
4036         return btrfs_reduce_alloc_profile(fs_info, flags);
4037 }
4038
4039 static u64 get_alloc_profile_by_root(struct btrfs_root *root, int data)
4040 {
4041         struct btrfs_fs_info *fs_info = root->fs_info;
4042         u64 flags;
4043         u64 ret;
4044
4045         if (data)
4046                 flags = BTRFS_BLOCK_GROUP_DATA;
4047         else if (root == fs_info->chunk_root)
4048                 flags = BTRFS_BLOCK_GROUP_SYSTEM;
4049         else
4050                 flags = BTRFS_BLOCK_GROUP_METADATA;
4051
4052         ret = get_alloc_profile(fs_info, flags);
4053         return ret;
4054 }
4055
4056 u64 btrfs_data_alloc_profile(struct btrfs_fs_info *fs_info)
4057 {
4058         return get_alloc_profile(fs_info, BTRFS_BLOCK_GROUP_DATA);
4059 }
4060
4061 u64 btrfs_metadata_alloc_profile(struct btrfs_fs_info *fs_info)
4062 {
4063         return get_alloc_profile(fs_info, BTRFS_BLOCK_GROUP_METADATA);
4064 }
4065
4066 u64 btrfs_system_alloc_profile(struct btrfs_fs_info *fs_info)
4067 {
4068         return get_alloc_profile(fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
4069 }
4070
4071 static u64 btrfs_space_info_used(struct btrfs_space_info *s_info,
4072                                  bool may_use_included)
4073 {
4074         ASSERT(s_info);
4075         return s_info->bytes_used + s_info->bytes_reserved +
4076                 s_info->bytes_pinned + s_info->bytes_readonly +
4077                 (may_use_included ? s_info->bytes_may_use : 0);
4078 }
4079
4080 int btrfs_alloc_data_chunk_ondemand(struct btrfs_inode *inode, u64 bytes)
4081 {
4082         struct btrfs_root *root = inode->root;
4083         struct btrfs_fs_info *fs_info = root->fs_info;
4084         struct btrfs_space_info *data_sinfo = fs_info->data_sinfo;
4085         u64 used;
4086         int ret = 0;
4087         int need_commit = 2;
4088         int have_pinned_space;
4089
4090         /* make sure bytes are sectorsize aligned */
4091         bytes = ALIGN(bytes, fs_info->sectorsize);
4092
4093         if (btrfs_is_free_space_inode(inode)) {
4094                 need_commit = 0;
4095                 ASSERT(current->journal_info);
4096         }
4097
4098 again:
4099         /* make sure we have enough space to handle the data first */
4100         spin_lock(&data_sinfo->lock);
4101         used = btrfs_space_info_used(data_sinfo, true);
4102
4103         if (used + bytes > data_sinfo->total_bytes) {
4104                 struct btrfs_trans_handle *trans;
4105
4106                 /*
4107                  * if we don't have enough free bytes in this space then we need
4108                  * to alloc a new chunk.
4109                  */
4110                 if (!data_sinfo->full) {
4111                         u64 alloc_target;
4112
4113                         data_sinfo->force_alloc = CHUNK_ALLOC_FORCE;
4114                         spin_unlock(&data_sinfo->lock);
4115
4116                         alloc_target = btrfs_data_alloc_profile(fs_info);
4117                         /*
4118                          * It is ugly that we don't call nolock join
4119                          * transaction for the free space inode case here.
4120                          * But it is safe because we only do the data space
4121                          * reservation for the free space cache in the
4122                          * transaction context, the common join transaction
4123                          * just increase the counter of the current transaction
4124                          * handler, doesn't try to acquire the trans_lock of
4125                          * the fs.
4126                          */
4127                         trans = btrfs_join_transaction(root);
4128                         if (IS_ERR(trans))
4129                                 return PTR_ERR(trans);
4130
4131                         ret = do_chunk_alloc(trans, alloc_target,
4132                                              CHUNK_ALLOC_NO_FORCE);
4133                         btrfs_end_transaction(trans);
4134                         if (ret < 0) {
4135                                 if (ret != -ENOSPC)
4136                                         return ret;
4137                                 else {
4138                                         have_pinned_space = 1;
4139                                         goto commit_trans;
4140                                 }
4141                         }
4142
4143                         goto again;
4144                 }
4145
4146                 /*
4147                  * If we don't have enough pinned space to deal with this
4148                  * allocation, and no removed chunk in current transaction,
4149                  * don't bother committing the transaction.
4150                  */
4151                 have_pinned_space = __percpu_counter_compare(
4152                         &data_sinfo->total_bytes_pinned,
4153                         used + bytes - data_sinfo->total_bytes,
4154                         BTRFS_TOTAL_BYTES_PINNED_BATCH);
4155                 spin_unlock(&data_sinfo->lock);
4156
4157                 /* commit the current transaction and try again */
4158 commit_trans:
4159                 if (need_commit) {
4160                         need_commit--;
4161
4162                         if (need_commit > 0) {
4163                                 btrfs_start_delalloc_roots(fs_info, -1);
4164                                 btrfs_wait_ordered_roots(fs_info, U64_MAX, 0,
4165                                                          (u64)-1);
4166                         }
4167
4168                         trans = btrfs_join_transaction(root);
4169                         if (IS_ERR(trans))
4170                                 return PTR_ERR(trans);
4171                         if (have_pinned_space >= 0 ||
4172                             test_bit(BTRFS_TRANS_HAVE_FREE_BGS,
4173                                      &trans->transaction->flags) ||
4174                             need_commit > 0) {
4175                                 ret = btrfs_commit_transaction(trans);
4176                                 if (ret)
4177                                         return ret;
4178                                 /*
4179                                  * The cleaner kthread might still be doing iput
4180                                  * operations. Wait for it to finish so that
4181                                  * more space is released.  We don't need to
4182                                  * explicitly run the delayed iputs here because
4183                                  * the commit_transaction would have woken up
4184                                  * the cleaner.
4185                                  */
4186                                 ret = btrfs_wait_on_delayed_iputs(fs_info);
4187                                 if (ret)
4188                                         return ret;
4189                                 goto again;
4190                         } else {
4191                                 btrfs_end_transaction(trans);
4192                         }
4193                 }
4194
4195                 trace_btrfs_space_reservation(fs_info,
4196                                               "space_info:enospc",
4197                                               data_sinfo->flags, bytes, 1);
4198                 return -ENOSPC;
4199         }
4200         update_bytes_may_use(data_sinfo, bytes);
4201         trace_btrfs_space_reservation(fs_info, "space_info",
4202                                       data_sinfo->flags, bytes, 1);
4203         spin_unlock(&data_sinfo->lock);
4204
4205         return 0;
4206 }
4207
4208 int btrfs_check_data_free_space(struct inode *inode,
4209                         struct extent_changeset **reserved, u64 start, u64 len)
4210 {
4211         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
4212         int ret;
4213
4214         /* align the range */
4215         len = round_up(start + len, fs_info->sectorsize) -
4216               round_down(start, fs_info->sectorsize);
4217         start = round_down(start, fs_info->sectorsize);
4218
4219         ret = btrfs_alloc_data_chunk_ondemand(BTRFS_I(inode), len);
4220         if (ret < 0)
4221                 return ret;
4222
4223         /* Use new btrfs_qgroup_reserve_data to reserve precious data space. */
4224         ret = btrfs_qgroup_reserve_data(inode, reserved, start, len);
4225         if (ret < 0)
4226                 btrfs_free_reserved_data_space_noquota(inode, start, len);
4227         else
4228                 ret = 0;
4229         return ret;
4230 }
4231
4232 /*
4233  * Called if we need to clear a data reservation for this inode
4234  * Normally in a error case.
4235  *
4236  * This one will *NOT* use accurate qgroup reserved space API, just for case
4237  * which we can't sleep and is sure it won't affect qgroup reserved space.
4238  * Like clear_bit_hook().
4239  */
4240 void btrfs_free_reserved_data_space_noquota(struct inode *inode, u64 start,
4241                                             u64 len)
4242 {
4243         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
4244         struct btrfs_space_info *data_sinfo;
4245
4246         /* Make sure the range is aligned to sectorsize */
4247         len = round_up(start + len, fs_info->sectorsize) -
4248               round_down(start, fs_info->sectorsize);
4249         start = round_down(start, fs_info->sectorsize);
4250
4251         data_sinfo = fs_info->data_sinfo;
4252         spin_lock(&data_sinfo->lock);
4253         update_bytes_may_use(data_sinfo, -len);
4254         trace_btrfs_space_reservation(fs_info, "space_info",
4255                                       data_sinfo->flags, len, 0);
4256         spin_unlock(&data_sinfo->lock);
4257 }
4258
4259 /*
4260  * Called if we need to clear a data reservation for this inode
4261  * Normally in a error case.
4262  *
4263  * This one will handle the per-inode data rsv map for accurate reserved
4264  * space framework.
4265  */
4266 void btrfs_free_reserved_data_space(struct inode *inode,
4267                         struct extent_changeset *reserved, u64 start, u64 len)
4268 {
4269         struct btrfs_root *root = BTRFS_I(inode)->root;
4270
4271         /* Make sure the range is aligned to sectorsize */
4272         len = round_up(start + len, root->fs_info->sectorsize) -
4273               round_down(start, root->fs_info->sectorsize);
4274         start = round_down(start, root->fs_info->sectorsize);
4275
4276         btrfs_free_reserved_data_space_noquota(inode, start, len);
4277         btrfs_qgroup_free_data(inode, reserved, start, len);
4278 }
4279
4280 static void force_metadata_allocation(struct btrfs_fs_info *info)
4281 {
4282         struct list_head *head = &info->space_info;
4283         struct btrfs_space_info *found;
4284
4285         rcu_read_lock();
4286         list_for_each_entry_rcu(found, head, list) {
4287                 if (found->flags & BTRFS_BLOCK_GROUP_METADATA)
4288                         found->force_alloc = CHUNK_ALLOC_FORCE;
4289         }
4290         rcu_read_unlock();
4291 }
4292
4293 static inline u64 calc_global_rsv_need_space(struct btrfs_block_rsv *global)
4294 {
4295         return (global->size << 1);
4296 }
4297
4298 static int should_alloc_chunk(struct btrfs_fs_info *fs_info,
4299                               struct btrfs_space_info *sinfo, int force)
4300 {
4301         u64 bytes_used = btrfs_space_info_used(sinfo, false);
4302         u64 thresh;
4303
4304         if (force == CHUNK_ALLOC_FORCE)
4305                 return 1;
4306
4307         /*
4308          * in limited mode, we want to have some free space up to
4309          * about 1% of the FS size.
4310          */
4311         if (force == CHUNK_ALLOC_LIMITED) {
4312                 thresh = btrfs_super_total_bytes(fs_info->super_copy);
4313                 thresh = max_t(u64, SZ_64M, div_factor_fine(thresh, 1));
4314
4315                 if (sinfo->total_bytes - bytes_used < thresh)
4316                         return 1;
4317         }
4318
4319         if (bytes_used + SZ_2M < div_factor(sinfo->total_bytes, 8))
4320                 return 0;
4321         return 1;
4322 }
4323
4324 static u64 get_profile_num_devs(struct btrfs_fs_info *fs_info, u64 type)
4325 {
4326         u64 num_dev;
4327
4328         if (type & (BTRFS_BLOCK_GROUP_RAID10 |
4329                     BTRFS_BLOCK_GROUP_RAID0 |
4330                     BTRFS_BLOCK_GROUP_RAID5 |
4331                     BTRFS_BLOCK_GROUP_RAID6))
4332                 num_dev = fs_info->fs_devices->rw_devices;
4333         else if (type & BTRFS_BLOCK_GROUP_RAID1)
4334                 num_dev = 2;
4335         else
4336                 num_dev = 1;    /* DUP or single */
4337
4338         return num_dev;
4339 }
4340
4341 /*
4342  * If @is_allocation is true, reserve space in the system space info necessary
4343  * for allocating a chunk, otherwise if it's false, reserve space necessary for
4344  * removing a chunk.
4345  */
4346 void check_system_chunk(struct btrfs_trans_handle *trans, u64 type)
4347 {
4348         struct btrfs_fs_info *fs_info = trans->fs_info;
4349         struct btrfs_space_info *info;
4350         u64 left;
4351         u64 thresh;
4352         int ret = 0;
4353         u64 num_devs;
4354
4355         /*
4356          * Needed because we can end up allocating a system chunk and for an
4357          * atomic and race free space reservation in the chunk block reserve.
4358          */
4359         lockdep_assert_held(&fs_info->chunk_mutex);
4360
4361         info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
4362         spin_lock(&info->lock);
4363         left = info->total_bytes - btrfs_space_info_used(info, true);
4364         spin_unlock(&info->lock);
4365
4366         num_devs = get_profile_num_devs(fs_info, type);
4367
4368         /* num_devs device items to update and 1 chunk item to add or remove */
4369         thresh = btrfs_calc_trunc_metadata_size(fs_info, num_devs) +
4370                 btrfs_calc_trans_metadata_size(fs_info, 1);
4371
4372         if (left < thresh && btrfs_test_opt(fs_info, ENOSPC_DEBUG)) {
4373                 btrfs_info(fs_info, "left=%llu, need=%llu, flags=%llu",
4374                            left, thresh, type);
4375                 dump_space_info(fs_info, info, 0, 0);
4376         }
4377
4378         if (left < thresh) {
4379                 u64 flags = btrfs_system_alloc_profile(fs_info);
4380
4381                 /*
4382                  * Ignore failure to create system chunk. We might end up not
4383                  * needing it, as we might not need to COW all nodes/leafs from
4384                  * the paths we visit in the chunk tree (they were already COWed
4385                  * or created in the current transaction for example).
4386                  */
4387                 ret = btrfs_alloc_chunk(trans, flags);
4388         }
4389
4390         if (!ret) {
4391                 ret = btrfs_block_rsv_add(fs_info->chunk_root,
4392                                           &fs_info->chunk_block_rsv,
4393                                           thresh, BTRFS_RESERVE_NO_FLUSH);
4394                 if (!ret)
4395                         trans->chunk_bytes_reserved += thresh;
4396         }
4397 }
4398
4399 /*
4400  * If force is CHUNK_ALLOC_FORCE:
4401  *    - return 1 if it successfully allocates a chunk,
4402  *    - return errors including -ENOSPC otherwise.
4403  * If force is NOT CHUNK_ALLOC_FORCE:
4404  *    - return 0 if it doesn't need to allocate a new chunk,
4405  *    - return 1 if it successfully allocates a chunk,
4406  *    - return errors including -ENOSPC otherwise.
4407  */
4408 static int do_chunk_alloc(struct btrfs_trans_handle *trans, u64 flags,
4409                           int force)
4410 {
4411         struct btrfs_fs_info *fs_info = trans->fs_info;
4412         struct btrfs_space_info *space_info;
4413         bool wait_for_alloc = false;
4414         bool should_alloc = false;
4415         int ret = 0;
4416
4417         /* Don't re-enter if we're already allocating a chunk */
4418         if (trans->allocating_chunk)
4419                 return -ENOSPC;
4420
4421         space_info = __find_space_info(fs_info, flags);
4422         ASSERT(space_info);
4423
4424         do {
4425                 spin_lock(&space_info->lock);
4426                 if (force < space_info->force_alloc)
4427                         force = space_info->force_alloc;
4428                 should_alloc = should_alloc_chunk(fs_info, space_info, force);
4429                 if (space_info->full) {
4430                         /* No more free physical space */
4431                         if (should_alloc)
4432                                 ret = -ENOSPC;
4433                         else
4434                                 ret = 0;
4435                         spin_unlock(&space_info->lock);
4436                         return ret;
4437                 } else if (!should_alloc) {
4438                         spin_unlock(&space_info->lock);
4439                         return 0;
4440                 } else if (space_info->chunk_alloc) {
4441                         /*
4442                          * Someone is already allocating, so we need to block
4443                          * until this someone is finished and then loop to
4444                          * recheck if we should continue with our allocation
4445                          * attempt.
4446                          */
4447                         wait_for_alloc = true;
4448                         spin_unlock(&space_info->lock);
4449                         mutex_lock(&fs_info->chunk_mutex);
4450                         mutex_unlock(&fs_info->chunk_mutex);
4451                 } else {
4452                         /* Proceed with allocation */
4453                         space_info->chunk_alloc = 1;
4454                         wait_for_alloc = false;
4455                         spin_unlock(&space_info->lock);
4456                 }
4457
4458                 cond_resched();
4459         } while (wait_for_alloc);
4460
4461         mutex_lock(&fs_info->chunk_mutex);
4462         trans->allocating_chunk = true;
4463
4464         /*
4465          * If we have mixed data/metadata chunks we want to make sure we keep
4466          * allocating mixed chunks instead of individual chunks.
4467          */
4468         if (btrfs_mixed_space_info(space_info))
4469                 flags |= (BTRFS_BLOCK_GROUP_DATA | BTRFS_BLOCK_GROUP_METADATA);
4470
4471         /*
4472          * if we're doing a data chunk, go ahead and make sure that
4473          * we keep a reasonable number of metadata chunks allocated in the
4474          * FS as well.
4475          */
4476         if (flags & BTRFS_BLOCK_GROUP_DATA && fs_info->metadata_ratio) {
4477                 fs_info->data_chunk_allocations++;
4478                 if (!(fs_info->data_chunk_allocations %
4479                       fs_info->metadata_ratio))
4480                         force_metadata_allocation(fs_info);
4481         }
4482
4483         /*
4484          * Check if we have enough space in SYSTEM chunk because we may need
4485          * to update devices.
4486          */
4487         check_system_chunk(trans, flags);
4488
4489         ret = btrfs_alloc_chunk(trans, flags);
4490         trans->allocating_chunk = false;
4491
4492         spin_lock(&space_info->lock);
4493         if (ret < 0) {
4494                 if (ret == -ENOSPC)
4495                         space_info->full = 1;
4496                 else
4497                         goto out;
4498         } else {
4499                 ret = 1;
4500                 space_info->max_extent_size = 0;
4501         }
4502
4503         space_info->force_alloc = CHUNK_ALLOC_NO_FORCE;
4504 out:
4505         space_info->chunk_alloc = 0;
4506         spin_unlock(&space_info->lock);
4507         mutex_unlock(&fs_info->chunk_mutex);
4508         /*
4509          * When we allocate a new chunk we reserve space in the chunk block
4510          * reserve to make sure we can COW nodes/leafs in the chunk tree or
4511          * add new nodes/leafs to it if we end up needing to do it when
4512          * inserting the chunk item and updating device items as part of the
4513          * second phase of chunk allocation, performed by
4514          * btrfs_finish_chunk_alloc(). So make sure we don't accumulate a
4515          * large number of new block groups to create in our transaction
4516          * handle's new_bgs list to avoid exhausting the chunk block reserve
4517          * in extreme cases - like having a single transaction create many new
4518          * block groups when starting to write out the free space caches of all
4519          * the block groups that were made dirty during the lifetime of the
4520          * transaction.
4521          */
4522         if (trans->chunk_bytes_reserved >= (u64)SZ_2M)
4523                 btrfs_create_pending_block_groups(trans);
4524
4525         return ret;
4526 }
4527
4528 static int can_overcommit(struct btrfs_fs_info *fs_info,
4529                           struct btrfs_space_info *space_info, u64 bytes,
4530                           enum btrfs_reserve_flush_enum flush,
4531                           bool system_chunk)
4532 {
4533         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
4534         u64 profile;
4535         u64 space_size;
4536         u64 avail;
4537         u64 used;
4538         int factor;
4539
4540         /* Don't overcommit when in mixed mode. */
4541         if (space_info->flags & BTRFS_BLOCK_GROUP_DATA)
4542                 return 0;
4543
4544         if (system_chunk)
4545                 profile = btrfs_system_alloc_profile(fs_info);
4546         else
4547                 profile = btrfs_metadata_alloc_profile(fs_info);
4548
4549         used = btrfs_space_info_used(space_info, false);
4550
4551         /*
4552          * We only want to allow over committing if we have lots of actual space
4553          * free, but if we don't have enough space to handle the global reserve
4554          * space then we could end up having a real enospc problem when trying
4555          * to allocate a chunk or some other such important allocation.
4556          */
4557         spin_lock(&global_rsv->lock);
4558         space_size = calc_global_rsv_need_space(global_rsv);
4559         spin_unlock(&global_rsv->lock);
4560         if (used + space_size >= space_info->total_bytes)
4561                 return 0;
4562
4563         used += space_info->bytes_may_use;
4564
4565         avail = atomic64_read(&fs_info->free_chunk_space);
4566
4567         /*
4568          * If we have dup, raid1 or raid10 then only half of the free
4569          * space is actually usable.  For raid56, the space info used
4570          * doesn't include the parity drive, so we don't have to
4571          * change the math
4572          */
4573         factor = btrfs_bg_type_to_factor(profile);
4574         avail = div_u64(avail, factor);
4575
4576         /*
4577          * If we aren't flushing all things, let us overcommit up to
4578          * 1/2th of the space. If we can flush, don't let us overcommit
4579          * too much, let it overcommit up to 1/8 of the space.
4580          */
4581         if (flush == BTRFS_RESERVE_FLUSH_ALL)
4582                 avail >>= 3;
4583         else
4584                 avail >>= 1;
4585
4586         if (used + bytes < space_info->total_bytes + avail)
4587                 return 1;
4588         return 0;
4589 }
4590
4591 static void btrfs_writeback_inodes_sb_nr(struct btrfs_fs_info *fs_info,
4592                                          unsigned long nr_pages, int nr_items)
4593 {
4594         struct super_block *sb = fs_info->sb;
4595
4596         if (down_read_trylock(&sb->s_umount)) {
4597                 writeback_inodes_sb_nr(sb, nr_pages, WB_REASON_FS_FREE_SPACE);
4598                 up_read(&sb->s_umount);
4599         } else {
4600                 /*
4601                  * We needn't worry the filesystem going from r/w to r/o though
4602                  * we don't acquire ->s_umount mutex, because the filesystem
4603                  * should guarantee the delalloc inodes list be empty after
4604                  * the filesystem is readonly(all dirty pages are written to
4605                  * the disk).
4606                  */
4607                 btrfs_start_delalloc_roots(fs_info, nr_items);
4608                 if (!current->journal_info)
4609                         btrfs_wait_ordered_roots(fs_info, nr_items, 0, (u64)-1);
4610         }
4611 }
4612
4613 static inline u64 calc_reclaim_items_nr(struct btrfs_fs_info *fs_info,
4614                                         u64 to_reclaim)
4615 {
4616         u64 bytes;
4617         u64 nr;
4618
4619         bytes = btrfs_calc_trans_metadata_size(fs_info, 1);
4620         nr = div64_u64(to_reclaim, bytes);
4621         if (!nr)
4622                 nr = 1;
4623         return nr;
4624 }
4625
4626 #define EXTENT_SIZE_PER_ITEM    SZ_256K
4627
4628 /*
4629  * shrink metadata reservation for delalloc
4630  */
4631 static void shrink_delalloc(struct btrfs_fs_info *fs_info, u64 to_reclaim,
4632                             u64 orig, bool wait_ordered)
4633 {
4634         struct btrfs_space_info *space_info;
4635         struct btrfs_trans_handle *trans;
4636         u64 delalloc_bytes;
4637         u64 dio_bytes;
4638         u64 async_pages;
4639         u64 items;
4640         long time_left;
4641         unsigned long nr_pages;
4642         int loops;
4643
4644         /* Calc the number of the pages we need flush for space reservation */
4645         items = calc_reclaim_items_nr(fs_info, to_reclaim);
4646         to_reclaim = items * EXTENT_SIZE_PER_ITEM;
4647
4648         trans = (struct btrfs_trans_handle *)current->journal_info;
4649         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
4650
4651         delalloc_bytes = percpu_counter_sum_positive(
4652                                                 &fs_info->delalloc_bytes);
4653         dio_bytes = percpu_counter_sum_positive(&fs_info->dio_bytes);
4654         if (delalloc_bytes == 0 && dio_bytes == 0) {
4655                 if (trans)
4656                         return;
4657                 if (wait_ordered)
4658                         btrfs_wait_ordered_roots(fs_info, items, 0, (u64)-1);
4659                 return;
4660         }
4661
4662         /*
4663          * If we are doing more ordered than delalloc we need to just wait on
4664          * ordered extents, otherwise we'll waste time trying to flush delalloc
4665          * that likely won't give us the space back we need.
4666          */
4667         if (dio_bytes > delalloc_bytes)
4668                 wait_ordered = true;
4669
4670         loops = 0;
4671         while ((delalloc_bytes || dio_bytes) && loops < 3) {
4672                 nr_pages = min(delalloc_bytes, to_reclaim) >> PAGE_SHIFT;
4673
4674                 /*
4675                  * Triggers inode writeback for up to nr_pages. This will invoke
4676                  * ->writepages callback and trigger delalloc filling
4677                  *  (btrfs_run_delalloc_range()).
4678                  */
4679                 btrfs_writeback_inodes_sb_nr(fs_info, nr_pages, items);
4680
4681                 /*
4682                  * We need to wait for the compressed pages to start before
4683                  * we continue.
4684                  */
4685                 async_pages = atomic_read(&fs_info->async_delalloc_pages);
4686                 if (!async_pages)
4687                         goto skip_async;
4688
4689                 /*
4690                  * Calculate how many compressed pages we want to be written
4691                  * before we continue. I.e if there are more async pages than we
4692                  * require wait_event will wait until nr_pages are written.
4693                  */
4694                 if (async_pages <= nr_pages)
4695                         async_pages = 0;
4696                 else
4697                         async_pages -= nr_pages;
4698
4699                 wait_event(fs_info->async_submit_wait,
4700                            atomic_read(&fs_info->async_delalloc_pages) <=
4701                            (int)async_pages);
4702 skip_async:
4703                 spin_lock(&space_info->lock);
4704                 if (list_empty(&space_info->tickets) &&
4705                     list_empty(&space_info->priority_tickets)) {
4706                         spin_unlock(&space_info->lock);
4707                         break;
4708                 }
4709                 spin_unlock(&space_info->lock);
4710
4711                 loops++;
4712                 if (wait_ordered && !trans) {
4713                         btrfs_wait_ordered_roots(fs_info, items, 0, (u64)-1);
4714                 } else {
4715                         time_left = schedule_timeout_killable(1);
4716                         if (time_left)
4717                                 break;
4718                 }
4719                 delalloc_bytes = percpu_counter_sum_positive(
4720                                                 &fs_info->delalloc_bytes);
4721                 dio_bytes = percpu_counter_sum_positive(&fs_info->dio_bytes);
4722         }
4723 }
4724
4725 struct reserve_ticket {
4726         u64 orig_bytes;
4727         u64 bytes;
4728         int error;
4729         struct list_head list;
4730         wait_queue_head_t wait;
4731 };
4732
4733 /**
4734  * maybe_commit_transaction - possibly commit the transaction if its ok to
4735  * @root - the root we're allocating for
4736  * @bytes - the number of bytes we want to reserve
4737  * @force - force the commit
4738  *
4739  * This will check to make sure that committing the transaction will actually
4740  * get us somewhere and then commit the transaction if it does.  Otherwise it
4741  * will return -ENOSPC.
4742  */
4743 static int may_commit_transaction(struct btrfs_fs_info *fs_info,
4744                                   struct btrfs_space_info *space_info)
4745 {
4746         struct reserve_ticket *ticket = NULL;
4747         struct btrfs_block_rsv *delayed_rsv = &fs_info->delayed_block_rsv;
4748         struct btrfs_block_rsv *delayed_refs_rsv = &fs_info->delayed_refs_rsv;
4749         struct btrfs_trans_handle *trans;
4750         u64 bytes_needed;
4751         u64 reclaim_bytes = 0;
4752
4753         trans = (struct btrfs_trans_handle *)current->journal_info;
4754         if (trans)
4755                 return -EAGAIN;
4756
4757         spin_lock(&space_info->lock);
4758         if (!list_empty(&space_info->priority_tickets))
4759                 ticket = list_first_entry(&space_info->priority_tickets,
4760                                           struct reserve_ticket, list);
4761         else if (!list_empty(&space_info->tickets))
4762                 ticket = list_first_entry(&space_info->tickets,
4763                                           struct reserve_ticket, list);
4764         bytes_needed = (ticket) ? ticket->bytes : 0;
4765         spin_unlock(&space_info->lock);
4766
4767         if (!bytes_needed)
4768                 return 0;
4769
4770         trans = btrfs_join_transaction(fs_info->extent_root);
4771         if (IS_ERR(trans))
4772                 return PTR_ERR(trans);
4773
4774         /*
4775          * See if there is enough pinned space to make this reservation, or if
4776          * we have block groups that are going to be freed, allowing us to
4777          * possibly do a chunk allocation the next loop through.
4778          */
4779         if (test_bit(BTRFS_TRANS_HAVE_FREE_BGS, &trans->transaction->flags) ||
4780             __percpu_counter_compare(&space_info->total_bytes_pinned,
4781                                      bytes_needed,
4782                                      BTRFS_TOTAL_BYTES_PINNED_BATCH) >= 0)
4783                 goto commit;
4784
4785         /*
4786          * See if there is some space in the delayed insertion reservation for
4787          * this reservation.
4788          */
4789         if (space_info != delayed_rsv->space_info)
4790                 goto enospc;
4791
4792         spin_lock(&delayed_rsv->lock);
4793         reclaim_bytes += delayed_rsv->reserved;
4794         spin_unlock(&delayed_rsv->lock);
4795
4796         spin_lock(&delayed_refs_rsv->lock);
4797         reclaim_bytes += delayed_refs_rsv->reserved;
4798         spin_unlock(&delayed_refs_rsv->lock);
4799         if (reclaim_bytes >= bytes_needed)
4800                 goto commit;
4801         bytes_needed -= reclaim_bytes;
4802
4803         if (__percpu_counter_compare(&space_info->total_bytes_pinned,
4804                                    bytes_needed,
4805                                    BTRFS_TOTAL_BYTES_PINNED_BATCH) < 0)
4806                 goto enospc;
4807
4808 commit:
4809         return btrfs_commit_transaction(trans);
4810 enospc:
4811         btrfs_end_transaction(trans);
4812         return -ENOSPC;
4813 }
4814
4815 /*
4816  * Try to flush some data based on policy set by @state. This is only advisory
4817  * and may fail for various reasons. The caller is supposed to examine the
4818  * state of @space_info to detect the outcome.
4819  */
4820 static void flush_space(struct btrfs_fs_info *fs_info,
4821                        struct btrfs_space_info *space_info, u64 num_bytes,
4822                        int state)
4823 {
4824         struct btrfs_root *root = fs_info->extent_root;
4825         struct btrfs_trans_handle *trans;
4826         int nr;
4827         int ret = 0;
4828
4829         switch (state) {
4830         case FLUSH_DELAYED_ITEMS_NR:
4831         case FLUSH_DELAYED_ITEMS:
4832                 if (state == FLUSH_DELAYED_ITEMS_NR)
4833                         nr = calc_reclaim_items_nr(fs_info, num_bytes) * 2;
4834                 else
4835                         nr = -1;
4836
4837                 trans = btrfs_join_transaction(root);
4838                 if (IS_ERR(trans)) {
4839                         ret = PTR_ERR(trans);
4840                         break;
4841                 }
4842                 ret = btrfs_run_delayed_items_nr(trans, nr);
4843                 btrfs_end_transaction(trans);
4844                 break;
4845         case FLUSH_DELALLOC:
4846         case FLUSH_DELALLOC_WAIT:
4847                 shrink_delalloc(fs_info, num_bytes * 2, num_bytes,
4848                                 state == FLUSH_DELALLOC_WAIT);
4849                 break;
4850         case FLUSH_DELAYED_REFS_NR:
4851         case FLUSH_DELAYED_REFS:
4852                 trans = btrfs_join_transaction(root);
4853                 if (IS_ERR(trans)) {
4854                         ret = PTR_ERR(trans);
4855                         break;
4856                 }
4857                 if (state == FLUSH_DELAYED_REFS_NR)
4858                         nr = calc_reclaim_items_nr(fs_info, num_bytes);
4859                 else
4860                         nr = 0;
4861                 btrfs_run_delayed_refs(trans, nr);
4862                 btrfs_end_transaction(trans);
4863                 break;
4864         case ALLOC_CHUNK:
4865         case ALLOC_CHUNK_FORCE:
4866                 trans = btrfs_join_transaction(root);
4867                 if (IS_ERR(trans)) {
4868                         ret = PTR_ERR(trans);
4869                         break;
4870                 }
4871                 ret = do_chunk_alloc(trans,
4872                                      btrfs_metadata_alloc_profile(fs_info),
4873                                      (state == ALLOC_CHUNK) ?
4874                                       CHUNK_ALLOC_NO_FORCE : CHUNK_ALLOC_FORCE);
4875                 btrfs_end_transaction(trans);
4876                 if (ret > 0 || ret == -ENOSPC)
4877                         ret = 0;
4878                 break;
4879         case COMMIT_TRANS:
4880                 /*
4881                  * If we have pending delayed iputs then we could free up a
4882                  * bunch of pinned space, so make sure we run the iputs before
4883                  * we do our pinned bytes check below.
4884                  */
4885                 btrfs_run_delayed_iputs(fs_info);
4886                 btrfs_wait_on_delayed_iputs(fs_info);
4887
4888                 ret = may_commit_transaction(fs_info, space_info);
4889                 break;
4890         default:
4891                 ret = -ENOSPC;
4892                 break;
4893         }
4894
4895         trace_btrfs_flush_space(fs_info, space_info->flags, num_bytes, state,
4896                                 ret);
4897         return;
4898 }
4899
4900 static inline u64
4901 btrfs_calc_reclaim_metadata_size(struct btrfs_fs_info *fs_info,
4902                                  struct btrfs_space_info *space_info,
4903                                  bool system_chunk)
4904 {
4905         struct reserve_ticket *ticket;
4906         u64 used;
4907         u64 expected;
4908         u64 to_reclaim = 0;
4909
4910         list_for_each_entry(ticket, &space_info->tickets, list)
4911                 to_reclaim += ticket->bytes;
4912         list_for_each_entry(ticket, &space_info->priority_tickets, list)
4913                 to_reclaim += ticket->bytes;
4914         if (to_reclaim)
4915                 return to_reclaim;
4916
4917         to_reclaim = min_t(u64, num_online_cpus() * SZ_1M, SZ_16M);
4918         if (can_overcommit(fs_info, space_info, to_reclaim,
4919                            BTRFS_RESERVE_FLUSH_ALL, system_chunk))
4920                 return 0;
4921
4922         used = btrfs_space_info_used(space_info, true);
4923
4924         if (can_overcommit(fs_info, space_info, SZ_1M,
4925                            BTRFS_RESERVE_FLUSH_ALL, system_chunk))
4926                 expected = div_factor_fine(space_info->total_bytes, 95);
4927         else
4928                 expected = div_factor_fine(space_info->total_bytes, 90);
4929
4930         if (used > expected)
4931                 to_reclaim = used - expected;
4932         else
4933                 to_reclaim = 0;
4934         to_reclaim = min(to_reclaim, space_info->bytes_may_use +
4935                                      space_info->bytes_reserved);
4936         return to_reclaim;
4937 }
4938
4939 static inline int need_do_async_reclaim(struct btrfs_fs_info *fs_info,
4940                                         struct btrfs_space_info *space_info,
4941                                         u64 used, bool system_chunk)
4942 {
4943         u64 thresh = div_factor_fine(space_info->total_bytes, 98);
4944
4945         /* If we're just plain full then async reclaim just slows us down. */
4946         if ((space_info->bytes_used + space_info->bytes_reserved) >= thresh)
4947                 return 0;
4948
4949         if (!btrfs_calc_reclaim_metadata_size(fs_info, space_info,
4950                                               system_chunk))
4951                 return 0;
4952
4953         return (used >= thresh && !btrfs_fs_closing(fs_info) &&
4954                 !test_bit(BTRFS_FS_STATE_REMOUNTING, &fs_info->fs_state));
4955 }
4956
4957 static bool wake_all_tickets(struct list_head *head)
4958 {
4959         struct reserve_ticket *ticket;
4960
4961         while (!list_empty(head)) {
4962                 ticket = list_first_entry(head, struct reserve_ticket, list);
4963                 list_del_init(&ticket->list);
4964                 ticket->error = -ENOSPC;
4965                 wake_up(&ticket->wait);
4966                 if (ticket->bytes != ticket->orig_bytes)
4967                         return true;
4968         }
4969         return false;
4970 }
4971
4972 /*
4973  * This is for normal flushers, we can wait all goddamned day if we want to.  We
4974  * will loop and continuously try to flush as long as we are making progress.
