fs/btrfs/extent-tree.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Copyright (C) 2007 Oracle.  All rights reserved.
   4  */
   5
   6 #include <linux/sched.h>
   7 #include <linux/sched/signal.h>
   8 #include <linux/pagemap.h>
   9 #include <linux/writeback.h>
  10 #include <linux/blkdev.h>
  11 #include <linux/sort.h>
  12 #include <linux/rcupdate.h>
  13 #include <linux/kthread.h>
  14 #include <linux/slab.h>
  15 #include <linux/ratelimit.h>
  16 #include <linux/percpu_counter.h>
  17 #include <linux/lockdep.h>
  18 #include <linux/crc32c.h>
  19 #include "tree-log.h"
  20 #include "disk-io.h"
  21 #include "print-tree.h"
  22 #include "volumes.h"
  23 #include "raid56.h"
  24 #include "locking.h"
  25 #include "free-space-cache.h"
  26 #include "free-space-tree.h"
  27 #include "math.h"
  28 #include "sysfs.h"
  29 #include "qgroup.h"
  30 #include "ref-verify.h"
  31
  32 #undef SCRAMBLE_DELAYED_REFS
  33
  34 /*
  35  * control flags for do_chunk_alloc's force field
  36  * CHUNK_ALLOC_NO_FORCE means to only allocate a chunk
  37  * if we really need one.
  38  *
  39  * CHUNK_ALLOC_LIMITED means to only try and allocate one
  40  * if we have very few chunks already allocated.  This is
  41  * used as part of the clustering code to help make sure
  42  * we have a good pool of storage to cluster in, without
  43  * filling the FS with empty chunks
  44  *
  45  * CHUNK_ALLOC_FORCE means it must try to allocate one
  46  *
  47  */
  48 enum {
  49         CHUNK_ALLOC_NO_FORCE = 0,
  50         CHUNK_ALLOC_LIMITED = 1,
  51         CHUNK_ALLOC_FORCE = 2,
  52 };
  53
  54 /*
  55  * Declare a helper function to detect underflow of various space info members
  56  */
  57 #define DECLARE_SPACE_INFO_UPDATE(name)                                 \
  58 static inline void update_##name(struct btrfs_space_info *sinfo,        \
  59                                  s64 bytes)                             \
  60 {                                                                       \
  61         if (bytes < 0 && sinfo->name < -bytes) {                        \
  62                 WARN_ON(1);                                             \
  63                 sinfo->name = 0;                                        \
  64                 return;                                                 \
  65         }                                                               \
  66         sinfo->name += bytes;                                           \
  67 }
  68
  69 DECLARE_SPACE_INFO_UPDATE(bytes_may_use);
  70 DECLARE_SPACE_INFO_UPDATE(bytes_pinned);
  71
  72 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
  73                                struct btrfs_delayed_ref_node *node, u64 parent,
  74                                u64 root_objectid, u64 owner_objectid,
  75                                u64 owner_offset, int refs_to_drop,
  76                                struct btrfs_delayed_extent_op *extra_op);
  77 static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op,
  78                                     struct extent_buffer *leaf,
  79                                     struct btrfs_extent_item *ei);
  80 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
  81                                       u64 parent, u64 root_objectid,
  82                                       u64 flags, u64 owner, u64 offset,
  83                                       struct btrfs_key *ins, int ref_mod);
  84 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
  85                                      struct btrfs_delayed_ref_node *node,
  86                                      struct btrfs_delayed_extent_op *extent_op);
  87 static int do_chunk_alloc(struct btrfs_trans_handle *trans, u64 flags,
  88                           int force);
  89 static int find_next_key(struct btrfs_path *path, int level,
  90                          struct btrfs_key *key);
  91 static void dump_space_info(struct btrfs_fs_info *fs_info,
  92                             struct btrfs_space_info *info, u64 bytes,
  93                             int dump_block_groups);
  94 static int block_rsv_use_bytes(struct btrfs_block_rsv *block_rsv,
  95                                u64 num_bytes);
  96 static void space_info_add_new_bytes(struct btrfs_fs_info *fs_info,
  97                                      struct btrfs_space_info *space_info,
  98                                      u64 num_bytes);
  99 static void space_info_add_old_bytes(struct btrfs_fs_info *fs_info,
 100                                      struct btrfs_space_info *space_info,
 101                                      u64 num_bytes);
 102
 103 static noinline int
 104 block_group_cache_done(struct btrfs_block_group_cache *cache)
 105 {
 106         smp_mb();
 107         return cache->cached == BTRFS_CACHE_FINISHED ||
 108                 cache->cached == BTRFS_CACHE_ERROR;
 109 }
 110
 111 static int block_group_bits(struct btrfs_block_group_cache *cache, u64 bits)
 112 {
 113         return (cache->flags & bits) == bits;
 114 }
 115
 116 void btrfs_get_block_group(struct btrfs_block_group_cache *cache)
 117 {
 118         atomic_inc(&cache->count);
 119 }
 120
 121 void btrfs_put_block_group(struct btrfs_block_group_cache *cache)
 122 {
 123         if (atomic_dec_and_test(&cache->count)) {
 124                 WARN_ON(cache->pinned > 0);
 125                 WARN_ON(cache->reserved > 0);
 126
 127                 /*
 128                  * If not empty, someone is still holding mutex of
 129                  * full_stripe_lock, which can only be released by caller.
 130                  * And it will definitely cause use-after-free when caller
 131                  * tries to release full stripe lock.
 132                  *
 133                  * No better way to resolve, but only to warn.
 134                  */
 135                 WARN_ON(!RB_EMPTY_ROOT(&cache->full_stripe_locks_root.root));
 136                 kfree(cache->free_space_ctl);
 137                 kfree(cache);
 138         }
 139 }
 140
 141 /*
 142  * this adds the block group to the fs_info rb tree for the block group
 143  * cache
 144  */
 145 static int btrfs_add_block_group_cache(struct btrfs_fs_info *info,
 146                                 struct btrfs_block_group_cache *block_group)
 147 {
 148         struct rb_node **p;
 149         struct rb_node *parent = NULL;
 150         struct btrfs_block_group_cache *cache;
 151
 152         spin_lock(&info->block_group_cache_lock);
 153         p = &info->block_group_cache_tree.rb_node;
 154
 155         while (*p) {
 156                 parent = *p;
 157                 cache = rb_entry(parent, struct btrfs_block_group_cache,
 158                                  cache_node);
 159                 if (block_group->key.objectid < cache->key.objectid) {
 160                         p = &(*p)->rb_left;
 161                 } else if (block_group->key.objectid > cache->key.objectid) {
 162                         p = &(*p)->rb_right;
 163                 } else {
 164                         spin_unlock(&info->block_group_cache_lock);
 165                         return -EEXIST;
 166                 }
 167         }
 168
 169         rb_link_node(&block_group->cache_node, parent, p);
 170         rb_insert_color(&block_group->cache_node,
 171                         &info->block_group_cache_tree);
 172
 173         if (info->first_logical_byte > block_group->key.objectid)
 174                 info->first_logical_byte = block_group->key.objectid;
 175
 176         spin_unlock(&info->block_group_cache_lock);
 177
 178         return 0;
 179 }
 180
 181 /*
 182  * This will return the block group at or after bytenr if contains is 0, else
 183  * it will return the block group that contains the bytenr
 184  */
 185 static struct btrfs_block_group_cache *
 186 block_group_cache_tree_search(struct btrfs_fs_info *info, u64 bytenr,
 187                               int contains)
 188 {
 189         struct btrfs_block_group_cache *cache, *ret = NULL;
 190         struct rb_node *n;
 191         u64 end, start;
 192
 193         spin_lock(&info->block_group_cache_lock);
 194         n = info->block_group_cache_tree.rb_node;
 195
 196         while (n) {
 197                 cache = rb_entry(n, struct btrfs_block_group_cache,
 198                                  cache_node);
 199                 end = cache->key.objectid + cache->key.offset - 1;
 200                 start = cache->key.objectid;
 201
 202                 if (bytenr < start) {
 203                         if (!contains && (!ret || start < ret->key.objectid))
 204                                 ret = cache;
 205                         n = n->rb_left;
 206                 } else if (bytenr > start) {
 207                         if (contains && bytenr <= end) {
 208                                 ret = cache;
 209                                 break;
 210                         }
 211                         n = n->rb_right;
 212                 } else {
 213                         ret = cache;
 214                         break;
 215                 }
 216         }
 217         if (ret) {
 218                 btrfs_get_block_group(ret);
 219                 if (bytenr == 0 && info->first_logical_byte > ret->key.objectid)
 220                         info->first_logical_byte = ret->key.objectid;
 221         }
 222         spin_unlock(&info->block_group_cache_lock);
 223
 224         return ret;
 225 }
 226
 227 static int add_excluded_extent(struct btrfs_fs_info *fs_info,
 228                                u64 start, u64 num_bytes)
 229 {
 230         u64 end = start + num_bytes - 1;
 231         set_extent_bits(&fs_info->freed_extents[0],
 232                         start, end, EXTENT_UPTODATE);
 233         set_extent_bits(&fs_info->freed_extents[1],
 234                         start, end, EXTENT_UPTODATE);
 235         return 0;
 236 }
 237
 238 static void free_excluded_extents(struct btrfs_block_group_cache *cache)
 239 {
 240         struct btrfs_fs_info *fs_info = cache->fs_info;
 241         u64 start, end;
 242
 243         start = cache->key.objectid;
 244         end = start + cache->key.offset - 1;
 245
 246         clear_extent_bits(&fs_info->freed_extents[0],
 247                           start, end, EXTENT_UPTODATE);
 248         clear_extent_bits(&fs_info->freed_extents[1],
 249                           start, end, EXTENT_UPTODATE);
 250 }
 251
 252 static int exclude_super_stripes(struct btrfs_block_group_cache *cache)
 253 {
 254         struct btrfs_fs_info *fs_info = cache->fs_info;
 255         u64 bytenr;
 256         u64 *logical;
 257         int stripe_len;
 258         int i, nr, ret;
 259
 260         if (cache->key.objectid < BTRFS_SUPER_INFO_OFFSET) {
 261                 stripe_len = BTRFS_SUPER_INFO_OFFSET - cache->key.objectid;
 262                 cache->bytes_super += stripe_len;
 263                 ret = add_excluded_extent(fs_info, cache->key.objectid,
 264                                           stripe_len);
 265                 if (ret)
 266                         return ret;
 267         }
 268
 269         for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
 270                 bytenr = btrfs_sb_offset(i);
 271                 ret = btrfs_rmap_block(fs_info, cache->key.objectid,
 272                                        bytenr, &logical, &nr, &stripe_len);
 273                 if (ret)
 274                         return ret;
 275
 276                 while (nr--) {
 277                         u64 start, len;
 278
 279                         if (logical[nr] > cache->key.objectid +
 280                             cache->key.offset)
 281                                 continue;
 282
 283                         if (logical[nr] + stripe_len <= cache->key.objectid)
 284                                 continue;
 285
 286                         start = logical[nr];
 287                         if (start < cache->key.objectid) {
 288                                 start = cache->key.objectid;
 289                                 len = (logical[nr] + stripe_len) - start;
 290                         } else {
 291                                 len = min_t(u64, stripe_len,
 292                                             cache->key.objectid +
 293                                             cache->key.offset - start);
 294                         }
 295
 296                         cache->bytes_super += len;
 297                         ret = add_excluded_extent(fs_info, start, len);
 298                         if (ret) {
 299                                 kfree(logical);
 300                                 return ret;
 301                         }
 302                 }
 303
 304                 kfree(logical);
 305         }
 306         return 0;
 307 }
 308
 309 static struct btrfs_caching_control *
 310 get_caching_control(struct btrfs_block_group_cache *cache)
 311 {
 312         struct btrfs_caching_control *ctl;
 313
 314         spin_lock(&cache->lock);
 315         if (!cache->caching_ctl) {
 316                 spin_unlock(&cache->lock);
 317                 return NULL;
 318         }
 319
 320         ctl = cache->caching_ctl;
 321         refcount_inc(&ctl->count);
 322         spin_unlock(&cache->lock);
 323         return ctl;
 324 }
 325
 326 static void put_caching_control(struct btrfs_caching_control *ctl)
 327 {
 328         if (refcount_dec_and_test(&ctl->count))
 329                 kfree(ctl);
 330 }
 331
 332 #ifdef CONFIG_BTRFS_DEBUG
 333 static void fragment_free_space(struct btrfs_block_group_cache *block_group)
 334 {
 335         struct btrfs_fs_info *fs_info = block_group->fs_info;
 336         u64 start = block_group->key.objectid;
 337         u64 len = block_group->key.offset;
 338         u64 chunk = block_group->flags & BTRFS_BLOCK_GROUP_METADATA ?
 339                 fs_info->nodesize : fs_info->sectorsize;
 340         u64 step = chunk << 1;
 341
 342         while (len > chunk) {
 343                 btrfs_remove_free_space(block_group, start, chunk);
 344                 start += step;
 345                 if (len < step)
 346                         len = 0;
 347                 else
 348                         len -= step;
 349         }
 350 }
 351 #endif
 352
 353 /*
 354  * this is only called by cache_block_group, since we could have freed extents
 355  * we need to check the pinned_extents for any extents that can't be used yet
 356  * since their free space will be released as soon as the transaction commits.
 357  */
 358 u64 add_new_free_space(struct btrfs_block_group_cache *block_group,
 359                        u64 start, u64 end)
 360 {
 361         struct btrfs_fs_info *info = block_group->fs_info;
 362         u64 extent_start, extent_end, size, total_added = 0;
 363         int ret;
 364
 365         while (start < end) {
 366                 ret = find_first_extent_bit(info->pinned_extents, start,
 367                                             &extent_start, &extent_end,
 368                                             EXTENT_DIRTY | EXTENT_UPTODATE,
 369                                             NULL);
 370                 if (ret)
 371                         break;
 372
 373                 if (extent_start <= start) {
 374                         start = extent_end + 1;
 375                 } else if (extent_start > start && extent_start < end) {
 376                         size = extent_start - start;
 377                         total_added += size;
 378                         ret = btrfs_add_free_space(block_group, start,
 379                                                    size);
 380                         BUG_ON(ret); /* -ENOMEM or logic error */
 381                         start = extent_end + 1;
 382                 } else {
 383                         break;
 384                 }
 385         }
 386
 387         if (start < end) {
 388                 size = end - start;
 389                 total_added += size;
 390                 ret = btrfs_add_free_space(block_group, start, size);
 391                 BUG_ON(ret); /* -ENOMEM or logic error */
 392         }
 393
 394         return total_added;
 395 }
 396
 397 static int load_extent_tree_free(struct btrfs_caching_control *caching_ctl)
 398 {
 399         struct btrfs_block_group_cache *block_group = caching_ctl->block_group;
 400         struct btrfs_fs_info *fs_info = block_group->fs_info;
 401         struct btrfs_root *extent_root = fs_info->extent_root;
 402         struct btrfs_path *path;
 403         struct extent_buffer *leaf;
 404         struct btrfs_key key;
 405         u64 total_found = 0;
 406         u64 last = 0;
 407         u32 nritems;
 408         int ret;
 409         bool wakeup = true;
 410
 411         path = btrfs_alloc_path();
 412         if (!path)
 413                 return -ENOMEM;
 414
 415         last = max_t(u64, block_group->key.objectid, BTRFS_SUPER_INFO_OFFSET);
 416
 417 #ifdef CONFIG_BTRFS_DEBUG
 418         /*
 419          * If we're fragmenting we don't want to make anybody think we can
 420          * allocate from this block group until we've had a chance to fragment
 421          * the free space.
 422          */
 423         if (btrfs_should_fragment_free_space(block_group))
 424                 wakeup = false;
 425 #endif
 426         /*
 427          * We don't want to deadlock with somebody trying to allocate a new
 428          * extent for the extent root while also trying to search the extent
 429          * root to add free space.  So we skip locking and search the commit
 430          * root, since its read-only
 431          */
 432         path->skip_locking = 1;
 433         path->search_commit_root = 1;
 434         path->reada = READA_FORWARD;
 435
 436         key.objectid = last;
 437         key.offset = 0;
 438         key.type = BTRFS_EXTENT_ITEM_KEY;
 439
 440 next:
 441         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
 442         if (ret < 0)
 443                 goto out;
 444
 445         leaf = path->nodes[0];
 446         nritems = btrfs_header_nritems(leaf);
 447
 448         while (1) {
 449                 if (btrfs_fs_closing(fs_info) > 1) {
 450                         last = (u64)-1;
 451                         break;
 452                 }
 453
 454                 if (path->slots[0] < nritems) {
 455                         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
 456                 } else {
 457                         ret = find_next_key(path, 0, &key);
 458                         if (ret)
 459                                 break;
 460
 461                         if (need_resched() ||
 462                             rwsem_is_contended(&fs_info->commit_root_sem)) {
 463                                 if (wakeup)
 464                                         caching_ctl->progress = last;
 465                                 btrfs_release_path(path);
 466                                 up_read(&fs_info->commit_root_sem);
 467                                 mutex_unlock(&caching_ctl->mutex);
 468                                 cond_resched();
 469                                 mutex_lock(&caching_ctl->mutex);
 470                                 down_read(&fs_info->commit_root_sem);
 471                                 goto next;
 472                         }
 473
 474                         ret = btrfs_next_leaf(extent_root, path);
 475                         if (ret < 0)
 476                                 goto out;
 477                         if (ret)
 478                                 break;
 479                         leaf = path->nodes[0];
 480                         nritems = btrfs_header_nritems(leaf);
 481                         continue;
 482                 }
 483
 484                 if (key.objectid < last) {
 485                         key.objectid = last;
 486                         key.offset = 0;
 487                         key.type = BTRFS_EXTENT_ITEM_KEY;
 488
 489                         if (wakeup)
 490                                 caching_ctl->progress = last;
 491                         btrfs_release_path(path);
 492                         goto next;
 493                 }
 494
 495                 if (key.objectid < block_group->key.objectid) {
 496                         path->slots[0]++;
 497                         continue;
 498                 }
 499
 500                 if (key.objectid >= block_group->key.objectid +
 501                     block_group->key.offset)
 502                         break;
 503
 504                 if (key.type == BTRFS_EXTENT_ITEM_KEY ||
 505                     key.type == BTRFS_METADATA_ITEM_KEY) {
 506                         total_found += add_new_free_space(block_group, last,
 507                                                           key.objectid);
 508                         if (key.type == BTRFS_METADATA_ITEM_KEY)
 509                                 last = key.objectid +
 510                                         fs_info->nodesize;
 511                         else
 512                                 last = key.objectid + key.offset;
 513
 514                         if (total_found > CACHING_CTL_WAKE_UP) {
 515                                 total_found = 0;
 516                                 if (wakeup)
 517                                         wake_up(&caching_ctl->wait);