4975  * We count progress as clearing off tickets each time we have to loop.
4976  */
4977 static void btrfs_async_reclaim_metadata_space(struct work_struct *work)
4978 {
4979         struct btrfs_fs_info *fs_info;
4980         struct btrfs_space_info *space_info;
4981         u64 to_reclaim;
4982         int flush_state;
4983         int commit_cycles = 0;
4984         u64 last_tickets_id;
4985
4986         fs_info = container_of(work, struct btrfs_fs_info, async_reclaim_work);
4987         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
4988
4989         spin_lock(&space_info->lock);
4990         to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info, space_info,
4991                                                       false);
4992         if (!to_reclaim) {
4993                 space_info->flush = 0;
4994                 spin_unlock(&space_info->lock);
4995                 return;
4996         }
4997         last_tickets_id = space_info->tickets_id;
4998         spin_unlock(&space_info->lock);
4999
5000         flush_state = FLUSH_DELAYED_ITEMS_NR;
5001         do {
5002                 flush_space(fs_info, space_info, to_reclaim, flush_state);
5003                 spin_lock(&space_info->lock);
5004                 if (list_empty(&space_info->tickets)) {
5005                         space_info->flush = 0;
5006                         spin_unlock(&space_info->lock);
5007                         return;
5008                 }
5009                 to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info,
5010                                                               space_info,
5011                                                               false);
5012                 if (last_tickets_id == space_info->tickets_id) {
5013                         flush_state++;
5014                 } else {
5015                         last_tickets_id = space_info->tickets_id;
5016                         flush_state = FLUSH_DELAYED_ITEMS_NR;
5017                         if (commit_cycles)
5018                                 commit_cycles--;
5019                 }
5020
5021                 /*
5022                  * We don't want to force a chunk allocation until we've tried
5023                  * pretty hard to reclaim space.  Think of the case where we
5024                  * freed up a bunch of space and so have a lot of pinned space
5025                  * to reclaim.  We would rather use that than possibly create a
5026                  * underutilized metadata chunk.  So if this is our first run
5027                  * through the flushing state machine skip ALLOC_CHUNK_FORCE and
5028                  * commit the transaction.  If nothing has changed the next go
5029                  * around then we can force a chunk allocation.
5030                  */
5031                 if (flush_state == ALLOC_CHUNK_FORCE && !commit_cycles)
5032                         flush_state++;
5033
5034                 if (flush_state > COMMIT_TRANS) {
5035                         commit_cycles++;
5036                         if (commit_cycles > 2) {
5037                                 if (wake_all_tickets(&space_info->tickets)) {
5038                                         flush_state = FLUSH_DELAYED_ITEMS_NR;
5039                                         commit_cycles--;
5040                                 } else {
5041                                         space_info->flush = 0;
5042                                 }
5043                         } else {
5044                                 flush_state = FLUSH_DELAYED_ITEMS_NR;
5045                         }
5046                 }
5047                 spin_unlock(&space_info->lock);
5048         } while (flush_state <= COMMIT_TRANS);
5049 }
5050
5051 void btrfs_init_async_reclaim_work(struct work_struct *work)
5052 {
5053         INIT_WORK(work, btrfs_async_reclaim_metadata_space);
5054 }
5055
5056 static const enum btrfs_flush_state priority_flush_states[] = {
5057         FLUSH_DELAYED_ITEMS_NR,
5058         FLUSH_DELAYED_ITEMS,
5059         ALLOC_CHUNK,
5060 };
5061
5062 static void priority_reclaim_metadata_space(struct btrfs_fs_info *fs_info,
5063                                             struct btrfs_space_info *space_info,
5064                                             struct reserve_ticket *ticket)
5065 {
5066         u64 to_reclaim;
5067         int flush_state;
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                 spin_unlock(&space_info->lock);
5074                 return;
5075         }
5076         spin_unlock(&space_info->lock);
5077
5078         flush_state = 0;
5079         do {
5080                 flush_space(fs_info, space_info, to_reclaim,
5081                             priority_flush_states[flush_state]);
5082                 flush_state++;
5083                 spin_lock(&space_info->lock);
5084                 if (ticket->bytes == 0) {
5085                         spin_unlock(&space_info->lock);
5086                         return;
5087                 }
5088                 spin_unlock(&space_info->lock);
5089         } while (flush_state < ARRAY_SIZE(priority_flush_states));
5090 }
5091
5092 static int wait_reserve_ticket(struct btrfs_fs_info *fs_info,
5093                                struct btrfs_space_info *space_info,
5094                                struct reserve_ticket *ticket)
5095
5096 {
5097         DEFINE_WAIT(wait);
5098         u64 reclaim_bytes = 0;
5099         int ret = 0;
5100
5101         spin_lock(&space_info->lock);
5102         while (ticket->bytes > 0 && ticket->error == 0) {
5103                 ret = prepare_to_wait_event(&ticket->wait, &wait, TASK_KILLABLE);
5104                 if (ret) {
5105                         ret = -EINTR;
5106                         break;
5107                 }
5108                 spin_unlock(&space_info->lock);
5109
5110                 schedule();
5111
5112                 finish_wait(&ticket->wait, &wait);
5113                 spin_lock(&space_info->lock);
5114         }
5115         if (!ret)
5116                 ret = ticket->error;
5117         if (!list_empty(&ticket->list))
5118                 list_del_init(&ticket->list);
5119         if (ticket->bytes && ticket->bytes < ticket->orig_bytes)
5120                 reclaim_bytes = ticket->orig_bytes - ticket->bytes;
5121         spin_unlock(&space_info->lock);
5122
5123         if (reclaim_bytes)
5124                 space_info_add_old_bytes(fs_info, space_info, reclaim_bytes);
5125         return ret;
5126 }
5127
5128 /**
5129  * reserve_metadata_bytes - try to reserve bytes from the block_rsv's space
5130  * @root - the root we're allocating for
5131  * @space_info - the space info we want to allocate from
5132  * @orig_bytes - the number of bytes we want
5133  * @flush - whether or not we can flush to make our reservation
5134  *
5135  * This will reserve orig_bytes number of bytes from the space info associated
5136  * with the block_rsv.  If there is not enough space it will make an attempt to
5137  * flush out space to make room.  It will do this by flushing delalloc if
5138  * possible or committing the transaction.  If flush is 0 then no attempts to
5139  * regain reservations will be made and this will fail if there is not enough
5140  * space already.
5141  */
5142 static int __reserve_metadata_bytes(struct btrfs_fs_info *fs_info,
5143                                     struct btrfs_space_info *space_info,
5144                                     u64 orig_bytes,
5145                                     enum btrfs_reserve_flush_enum flush,
5146                                     bool system_chunk)
5147 {
5148         struct reserve_ticket ticket;
5149         u64 used;
5150         u64 reclaim_bytes = 0;
5151         int ret = 0;
5152
5153         ASSERT(orig_bytes);
5154         ASSERT(!current->journal_info || flush != BTRFS_RESERVE_FLUSH_ALL);
5155
5156         spin_lock(&space_info->lock);
5157         ret = -ENOSPC;
5158         used = btrfs_space_info_used(space_info, true);
5159
5160         /*
5161          * If we have enough space then hooray, make our reservation and carry
5162          * on.  If not see if we can overcommit, and if we can, hooray carry on.
5163          * If not things get more complicated.
5164          */
5165         if (used + orig_bytes <= space_info->total_bytes) {
5166                 update_bytes_may_use(space_info, orig_bytes);
5167                 trace_btrfs_space_reservation(fs_info, "space_info",
5168                                               space_info->flags, orig_bytes, 1);
5169                 ret = 0;
5170         } else if (can_overcommit(fs_info, space_info, orig_bytes, flush,
5171                                   system_chunk)) {
5172                 update_bytes_may_use(space_info, orig_bytes);
5173                 trace_btrfs_space_reservation(fs_info, "space_info",
5174                                               space_info->flags, orig_bytes, 1);
5175                 ret = 0;
5176         }
5177
5178         /*
5179          * If we couldn't make a reservation then setup our reservation ticket
5180          * and kick the async worker if it's not already running.
5181          *
5182          * If we are a priority flusher then we just need to add our ticket to
5183          * the list and we will do our own flushing further down.
5184          */
5185         if (ret && flush != BTRFS_RESERVE_NO_FLUSH) {
5186                 ticket.orig_bytes = orig_bytes;
5187                 ticket.bytes = orig_bytes;
5188                 ticket.error = 0;
5189                 init_waitqueue_head(&ticket.wait);
5190                 if (flush == BTRFS_RESERVE_FLUSH_ALL) {
5191                         list_add_tail(&ticket.list, &space_info->tickets);
5192                         if (!space_info->flush) {
5193                                 space_info->flush = 1;
5194                                 trace_btrfs_trigger_flush(fs_info,
5195                                                           space_info->flags,
5196                                                           orig_bytes, flush,
5197                                                           "enospc");
5198                                 queue_work(system_unbound_wq,
5199                                            &fs_info->async_reclaim_work);
5200                         }
5201                 } else {
5202                         list_add_tail(&ticket.list,
5203                                       &space_info->priority_tickets);
5204                 }
5205         } else if (!ret && space_info->flags & BTRFS_BLOCK_GROUP_METADATA) {
5206                 used += orig_bytes;
5207                 /*
5208                  * We will do the space reservation dance during log replay,
5209                  * which means we won't have fs_info->fs_root set, so don't do
5210                  * the async reclaim as we will panic.
5211                  */
5212                 if (!test_bit(BTRFS_FS_LOG_RECOVERING, &fs_info->flags) &&
5213                     need_do_async_reclaim(fs_info, space_info,
5214                                           used, system_chunk) &&
5215                     !work_busy(&fs_info->async_reclaim_work)) {
5216                         trace_btrfs_trigger_flush(fs_info, space_info->flags,
5217                                                   orig_bytes, flush, "preempt");
5218                         queue_work(system_unbound_wq,
5219                                    &fs_info->async_reclaim_work);
5220                 }
5221         }
5222         spin_unlock(&space_info->lock);
5223         if (!ret || flush == BTRFS_RESERVE_NO_FLUSH)
5224                 return ret;
5225
5226         if (flush == BTRFS_RESERVE_FLUSH_ALL)
5227                 return wait_reserve_ticket(fs_info, space_info, &ticket);
5228
5229         ret = 0;
5230         priority_reclaim_metadata_space(fs_info, space_info, &ticket);
5231         spin_lock(&space_info->lock);
5232         if (ticket.bytes) {
5233                 if (ticket.bytes < orig_bytes)
5234                         reclaim_bytes = orig_bytes - ticket.bytes;
5235                 list_del_init(&ticket.list);
5236                 ret = -ENOSPC;
5237         }
5238         spin_unlock(&space_info->lock);
5239
5240         if (reclaim_bytes)
5241                 space_info_add_old_bytes(fs_info, space_info, reclaim_bytes);
5242         ASSERT(list_empty(&ticket.list));
5243         return ret;
5244 }
5245
5246 /**
5247  * reserve_metadata_bytes - try to reserve bytes from the block_rsv's space
5248  * @root - the root we're allocating for
5249  * @block_rsv - the block_rsv we're allocating for
5250  * @orig_bytes - the number of bytes we want
5251  * @flush - whether or not we can flush to make our reservation
5252  *
5253  * This will reserve orig_bytes number of bytes from the space info associated
5254  * with the block_rsv.  If there is not enough space it will make an attempt to
5255  * flush out space to make room.  It will do this by flushing delalloc if
5256  * possible or committing the transaction.  If flush is 0 then no attempts to
5257  * regain reservations will be made and this will fail if there is not enough
5258  * space already.
5259  */
5260 static int reserve_metadata_bytes(struct btrfs_root *root,
5261                                   struct btrfs_block_rsv *block_rsv,
5262                                   u64 orig_bytes,
5263                                   enum btrfs_reserve_flush_enum flush)
5264 {
5265         struct btrfs_fs_info *fs_info = root->fs_info;
5266         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
5267         int ret;
5268         bool system_chunk = (root == fs_info->chunk_root);
5269
5270         ret = __reserve_metadata_bytes(fs_info, block_rsv->space_info,
5271                                        orig_bytes, flush, system_chunk);
5272         if (ret == -ENOSPC &&
5273             unlikely(root->orphan_cleanup_state == ORPHAN_CLEANUP_STARTED)) {
5274                 if (block_rsv != global_rsv &&
5275                     !block_rsv_use_bytes(global_rsv, orig_bytes))
5276                         ret = 0;
5277         }
5278         if (ret == -ENOSPC) {
5279                 trace_btrfs_space_reservation(fs_info, "space_info:enospc",
5280                                               block_rsv->space_info->flags,
5281                                               orig_bytes, 1);
5282
5283                 if (btrfs_test_opt(fs_info, ENOSPC_DEBUG))
5284                         dump_space_info(fs_info, block_rsv->space_info,
5285                                         orig_bytes, 0);
5286         }
5287         return ret;
5288 }
5289
5290 static struct btrfs_block_rsv *get_block_rsv(
5291                                         const struct btrfs_trans_handle *trans,
5292                                         const struct btrfs_root *root)
5293 {
5294         struct btrfs_fs_info *fs_info = root->fs_info;
5295         struct btrfs_block_rsv *block_rsv = NULL;
5296
5297         if (test_bit(BTRFS_ROOT_REF_COWS, &root->state) ||
5298             (root == fs_info->csum_root && trans->adding_csums) ||
5299             (root == fs_info->uuid_root))
5300                 block_rsv = trans->block_rsv;
5301
5302         if (!block_rsv)
5303                 block_rsv = root->block_rsv;
5304
5305         if (!block_rsv)
5306                 block_rsv = &fs_info->empty_block_rsv;
5307
5308         return block_rsv;
5309 }
5310
5311 static int block_rsv_use_bytes(struct btrfs_block_rsv *block_rsv,
5312                                u64 num_bytes)
5313 {
5314         int ret = -ENOSPC;
5315         spin_lock(&block_rsv->lock);
5316         if (block_rsv->reserved >= num_bytes) {
5317                 block_rsv->reserved -= num_bytes;
5318                 if (block_rsv->reserved < block_rsv->size)
5319                         block_rsv->full = 0;
5320                 ret = 0;
5321         }
5322         spin_unlock(&block_rsv->lock);
5323         return ret;
5324 }
5325
5326 static void block_rsv_add_bytes(struct btrfs_block_rsv *block_rsv,
5327                                 u64 num_bytes, bool update_size)
5328 {
5329         spin_lock(&block_rsv->lock);
5330         block_rsv->reserved += num_bytes;
5331         if (update_size)
5332                 block_rsv->size += num_bytes;
5333         else if (block_rsv->reserved >= block_rsv->size)
5334                 block_rsv->full = 1;
5335         spin_unlock(&block_rsv->lock);
5336 }
5337
5338 int btrfs_cond_migrate_bytes(struct btrfs_fs_info *fs_info,
5339                              struct btrfs_block_rsv *dest, u64 num_bytes,
5340                              int min_factor)
5341 {
5342         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
5343         u64 min_bytes;
5344
5345         if (global_rsv->space_info != dest->space_info)
5346                 return -ENOSPC;
5347
5348         spin_lock(&global_rsv->lock);
5349         min_bytes = div_factor(global_rsv->size, min_factor);
5350         if (global_rsv->reserved < min_bytes + num_bytes) {
5351                 spin_unlock(&global_rsv->lock);
5352                 return -ENOSPC;
5353         }
5354         global_rsv->reserved -= num_bytes;
5355         if (global_rsv->reserved < global_rsv->size)
5356                 global_rsv->full = 0;
5357         spin_unlock(&global_rsv->lock);
5358
5359         block_rsv_add_bytes(dest, num_bytes, true);
5360         return 0;
5361 }
5362
5363 /**
5364  * btrfs_migrate_to_delayed_refs_rsv - transfer bytes to our delayed refs rsv.
5365  * @fs_info - the fs info for our fs.
5366  * @src - the source block rsv to transfer from.
5367  * @num_bytes - the number of bytes to transfer.
5368  *
5369  * This transfers up to the num_bytes amount from the src rsv to the
5370  * delayed_refs_rsv.  Any extra bytes are returned to the space info.
5371  */
5372 void btrfs_migrate_to_delayed_refs_rsv(struct btrfs_fs_info *fs_info,
5373                                        struct btrfs_block_rsv *src,
5374                                        u64 num_bytes)
5375 {
5376         struct btrfs_block_rsv *delayed_refs_rsv = &fs_info->delayed_refs_rsv;
5377         u64 to_free = 0;
5378
5379         spin_lock(&src->lock);
5380         src->reserved -= num_bytes;
5381         src->size -= num_bytes;
5382         spin_unlock(&src->lock);
5383
5384         spin_lock(&delayed_refs_rsv->lock);
5385         if (delayed_refs_rsv->size > delayed_refs_rsv->reserved) {
5386                 u64 delta = delayed_refs_rsv->size -
5387                         delayed_refs_rsv->reserved;
5388                 if (num_bytes > delta) {
5389                         to_free = num_bytes - delta;
5390                         num_bytes = delta;
5391                 }
5392         } else {
5393                 to_free = num_bytes;
5394                 num_bytes = 0;
5395         }
5396
5397         if (num_bytes)
5398                 delayed_refs_rsv->reserved += num_bytes;
5399         if (delayed_refs_rsv->reserved >= delayed_refs_rsv->size)
5400                 delayed_refs_rsv->full = 1;
5401         spin_unlock(&delayed_refs_rsv->lock);
5402
5403         if (num_bytes)
5404                 trace_btrfs_space_reservation(fs_info, "delayed_refs_rsv",
5405                                               0, num_bytes, 1);
5406         if (to_free)
5407                 space_info_add_old_bytes(fs_info, delayed_refs_rsv->space_info,
5408                                          to_free);
5409 }
5410
5411 /**
5412  * btrfs_delayed_refs_rsv_refill - refill based on our delayed refs usage.
5413  * @fs_info - the fs_info for our fs.
5414  * @flush - control how we can flush for this reservation.
5415  *
5416  * This will refill the delayed block_rsv up to 1 items size worth of space and
5417  * will return -ENOSPC if we can't make the reservation.
5418  */
5419 int btrfs_delayed_refs_rsv_refill(struct btrfs_fs_info *fs_info,
5420                                   enum btrfs_reserve_flush_enum flush)
5421 {
5422         struct btrfs_block_rsv *block_rsv = &fs_info->delayed_refs_rsv;
5423         u64 limit = btrfs_calc_trans_metadata_size(fs_info, 1);
5424         u64 num_bytes = 0;
5425         int ret = -ENOSPC;
5426
5427         spin_lock(&block_rsv->lock);
5428         if (block_rsv->reserved < block_rsv->size) {
5429                 num_bytes = block_rsv->size - block_rsv->reserved;
5430                 num_bytes = min(num_bytes, limit);
5431         }
5432         spin_unlock(&block_rsv->lock);
5433
5434         if (!num_bytes)
5435                 return 0;
5436
5437         ret = reserve_metadata_bytes(fs_info->extent_root, block_rsv,
5438                                      num_bytes, flush);
5439         if (ret)
5440                 return ret;
5441         block_rsv_add_bytes(block_rsv, num_bytes, 0);
5442         trace_btrfs_space_reservation(fs_info, "delayed_refs_rsv",
5443                                       0, num_bytes, 1);
5444         return 0;
5445 }
5446
5447 /*
5448  * This is for space we already have accounted in space_info->bytes_may_use, so
5449  * basically when we're returning space from block_rsv's.
5450  */
5451 static void space_info_add_old_bytes(struct btrfs_fs_info *fs_info,
5452                                      struct btrfs_space_info *space_info,
5453                                      u64 num_bytes)
5454 {
5455         struct reserve_ticket *ticket;
5456         struct list_head *head;
5457         u64 used;
5458         enum btrfs_reserve_flush_enum flush = BTRFS_RESERVE_NO_FLUSH;
5459         bool check_overcommit = false;
5460
5461         spin_lock(&space_info->lock);
5462         head = &space_info->priority_tickets;
5463
5464         /*
5465          * If we are over our limit then we need to check and see if we can
5466          * overcommit, and if we can't then we just need to free up our space
5467          * and not satisfy any requests.
5468          */
5469         used = btrfs_space_info_used(space_info, true);
5470         if (used - num_bytes >= space_info->total_bytes)
5471                 check_overcommit = true;
5472 again:
5473         while (!list_empty(head) && num_bytes) {
5474                 ticket = list_first_entry(head, struct reserve_ticket,
5475                                           list);
5476                 /*
5477                  * We use 0 bytes because this space is already reserved, so
5478                  * adding the ticket space would be a double count.
5479                  */
5480                 if (check_overcommit &&
5481                     !can_overcommit(fs_info, space_info, 0, flush, false))
5482                         break;
5483                 if (num_bytes >= ticket->bytes) {
5484                         list_del_init(&ticket->list);
5485                         num_bytes -= ticket->bytes;
5486                         ticket->bytes = 0;
5487                         space_info->tickets_id++;
5488                         wake_up(&ticket->wait);
5489                 } else {
5490                         ticket->bytes -= num_bytes;
5491                         num_bytes = 0;
5492                 }
5493         }
5494
5495         if (num_bytes && head == &space_info->priority_tickets) {
5496                 head = &space_info->tickets;
5497                 flush = BTRFS_RESERVE_FLUSH_ALL;
5498                 goto again;
5499         }
5500         update_bytes_may_use(space_info, -num_bytes);
5501         trace_btrfs_space_reservation(fs_info, "space_info",
5502                                       space_info->flags, num_bytes, 0);
5503         spin_unlock(&space_info->lock);
5504 }
5505
5506 /*
5507  * This is for newly allocated space that isn't accounted in
5508  * space_info->bytes_may_use yet.  So if we allocate a chunk or unpin an extent
5509  * we use this helper.
5510  */
5511 static void space_info_add_new_bytes(struct btrfs_fs_info *fs_info,
5512                                      struct btrfs_space_info *space_info,
5513                                      u64 num_bytes)
5514 {
5515         struct reserve_ticket *ticket;
5516         struct list_head *head = &space_info->priority_tickets;
5517
5518 again:
5519         while (!list_empty(head) && num_bytes) {
5520                 ticket = list_first_entry(head, struct reserve_ticket,
5521                                           list);
5522                 if (num_bytes >= ticket->bytes) {
5523                         trace_btrfs_space_reservation(fs_info, "space_info",
5524                                                       space_info->flags,
5525                                                       ticket->bytes, 1);
5526                         list_del_init(&ticket->list);
5527                         num_bytes -= ticket->bytes;
5528                         update_bytes_may_use(space_info, ticket->bytes);
5529                         ticket->bytes = 0;
5530                         space_info->tickets_id++;
5531                         wake_up(&ticket->wait);
5532                 } else {
5533                         trace_btrfs_space_reservation(fs_info, "space_info",
5534                                                       space_info->flags,
5535                                                       num_bytes, 1);
5536                         update_bytes_may_use(space_info, num_bytes);
5537                         ticket->bytes -= num_bytes;
5538                         num_bytes = 0;
5539                 }
5540         }
5541
5542         if (num_bytes && head == &space_info->priority_tickets) {
5543                 head = &space_info->tickets;
5544                 goto again;
5545         }
5546 }
5547
5548 static u64 block_rsv_release_bytes(struct btrfs_fs_info *fs_info,
5549                                     struct btrfs_block_rsv *block_rsv,
5550                                     struct btrfs_block_rsv *dest, u64 num_bytes,
5551                                     u64 *qgroup_to_release_ret)
5552 {
5553         struct btrfs_space_info *space_info = block_rsv->space_info;
5554         u64 qgroup_to_release = 0;
5555         u64 ret;
5556
5557         spin_lock(&block_rsv->lock);
5558         if (num_bytes == (u64)-1) {
5559                 num_bytes = block_rsv->size;
5560                 qgroup_to_release = block_rsv->qgroup_rsv_size;
5561         }
5562         block_rsv->size -= num_bytes;
5563         if (block_rsv->reserved >= block_rsv->size) {
5564                 num_bytes = block_rsv->reserved - block_rsv->size;
5565                 block_rsv->reserved = block_rsv->size;
5566                 block_rsv->full = 1;
5567         } else {
5568                 num_bytes = 0;
5569         }
5570         if (block_rsv->qgroup_rsv_reserved >= block_rsv->qgroup_rsv_size) {
5571                 qgroup_to_release = block_rsv->qgroup_rsv_reserved -
5572                                     block_rsv->qgroup_rsv_size;
5573                 block_rsv->qgroup_rsv_reserved = block_rsv->qgroup_rsv_size;
5574         } else {
5575                 qgroup_to_release = 0;
5576         }
5577         spin_unlock(&block_rsv->lock);
5578
5579         ret = num_bytes;
5580         if (num_bytes > 0) {
5581                 if (dest) {
5582                         spin_lock(&dest->lock);
5583                         if (!dest->full) {
5584                                 u64 bytes_to_add;
5585
5586                                 bytes_to_add = dest->size - dest->reserved;
5587                                 bytes_to_add = min(num_bytes, bytes_to_add);
5588                                 dest->reserved += bytes_to_add;
5589                                 if (dest->reserved >= dest->size)
5590                                         dest->full = 1;
5591                                 num_bytes -= bytes_to_add;
5592                         }
5593                         spin_unlock(&dest->lock);
5594                 }
5595                 if (num_bytes)
5596                         space_info_add_old_bytes(fs_info, space_info,
5597                                                  num_bytes);
5598         }
5599         if (qgroup_to_release_ret)
5600                 *qgroup_to_release_ret = qgroup_to_release;
5601         return ret;
5602 }
5603
5604 int btrfs_block_rsv_migrate(struct btrfs_block_rsv *src,
5605                             struct btrfs_block_rsv *dst, u64 num_bytes,
5606                             bool update_size)
5607 {
5608         int ret;
5609
5610         ret = block_rsv_use_bytes(src, num_bytes);
5611         if (ret)
5612                 return ret;
5613
5614         block_rsv_add_bytes(dst, num_bytes, update_size);
5615         return 0;
5616 }
5617
5618 void btrfs_init_block_rsv(struct btrfs_block_rsv *rsv, unsigned short type)
5619 {
5620         memset(rsv, 0, sizeof(*rsv));
5621         spin_lock_init(&rsv->lock);
5622         rsv->type = type;
5623 }
5624
5625 void btrfs_init_metadata_block_rsv(struct btrfs_fs_info *fs_info,
5626                                    struct btrfs_block_rsv *rsv,
5627                                    unsigned short type)
5628 {
5629         btrfs_init_block_rsv(rsv, type);
5630         rsv->space_info = __find_space_info(fs_info,
5631                                             BTRFS_BLOCK_GROUP_METADATA);
5632 }
5633
5634 struct btrfs_block_rsv *btrfs_alloc_block_rsv(struct btrfs_fs_info *fs_info,
5635                                               unsigned short type)
5636 {
5637         struct btrfs_block_rsv *block_rsv;
5638
5639         block_rsv = kmalloc(sizeof(*block_rsv), GFP_NOFS);
5640         if (!block_rsv)
5641                 return NULL;
5642
5643         btrfs_init_metadata_block_rsv(fs_info, block_rsv, type);
5644         return block_rsv;
5645 }
5646
5647 void btrfs_free_block_rsv(struct btrfs_fs_info *fs_info,
5648                           struct btrfs_block_rsv *rsv)
5649 {
5650         if (!rsv)
5651                 return;
5652         btrfs_block_rsv_release(fs_info, rsv, (u64)-1);
5653         kfree(rsv);
5654 }
5655
5656 int btrfs_block_rsv_add(struct btrfs_root *root,
5657                         struct btrfs_block_rsv *block_rsv, u64 num_bytes,
5658                         enum btrfs_reserve_flush_enum flush)
5659 {
5660         int ret;
5661
5662         if (num_bytes == 0)
5663                 return 0;
5664
5665         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
5666         if (!ret)
5667                 block_rsv_add_bytes(block_rsv, num_bytes, true);
5668
5669         return ret;
5670 }
5671
5672 int btrfs_block_rsv_check(struct btrfs_block_rsv *block_rsv, int min_factor)
5673 {
5674         u64 num_bytes = 0;
5675         int ret = -ENOSPC;
5676
5677         if (!block_rsv)
5678                 return 0;
5679
5680         spin_lock(&block_rsv->lock);
5681         num_bytes = div_factor(block_rsv->size, min_factor);
5682         if (block_rsv->reserved >= num_bytes)
5683                 ret = 0;
5684         spin_unlock(&block_rsv->lock);
5685
5686         return ret;
5687 }
5688
5689 int btrfs_block_rsv_refill(struct btrfs_root *root,
5690                            struct btrfs_block_rsv *block_rsv, u64 min_reserved,
5691                            enum btrfs_reserve_flush_enum flush)
5692 {
5693         u64 num_bytes = 0;
5694         int ret = -ENOSPC;
5695
5696         if (!block_rsv)
5697                 return 0;
5698
5699         spin_lock(&block_rsv->lock);
5700         num_bytes = min_reserved;
5701         if (block_rsv->reserved >= num_bytes)
5702                 ret = 0;
5703         else
5704                 num_bytes -= block_rsv->reserved;
5705         spin_unlock(&block_rsv->lock);
5706
5707         if (!ret)
5708                 return 0;
5709
5710         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
5711         if (!ret) {
5712                 block_rsv_add_bytes(block_rsv, num_bytes, false);
5713                 return 0;
5714         }
5715
5716         return ret;
5717 }
5718
5719 static u64 __btrfs_block_rsv_release(struct btrfs_fs_info *fs_info,
5720                                      struct btrfs_block_rsv *block_rsv,
5721                                      u64 num_bytes, u64 *qgroup_to_release)
5722 {
5723         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
5724         struct btrfs_block_rsv *delayed_rsv = &fs_info->delayed_refs_rsv;
5725         struct btrfs_block_rsv *target = delayed_rsv;
5726
5727         if (target->full || target == block_rsv)
5728                 target = global_rsv;
5729
5730         if (block_rsv->space_info != target->space_info)
5731                 target = NULL;
5732
5733         return block_rsv_release_bytes(fs_info, block_rsv, target, num_bytes,
5734                                        qgroup_to_release);
5735 }
5736
5737 void btrfs_block_rsv_release(struct btrfs_fs_info *fs_info,
5738                              struct btrfs_block_rsv *block_rsv,
5739                              u64 num_bytes)
5740 {
5741         __btrfs_block_rsv_release(fs_info, block_rsv, num_bytes, NULL);
5742 }
5743
5744 /**
5745  * btrfs_inode_rsv_release - release any excessive reservation.
5746  * @inode - the inode we need to release from.
5747  * @qgroup_free - free or convert qgroup meta.
5748  *   Unlike normal operation, qgroup meta reservation needs to know if we are
5749  *   freeing qgroup reservation or just converting it into per-trans.  Normally
5750  *   @qgroup_free is true for error handling, and false for normal release.
5751  *
5752  * This is the same as btrfs_block_rsv_release, except that it handles the
5753  * tracepoint for the reservation.
5754  */
5755 static void btrfs_inode_rsv_release(struct btrfs_inode *inode, bool qgroup_free)
5756 {
5757         struct btrfs_fs_info *fs_info = inode->root->fs_info;
5758         struct btrfs_block_rsv *block_rsv = &inode->block_rsv;
5759         u64 released = 0;
5760         u64 qgroup_to_release = 0;
5761
5762         /*
5763          * Since we statically set the block_rsv->size we just want to say we
5764          * are releasing 0 bytes, and then we'll just get the reservation over
5765          * the size free'd.
5766          */
5767         released = __btrfs_block_rsv_release(fs_info, block_rsv, 0,
5768                                              &qgroup_to_release);
5769         if (released > 0)
5770                 trace_btrfs_space_reservation(fs_info, "delalloc",
5771                                               btrfs_ino(inode), released, 0);
5772         if (qgroup_free)
5773                 btrfs_qgroup_free_meta_prealloc(inode->root, qgroup_to_release);
5774         else
5775                 btrfs_qgroup_convert_reserved_meta(inode->root,
5776                                                    qgroup_to_release);
5777 }
5778
5779 /**
5780  * btrfs_delayed_refs_rsv_release - release a ref head's reservation.
5781  * @fs_info - the fs_info for our fs.
5782  * @nr - the number of items to drop.
5783  *
5784  * This drops the delayed ref head's count from the delayed refs rsv and frees
5785  * any excess reservation we had.
5786  */
5787 void btrfs_delayed_refs_rsv_release(struct btrfs_fs_info *fs_info, int nr)
5788 {
5789         struct btrfs_block_rsv *block_rsv = &fs_info->delayed_refs_rsv;
5790         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
5791         u64 num_bytes = btrfs_calc_trans_metadata_size(fs_info, nr);
5792         u64 released = 0;
5793
5794         released = block_rsv_release_bytes(fs_info, block_rsv, global_rsv,
5795                                            num_bytes, NULL);
5796         if (released)
5797                 trace_btrfs_space_reservation(fs_info, "delayed_refs_rsv",
5798                                               0, released, 0);
5799 }
5800
5801 static void update_global_block_rsv(struct btrfs_fs_info *fs_info)
5802 {
5803         struct btrfs_block_rsv *block_rsv = &fs_info->global_block_rsv;
5804         struct btrfs_space_info *sinfo = block_rsv->space_info;
5805         u64 num_bytes;
5806
5807         /*
5808          * The global block rsv is based on the size of the extent tree, the
5809          * checksum tree and the root tree.  If the fs is empty we want to set
5810          * it to a minimal amount for safety.
5811          */
5812         num_bytes = btrfs_root_used(&fs_info->extent_root->root_item) +
5813                 btrfs_root_used(&fs_info->csum_root->root_item) +
5814                 btrfs_root_used(&fs_info->tree_root->root_item);
5815         num_bytes = max_t(u64, num_bytes, SZ_16M);
5816
5817         spin_lock(&sinfo->lock);
5818         spin_lock(&block_rsv->lock);
5819
5820         block_rsv->size = min_t(u64, num_bytes, SZ_512M);
5821
5822         if (block_rsv->reserved < block_rsv->size) {
5823                 num_bytes = btrfs_space_info_used(sinfo, true);
5824                 if (sinfo->total_bytes > num_bytes) {
5825                         num_bytes = sinfo->total_bytes - num_bytes;
5826                         num_bytes = min(num_bytes,
5827                                         block_rsv->size - block_rsv->reserved);
5828                         block_rsv->reserved += num_bytes;
5829                         update_bytes_may_use(sinfo, num_bytes);
5830                         trace_btrfs_space_reservation(fs_info, "space_info",
5831                                                       sinfo->flags, num_bytes,
5832                                                       1);
5833                 }
5834         } else if (block_rsv->reserved > block_rsv->size) {
5835                 num_bytes = block_rsv->reserved - block_rsv->size;
5836                 update_bytes_may_use(sinfo, -num_bytes);
5837                 trace_btrfs_space_reservation(fs_info, "space_info",
5838                                       sinfo->flags, num_bytes, 0);
5839                 block_rsv->reserved = block_rsv->size;
5840         }
5841
5842         if (block_rsv->reserved == block_rsv->size)
5843                 block_rsv->full = 1;
5844         else
5845                 block_rsv->full = 0;
5846
5847         spin_unlock(&block_rsv->lock);
5848         spin_unlock(&sinfo->lock);
5849 }
5850
5851 static void init_global_block_rsv(struct btrfs_fs_info *fs_info)
5852 {
5853         struct btrfs_space_info *space_info;
5854
5855         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
5856         fs_info->chunk_block_rsv.space_info = space_info;
5857
5858         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
5859         fs_info->global_block_rsv.space_info = space_info;
5860         fs_info->trans_block_rsv.space_info = space_info;
5861         fs_info->empty_block_rsv.space_info = space_info;
5862         fs_info->delayed_block_rsv.space_info = space_info;
5863         fs_info->delayed_refs_rsv.space_info = space_info;
5864
5865         fs_info->extent_root->block_rsv = &fs_info->delayed_refs_rsv;
5866         fs_info->csum_root->block_rsv = &fs_info->delayed_refs_rsv;
5867         fs_info->dev_root->block_rsv = &fs_info->global_block_rsv;
5868         fs_info->tree_root->block_rsv = &fs_info->global_block_rsv;
5869         if (fs_info->quota_root)
5870                 fs_info->quota_root->block_rsv = &fs_info->global_block_rsv;
5871         fs_info->chunk_root->block_rsv = &fs_info->chunk_block_rsv;
5872
5873         update_global_block_rsv(fs_info);
5874 }
5875
5876 static void release_global_block_rsv(struct btrfs_fs_info *fs_info)
5877 {
5878         block_rsv_release_bytes(fs_info, &fs_info->global_block_rsv, NULL,
5879                                 (u64)-1, NULL);
5880         WARN_ON(fs_info->trans_block_rsv.size > 0);
5881         WARN_ON(fs_info->trans_block_rsv.reserved > 0);
5882         WARN_ON(fs_info->chunk_block_rsv.size > 0);
5883         WARN_ON(fs_info->chunk_block_rsv.reserved > 0);
5884         WARN_ON(fs_info->delayed_block_rsv.size > 0);
5885         WARN_ON(fs_info->delayed_block_rsv.reserved > 0);
5886         WARN_ON(fs_info->delayed_refs_rsv.reserved > 0);
5887         WARN_ON(fs_info->delayed_refs_rsv.size > 0);
5888 }
5889
5890 /*
5891  * btrfs_update_delayed_refs_rsv - adjust the size of the delayed refs rsv
5892  * @trans - the trans that may have generated delayed refs
5893  *
5894  * This is to be called anytime we may have adjusted trans->delayed_ref_updates,
5895  * it'll calculate the additional size and add it to the delayed_refs_rsv.