 518                         }
 519                 }
 520                 path->slots[0]++;
 521         }
 522         ret = 0;
 523
 524         total_found += add_new_free_space(block_group, last,
 525                                           block_group->key.objectid +
 526                                           block_group->key.offset);
 527         caching_ctl->progress = (u64)-1;
 528
 529 out:
 530         btrfs_free_path(path);
 531         return ret;
 532 }
 533
 534 static noinline void caching_thread(struct btrfs_work *work)
 535 {
 536         struct btrfs_block_group_cache *block_group;
 537         struct btrfs_fs_info *fs_info;
 538         struct btrfs_caching_control *caching_ctl;
 539         int ret;
 540
 541         caching_ctl = container_of(work, struct btrfs_caching_control, work);
 542         block_group = caching_ctl->block_group;
 543         fs_info = block_group->fs_info;
 544
 545         mutex_lock(&caching_ctl->mutex);
 546         down_read(&fs_info->commit_root_sem);
 547
 548         if (btrfs_fs_compat_ro(fs_info, FREE_SPACE_TREE))
 549                 ret = load_free_space_tree(caching_ctl);
 550         else
 551                 ret = load_extent_tree_free(caching_ctl);
 552
 553         spin_lock(&block_group->lock);
 554         block_group->caching_ctl = NULL;
 555         block_group->cached = ret ? BTRFS_CACHE_ERROR : BTRFS_CACHE_FINISHED;
 556         spin_unlock(&block_group->lock);
 557
 558 #ifdef CONFIG_BTRFS_DEBUG
 559         if (btrfs_should_fragment_free_space(block_group)) {
 560                 u64 bytes_used;
 561
 562                 spin_lock(&block_group->space_info->lock);
 563                 spin_lock(&block_group->lock);
 564                 bytes_used = block_group->key.offset -
 565                         btrfs_block_group_used(&block_group->item);
 566                 block_group->space_info->bytes_used += bytes_used >> 1;
 567                 spin_unlock(&block_group->lock);
 568                 spin_unlock(&block_group->space_info->lock);
 569                 fragment_free_space(block_group);
 570         }
 571 #endif
 572
 573         caching_ctl->progress = (u64)-1;
 574
 575         up_read(&fs_info->commit_root_sem);
 576         free_excluded_extents(block_group);
 577         mutex_unlock(&caching_ctl->mutex);
 578
 579         wake_up(&caching_ctl->wait);
 580
 581         put_caching_control(caching_ctl);
 582         btrfs_put_block_group(block_group);
 583 }
 584
 585 static int cache_block_group(struct btrfs_block_group_cache *cache,
 586                              int load_cache_only)
 587 {
 588         DEFINE_WAIT(wait);
 589         struct btrfs_fs_info *fs_info = cache->fs_info;
 590         struct btrfs_caching_control *caching_ctl;
 591         int ret = 0;
 592
 593         caching_ctl = kzalloc(sizeof(*caching_ctl), GFP_NOFS);
 594         if (!caching_ctl)
 595                 return -ENOMEM;
 596
 597         INIT_LIST_HEAD(&caching_ctl->list);
 598         mutex_init(&caching_ctl->mutex);
 599         init_waitqueue_head(&caching_ctl->wait);
 600         caching_ctl->block_group = cache;
 601         caching_ctl->progress = cache->key.objectid;
 602         refcount_set(&caching_ctl->count, 1);
 603         btrfs_init_work(&caching_ctl->work, btrfs_cache_helper,
 604                         caching_thread, NULL, NULL);
 605
 606         spin_lock(&cache->lock);
 607         /*
 608          * This should be a rare occasion, but this could happen I think in the
 609          * case where one thread starts to load the space cache info, and then
 610          * some other thread starts a transaction commit which tries to do an
 611          * allocation while the other thread is still loading the space cache
 612          * info.  The previous loop should have kept us from choosing this block
 613          * group, but if we've moved to the state where we will wait on caching
 614          * block groups we need to first check if we're doing a fast load here,
 615          * so we can wait for it to finish, otherwise we could end up allocating
 616          * from a block group who's cache gets evicted for one reason or
 617          * another.
 618          */
 619         while (cache->cached == BTRFS_CACHE_FAST) {
 620                 struct btrfs_caching_control *ctl;
 621
 622                 ctl = cache->caching_ctl;
 623                 refcount_inc(&ctl->count);
 624                 prepare_to_wait(&ctl->wait, &wait, TASK_UNINTERRUPTIBLE);
 625                 spin_unlock(&cache->lock);
 626
 627                 schedule();
 628
 629                 finish_wait(&ctl->wait, &wait);
 630                 put_caching_control(ctl);
 631                 spin_lock(&cache->lock);
 632         }
 633
 634         if (cache->cached != BTRFS_CACHE_NO) {
 635                 spin_unlock(&cache->lock);
 636                 kfree(caching_ctl);
 637                 return 0;
 638         }
 639         WARN_ON(cache->caching_ctl);
 640         cache->caching_ctl = caching_ctl;
 641         cache->cached = BTRFS_CACHE_FAST;
 642         spin_unlock(&cache->lock);
 643
 644         if (btrfs_test_opt(fs_info, SPACE_CACHE)) {
 645                 mutex_lock(&caching_ctl->mutex);
 646                 ret = load_free_space_cache(fs_info, cache);
 647
 648                 spin_lock(&cache->lock);
 649                 if (ret == 1) {
 650                         cache->caching_ctl = NULL;
 651                         cache->cached = BTRFS_CACHE_FINISHED;
 652                         cache->last_byte_to_unpin = (u64)-1;
 653                         caching_ctl->progress = (u64)-1;
 654                 } else {
 655                         if (load_cache_only) {
 656                                 cache->caching_ctl = NULL;
 657                                 cache->cached = BTRFS_CACHE_NO;
 658                         } else {
 659                                 cache->cached = BTRFS_CACHE_STARTED;
 660                                 cache->has_caching_ctl = 1;
 661                         }
 662                 }
 663                 spin_unlock(&cache->lock);
 664 #ifdef CONFIG_BTRFS_DEBUG
 665                 if (ret == 1 &&
 666                     btrfs_should_fragment_free_space(cache)) {
 667                         u64 bytes_used;
 668
 669                         spin_lock(&cache->space_info->lock);
 670                         spin_lock(&cache->lock);
 671                         bytes_used = cache->key.offset -
 672                                 btrfs_block_group_used(&cache->item);
 673                         cache->space_info->bytes_used += bytes_used >> 1;
 674                         spin_unlock(&cache->lock);
 675                         spin_unlock(&cache->space_info->lock);
 676                         fragment_free_space(cache);
 677                 }
 678 #endif
 679                 mutex_unlock(&caching_ctl->mutex);
 680
 681                 wake_up(&caching_ctl->wait);
 682                 if (ret == 1) {
 683                         put_caching_control(caching_ctl);
 684                         free_excluded_extents(cache);
 685                         return 0;
 686                 }
 687         } else {
 688                 /*
 689                  * We're either using the free space tree or no caching at all.
 690                  * Set cached to the appropriate value and wakeup any waiters.
 691                  */
 692                 spin_lock(&cache->lock);
 693                 if (load_cache_only) {
 694                         cache->caching_ctl = NULL;
 695                         cache->cached = BTRFS_CACHE_NO;
 696                 } else {
 697                         cache->cached = BTRFS_CACHE_STARTED;
 698                         cache->has_caching_ctl = 1;
 699                 }
 700                 spin_unlock(&cache->lock);
 701                 wake_up(&caching_ctl->wait);
 702         }
 703
 704         if (load_cache_only) {
 705                 put_caching_control(caching_ctl);
 706                 return 0;
 707         }
 708
 709         down_write(&fs_info->commit_root_sem);
 710         refcount_inc(&caching_ctl->count);
 711         list_add_tail(&caching_ctl->list, &fs_info->caching_block_groups);
 712         up_write(&fs_info->commit_root_sem);
 713
 714         btrfs_get_block_group(cache);
 715
 716         btrfs_queue_work(fs_info->caching_workers, &caching_ctl->work);
 717
 718         return ret;
 719 }
 720
 721 /*
 722  * return the block group that starts at or after bytenr
 723  */
 724 static struct btrfs_block_group_cache *
 725 btrfs_lookup_first_block_group(struct btrfs_fs_info *info, u64 bytenr)
 726 {
 727         return block_group_cache_tree_search(info, bytenr, 0);
 728 }
 729
 730 /*
 731  * return the block group that contains the given bytenr
 732  */
 733 struct btrfs_block_group_cache *btrfs_lookup_block_group(
 734                                                  struct btrfs_fs_info *info,
 735                                                  u64 bytenr)
 736 {
 737         return block_group_cache_tree_search(info, bytenr, 1);
 738 }
 739
 740 static struct btrfs_space_info *__find_space_info(struct btrfs_fs_info *info,
 741                                                   u64 flags)
 742 {
 743         struct list_head *head = &info->space_info;
 744         struct btrfs_space_info *found;
 745
 746         flags &= BTRFS_BLOCK_GROUP_TYPE_MASK;
 747
 748         rcu_read_lock();
 749         list_for_each_entry_rcu(found, head, list) {
 750                 if (found->flags & flags) {
 751                         rcu_read_unlock();
 752                         return found;
 753                 }
 754         }
 755         rcu_read_unlock();
 756         return NULL;
 757 }
 758
 759 static void add_pinned_bytes(struct btrfs_fs_info *fs_info, s64 num_bytes,
 760                              bool metadata, u64 root_objectid)
 761 {
 762         struct btrfs_space_info *space_info;
 763         u64 flags;
 764
 765         if (metadata) {
 766                 if (root_objectid == BTRFS_CHUNK_TREE_OBJECTID)
 767                         flags = BTRFS_BLOCK_GROUP_SYSTEM;
 768                 else
 769                         flags = BTRFS_BLOCK_GROUP_METADATA;
 770         } else {
 771                 flags = BTRFS_BLOCK_GROUP_DATA;
 772         }
 773
 774         space_info = __find_space_info(fs_info, flags);
 775         ASSERT(space_info);
 776         percpu_counter_add_batch(&space_info->total_bytes_pinned, num_bytes,
 777                     BTRFS_TOTAL_BYTES_PINNED_BATCH);
 778 }
 779
 780 /*
 781  * after adding space to the filesystem, we need to clear the full flags
 782  * on all the space infos.