5896  */
5897 void btrfs_update_delayed_refs_rsv(struct btrfs_trans_handle *trans)
5898 {
5899         struct btrfs_fs_info *fs_info = trans->fs_info;
5900         struct btrfs_block_rsv *delayed_rsv = &fs_info->delayed_refs_rsv;
5901         u64 num_bytes;
5902
5903         if (!trans->delayed_ref_updates)
5904                 return;
5905
5906         num_bytes = btrfs_calc_trans_metadata_size(fs_info,
5907                                                    trans->delayed_ref_updates);
5908         spin_lock(&delayed_rsv->lock);
5909         delayed_rsv->size += num_bytes;
5910         delayed_rsv->full = 0;
5911         spin_unlock(&delayed_rsv->lock);
5912         trans->delayed_ref_updates = 0;
5913 }
5914
5915 /*
5916  * To be called after all the new block groups attached to the transaction
5917  * handle have been created (btrfs_create_pending_block_groups()).
5918  */
5919 void btrfs_trans_release_chunk_metadata(struct btrfs_trans_handle *trans)
5920 {
5921         struct btrfs_fs_info *fs_info = trans->fs_info;
5922
5923         if (!trans->chunk_bytes_reserved)
5924                 return;
5925
5926         WARN_ON_ONCE(!list_empty(&trans->new_bgs));
5927
5928         block_rsv_release_bytes(fs_info, &fs_info->chunk_block_rsv, NULL,
5929                                 trans->chunk_bytes_reserved, NULL);
5930         trans->chunk_bytes_reserved = 0;
5931 }
5932
5933 /*
5934  * btrfs_subvolume_reserve_metadata() - reserve space for subvolume operation
5935  * root: the root of the parent directory
5936  * rsv: block reservation
5937  * items: the number of items that we need do reservation
5938  * use_global_rsv: allow fallback to the global block reservation
5939  *
5940  * This function is used to reserve the space for snapshot/subvolume
5941  * creation and deletion. Those operations are different with the
5942  * common file/directory operations, they change two fs/file trees
5943  * and root tree, the number of items that the qgroup reserves is
5944  * different with the free space reservation. So we can not use
5945  * the space reservation mechanism in start_transaction().
5946  */
5947 int btrfs_subvolume_reserve_metadata(struct btrfs_root *root,
5948                                      struct btrfs_block_rsv *rsv, int items,
5949                                      bool use_global_rsv)
5950 {
5951         u64 qgroup_num_bytes = 0;
5952         u64 num_bytes;
5953         int ret;
5954         struct btrfs_fs_info *fs_info = root->fs_info;
5955         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
5956
5957         if (test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags)) {
5958                 /* One for parent inode, two for dir entries */
5959                 qgroup_num_bytes = 3 * fs_info->nodesize;
5960                 ret = btrfs_qgroup_reserve_meta_prealloc(root,
5961                                 qgroup_num_bytes, true);
5962                 if (ret)
5963                         return ret;
5964         }
5965
5966         num_bytes = btrfs_calc_trans_metadata_size(fs_info, items);
5967         rsv->space_info = __find_space_info(fs_info,
5968                                             BTRFS_BLOCK_GROUP_METADATA);
5969         ret = btrfs_block_rsv_add(root, rsv, num_bytes,
5970                                   BTRFS_RESERVE_FLUSH_ALL);
5971
5972         if (ret == -ENOSPC && use_global_rsv)
5973                 ret = btrfs_block_rsv_migrate(global_rsv, rsv, num_bytes, true);
5974
5975         if (ret && qgroup_num_bytes)
5976                 btrfs_qgroup_free_meta_prealloc(root, qgroup_num_bytes);
5977
5978         return ret;
5979 }
5980
5981 void btrfs_subvolume_release_metadata(struct btrfs_fs_info *fs_info,
5982                                       struct btrfs_block_rsv *rsv)
5983 {
5984         btrfs_block_rsv_release(fs_info, rsv, (u64)-1);
5985 }
5986
5987 static void btrfs_calculate_inode_block_rsv_size(struct btrfs_fs_info *fs_info,
5988                                                  struct btrfs_inode *inode)
5989 {
5990         struct btrfs_block_rsv *block_rsv = &inode->block_rsv;
5991         u64 reserve_size = 0;
5992         u64 qgroup_rsv_size = 0;
5993         u64 csum_leaves;
5994         unsigned outstanding_extents;
5995
5996         lockdep_assert_held(&inode->lock);
5997         outstanding_extents = inode->outstanding_extents;
5998         if (outstanding_extents)
5999                 reserve_size = btrfs_calc_trans_metadata_size(fs_info,
6000                                                 outstanding_extents + 1);
6001         csum_leaves = btrfs_csum_bytes_to_leaves(fs_info,
6002                                                  inode->csum_bytes);
6003         reserve_size += btrfs_calc_trans_metadata_size(fs_info,
6004                                                        csum_leaves);
6005         /*
6006          * For qgroup rsv, the calculation is very simple:
6007          * account one nodesize for each outstanding extent
6008          *
6009          * This is overestimating in most cases.
6010          */
6011         qgroup_rsv_size = (u64)outstanding_extents * fs_info->nodesize;
6012
6013         spin_lock(&block_rsv->lock);
6014         block_rsv->size = reserve_size;
6015         block_rsv->qgroup_rsv_size = qgroup_rsv_size;
6016         spin_unlock(&block_rsv->lock);
6017 }
6018
6019 static void calc_inode_reservations(struct btrfs_fs_info *fs_info,
6020                                     u64 num_bytes, u64 *meta_reserve,
6021                                     u64 *qgroup_reserve)
6022 {
6023         u64 nr_extents = count_max_extents(num_bytes);
6024         u64 csum_leaves = btrfs_csum_bytes_to_leaves(fs_info, num_bytes);
6025
6026         /* We add one for the inode update at finish ordered time */
6027         *meta_reserve = btrfs_calc_trans_metadata_size(fs_info,
6028                                                 nr_extents + csum_leaves + 1);
6029         *qgroup_reserve = nr_extents * fs_info->nodesize;
6030 }
6031
6032 int btrfs_delalloc_reserve_metadata(struct btrfs_inode *inode, u64 num_bytes)
6033 {
6034         struct btrfs_root *root = inode->root;
6035         struct btrfs_fs_info *fs_info = root->fs_info;
6036         struct btrfs_block_rsv *block_rsv = &inode->block_rsv;
6037         u64 meta_reserve, qgroup_reserve;
6038         unsigned nr_extents;
6039         enum btrfs_reserve_flush_enum flush = BTRFS_RESERVE_FLUSH_ALL;
6040         int ret = 0;
6041         bool delalloc_lock = true;
6042
6043         /* If we are a free space inode we need to not flush since we will be in
6044          * the middle of a transaction commit.  We also don't need the delalloc
6045          * mutex since we won't race with anybody.  We need this mostly to make
6046          * lockdep shut its filthy mouth.
6047          *
6048          * If we have a transaction open (can happen if we call truncate_block
6049          * from truncate), then we need FLUSH_LIMIT so we don't deadlock.
6050          */
6051         if (btrfs_is_free_space_inode(inode)) {
6052                 flush = BTRFS_RESERVE_NO_FLUSH;
6053                 delalloc_lock = false;
6054         } else {
6055                 if (current->journal_info)
6056                         flush = BTRFS_RESERVE_FLUSH_LIMIT;
6057
6058                 if (btrfs_transaction_in_commit(fs_info))
6059                         schedule_timeout(1);
6060         }
6061
6062         if (delalloc_lock)
6063                 mutex_lock(&inode->delalloc_mutex);
6064
6065         num_bytes = ALIGN(num_bytes, fs_info->sectorsize);
6066
6067         /*
6068          * We always want to do it this way, every other way is wrong and ends
6069          * in tears.  Pre-reserving the amount we are going to add will always
6070          * be the right way, because otherwise if we have enough parallelism we
6071          * could end up with thousands of inodes all holding little bits of
6072          * reservations they were able to make previously and the only way to
6073          * reclaim that space is to ENOSPC out the operations and clear
6074          * everything out and try again, which is bad.  This way we just
6075          * over-reserve slightly, and clean up the mess when we are done.
6076          */
6077         calc_inode_reservations(fs_info, num_bytes, &meta_reserve,
6078                                 &qgroup_reserve);
6079         ret = btrfs_qgroup_reserve_meta_prealloc(root, qgroup_reserve, true);
6080         if (ret)
6081                 goto out_fail;
6082         ret = reserve_metadata_bytes(root, block_rsv, meta_reserve, flush);
6083         if (ret)
6084                 goto out_qgroup;
6085
6086         /*
6087          * Now we need to update our outstanding extents and csum bytes _first_
6088          * and then add the reservation to the block_rsv.  This keeps us from
6089          * racing with an ordered completion or some such that would think it
6090          * needs to free the reservation we just made.
6091          */
6092         spin_lock(&inode->lock);
6093         nr_extents = count_max_extents(num_bytes);
6094         btrfs_mod_outstanding_extents(inode, nr_extents);
6095         inode->csum_bytes += num_bytes;
6096         btrfs_calculate_inode_block_rsv_size(fs_info, inode);
6097         spin_unlock(&inode->lock);
6098
6099         /* Now we can safely add our space to our block rsv */
6100         block_rsv_add_bytes(block_rsv, meta_reserve, false);
6101         trace_btrfs_space_reservation(root->fs_info, "delalloc",
6102                                       btrfs_ino(inode), meta_reserve, 1);
6103
6104         spin_lock(&block_rsv->lock);
6105         block_rsv->qgroup_rsv_reserved += qgroup_reserve;
6106         spin_unlock(&block_rsv->lock);
6107
6108         if (delalloc_lock)
6109                 mutex_unlock(&inode->delalloc_mutex);
6110         return 0;
6111 out_qgroup:
6112         btrfs_qgroup_free_meta_prealloc(root, qgroup_reserve);
6113 out_fail:
6114         btrfs_inode_rsv_release(inode, true);
6115         if (delalloc_lock)
6116                 mutex_unlock(&inode->delalloc_mutex);
6117         return ret;
6118 }
6119
6120 /**
6121  * btrfs_delalloc_release_metadata - release a metadata reservation for an inode
6122  * @inode: the inode to release the reservation for.
6123  * @num_bytes: the number of bytes we are releasing.
6124  * @qgroup_free: free qgroup reservation or convert it to per-trans reservation
6125  *
6126  * This will release the metadata reservation for an inode.  This can be called
6127  * once we complete IO for a given set of bytes to release their metadata
6128  * reservations, or on error for the same reason.
6129  */
6130 void btrfs_delalloc_release_metadata(struct btrfs_inode *inode, u64 num_bytes,
6131                                      bool qgroup_free)
6132 {
6133         struct btrfs_fs_info *fs_info = inode->root->fs_info;
6134
6135         num_bytes = ALIGN(num_bytes, fs_info->sectorsize);
6136         spin_lock(&inode->lock);
6137         inode->csum_bytes -= num_bytes;
6138         btrfs_calculate_inode_block_rsv_size(fs_info, inode);
6139         spin_unlock(&inode->lock);
6140
6141         if (btrfs_is_testing(fs_info))
6142                 return;
6143
6144         btrfs_inode_rsv_release(inode, qgroup_free);
6145 }
6146
6147 /**
6148  * btrfs_delalloc_release_extents - release our outstanding_extents
6149  * @inode: the inode to balance the reservation for.
6150  * @num_bytes: the number of bytes we originally reserved with
6151  * @qgroup_free: do we need to free qgroup meta reservation or convert them.
6152  *
6153  * When we reserve space we increase outstanding_extents for the extents we may
6154  * add.  Once we've set the range as delalloc or created our ordered extents we
6155  * have outstanding_extents to track the real usage, so we use this to free our
6156  * temporarily tracked outstanding_extents.  This _must_ be used in conjunction
6157  * with btrfs_delalloc_reserve_metadata.
6158  */
6159 void btrfs_delalloc_release_extents(struct btrfs_inode *inode, u64 num_bytes,
6160                                     bool qgroup_free)
6161 {
6162         struct btrfs_fs_info *fs_info = inode->root->fs_info;
6163         unsigned num_extents;
6164
6165         spin_lock(&inode->lock);
6166         num_extents = count_max_extents(num_bytes);
6167         btrfs_mod_outstanding_extents(inode, -num_extents);
6168         btrfs_calculate_inode_block_rsv_size(fs_info, inode);
6169         spin_unlock(&inode->lock);
6170
6171         if (btrfs_is_testing(fs_info))
6172                 return;
6173
6174         btrfs_inode_rsv_release(inode, qgroup_free);
6175 }
6176
6177 /**
6178  * btrfs_delalloc_reserve_space - reserve data and metadata space for
6179  * delalloc
6180  * @inode: inode we're writing to
6181  * @start: start range we are writing to
6182  * @len: how long the range we are writing to
6183  * @reserved: mandatory parameter, record actually reserved qgroup ranges of
6184  *            current reservation.
6185  *
6186  * This will do the following things
6187  *
6188  * o reserve space in data space info for num bytes
6189  *   and reserve precious corresponding qgroup space
6190  *   (Done in check_data_free_space)
6191  *
6192  * o reserve space for metadata space, based on the number of outstanding
6193  *   extents and how much csums will be needed
6194  *   also reserve metadata space in a per root over-reserve method.
6195  * o add to the inodes->delalloc_bytes
6196  * o add it to the fs_info's delalloc inodes list.
6197  *   (Above 3 all done in delalloc_reserve_metadata)
6198  *
6199  * Return 0 for success
6200  * Return <0 for error(-ENOSPC or -EQUOT)
6201  */
6202 int btrfs_delalloc_reserve_space(struct inode *inode,
6203                         struct extent_changeset **reserved, u64 start, u64 len)
6204 {
6205         int ret;
6206
6207         ret = btrfs_check_data_free_space(inode, reserved, start, len);
6208         if (ret < 0)
6209                 return ret;
6210         ret = btrfs_delalloc_reserve_metadata(BTRFS_I(inode), len);
6211         if (ret < 0)
6212                 btrfs_free_reserved_data_space(inode, *reserved, start, len);
6213         return ret;
6214 }
6215
6216 /**
6217  * btrfs_delalloc_release_space - release data and metadata space for delalloc
6218  * @inode: inode we're releasing space for
6219  * @start: start position of the space already reserved
6220  * @len: the len of the space already reserved
6221  * @release_bytes: the len of the space we consumed or didn't use
6222  *
6223  * This function will release the metadata space that was not used and will
6224  * decrement ->delalloc_bytes and remove it from the fs_info delalloc_inodes
6225  * list if there are no delalloc bytes left.
6226  * Also it will handle the qgroup reserved space.
6227  */
6228 void btrfs_delalloc_release_space(struct inode *inode,
6229                                   struct extent_changeset *reserved,
6230                                   u64 start, u64 len, bool qgroup_free)
6231 {
6232         btrfs_delalloc_release_metadata(BTRFS_I(inode), len, qgroup_free);
6233         btrfs_free_reserved_data_space(inode, reserved, start, len);
6234 }
6235
6236 static int update_block_group(struct btrfs_trans_handle *trans,
6237                               u64 bytenr, u64 num_bytes, int alloc)
6238 {
6239         struct btrfs_fs_info *info = trans->fs_info;
6240         struct btrfs_block_group_cache *cache = NULL;
6241         u64 total = num_bytes;
6242         u64 old_val;
6243         u64 byte_in_group;
6244         int factor;
6245         int ret = 0;
6246
6247         /* block accounting for super block */
6248         spin_lock(&info->delalloc_root_lock);
6249         old_val = btrfs_super_bytes_used(info->super_copy);
6250         if (alloc)
6251                 old_val += num_bytes;
6252         else
6253                 old_val -= num_bytes;
6254         btrfs_set_super_bytes_used(info->super_copy, old_val);
6255         spin_unlock(&info->delalloc_root_lock);
6256
6257         while (total) {
6258                 cache = btrfs_lookup_block_group(info, bytenr);
6259                 if (!cache) {
6260                         ret = -ENOENT;
6261                         break;
6262                 }
6263                 factor = btrfs_bg_type_to_factor(cache->flags);
6264
6265                 /*
6266                  * If this block group has free space cache written out, we
6267                  * need to make sure to load it if we are removing space.  This
6268                  * is because we need the unpinning stage to actually add the
6269                  * space back to the block group, otherwise we will leak space.
6270                  */
6271                 if (!alloc && cache->cached == BTRFS_CACHE_NO)
6272                         cache_block_group(cache, 1);
6273
6274                 byte_in_group = bytenr - cache->key.objectid;
6275                 WARN_ON(byte_in_group > cache->key.offset);
6276
6277                 spin_lock(&cache->space_info->lock);
6278                 spin_lock(&cache->lock);
6279
6280                 if (btrfs_test_opt(info, SPACE_CACHE) &&
6281                     cache->disk_cache_state < BTRFS_DC_CLEAR)
6282                         cache->disk_cache_state = BTRFS_DC_CLEAR;
6283
6284                 old_val = btrfs_block_group_used(&cache->item);
6285                 num_bytes = min(total, cache->key.offset - byte_in_group);
6286                 if (alloc) {
6287                         old_val += num_bytes;
6288                         btrfs_set_block_group_used(&cache->item, old_val);
6289                         cache->reserved -= num_bytes;
6290                         cache->space_info->bytes_reserved -= num_bytes;
6291                         cache->space_info->bytes_used += num_bytes;
6292                         cache->space_info->disk_used += num_bytes * factor;
6293                         spin_unlock(&cache->lock);
6294                         spin_unlock(&cache->space_info->lock);
6295                 } else {
6296                         old_val -= num_bytes;
6297                         btrfs_set_block_group_used(&cache->item, old_val);
6298                         cache->pinned += num_bytes;
6299                         update_bytes_pinned(cache->space_info, num_bytes);
6300                         cache->space_info->bytes_used -= num_bytes;
6301                         cache->space_info->disk_used -= num_bytes * factor;
6302                         spin_unlock(&cache->lock);
6303                         spin_unlock(&cache->space_info->lock);
6304
6305                         trace_btrfs_space_reservation(info, "pinned",
6306                                                       cache->space_info->flags,
6307                                                       num_bytes, 1);
6308                         percpu_counter_add_batch(&cache->space_info->total_bytes_pinned,
6309                                            num_bytes,
6310                                            BTRFS_TOTAL_BYTES_PINNED_BATCH);
6311                         set_extent_dirty(info->pinned_extents,
6312                                          bytenr, bytenr + num_bytes - 1,
6313                                          GFP_NOFS | __GFP_NOFAIL);
6314                 }
6315
6316                 spin_lock(&trans->transaction->dirty_bgs_lock);
6317                 if (list_empty(&cache->dirty_list)) {
6318                         list_add_tail(&cache->dirty_list,
6319                                       &trans->transaction->dirty_bgs);
6320                         trans->delayed_ref_updates++;
6321                         btrfs_get_block_group(cache);
6322                 }
6323                 spin_unlock(&trans->transaction->dirty_bgs_lock);
6324
6325                 /*
6326                  * No longer have used bytes in this block group, queue it for
6327                  * deletion. We do this after adding the block group to the
6328                  * dirty list to avoid races between cleaner kthread and space
6329                  * cache writeout.
6330                  */
6331                 if (!alloc && old_val == 0)
6332                         btrfs_mark_bg_unused(cache);
6333
6334                 btrfs_put_block_group(cache);
6335                 total -= num_bytes;
6336                 bytenr += num_bytes;
6337         }
6338
6339         /* Modified block groups are accounted for in the delayed_refs_rsv. */
6340         btrfs_update_delayed_refs_rsv(trans);
6341         return ret;
6342 }
6343
6344 static u64 first_logical_byte(struct btrfs_fs_info *fs_info, u64 search_start)
6345 {
6346         struct btrfs_block_group_cache *cache;
6347         u64 bytenr;
6348
6349         spin_lock(&fs_info->block_group_cache_lock);
6350         bytenr = fs_info->first_logical_byte;
6351         spin_unlock(&fs_info->block_group_cache_lock);
6352
6353         if (bytenr < (u64)-1)
6354                 return bytenr;
6355
6356         cache = btrfs_lookup_first_block_group(fs_info, search_start);
6357         if (!cache)
6358                 return 0;
6359
6360         bytenr = cache->key.objectid;
6361         btrfs_put_block_group(cache);
6362
6363         return bytenr;
6364 }
6365
6366 static int pin_down_extent(struct btrfs_block_group_cache *cache,
6367                            u64 bytenr, u64 num_bytes, int reserved)
6368 {
6369         struct btrfs_fs_info *fs_info = cache->fs_info;
6370
6371         spin_lock(&cache->space_info->lock);
6372         spin_lock(&cache->lock);
6373         cache->pinned += num_bytes;
6374         update_bytes_pinned(cache->space_info, num_bytes);
6375         if (reserved) {
6376                 cache->reserved -= num_bytes;
6377                 cache->space_info->bytes_reserved -= num_bytes;
6378         }
6379         spin_unlock(&cache->lock);
6380         spin_unlock(&cache->space_info->lock);
6381
6382         trace_btrfs_space_reservation(fs_info, "pinned",
6383                                       cache->space_info->flags, num_bytes, 1);
6384         percpu_counter_add_batch(&cache->space_info->total_bytes_pinned,
6385                     num_bytes, BTRFS_TOTAL_BYTES_PINNED_BATCH);
6386         set_extent_dirty(fs_info->pinned_extents, bytenr,
6387                          bytenr + num_bytes - 1, GFP_NOFS | __GFP_NOFAIL);
6388         return 0;
6389 }
6390
6391 /*
6392  * this function must be called within transaction
6393  */
6394 int btrfs_pin_extent(struct btrfs_fs_info *fs_info,
6395                      u64 bytenr, u64 num_bytes, int reserved)
6396 {
6397         struct btrfs_block_group_cache *cache;
6398
6399         cache = btrfs_lookup_block_group(fs_info, bytenr);
6400         BUG_ON(!cache); /* Logic error */
6401
6402         pin_down_extent(cache, bytenr, num_bytes, reserved);
6403
6404         btrfs_put_block_group(cache);
6405         return 0;
6406 }
6407
6408 /*
6409  * this function must be called within transaction
6410  */
6411 int btrfs_pin_extent_for_log_replay(struct btrfs_fs_info *fs_info,
6412                                     u64 bytenr, u64 num_bytes)
6413 {
6414         struct btrfs_block_group_cache *cache;
6415         int ret;
6416
6417         cache = btrfs_lookup_block_group(fs_info, bytenr);
6418         if (!cache)
6419                 return -EINVAL;
6420
6421         /*
6422          * pull in the free space cache (if any) so that our pin
6423          * removes the free space from the cache.  We have load_only set
6424          * to one because the slow code to read in the free extents does check
6425          * the pinned extents.
6426          */
6427         cache_block_group(cache, 1);
6428
6429         pin_down_extent(cache, bytenr, num_bytes, 0);
6430
6431         /* remove us from the free space cache (if we're there at all) */
6432         ret = btrfs_remove_free_space(cache, bytenr, num_bytes);
6433         btrfs_put_block_group(cache);
6434         return ret;
6435 }
6436
6437 static int __exclude_logged_extent(struct btrfs_fs_info *fs_info,
6438                                    u64 start, u64 num_bytes)
6439 {
6440         int ret;
6441         struct btrfs_block_group_cache *block_group;
6442         struct btrfs_caching_control *caching_ctl;
6443
6444         block_group = btrfs_lookup_block_group(fs_info, start);
6445         if (!block_group)
6446                 return -EINVAL;
6447
6448         cache_block_group(block_group, 0);
6449         caching_ctl = get_caching_control(block_group);
6450
6451         if (!caching_ctl) {
6452                 /* Logic error */
6453                 BUG_ON(!block_group_cache_done(block_group));
6454                 ret = btrfs_remove_free_space(block_group, start, num_bytes);
6455         } else {
6456                 mutex_lock(&caching_ctl->mutex);
6457
6458                 if (start >= caching_ctl->progress) {
6459                         ret = add_excluded_extent(fs_info, start, num_bytes);
6460                 } else if (start + num_bytes <= caching_ctl->progress) {
6461                         ret = btrfs_remove_free_space(block_group,
6462                                                       start, num_bytes);
6463                 } else {
6464                         num_bytes = caching_ctl->progress - start;
6465                         ret = btrfs_remove_free_space(block_group,
6466                                                       start, num_bytes);
6467                         if (ret)
6468                                 goto out_lock;
6469
6470                         num_bytes = (start + num_bytes) -
6471                                 caching_ctl->progress;
6472                         start = caching_ctl->progress;
6473                         ret = add_excluded_extent(fs_info, start, num_bytes);
6474                 }
6475 out_lock:
6476                 mutex_unlock(&caching_ctl->mutex);
6477                 put_caching_control(caching_ctl);
6478         }
6479         btrfs_put_block_group(block_group);
6480         return ret;
6481 }
6482
6483 int btrfs_exclude_logged_extents(struct extent_buffer *eb)
6484 {
6485         struct btrfs_fs_info *fs_info = eb->fs_info;
6486         struct btrfs_file_extent_item *item;
6487         struct btrfs_key key;
6488         int found_type;
6489         int i;
6490         int ret = 0;
6491
6492         if (!btrfs_fs_incompat(fs_info, MIXED_GROUPS))
6493                 return 0;
6494
6495         for (i = 0; i < btrfs_header_nritems(eb); i++) {
6496                 btrfs_item_key_to_cpu(eb, &key, i);
6497                 if (key.type != BTRFS_EXTENT_DATA_KEY)
6498                         continue;
6499                 item = btrfs_item_ptr(eb, i, struct btrfs_file_extent_item);
6500                 found_type = btrfs_file_extent_type(eb, item);
6501                 if (found_type == BTRFS_FILE_EXTENT_INLINE)
6502                         continue;
6503                 if (btrfs_file_extent_disk_bytenr(eb, item) == 0)
6504                         continue;
6505                 key.objectid = btrfs_file_extent_disk_bytenr(eb, item);
6506                 key.offset = btrfs_file_extent_disk_num_bytes(eb, item);
6507                 ret = __exclude_logged_extent(fs_info, key.objectid, key.offset);
6508                 if (ret)
6509                         break;
6510         }
6511
6512         return ret;
6513 }
6514
6515 static void
6516 btrfs_inc_block_group_reservations(struct btrfs_block_group_cache *bg)
6517 {
6518         atomic_inc(&bg->reservations);
6519 }
6520
6521 void btrfs_dec_block_group_reservations(struct btrfs_fs_info *fs_info,
6522                                         const u64 start)
6523 {
6524         struct btrfs_block_group_cache *bg;
6525
6526         bg = btrfs_lookup_block_group(fs_info, start);
6527         ASSERT(bg);
6528         if (atomic_dec_and_test(&bg->reservations))
6529                 wake_up_var(&bg->reservations);
6530         btrfs_put_block_group(bg);
6531 }
6532
6533 void btrfs_wait_block_group_reservations(struct btrfs_block_group_cache *bg)
6534 {
6535         struct btrfs_space_info *space_info = bg->space_info;
6536
6537         ASSERT(bg->ro);
6538
6539         if (!(bg->flags & BTRFS_BLOCK_GROUP_DATA))
6540                 return;
6541
6542         /*
6543          * Our block group is read only but before we set it to read only,
6544          * some task might have had allocated an extent from it already, but it
6545          * has not yet created a respective ordered extent (and added it to a
6546          * root's list of ordered extents).
6547          * Therefore wait for any task currently allocating extents, since the
6548          * block group's reservations counter is incremented while a read lock
6549          * on the groups' semaphore is held and decremented after releasing
6550          * the read access on that semaphore and creating the ordered extent.
6551          */
6552         down_write(&space_info->groups_sem);
6553         up_write(&space_info->groups_sem);
6554
6555         wait_var_event(&bg->reservations, !atomic_read(&bg->reservations));
6556 }
6557
6558 /**
6559  * btrfs_add_reserved_bytes - update the block_group and space info counters
6560  * @cache:      The cache we are manipulating
6561  * @ram_bytes:  The number of bytes of file content, and will be same to
6562  *              @num_bytes except for the compress path.
6563  * @num_bytes:  The number of bytes in question
6564  * @delalloc:   The blocks are allocated for the delalloc write
6565  *
6566  * This is called by the allocator when it reserves space. If this is a
6567  * reservation and the block group has become read only we cannot make the
6568  * reservation and return -EAGAIN, otherwise this function always succeeds.
6569  */
6570 static int btrfs_add_reserved_bytes(struct btrfs_block_group_cache *cache,
6571                                     u64 ram_bytes, u64 num_bytes, int delalloc)
6572 {
6573         struct btrfs_space_info *space_info = cache->space_info;
6574         int ret = 0;
6575
6576         spin_lock(&space_info->lock);
6577         spin_lock(&cache->lock);
6578         if (cache->ro) {
6579                 ret = -EAGAIN;
6580         } else {
6581                 cache->reserved += num_bytes;
6582                 space_info->bytes_reserved += num_bytes;
6583                 update_bytes_may_use(space_info, -ram_bytes);
6584                 if (delalloc)
6585                         cache->delalloc_bytes += num_bytes;
6586         }
6587         spin_unlock(&cache->lock);
6588         spin_unlock(&space_info->lock);
6589         return ret;
6590 }
6591
6592 /**
6593  * btrfs_free_reserved_bytes - update the block_group and space info counters
6594  * @cache:      The cache we are manipulating
6595  * @num_bytes:  The number of bytes in question
6596  * @delalloc:   The blocks are allocated for the delalloc write
6597  *
6598  * This is called by somebody who is freeing space that was never actually used
6599  * on disk.  For example if you reserve some space for a new leaf in transaction
6600  * A and before transaction A commits you free that leaf, you call this with
6601  * reserve set to 0 in order to clear the reservation.
6602  */
6603
6604 static void btrfs_free_reserved_bytes(struct btrfs_block_group_cache *cache,
6605                                       u64 num_bytes, int delalloc)
6606 {
6607         struct btrfs_space_info *space_info = cache->space_info;
6608
6609         spin_lock(&space_info->lock);
6610         spin_lock(&cache->lock);
6611         if (cache->ro)
6612                 space_info->bytes_readonly += num_bytes;
6613         cache->reserved -= num_bytes;
6614         space_info->bytes_reserved -= num_bytes;
6615         space_info->max_extent_size = 0;
6616
6617         if (delalloc)
6618                 cache->delalloc_bytes -= num_bytes;
6619         spin_unlock(&cache->lock);
6620         spin_unlock(&space_info->lock);
6621 }
6622 void btrfs_prepare_extent_commit(struct btrfs_fs_info *fs_info)
6623 {
6624         struct btrfs_caching_control *next;
6625         struct btrfs_caching_control *caching_ctl;
6626         struct btrfs_block_group_cache *cache;
6627
6628         down_write(&fs_info->commit_root_sem);
6629
6630         list_for_each_entry_safe(caching_ctl, next,
6631                                  &fs_info->caching_block_groups, list) {
6632                 cache = caching_ctl->block_group;
6633                 if (block_group_cache_done(cache)) {
6634                         cache->last_byte_to_unpin = (u64)-1;
6635                         list_del_init(&caching_ctl->list);
6636                         put_caching_control(caching_ctl);
6637                 } else {
6638                         cache->last_byte_to_unpin = caching_ctl->progress;
6639                 }
6640         }
6641
6642         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
6643                 fs_info->pinned_extents = &fs_info->freed_extents[1];
6644         else
6645                 fs_info->pinned_extents = &fs_info->freed_extents[0];
6646
6647         up_write(&fs_info->commit_root_sem);
6648
6649         update_global_block_rsv(fs_info);
6650 }
6651
6652 /*
6653  * Returns the free cluster for the given space info and sets empty_cluster to
6654  * what it should be based on the mount options.
6655  */
6656 static struct btrfs_free_cluster *
6657 fetch_cluster_info(struct btrfs_fs_info *fs_info,
6658                    struct btrfs_space_info *space_info, u64 *empty_cluster)
6659 {
6660         struct btrfs_free_cluster *ret = NULL;
6661
6662         *empty_cluster = 0;
6663         if (btrfs_mixed_space_info(space_info))
6664                 return ret;
6665
6666         if (space_info->flags & BTRFS_BLOCK_GROUP_METADATA) {
6667                 ret = &fs_info->meta_alloc_cluster;
6668                 if (btrfs_test_opt(fs_info, SSD))
6669                         *empty_cluster = SZ_2M;
6670                 else
6671                         *empty_cluster = SZ_64K;
6672         } else if ((space_info->flags & BTRFS_BLOCK_GROUP_DATA) &&
6673                    btrfs_test_opt(fs_info, SSD_SPREAD)) {
6674                 *empty_cluster = SZ_2M;
6675                 ret = &fs_info->data_alloc_cluster;
6676         }
6677
6678         return ret;
6679 }
6680
6681 static int unpin_extent_range(struct btrfs_fs_info *fs_info,
6682                               u64 start, u64 end,
6683                               const bool return_free_space)
6684 {
6685         struct btrfs_block_group_cache *cache = NULL;
6686         struct btrfs_space_info *space_info;
6687         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
6688         struct btrfs_free_cluster *cluster = NULL;
6689         u64 len;
6690         u64 total_unpinned = 0;
6691         u64 empty_cluster = 0;
6692         bool readonly;
6693
6694         while (start <= end) {
6695                 readonly = false;
6696                 if (!cache ||
6697                     start >= cache->key.objectid + cache->key.offset) {
6698                         if (cache)
6699                                 btrfs_put_block_group(cache);
6700                         total_unpinned = 0;
6701                         cache = btrfs_lookup_block_group(fs_info, start);
6702                         BUG_ON(!cache); /* Logic error */
6703
6704                         cluster = fetch_cluster_info(fs_info,
6705                                                      cache->space_info,
6706                                                      &empty_cluster);
6707                         empty_cluster <<= 1;
6708                 }
6709
6710                 len = cache->key.objectid + cache->key.offset - start;
6711                 len = min(len, end + 1 - start);
6712
6713                 if (start < cache->last_byte_to_unpin) {
6714                         len = min(len, cache->last_byte_to_unpin - start);
6715                         if (return_free_space)
6716                                 btrfs_add_free_space(cache, start, len);
6717                 }
6718
6719                 start += len;
6720                 total_unpinned += len;
6721                 space_info = cache->space_info;
6722
6723                 /*
6724                  * If this space cluster has been marked as fragmented and we've
6725                  * unpinned enough in this block group to potentially allow a
6726                  * cluster to be created inside of it go ahead and clear the
6727                  * fragmented check.
6728                  */
6729                 if (cluster && cluster->fragmented &&
6730                     total_unpinned > empty_cluster) {
6731                         spin_lock(&cluster->lock);
6732                         cluster->fragmented = 0;
6733                         spin_unlock(&cluster->lock);
6734                 }
6735
6736                 spin_lock(&space_info->lock);
6737                 spin_lock(&cache->lock);
6738                 cache->pinned -= len;
6739                 update_bytes_pinned(space_info, -len);
6740
6741                 trace_btrfs_space_reservation(fs_info, "pinned",
6742                                               space_info->flags, len, 0);
6743                 space_info->max_extent_size = 0;
6744                 percpu_counter_add_batch(&space_info->total_bytes_pinned,
6745                             -len, BTRFS_TOTAL_BYTES_PINNED_BATCH);
6746                 if (cache->ro) {
6747                         space_info->bytes_readonly += len;
6748                         readonly = true;
6749                 }
6750                 spin_unlock(&cache->lock);
6751                 if (!readonly && return_free_space &&
6752                     global_rsv->space_info == space_info) {
6753                         u64 to_add = len;
6754
6755                         spin_lock(&global_rsv->lock);
6756                         if (!global_rsv->full) {
6757                                 to_add = min(len, global_rsv->size -
6758                                              global_rsv->reserved);
6759                                 global_rsv->reserved += to_add;
6760                                 update_bytes_may_use(space_info, to_add);
6761                                 if (global_rsv->reserved >= global_rsv->size)
6762                                         global_rsv->full = 1;
6763                                 trace_btrfs_space_reservation(fs_info,
6764                                                               "space_info",
6765                                                               space_info->flags,
6766                                                               to_add, 1);
6767                                 len -= to_add;
6768                         }
6769                         spin_unlock(&global_rsv->lock);
6770                         /* Add to any tickets we may have */
6771                         if (len)
6772                                 space_info_add_new_bytes(fs_info, space_info,
6773                                                          len);
6774                 }
6775                 spin_unlock(&space_info->lock);
6776         }
6777
6778         if (cache)
6779                 btrfs_put_block_group(cache);
6780         return 0;
6781 }
6782
6783 int btrfs_finish_extent_commit(struct btrfs_trans_handle *trans)
6784 {
6785         struct btrfs_fs_info *fs_info = trans->fs_info;
6786         struct btrfs_block_group_cache *block_group, *tmp;
6787         struct list_head *deleted_bgs;
6788         struct extent_io_tree *unpin;
6789         u64 start;
6790         u64 end;
6791         int ret;
6792
6793         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
6794                 unpin = &fs_info->freed_extents[1];
6795         else
6796                 unpin = &fs_info->freed_extents[0];
6797
6798         while (!trans->aborted) {
6799                 struct extent_state *cached_state = NULL;
6800
6801                 mutex_lock(&fs_info->unused_bg_unpin_mutex);
6802                 ret = find_first_extent_bit(unpin, 0, &start, &end,
6803                                             EXTENT_DIRTY, &cached_state);
6804                 if (ret) {
6805                         mutex_unlock(&fs_info->unused_bg_unpin_mutex);
6806                         break;
6807                 }
6808
6809                 if (btrfs_test_opt(fs_info, DISCARD))
6810                         ret = btrfs_discard_extent(fs_info, start,
6811                                                    end + 1 - start, NULL);
6812
6813                 clear_extent_dirty(unpin, start, end, &cached_state);
6814                 unpin_extent_range(fs_info, start, end, true);
6815                 mutex_unlock(&fs_info->unused_bg_unpin_mutex);
6816                 free_extent_state(cached_state);
6817                 cond_resched();
6818         }
6819
6820         /*
6821          * Transaction is finished.  We don't need the lock anymore.  We
6822          * do need to clean up the block groups in case of a transaction
6823          * abort.
6824          */
6825         deleted_bgs = &trans->transaction->deleted_bgs;
6826         list_for_each_entry_safe(block_group, tmp, deleted_bgs, bg_list) {
6827                 u64 trimmed = 0;
6828
6829                 ret = -EROFS;
6830                 if (!trans->aborted)
6831                         ret = btrfs_discard_extent(fs_info,
6832                                                    block_group->key.objectid,
6833                                                    block_group->key.offset,
6834                                                    &trimmed);
6835
6836                 list_del_init(&block_group->bg_list);
6837                 btrfs_put_block_group_trimming(block_group);
6838                 btrfs_put_block_group(block_group);
6839
6840                 if (ret) {
6841                         const char *errstr = btrfs_decode_error(ret);
6842                         btrfs_warn(fs_info,
6843                            "discard failed while removing blockgroup: errno=%d %s",
6844                                    ret, errstr);
6845                 }
6846         }
6847
6848         return 0;
6849 }
6850
6851 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
6852                                struct btrfs_delayed_ref_node *node, u64 parent,
6853                                u64 root_objectid, u64 owner_objectid,
6854                                u64 owner_offset, int refs_to_drop,
6855                                struct btrfs_delayed_extent_op *extent_op)
6856 {
6857         struct btrfs_fs_info *info = trans->fs_info;
6858         struct btrfs_key key;
6859         struct btrfs_path *path;
6860         struct btrfs_root *extent_root = info->extent_root;
6861         struct extent_buffer *leaf;
6862         struct btrfs_extent_item *ei;
6863         struct btrfs_extent_inline_ref *iref;
6864         int ret;
6865         int is_data;
6866         int extent_slot = 0;
6867         int found_extent = 0;
6868         int num_to_del = 1;
6869         u32 item_size;
6870         u64 refs;
6871         u64 bytenr = node->bytenr;
6872         u64 num_bytes = node->num_bytes;
6873         int last_ref = 0;
6874         bool skinny_metadata = btrfs_fs_incompat(info, SKINNY_METADATA);
6875
6876         path = btrfs_alloc_path();
6877         if (!path)
6878                 return -ENOMEM;
6879
6880         path->reada = READA_FORWARD;
6881         path->leave_spinning = 1;
6882
6883         is_data = owner_objectid >= BTRFS_FIRST_FREE_OBJECTID;
6884         BUG_ON(!is_data && refs_to_drop != 1);
6885
6886         if (is_data)
6887                 skinny_metadata = false;
6888
6889         ret = lookup_extent_backref(trans, path, &iref, bytenr, num_bytes,
6890                                     parent, root_objectid, owner_objectid,
6891                                     owner_offset);
6892         if (ret == 0) {
6893                 extent_slot = path->slots[0];
6894                 while (extent_slot >= 0) {
6895                         btrfs_item_key_to_cpu(path->nodes[0], &key,
6896                                               extent_slot);
6897                         if (key.objectid != bytenr)
6898                                 break;
6899                         if (key.type == BTRFS_EXTENT_ITEM_KEY &&
6900                             key.offset == num_bytes) {
6901                                 found_extent = 1;
6902                                 break;
6903                         }
6904                         if (key.type == BTRFS_METADATA_ITEM_KEY &&
6905                             key.offset == owner_objectid) {
6906                                 found_extent = 1;
6907                                 break;
6908                         }
6909                         if (path->slots[0] - extent_slot > 5)
6910                                 break;
6911                         extent_slot--;
6912                 }
6913
6914                 if (!found_extent) {
6915                         BUG_ON(iref);
6916                         ret = remove_extent_backref(trans, path, NULL,
6917                                                     refs_to_drop,
6918                                                     is_data, &last_ref);
6919                         if (ret) {
6920                                 btrfs_abort_transaction(trans, ret);
6921                                 goto out;
6922                         }
6923                         btrfs_release_path(path);
6924                         path->leave_spinning = 1;
6925
6926                         key.objectid = bytenr;
6927                         key.type = BTRFS_EXTENT_ITEM_KEY;
6928                         key.offset = num_bytes;
6929
6930                         if (!is_data && skinny_metadata) {
6931                                 key.type = BTRFS_METADATA_ITEM_KEY;
6932                                 key.offset = owner_objectid;
6933                         }
6934
6935                         ret = btrfs_search_slot(trans, extent_root,
6936                                                 &key, path, -1, 1);
6937                         if (ret > 0 && skinny_metadata && path->slots[0]) {
6938                                 /*
6939                                  * Couldn't find our skinny metadata item,
6940                                  * see if we have ye olde extent item.
6941                                  */
6942                                 path->slots[0]--;
6943                                 btrfs_item_key_to_cpu(path->nodes[0], &key,
6944                                                       path->slots[0]);
6945                                 if (key.objectid == bytenr &&
6946                                     key.type == BTRFS_EXTENT_ITEM_KEY &&
6947                                     key.offset == num_bytes)
6948                                         ret = 0;
6949                         }
6950
6951                         if (ret > 0 && skinny_metadata) {
6952                                 skinny_metadata = false;
6953                                 key.objectid = bytenr;
6954                                 key.type = BTRFS_EXTENT_ITEM_KEY;
6955                                 key.offset = num_bytes;
6956                                 btrfs_release_path(path);
6957                                 ret = btrfs_search_slot(trans, extent_root,
6958                                                         &key, path, -1, 1);
6959                         }
6960
6961                         if (ret) {
6962                                 btrfs_err(info,
6963                                           "umm, got %d back from search, was looking for %llu",
6964                                           ret, bytenr);
6965                                 if (ret > 0)
6966                                         btrfs_print_leaf(path->nodes[0]);
6967                         }
6968                         if (ret < 0) {
6969                                 btrfs_abort_transaction(trans, ret);
6970                                 goto out;
6971                         }
6972                         extent_slot = path->slots[0];
6973                 }
6974         } else if (WARN_ON(ret == -ENOENT)) {
6975                 btrfs_print_leaf(path->nodes[0]);
6976                 btrfs_err(info,
6977                         "unable to find ref byte nr %llu parent %llu root %llu  owner %llu offset %llu",
6978                         bytenr, parent, root_objectid, owner_objectid,
6979                         owner_offset);
6980                 btrfs_abort_transaction(trans, ret);
6981                 goto out;
6982         } else {
6983                 btrfs_abort_transaction(trans, ret);
6984                 goto out;
6985         }
6986
6987         leaf = path->nodes[0];
6988         item_size = btrfs_item_size_nr(leaf, extent_slot);
6989         if (unlikely(item_size < sizeof(*ei))) {
6990                 ret = -EINVAL;
6991                 btrfs_print_v0_err(info);
6992                 btrfs_abort_transaction(trans, ret);
6993                 goto out;
6994         }
6995         ei = btrfs_item_ptr(leaf, extent_slot,
6996                             struct btrfs_extent_item);
6997         if (owner_objectid < BTRFS_FIRST_FREE_OBJECTID &&
6998             key.type == BTRFS_EXTENT_ITEM_KEY) {
6999                 struct btrfs_tree_block_info *bi;
7000                 BUG_ON(item_size < sizeof(*ei) + sizeof(*bi));
7001                 bi = (struct btrfs_tree_block_info *)(ei + 1);
7002                 WARN_ON(owner_objectid != btrfs_tree_block_level(leaf, bi));
7003         }
7004
7005         refs = btrfs_extent_refs(leaf, ei);
7006         if (refs < refs_to_drop) {
7007                 btrfs_err(info,
7008                           "trying to drop %d refs but we only have %Lu for bytenr %Lu",
7009                           refs_to_drop, refs, bytenr);
7010                 ret = -EINVAL;
7011                 btrfs_abort_transaction(trans, ret);
7012                 goto out;
7013         }
7014         refs -= refs_to_drop;
7015
7016         if (refs > 0) {
7017                 if (extent_op)
7018                         __run_delayed_extent_op(extent_op, leaf, ei);
7019                 /*
7020                  * In the case of inline back ref, reference count will
7021                  * be updated by remove_extent_backref
7022                  */
7023                 if (iref) {
7024                         BUG_ON(!found_extent);
7025                 } else {
7026                         btrfs_set_extent_refs(leaf, ei, refs);
7027                         btrfs_mark_buffer_dirty(leaf);
7028                 }
7029                 if (found_extent) {
7030                         ret = remove_extent_backref(trans, path, iref,
7031                                                     refs_to_drop, is_data,
7032                                                     &last_ref);
7033                         if (ret) {
7034                                 btrfs_abort_transaction(trans, ret);
7035                                 goto out;
7036                         }
7037                 }
7038         } else {
7039                 if (found_extent) {
7040                         BUG_ON(is_data && refs_to_drop !=
7041                                extent_data_ref_count(path, iref));
7042                         if (iref) {
7043                                 BUG_ON(path->slots[0] != extent_slot);
7044                         } else {
7045                                 BUG_ON(path->slots[0] != extent_slot + 1);
7046                                 path->slots[0] = extent_slot;
7047                                 num_to_del = 2;
7048                         }
7049                 }
7050
7051                 last_ref = 1;
7052                 ret = btrfs_del_items(trans, extent_root, path, path->slots[0],
7053                                       num_to_del);
7054                 if (ret) {
7055                         btrfs_abort_transaction(trans, ret);
7056                         goto out;
7057                 }
7058                 btrfs_release_path(path);
7059
7060                 if (is_data) {
7061                         ret = btrfs_del_csums(trans, info, bytenr, num_bytes);
7062                         if (ret) {
7063                                 btrfs_abort_transaction(trans, ret);
7064                                 goto out;
7065                         }
7066                 }
7067
7068                 ret = add_to_free_space_tree(trans, bytenr, num_bytes);
7069                 if (ret) {
7070                         btrfs_abort_transaction(trans, ret);
7071                         goto out;
7072                 }
7073
7074                 ret = update_block_group(trans, bytenr, num_bytes, 0);
7075                 if (ret) {
7076                         btrfs_abort_transaction(trans, ret);
7077                         goto out;
7078                 }
7079         }
7080         btrfs_release_path(path);
7081
7082 out:
7083         btrfs_free_path(path);
7084         return ret;
7085 }
7086
7087 /*
7088  * when we free an block, it is possible (and likely) that we free the last
7089  * delayed ref for that extent as well.  This searches the delayed ref tree for
7090  * a given extent, and if there are no other delayed refs to be processed, it
7091  * removes it from the tree.
7092  */
7093 static noinline int check_ref_cleanup(struct btrfs_trans_handle *trans,
7094                                       u64 bytenr)
7095 {
7096         struct btrfs_delayed_ref_head *head;
7097         struct btrfs_delayed_ref_root *delayed_refs;
7098         int ret = 0;
7099
7100         delayed_refs = &trans->transaction->delayed_refs;
7101         spin_lock(&delayed_refs->lock);
7102         head = btrfs_find_delayed_ref_head(delayed_refs, bytenr);
7103         if (!head)
7104                 goto out_delayed_unlock;
7105
7106         spin_lock(&head->lock);
7107         if (!RB_EMPTY_ROOT(&head->ref_tree.rb_root))
7108                 goto out;
7109
7110         if (cleanup_extent_op(head) != NULL)
7111                 goto out;
7112
7113         /*
7114          * waiting for the lock here would deadlock.  If someone else has it
7115          * locked they are already in the process of dropping it anyway
7116          */
7117         if (!mutex_trylock(&head->mutex))
7118                 goto out;
7119
7120         btrfs_delete_ref_head(delayed_refs, head);
7121         head->processing = 0;
7122
7123         spin_unlock(&head->lock);
7124         spin_unlock(&delayed_refs->lock);
7125
7126         BUG_ON(head->extent_op);
7127         if (head->must_insert_reserved)
7128                 ret = 1;
7129
7130         btrfs_cleanup_ref_head_accounting(trans->fs_info, delayed_refs, head);
7131         mutex_unlock(&head->mutex);
7132         btrfs_put_delayed_ref_head(head);
7133         return ret;
7134 out:
7135         spin_unlock(&head->lock);
7136
7137 out_delayed_unlock:
7138         spin_unlock(&delayed_refs->lock);
7139         return 0;
7140 }
7141
7142 void btrfs_free_tree_block(struct btrfs_trans_handle *trans,
7143                            struct btrfs_root *root,
7144                            struct extent_buffer *buf,
7145                            u64 parent, int last_ref)
7146 {
7147         struct btrfs_fs_info *fs_info = root->fs_info;
7148         struct btrfs_ref generic_ref = { 0 };
7149         int pin = 1;
7150         int ret;
7151
7152         btrfs_init_generic_ref(&generic_ref, BTRFS_DROP_DELAYED_REF,
7153                                buf->start, buf->len, parent);
7154         btrfs_init_tree_ref(&generic_ref, btrfs_header_level(buf),
7155                             root->root_key.objectid);
7156
7157         if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
7158                 int old_ref_mod, new_ref_mod;
7159
7160                 btrfs_ref_tree_mod(fs_info, &generic_ref);
7161                 ret = btrfs_add_delayed_tree_ref(trans, &generic_ref, NULL,
7162                                                  &old_ref_mod, &new_ref_mod);
7163                 BUG_ON(ret); /* -ENOMEM */
7164                 pin = old_ref_mod >= 0 && new_ref_mod < 0;
7165         }
7166
7167         if (last_ref && btrfs_header_generation(buf) == trans->transid) {
7168                 struct btrfs_block_group_cache *cache;
7169
7170                 if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
7171                         ret = check_ref_cleanup(trans, buf->start);
7172                         if (!ret)
7173                                 goto out;
7174                 }
7175
7176                 pin = 0;
7177                 cache = btrfs_lookup_block_group(fs_info, buf->start);
7178
7179                 if (btrfs_header_flag(buf, BTRFS_HEADER_FLAG_WRITTEN)) {
7180                         pin_down_extent(cache, buf->start, buf->len, 1);
7181                         btrfs_put_block_group(cache);
7182                         goto out;
7183                 }
7184
7185                 WARN_ON(test_bit(EXTENT_BUFFER_DIRTY, &buf->bflags));
7186
7187                 btrfs_add_free_space(cache, buf->start, buf->len);
7188                 btrfs_free_reserved_bytes(cache, buf->len, 0);
7189                 btrfs_put_block_group(cache);
7190                 trace_btrfs_reserved_extent_free(fs_info, buf->start, buf->len);
7191         }
7192 out:
7193         if (pin)
7194                 add_pinned_bytes(fs_info, &generic_ref, 1);
7195
7196         if (last_ref) {
7197                 /*
7198                  * Deleting the buffer, clear the corrupt flag since it doesn't
7199                  * matter anymore.
7200                  */
7201                 clear_bit(EXTENT_BUFFER_CORRUPT, &buf->bflags);
7202         }
7203 }
7204
7205 /* Can return -ENOMEM */
7206 int btrfs_free_extent(struct btrfs_trans_handle *trans, struct btrfs_ref *ref)
7207 {
7208         struct btrfs_fs_info *fs_info = trans->fs_info;
7209         int old_ref_mod, new_ref_mod;
7210         int ret;
7211
7212         if (btrfs_is_testing(fs_info))
7213                 return 0;
7214
7215         /*
7216          * tree log blocks never actually go into the extent allocation
7217          * tree, just update pinning info and exit early.
7218          */
7219         if ((ref->type == BTRFS_REF_METADATA &&
7220              ref->tree_ref.root == BTRFS_TREE_LOG_OBJECTID) ||
7221             (ref->type == BTRFS_REF_DATA &&
7222              ref->data_ref.ref_root == BTRFS_TREE_LOG_OBJECTID)) {
7223                 /* unlocks the pinned mutex */
7224                 btrfs_pin_extent(fs_info, ref->bytenr, ref->len, 1);
7225                 old_ref_mod = new_ref_mod = 0;
7226                 ret = 0;
7227         } else if (ref->type == BTRFS_REF_METADATA) {
7228                 ret = btrfs_add_delayed_tree_ref(trans, ref, NULL,
7229                                                  &old_ref_mod, &new_ref_mod);
7230         } else {
7231                 ret = btrfs_add_delayed_data_ref(trans, ref, 0,
7232                                                  &old_ref_mod, &new_ref_mod);
7233         }
7234
7235         if (!((ref->type == BTRFS_REF_METADATA &&
7236                ref->tree_ref.root == BTRFS_TREE_LOG_OBJECTID) ||
7237               (ref->type == BTRFS_REF_DATA &&
7238                ref->data_ref.ref_root == BTRFS_TREE_LOG_OBJECTID)))
7239                 btrfs_ref_tree_mod(fs_info, ref);
7240
7241         if (ret == 0 && old_ref_mod >= 0 && new_ref_mod < 0)
7242                 add_pinned_bytes(fs_info, ref, 1);
7243
7244         return ret;
7245 }
7246
7247 /*
7248  * when we wait for progress in the block group caching, its because
7249  * our allocation attempt failed at least once.  So, we must sleep
7250  * and let some progress happen before we try again.
7251  *
7252  * This function will sleep at least once waiting for new free space to
7253  * show up, and then it will check the block group free space numbers
7254  * for our min num_bytes.  Another option is to have it go ahead
7255  * and look in the rbtree for a free extent of a given size, but this
7256  * is a good start.
7257  *
7258  * Callers of this must check if cache->cached == BTRFS_CACHE_ERROR before using
7259  * any of the information in this block group.
7260  */
7261 static noinline void
7262 wait_block_group_cache_progress(struct btrfs_block_group_cache *cache,
7263                                 u64 num_bytes)
7264 {
7265         struct btrfs_caching_control *caching_ctl;
7266
7267         caching_ctl = get_caching_control(cache);
7268         if (!caching_ctl)
7269                 return;
7270
7271         wait_event(caching_ctl->wait, block_group_cache_done(cache) ||
7272                    (cache->free_space_ctl->free_space >= num_bytes));
7273
7274         put_caching_control(caching_ctl);
7275 }
7276
7277 static noinline int
7278 wait_block_group_cache_done(struct btrfs_block_group_cache *cache)
7279 {
7280         struct btrfs_caching_control *caching_ctl;
7281         int ret = 0;
7282
7283         caching_ctl = get_caching_control(cache);
7284         if (!caching_ctl)
7285                 return (cache->cached == BTRFS_CACHE_ERROR) ? -EIO : 0;
7286
7287         wait_event(caching_ctl->wait, block_group_cache_done(cache));
7288         if (cache->cached == BTRFS_CACHE_ERROR)
7289                 ret = -EIO;
7290         put_caching_control(caching_ctl);
7291         return ret;
7292 }
7293
7294 enum btrfs_loop_type {
7295         LOOP_CACHING_NOWAIT = 0,
7296         LOOP_CACHING_WAIT = 1,
7297         LOOP_ALLOC_CHUNK = 2,
7298         LOOP_NO_EMPTY_SIZE = 3,
7299 };
7300
7301 static inline void
7302 btrfs_lock_block_group(struct btrfs_block_group_cache *cache,
7303                        int delalloc)
7304 {
7305         if (delalloc)
7306                 down_read(&cache->data_rwsem);
7307 }
7308
7309 static inline void
7310 btrfs_grab_block_group(struct btrfs_block_group_cache *cache,
7311                        int delalloc)
7312 {
7313         btrfs_get_block_group(cache);
7314         if (delalloc)
7315                 down_read(&cache->data_rwsem);
7316 }
7317
7318 static struct btrfs_block_group_cache *
7319 btrfs_lock_cluster(struct btrfs_block_group_cache *block_group,
7320                    struct btrfs_free_cluster *cluster,
7321                    int delalloc)
7322 {
7323         struct btrfs_block_group_cache *used_bg = NULL;
7324
7325         spin_lock(&cluster->refill_lock);
7326         while (1) {
7327                 used_bg = cluster->block_group;
7328                 if (!used_bg)
7329                         return NULL;
7330
7331                 if (used_bg == block_group)
7332                         return used_bg;
7333
7334                 btrfs_get_block_group(used_bg);
7335
7336                 if (!delalloc)
7337                         return used_bg;
7338
7339                 if (down_read_trylock(&used_bg->data_rwsem))
7340                         return used_bg;
7341
7342                 spin_unlock(&cluster->refill_lock);
7343
7344                 /* We should only have one-level nested. */
7345                 down_read_nested(&used_bg->data_rwsem, SINGLE_DEPTH_NESTING);
7346
7347                 spin_lock(&cluster->refill_lock);
7348                 if (used_bg == cluster->block_group)
7349                         return used_bg;
7350
7351                 up_read(&used_bg->data_rwsem);
7352                 btrfs_put_block_group(used_bg);
7353         }
7354 }
7355
7356 static inline void
7357 btrfs_release_block_group(struct btrfs_block_group_cache *cache,
7358                          int delalloc)
7359 {
7360         if (delalloc)
7361                 up_read(&cache->data_rwsem);
7362         btrfs_put_block_group(cache);
7363 }
7364
7365 /*
7366  * Structure used internally for find_free_extent() function.  Wraps needed
7367  * parameters.
7368  */
7369 struct find_free_extent_ctl {
7370         /* Basic allocation info */
7371         u64 ram_bytes;
7372         u64 num_bytes;
7373         u64 empty_size;
7374         u64 flags;
7375         int delalloc;
7376
7377         /* Where to start the search inside the bg */
7378         u64 search_start;
7379
7380         /* For clustered allocation */
7381         u64 empty_cluster;
7382
7383         bool have_caching_bg;
7384         bool orig_have_caching_bg;
7385
7386         /* RAID index, converted from flags */
7387         int index;
7388
7389         /*
7390          * Current loop number, check find_free_extent_update_loop() for details
7391          */
7392         int loop;
7393
7394         /*
7395          * Whether we're refilling a cluster, if true we need to re-search
7396          * current block group but don't try to refill the cluster again.
7397          */
7398         bool retry_clustered;
7399
7400         /*
7401          * Whether we're updating free space cache, if true we need to re-search
7402          * current block group but don't try updating free space cache again.
7403          */
7404         bool retry_unclustered;
7405
7406         /* If current block group is cached */
7407         int cached;
7408
7409         /* Max contiguous hole found */
7410         u64 max_extent_size;
7411
7412         /* Total free space from free space cache, not always contiguous */
7413         u64 total_free_space;
7414
7415         /* Found result */
7416         u64 found_offset;
7417 };
7418
7419
7420 /*
7421  * Helper function for find_free_extent().
7422  *
7423  * Return -ENOENT to inform caller that we need fallback to unclustered mode.
7424  * Return -EAGAIN to inform caller that we need to re-search this block group
7425  * Return >0 to inform caller that we find nothing
7426  * Return 0 means we have found a location and set ffe_ctl->found_offset.
7427  */
7428 static int find_free_extent_clustered(struct btrfs_block_group_cache *bg,
7429                 struct btrfs_free_cluster *last_ptr,
7430                 struct find_free_extent_ctl *ffe_ctl,
7431                 struct btrfs_block_group_cache **cluster_bg_ret)
7432 {
7433         struct btrfs_block_group_cache *cluster_bg;
7434         u64 aligned_cluster;
7435         u64 offset;
7436         int ret;
7437
7438         cluster_bg = btrfs_lock_cluster(bg, last_ptr, ffe_ctl->delalloc);
7439         if (!cluster_bg)
7440                 goto refill_cluster;
7441         if (cluster_bg != bg && (cluster_bg->ro ||
7442             !block_group_bits(cluster_bg, ffe_ctl->flags)))
7443                 goto release_cluster;
7444
7445         offset = btrfs_alloc_from_cluster(cluster_bg, last_ptr,
7446                         ffe_ctl->num_bytes, cluster_bg->key.objectid,
7447                         &ffe_ctl->max_extent_size);
7448         if (offset) {
7449                 /* We have a block, we're done */
7450                 spin_unlock(&last_ptr->refill_lock);
7451                 trace_btrfs_reserve_extent_cluster(cluster_bg,
7452                                 ffe_ctl->search_start, ffe_ctl->num_bytes);
7453                 *cluster_bg_ret = cluster_bg;
7454                 ffe_ctl->found_offset = offset;
7455                 return 0;
7456         }
7457         WARN_ON(last_ptr->block_group != cluster_bg);
7458
7459 release_cluster:
7460         /*
7461          * If we are on LOOP_NO_EMPTY_SIZE, we can't set up a new clusters, so
7462          * lets just skip it and let the allocator find whatever block it can
7463          * find. If we reach this point, we will have tried the cluster
7464          * allocator plenty of times and not have found anything, so we are
7465          * likely way too fragmented for the clustering stuff to find anything.
7466          *
7467          * However, if the cluster is taken from the current block group,
7468          * release the cluster first, so that we stand a better chance of
7469          * succeeding in the unclustered allocation.
7470          */
7471         if (ffe_ctl->loop >= LOOP_NO_EMPTY_SIZE && cluster_bg != bg) {
7472                 spin_unlock(&last_ptr->refill_lock);
7473                 btrfs_release_block_group(cluster_bg, ffe_ctl->delalloc);
7474                 return -ENOENT;
7475         }
7476
7477         /* This cluster didn't work out, free it and start over */
7478         btrfs_return_cluster_to_free_space(NULL, last_ptr);
7479
7480         if (cluster_bg != bg)
7481                 btrfs_release_block_group(cluster_bg, ffe_ctl->delalloc);
7482
7483 refill_cluster:
7484         if (ffe_ctl->loop >= LOOP_NO_EMPTY_SIZE) {
7485                 spin_unlock(&last_ptr->refill_lock);
7486                 return -ENOENT;
7487         }
7488
7489         aligned_cluster = max_t(u64,
7490                         ffe_ctl->empty_cluster + ffe_ctl->empty_size,
7491                         bg->full_stripe_len);
7492         ret = btrfs_find_space_cluster(bg, last_ptr, ffe_ctl->search_start,
7493                         ffe_ctl->num_bytes, aligned_cluster);
7494         if (ret == 0) {
7495                 /* Now pull our allocation out of this cluster */
7496                 offset = btrfs_alloc_from_cluster(bg, last_ptr,
7497                                 ffe_ctl->num_bytes, ffe_ctl->search_start,
7498                                 &ffe_ctl->max_extent_size);
7499                 if (offset) {
7500                         /* We found one, proceed */
7501                         spin_unlock(&last_ptr->refill_lock);
7502                         trace_btrfs_reserve_extent_cluster(bg,
7503                                         ffe_ctl->search_start,
7504                                         ffe_ctl->num_bytes);
7505                         ffe_ctl->found_offset = offset;
7506                         return 0;
7507                 }
7508         } else if (!ffe_ctl->cached && ffe_ctl->loop > LOOP_CACHING_NOWAIT &&
7509                    !ffe_ctl->retry_clustered) {
7510                 spin_unlock(&last_ptr->refill_lock);
7511
7512                 ffe_ctl->retry_clustered = true;
7513                 wait_block_group_cache_progress(bg, ffe_ctl->num_bytes +
7514                                 ffe_ctl->empty_cluster + ffe_ctl->empty_size);
7515                 return -EAGAIN;
7516         }
7517         /*
7518          * At this point we either didn't find a cluster or we weren't able to
7519          * allocate a block from our cluster.  Free the cluster we've been
7520          * trying to use, and go to the next block group.
7521          */
7522         btrfs_return_cluster_to_free_space(NULL, last_ptr);
7523         spin_unlock(&last_ptr->refill_lock);
7524         return 1;
7525 }
7526
7527 /*
7528  * Return >0 to inform caller that we find nothing
7529  * Return 0 when we found an free extent and set ffe_ctrl->found_offset
7530  * Return -EAGAIN to inform caller that we need to re-search this block group
7531  */
7532 static int find_free_extent_unclustered(struct btrfs_block_group_cache *bg,
7533                 struct btrfs_free_cluster *last_ptr,
7534                 struct find_free_extent_ctl *ffe_ctl)
7535 {
7536         u64 offset;
7537
7538         /*
7539          * We are doing an unclustered allocation, set the fragmented flag so
7540          * we don't bother trying to setup a cluster again until we get more
7541          * space.
7542          */
7543         if (unlikely(last_ptr)) {
7544                 spin_lock(&last_ptr->lock);
7545                 last_ptr->fragmented = 1;
7546                 spin_unlock(&last_ptr->lock);
7547         }
7548         if (ffe_ctl->cached) {
7549                 struct btrfs_free_space_ctl *free_space_ctl;
7550
7551                 free_space_ctl = bg->free_space_ctl;
7552                 spin_lock(&free_space_ctl->tree_lock);
7553                 if (free_space_ctl->free_space <
7554                     ffe_ctl->num_bytes + ffe_ctl->empty_cluster +
7555                     ffe_ctl->empty_size) {
7556                         ffe_ctl->total_free_space = max_t(u64,
7557                                         ffe_ctl->total_free_space,
7558                                         free_space_ctl->free_space);
7559                         spin_unlock(&free_space_ctl->tree_lock);
7560                         return 1;
7561                 }
7562                 spin_unlock(&free_space_ctl->tree_lock);
7563         }
7564
7565         offset = btrfs_find_space_for_alloc(bg, ffe_ctl->search_start,
7566                         ffe_ctl->num_bytes, ffe_ctl->empty_size,
7567                         &ffe_ctl->max_extent_size);
7568
7569         /*
7570          * If we didn't find a chunk, and we haven't failed on this block group
7571          * before, and this block group is in the middle of caching and we are
7572          * ok with waiting, then go ahead and wait for progress to be made, and
7573          * set @retry_unclustered to true.
7574          *
7575          * If @retry_unclustered is true then we've already waited on this
7576          * block group once and should move on to the next block group.
7577          */
7578         if (!offset && !ffe_ctl->retry_unclustered && !ffe_ctl->cached &&
7579             ffe_ctl->loop > LOOP_CACHING_NOWAIT) {
7580                 wait_block_group_cache_progress(bg, ffe_ctl->num_bytes +
7581                                                 ffe_ctl->empty_size);
7582                 ffe_ctl->retry_unclustered = true;
7583                 return -EAGAIN;
7584         } else if (!offset) {
7585                 return 1;
7586         }
7587         ffe_ctl->found_offset = offset;
7588         return 0;
7589 }
7590
7591 /*
7592  * Return >0 means caller needs to re-search for free extent
7593  * Return 0 means we have the needed free extent.
7594  * Return <0 means we failed to locate any free extent.
7595  */
7596 static int find_free_extent_update_loop(struct btrfs_fs_info *fs_info,
7597                                         struct btrfs_free_cluster *last_ptr,
7598                                         struct btrfs_key *ins,
7599                                         struct find_free_extent_ctl *ffe_ctl,
7600                                         int full_search, bool use_cluster)
7601 {
7602         struct btrfs_root *root = fs_info->extent_root;
7603         int ret;
7604
7605         if ((ffe_ctl->loop == LOOP_CACHING_NOWAIT) &&
7606             ffe_ctl->have_caching_bg && !ffe_ctl->orig_have_caching_bg)
7607                 ffe_ctl->orig_have_caching_bg = true;
7608
7609         if (!ins->objectid && ffe_ctl->loop >= LOOP_CACHING_WAIT &&
7610             ffe_ctl->have_caching_bg)
7611                 return 1;
7612
7613         if (!ins->objectid && ++(ffe_ctl->index) < BTRFS_NR_RAID_TYPES)
7614                 return 1;
7615
7616         if (ins->objectid) {
7617                 if (!use_cluster && last_ptr) {
7618                         spin_lock(&last_ptr->lock);
7619                         last_ptr->window_start = ins->objectid;
7620                         spin_unlock(&last_ptr->lock);
7621                 }
7622                 return 0;
7623         }
7624
7625         /*
7626          * LOOP_CACHING_NOWAIT, search partially cached block groups, kicking
7627          *                      caching kthreads as we move along
7628          * LOOP_CACHING_WAIT, search everything, and wait if our bg is caching
7629          * LOOP_ALLOC_CHUNK, force a chunk allocation and try again
7630          * LOOP_NO_EMPTY_SIZE, set empty_size and empty_cluster to 0 and try
7631          *                     again
7632          */
7633         if (ffe_ctl->loop < LOOP_NO_EMPTY_SIZE) {
7634                 ffe_ctl->index = 0;
7635                 if (ffe_ctl->loop == LOOP_CACHING_NOWAIT) {
7636                         /*
7637                          * We want to skip the LOOP_CACHING_WAIT step if we
7638                          * don't have any uncached bgs and we've already done a
7639                          * full search through.