 783  */
 784 void btrfs_clear_space_info_full(struct btrfs_fs_info *info)
 785 {
 786         struct list_head *head = &info->space_info;
 787         struct btrfs_space_info *found;
 788
 789         rcu_read_lock();
 790         list_for_each_entry_rcu(found, head, list)
 791                 found->full = 0;
 792         rcu_read_unlock();
 793 }
 794
 795 /* simple helper to search for an existing data extent at a given offset */
 796 int btrfs_lookup_data_extent(struct btrfs_fs_info *fs_info, u64 start, u64 len)
 797 {
 798         int ret;
 799         struct btrfs_key key;
 800         struct btrfs_path *path;
 801
 802         path = btrfs_alloc_path();
 803         if (!path)
 804                 return -ENOMEM;
 805
 806         key.objectid = start;
 807         key.offset = len;
 808         key.type = BTRFS_EXTENT_ITEM_KEY;
 809         ret = btrfs_search_slot(NULL, fs_info->extent_root, &key, path, 0, 0);
 810         btrfs_free_path(path);
 811         return ret;
 812 }
 813
 814 /*
 815  * helper function to lookup reference count and flags of a tree block.
 816  *
 817  * the head node for delayed ref is used to store the sum of all the
 818  * reference count modifications queued up in the rbtree. the head
 819  * node may also store the extent flags to set. This way you can check
 820  * to see what the reference count and extent flags would be if all of
 821  * the delayed refs are not processed.
 822  */
 823 int btrfs_lookup_extent_info(struct btrfs_trans_handle *trans,
 824                              struct btrfs_fs_info *fs_info, u64 bytenr,
 825                              u64 offset, int metadata, u64 *refs, u64 *flags)
 826 {
 827         struct btrfs_delayed_ref_head *head;
 828         struct btrfs_delayed_ref_root *delayed_refs;
 829         struct btrfs_path *path;
 830         struct btrfs_extent_item *ei;
 831         struct extent_buffer *leaf;
 832         struct btrfs_key key;
 833         u32 item_size;
 834         u64 num_refs;
 835         u64 extent_flags;
 836         int ret;
 837
 838         /*
 839          * If we don't have skinny metadata, don't bother doing anything
 840          * different
 841          */
 842         if (metadata && !btrfs_fs_incompat(fs_info, SKINNY_METADATA)) {
 843                 offset = fs_info->nodesize;
 844                 metadata = 0;
 845         }
 846
 847         path = btrfs_alloc_path();
 848         if (!path)
 849                 return -ENOMEM;
 850
 851         if (!trans) {
 852                 path->skip_locking = 1;
 853                 path->search_commit_root = 1;
 854         }
 855
 856 search_again:
 857         key.objectid = bytenr;
 858         key.offset = offset;
 859         if (metadata)
 860                 key.type = BTRFS_METADATA_ITEM_KEY;
 861         else
 862                 key.type = BTRFS_EXTENT_ITEM_KEY;
 863
 864         ret = btrfs_search_slot(trans, fs_info->extent_root, &key, path, 0, 0);
 865         if (ret < 0)
 866                 goto out_free;
 867
 868         if (ret > 0 && metadata && key.type == BTRFS_METADATA_ITEM_KEY) {
 869                 if (path->slots[0]) {
 870                         path->slots[0]--;
 871                         btrfs_item_key_to_cpu(path->nodes[0], &key,
 872                                               path->slots[0]);
 873                         if (key.objectid == bytenr &&
 874                             key.type == BTRFS_EXTENT_ITEM_KEY &&
 875                             key.offset == fs_info->nodesize)
 876                                 ret = 0;
 877                 }
 878         }
 879
 880         if (ret == 0) {
 881                 leaf = path->nodes[0];
 882                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
 883                 if (item_size >= sizeof(*ei)) {
 884                         ei = btrfs_item_ptr(leaf, path->slots[0],
 885                                             struct btrfs_extent_item);
 886                         num_refs = btrfs_extent_refs(leaf, ei);
 887                         extent_flags = btrfs_extent_flags(leaf, ei);
 888                 } else {
 889                         ret = -EINVAL;
 890                         btrfs_print_v0_err(fs_info);
 891                         if (trans)
 892                                 btrfs_abort_transaction(trans, ret);
 893                         else
 894                                 btrfs_handle_fs_error(fs_info, ret, NULL);
 895
 896                         goto out_free;
 897                 }
 898
 899                 BUG_ON(num_refs == 0);
 900         } else {
 901                 num_refs = 0;
 902                 extent_flags = 0;
 903                 ret = 0;
 904         }
 905
 906         if (!trans)
 907                 goto out;
 908
 909         delayed_refs = &trans->transaction->delayed_refs;
 910         spin_lock(&delayed_refs->lock);
 911         head = btrfs_find_delayed_ref_head(delayed_refs, bytenr);
 912         if (head) {
 913                 if (!mutex_trylock(&head->mutex)) {
 914                         refcount_inc(&head->refs);
 915                         spin_unlock(&delayed_refs->lock);
 916
 917                         btrfs_release_path(path);
 918
 919                         /*
 920                          * Mutex was contended, block until it's released and try
 921                          * again
 922                          */
 923                         mutex_lock(&head->mutex);
 924                         mutex_unlock(&head->mutex);
 925                         btrfs_put_delayed_ref_head(head);
 926                         goto search_again;
 927                 }
 928                 spin_lock(&head->lock);
 929                 if (head->extent_op && head->extent_op->update_flags)
 930                         extent_flags |= head->extent_op->flags_to_set;
 931                 else
 932                         BUG_ON(num_refs == 0);
 933
 934                 num_refs += head->ref_mod;
 935                 spin_unlock(&head->lock);
 936                 mutex_unlock(&head->mutex);
 937         }
 938         spin_unlock(&delayed_refs->lock);
 939 out:
 940         WARN_ON(num_refs == 0);
 941         if (refs)
 942                 *refs = num_refs;
 943         if (flags)
 944                 *flags = extent_flags;
 945 out_free:
 946         btrfs_free_path(path);
 947         return ret;
 948 }
 949
 950 /*
 951  * Back reference rules.  Back refs have three main goals:
 952  *
 953  * 1) differentiate between all holders of references to an extent so that
 954  *    when a reference is dropped we can make sure it was a valid reference
 955  *    before freeing the extent.
 956  *
 957  * 2) Provide enough information to quickly find the holders of an extent
 958  *    if we notice a given block is corrupted or bad.
 959  *
 960  * 3) Make it easy to migrate blocks for FS shrinking or storage pool
 961  *    maintenance.  This is actually the same as #2, but with a slightly
 962  *    different use case.
 963  *
 964  * There are two kinds of back refs. The implicit back refs is optimized
 965  * for pointers in non-shared tree blocks. For a given pointer in a block,
 966  * back refs of this kind provide information about the block's owner tree
 967  * and the pointer's key. These information allow us to find the block by
 968  * b-tree searching. The full back refs is for pointers in tree blocks not
 969  * referenced by their owner trees. The location of tree block is recorded
 970  * in the back refs. Actually the full back refs is generic, and can be
 971  * used in all cases the implicit back refs is used. The major shortcoming
 972  * of the full back refs is its overhead. Every time a tree block gets
 973  * COWed, we have to update back refs entry for all pointers in it.
 974  *
 975  * For a newly allocated tree block, we use implicit back refs for
 976  * pointers in it. This means most tree related operations only involve
 977  * implicit back refs. For a tree block created in old transaction, the
 978  * only way to drop a reference to it is COW it. So we can detect the
 979  * event that tree block loses its owner tree's reference and do the
 980  * back refs conversion.
 981  *
 982  * When a tree block is COWed through a tree, there are four cases:
 983  *
 984  * The reference count of the block is one and the tree is the block's
 985  * owner tree. Nothing to do in this case.
 986  *
 987  * The reference count of the block is one and the tree is not the
 988  * block's owner tree. In this case, full back refs is used for pointers
 989  * in the block. Remove these full back refs, add implicit back refs for
 990  * every pointers in the new block.
 991  *
 992  * The reference count of the block is greater than one and the tree is
 993  * the block's owner tree. In this case, implicit back refs is used for
 994  * pointers in the block. Add full back refs for every pointers in the
 995  * block, increase lower level extents' reference counts. The original
 996  * implicit back refs are entailed to the new block.
 997  *
 998  * The reference count of the block is greater than one and the tree is
 999  * not the block's owner tree. Add implicit back refs for every pointer in
1000  * the new block, increase lower level extents' reference count.