7640                          */
7641                         if (ffe_ctl->orig_have_caching_bg || !full_search)
7642                                 ffe_ctl->loop = LOOP_CACHING_WAIT;
7643                         else
7644                                 ffe_ctl->loop = LOOP_ALLOC_CHUNK;
7645                 } else {
7646                         ffe_ctl->loop++;
7647                 }
7648
7649                 if (ffe_ctl->loop == LOOP_ALLOC_CHUNK) {
7650                         struct btrfs_trans_handle *trans;
7651                         int exist = 0;
7652
7653                         trans = current->journal_info;
7654                         if (trans)
7655                                 exist = 1;
7656                         else
7657                                 trans = btrfs_join_transaction(root);
7658
7659                         if (IS_ERR(trans)) {
7660                                 ret = PTR_ERR(trans);
7661                                 return ret;
7662                         }
7663
7664                         ret = do_chunk_alloc(trans, ffe_ctl->flags,
7665                                              CHUNK_ALLOC_FORCE);
7666
7667                         /*
7668                          * If we can't allocate a new chunk we've already looped
7669                          * through at least once, move on to the NO_EMPTY_SIZE
7670                          * case.
7671                          */
7672                         if (ret == -ENOSPC)
7673                                 ffe_ctl->loop = LOOP_NO_EMPTY_SIZE;
7674
7675                         /* Do not bail out on ENOSPC since we can do more. */
7676                         if (ret < 0 && ret != -ENOSPC)
7677                                 btrfs_abort_transaction(trans, ret);
7678                         else
7679                                 ret = 0;
7680                         if (!exist)
7681                                 btrfs_end_transaction(trans);
7682                         if (ret)
7683                                 return ret;
7684                 }
7685
7686                 if (ffe_ctl->loop == LOOP_NO_EMPTY_SIZE) {
7687                         /*
7688                          * Don't loop again if we already have no empty_size and
7689                          * no empty_cluster.
7690                          */
7691                         if (ffe_ctl->empty_size == 0 &&
7692                             ffe_ctl->empty_cluster == 0)
7693                                 return -ENOSPC;
7694                         ffe_ctl->empty_size = 0;
7695                         ffe_ctl->empty_cluster = 0;
7696                 }
7697                 return 1;
7698         }
7699         return -ENOSPC;
7700 }
7701
7702 /*
7703  * walks the btree of allocated extents and find a hole of a given size.
7704  * The key ins is changed to record the hole:
7705  * ins->objectid == start position
7706  * ins->flags = BTRFS_EXTENT_ITEM_KEY
7707  * ins->offset == the size of the hole.
7708  * Any available blocks before search_start are skipped.
7709  *
7710  * If there is no suitable free space, we will record the max size of
7711  * the free space extent currently.
7712  *
7713  * The overall logic and call chain:
7714  *
7715  * find_free_extent()
7716  * |- Iterate through all block groups
7717  * |  |- Get a valid block group
7718  * |  |- Try to do clustered allocation in that block group
7719  * |  |- Try to do unclustered allocation in that block group
7720  * |  |- Check if the result is valid
7721  * |  |  |- If valid, then exit
7722  * |  |- Jump to next block group
7723  * |
7724  * |- Push harder to find free extents
7725  *    |- If not found, re-iterate all block groups
7726  */
7727 static noinline int find_free_extent(struct btrfs_fs_info *fs_info,
7728                                 u64 ram_bytes, u64 num_bytes, u64 empty_size,
7729                                 u64 hint_byte, struct btrfs_key *ins,
7730                                 u64 flags, int delalloc)
7731 {
7732         int ret = 0;
7733         struct btrfs_free_cluster *last_ptr = NULL;
7734         struct btrfs_block_group_cache *block_group = NULL;
7735         struct find_free_extent_ctl ffe_ctl = {0};
7736         struct btrfs_space_info *space_info;
7737         bool use_cluster = true;
7738         bool full_search = false;
7739
7740         WARN_ON(num_bytes < fs_info->sectorsize);
7741
7742         ffe_ctl.ram_bytes = ram_bytes;
7743         ffe_ctl.num_bytes = num_bytes;
7744         ffe_ctl.empty_size = empty_size;
7745         ffe_ctl.flags = flags;
7746         ffe_ctl.search_start = 0;
7747         ffe_ctl.retry_clustered = false;
7748         ffe_ctl.retry_unclustered = false;
7749         ffe_ctl.delalloc = delalloc;
7750         ffe_ctl.index = btrfs_bg_flags_to_raid_index(flags);
7751         ffe_ctl.have_caching_bg = false;
7752         ffe_ctl.orig_have_caching_bg = false;
7753         ffe_ctl.found_offset = 0;
7754
7755         ins->type = BTRFS_EXTENT_ITEM_KEY;
7756         ins->objectid = 0;
7757         ins->offset = 0;
7758
7759         trace_find_free_extent(fs_info, num_bytes, empty_size, flags);
7760
7761         space_info = __find_space_info(fs_info, flags);
7762         if (!space_info) {
7763                 btrfs_err(fs_info, "No space info for %llu", flags);
7764                 return -ENOSPC;
7765         }
7766
7767         /*
7768          * If our free space is heavily fragmented we may not be able to make
7769          * big contiguous allocations, so instead of doing the expensive search
7770          * for free space, simply return ENOSPC with our max_extent_size so we
7771          * can go ahead and search for a more manageable chunk.
7772          *
7773          * If our max_extent_size is large enough for our allocation simply
7774          * disable clustering since we will likely not be able to find enough
7775          * space to create a cluster and induce latency trying.
7776          */
7777         if (unlikely(space_info->max_extent_size)) {
7778                 spin_lock(&space_info->lock);
7779                 if (space_info->max_extent_size &&
7780                     num_bytes > space_info->max_extent_size) {
7781                         ins->offset = space_info->max_extent_size;
7782                         spin_unlock(&space_info->lock);
7783                         return -ENOSPC;
7784                 } else if (space_info->max_extent_size) {
7785                         use_cluster = false;
7786                 }
7787                 spin_unlock(&space_info->lock);
7788         }
7789
7790         last_ptr = fetch_cluster_info(fs_info, space_info,
7791                                       &ffe_ctl.empty_cluster);
7792         if (last_ptr) {
7793                 spin_lock(&last_ptr->lock);
7794                 if (last_ptr->block_group)
7795                         hint_byte = last_ptr->window_start;
7796                 if (last_ptr->fragmented) {
7797                         /*
7798                          * We still set window_start so we can keep track of the
7799                          * last place we found an allocation to try and save
7800                          * some time.
7801                          */
7802                         hint_byte = last_ptr->window_start;
7803                         use_cluster = false;
7804                 }
7805                 spin_unlock(&last_ptr->lock);
7806         }
7807
7808         ffe_ctl.search_start = max(ffe_ctl.search_start,
7809                                    first_logical_byte(fs_info, 0));
7810         ffe_ctl.search_start = max(ffe_ctl.search_start, hint_byte);
7811         if (ffe_ctl.search_start == hint_byte) {
7812                 block_group = btrfs_lookup_block_group(fs_info,
7813                                                        ffe_ctl.search_start);
7814                 /*
7815                  * we don't want to use the block group if it doesn't match our
7816                  * allocation bits, or if its not cached.
7817                  *
7818                  * However if we are re-searching with an ideal block group
7819                  * picked out then we don't care that the block group is cached.
7820                  */
7821                 if (block_group && block_group_bits(block_group, flags) &&
7822                     block_group->cached != BTRFS_CACHE_NO) {
7823                         down_read(&space_info->groups_sem);
7824                         if (list_empty(&block_group->list) ||
7825                             block_group->ro) {
7826                                 /*
7827                                  * someone is removing this block group,
7828                                  * we can't jump into the have_block_group
7829                                  * target because our list pointers are not
7830                                  * valid
7831                                  */
7832                                 btrfs_put_block_group(block_group);
7833                                 up_read(&space_info->groups_sem);
7834                         } else {
7835                                 ffe_ctl.index = btrfs_bg_flags_to_raid_index(
7836                                                 block_group->flags);
7837                                 btrfs_lock_block_group(block_group, delalloc);
7838                                 goto have_block_group;
7839                         }
7840                 } else if (block_group) {
7841                         btrfs_put_block_group(block_group);
7842                 }
7843         }
7844 search:
7845         ffe_ctl.have_caching_bg = false;
7846         if (ffe_ctl.index == btrfs_bg_flags_to_raid_index(flags) ||
7847             ffe_ctl.index == 0)
7848                 full_search = true;
7849         down_read(&space_info->groups_sem);
7850         list_for_each_entry(block_group,
7851                             &space_info->block_groups[ffe_ctl.index], list) {
7852                 /* If the block group is read-only, we can skip it entirely. */
7853                 if (unlikely(block_group->ro))
7854                         continue;
7855
7856                 btrfs_grab_block_group(block_group, delalloc);
7857                 ffe_ctl.search_start = block_group->key.objectid;
7858
7859                 /*
7860                  * this can happen if we end up cycling through all the
7861                  * raid types, but we want to make sure we only allocate
7862                  * for the proper type.
7863                  */
7864                 if (!block_group_bits(block_group, flags)) {
7865                         u64 extra = BTRFS_BLOCK_GROUP_DUP |
7866                                 BTRFS_BLOCK_GROUP_RAID1 |
7867                                 BTRFS_BLOCK_GROUP_RAID5 |
7868                                 BTRFS_BLOCK_GROUP_RAID6 |
7869                                 BTRFS_BLOCK_GROUP_RAID10;
7870
7871                         /*
7872                          * if they asked for extra copies and this block group
7873                          * doesn't provide them, bail.  This does allow us to
7874                          * fill raid0 from raid1.
7875                          */
7876                         if ((flags & extra) && !(block_group->flags & extra))
7877                                 goto loop;
7878                 }
7879
7880 have_block_group:
7881                 ffe_ctl.cached = block_group_cache_done(block_group);
7882                 if (unlikely(!ffe_ctl.cached)) {
7883                         ffe_ctl.have_caching_bg = true;
7884                         ret = cache_block_group(block_group, 0);
7885                         BUG_ON(ret < 0);
7886                         ret = 0;
7887                 }
7888
7889                 if (unlikely(block_group->cached == BTRFS_CACHE_ERROR))
7890                         goto loop;
7891
7892                 /*
7893                  * Ok we want to try and use the cluster allocator, so
7894                  * lets look there
7895                  */
7896                 if (last_ptr && use_cluster) {
7897                         struct btrfs_block_group_cache *cluster_bg = NULL;
7898
7899                         ret = find_free_extent_clustered(block_group, last_ptr,
7900                                                          &ffe_ctl, &cluster_bg);
7901
7902                         if (ret == 0) {
7903                                 if (cluster_bg && cluster_bg != block_group) {
7904                                         btrfs_release_block_group(block_group,
7905                                                                   delalloc);
7906                                         block_group = cluster_bg;
7907                                 }
7908                                 goto checks;
7909                         } else if (ret == -EAGAIN) {
7910                                 goto have_block_group;
7911                         } else if (ret > 0) {
7912                                 goto loop;
7913                         }
7914                         /* ret == -ENOENT case falls through */
7915                 }
7916
7917                 ret = find_free_extent_unclustered(block_group, last_ptr,
7918                                                    &ffe_ctl);
7919                 if (ret == -EAGAIN)
7920                         goto have_block_group;
7921                 else if (ret > 0)
7922                         goto loop;
7923                 /* ret == 0 case falls through */
7924 checks:
7925                 ffe_ctl.search_start = round_up(ffe_ctl.found_offset,
7926                                              fs_info->stripesize);
7927
7928                 /* move on to the next group */
7929                 if (ffe_ctl.search_start + num_bytes >
7930                     block_group->key.objectid + block_group->key.offset) {
7931                         btrfs_add_free_space(block_group, ffe_ctl.found_offset,
7932                                              num_bytes);
7933                         goto loop;
7934                 }
7935
7936                 if (ffe_ctl.found_offset < ffe_ctl.search_start)
7937                         btrfs_add_free_space(block_group, ffe_ctl.found_offset,
7938                                 ffe_ctl.search_start - ffe_ctl.found_offset);
7939
7940                 ret = btrfs_add_reserved_bytes(block_group, ram_bytes,
7941                                 num_bytes, delalloc);
7942                 if (ret == -EAGAIN) {
7943                         btrfs_add_free_space(block_group, ffe_ctl.found_offset,
7944                                              num_bytes);
7945                         goto loop;
7946                 }
7947                 btrfs_inc_block_group_reservations(block_group);
7948
7949                 /* we are all good, lets return */
7950                 ins->objectid = ffe_ctl.search_start;
7951                 ins->offset = num_bytes;
7952
7953                 trace_btrfs_reserve_extent(block_group, ffe_ctl.search_start,
7954                                            num_bytes);
7955                 btrfs_release_block_group(block_group, delalloc);
7956                 break;
7957 loop:
7958                 ffe_ctl.retry_clustered = false;
7959                 ffe_ctl.retry_unclustered = false;
7960                 BUG_ON(btrfs_bg_flags_to_raid_index(block_group->flags) !=
7961                        ffe_ctl.index);
7962                 btrfs_release_block_group(block_group, delalloc);
7963                 cond_resched();
7964         }
7965         up_read(&space_info->groups_sem);
7966
7967         ret = find_free_extent_update_loop(fs_info, last_ptr, ins, &ffe_ctl,
7968                                            full_search, use_cluster);
7969         if (ret > 0)
7970                 goto search;
7971
7972         if (ret == -ENOSPC) {
7973                 /*
7974                  * Use ffe_ctl->total_free_space as fallback if we can't find
7975                  * any contiguous hole.
7976                  */
7977                 if (!ffe_ctl.max_extent_size)
7978                         ffe_ctl.max_extent_size = ffe_ctl.total_free_space;
7979                 spin_lock(&space_info->lock);
7980                 space_info->max_extent_size = ffe_ctl.max_extent_size;
7981                 spin_unlock(&space_info->lock);
7982                 ins->offset = ffe_ctl.max_extent_size;
7983         }
7984         return ret;
7985 }
7986
7987 #define DUMP_BLOCK_RSV(fs_info, rsv_name)                               \
7988 do {                                                                    \
7989         struct btrfs_block_rsv *__rsv = &(fs_info)->rsv_name;           \
7990         spin_lock(&__rsv->lock);                                        \
7991         btrfs_info(fs_info, #rsv_name ": size %llu reserved %llu",      \
7992                    __rsv->size, __rsv->reserved);                       \
7993         spin_unlock(&__rsv->lock);                                      \
7994 } while (0)
7995
7996 static void dump_space_info(struct btrfs_fs_info *fs_info,
7997                             struct btrfs_space_info *info, u64 bytes,
7998                             int dump_block_groups)
7999 {
8000         struct btrfs_block_group_cache *cache;
8001         int index = 0;
8002
8003         spin_lock(&info->lock);
8004         btrfs_info(fs_info, "space_info %llu has %llu free, is %sfull",
8005                    info->flags,
8006                    info->total_bytes - btrfs_space_info_used(info, true),
8007                    info->full ? "" : "not ");
8008         btrfs_info(fs_info,
8009                 "space_info total=%llu, used=%llu, pinned=%llu, reserved=%llu, may_use=%llu, readonly=%llu",
8010                 info->total_bytes, info->bytes_used, info->bytes_pinned,
8011                 info->bytes_reserved, info->bytes_may_use,
8012                 info->bytes_readonly);
8013         spin_unlock(&info->lock);
8014
8015         DUMP_BLOCK_RSV(fs_info, global_block_rsv);
8016         DUMP_BLOCK_RSV(fs_info, trans_block_rsv);
8017         DUMP_BLOCK_RSV(fs_info, chunk_block_rsv);
8018         DUMP_BLOCK_RSV(fs_info, delayed_block_rsv);
8019         DUMP_BLOCK_RSV(fs_info, delayed_refs_rsv);
8020
8021         if (!dump_block_groups)
8022                 return;
8023
8024         down_read(&info->groups_sem);
8025 again:
8026         list_for_each_entry(cache, &info->block_groups[index], list) {
8027                 spin_lock(&cache->lock);
8028                 btrfs_info(fs_info,
8029                         "block group %llu has %llu bytes, %llu used %llu pinned %llu reserved %s",
8030                         cache->key.objectid, cache->key.offset,
8031                         btrfs_block_group_used(&cache->item), cache->pinned,
8032                         cache->reserved, cache->ro ? "[readonly]" : "");
8033                 btrfs_dump_free_space(cache, bytes);
8034                 spin_unlock(&cache->lock);
8035         }
8036         if (++index < BTRFS_NR_RAID_TYPES)
8037                 goto again;
8038         up_read(&info->groups_sem);
8039 }
8040
8041 /*
8042  * btrfs_reserve_extent - entry point to the extent allocator. Tries to find a
8043  *                        hole that is at least as big as @num_bytes.
8044  *
8045  * @root           -    The root that will contain this extent
8046  *
8047  * @ram_bytes      -    The amount of space in ram that @num_bytes take. This
8048  *                      is used for accounting purposes. This value differs
8049  *                      from @num_bytes only in the case of compressed extents.
8050  *
8051  * @num_bytes      -    Number of bytes to allocate on-disk.
8052  *
8053  * @min_alloc_size -    Indicates the minimum amount of space that the
8054  *                      allocator should try to satisfy. In some cases
8055  *                      @num_bytes may be larger than what is required and if
8056  *                      the filesystem is fragmented then allocation fails.
8057  *                      However, the presence of @min_alloc_size gives a
8058  *                      chance to try and satisfy the smaller allocation.
8059  *
8060  * @empty_size     -    A hint that you plan on doing more COW. This is the
8061  *                      size in bytes the allocator should try to find free
8062  *                      next to the block it returns.  This is just a hint and
8063  *                      may be ignored by the allocator.
8064  *
8065  * @hint_byte      -    Hint to the allocator to start searching above the byte
8066  *                      address passed. It might be ignored.
8067  *
8068  * @ins            -    This key is modified to record the found hole. It will
8069  *                      have the following values:
8070  *                      ins->objectid == start position
8071  *                      ins->flags = BTRFS_EXTENT_ITEM_KEY
8072  *                      ins->offset == the size of the hole.
8073  *
8074  * @is_data        -    Boolean flag indicating whether an extent is
8075  *                      allocated for data (true) or metadata (false)
8076  *
8077  * @delalloc       -    Boolean flag indicating whether this allocation is for
8078  *                      delalloc or not. If 'true' data_rwsem of block groups
8079  *                      is going to be acquired.
8080  *
8081  *
8082  * Returns 0 when an allocation succeeded or < 0 when an error occurred. In
8083  * case -ENOSPC is returned then @ins->offset will contain the size of the
8084  * largest available hole the allocator managed to find.
8085  */
8086 int btrfs_reserve_extent(struct btrfs_root *root, u64 ram_bytes,
8087                          u64 num_bytes, u64 min_alloc_size,
8088                          u64 empty_size, u64 hint_byte,
8089                          struct btrfs_key *ins, int is_data, int delalloc)
8090 {
8091         struct btrfs_fs_info *fs_info = root->fs_info;
8092         bool final_tried = num_bytes == min_alloc_size;
8093         u64 flags;
8094         int ret;
8095
8096         flags = get_alloc_profile_by_root(root, is_data);
8097 again:
8098         WARN_ON(num_bytes < fs_info->sectorsize);
8099         ret = find_free_extent(fs_info, ram_bytes, num_bytes, empty_size,
8100                                hint_byte, ins, flags, delalloc);
8101         if (!ret && !is_data) {
8102                 btrfs_dec_block_group_reservations(fs_info, ins->objectid);
8103         } else if (ret == -ENOSPC) {
8104                 if (!final_tried && ins->offset) {
8105                         num_bytes = min(num_bytes >> 1, ins->offset);
8106                         num_bytes = round_down(num_bytes,
8107                                                fs_info->sectorsize);
8108                         num_bytes = max(num_bytes, min_alloc_size);
8109                         ram_bytes = num_bytes;
8110                         if (num_bytes == min_alloc_size)
8111                                 final_tried = true;
8112                         goto again;
8113                 } else if (btrfs_test_opt(fs_info, ENOSPC_DEBUG)) {
8114                         struct btrfs_space_info *sinfo;
8115
8116                         sinfo = __find_space_info(fs_info, flags);
8117                         btrfs_err(fs_info,
8118                                   "allocation failed flags %llu, wanted %llu",
8119                                   flags, num_bytes);
8120                         if (sinfo)
8121                                 dump_space_info(fs_info, sinfo, num_bytes, 1);
8122                 }
8123         }
8124
8125         return ret;
8126 }
8127
8128 static int __btrfs_free_reserved_extent(struct btrfs_fs_info *fs_info,
8129                                         u64 start, u64 len,
8130                                         int pin, int delalloc)
8131 {
8132         struct btrfs_block_group_cache *cache;
8133         int ret = 0;
8134
8135         cache = btrfs_lookup_block_group(fs_info, start);
8136         if (!cache) {
8137                 btrfs_err(fs_info, "Unable to find block group for %llu",
8138                           start);
8139                 return -ENOSPC;
8140         }
8141
8142         if (pin)
8143                 pin_down_extent(cache, start, len, 1);
8144         else {
8145                 if (btrfs_test_opt(fs_info, DISCARD))
8146                         ret = btrfs_discard_extent(fs_info, start, len, NULL);
8147                 btrfs_add_free_space(cache, start, len);
8148                 btrfs_free_reserved_bytes(cache, len, delalloc);
8149                 trace_btrfs_reserved_extent_free(fs_info, start, len);
8150         }
8151
8152         btrfs_put_block_group(cache);
8153         return ret;
8154 }
8155
8156 int btrfs_free_reserved_extent(struct btrfs_fs_info *fs_info,
8157                                u64 start, u64 len, int delalloc)
8158 {
8159         return __btrfs_free_reserved_extent(fs_info, start, len, 0, delalloc);
8160 }
8161
8162 int btrfs_free_and_pin_reserved_extent(struct btrfs_fs_info *fs_info,
8163                                        u64 start, u64 len)
8164 {
8165         return __btrfs_free_reserved_extent(fs_info, start, len, 1, 0);
8166 }
8167
8168 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
8169                                       u64 parent, u64 root_objectid,
8170                                       u64 flags, u64 owner, u64 offset,
8171                                       struct btrfs_key *ins, int ref_mod)
8172 {
8173         struct btrfs_fs_info *fs_info = trans->fs_info;
8174         int ret;
8175         struct btrfs_extent_item *extent_item;
8176         struct btrfs_extent_inline_ref *iref;
8177         struct btrfs_path *path;
8178         struct extent_buffer *leaf;
8179         int type;
8180         u32 size;
8181
8182         if (parent > 0)
8183                 type = BTRFS_SHARED_DATA_REF_KEY;
8184         else
8185                 type = BTRFS_EXTENT_DATA_REF_KEY;
8186
8187         size = sizeof(*extent_item) + btrfs_extent_inline_ref_size(type);
8188
8189         path = btrfs_alloc_path();
8190         if (!path)
8191                 return -ENOMEM;
8192
8193         path->leave_spinning = 1;
8194         ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path,
8195                                       ins, size);
8196         if (ret) {
8197                 btrfs_free_path(path);
8198                 return ret;
8199         }
8200
8201         leaf = path->nodes[0];
8202         extent_item = btrfs_item_ptr(leaf, path->slots[0],
8203                                      struct btrfs_extent_item);
8204         btrfs_set_extent_refs(leaf, extent_item, ref_mod);
8205         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
8206         btrfs_set_extent_flags(leaf, extent_item,
8207                                flags | BTRFS_EXTENT_FLAG_DATA);
8208
8209         iref = (struct btrfs_extent_inline_ref *)(extent_item + 1);
8210         btrfs_set_extent_inline_ref_type(leaf, iref, type);
8211         if (parent > 0) {
8212                 struct btrfs_shared_data_ref *ref;
8213                 ref = (struct btrfs_shared_data_ref *)(iref + 1);
8214                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
8215                 btrfs_set_shared_data_ref_count(leaf, ref, ref_mod);
8216         } else {
8217                 struct btrfs_extent_data_ref *ref;
8218                 ref = (struct btrfs_extent_data_ref *)(&iref->offset);
8219                 btrfs_set_extent_data_ref_root(leaf, ref, root_objectid);
8220                 btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
8221                 btrfs_set_extent_data_ref_offset(leaf, ref, offset);
8222                 btrfs_set_extent_data_ref_count(leaf, ref, ref_mod);
8223         }
8224
8225         btrfs_mark_buffer_dirty(path->nodes[0]);
8226         btrfs_free_path(path);
8227
8228         ret = remove_from_free_space_tree(trans, ins->objectid, ins->offset);
8229         if (ret)
8230                 return ret;
8231
8232         ret = update_block_group(trans, ins->objectid, ins->offset, 1);
8233         if (ret) { /* -ENOENT, logic error */
8234                 btrfs_err(fs_info, "update block group failed for %llu %llu",
8235                         ins->objectid, ins->offset);
8236                 BUG();
8237         }
8238         trace_btrfs_reserved_extent_alloc(fs_info, ins->objectid, ins->offset);
8239         return ret;
8240 }
8241
8242 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
8243                                      struct btrfs_delayed_ref_node *node,
8244                                      struct btrfs_delayed_extent_op *extent_op)
8245 {
8246         struct btrfs_fs_info *fs_info = trans->fs_info;
8247         int ret;
8248         struct btrfs_extent_item *extent_item;
8249         struct btrfs_key extent_key;
8250         struct btrfs_tree_block_info *block_info;
8251         struct btrfs_extent_inline_ref *iref;
8252         struct btrfs_path *path;
8253         struct extent_buffer *leaf;
8254         struct btrfs_delayed_tree_ref *ref;
8255         u32 size = sizeof(*extent_item) + sizeof(*iref);
8256         u64 num_bytes;
8257         u64 flags = extent_op->flags_to_set;
8258         bool skinny_metadata = btrfs_fs_incompat(fs_info, SKINNY_METADATA);
8259
8260         ref = btrfs_delayed_node_to_tree_ref(node);
8261
8262         extent_key.objectid = node->bytenr;
8263         if (skinny_metadata) {
8264                 extent_key.offset = ref->level;
8265                 extent_key.type = BTRFS_METADATA_ITEM_KEY;
8266                 num_bytes = fs_info->nodesize;
8267         } else {
8268                 extent_key.offset = node->num_bytes;
8269                 extent_key.type = BTRFS_EXTENT_ITEM_KEY;
8270                 size += sizeof(*block_info);
8271                 num_bytes = node->num_bytes;
8272         }
8273
8274         path = btrfs_alloc_path();
8275         if (!path)
8276                 return -ENOMEM;
8277
8278         path->leave_spinning = 1;
8279         ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path,
8280                                       &extent_key, size);
8281         if (ret) {
8282                 btrfs_free_path(path);
8283                 return ret;
8284         }
8285
8286         leaf = path->nodes[0];
8287         extent_item = btrfs_item_ptr(leaf, path->slots[0],
8288                                      struct btrfs_extent_item);
8289         btrfs_set_extent_refs(leaf, extent_item, 1);
8290         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
8291         btrfs_set_extent_flags(leaf, extent_item,
8292                                flags | BTRFS_EXTENT_FLAG_TREE_BLOCK);
8293
8294         if (skinny_metadata) {
8295                 iref = (struct btrfs_extent_inline_ref *)(extent_item + 1);
8296         } else {
8297                 block_info = (struct btrfs_tree_block_info *)(extent_item + 1);
8298                 btrfs_set_tree_block_key(leaf, block_info, &extent_op->key);
8299                 btrfs_set_tree_block_level(leaf, block_info, ref->level);
8300                 iref = (struct btrfs_extent_inline_ref *)(block_info + 1);
8301         }
8302
8303         if (node->type == BTRFS_SHARED_BLOCK_REF_KEY) {
8304                 BUG_ON(!(flags & BTRFS_BLOCK_FLAG_FULL_BACKREF));
8305                 btrfs_set_extent_inline_ref_type(leaf, iref,
8306                                                  BTRFS_SHARED_BLOCK_REF_KEY);
8307                 btrfs_set_extent_inline_ref_offset(leaf, iref, ref->parent);
8308         } else {
8309                 btrfs_set_extent_inline_ref_type(leaf, iref,
8310                                                  BTRFS_TREE_BLOCK_REF_KEY);
8311                 btrfs_set_extent_inline_ref_offset(leaf, iref, ref->root);
8312         }
8313
8314         btrfs_mark_buffer_dirty(leaf);
8315         btrfs_free_path(path);
8316
8317         ret = remove_from_free_space_tree(trans, extent_key.objectid,
8318                                           num_bytes);
8319         if (ret)
8320                 return ret;
8321
8322         ret = update_block_group(trans, extent_key.objectid,
8323                                  fs_info->nodesize, 1);
8324         if (ret) { /* -ENOENT, logic error */
8325                 btrfs_err(fs_info, "update block group failed for %llu %llu",
8326                         extent_key.objectid, extent_key.offset);
8327                 BUG();
8328         }
8329
8330         trace_btrfs_reserved_extent_alloc(fs_info, extent_key.objectid,
8331                                           fs_info->nodesize);
8332         return ret;
8333 }
8334
8335 int btrfs_alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
8336                                      struct btrfs_root *root, u64 owner,
8337                                      u64 offset, u64 ram_bytes,
8338                                      struct btrfs_key *ins)
8339 {
8340         struct btrfs_ref generic_ref = { 0 };
8341         int ret;
8342
8343         BUG_ON(root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID);
8344
8345         btrfs_init_generic_ref(&generic_ref, BTRFS_ADD_DELAYED_EXTENT,
8346                                ins->objectid, ins->offset, 0);
8347         btrfs_init_data_ref(&generic_ref, root->root_key.objectid, owner, offset);
8348         btrfs_ref_tree_mod(root->fs_info, &generic_ref);
8349         ret = btrfs_add_delayed_data_ref(trans, &generic_ref,
8350                                          ram_bytes, NULL, NULL);
8351         return ret;
8352 }
8353
8354 /*
8355  * this is used by the tree logging recovery code.  It records that
8356  * an extent has been allocated and makes sure to clear the free
8357  * space cache bits as well
8358  */
8359 int btrfs_alloc_logged_file_extent(struct btrfs_trans_handle *trans,
8360                                    u64 root_objectid, u64 owner, u64 offset,
8361                                    struct btrfs_key *ins)
8362 {
8363         struct btrfs_fs_info *fs_info = trans->fs_info;
8364         int ret;
8365         struct btrfs_block_group_cache *block_group;
8366         struct btrfs_space_info *space_info;
8367
8368         /*
8369          * Mixed block groups will exclude before processing the log so we only
8370          * need to do the exclude dance if this fs isn't mixed.
8371          */
8372         if (!btrfs_fs_incompat(fs_info, MIXED_GROUPS)) {
8373                 ret = __exclude_logged_extent(fs_info, ins->objectid,
8374                                               ins->offset);
8375                 if (ret)
8376                         return ret;
8377         }
8378
8379         block_group = btrfs_lookup_block_group(fs_info, ins->objectid);
8380         if (!block_group)
8381                 return -EINVAL;
8382
8383         space_info = block_group->space_info;
8384         spin_lock(&space_info->lock);
8385         spin_lock(&block_group->lock);
8386         space_info->bytes_reserved += ins->offset;
8387         block_group->reserved += ins->offset;
8388         spin_unlock(&block_group->lock);
8389         spin_unlock(&space_info->lock);
8390
8391         ret = alloc_reserved_file_extent(trans, 0, root_objectid, 0, owner,
8392                                          offset, ins, 1);
8393         btrfs_put_block_group(block_group);
8394         return ret;
8395 }
8396
8397 static struct extent_buffer *
8398 btrfs_init_new_buffer(struct btrfs_trans_handle *trans, struct btrfs_root *root,
8399                       u64 bytenr, int level, u64 owner)
8400 {
8401         struct btrfs_fs_info *fs_info = root->fs_info;
8402         struct extent_buffer *buf;
8403
8404         buf = btrfs_find_create_tree_block(fs_info, bytenr);
8405         if (IS_ERR(buf))
8406                 return buf;
8407
8408         /*
8409          * Extra safety check in case the extent tree is corrupted and extent
8410          * allocator chooses to use a tree block which is already used and
8411          * locked.
8412          */
8413         if (buf->lock_owner == current->pid) {
8414                 btrfs_err_rl(fs_info,
8415 "tree block %llu owner %llu already locked by pid=%d, extent tree corruption detected",
8416                         buf->start, btrfs_header_owner(buf), current->pid);
8417                 free_extent_buffer(buf);
8418                 return ERR_PTR(-EUCLEAN);
8419         }
8420
8421         btrfs_set_buffer_lockdep_class(root->root_key.objectid, buf, level);
8422         btrfs_tree_lock(buf);
8423         btrfs_clean_tree_block(buf);
8424         clear_bit(EXTENT_BUFFER_STALE, &buf->bflags);
8425
8426         btrfs_set_lock_blocking_write(buf);
8427         set_extent_buffer_uptodate(buf);
8428
8429         memzero_extent_buffer(buf, 0, sizeof(struct btrfs_header));
8430         btrfs_set_header_level(buf, level);
8431         btrfs_set_header_bytenr(buf, buf->start);
8432         btrfs_set_header_generation(buf, trans->transid);
8433         btrfs_set_header_backref_rev(buf, BTRFS_MIXED_BACKREF_REV);
8434         btrfs_set_header_owner(buf, owner);
8435         write_extent_buffer_fsid(buf, fs_info->fs_devices->metadata_uuid);
8436         write_extent_buffer_chunk_tree_uuid(buf, fs_info->chunk_tree_uuid);
8437         if (root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID) {
8438                 buf->log_index = root->log_transid % 2;
8439                 /*
8440                  * we allow two log transactions at a time, use different
8441                  * EXTENT bit to differentiate dirty pages.
8442                  */
8443                 if (buf->log_index == 0)
8444                         set_extent_dirty(&root->dirty_log_pages, buf->start,
8445                                         buf->start + buf->len - 1, GFP_NOFS);
8446                 else
8447                         set_extent_new(&root->dirty_log_pages, buf->start,
8448                                         buf->start + buf->len - 1);
8449         } else {
8450                 buf->log_index = -1;
8451                 set_extent_dirty(&trans->transaction->dirty_pages, buf->start,
8452                          buf->start + buf->len - 1, GFP_NOFS);
8453         }
8454         trans->dirty = true;
8455         /* this returns a buffer locked for blocking */
8456         return buf;
8457 }
8458
8459 static struct btrfs_block_rsv *
8460 use_block_rsv(struct btrfs_trans_handle *trans,
8461               struct btrfs_root *root, u32 blocksize)
8462 {
8463         struct btrfs_fs_info *fs_info = root->fs_info;
8464         struct btrfs_block_rsv *block_rsv;
8465         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
8466         int ret;
8467         bool global_updated = false;
8468
8469         block_rsv = get_block_rsv(trans, root);
8470
8471         if (unlikely(block_rsv->size == 0))
8472                 goto try_reserve;
8473 again:
8474         ret = block_rsv_use_bytes(block_rsv, blocksize);
8475         if (!ret)
8476                 return block_rsv;
8477
8478         if (block_rsv->failfast)
8479                 return ERR_PTR(ret);
8480
8481         if (block_rsv->type == BTRFS_BLOCK_RSV_GLOBAL && !global_updated) {
8482                 global_updated = true;
8483                 update_global_block_rsv(fs_info);
8484                 goto again;
8485         }
8486
8487         /*
8488          * The global reserve still exists to save us from ourselves, so don't
8489          * warn_on if we are short on our delayed refs reserve.
8490          */
8491         if (block_rsv->type != BTRFS_BLOCK_RSV_DELREFS &&
8492             btrfs_test_opt(fs_info, ENOSPC_DEBUG)) {
8493                 static DEFINE_RATELIMIT_STATE(_rs,
8494                                 DEFAULT_RATELIMIT_INTERVAL * 10,
8495                                 /*DEFAULT_RATELIMIT_BURST*/ 1);
8496                 if (__ratelimit(&_rs))
8497                         WARN(1, KERN_DEBUG
8498                                 "BTRFS: block rsv returned %d\n", ret);
8499         }
8500 try_reserve:
8501         ret = reserve_metadata_bytes(root, block_rsv, blocksize,
8502                                      BTRFS_RESERVE_NO_FLUSH);
8503         if (!ret)
8504                 return block_rsv;
8505         /*
8506          * If we couldn't reserve metadata bytes try and use some from
8507          * the global reserve if its space type is the same as the global
8508          * reservation.
8509          */
8510         if (block_rsv->type != BTRFS_BLOCK_RSV_GLOBAL &&
8511             block_rsv->space_info == global_rsv->space_info) {
8512                 ret = block_rsv_use_bytes(global_rsv, blocksize);
8513                 if (!ret)
8514                         return global_rsv;
8515         }
8516         return ERR_PTR(ret);
8517 }
8518
8519 static void unuse_block_rsv(struct btrfs_fs_info *fs_info,
8520                             struct btrfs_block_rsv *block_rsv, u32 blocksize)
8521 {
8522         block_rsv_add_bytes(block_rsv, blocksize, false);
8523         block_rsv_release_bytes(fs_info, block_rsv, NULL, 0, NULL);
8524 }
8525
8526 /*
8527  * finds a free extent and does all the dirty work required for allocation
8528  * returns the tree buffer or an ERR_PTR on error.
8529  */
8530 struct extent_buffer *btrfs_alloc_tree_block(struct btrfs_trans_handle *trans,
8531                                              struct btrfs_root *root,
8532                                              u64 parent, u64 root_objectid,
8533                                              const struct btrfs_disk_key *key,
8534                                              int level, u64 hint,
8535                                              u64 empty_size)
8536 {
8537         struct btrfs_fs_info *fs_info = root->fs_info;
8538         struct btrfs_key ins;
8539         struct btrfs_block_rsv *block_rsv;
8540         struct extent_buffer *buf;
8541         struct btrfs_delayed_extent_op *extent_op;
8542         struct btrfs_ref generic_ref = { 0 };
8543         u64 flags = 0;
8544         int ret;
8545         u32 blocksize = fs_info->nodesize;
8546         bool skinny_metadata = btrfs_fs_incompat(fs_info, SKINNY_METADATA);
8547
8548 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
8549         if (btrfs_is_testing(fs_info)) {
8550                 buf = btrfs_init_new_buffer(trans, root, root->alloc_bytenr,
8551                                             level, root_objectid);
8552                 if (!IS_ERR(buf))
8553                         root->alloc_bytenr += blocksize;
8554                 return buf;
8555         }
8556 #endif
8557
8558         block_rsv = use_block_rsv(trans, root, blocksize);
8559         if (IS_ERR(block_rsv))
8560                 return ERR_CAST(block_rsv);
8561
8562         ret = btrfs_reserve_extent(root, blocksize, blocksize, blocksize,
8563                                    empty_size, hint, &ins, 0, 0);
8564         if (ret)
8565                 goto out_unuse;
8566
8567         buf = btrfs_init_new_buffer(trans, root, ins.objectid, level,
8568                                     root_objectid);
8569         if (IS_ERR(buf)) {
8570                 ret = PTR_ERR(buf);
8571                 goto out_free_reserved;
8572         }
8573
8574         if (root_objectid == BTRFS_TREE_RELOC_OBJECTID) {
8575                 if (parent == 0)
8576                         parent = ins.objectid;
8577                 flags |= BTRFS_BLOCK_FLAG_FULL_BACKREF;
8578         } else
8579                 BUG_ON(parent > 0);
8580
8581         if (root_objectid != BTRFS_TREE_LOG_OBJECTID) {
8582                 extent_op = btrfs_alloc_delayed_extent_op();
8583                 if (!extent_op) {
8584                         ret = -ENOMEM;
8585                         goto out_free_buf;
8586                 }
8587                 if (key)
8588                         memcpy(&extent_op->key, key, sizeof(extent_op->key));
8589                 else
8590                         memset(&extent_op->key, 0, sizeof(extent_op->key));
8591                 extent_op->flags_to_set = flags;
8592                 extent_op->update_key = skinny_metadata ? false : true;
8593                 extent_op->update_flags = true;
8594                 extent_op->is_data = false;
8595                 extent_op->level = level;
8596
8597                 btrfs_init_generic_ref(&generic_ref, BTRFS_ADD_DELAYED_EXTENT,
8598                                        ins.objectid, ins.offset, parent);
8599                 generic_ref.real_root = root->root_key.objectid;
8600                 btrfs_init_tree_ref(&generic_ref, level, root_objectid);
8601                 btrfs_ref_tree_mod(fs_info, &generic_ref);
8602                 ret = btrfs_add_delayed_tree_ref(trans, &generic_ref,
8603                                                  extent_op, NULL, NULL);
8604                 if (ret)
8605                         goto out_free_delayed;
8606         }
8607         return buf;
8608
8609 out_free_delayed:
8610         btrfs_free_delayed_extent_op(extent_op);
8611 out_free_buf:
8612         free_extent_buffer(buf);
8613 out_free_reserved:
8614         btrfs_free_reserved_extent(fs_info, ins.objectid, ins.offset, 0);
8615 out_unuse:
8616         unuse_block_rsv(fs_info, block_rsv, blocksize);
8617         return ERR_PTR(ret);
8618 }
8619
8620 struct walk_control {
8621         u64 refs[BTRFS_MAX_LEVEL];
8622         u64 flags[BTRFS_MAX_LEVEL];
8623         struct btrfs_key update_progress;
8624         struct btrfs_key drop_progress;
8625         int drop_level;
8626         int stage;
8627         int level;
8628         int shared_level;
8629         int update_ref;
8630         int keep_locks;
8631         int reada_slot;
8632         int reada_count;
8633         int restarted;
8634 };
8635
8636 #define DROP_REFERENCE  1
8637 #define UPDATE_BACKREF  2
8638
8639 static noinline void reada_walk_down(struct btrfs_trans_handle *trans,
8640                                      struct btrfs_root *root,
8641                                      struct walk_control *wc,
8642                                      struct btrfs_path *path)
8643 {
8644         struct btrfs_fs_info *fs_info = root->fs_info;
8645         u64 bytenr;
8646         u64 generation;
8647         u64 refs;
8648         u64 flags;
8649         u32 nritems;
8650         struct btrfs_key key;
8651         struct extent_buffer *eb;
8652         int ret;
8653         int slot;
8654         int nread = 0;
8655
8656         if (path->slots[wc->level] < wc->reada_slot) {
8657                 wc->reada_count = wc->reada_count * 2 / 3;
8658                 wc->reada_count = max(wc->reada_count, 2);
8659         } else {
8660                 wc->reada_count = wc->reada_count * 3 / 2;
8661                 wc->reada_count = min_t(int, wc->reada_count,
8662                                         BTRFS_NODEPTRS_PER_BLOCK(fs_info));
8663         }
8664
8665         eb = path->nodes[wc->level];
8666         nritems = btrfs_header_nritems(eb);
8667
8668         for (slot = path->slots[wc->level]; slot < nritems; slot++) {
8669                 if (nread >= wc->reada_count)
8670                         break;
8671
8672                 cond_resched();
8673                 bytenr = btrfs_node_blockptr(eb, slot);
8674                 generation = btrfs_node_ptr_generation(eb, slot);
8675
8676                 if (slot == path->slots[wc->level])
8677                         goto reada;
8678
8679                 if (wc->stage == UPDATE_BACKREF &&
8680                     generation <= root->root_key.offset)
8681                         continue;
8682
8683                 /* We don't lock the tree block, it's OK to be racy here */
8684                 ret = btrfs_lookup_extent_info(trans, fs_info, bytenr,
8685                                                wc->level - 1, 1, &refs,
8686                                                &flags);
8687                 /* We don't care about errors in readahead. */
8688                 if (ret < 0)
8689                         continue;
8690                 BUG_ON(refs == 0);
8691
8692                 if (wc->stage == DROP_REFERENCE) {
8693                         if (refs == 1)
8694                                 goto reada;
8695
8696                         if (wc->level == 1 &&
8697                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
8698                                 continue;
8699                         if (!wc->update_ref ||
8700                             generation <= root->root_key.offset)
8701                                 continue;
8702                         btrfs_node_key_to_cpu(eb, &key, slot);
8703                         ret = btrfs_comp_cpu_keys(&key,
8704                                                   &wc->update_progress);
8705                         if (ret < 0)
8706                                 continue;
8707                 } else {
8708                         if (wc->level == 1 &&
8709                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
8710                                 continue;
8711                 }
8712 reada:
8713                 readahead_tree_block(fs_info, bytenr);
8714                 nread++;
8715         }
8716         wc->reada_slot = slot;
8717 }
8718
8719 /*
8720  * helper to process tree block while walking down the tree.
8721  *
8722  * when wc->stage == UPDATE_BACKREF, this function updates
8723  * back refs for pointers in the block.
8724  *
8725  * NOTE: return value 1 means we should stop walking down.
8726  */
8727 static noinline int walk_down_proc(struct btrfs_trans_handle *trans,
8728                                    struct btrfs_root *root,
8729                                    struct btrfs_path *path,
8730                                    struct walk_control *wc, int lookup_info)
8731 {
8732         struct btrfs_fs_info *fs_info = root->fs_info;
8733         int level = wc->level;
8734         struct extent_buffer *eb = path->nodes[level];
8735         u64 flag = BTRFS_BLOCK_FLAG_FULL_BACKREF;
8736         int ret;
8737
8738         if (wc->stage == UPDATE_BACKREF &&
8739             btrfs_header_owner(eb) != root->root_key.objectid)
8740                 return 1;
8741
8742         /*
8743          * when reference count of tree block is 1, it won't increase
8744          * again. once full backref flag is set, we never clear it.
8745          */
8746         if (lookup_info &&
8747             ((wc->stage == DROP_REFERENCE && wc->refs[level] != 1) ||
8748              (wc->stage == UPDATE_BACKREF && !(wc->flags[level] & flag)))) {
8749                 BUG_ON(!path->locks[level]);
8750                 ret = btrfs_lookup_extent_info(trans, fs_info,
8751                                                eb->start, level, 1,
8752                                                &wc->refs[level],
8753                                                &wc->flags[level]);
8754                 BUG_ON(ret == -ENOMEM);
8755                 if (ret)
8756                         return ret;
8757                 BUG_ON(wc->refs[level] == 0);
8758         }
8759
8760         if (wc->stage == DROP_REFERENCE) {
8761                 if (wc->refs[level] > 1)
8762                         return 1;
8763
8764                 if (path->locks[level] && !wc->keep_locks) {
8765                         btrfs_tree_unlock_rw(eb, path->locks[level]);
8766                         path->locks[level] = 0;
8767                 }
8768                 return 0;
8769         }
8770
8771         /* wc->stage == UPDATE_BACKREF */
8772         if (!(wc->flags[level] & flag)) {
8773                 BUG_ON(!path->locks[level]);
8774                 ret = btrfs_inc_ref(trans, root, eb, 1);
8775                 BUG_ON(ret); /* -ENOMEM */
8776                 ret = btrfs_dec_ref(trans, root, eb, 0);
8777                 BUG_ON(ret); /* -ENOMEM */
8778                 ret = btrfs_set_disk_extent_flags(trans, eb->start,
8779                                                   eb->len, flag,
8780                                                   btrfs_header_level(eb), 0);
8781                 BUG_ON(ret); /* -ENOMEM */
8782                 wc->flags[level] |= flag;
8783         }
8784
8785         /*
8786          * the block is shared by multiple trees, so it's not good to
8787          * keep the tree lock
8788          */
8789         if (path->locks[level] && level > 0) {
8790                 btrfs_tree_unlock_rw(eb, path->locks[level]);
8791                 path->locks[level] = 0;
8792         }
8793         return 0;
8794 }
8795
8796 /*
8797  * This is used to verify a ref exists for this root to deal with a bug where we
8798  * would have a drop_progress key that hadn't been updated properly.
8799  */
8800 static int check_ref_exists(struct btrfs_trans_handle *trans,
8801                             struct btrfs_root *root, u64 bytenr, u64 parent,
8802                             int level)
8803 {
8804         struct btrfs_path *path;
8805         struct btrfs_extent_inline_ref *iref;
8806         int ret;
8807
8808         path = btrfs_alloc_path();
8809         if (!path)
8810                 return -ENOMEM;
8811
8812         ret = lookup_extent_backref(trans, path, &iref, bytenr,
8813                                     root->fs_info->nodesize, parent,
8814                                     root->root_key.objectid, level, 0);
8815         btrfs_free_path(path);
8816         if (ret == -ENOENT)
8817                 return 0;
8818         if (ret < 0)
8819                 return ret;
8820         return 1;
8821 }
8822
8823 /*
8824  * helper to process tree block pointer.
8825  *
8826  * when wc->stage == DROP_REFERENCE, this function checks
8827  * reference count of the block pointed to. if the block
8828  * is shared and we need update back refs for the subtree
8829  * rooted at the block, this function changes wc->stage to
8830  * UPDATE_BACKREF. if the block is shared and there is no
8831  * need to update back, this function drops the reference
8832  * to the block.
8833  *
8834  * NOTE: return value 1 means we should stop walking down.
8835  */
8836 static noinline int do_walk_down(struct btrfs_trans_handle *trans,
8837                                  struct btrfs_root *root,
8838                                  struct btrfs_path *path,
8839                                  struct walk_control *wc, int *lookup_info)
8840 {
8841         struct btrfs_fs_info *fs_info = root->fs_info;
8842         u64 bytenr;
8843         u64 generation;
8844         u64 parent;
8845         struct btrfs_key key;
8846         struct btrfs_key first_key;
8847         struct btrfs_ref ref = { 0 };
8848         struct extent_buffer *next;
8849         int level = wc->level;
8850         int reada = 0;
8851         int ret = 0;
8852         bool need_account = false;
8853
8854         generation = btrfs_node_ptr_generation(path->nodes[level],
8855                                                path->slots[level]);
8856         /*
8857          * if the lower level block was created before the snapshot
8858          * was created, we know there is no need to update back refs
8859          * for the subtree
8860          */
8861         if (wc->stage == UPDATE_BACKREF &&
8862             generation <= root->root_key.offset) {
8863                 *lookup_info = 1;
8864                 return 1;
8865         }
8866
8867         bytenr = btrfs_node_blockptr(path->nodes[level], path->slots[level]);
8868         btrfs_node_key_to_cpu(path->nodes[level], &first_key,
8869                               path->slots[level]);
8870
8871         next = find_extent_buffer(fs_info, bytenr);
8872         if (!next) {
8873                 next = btrfs_find_create_tree_block(fs_info, bytenr);
8874                 if (IS_ERR(next))
8875                         return PTR_ERR(next);
8876
8877                 btrfs_set_buffer_lockdep_class(root->root_key.objectid, next,
8878                                                level - 1);
8879                 reada = 1;
8880         }
8881         btrfs_tree_lock(next);
8882         btrfs_set_lock_blocking_write(next);
8883
8884         ret = btrfs_lookup_extent_info(trans, fs_info, bytenr, level - 1, 1,
8885                                        &wc->refs[level - 1],
8886                                        &wc->flags[level - 1]);
8887         if (ret < 0)
8888                 goto out_unlock;
8889
8890         if (unlikely(wc->refs[level - 1] == 0)) {
8891                 btrfs_err(fs_info, "Missing references.");
8892                 ret = -EIO;
8893                 goto out_unlock;
8894         }
8895         *lookup_info = 0;
8896
8897         if (wc->stage == DROP_REFERENCE) {
8898                 if (wc->refs[level - 1] > 1) {
8899                         need_account = true;
8900                         if (level == 1 &&
8901                             (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
8902                                 goto skip;
8903
8904                         if (!wc->update_ref ||
8905                             generation <= root->root_key.offset)
8906                                 goto skip;
8907
8908                         btrfs_node_key_to_cpu(path->nodes[level], &key,
8909                                               path->slots[level]);
8910                         ret = btrfs_comp_cpu_keys(&key, &wc->update_progress);
8911                         if (ret < 0)
8912                                 goto skip;
8913
8914                         wc->stage = UPDATE_BACKREF;
8915                         wc->shared_level = level - 1;
8916                 }
8917         } else {
8918                 if (level == 1 &&
8919                     (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
8920                         goto skip;
8921         }
8922
8923         if (!btrfs_buffer_uptodate(next, generation, 0)) {
8924                 btrfs_tree_unlock(next);
8925                 free_extent_buffer(next);
8926                 next = NULL;
8927                 *lookup_info = 1;
8928         }
8929
8930         if (!next) {
8931                 if (reada && level == 1)
8932                         reada_walk_down(trans, root, wc, path);
8933                 next = read_tree_block(fs_info, bytenr, generation, level - 1,
8934                                        &first_key);
8935                 if (IS_ERR(next)) {
8936                         return PTR_ERR(next);
8937                 } else if (!extent_buffer_uptodate(next)) {
8938                         free_extent_buffer(next);
8939                         return -EIO;
8940                 }
8941                 btrfs_tree_lock(next);
8942                 btrfs_set_lock_blocking_write(next);
8943         }
8944
8945         level--;
8946         ASSERT(level == btrfs_header_level(next));
8947         if (level != btrfs_header_level(next)) {
8948                 btrfs_err(root->fs_info, "mismatched level");
8949                 ret = -EIO;
8950                 goto out_unlock;
8951         }
8952         path->nodes[level] = next;
8953         path->slots[level] = 0;
8954         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
8955         wc->level = level;
8956         if (wc->level == 1)
8957                 wc->reada_slot = 0;
8958         return 0;
8959 skip:
8960         wc->refs[level - 1] = 0;
8961         wc->flags[level - 1] = 0;
8962         if (wc->stage == DROP_REFERENCE) {
8963                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF) {
8964                         parent = path->nodes[level]->start;
8965                 } else {
8966                         ASSERT(root->root_key.objectid ==
8967                                btrfs_header_owner(path->nodes[level]));
8968                         if (root->root_key.objectid !=
8969                             btrfs_header_owner(path->nodes[level])) {
8970                                 btrfs_err(root->fs_info,
8971                                                 "mismatched block owner");
8972                                 ret = -EIO;
8973                                 goto out_unlock;
8974                         }
8975                         parent = 0;
8976                 }
8977
8978                 /*
8979                  * If we had a drop_progress we need to verify the refs are set
8980                  * as expected.  If we find our ref then we know that from here
8981                  * on out everything should be correct, and we can clear the
8982                  * ->restarted flag.
8983                  */
8984                 if (wc->restarted) {
8985                         ret = check_ref_exists(trans, root, bytenr, parent,
8986                                                level - 1);
8987                         if (ret < 0)
8988                                 goto out_unlock;
8989                         if (ret == 0)
8990                                 goto no_delete;
8991                         ret = 0;
8992                         wc->restarted = 0;
8993                 }
8994
8995                 /*
8996                  * Reloc tree doesn't contribute to qgroup numbers, and we have
8997                  * already accounted them at merge time (replace_path),
8998                  * thus we could skip expensive subtree trace here.
8999                  */
9000                 if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID &&
9001                     need_account) {
9002                         ret = btrfs_qgroup_trace_subtree(trans, next,
9003                                                          generation, level - 1);
9004                         if (ret) {
9005                                 btrfs_err_rl(fs_info,
9006                                              "Error %d accounting shared subtree. Quota is out of sync, rescan required.",
9007                                              ret);
9008                         }
9009                 }
9010
9011                 /*
9012                  * We need to update the next key in our walk control so we can
9013                  * update the drop_progress key accordingly.  We don't care if
9014                  * find_next_key doesn't find a key because that means we're at
9015                  * the end and are going to clean up now.
9016                  */
9017                 wc->drop_level = level;
9018                 find_next_key(path, level, &wc->drop_progress);
9019
9020                 btrfs_init_generic_ref(&ref, BTRFS_DROP_DELAYED_REF, bytenr,
9021                                        fs_info->nodesize, parent);
9022                 btrfs_init_tree_ref(&ref, level - 1, root->root_key.objectid);
9023                 ret = btrfs_free_extent(trans, &ref);
9024                 if (ret)
9025                         goto out_unlock;
9026         }
9027 no_delete:
9028         *lookup_info = 1;
9029         ret = 1;
9030
9031 out_unlock:
9032         btrfs_tree_unlock(next);
9033         free_extent_buffer(next);
9034
9035         return ret;
9036 }
9037
9038 /*
9039  * helper to process tree block while walking up the tree.
9040  *
9041  * when wc->stage == DROP_REFERENCE, this function drops
9042  * reference count on the block.
9043  *
9044  * when wc->stage == UPDATE_BACKREF, this function changes
9045  * wc->stage back to DROP_REFERENCE if we changed wc->stage
9046  * to UPDATE_BACKREF previously while processing the block.
9047  *
9048  * NOTE: return value 1 means we should stop walking up.
9049  */
9050 static noinline int walk_up_proc(struct btrfs_trans_handle *trans,
9051                                  struct btrfs_root *root,
9052                                  struct btrfs_path *path,
9053                                  struct walk_control *wc)
9054 {
9055         struct btrfs_fs_info *fs_info = root->fs_info;
9056         int ret;
9057         int level = wc->level;
9058         struct extent_buffer *eb = path->nodes[level];
9059         u64 parent = 0;
9060
9061         if (wc->stage == UPDATE_BACKREF) {
9062                 BUG_ON(wc->shared_level < level);
9063                 if (level < wc->shared_level)
9064                         goto out;
9065
9066                 ret = find_next_key(path, level + 1, &wc->update_progress);
9067                 if (ret > 0)
9068                         wc->update_ref = 0;
9069
9070                 wc->stage = DROP_REFERENCE;
9071                 wc->shared_level = -1;
9072                 path->slots[level] = 0;
9073
9074                 /*
9075                  * check reference count again if the block isn't locked.
9076                  * we should start walking down the tree again if reference
9077                  * count is one.
9078                  */
9079                 if (!path->locks[level]) {
9080                         BUG_ON(level == 0);
9081                         btrfs_tree_lock(eb);
9082                         btrfs_set_lock_blocking_write(eb);
9083                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
9084
9085                         ret = btrfs_lookup_extent_info(trans, fs_info,
9086                                                        eb->start, level, 1,
9087                                                        &wc->refs[level],
9088                                                        &wc->flags[level]);
9089                         if (ret < 0) {
9090                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
9091                                 path->locks[level] = 0;
9092                                 return ret;
9093                         }
9094                         BUG_ON(wc->refs[level] == 0);
9095                         if (wc->refs[level] == 1) {
9096                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
9097                                 path->locks[level] = 0;
9098                                 return 1;
9099                         }
9100                 }
9101         }
9102
9103         /* wc->stage == DROP_REFERENCE */
9104         BUG_ON(wc->refs[level] > 1 && !path->locks[level]);
9105
9106         if (wc->refs[level] == 1) {
9107                 if (level == 0) {
9108                         if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
9109                                 ret = btrfs_dec_ref(trans, root, eb, 1);
9110                         else
9111                                 ret = btrfs_dec_ref(trans, root, eb, 0);
9112                         BUG_ON(ret); /* -ENOMEM */
9113                         if (is_fstree(root->root_key.objectid)) {
9114                                 ret = btrfs_qgroup_trace_leaf_items(trans, eb);
9115                                 if (ret) {
9116                                         btrfs_err_rl(fs_info,
9117         "error %d accounting leaf items, quota is out of sync, rescan required",
9118                                              ret);
9119                                 }
9120                         }
9121                 }
9122                 /* make block locked assertion in btrfs_clean_tree_block happy */
9123                 if (!path->locks[level] &&
9124                     btrfs_header_generation(eb) == trans->transid) {
9125                         btrfs_tree_lock(eb);
9126                         btrfs_set_lock_blocking_write(eb);
9127                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
9128                 }
9129                 btrfs_clean_tree_block(eb);
9130         }
9131
9132         if (eb == root->node) {
9133                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
9134                         parent = eb->start;
9135                 else if (root->root_key.objectid != btrfs_header_owner(eb))
9136                         goto owner_mismatch;
9137         } else {
9138                 if (wc->flags[level + 1] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
9139                         parent = path->nodes[level + 1]->start;
9140                 else if (root->root_key.objectid !=
9141                          btrfs_header_owner(path->nodes[level + 1]))
9142                         goto owner_mismatch;
9143         }
9144
9145         btrfs_free_tree_block(trans, root, eb, parent, wc->refs[level] == 1);
9146 out:
9147         wc->refs[level] = 0;
9148         wc->flags[level] = 0;
9149         return 0;
9150
9151 owner_mismatch:
9152         btrfs_err_rl(fs_info, "unexpected tree owner, have %llu expect %llu",
9153                      btrfs_header_owner(eb), root->root_key.objectid);
9154         return -EUCLEAN;
9155 }
9156
9157 static noinline int walk_down_tree(struct btrfs_trans_handle *trans,
9158                                    struct btrfs_root *root,
9159                                    struct btrfs_path *path,
9160                                    struct walk_control *wc)
9161 {
9162         int level = wc->level;
9163         int lookup_info = 1;
9164         int ret;
9165
9166         while (level >= 0) {
9167                 ret = walk_down_proc(trans, root, path, wc, lookup_info);
9168                 if (ret > 0)
9169                         break;
9170
9171                 if (level == 0)
9172                         break;
9173
9174                 if (path->slots[level] >=
9175                     btrfs_header_nritems(path->nodes[level]))
9176                         break;
9177
9178                 ret = do_walk_down(trans, root, path, wc, &lookup_info);
9179                 if (ret > 0) {
9180                         path->slots[level]++;
9181                         continue;
9182                 } else if (ret < 0)
9183                         return ret;
9184                 level = wc->level;
9185         }
9186         return 0;
9187 }
9188
9189 static noinline int walk_up_tree(struct btrfs_trans_handle *trans,
9190                                  struct btrfs_root *root,
9191                                  struct btrfs_path *path,
9192                                  struct walk_control *wc, int max_level)
9193 {
9194         int level = wc->level;
9195         int ret;
9196
9197         path->slots[level] = btrfs_header_nritems(path->nodes[level]);
9198         while (level < max_level && path->nodes[level]) {
9199                 wc->level = level;
9200                 if (path->slots[level] + 1 <
9201                     btrfs_header_nritems(path->nodes[level])) {
9202                         path->slots[level]++;
9203                         return 0;
9204                 } else {
9205                         ret = walk_up_proc(trans, root, path, wc);
9206                         if (ret > 0)
9207                                 return 0;
9208                         if (ret < 0)
9209                                 return ret;
9210
9211                         if (path->locks[level]) {
9212                                 btrfs_tree_unlock_rw(path->nodes[level],
9213                                                      path->locks[level]);
9214                                 path->locks[level] = 0;
9215                         }
9216                         free_extent_buffer(path->nodes[level]);
9217                         path->nodes[level] = NULL;
9218                         level++;
9219                 }
9220         }
9221         return 1;
9222 }
9223
9224 /*
9225  * drop a subvolume tree.
9226  *
9227  * this function traverses the tree freeing any blocks that only
9228  * referenced by the tree.
9229  *
9230  * when a shared tree block is found. this function decreases its
9231  * reference count by one. if update_ref is true, this function
9232  * also make sure backrefs for the shared block and all lower level
9233  * blocks are properly updated.
9234  *
9235  * If called with for_reloc == 0, may exit early with -EAGAIN
9236  */
9237 int btrfs_drop_snapshot(struct btrfs_root *root,
9238                          struct btrfs_block_rsv *block_rsv, int update_ref,
9239                          int for_reloc)
9240 {
9241         struct btrfs_fs_info *fs_info = root->fs_info;
9242         struct btrfs_path *path;
9243         struct btrfs_trans_handle *trans;
9244         struct btrfs_root *tree_root = fs_info->tree_root;
9245         struct btrfs_root_item *root_item = &root->root_item;
9246         struct walk_control *wc;
9247         struct btrfs_key key;
9248         int err = 0;
9249         int ret;
9250         int level;
9251         bool root_dropped = false;
9252
9253         btrfs_debug(fs_info, "Drop subvolume %llu", root->root_key.objectid);
9254
9255         path = btrfs_alloc_path();
9256         if (!path) {
9257                 err = -ENOMEM;
9258                 goto out;
9259         }
9260
9261         wc = kzalloc(sizeof(*wc), GFP_NOFS);
9262         if (!wc) {
9263                 btrfs_free_path(path);
9264                 err = -ENOMEM;
9265                 goto out;
9266         }
9267
9268         trans = btrfs_start_transaction(tree_root, 0);
9269         if (IS_ERR(trans)) {
9270                 err = PTR_ERR(trans);
9271                 goto out_free;
9272         }
9273
9274         err = btrfs_run_delayed_items(trans);
9275         if (err)
9276                 goto out_end_trans;
9277
9278         if (block_rsv)
9279                 trans->block_rsv = block_rsv;
9280
9281         /*
9282          * This will help us catch people modifying the fs tree while we're
9283          * dropping it.  It is unsafe to mess with the fs tree while it's being
9284          * dropped as we unlock the root node and parent nodes as we walk down
9285          * the tree, assuming nothing will change.  If something does change
9286          * then we'll have stale information and drop references to blocks we've
9287          * already dropped.
9288          */
9289         set_bit(BTRFS_ROOT_DELETING, &root->state);
9290         if (btrfs_disk_key_objectid(&root_item->drop_progress) == 0) {
9291                 level = btrfs_header_level(root->node);
9292                 path->nodes[level] = btrfs_lock_root_node(root);
9293                 btrfs_set_lock_blocking_write(path->nodes[level]);
9294                 path->slots[level] = 0;
9295                 path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
9296                 memset(&wc->update_progress, 0,
9297                        sizeof(wc->update_progress));
9298         } else {
9299                 btrfs_disk_key_to_cpu(&key, &root_item->drop_progress);
9300                 memcpy(&wc->update_progress, &key,
9301                        sizeof(wc->update_progress));
9302
9303                 level = root_item->drop_level;
9304                 BUG_ON(level == 0);
9305                 path->lowest_level = level;
9306                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
9307                 path->lowest_level = 0;
9308                 if (ret < 0) {
9309                         err = ret;
9310                         goto out_end_trans;
9311                 }
9312                 WARN_ON(ret > 0);
9313
9314                 /*
9315                  * unlock our path, this is safe because only this
9316                  * function is allowed to delete this snapshot
9317                  */
9318                 btrfs_unlock_up_safe(path, 0);
9319
9320                 level = btrfs_header_level(root->node);
9321                 while (1) {
9322                         btrfs_tree_lock(path->nodes[level]);
9323                         btrfs_set_lock_blocking_write(path->nodes[level]);
9324                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
9325
9326                         ret = btrfs_lookup_extent_info(trans, fs_info,
9327                                                 path->nodes[level]->start,
9328                                                 level, 1, &wc->refs[level],
9329                                                 &wc->flags[level]);
9330                         if (ret < 0) {
9331                                 err = ret;
9332                                 goto out_end_trans;
9333                         }
9334                         BUG_ON(wc->refs[level] == 0);
9335
9336                         if (level == root_item->drop_level)
9337                                 break;
9338
9339                         btrfs_tree_unlock(path->nodes[level]);
9340                         path->locks[level] = 0;
9341                         WARN_ON(wc->refs[level] != 1);
9342                         level--;
9343                 }
9344         }
9345
9346         wc->restarted = test_bit(BTRFS_ROOT_DEAD_TREE, &root->state);
9347         wc->level = level;
9348         wc->shared_level = -1;
9349         wc->stage = DROP_REFERENCE;
9350         wc->update_ref = update_ref;
9351         wc->keep_locks = 0;
9352         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(fs_info);
9353
9354         while (1) {
9355
9356                 ret = walk_down_tree(trans, root, path, wc);
9357                 if (ret < 0) {
9358                         err = ret;
9359                         break;
9360                 }
9361
9362                 ret = walk_up_tree(trans, root, path, wc, BTRFS_MAX_LEVEL);
9363                 if (ret < 0) {
9364                         err = ret;
9365                         break;
9366                 }
9367
9368                 if (ret > 0) {
9369                         BUG_ON(wc->stage != DROP_REFERENCE);
9370                         break;
9371                 }
9372
9373                 if (wc->stage == DROP_REFERENCE) {
9374                         wc->drop_level = wc->level;
9375                         btrfs_node_key_to_cpu(path->nodes[wc->drop_level],
9376                                               &wc->drop_progress,
9377                                               path->slots[wc->drop_level]);
9378                 }
9379                 btrfs_cpu_key_to_disk(&root_item->drop_progress,
9380                                       &wc->drop_progress);
9381                 root_item->drop_level = wc->drop_level;
9382
9383                 BUG_ON(wc->level == 0);
9384                 if (btrfs_should_end_transaction(trans) ||
9385                     (!for_reloc && btrfs_need_cleaner_sleep(fs_info))) {
9386                         ret = btrfs_update_root(trans, tree_root,
9387                                                 &root->root_key,
9388                                                 root_item);
9389                         if (ret) {
9390                                 btrfs_abort_transaction(trans, ret);
9391                                 err = ret;
9392                                 goto out_end_trans;
9393                         }
9394
9395                         btrfs_end_transaction_throttle(trans);
9396                         if (!for_reloc && btrfs_need_cleaner_sleep(fs_info)) {
9397                                 btrfs_debug(fs_info,
9398                                             "drop snapshot early exit");
9399                                 err = -EAGAIN;
9400                                 goto out_free;
9401                         }
9402
9403                         trans = btrfs_start_transaction(tree_root, 0);
9404                         if (IS_ERR(trans)) {
9405                                 err = PTR_ERR(trans);
9406                                 goto out_free;
9407                         }
9408                         if (block_rsv)
9409                                 trans->block_rsv = block_rsv;
9410                 }
9411         }
9412         btrfs_release_path(path);
9413         if (err)
9414                 goto out_end_trans;
9415
9416         ret = btrfs_del_root(trans, &root->root_key);
9417         if (ret) {
9418                 btrfs_abort_transaction(trans, ret);
9419                 err = ret;
9420                 goto out_end_trans;
9421         }
9422
9423         if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) {
9424                 ret = btrfs_find_root(tree_root, &root->root_key, path,
9425                                       NULL, NULL);
9426                 if (ret < 0) {
9427                         btrfs_abort_transaction(trans, ret);
9428                         err = ret;
9429                         goto out_end_trans;
9430                 } else if (ret > 0) {
9431                         /* if we fail to delete the orphan item this time
9432                          * around, it'll get picked up the next time.
9433                          *
9434                          * The most common failure here is just -ENOENT.
9435                          */
9436                         btrfs_del_orphan_item(trans, tree_root,
9437                                               root->root_key.objectid);
9438                 }
9439         }
9440
9441         if (test_bit(BTRFS_ROOT_IN_RADIX, &root->state)) {
9442                 btrfs_add_dropped_root(trans, root);
9443         } else {
9444                 free_extent_buffer(root->node);
9445                 free_extent_buffer(root->commit_root);
9446                 btrfs_put_fs_root(root);
9447         }
9448         root_dropped = true;
9449 out_end_trans:
9450         btrfs_end_transaction_throttle(trans);
9451 out_free:
9452         kfree(wc);
9453         btrfs_free_path(path);
9454 out:
9455         /*
9456          * So if we need to stop dropping the snapshot for whatever reason we
9457          * need to make sure to add it back to the dead root list so that we
9458          * keep trying to do the work later.  This also cleans up roots if we
9459          * don't have it in the radix (like when we recover after a power fail
9460          * or unmount) so we don't leak memory.
9461          */
9462         if (!for_reloc && !root_dropped)
9463                 btrfs_add_dead_root(root);
9464         if (err && err != -EAGAIN)
9465                 btrfs_handle_fs_error(fs_info, err, NULL);
9466         return err;
9467 }
9468
9469 /*
9470  * drop subtree rooted at tree block 'node'.