1001  *
1002  * Back Reference Key composing:
1003  *
1004  * The key objectid corresponds to the first byte in the extent,
1005  * The key type is used to differentiate between types of back refs.
1006  * There are different meanings of the key offset for different types
1007  * of back refs.
1008  *
1009  * File extents can be referenced by:
1010  *
1011  * - multiple snapshots, subvolumes, or different generations in one subvol
1012  * - different files inside a single subvolume
1013  * - different offsets inside a file (bookend extents in file.c)
1014  *
1015  * The extent ref structure for the implicit back refs has fields for:
1016  *
1017  * - Objectid of the subvolume root
1018  * - objectid of the file holding the reference
1019  * - original offset in the file
1020  * - how many bookend extents
1021  *
1022  * The key offset for the implicit back refs is hash of the first
1023  * three fields.
1024  *
1025  * The extent ref structure for the full back refs has field for:
1026  *
1027  * - number of pointers in the tree leaf
1028  *
1029  * The key offset for the implicit back refs is the first byte of
1030  * the tree leaf
1031  *
1032  * When a file extent is allocated, The implicit back refs is used.
1033  * the fields are filled in:
1034  *
1035  *     (root_key.objectid, inode objectid, offset in file, 1)
1036  *
1037  * When a file extent is removed file truncation, we find the
1038  * corresponding implicit back refs and check the following fields:
1039  *
1040  *     (btrfs_header_owner(leaf), inode objectid, offset in file)
1041  *
1042  * Btree extents can be referenced by:
1043  *
1044  * - Different subvolumes
1045  *
1046  * Both the implicit back refs and the full back refs for tree blocks
1047  * only consist of key. The key offset for the implicit back refs is
1048  * objectid of block's owner tree. The key offset for the full back refs
1049  * is the first byte of parent block.
1050  *
1051  * When implicit back refs is used, information about the lowest key and
1052  * level of the tree block are required. These information are stored in
1053  * tree block info structure.
1054  */
1055
1056 /*
1057  * is_data == BTRFS_REF_TYPE_BLOCK, tree block type is required,
1058  * is_data == BTRFS_REF_TYPE_DATA, data type is requiried,
1059  * is_data == BTRFS_REF_TYPE_ANY, either type is OK.
1060  */
1061 int btrfs_get_extent_inline_ref_type(const struct extent_buffer *eb,
1062                                      struct btrfs_extent_inline_ref *iref,
1063                                      enum btrfs_inline_ref_type is_data)
1064 {
1065         int type = btrfs_extent_inline_ref_type(eb, iref);
1066         u64 offset = btrfs_extent_inline_ref_offset(eb, iref);
1067
1068         if (type == BTRFS_TREE_BLOCK_REF_KEY ||
1069             type == BTRFS_SHARED_BLOCK_REF_KEY ||
1070             type == BTRFS_SHARED_DATA_REF_KEY ||
1071             type == BTRFS_EXTENT_DATA_REF_KEY) {
1072                 if (is_data == BTRFS_REF_TYPE_BLOCK) {
1073                         if (type == BTRFS_TREE_BLOCK_REF_KEY)
1074                                 return type;
1075                         if (type == BTRFS_SHARED_BLOCK_REF_KEY) {
1076                                 ASSERT(eb->fs_info);
1077                                 /*
1078                                  * Every shared one has parent tree
1079                                  * block, which must be aligned to
1080                                  * nodesize.
1081                                  */
1082                                 if (offset &&
1083                                     IS_ALIGNED(offset, eb->fs_info->nodesize))
1084                                         return type;
1085                         }
1086                 } else if (is_data == BTRFS_REF_TYPE_DATA) {
1087                         if (type == BTRFS_EXTENT_DATA_REF_KEY)
1088                                 return type;
1089                         if (type == BTRFS_SHARED_DATA_REF_KEY) {
1090                                 ASSERT(eb->fs_info);
1091                                 /*
1092                                  * Every shared one has parent tree
1093                                  * block, which must be aligned to
1094                                  * nodesize.
1095                                  */
1096                                 if (offset &&
1097                                     IS_ALIGNED(offset, eb->fs_info->nodesize))
1098                                         return type;
1099                         }
1100                 } else {
1101                         ASSERT(is_data == BTRFS_REF_TYPE_ANY);
1102                         return type;
1103                 }
1104         }
1105
1106         btrfs_print_leaf((struct extent_buffer *)eb);
1107         btrfs_err(eb->fs_info, "eb %llu invalid extent inline ref type %d",
1108                   eb->start, type);
1109         WARN_ON(1);
1110
1111         return BTRFS_REF_TYPE_INVALID;
1112 }
1113
1114 static u64 hash_extent_data_ref(u64 root_objectid, u64 owner, u64 offset)
1115 {
1116         u32 high_crc = ~(u32)0;
1117         u32 low_crc = ~(u32)0;
1118         __le64 lenum;
1119
1120         lenum = cpu_to_le64(root_objectid);
1121         high_crc = crc32c(high_crc, &lenum, sizeof(lenum));
1122         lenum = cpu_to_le64(owner);
1123         low_crc = crc32c(low_crc, &lenum, sizeof(lenum));
1124         lenum = cpu_to_le64(offset);
1125         low_crc = crc32c(low_crc, &lenum, sizeof(lenum));
1126
1127         return ((u64)high_crc << 31) ^ (u64)low_crc;
1128 }
1129
1130 static u64 hash_extent_data_ref_item(struct extent_buffer *leaf,
1131                                      struct btrfs_extent_data_ref *ref)
1132 {
1133         return hash_extent_data_ref(btrfs_extent_data_ref_root(leaf, ref),
1134                                     btrfs_extent_data_ref_objectid(leaf, ref),
1135                                     btrfs_extent_data_ref_offset(leaf, ref));
1136 }
1137
1138 static int match_extent_data_ref(struct extent_buffer *leaf,
1139                                  struct btrfs_extent_data_ref *ref,
1140                                  u64 root_objectid, u64 owner, u64 offset)
1141 {
1142         if (btrfs_extent_data_ref_root(leaf, ref) != root_objectid ||
1143             btrfs_extent_data_ref_objectid(leaf, ref) != owner ||
1144             btrfs_extent_data_ref_offset(leaf, ref) != offset)
1145                 return 0;
1146         return 1;
1147 }
1148
1149 static noinline int lookup_extent_data_ref(struct btrfs_trans_handle *trans,
1150                                            struct btrfs_path *path,
1151                                            u64 bytenr, u64 parent,
1152                                            u64 root_objectid,
1153                                            u64 owner, u64 offset)
1154 {
1155         struct btrfs_root *root = trans->fs_info->extent_root;
1156         struct btrfs_key key;
1157         struct btrfs_extent_data_ref *ref;
1158         struct extent_buffer *leaf;
1159         u32 nritems;
1160         int ret;
1161         int recow;
1162         int err = -ENOENT;
1163
1164         key.objectid = bytenr;
1165         if (parent) {
1166                 key.type = BTRFS_SHARED_DATA_REF_KEY;
1167                 key.offset = parent;
1168         } else {
1169                 key.type = BTRFS_EXTENT_DATA_REF_KEY;
1170                 key.offset = hash_extent_data_ref(root_objectid,
1171                                                   owner, offset);
1172         }
1173 again:
1174         recow = 0;
1175         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1176         if (ret < 0) {
1177                 err = ret;
1178                 goto fail;
1179         }
1180
1181         if (parent) {
1182                 if (!ret)
1183                         return 0;
1184                 goto fail;
1185         }
1186
1187         leaf = path->nodes[0];
1188         nritems = btrfs_header_nritems(leaf);
1189         while (1) {
1190                 if (path->slots[0] >= nritems) {
1191                         ret = btrfs_next_leaf(root, path);
1192                         if (ret < 0)
1193                                 err = ret;
1194                         if (ret)
1195                                 goto fail;
1196
1197                         leaf = path->nodes[0];
1198                         nritems = btrfs_header_nritems(leaf);
1199                         recow = 1;
1200                 }
1201
1202                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1203                 if (key.objectid != bytenr ||
1204                     key.type != BTRFS_EXTENT_DATA_REF_KEY)
1205                         goto fail;
1206
1207                 ref = btrfs_item_ptr(leaf, path->slots[0],
1208                                      struct btrfs_extent_data_ref);
1209
1210                 if (match_extent_data_ref(leaf, ref, root_objectid,
1211                                           owner, offset)) {
1212                         if (recow) {
1213                                 btrfs_release_path(path);
1214                                 goto again;
1215                         }
1216                         err = 0;
1217                         break;
1218                 }
1219                 path->slots[0]++;
1220         }
1221 fail:
1222         return err;
1223 }
1224
1225 static noinline int insert_extent_data_ref(struct btrfs_trans_handle *trans,
1226                                            struct btrfs_path *path,
1227                                            u64 bytenr, u64 parent,
1228                                            u64 root_objectid, u64 owner,
1229                                            u64 offset, int refs_to_add)
1230 {
1231         struct btrfs_root *root = trans->fs_info->extent_root;
1232         struct btrfs_key key;
1233         struct extent_buffer *leaf;
1234         u32 size;
1235         u32 num_refs;
1236         int ret;
1237
1238         key.objectid = bytenr;
1239         if (parent) {
1240                 key.type = BTRFS_SHARED_DATA_REF_KEY;
1241                 key.offset = parent;
1242                 size = sizeof(struct btrfs_shared_data_ref);
1243         } else {
1244                 key.type = BTRFS_EXTENT_DATA_REF_KEY;
1245                 key.offset = hash_extent_data_ref(root_objectid,
1246                                                   owner, offset);
1247                 size = sizeof(struct btrfs_extent_data_ref);
1248         }
1249