9471  *
9472  * NOTE: this function will unlock and release tree block 'node'
9473  * only used by relocation code
9474  */
9475 int btrfs_drop_subtree(struct btrfs_trans_handle *trans,
9476                         struct btrfs_root *root,
9477                         struct extent_buffer *node,
9478                         struct extent_buffer *parent)
9479 {
9480         struct btrfs_fs_info *fs_info = root->fs_info;
9481         struct btrfs_path *path;
9482         struct walk_control *wc;
9483         int level;
9484         int parent_level;
9485         int ret = 0;
9486         int wret;
9487
9488         BUG_ON(root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID);
9489
9490         path = btrfs_alloc_path();
9491         if (!path)
9492                 return -ENOMEM;
9493
9494         wc = kzalloc(sizeof(*wc), GFP_NOFS);
9495         if (!wc) {
9496                 btrfs_free_path(path);
9497                 return -ENOMEM;
9498         }
9499
9500         btrfs_assert_tree_locked(parent);
9501         parent_level = btrfs_header_level(parent);
9502         extent_buffer_get(parent);
9503         path->nodes[parent_level] = parent;
9504         path->slots[parent_level] = btrfs_header_nritems(parent);
9505
9506         btrfs_assert_tree_locked(node);
9507         level = btrfs_header_level(node);
9508         path->nodes[level] = node;
9509         path->slots[level] = 0;
9510         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
9511
9512         wc->refs[parent_level] = 1;
9513         wc->flags[parent_level] = BTRFS_BLOCK_FLAG_FULL_BACKREF;
9514         wc->level = level;
9515         wc->shared_level = -1;
9516         wc->stage = DROP_REFERENCE;
9517         wc->update_ref = 0;
9518         wc->keep_locks = 1;
9519         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(fs_info);
9520
9521         while (1) {
9522                 wret = walk_down_tree(trans, root, path, wc);
9523                 if (wret < 0) {
9524                         ret = wret;
9525                         break;
9526                 }
9527
9528                 wret = walk_up_tree(trans, root, path, wc, parent_level);
9529                 if (wret < 0)
9530                         ret = wret;
9531                 if (wret != 0)
9532                         break;
9533         }
9534
9535         kfree(wc);
9536         btrfs_free_path(path);
9537         return ret;
9538 }
9539
9540 static u64 update_block_group_flags(struct btrfs_fs_info *fs_info, u64 flags)
9541 {
9542         u64 num_devices;
9543         u64 stripped;
9544
9545         /*
9546          * if restripe for this chunk_type is on pick target profile and
9547          * return, otherwise do the usual balance
9548          */
9549         stripped = get_restripe_target(fs_info, flags);
9550         if (stripped)
9551                 return extended_to_chunk(stripped);
9552
9553         num_devices = fs_info->fs_devices->rw_devices;
9554
9555         stripped = BTRFS_BLOCK_GROUP_RAID0 |
9556                 BTRFS_BLOCK_GROUP_RAID5 | BTRFS_BLOCK_GROUP_RAID6 |
9557                 BTRFS_BLOCK_GROUP_RAID1 | BTRFS_BLOCK_GROUP_RAID10;
9558
9559         if (num_devices == 1) {
9560                 stripped |= BTRFS_BLOCK_GROUP_DUP;
9561                 stripped = flags & ~stripped;
9562
9563                 /* turn raid0 into single device chunks */
9564                 if (flags & BTRFS_BLOCK_GROUP_RAID0)
9565                         return stripped;
9566
9567                 /* turn mirroring into duplication */
9568                 if (flags & (BTRFS_BLOCK_GROUP_RAID1 |
9569                              BTRFS_BLOCK_GROUP_RAID10))
9570                         return stripped | BTRFS_BLOCK_GROUP_DUP;
9571         } else {
9572                 /* they already had raid on here, just return */
9573                 if (flags & stripped)
9574                         return flags;
9575
9576                 stripped |= BTRFS_BLOCK_GROUP_DUP;
9577                 stripped = flags & ~stripped;
9578
9579                 /* switch duplicated blocks with raid1 */
9580                 if (flags & BTRFS_BLOCK_GROUP_DUP)
9581                         return stripped | BTRFS_BLOCK_GROUP_RAID1;
9582
9583                 /* this is drive concat, leave it alone */
9584         }
9585
9586         return flags;
9587 }
9588
9589 static int inc_block_group_ro(struct btrfs_block_group_cache *cache, int force)
9590 {
9591         struct btrfs_space_info *sinfo = cache->space_info;
9592         u64 num_bytes;
9593         u64 sinfo_used;
9594         u64 min_allocable_bytes;
9595         int ret = -ENOSPC;
9596
9597         /*
9598          * We need some metadata space and system metadata space for
9599          * allocating chunks in some corner cases until we force to set
9600          * it to be readonly.
9601          */
9602         if ((sinfo->flags &
9603              (BTRFS_BLOCK_GROUP_SYSTEM | BTRFS_BLOCK_GROUP_METADATA)) &&
9604             !force)
9605                 min_allocable_bytes = SZ_1M;
9606         else
9607                 min_allocable_bytes = 0;
9608
9609         spin_lock(&sinfo->lock);
9610         spin_lock(&cache->lock);
9611
9612         if (cache->ro) {
9613                 cache->ro++;
9614                 ret = 0;
9615                 goto out;
9616         }
9617
9618         num_bytes = cache->key.offset - cache->reserved - cache->pinned -
9619                     cache->bytes_super - btrfs_block_group_used(&cache->item);
9620         sinfo_used = btrfs_space_info_used(sinfo, true);
9621
9622         if (sinfo_used + num_bytes + min_allocable_bytes <=
9623             sinfo->total_bytes) {
9624                 sinfo->bytes_readonly += num_bytes;
9625                 cache->ro++;
9626                 list_add_tail(&cache->ro_list, &sinfo->ro_bgs);
9627                 ret = 0;
9628         }
9629 out:
9630         spin_unlock(&cache->lock);
9631         spin_unlock(&sinfo->lock);
9632         if (ret == -ENOSPC && btrfs_test_opt(cache->fs_info, ENOSPC_DEBUG)) {
9633                 btrfs_info(cache->fs_info,
9634                         "unable to make block group %llu ro",
9635                         cache->key.objectid);
9636                 btrfs_info(cache->fs_info,
9637                         "sinfo_used=%llu bg_num_bytes=%llu min_allocable=%llu",
9638                         sinfo_used, num_bytes, min_allocable_bytes);
9639                 dump_space_info(cache->fs_info, cache->space_info, 0, 0);
9640         }
9641         return ret;
9642 }
9643
9644 int btrfs_inc_block_group_ro(struct btrfs_block_group_cache *cache)
9645
9646 {
9647         struct btrfs_fs_info *fs_info = cache->fs_info;
9648         struct btrfs_trans_handle *trans;
9649         u64 alloc_flags;
9650         int ret;
9651
9652 again:
9653         trans = btrfs_join_transaction(fs_info->extent_root);
9654         if (IS_ERR(trans))
9655                 return PTR_ERR(trans);
9656
9657         /*
9658          * we're not allowed to set block groups readonly after the dirty
9659          * block groups cache has started writing.  If it already started,
9660          * back off and let this transaction commit
9661          */
9662         mutex_lock(&fs_info->ro_block_group_mutex);
9663         if (test_bit(BTRFS_TRANS_DIRTY_BG_RUN, &trans->transaction->flags)) {
9664                 u64 transid = trans->transid;
9665
9666                 mutex_unlock(&fs_info->ro_block_group_mutex);
9667                 btrfs_end_transaction(trans);
9668
9669                 ret = btrfs_wait_for_commit(fs_info, transid);
9670                 if (ret)
9671                         return ret;
9672                 goto again;
9673         }
9674
9675         /*
9676          * if we are changing raid levels, try to allocate a corresponding
9677          * block group with the new raid level.
9678          */
9679         alloc_flags = update_block_group_flags(fs_info, cache->flags);
9680         if (alloc_flags != cache->flags) {
9681                 ret = do_chunk_alloc(trans, alloc_flags,
9682                                      CHUNK_ALLOC_FORCE);
9683                 /*
9684                  * ENOSPC is allowed here, we may have enough space
9685                  * already allocated at the new raid level to
9686                  * carry on
9687                  */
9688                 if (ret == -ENOSPC)
9689                         ret = 0;
9690                 if (ret < 0)
9691                         goto out;
9692         }
9693
9694         ret = inc_block_group_ro(cache, 0);
9695         if (!ret)
9696                 goto out;
9697         alloc_flags = get_alloc_profile(fs_info, cache->space_info->flags);
9698         ret = do_chunk_alloc(trans, alloc_flags, CHUNK_ALLOC_FORCE);
9699         if (ret < 0)
9700                 goto out;
9701         ret = inc_block_group_ro(cache, 0);
9702 out:
9703         if (cache->flags & BTRFS_BLOCK_GROUP_SYSTEM) {
9704                 alloc_flags = update_block_group_flags(fs_info, cache->flags);
9705                 mutex_lock(&fs_info->chunk_mutex);
9706                 check_system_chunk(trans, alloc_flags);
9707                 mutex_unlock(&fs_info->chunk_mutex);
9708         }
9709         mutex_unlock(&fs_info->ro_block_group_mutex);
9710
9711         btrfs_end_transaction(trans);
9712         return ret;
9713 }
9714
9715 int btrfs_force_chunk_alloc(struct btrfs_trans_handle *trans, u64 type)
9716 {
9717         u64 alloc_flags = get_alloc_profile(trans->fs_info, type);
9718
9719         return do_chunk_alloc(trans, alloc_flags, CHUNK_ALLOC_FORCE);
9720 }
9721
9722 /*
9723  * helper to account the unused space of all the readonly block group in the
9724  * space_info. takes mirrors into account.
9725  */
9726 u64 btrfs_account_ro_block_groups_free_space(struct btrfs_space_info *sinfo)
9727 {
9728         struct btrfs_block_group_cache *block_group;
9729         u64 free_bytes = 0;
9730         int factor;
9731
9732         /* It's df, we don't care if it's racy */
9733         if (list_empty(&sinfo->ro_bgs))
9734                 return 0;
9735
9736         spin_lock(&sinfo->lock);
9737         list_for_each_entry(block_group, &sinfo->ro_bgs, ro_list) {
9738                 spin_lock(&block_group->lock);
9739
9740                 if (!block_group->ro) {
9741                         spin_unlock(&block_group->lock);
9742                         continue;
9743                 }
9744
9745                 factor = btrfs_bg_type_to_factor(block_group->flags);
9746                 free_bytes += (block_group->key.offset -
9747                                btrfs_block_group_used(&block_group->item)) *
9748                                factor;
9749
9750                 spin_unlock(&block_group->lock);
9751         }
9752         spin_unlock(&sinfo->lock);
9753
9754         return free_bytes;
9755 }
9756
9757 void btrfs_dec_block_group_ro(struct btrfs_block_group_cache *cache)
9758 {
9759         struct btrfs_space_info *sinfo = cache->space_info;
9760         u64 num_bytes;
9761
9762         BUG_ON(!cache->ro);
9763
9764         spin_lock(&sinfo->lock);
9765         spin_lock(&cache->lock);
9766         if (!--cache->ro) {
9767                 num_bytes = cache->key.offset - cache->reserved -
9768                             cache->pinned - cache->bytes_super -
9769                             btrfs_block_group_used(&cache->item);
9770                 sinfo->bytes_readonly -= num_bytes;
9771                 list_del_init(&cache->ro_list);
9772         }
9773         spin_unlock(&cache->lock);
9774         spin_unlock(&sinfo->lock);
9775 }
9776
9777 /*
9778  * Checks to see if it's even possible to relocate this block group.
9779  *
9780  * @return - -1 if it's not a good idea to relocate this block group, 0 if its
9781  * ok to go ahead and try.
9782  */
9783 int btrfs_can_relocate(struct btrfs_fs_info *fs_info, u64 bytenr)
9784 {
9785         struct btrfs_block_group_cache *block_group;
9786         struct btrfs_space_info *space_info;
9787         struct btrfs_fs_devices *fs_devices = fs_info->fs_devices;
9788         struct btrfs_device *device;
9789         u64 min_free;
9790         u64 dev_min = 1;
9791         u64 dev_nr = 0;
9792         u64 target;
9793         int debug;
9794         int index;
9795         int full = 0;
9796         int ret = 0;
9797
9798         debug = btrfs_test_opt(fs_info, ENOSPC_DEBUG);
9799
9800         block_group = btrfs_lookup_block_group(fs_info, bytenr);
9801
9802         /* odd, couldn't find the block group, leave it alone */
9803         if (!block_group) {
9804                 if (debug)
9805                         btrfs_warn(fs_info,
9806                                    "can't find block group for bytenr %llu",
9807                                    bytenr);
9808                 return -1;
9809         }
9810
9811         min_free = btrfs_block_group_used(&block_group->item);
9812
9813         /* no bytes used, we're good */
9814         if (!min_free)
9815                 goto out;
9816
9817         space_info = block_group->space_info;
9818         spin_lock(&space_info->lock);
9819
9820         full = space_info->full;
9821
9822         /*
9823          * if this is the last block group we have in this space, we can't
9824          * relocate it unless we're able to allocate a new chunk below.
9825          *
9826          * Otherwise, we need to make sure we have room in the space to handle
9827          * all of the extents from this block group.  If we can, we're good
9828          */
9829         if ((space_info->total_bytes != block_group->key.offset) &&
9830             (btrfs_space_info_used(space_info, false) + min_free <
9831              space_info->total_bytes)) {
9832                 spin_unlock(&space_info->lock);
9833                 goto out;
9834         }
9835         spin_unlock(&space_info->lock);
9836
9837         /*
9838          * ok we don't have enough space, but maybe we have free space on our
9839          * devices to allocate new chunks for relocation, so loop through our
9840          * alloc devices and guess if we have enough space.  if this block
9841          * group is going to be restriped, run checks against the target
9842          * profile instead of the current one.
9843          */
9844         ret = -1;
9845
9846         /*
9847          * index:
9848          *      0: raid10
9849          *      1: raid1
9850          *      2: dup
9851          *      3: raid0
9852          *      4: single
9853          */
9854         target = get_restripe_target(fs_info, block_group->flags);
9855         if (target) {
9856                 index = btrfs_bg_flags_to_raid_index(extended_to_chunk(target));
9857         } else {
9858                 /*
9859                  * this is just a balance, so if we were marked as full
9860                  * we know there is no space for a new chunk
9861                  */
9862                 if (full) {
9863                         if (debug)
9864                                 btrfs_warn(fs_info,
9865                                            "no space to alloc new chunk for block group %llu",
9866                                            block_group->key.objectid);
9867                         goto out;
9868                 }
9869
9870                 index = btrfs_bg_flags_to_raid_index(block_group->flags);
9871         }
9872
9873         if (index == BTRFS_RAID_RAID10) {
9874                 dev_min = 4;
9875                 /* Divide by 2 */
9876                 min_free >>= 1;
9877         } else if (index == BTRFS_RAID_RAID1) {
9878                 dev_min = 2;
9879         } else if (index == BTRFS_RAID_DUP) {
9880                 /* Multiply by 2 */
9881                 min_free <<= 1;
9882         } else if (index == BTRFS_RAID_RAID0) {
9883                 dev_min = fs_devices->rw_devices;
9884                 min_free = div64_u64(min_free, dev_min);
9885         }
9886
9887         mutex_lock(&fs_info->chunk_mutex);
9888         list_for_each_entry(device, &fs_devices->alloc_list, dev_alloc_list) {
9889                 u64 dev_offset;
9890
9891                 /*
9892                  * check to make sure we can actually find a chunk with enough
9893                  * space to fit our block group in.
9894                  */
9895                 if (device->total_bytes > device->bytes_used + min_free &&
9896                     !test_bit(BTRFS_DEV_STATE_REPLACE_TGT, &device->dev_state)) {
9897                         ret = find_free_dev_extent(device, min_free,
9898                                                    &dev_offset, NULL);
9899                         if (!ret)
9900                                 dev_nr++;
9901
9902                         if (dev_nr >= dev_min)
9903                                 break;
9904
9905                         ret = -1;
9906                 }
9907         }
9908         if (debug && ret == -1)
9909                 btrfs_warn(fs_info,
9910                            "no space to allocate a new chunk for block group %llu",
9911                            block_group->key.objectid);
9912         mutex_unlock(&fs_info->chunk_mutex);
9913 out:
9914         btrfs_put_block_group(block_group);
9915         return ret;
9916 }
9917
9918 static int find_first_block_group(struct btrfs_fs_info *fs_info,
9919                                   struct btrfs_path *path,
9920                                   struct btrfs_key *key)
9921 {
9922         struct btrfs_root *root = fs_info->extent_root;
9923         int ret = 0;
9924         struct btrfs_key found_key;
9925         struct extent_buffer *leaf;
9926         struct btrfs_block_group_item bg;
9927         u64 flags;
9928         int slot;
9929
9930         ret = btrfs_search_slot(NULL, root, key, path, 0, 0);
9931         if (ret < 0)
9932                 goto out;
9933
9934         while (1) {
9935                 slot = path->slots[0];
9936                 leaf = path->nodes[0];
9937                 if (slot >= btrfs_header_nritems(leaf)) {
9938                         ret = btrfs_next_leaf(root, path);
9939                         if (ret == 0)
9940                                 continue;
9941                         if (ret < 0)
9942                                 goto out;
9943                         break;
9944                 }
9945                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
9946
9947                 if (found_key.objectid >= key->objectid &&
9948                     found_key.type == BTRFS_BLOCK_GROUP_ITEM_KEY) {
9949                         struct extent_map_tree *em_tree;
9950                         struct extent_map *em;
9951
9952                         em_tree = &root->fs_info->mapping_tree.map_tree;
9953                         read_lock(&em_tree->lock);
9954                         em = lookup_extent_mapping(em_tree, found_key.objectid,
9955                                                    found_key.offset);
9956                         read_unlock(&em_tree->lock);
9957                         if (!em) {
9958                                 btrfs_err(fs_info,
9959                         "logical %llu len %llu found bg but no related chunk",
9960                                           found_key.objectid, found_key.offset);
9961                                 ret = -ENOENT;
9962                         } else if (em->start != found_key.objectid ||
9963                                    em->len != found_key.offset) {
9964                                 btrfs_err(fs_info,
9965                 "block group %llu len %llu mismatch with chunk %llu len %llu",
9966                                           found_key.objectid, found_key.offset,
9967                                           em->start, em->len);
9968                                 ret = -EUCLEAN;
9969                         } else {
9970                                 read_extent_buffer(leaf, &bg,
9971                                         btrfs_item_ptr_offset(leaf, slot),
9972                                         sizeof(bg));
9973                                 flags = btrfs_block_group_flags(&bg) &
9974                                         BTRFS_BLOCK_GROUP_TYPE_MASK;
9975
9976                                 if (flags != (em->map_lookup->type &
9977                                               BTRFS_BLOCK_GROUP_TYPE_MASK)) {
9978                                         btrfs_err(fs_info,
9979 "block group %llu len %llu type flags 0x%llx mismatch with chunk type flags 0x%llx",
9980                                                 found_key.objectid,
9981                                                 found_key.offset, flags,
9982                                                 (BTRFS_BLOCK_GROUP_TYPE_MASK &
9983                                                  em->map_lookup->type));
9984                                         ret = -EUCLEAN;
9985                                 } else {
9986                                         ret = 0;
9987                                 }
9988                         }
9989                         free_extent_map(em);
9990                         goto out;
9991                 }
9992                 path->slots[0]++;
9993         }
9994 out:
9995         return ret;
9996 }
9997
9998 void btrfs_put_block_group_cache(struct btrfs_fs_info *info)
9999 {
10000         struct btrfs_block_group_cache *block_group;
10001         u64 last = 0;
10002
10003         while (1) {
10004                 struct inode *inode;
10005
10006                 block_group = btrfs_lookup_first_block_group(info, last);
10007                 while (block_group) {
10008                         wait_block_group_cache_done(block_group);
10009                         spin_lock(&block_group->lock);
10010                         if (block_group->iref)
10011                                 break;
10012                         spin_unlock(&block_group->lock);
10013                         block_group = next_block_group(block_group);
10014                 }
10015                 if (!block_group) {
10016                         if (last == 0)
10017                                 break;
10018                         last = 0;
10019                         continue;
10020                 }
10021
10022                 inode = block_group->inode;
10023                 block_group->iref = 0;
10024                 block_group->inode = NULL;
10025                 spin_unlock(&block_group->lock);
10026                 ASSERT(block_group->io_ctl.inode == NULL);
10027                 iput(inode);
10028                 last = block_group->key.objectid + block_group->key.offset;
10029                 btrfs_put_block_group(block_group);
10030         }
10031 }
10032
10033 /*
10034  * Must be called only after stopping all workers, since we could have block
10035  * group caching kthreads running, and therefore they could race with us if we
10036  * freed the block groups before stopping them.
10037  */
10038 int btrfs_free_block_groups(struct btrfs_fs_info *info)
10039 {
10040         struct btrfs_block_group_cache *block_group;
10041         struct btrfs_space_info *space_info;
10042         struct btrfs_caching_control *caching_ctl;
10043         struct rb_node *n;
10044
10045         down_write(&info->commit_root_sem);
10046         while (!list_empty(&info->caching_block_groups)) {
10047                 caching_ctl = list_entry(info->caching_block_groups.next,
10048                                          struct btrfs_caching_control, list);
10049                 list_del(&caching_ctl->list);
10050                 put_caching_control(caching_ctl);
10051         }
10052         up_write(&info->commit_root_sem);
10053
10054         spin_lock(&info->unused_bgs_lock);
10055         while (!list_empty(&info->unused_bgs)) {
10056                 block_group = list_first_entry(&info->unused_bgs,
10057                                                struct btrfs_block_group_cache,
10058                                                bg_list);
10059                 list_del_init(&block_group->bg_list);
10060                 btrfs_put_block_group(block_group);
10061         }
10062         spin_unlock(&info->unused_bgs_lock);
10063
10064         spin_lock(&info->block_group_cache_lock);
10065         while ((n = rb_last(&info->block_group_cache_tree)) != NULL) {
10066                 block_group = rb_entry(n, struct btrfs_block_group_cache,
10067                                        cache_node);
10068                 rb_erase(&block_group->cache_node,
10069                          &info->block_group_cache_tree);
10070                 RB_CLEAR_NODE(&block_group->cache_node);
10071                 spin_unlock(&info->block_group_cache_lock);
10072
10073                 down_write(&block_group->space_info->groups_sem);
10074                 list_del(&block_group->list);
10075                 up_write(&block_group->space_info->groups_sem);
10076
10077                 /*
10078                  * We haven't cached this block group, which means we could
10079                  * possibly have excluded extents on this block group.
10080                  */
10081                 if (block_group->cached == BTRFS_CACHE_NO ||
10082                     block_group->cached == BTRFS_CACHE_ERROR)
10083                         free_excluded_extents(block_group);
10084
10085                 btrfs_remove_free_space_cache(block_group);
10086                 ASSERT(block_group->cached != BTRFS_CACHE_STARTED);
10087                 ASSERT(list_empty(&block_group->dirty_list));
10088                 ASSERT(list_empty(&block_group->io_list));
10089                 ASSERT(list_empty(&block_group->bg_list));
10090                 ASSERT(atomic_read(&block_group->count) == 1);
10091                 btrfs_put_block_group(block_group);
10092
10093                 spin_lock(&info->block_group_cache_lock);
10094         }
10095         spin_unlock(&info->block_group_cache_lock);
10096
10097         /* now that all the block groups are freed, go through and
10098          * free all the space_info structs.  This is only called during
10099          * the final stages of unmount, and so we know nobody is
10100          * using them.  We call synchronize_rcu() once before we start,
10101          * just to be on the safe side.
10102          */
10103         synchronize_rcu();
10104
10105         release_global_block_rsv(info);
10106
10107         while (!list_empty(&info->space_info)) {
10108                 int i;
10109
10110                 space_info = list_entry(info->space_info.next,
10111                                         struct btrfs_space_info,
10112                                         list);
10113
10114                 /*
10115                  * Do not hide this behind enospc_debug, this is actually
10116                  * important and indicates a real bug if this happens.
10117                  */
10118                 if (WARN_ON(space_info->bytes_pinned > 0 ||
10119                             space_info->bytes_reserved > 0 ||
10120                             space_info->bytes_may_use > 0))
10121                         dump_space_info(info, space_info, 0, 0);
10122                 list_del(&space_info->list);
10123                 for (i = 0; i < BTRFS_NR_RAID_TYPES; i++) {
10124                         struct kobject *kobj;
10125                         kobj = space_info->block_group_kobjs[i];
10126                         space_info->block_group_kobjs[i] = NULL;
10127                         if (kobj) {
10128                                 kobject_del(kobj);
10129                                 kobject_put(kobj);
10130                         }
10131                 }
10132                 kobject_del(&space_info->kobj);
10133                 kobject_put(&space_info->kobj);
10134         }
10135         return 0;
10136 }
10137
10138 /* link_block_group will queue up kobjects to add when we're reclaim-safe */
10139 void btrfs_add_raid_kobjects(struct btrfs_fs_info *fs_info)
10140 {
10141         struct btrfs_space_info *space_info;
10142         struct raid_kobject *rkobj;
10143         LIST_HEAD(list);
10144         int index;
10145         int ret = 0;
10146
10147         spin_lock(&fs_info->pending_raid_kobjs_lock);
10148         list_splice_init(&fs_info->pending_raid_kobjs, &list);
10149         spin_unlock(&fs_info->pending_raid_kobjs_lock);
10150
10151         list_for_each_entry(rkobj, &list, list) {
10152                 space_info = __find_space_info(fs_info, rkobj->flags);
10153                 index = btrfs_bg_flags_to_raid_index(rkobj->flags);
10154
10155                 ret = kobject_add(&rkobj->kobj, &space_info->kobj,
10156                                   "%s", get_raid_name(index));
10157                 if (ret) {
10158                         kobject_put(&rkobj->kobj);
10159                         break;
10160                 }
10161         }
10162         if (ret)
10163                 btrfs_warn(fs_info,
10164                            "failed to add kobject for block cache, ignoring");
10165 }
10166
10167 static void link_block_group(struct btrfs_block_group_cache *cache)
10168 {
10169         struct btrfs_space_info *space_info = cache->space_info;
10170         struct btrfs_fs_info *fs_info = cache->fs_info;
10171         int index = btrfs_bg_flags_to_raid_index(cache->flags);
10172         bool first = false;
10173
10174         down_write(&space_info->groups_sem);
10175         if (list_empty(&space_info->block_groups[index]))
10176                 first = true;
10177         list_add_tail(&cache->list, &space_info->block_groups[index]);
10178         up_write(&space_info->groups_sem);
10179
10180         if (first) {
10181                 struct raid_kobject *rkobj = kzalloc(sizeof(*rkobj), GFP_NOFS);
10182                 if (!rkobj) {
10183                         btrfs_warn(cache->fs_info,
10184                                 "couldn't alloc memory for raid level kobject");
10185                         return;
10186                 }
10187                 rkobj->flags = cache->flags;
10188                 kobject_init(&rkobj->kobj, &btrfs_raid_ktype);
10189
10190                 spin_lock(&fs_info->pending_raid_kobjs_lock);
10191                 list_add_tail(&rkobj->list, &fs_info->pending_raid_kobjs);
10192                 spin_unlock(&fs_info->pending_raid_kobjs_lock);
10193                 space_info->block_group_kobjs[index] = &rkobj->kobj;
10194         }
10195 }
10196
10197 static struct btrfs_block_group_cache *
10198 btrfs_create_block_group_cache(struct btrfs_fs_info *fs_info,
10199                                u64 start, u64 size)
10200 {
10201         struct btrfs_block_group_cache *cache;
10202
10203         cache = kzalloc(sizeof(*cache), GFP_NOFS);
10204         if (!cache)
10205                 return NULL;
10206
10207         cache->free_space_ctl = kzalloc(sizeof(*cache->free_space_ctl),
10208                                         GFP_NOFS);
10209         if (!cache->free_space_ctl) {
10210                 kfree(cache);
10211                 return NULL;
10212         }
10213
10214         cache->key.objectid = start;
10215         cache->key.offset = size;
10216         cache->key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
10217
10218         cache->fs_info = fs_info;
10219         cache->full_stripe_len = btrfs_full_stripe_len(fs_info, start);
10220         set_free_space_tree_thresholds(cache);
10221
10222         atomic_set(&cache->count, 1);
10223         spin_lock_init(&cache->lock);
10224         init_rwsem(&cache->data_rwsem);
10225         INIT_LIST_HEAD(&cache->list);
10226         INIT_LIST_HEAD(&cache->cluster_list);
10227         INIT_LIST_HEAD(&cache->bg_list);
10228         INIT_LIST_HEAD(&cache->ro_list);
10229         INIT_LIST_HEAD(&cache->dirty_list);
10230         INIT_LIST_HEAD(&cache->io_list);
10231         btrfs_init_free_space_ctl(cache);
10232         atomic_set(&cache->trimming, 0);
10233         mutex_init(&cache->free_space_lock);
10234         btrfs_init_full_stripe_locks_tree(&cache->full_stripe_locks_root);
10235
10236         return cache;
10237 }
10238
10239
10240 /*
10241  * Iterate all chunks and verify that each of them has the corresponding block
10242  * group
10243  */
10244 static int check_chunk_block_group_mappings(struct btrfs_fs_info *fs_info)
10245 {
10246         struct btrfs_mapping_tree *map_tree = &fs_info->mapping_tree;
10247         struct extent_map *em;
10248         struct btrfs_block_group_cache *bg;
10249         u64 start = 0;
10250         int ret = 0;
10251
10252         while (1) {
10253                 read_lock(&map_tree->map_tree.lock);
10254                 /*
10255                  * lookup_extent_mapping will return the first extent map
10256                  * intersecting the range, so setting @len to 1 is enough to
10257                  * get the first chunk.
10258                  */
10259                 em = lookup_extent_mapping(&map_tree->map_tree, start, 1);
10260                 read_unlock(&map_tree->map_tree.lock);
10261                 if (!em)
10262                         break;
10263
10264                 bg = btrfs_lookup_block_group(fs_info, em->start);
10265                 if (!bg) {
10266                         btrfs_err(fs_info,
10267         "chunk start=%llu len=%llu doesn't have corresponding block group",
10268                                      em->start, em->len);
10269                         ret = -EUCLEAN;
10270                         free_extent_map(em);
10271                         break;
10272                 }
10273                 if (bg->key.objectid != em->start ||
10274                     bg->key.offset != em->len ||
10275                     (bg->flags & BTRFS_BLOCK_GROUP_TYPE_MASK) !=
10276                     (em->map_lookup->type & BTRFS_BLOCK_GROUP_TYPE_MASK)) {
10277                         btrfs_err(fs_info,
10278 "chunk start=%llu len=%llu flags=0x%llx doesn't match block group start=%llu len=%llu flags=0x%llx",
10279                                 em->start, em->len,
10280                                 em->map_lookup->type & BTRFS_BLOCK_GROUP_TYPE_MASK,
10281                                 bg->key.objectid, bg->key.offset,
10282                                 bg->flags & BTRFS_BLOCK_GROUP_TYPE_MASK);
10283                         ret = -EUCLEAN;
10284                         free_extent_map(em);
10285                         btrfs_put_block_group(bg);
10286                         break;
10287                 }
10288                 start = em->start + em->len;
10289                 free_extent_map(em);
10290                 btrfs_put_block_group(bg);
10291         }
10292         return ret;
10293 }
10294
10295 int btrfs_read_block_groups(struct btrfs_fs_info *info)
10296 {
10297         struct btrfs_path *path;
10298         int ret;
10299         struct btrfs_block_group_cache *cache;
10300         struct btrfs_space_info *space_info;
10301         struct btrfs_key key;
10302         struct btrfs_key found_key;
10303         struct extent_buffer *leaf;
10304         int need_clear = 0;
10305         u64 cache_gen;
10306         u64 feature;
10307         int mixed;
10308
10309         feature = btrfs_super_incompat_flags(info->super_copy);
10310         mixed = !!(feature & BTRFS_FEATURE_INCOMPAT_MIXED_GROUPS);
10311
10312         key.objectid = 0;
10313         key.offset = 0;
10314         key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
10315         path = btrfs_alloc_path();
10316         if (!path)
10317                 return -ENOMEM;
10318         path->reada = READA_FORWARD;
10319
10320         cache_gen = btrfs_super_cache_generation(info->super_copy);
10321         if (btrfs_test_opt(info, SPACE_CACHE) &&
10322             btrfs_super_generation(info->super_copy) != cache_gen)
10323                 need_clear = 1;
10324         if (btrfs_test_opt(info, CLEAR_CACHE))
10325                 need_clear = 1;
10326
10327         while (1) {
10328                 ret = find_first_block_group(info, path, &key);
10329                 if (ret > 0)
10330                         break;
10331                 if (ret != 0)
10332                         goto error;
10333
10334                 leaf = path->nodes[0];
10335                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
10336
10337                 cache = btrfs_create_block_group_cache(info, found_key.objectid,
10338                                                        found_key.offset);
10339                 if (!cache) {
10340                         ret = -ENOMEM;
10341                         goto error;
10342                 }
10343
10344                 if (need_clear) {
10345                         /*
10346                          * When we mount with old space cache, we need to
10347                          * set BTRFS_DC_CLEAR and set dirty flag.
10348                          *
10349                          * a) Setting 'BTRFS_DC_CLEAR' makes sure that we
10350                          *    truncate the old free space cache inode and
10351                          *    setup a new one.
10352                          * b) Setting 'dirty flag' makes sure that we flush
10353                          *    the new space cache info onto disk.
10354                          */
10355                         if (btrfs_test_opt(info, SPACE_CACHE))
10356                                 cache->disk_cache_state = BTRFS_DC_CLEAR;
10357                 }
10358
10359                 read_extent_buffer(leaf, &cache->item,
10360                                    btrfs_item_ptr_offset(leaf, path->slots[0]),
10361                                    sizeof(cache->item));
10362                 cache->flags = btrfs_block_group_flags(&cache->item);
10363                 if (!mixed &&
10364                     ((cache->flags & BTRFS_BLOCK_GROUP_METADATA) &&
10365                     (cache->flags & BTRFS_BLOCK_GROUP_DATA))) {
10366                         btrfs_err(info,
10367 "bg %llu is a mixed block group but filesystem hasn't enabled mixed block groups",
10368                                   cache->key.objectid);
10369                         ret = -EINVAL;
10370                         goto error;
10371                 }
10372
10373                 key.objectid = found_key.objectid + found_key.offset;
10374                 btrfs_release_path(path);
10375
10376                 /*
10377                  * We need to exclude the super stripes now so that the space
10378                  * info has super bytes accounted for, otherwise we'll think
10379                  * we have more space than we actually do.
10380                  */
10381                 ret = exclude_super_stripes(cache);
10382                 if (ret) {
10383                         /*
10384                          * We may have excluded something, so call this just in
10385                          * case.
10386                          */
10387                         free_excluded_extents(cache);
10388                         btrfs_put_block_group(cache);
10389                         goto error;
10390                 }
10391
10392                 /*
10393                  * check for two cases, either we are full, and therefore
10394                  * don't need to bother with the caching work since we won't
10395                  * find any space, or we are empty, and we can just add all
10396                  * the space in and be done with it.  This saves us _a_lot_ of
10397                  * time, particularly in the full case.
10398                  */
10399                 if (found_key.offset == btrfs_block_group_used(&cache->item)) {
10400                         cache->last_byte_to_unpin = (u64)-1;
10401                         cache->cached = BTRFS_CACHE_FINISHED;
10402                         free_excluded_extents(cache);
10403                 } else if (btrfs_block_group_used(&cache->item) == 0) {
10404                         cache->last_byte_to_unpin = (u64)-1;
10405                         cache->cached = BTRFS_CACHE_FINISHED;
10406                         add_new_free_space(cache, found_key.objectid,
10407                                            found_key.objectid +
10408                                            found_key.offset);
10409                         free_excluded_extents(cache);
10410                 }
10411
10412                 ret = btrfs_add_block_group_cache(info, cache);
10413                 if (ret) {
10414                         btrfs_remove_free_space_cache(cache);
10415                         btrfs_put_block_group(cache);
10416                         goto error;
10417                 }
10418
10419                 trace_btrfs_add_block_group(info, cache, 0);
10420                 update_space_info(info, cache->flags, found_key.offset,
10421                                   btrfs_block_group_used(&cache->item),
10422                                   cache->bytes_super, &space_info);
10423
10424                 cache->space_info = space_info;
10425
10426                 link_block_group(cache);
10427
10428                 set_avail_alloc_bits(info, cache->flags);
10429                 if (btrfs_chunk_readonly(info, cache->key.objectid)) {
10430                         inc_block_group_ro(cache, 1);
10431                 } else if (btrfs_block_group_used(&cache->item) == 0) {
10432                         ASSERT(list_empty(&cache->bg_list));
10433                         btrfs_mark_bg_unused(cache);
10434                 }
10435         }
10436
10437         list_for_each_entry_rcu(space_info, &info->space_info, list) {
10438                 if (!(get_alloc_profile(info, space_info->flags) &
10439                       (BTRFS_BLOCK_GROUP_RAID10 |
10440                        BTRFS_BLOCK_GROUP_RAID1 |
10441                        BTRFS_BLOCK_GROUP_RAID5 |
10442                        BTRFS_BLOCK_GROUP_RAID6 |
10443                        BTRFS_BLOCK_GROUP_DUP)))
10444                         continue;
10445                 /*
10446                  * avoid allocating from un-mirrored block group if there are
10447                  * mirrored block groups.
10448                  */
10449                 list_for_each_entry(cache,
10450                                 &space_info->block_groups[BTRFS_RAID_RAID0],
10451                                 list)
10452                         inc_block_group_ro(cache, 1);
10453                 list_for_each_entry(cache,
10454                                 &space_info->block_groups[BTRFS_RAID_SINGLE],
10455                                 list)
10456                         inc_block_group_ro(cache, 1);
10457         }
10458
10459         btrfs_add_raid_kobjects(info);
10460         init_global_block_rsv(info);
10461         ret = check_chunk_block_group_mappings(info);
10462 error:
10463         btrfs_free_path(path);
10464         return ret;
10465 }
10466
10467 void btrfs_create_pending_block_groups(struct btrfs_trans_handle *trans)
10468 {
10469         struct btrfs_fs_info *fs_info = trans->fs_info;
10470         struct btrfs_block_group_cache *block_group;
10471         struct btrfs_root *extent_root = fs_info->extent_root;
10472         struct btrfs_block_group_item item;
10473         struct btrfs_key key;
10474         int ret = 0;
10475
10476         if (!trans->can_flush_pending_bgs)
10477                 return;
10478
10479         while (!list_empty(&trans->new_bgs)) {
10480                 block_group = list_first_entry(&trans->new_bgs,
10481                                                struct btrfs_block_group_cache,
10482                                                bg_list);
10483                 if (ret)
10484                         goto next;
10485
10486                 spin_lock(&block_group->lock);
10487                 memcpy(&item, &block_group->item, sizeof(item));
10488                 memcpy(&key, &block_group->key, sizeof(key));
10489                 spin_unlock(&block_group->lock);
10490
10491                 ret = btrfs_insert_item(trans, extent_root, &key, &item,
10492                                         sizeof(item));
10493                 if (ret)
10494                         btrfs_abort_transaction(trans, ret);
10495                 ret = btrfs_finish_chunk_alloc(trans, key.objectid, key.offset);
10496                 if (ret)
10497                         btrfs_abort_transaction(trans, ret);
10498                 add_block_group_free_space(trans, block_group);
10499                 /* already aborted the transaction if it failed. */
10500 next:
10501                 btrfs_delayed_refs_rsv_release(fs_info, 1);
10502                 list_del_init(&block_group->bg_list);
10503         }
10504         btrfs_trans_release_chunk_metadata(trans);
10505 }
10506
10507 int btrfs_make_block_group(struct btrfs_trans_handle *trans, u64 bytes_used,
10508                            u64 type, u64 chunk_offset, u64 size)
10509 {
10510         struct btrfs_fs_info *fs_info = trans->fs_info;
10511         struct btrfs_block_group_cache *cache;
10512         int ret;
10513
10514         btrfs_set_log_full_commit(trans);
10515
10516         cache = btrfs_create_block_group_cache(fs_info, chunk_offset, size);
10517         if (!cache)
10518                 return -ENOMEM;
10519
10520         btrfs_set_block_group_used(&cache->item, bytes_used);
10521         btrfs_set_block_group_chunk_objectid(&cache->item,
10522                                              BTRFS_FIRST_CHUNK_TREE_OBJECTID);
10523         btrfs_set_block_group_flags(&cache->item, type);
10524
10525         cache->flags = type;
10526         cache->last_byte_to_unpin = (u64)-1;
10527         cache->cached = BTRFS_CACHE_FINISHED;
10528         cache->needs_free_space = 1;
10529         ret = exclude_super_stripes(cache);
10530         if (ret) {
10531                 /*
10532                  * We may have excluded something, so call this just in
10533                  * case.
10534                  */
10535                 free_excluded_extents(cache);
10536                 btrfs_put_block_group(cache);
10537                 return ret;
10538         }
10539
10540         add_new_free_space(cache, chunk_offset, chunk_offset + size);
10541
10542         free_excluded_extents(cache);
10543
10544 #ifdef CONFIG_BTRFS_DEBUG
10545         if (btrfs_should_fragment_free_space(cache)) {
10546                 u64 new_bytes_used = size - bytes_used;
10547
10548                 bytes_used += new_bytes_used >> 1;
10549                 fragment_free_space(cache);
10550         }
10551 #endif
10552         /*
10553          * Ensure the corresponding space_info object is created and
10554          * assigned to our block group. We want our bg to be added to the rbtree
10555          * with its ->space_info set.
10556          */
10557         cache->space_info = __find_space_info(fs_info, cache->flags);
10558         ASSERT(cache->space_info);
10559
10560         ret = btrfs_add_block_group_cache(fs_info, cache);
10561         if (ret) {
10562                 btrfs_remove_free_space_cache(cache);
10563                 btrfs_put_block_group(cache);
10564                 return ret;
10565         }
10566
10567         /*
10568          * Now that our block group has its ->space_info set and is inserted in
10569          * the rbtree, update the space info's counters.
10570          */
10571         trace_btrfs_add_block_group(fs_info, cache, 1);
10572         update_space_info(fs_info, cache->flags, size, bytes_used,
10573                                 cache->bytes_super, &cache->space_info);
10574         update_global_block_rsv(fs_info);
10575
10576         link_block_group(cache);
10577
10578         list_add_tail(&cache->bg_list, &trans->new_bgs);
10579         trans->delayed_ref_updates++;
10580         btrfs_update_delayed_refs_rsv(trans);
10581
10582         set_avail_alloc_bits(fs_info, type);
10583         return 0;
10584 }
10585
10586 static void clear_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
10587 {
10588         u64 extra_flags = chunk_to_extended(flags) &
10589                                 BTRFS_EXTENDED_PROFILE_MASK;
10590
10591         write_seqlock(&fs_info->profiles_lock);
10592         if (flags & BTRFS_BLOCK_GROUP_DATA)
10593                 fs_info->avail_data_alloc_bits &= ~extra_flags;
10594         if (flags & BTRFS_BLOCK_GROUP_METADATA)
10595                 fs_info->avail_metadata_alloc_bits &= ~extra_flags;
10596         if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
10597                 fs_info->avail_system_alloc_bits &= ~extra_flags;
10598         write_sequnlock(&fs_info->profiles_lock);
10599 }
10600
10601 int btrfs_remove_block_group(struct btrfs_trans_handle *trans,
10602                              u64 group_start, struct extent_map *em)
10603 {
10604         struct btrfs_fs_info *fs_info = trans->fs_info;
10605         struct btrfs_root *root = fs_info->extent_root;
10606         struct btrfs_path *path;
10607         struct btrfs_block_group_cache *block_group;
10608         struct btrfs_free_cluster *cluster;
10609         struct btrfs_root *tree_root = fs_info->tree_root;
10610         struct btrfs_key key;
10611         struct inode *inode;
10612         struct kobject *kobj = NULL;
10613         int ret;
10614         int index;
10615         int factor;
10616         struct btrfs_caching_control *caching_ctl = NULL;
10617         bool remove_em;
10618         bool remove_rsv = false;
10619
10620         block_group = btrfs_lookup_block_group(fs_info, group_start);
10621         BUG_ON(!block_group);
10622         BUG_ON(!block_group->ro);
10623
10624         trace_btrfs_remove_block_group(block_group);
10625         /*
10626          * Free the reserved super bytes from this block group before
10627          * remove it.
10628          */
10629         free_excluded_extents(block_group);
10630         btrfs_free_ref_tree_range(fs_info, block_group->key.objectid,
10631                                   block_group->key.offset);
10632
10633         memcpy(&key, &block_group->key, sizeof(key));
10634         index = btrfs_bg_flags_to_raid_index(block_group->flags);
10635         factor = btrfs_bg_type_to_factor(block_group->flags);
10636
10637         /* make sure this block group isn't part of an allocation cluster */
10638         cluster = &fs_info->data_alloc_cluster;
10639         spin_lock(&cluster->refill_lock);
10640         btrfs_return_cluster_to_free_space(block_group, cluster);
10641         spin_unlock(&cluster->refill_lock);
10642
10643         /*
10644          * make sure this block group isn't part of a metadata
10645          * allocation cluster
10646          */
10647         cluster = &fs_info->meta_alloc_cluster;
10648         spin_lock(&cluster->refill_lock);
10649         btrfs_return_cluster_to_free_space(block_group, cluster);
10650         spin_unlock(&cluster->refill_lock);
10651
10652         path = btrfs_alloc_path();
10653         if (!path) {
10654                 ret = -ENOMEM;
10655                 goto out;
10656         }
10657
10658         /*
10659          * get the inode first so any iput calls done for the io_list
10660          * aren't the final iput (no unlinks allowed now)
10661          */
10662         inode = lookup_free_space_inode(block_group, path);
10663
10664         mutex_lock(&trans->transaction->cache_write_mutex);
10665         /*
10666          * Make sure our free space cache IO is done before removing the
10667          * free space inode
10668          */
10669         spin_lock(&trans->transaction->dirty_bgs_lock);
10670         if (!list_empty(&block_group->io_list)) {
10671                 list_del_init(&block_group->io_list);
10672
10673                 WARN_ON(!IS_ERR(inode) && inode != block_group->io_ctl.inode);
10674
10675                 spin_unlock(&trans->transaction->dirty_bgs_lock);
10676                 btrfs_wait_cache_io(trans, block_group, path);
10677                 btrfs_put_block_group(block_group);
10678                 spin_lock(&trans->transaction->dirty_bgs_lock);
10679         }
10680
10681         if (!list_empty(&block_group->dirty_list)) {
10682                 list_del_init(&block_group->dirty_list);
10683                 remove_rsv = true;
10684                 btrfs_put_block_group(block_group);
10685         }
10686         spin_unlock(&trans->transaction->dirty_bgs_lock);
10687         mutex_unlock(&trans->transaction->cache_write_mutex);
10688
10689         if (!IS_ERR(inode)) {
10690                 ret = btrfs_orphan_add(trans, BTRFS_I(inode));
10691                 if (ret) {
10692                         btrfs_add_delayed_iput(inode);
10693                         goto out;
10694                 }
10695                 clear_nlink(inode);
10696                 /* One for the block groups ref */
10697                 spin_lock(&block_group->lock);
10698                 if (block_group->iref) {
10699                         block_group->iref = 0;
10700                         block_group->inode = NULL;
10701                         spin_unlock(&block_group->lock);
10702                         iput(inode);
10703                 } else {
10704                         spin_unlock(&block_group->lock);
10705                 }
10706                 /* One for our lookup ref */
10707                 btrfs_add_delayed_iput(inode);
10708         }
10709
10710         key.objectid = BTRFS_FREE_SPACE_OBJECTID;
10711         key.offset = block_group->key.objectid;
10712         key.type = 0;
10713
10714         ret = btrfs_search_slot(trans, tree_root, &key, path, -1, 1);
10715         if (ret < 0)
10716                 goto out;
10717         if (ret > 0)
10718                 btrfs_release_path(path);
10719         if (ret == 0) {
10720                 ret = btrfs_del_item(trans, tree_root, path);
10721                 if (ret)
10722                         goto out;
10723                 btrfs_release_path(path);
10724         }
10725
10726         spin_lock(&fs_info->block_group_cache_lock);
10727         rb_erase(&block_group->cache_node,
10728                  &fs_info->block_group_cache_tree);
10729         RB_CLEAR_NODE(&block_group->cache_node);
10730
10731         if (fs_info->first_logical_byte == block_group->key.objectid)
10732                 fs_info->first_logical_byte = (u64)-1;
10733         spin_unlock(&fs_info->block_group_cache_lock);
10734
10735         down_write(&block_group->space_info->groups_sem);
10736         /*
10737          * we must use list_del_init so people can check to see if they
10738          * are still on the list after taking the semaphore
10739          */
10740         list_del_init(&block_group->list);
10741         if (list_empty(&block_group->space_info->block_groups[index])) {
10742                 kobj = block_group->space_info->block_group_kobjs[index];
10743                 block_group->space_info->block_group_kobjs[index] = NULL;
10744                 clear_avail_alloc_bits(fs_info, block_group->flags);
10745         }
10746         up_write(&block_group->space_info->groups_sem);
10747         if (kobj) {
10748                 kobject_del(kobj);
10749                 kobject_put(kobj);
10750         }
10751
10752         if (block_group->has_caching_ctl)
10753                 caching_ctl = get_caching_control(block_group);
10754         if (block_group->cached == BTRFS_CACHE_STARTED)
10755                 wait_block_group_cache_done(block_group);
10756         if (block_group->has_caching_ctl) {
10757                 down_write(&fs_info->commit_root_sem);
10758                 if (!caching_ctl) {
10759                         struct btrfs_caching_control *ctl;
10760
10761                         list_for_each_entry(ctl,
10762                                     &fs_info->caching_block_groups, list)
10763                                 if (ctl->block_group == block_group) {
10764                                         caching_ctl = ctl;
10765                                         refcount_inc(&caching_ctl->count);
10766                                         break;
10767                                 }
10768                 }
10769                 if (caching_ctl)
10770                         list_del_init(&caching_ctl->list);
10771                 up_write(&fs_info->commit_root_sem);
10772                 if (caching_ctl) {
10773                         /* Once for the caching bgs list and once for us. */
10774                         put_caching_control(caching_ctl);
10775                         put_caching_control(caching_ctl);
10776                 }
10777         }
10778
10779         spin_lock(&trans->transaction->dirty_bgs_lock);
10780         WARN_ON(!list_empty(&block_group->dirty_list));
10781         WARN_ON(!list_empty(&block_group->io_list));
10782         spin_unlock(&trans->transaction->dirty_bgs_lock);
10783
10784         btrfs_remove_free_space_cache(block_group);
10785
10786         spin_lock(&block_group->space_info->lock);
10787         list_del_init(&block_group->ro_list);
10788
10789         if (btrfs_test_opt(fs_info, ENOSPC_DEBUG)) {
10790                 WARN_ON(block_group->space_info->total_bytes
10791                         < block_group->key.offset);
10792                 WARN_ON(block_group->space_info->bytes_readonly
10793                         < block_group->key.offset);
10794                 WARN_ON(block_group->space_info->disk_total
10795                         < block_group->key.offset * factor);
10796         }
10797         block_group->space_info->total_bytes -= block_group->key.offset;
10798         block_group->space_info->bytes_readonly -= block_group->key.offset;
10799         block_group->space_info->disk_total -= block_group->key.offset * factor;
10800
10801         spin_unlock(&block_group->space_info->lock);
10802
10803         memcpy(&key, &block_group->key, sizeof(key));
10804
10805         mutex_lock(&fs_info->chunk_mutex);
10806         spin_lock(&block_group->lock);
10807         block_group->removed = 1;
10808         /*
10809          * At this point trimming can't start on this block group, because we
10810          * removed the block group from the tree fs_info->block_group_cache_tree
10811          * so no one can't find it anymore and even if someone already got this
10812          * block group before we removed it from the rbtree, they have already
10813          * incremented block_group->trimming - if they didn't, they won't find
10814          * any free space entries because we already removed them all when we
10815          * called btrfs_remove_free_space_cache().
10816          *
10817          * And we must not remove the extent map from the fs_info->mapping_tree
10818          * to prevent the same logical address range and physical device space
10819          * ranges from being reused for a new block group. This is because our
10820          * fs trim operation (btrfs_trim_fs() / btrfs_ioctl_fitrim()) is
10821          * completely transactionless, so while it is trimming a range the
10822          * currently running transaction might finish and a new one start,
10823          * allowing for new block groups to be created that can reuse the same
10824          * physical device locations unless we take this special care.
10825          *
10826          * There may also be an implicit trim operation if the file system
10827          * is mounted with -odiscard. The same protections must remain
10828          * in place until the extents have been discarded completely when
10829          * the transaction commit has completed.
10830          */
10831         remove_em = (atomic_read(&block_group->trimming) == 0);
10832         spin_unlock(&block_group->lock);
10833
10834         mutex_unlock(&fs_info->chunk_mutex);
10835
10836         ret = remove_block_group_free_space(trans, block_group);
10837         if (ret)
10838                 goto out;
10839
10840         btrfs_put_block_group(block_group);
10841         btrfs_put_block_group(block_group);
10842
10843         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
10844         if (ret > 0)
10845                 ret = -EIO;
10846         if (ret < 0)
10847                 goto out;
10848
10849         ret = btrfs_del_item(trans, root, path);
10850         if (ret)
10851                 goto out;
10852
10853         if (remove_em) {
10854                 struct extent_map_tree *em_tree;
10855
10856                 em_tree = &fs_info->mapping_tree.map_tree;
10857                 write_lock(&em_tree->lock);
10858                 remove_extent_mapping(em_tree, em);
10859                 write_unlock(&em_tree->lock);
10860                 /* once for the tree */
10861                 free_extent_map(em);
10862         }
10863 out:
10864         if (remove_rsv)
10865                 btrfs_delayed_refs_rsv_release(fs_info, 1);
10866         btrfs_free_path(path);
10867         return ret;
10868 }
10869
10870 struct btrfs_trans_handle *
10871 btrfs_start_trans_remove_block_group(struct btrfs_fs_info *fs_info,
10872                                      const u64 chunk_offset)
10873 {
10874         struct extent_map_tree *em_tree = &fs_info->mapping_tree.map_tree;
10875         struct extent_map *em;
10876         struct map_lookup *map;
10877         unsigned int num_items;
10878
10879         read_lock(&em_tree->lock);
10880         em = lookup_extent_mapping(em_tree, chunk_offset, 1);
10881         read_unlock(&em_tree->lock);
10882         ASSERT(em && em->start == chunk_offset);
10883
10884         /*
10885          * We need to reserve 3 + N units from the metadata space info in order
10886          * to remove a block group (done at btrfs_remove_chunk() and at
10887          * btrfs_remove_block_group()), which are used for:
10888          *
10889          * 1 unit for adding the free space inode's orphan (located in the tree
10890          * of tree roots).
10891          * 1 unit for deleting the block group item (located in the extent
10892          * tree).
10893          * 1 unit for deleting the free space item (located in tree of tree
10894          * roots).
10895          * N units for deleting N device extent items corresponding to each
10896          * stripe (located in the device tree).
10897          *
10898          * In order to remove a block group we also need to reserve units in the
10899          * system space info in order to update the chunk tree (update one or
10900          * more device items and remove one chunk item), but this is done at
10901          * btrfs_remove_chunk() through a call to check_system_chunk().
10902          */
10903         map = em->map_lookup;
10904         num_items = 3 + map->num_stripes;
10905         free_extent_map(em);
10906
10907         return btrfs_start_transaction_fallback_global_rsv(fs_info->extent_root,
10908                                                            num_items, 1);
10909 }
10910
10911 /*
10912  * Process the unused_bgs list and remove any that don't have any allocated
10913  * space inside of them.
10914  */
10915 void btrfs_delete_unused_bgs(struct btrfs_fs_info *fs_info)
10916 {
10917         struct btrfs_block_group_cache *block_group;
10918         struct btrfs_space_info *space_info;
10919         struct btrfs_trans_handle *trans;
10920         int ret = 0;
10921
10922         if (!test_bit(BTRFS_FS_OPEN, &fs_info->flags))
10923                 return;
10924
10925         spin_lock(&fs_info->unused_bgs_lock);
10926         while (!list_empty(&fs_info->unused_bgs)) {
10927                 u64 start, end;
10928                 int trimming;
10929
10930                 block_group = list_first_entry(&fs_info->unused_bgs,
10931                                                struct btrfs_block_group_cache,
10932                                                bg_list);
10933                 list_del_init(&block_group->bg_list);
10934
10935                 space_info = block_group->space_info;
10936
10937                 if (ret || btrfs_mixed_space_info(space_info)) {
10938                         btrfs_put_block_group(block_group);
10939                         continue;
10940                 }
10941                 spin_unlock(&fs_info->unused_bgs_lock);
10942
10943                 mutex_lock(&fs_info->delete_unused_bgs_mutex);
10944
10945                 /* Don't want to race with allocators so take the groups_sem */
10946                 down_write(&space_info->groups_sem);
10947                 spin_lock(&block_group->lock);
10948                 if (block_group->reserved || block_group->pinned ||
10949                     btrfs_block_group_used(&block_group->item) ||
10950                     block_group->ro ||
10951                     list_is_singular(&block_group->list)) {
10952                         /*
10953                          * We want to bail if we made new allocations or have
10954                          * outstanding allocations in this block group.  We do
10955                          * the ro check in case balance is currently acting on
10956                          * this block group.
10957                          */
10958                         trace_btrfs_skip_unused_block_group(block_group);
10959                         spin_unlock(&block_group->lock);
10960                         up_write(&space_info->groups_sem);
10961                         goto next;
10962                 }
10963                 spin_unlock(&block_group->lock);
10964
10965                 /* We don't want to force the issue, only flip if it's ok. */
10966                 ret = inc_block_group_ro(block_group, 0);
10967                 up_write(&space_info->groups_sem);
10968                 if (ret < 0) {
10969                         ret = 0;
10970                         goto next;
10971                 }
10972
10973                 /*
10974                  * Want to do this before we do anything else so we can recover
10975                  * properly if we fail to join the transaction.
10976                  */
10977                 trans = btrfs_start_trans_remove_block_group(fs_info,
10978                                                      block_group->key.objectid);
10979                 if (IS_ERR(trans)) {
10980                         btrfs_dec_block_group_ro(block_group);
10981                         ret = PTR_ERR(trans);
10982                         goto next;
10983                 }
10984
10985                 /*
10986                  * We could have pending pinned extents for this block group,
10987                  * just delete them, we don't care about them anymore.
10988                  */
10989                 start = block_group->key.objectid;
10990                 end = start + block_group->key.offset - 1;
10991                 /*
10992                  * Hold the unused_bg_unpin_mutex lock to avoid racing with
10993                  * btrfs_finish_extent_commit(). If we are at transaction N,
10994                  * another task might be running finish_extent_commit() for the
10995                  * previous transaction N - 1, and have seen a range belonging
10996                  * to the block group in freed_extents[] before we were able to
10997                  * clear the whole block group range from freed_extents[]. This
10998                  * means that task can lookup for the block group after we
10999                  * unpinned it from freed_extents[] and removed it, leading to
11000                  * a BUG_ON() at btrfs_unpin_extent_range().
11001                  */
11002                 mutex_lock(&fs_info->unused_bg_unpin_mutex);
11003                 ret = clear_extent_bits(&fs_info->freed_extents[0], start, end,
11004                                   EXTENT_DIRTY);
11005                 if (ret) {
11006                         mutex_unlock(&fs_info->unused_bg_unpin_mutex);
11007                         btrfs_dec_block_group_ro(block_group);
11008                         goto end_trans;
11009                 }
11010                 ret = clear_extent_bits(&fs_info->freed_extents[1], start, end,
11011                                   EXTENT_DIRTY);
11012                 if (ret) {
11013                         mutex_unlock(&fs_info->unused_bg_unpin_mutex);
11014                         btrfs_dec_block_group_ro(block_group);
11015                         goto end_trans;
11016                 }
11017                 mutex_unlock(&fs_info->unused_bg_unpin_mutex);
11018
11019                 /* Reset pinned so btrfs_put_block_group doesn't complain */
11020                 spin_lock(&space_info->lock);
11021                 spin_lock(&block_group->lock);
11022
11023                 update_bytes_pinned(space_info, -block_group->pinned);
11024                 space_info->bytes_readonly += block_group->pinned;
11025                 percpu_counter_add_batch(&space_info->total_bytes_pinned,
11026                                    -block_group->pinned,
11027                                    BTRFS_TOTAL_BYTES_PINNED_BATCH);
11028                 block_group->pinned = 0;
11029
11030                 spin_unlock(&block_group->lock);
11031                 spin_unlock(&space_info->lock);
11032
11033                 /* DISCARD can flip during remount */
11034                 trimming = btrfs_test_opt(fs_info, DISCARD);
11035
11036                 /* Implicit trim during transaction commit. */
11037                 if (trimming)
11038                         btrfs_get_block_group_trimming(block_group);
11039
11040                 /*
11041                  * Btrfs_remove_chunk will abort the transaction if things go
11042                  * horribly wrong.
11043                  */
11044                 ret = btrfs_remove_chunk(trans, block_group->key.objectid);
11045
11046                 if (ret) {
11047                         if (trimming)
11048                                 btrfs_put_block_group_trimming(block_group);
11049                         goto end_trans;
11050                 }
11051
11052                 /*
11053                  * If we're not mounted with -odiscard, we can just forget
11054                  * about this block group. Otherwise we'll need to wait
11055                  * until transaction commit to do the actual discard.
11056                  */
11057                 if (trimming) {
11058                         spin_lock(&fs_info->unused_bgs_lock);
11059                         /*
11060                          * A concurrent scrub might have added us to the list
11061                          * fs_info->unused_bgs, so use a list_move operation
11062                          * to add the block group to the deleted_bgs list.
11063                          */
11064                         list_move(&block_group->bg_list,
11065                                   &trans->transaction->deleted_bgs);
11066                         spin_unlock(&fs_info->unused_bgs_lock);
11067                         btrfs_get_block_group(block_group);
11068                 }
11069 end_trans:
11070                 btrfs_end_transaction(trans);
11071 next:
11072                 mutex_unlock(&fs_info->delete_unused_bgs_mutex);
11073                 btrfs_put_block_group(block_group);
11074                 spin_lock(&fs_info->unused_bgs_lock);
11075         }
11076         spin_unlock(&fs_info->unused_bgs_lock);
11077 }
11078
11079 int btrfs_init_space_info(struct btrfs_fs_info *fs_info)
11080 {
11081         struct btrfs_super_block *disk_super;
11082         u64 features;
11083         u64 flags;
11084         int mixed = 0;
11085         int ret;
11086
11087         disk_super = fs_info->super_copy;
11088         if (!btrfs_super_root(disk_super))
11089                 return -EINVAL;
11090
11091         features = btrfs_super_incompat_flags(disk_super);
11092         if (features & BTRFS_FEATURE_INCOMPAT_MIXED_GROUPS)
11093                 mixed = 1;
11094
11095         flags = BTRFS_BLOCK_GROUP_SYSTEM;
11096         ret = create_space_info(fs_info, flags);
11097         if (ret)
11098                 goto out;
11099
11100         if (mixed) {
11101                 flags = BTRFS_BLOCK_GROUP_METADATA | BTRFS_BLOCK_GROUP_DATA;
11102                 ret = create_space_info(fs_info, flags);
11103         } else {
11104                 flags = BTRFS_BLOCK_GROUP_METADATA;
11105                 ret = create_space_info(fs_info, flags);
11106                 if (ret)
11107                         goto out;
11108
11109                 flags = BTRFS_BLOCK_GROUP_DATA;
11110                 ret = create_space_info(fs_info, flags);
11111         }
11112 out:
11113         return ret;
11114 }
11115
11116 int btrfs_error_unpin_extent_range(struct btrfs_fs_info *fs_info,
11117                                    u64 start, u64 end)
11118 {
11119         return unpin_extent_range(fs_info, start, end, false);
11120 }
11121
11122 /*
11123  * It used to be that old block groups would be left around forever.
11124  * Iterating over them would be enough to trim unused space.  Since we
11125  * now automatically remove them, we also need to iterate over unallocated
11126  * space.
11127  *
11128  * We don't want a transaction for this since the discard may take a
11129  * substantial amount of time.  We don't require that a transaction be
11130  * running, but we do need to take a running transaction into account
11131  * to ensure that we're not discarding chunks that were released or
11132  * allocated in the current transaction.
11133  *
11134  * Holding the chunks lock will prevent other threads from allocating
11135  * or releasing chunks, but it won't prevent a running transaction
11136  * from committing and releasing the memory that the pending chunks
11137  * list head uses.  For that, we need to take a reference to the
11138  * transaction and hold the commit root sem.  We only need to hold
11139  * it while performing the free space search since we have already
11140  * held back allocations.
11141  */
11142 static int btrfs_trim_free_extents(struct btrfs_device *device, u64 *trimmed)
11143 {
11144         u64 start = SZ_1M, len = 0, end = 0;
11145         int ret;
11146
11147         *trimmed = 0;
11148
11149         /* Discard not supported = nothing to do. */
11150         if (!blk_queue_discard(bdev_get_queue(device->bdev)))
11151                 return 0;
11152
11153         /* Not writable = nothing to do. */
11154         if (!test_bit(BTRFS_DEV_STATE_WRITEABLE, &device->dev_state))
11155                 return 0;
11156
11157         /* No free space = nothing to do. */
11158         if (device->total_bytes <= device->bytes_used)
11159                 return 0;
11160
11161         ret = 0;
11162
11163         while (1) {
11164                 struct btrfs_fs_info *fs_info = device->fs_info;
11165                 u64 bytes;
11166
11167                 ret = mutex_lock_interruptible(&fs_info->chunk_mutex);
11168                 if (ret)
11169                         break;
11170
11171                 find_first_clear_extent_bit(&device->alloc_state, start,
11172                                             &start, &end,
11173                                             CHUNK_TRIMMED | CHUNK_ALLOCATED);
11174                 /*
11175                  * If find_first_clear_extent_bit find a range that spans the
11176                  * end of the device it will set end to -1, in this case it's up
11177                  * to the caller to trim the value to the size of the device.
11178                  */
11179                 end = min(end, device->total_bytes - 1);
11180                 len = end - start + 1;
11181
11182                 /* We didn't find any extents */
11183                 if (!len) {
11184                         mutex_unlock(&fs_info->chunk_mutex);
11185                         ret = 0;
11186                         break;
11187                 }
11188
11189                 ret = btrfs_issue_discard(device->bdev, start, len,
11190                                           &bytes);
11191                 if (!ret)
11192                         set_extent_bits(&device->alloc_state, start,
11193                                         start + bytes - 1,
11194                                         CHUNK_TRIMMED);
11195                 mutex_unlock(&fs_info->chunk_mutex);
11196
11197                 if (ret)
11198                         break;
11199
11200                 start += len;
11201                 *trimmed += bytes;
11202
11203                 if (fatal_signal_pending(current)) {
11204                         ret = -ERESTARTSYS;
11205                         break;
11206                 }
11207
11208                 cond_resched();
11209         }
11210
11211         return ret;
11212 }
11213
11214 /*
11215  * Trim the whole filesystem by:
11216  * 1) trimming the free space in each block group
11217  * 2) trimming the unallocated space on each device
11218  *
11219  * This will also continue trimming even if a block group or device encounters
11220  * an error.  The return value will be the last error, or 0 if nothing bad
11221  * happens.
11222  */
11223 int btrfs_trim_fs(struct btrfs_fs_info *fs_info, struct fstrim_range *range)
11224 {
11225         struct btrfs_block_group_cache *cache = NULL;
11226         struct btrfs_device *device;
11227         struct list_head *devices;
11228         u64 group_trimmed;
11229         u64 start;
11230         u64 end;
11231         u64 trimmed = 0;
11232         u64 bg_failed = 0;
11233         u64 dev_failed = 0;
11234         int bg_ret = 0;
11235         int dev_ret = 0;
11236         int ret = 0;
11237
11238         cache = btrfs_lookup_first_block_group(fs_info, range->start);
11239         for (; cache; cache = next_block_group(cache)) {
11240                 if (cache->key.objectid >= (range->start + range->len)) {
11241                         btrfs_put_block_group(cache);
11242                         break;
11243                 }
11244
11245                 start = max(range->start, cache->key.objectid);
11246                 end = min(range->start + range->len,
11247                                 cache->key.objectid + cache->key.offset);
11248
11249                 if (end - start >= range->minlen) {
11250                         if (!block_group_cache_done(cache)) {
11251                                 ret = cache_block_group(cache, 0);
11252                                 if (ret) {
11253                                         bg_failed++;
11254                                         bg_ret = ret;
11255                                         continue;
11256                                 }
11257                                 ret = wait_block_group_cache_done(cache);
11258                                 if (ret) {
11259                                         bg_failed++;
11260                                         bg_ret = ret;
11261                                         continue;
11262                                 }
11263                         }
11264                         ret = btrfs_trim_block_group(cache,
11265                                                      &group_trimmed,
11266                                                      start,
11267                                                      end,
11268                                                      range->minlen);
11269
11270                         trimmed += group_trimmed;
11271                         if (ret) {
11272                                 bg_failed++;
11273                                 bg_ret = ret;
11274                                 continue;
11275                         }
11276                 }
11277         }
11278
11279         if (bg_failed)
11280                 btrfs_warn(fs_info,
11281                         "failed to trim %llu block group(s), last error %d",
11282                         bg_failed, bg_ret);
11283         mutex_lock(&fs_info->fs_devices->device_list_mutex);
11284         devices = &fs_info->fs_devices->devices;
11285         list_for_each_entry(device, devices, dev_list) {
11286                 ret = btrfs_trim_free_extents(device, &group_trimmed);
11287                 if (ret) {
11288                         dev_failed++;
11289                         dev_ret = ret;
11290                         break;
11291                 }
11292
11293                 trimmed += group_trimmed;
11294         }
11295         mutex_unlock(&fs_info->fs_devices->device_list_mutex);
11296
11297         if (dev_failed)
11298                 btrfs_warn(fs_info,
11299                         "failed to trim %llu device(s), last error %d",
11300                         dev_failed, dev_ret);
11301         range->len = trimmed;
11302         if (bg_ret)
11303                 return bg_ret;
11304         return dev_ret;
11305 }
11306
11307 /*
11308  * btrfs_{start,end}_write_no_snapshotting() are similar to
11309  * mnt_{want,drop}_write(), they are used to prevent some tasks from writing
11310  * data into the page cache through nocow before the subvolume is snapshoted,
11311  * but flush the data into disk after the snapshot creation, or to prevent
11312  * operations while snapshotting is ongoing and that cause the snapshot to be
11313  * inconsistent (writes followed by expanding truncates for example).
11314  */
11315 void btrfs_end_write_no_snapshotting(struct btrfs_root *root)
11316 {
11317         percpu_counter_dec(&root->subv_writers->counter);
11318         cond_wake_up(&root->subv_writers->wait);
11319 }
11320
11321 int btrfs_start_write_no_snapshotting(struct btrfs_root *root)
11322 {
11323         if (atomic_read(&root->will_be_snapshotted))
11324                 return 0;
11325
11326         percpu_counter_inc(&root->subv_writers->counter);
11327         /*
11328          * Make sure counter is updated before we check for snapshot creation.
11329          */
11330         smp_mb();
11331         if (atomic_read(&root->will_be_snapshotted)) {
11332                 btrfs_end_write_no_snapshotting(root);
11333                 return 0;
11334         }
11335         return 1;
11336 }
11337
11338 void btrfs_wait_for_snapshot_creation(struct btrfs_root *root)
11339 {
11340         while (true) {
11341                 int ret;
11342
11343                 ret = btrfs_start_write_no_snapshotting(root);
11344                 if (ret)
11345                         break;
11346                 wait_var_event(&root->will_be_snapshotted,
11347                                !atomic_read(&root->will_be_snapshotted));
11348         }
11349 }
11350
11351 void btrfs_mark_bg_unused(struct btrfs_block_group_cache *bg)
11352 {
11353         struct btrfs_fs_info *fs_info = bg->fs_info;
11354
11355         spin_lock(&fs_info->unused_bgs_lock);
11356         if (list_empty(&bg->bg_list)) {
11357                 btrfs_get_block_group(bg);
11358                 trace_btrfs_add_unused_block_group(bg);
11359                 list_add_tail(&bg->bg_list, &fs_info->unused_bgs);
11360         }
11361         spin_unlock(&fs_info->unused_bgs_lock);
11362 }