1250         ret = btrfs_insert_empty_item(trans, root, path, &key, size);
1251         if (ret && ret != -EEXIST)
1252                 goto fail;
1253
1254         leaf = path->nodes[0];
1255         if (parent) {
1256                 struct btrfs_shared_data_ref *ref;
1257                 ref = btrfs_item_ptr(leaf, path->slots[0],
1258                                      struct btrfs_shared_data_ref);
1259                 if (ret == 0) {
1260                         btrfs_set_shared_data_ref_count(leaf, ref, refs_to_add);
1261                 } else {
1262                         num_refs = btrfs_shared_data_ref_count(leaf, ref);
1263                         num_refs += refs_to_add;
1264                         btrfs_set_shared_data_ref_count(leaf, ref, num_refs);
1265                 }
1266         } else {
1267                 struct btrfs_extent_data_ref *ref;
1268                 while (ret == -EEXIST) {
1269                         ref = btrfs_item_ptr(leaf, path->slots[0],
1270                                              struct btrfs_extent_data_ref);
1271                         if (match_extent_data_ref(leaf, ref, root_objectid,
1272                                                   owner, offset))
1273                                 break;
1274                         btrfs_release_path(path);
1275                         key.offset++;
1276                         ret = btrfs_insert_empty_item(trans, root, path, &key,
1277                                                       size);
1278                         if (ret && ret != -EEXIST)
1279                                 goto fail;
1280
1281                         leaf = path->nodes[0];
1282                 }
1283                 ref = btrfs_item_ptr(leaf, path->slots[0],
1284                                      struct btrfs_extent_data_ref);
1285                 if (ret == 0) {
1286                         btrfs_set_extent_data_ref_root(leaf, ref,
1287                                                        root_objectid);
1288                         btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
1289                         btrfs_set_extent_data_ref_offset(leaf, ref, offset);
1290                         btrfs_set_extent_data_ref_count(leaf, ref, refs_to_add);
1291                 } else {
1292                         num_refs = btrfs_extent_data_ref_count(leaf, ref);
1293                         num_refs += refs_to_add;
1294                         btrfs_set_extent_data_ref_count(leaf, ref, num_refs);
1295                 }
1296         }
1297         btrfs_mark_buffer_dirty(leaf);
1298         ret = 0;
1299 fail:
1300         btrfs_release_path(path);
1301         return ret;
1302 }
1303
1304 static noinline int remove_extent_data_ref(struct btrfs_trans_handle *trans,
1305                                            struct btrfs_path *path,
1306                                            int refs_to_drop, int *last_ref)
1307 {
1308         struct btrfs_key key;
1309         struct btrfs_extent_data_ref *ref1 = NULL;
1310         struct btrfs_shared_data_ref *ref2 = NULL;
1311         struct extent_buffer *leaf;
1312         u32 num_refs = 0;
1313         int ret = 0;
1314
1315         leaf = path->nodes[0];
1316         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1317
1318         if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
1319                 ref1 = btrfs_item_ptr(leaf, path->slots[0],
1320                                       struct btrfs_extent_data_ref);
1321                 num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1322         } else if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
1323                 ref2 = btrfs_item_ptr(leaf, path->slots[0],
1324                                       struct btrfs_shared_data_ref);
1325                 num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1326         } else if (unlikely(key.type == BTRFS_EXTENT_REF_V0_KEY)) {
1327                 btrfs_print_v0_err(trans->fs_info);
1328                 btrfs_abort_transaction(trans, -EINVAL);
1329                 return -EINVAL;
1330         } else {
1331                 BUG();
1332         }
1333
1334         BUG_ON(num_refs < refs_to_drop);
1335         num_refs -= refs_to_drop;
1336
1337         if (num_refs == 0) {
1338                 ret = btrfs_del_item(trans, trans->fs_info->extent_root, path);
1339                 *last_ref = 1;
1340         } else {
1341                 if (key.type == BTRFS_EXTENT_DATA_REF_KEY)
1342                         btrfs_set_extent_data_ref_count(leaf, ref1, num_refs);
1343                 else if (key.type == BTRFS_SHARED_DATA_REF_KEY)
1344                         btrfs_set_shared_data_ref_count(leaf, ref2, num_refs);
1345                 btrfs_mark_buffer_dirty(leaf);
1346         }
1347         return ret;
1348 }
1349
1350 static noinline u32 extent_data_ref_count(struct btrfs_path *path,
1351                                           struct btrfs_extent_inline_ref *iref)
1352 {
1353         struct btrfs_key key;
1354         struct extent_buffer *leaf;
1355         struct btrfs_extent_data_ref *ref1;
1356         struct btrfs_shared_data_ref *ref2;
1357         u32 num_refs = 0;
1358         int type;
1359
1360         leaf = path->nodes[0];
1361         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1362
1363         BUG_ON(key.type == BTRFS_EXTENT_REF_V0_KEY);
1364         if (iref) {
1365                 /*
1366                  * If type is invalid, we should have bailed out earlier than
1367                  * this call.
1368                  */
1369                 type = btrfs_get_extent_inline_ref_type(leaf, iref, BTRFS_REF_TYPE_DATA);
1370                 ASSERT(type != BTRFS_REF_TYPE_INVALID);
1371                 if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1372                         ref1 = (struct btrfs_extent_data_ref *)(&iref->offset);
1373                         num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1374                 } else {
1375                         ref2 = (struct btrfs_shared_data_ref *)(iref + 1);
1376                         num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1377                 }
1378         } else if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
1379                 ref1 = btrfs_item_ptr(leaf, path->slots[0],
1380                                       struct btrfs_extent_data_ref);
1381                 num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1382         } else if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
1383                 ref2 = btrfs_item_ptr(leaf, path->slots[0],
1384                                       struct btrfs_shared_data_ref);
1385                 num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1386         } else {
1387                 WARN_ON(1);
1388         }
1389         return num_refs;
1390 }
1391
1392 static noinline int lookup_tree_block_ref(struct btrfs_trans_handle *trans,
1393                                           struct btrfs_path *path,
1394                                           u64 bytenr, u64 parent,
1395                                           u64 root_objectid)
1396 {
1397         struct btrfs_root *root = trans->fs_info->extent_root;
1398         struct btrfs_key key;
1399         int ret;
1400
1401         key.objectid = bytenr;
1402         if (parent) {
1403                 key.type = BTRFS_SHARED_BLOCK_REF_KEY;
1404                 key.offset = parent;
1405         } else {
1406                 key.type = BTRFS_TREE_BLOCK_REF_KEY;
1407                 key.offset = root_objectid;
1408         }
1409
1410         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1411         if (ret > 0)
1412                 ret = -ENOENT;
1413         return ret;
1414 }
1415
1416 static noinline int insert_tree_block_ref(struct btrfs_trans_handle *trans,
1417                                           struct btrfs_path *path,
1418                                           u64 bytenr, u64 parent,
1419                                           u64 root_objectid)
1420 {
1421         struct btrfs_key key;
1422         int ret;
1423
1424         key.objectid = bytenr;
1425         if (parent) {
1426                 key.type = BTRFS_SHARED_BLOCK_REF_KEY;
1427                 key.offset = parent;
1428         } else {
1429                 key.type = BTRFS_TREE_BLOCK_REF_KEY;
1430                 key.offset = root_objectid;
1431         }
1432
1433         ret = btrfs_insert_empty_item(trans, trans->fs_info->extent_root,
1434                                       path, &key, 0);
1435         btrfs_release_path(path);
1436         return ret;
1437 }
1438
1439 static inline int extent_ref_type(u64 parent, u64 owner)
1440 {
1441         int type;
1442         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1443                 if (parent > 0)
1444                         type = BTRFS_SHARED_BLOCK_REF_KEY;
1445                 else
1446                         type = BTRFS_TREE_BLOCK_REF_KEY;
1447         } else {
1448                 if (parent > 0)
1449                         type = BTRFS_SHARED_DATA_REF_KEY;
1450                 else
1451                         type = BTRFS_EXTENT_DATA_REF_KEY;
1452         }
1453         return type;
1454 }
1455
1456 static int find_next_key(struct btrfs_path *path, int level,
1457                          struct btrfs_key *key)
1458
1459 {
1460         for (; level < BTRFS_MAX_LEVEL; level++) {
1461                 if (!path->nodes[level])
1462                         break;
1463                 if (path->slots[level] + 1 >=
1464                     btrfs_header_nritems(path->nodes[level]))
1465                         continue;
1466                 if (level == 0)
1467                         btrfs_item_key_to_cpu(path->nodes[level], key,
1468                                               path->slots[level] + 1);
1469                 else
1470                         btrfs_node_key_to_cpu(path->nodes[level], key,
1471                                               path->slots[level] + 1);
1472                 return 0;
1473         }
1474         return 1;
1475 }
1476
1477 /*
1478  * look for inline back ref. if back ref is found, *ref_ret is set
1479  * to the address of inline back ref, and 0 is returned.
1480  *
1481  * if back ref isn't found, *ref_ret is set to the address where it
1482  * should be inserted, and -ENOENT is returned.
1483  *
1484  * if insert is true and there are too many inline back refs, the path
1485  * points to the extent item, and -EAGAIN is returned.
1486  *
1487  * NOTE: inline back refs are ordered in the same way that back ref
1488  *       items in the tree are ordered.
1489  */
1490 static noinline_for_stack
1491 int lookup_inline_extent_backref(struct btrfs_trans_handle *trans,
1492                                  struct btrfs_path *path,
1493                                  struct btrfs_extent_inline_ref **ref_ret,
1494                                  u64 bytenr, u64 num_bytes,
1495                                  u64 parent, u64 root_objectid,
1496                                  u64 owner, u64 offset, int insert)
1497 {
1498         struct btrfs_fs_info *fs_info = trans->fs_info;
1499         struct btrfs_root *root = fs_info->extent_root;
1500         struct btrfs_key key;
1501         struct extent_buffer *leaf;
1502         struct btrfs_extent_item *ei;
1503         struct btrfs_extent_inline_ref *iref;
1504         u64 flags;
1505         u64 item_size;
1506         unsigned long ptr;
1507         unsigned long end;
1508         int extra_size;
1509         int type;
1510         int want;
1511         int ret;
1512         int err = 0;
1513         bool skinny_metadata = btrfs_fs_incompat(fs_info, SKINNY_METADATA);
1514         int needed;
1515
1516         key.objectid = bytenr;
1517         key.type = BTRFS_EXTENT_ITEM_KEY;
1518         key.offset = num_bytes;
1519
1520         want = extent_ref_type(parent, owner);
1521         if (insert) {
1522                 extra_size = btrfs_extent_inline_ref_size(want);
1523                 path->keep_locks = 1;
1524         } else
1525                 extra_size = -1;
1526
1527         /*
1528          * Owner is our level, so we can just add one to get the level for the
1529          * block we are interested in.
1530          */
1531         if (skinny_metadata && owner < BTRFS_FIRST_FREE_OBJECTID) {
1532                 key.type = BTRFS_METADATA_ITEM_KEY;
1533                 key.offset = owner;
1534         }
1535
1536 again:
1537         ret = btrfs_search_slot(trans, root, &key, path, extra_size, 1);
1538         if (ret < 0) {
1539                 err = ret;
1540                 goto out;
1541         }
1542
1543         /*
1544          * We may be a newly converted file system which still has the old fat
1545          * extent entries for metadata, so try and see if we have one of those.
1546          */
1547         if (ret > 0 && skinny_metadata) {
1548                 skinny_metadata = false;
1549                 if (path->slots[0]) {
1550                         path->slots[0]--;
1551                         btrfs_item_key_to_cpu(path->nodes[0], &key,
1552                                               path->slots[0]);
1553                         if (key.objectid == bytenr &&
1554                             key.type == BTRFS_EXTENT_ITEM_KEY &&
1555                             key.offset == num_bytes)
1556                                 ret = 0;
1557                 }
1558                 if (ret) {
1559                         key.objectid = bytenr;
1560                         key.type = BTRFS_EXTENT_ITEM_KEY;
1561                         key.offset = num_bytes;
1562                         btrfs_release_path(path);
1563                         goto again;
1564                 }
1565         }
1566
1567         if (ret && !insert) {
1568                 err = -ENOENT;
1569                 goto out;
1570         } else if (WARN_ON(ret)) {
1571                 err = -EIO;
1572                 goto out;
1573         }
1574
1575         leaf = path->nodes[0];
1576         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1577         if (unlikely(item_size < sizeof(*ei))) {
1578                 err = -EINVAL;
1579                 btrfs_print_v0_err(fs_info);
1580                 btrfs_abort_transaction(trans, err);
1581                 goto out;
1582         }
1583
1584         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1585         flags = btrfs_extent_flags(leaf, ei);
1586
1587         ptr = (unsigned long)(ei + 1);
1588         end = (unsigned long)ei + item_size;
1589
1590         if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK && !skinny_metadata) {
1591                 ptr += sizeof(struct btrfs_tree_block_info);
1592                 BUG_ON(ptr > end);
1593         }
1594
1595         if (owner >= BTRFS_FIRST_FREE_OBJECTID)
1596                 needed = BTRFS_REF_TYPE_DATA;
1597         else
1598                 needed = BTRFS_REF_TYPE_BLOCK;
1599
1600         err = -ENOENT;
1601         while (1) {
1602                 if (ptr >= end) {
1603                         WARN_ON(ptr > end);
1604                         break;
1605                 }
1606                 iref = (struct btrfs_extent_inline_ref *)ptr;
1607                 type = btrfs_get_extent_inline_ref_type(leaf, iref, needed);
1608                 if (type == BTRFS_REF_TYPE_INVALID) {
1609                         err = -EUCLEAN;
1610                         goto out;
1611                 }
1612
1613                 if (want < type)
1614                         break;
1615                 if (want > type) {
1616                         ptr += btrfs_extent_inline_ref_size(type);
1617                         continue;
1618                 }
1619
1620                 if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1621                         struct btrfs_extent_data_ref *dref;
1622                         dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1623                         if (match_extent_data_ref(leaf, dref, root_objectid,
1624                                                   owner, offset)) {
1625                                 err = 0;
1626                                 break;
1627                         }
1628                         if (hash_extent_data_ref_item(leaf, dref) <
1629                             hash_extent_data_ref(root_objectid, owner, offset))
1630                                 break;
1631                 } else {
1632                         u64 ref_offset;
1633                         ref_offset = btrfs_extent_inline_ref_offset(leaf, iref);
1634                         if (parent > 0) {
1635                                 if (parent == ref_offset) {
1636                                         err = 0;
1637                                         break;
1638                                 }
1639                                 if (ref_offset < parent)
1640                                         break;
1641                         } else {
1642                                 if (root_objectid == ref_offset) {
1643                                         err = 0;
1644                                         break;
1645                                 }
1646                                 if (ref_offset < root_objectid)
1647                                         break;
1648                         }
1649                 }
1650                 ptr += btrfs_extent_inline_ref_size(type);
1651         }
1652         if (err == -ENOENT && insert) {
1653                 if (item_size + extra_size >=
1654                     BTRFS_MAX_EXTENT_ITEM_SIZE(root)) {
1655                         err = -EAGAIN;
1656                         goto out;
1657                 }
1658                 /*
1659                  * To add new inline back ref, we have to make sure
1660                  * there is no corresponding back ref item.
1661                  * For simplicity, we just do not add new inline back
1662                  * ref if there is any kind of item for this block
1663                  */
1664                 if (find_next_key(path, 0, &key) == 0 &&
1665                     key.objectid == bytenr &&
1666                     key.type < BTRFS_BLOCK_GROUP_ITEM_KEY) {
1667                         err = -EAGAIN;
1668                         goto out;
1669                 }
1670         }
1671         *ref_ret = (struct btrfs_extent_inline_ref *)ptr;
1672 out:
1673         if (insert) {
1674                 path->keep_locks = 0;
1675                 btrfs_unlock_up_safe(path, 1);
1676         }
1677         return err;
1678 }
1679
1680 /*
1681  * helper to add new inline back ref
1682  */
1683 static noinline_for_stack
1684 void setup_inline_extent_backref(struct btrfs_fs_info *fs_info,
1685                                  struct btrfs_path *path,
1686                                  struct btrfs_extent_inline_ref *iref,
1687                                  u64 parent, u64 root_objectid,
1688                                  u64 owner, u64 offset, int refs_to_add,
1689                                  struct btrfs_delayed_extent_op *extent_op)
1690 {
1691         struct extent_buffer *leaf;
1692         struct btrfs_extent_item *ei;
1693         unsigned long ptr;
1694         unsigned long end;
1695         unsigned long item_offset;
1696         u64 refs;
1697         int size;
1698         int type;
1699
1700         leaf = path->nodes[0];
1701         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1702         item_offset = (unsigned long)iref - (unsigned long)ei;
1703
1704         type = extent_ref_type(parent, owner);
1705         size = btrfs_extent_inline_ref_size(type);
1706
1707         btrfs_extend_item(fs_info, path, size);
1708
1709         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1710         refs = btrfs_extent_refs(leaf, ei);
1711         refs += refs_to_add;
1712         btrfs_set_extent_refs(leaf, ei, refs);
1713         if (extent_op)
1714                 __run_delayed_extent_op(extent_op, leaf, ei);
1715
1716         ptr = (unsigned long)ei + item_offset;
1717         end = (unsigned long)ei + btrfs_item_size_nr(leaf, path->slots[0]);
1718         if (ptr < end - size)
1719                 memmove_extent_buffer(leaf, ptr + size, ptr,
1720                                       end - size - ptr);
1721
1722         iref = (struct btrfs_extent_inline_ref *)ptr;
1723         btrfs_set_extent_inline_ref_type(leaf, iref, type);
1724         if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1725                 struct btrfs_extent_data_ref *dref;
1726                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1727                 btrfs_set_extent_data_ref_root(leaf, dref, root_objectid);
1728                 btrfs_set_extent_data_ref_objectid(leaf, dref, owner);
1729                 btrfs_set_extent_data_ref_offset(leaf, dref, offset);
1730                 btrfs_set_extent_data_ref_count(leaf, dref, refs_to_add);
1731         } else if (type == BTRFS_SHARED_DATA_REF_KEY) {
1732                 struct btrfs_shared_data_ref *sref;
1733                 sref = (struct btrfs_shared_data_ref *)(iref + 1);
1734                 btrfs_set_shared_data_ref_count(leaf, sref, refs_to_add);
1735                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
1736         } else if (type == BTRFS_SHARED_BLOCK_REF_KEY) {
1737                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
1738         } else {
1739                 btrfs_set_extent_inline_ref_offset(leaf, iref, root_objectid);
1740         }
1741         btrfs_mark_buffer_dirty(leaf);
1742 }
1743
1744 static int lookup_extent_backref(struct btrfs_trans_handle *trans,
1745                                  struct btrfs_path *path,
1746                                  struct btrfs_extent_inline_ref **ref_ret,
1747                                  u64 bytenr, u64 num_bytes, u64 parent,
1748                                  u64 root_objectid, u64 owner, u64 offset)
1749 {
1750         int ret;
1751
1752         ret = lookup_inline_extent_backref(trans, path, ref_ret, bytenr,
1753                                            num_bytes, parent, root_objectid,
1754                                            owner, offset, 0);
1755         if (ret != -ENOENT)
1756                 return ret;
1757
1758         btrfs_release_path(path);
1759         *ref_ret = NULL;
1760
1761         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1762                 ret = lookup_tree_block_ref(trans, path, bytenr, parent,
1763                                             root_objectid);
1764         } else {
1765                 ret = lookup_extent_data_ref(trans, path, bytenr, parent,
1766                                              root_objectid, owner, offset);
1767         }
1768         return ret;
1769 }
1770
1771 /*
1772  * helper to update/remove inline back ref
1773  */
1774 static noinline_for_stack
1775 void update_inline_extent_backref(struct btrfs_path *path,
1776                                   struct btrfs_extent_inline_ref *iref,
1777                                   int refs_to_mod,
1778                                   struct btrfs_delayed_extent_op *extent_op,
1779                                   int *last_ref)
1780 {
1781         struct extent_buffer *leaf = path->nodes[0];
1782         struct btrfs_fs_info *fs_info = leaf->fs_info;
1783         struct btrfs_extent_item *ei;
1784         struct btrfs_extent_data_ref *dref = NULL;
1785         struct btrfs_shared_data_ref *sref = NULL;
1786         unsigned long ptr;
1787         unsigned long end;
1788         u32 item_size;
1789         int size;
1790         int type;
1791         u64 refs;
1792
1793         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1794         refs = btrfs_extent_refs(leaf, ei);
1795         WARN_ON(refs_to_mod < 0 && refs + refs_to_mod <= 0);
1796         refs += refs_to_mod;
1797         btrfs_set_extent_refs(leaf, ei, refs);
1798         if (extent_op)
1799                 __run_delayed_extent_op(extent_op, leaf, ei);
1800
1801         /*
1802          * If type is invalid, we should have bailed out after
1803          * lookup_inline_extent_backref().
1804          */
1805         type = btrfs_get_extent_inline_ref_type(leaf, iref, BTRFS_REF_TYPE_ANY);
1806         ASSERT(type != BTRFS_REF_TYPE_INVALID);
1807
1808         if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1809                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1810                 refs = btrfs_extent_data_ref_count(leaf, dref);
1811         } else if (type == BTRFS_SHARED_DATA_REF_KEY) {
1812                 sref = (struct btrfs_shared_data_ref *)(iref + 1);
1813                 refs = btrfs_shared_data_ref_count(leaf, sref);
1814         } else {
1815                 refs = 1;
1816                 BUG_ON(refs_to_mod != -1);
1817         }
1818
1819         BUG_ON(refs_to_mod < 0 && refs < -refs_to_mod);
1820         refs += refs_to_mod;
1821
1822         if (refs > 0) {
1823                 if (type == BTRFS_EXTENT_DATA_REF_KEY)
1824                         btrfs_set_extent_data_ref_count(leaf, dref, refs);
1825                 else
1826                         btrfs_set_shared_data_ref_count(leaf, sref, refs);
1827         } else {
1828                 *last_ref = 1;
1829                 size =  btrfs_extent_inline_ref_size(type);
1830                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1831                 ptr = (unsigned long)iref;
1832                 end = (unsigned long)ei + item_size;
1833                 if (ptr + size < end)
1834                         memmove_extent_buffer(leaf, ptr, ptr + size,
1835                                               end - ptr - size);
1836                 item_size -= size;
1837                 btrfs_truncate_item(fs_info, path, item_size, 1);
1838         }
1839         btrfs_mark_buffer_dirty(leaf);
1840 }
1841
1842 static noinline_for_stack
1843 int insert_inline_extent_backref(struct btrfs_trans_handle *trans,
1844                                  struct btrfs_path *path,
1845                                  u64 bytenr, u64 num_bytes, u64 parent,
1846                                  u64 root_objectid, u64 owner,
1847                                  u64 offset, int refs_to_add,
1848                                  struct btrfs_delayed_extent_op *extent_op)
1849 {
1850         struct btrfs_extent_inline_ref *iref;
1851         int ret;
1852
1853         ret = lookup_inline_extent_backref(trans, path, &iref, bytenr,
1854                                            num_bytes, parent, root_objectid,
1855                                            owner, offset, 1);
1856         if (ret == 0) {
1857                 BUG_ON(owner < BTRFS_FIRST_FREE_OBJECTID);
1858                 update_inline_extent_backref(path, iref, refs_to_add,
1859                                              extent_op, NULL);
1860         } else if (ret == -ENOENT) {
1861                 setup_inline_extent_backref(trans->fs_info, path, iref, parent,
1862                                             root_objectid, owner, offset,
1863                                             refs_to_add, extent_op);
1864                 ret = 0;
1865         }
1866         return ret;
1867 }
1868
1869 static int insert_extent_backref(struct btrfs_trans_handle *trans,
1870                                  struct btrfs_path *path,
1871                                  u64 bytenr, u64 parent, u64 root_objectid,
1872                                  u64 owner, u64 offset, int refs_to_add)
1873 {
1874         int ret;
1875         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1876                 BUG_ON(refs_to_add != 1);
1877                 ret = insert_tree_block_ref(trans, path, bytenr, parent,
1878                                             root_objectid);
1879         } else {
1880                 ret = insert_extent_data_ref(trans, path, bytenr, parent,
1881                                              root_objectid, owner, offset,
1882                                              refs_to_add);
1883         }
1884         return ret;
1885 }
1886
1887 static int remove_extent_backref(struct btrfs_trans_handle *trans,
1888                                  struct btrfs_path *path,
1889                                  struct btrfs_extent_inline_ref *iref,
1890                                  int refs_to_drop, int is_data, int *last_ref)
1891 {
1892         int ret = 0;
1893
1894         BUG_ON(!is_data && refs_to_drop != 1);
1895         if (iref) {
1896                 update_inline_extent_backref(path, iref, -refs_to_drop, NULL,
1897                                              last_ref);
1898         } else if (is_data) {
1899                 ret = remove_extent_data_ref(trans, path, refs_to_drop,
1900                                              last_ref);
1901         } else {
1902                 *last_ref = 1;
1903                 ret = btrfs_del_item(trans, trans->fs_info->extent_root, path);
1904         }
1905         return ret;
1906 }
1907
1908 #define in_range(b, first, len)        ((b) >= (first) && (b) < (first) + (len))
1909 static int btrfs_issue_discard(struct block_device *bdev, u64 start, u64 len,
1910                                u64 *discarded_bytes)
1911 {
1912         int j, ret = 0;
1913         u64 bytes_left, end;
1914         u64 aligned_start = ALIGN(start, 1 << 9);
1915
1916         if (WARN_ON(start != aligned_start)) {
1917                 len -= aligned_start - start;
1918                 len = round_down(len, 1 << 9);
1919                 start = aligned_start;
1920         }
1921
1922         *discarded_bytes = 0;
1923
1924         if (!len)
1925                 return 0;
1926
1927         end = start + len;
1928         bytes_left = len;
1929
1930         /* Skip any superblocks on this device. */
1931         for (j = 0; j < BTRFS_SUPER_MIRROR_MAX; j++) {
1932                 u64 sb_start = btrfs_sb_offset(j);
1933                 u64 sb_end = sb_start + BTRFS_SUPER_INFO_SIZE;
1934                 u64 size = sb_start - start;
1935
1936                 if (!in_range(sb_start, start, bytes_left) &&
1937                     !in_range(sb_end, start, bytes_left) &&
1938                     !in_range(start, sb_start, BTRFS_SUPER_INFO_SIZE))
1939                         continue;
1940
1941                 /*
1942                  * Superblock spans beginning of range.  Adjust start and
1943                  * try again.
1944                  */
1945                 if (sb_start <= start) {
1946                         start += sb_end - start;
1947                         if (start > end) {
1948                                 bytes_left = 0;
1949                                 break;
1950                         }
1951                         bytes_left = end - start;
1952                         continue;
1953                 }
1954
1955                 if (size) {
1956                         ret = blkdev_issue_discard(bdev, start >> 9, size >> 9,
1957                                                    GFP_NOFS, 0);
1958                         if (!ret)
1959                                 *discarded_bytes += size;
1960                         else if (ret != -EOPNOTSUPP)
1961                                 return ret;
1962                 }
1963
1964                 start = sb_end;
1965                 if (start > end) {
1966                         bytes_left = 0;
1967                         break;
1968                 }
1969                 bytes_left = end - start;
1970         }
1971
1972         if (bytes_left) {
1973                 ret = blkdev_issue_discard(bdev, start >> 9, bytes_left >> 9,
1974                                            GFP_NOFS, 0);
1975                 if (!ret)
1976                         *discarded_bytes += bytes_left;
1977         }
1978         return ret;
1979 }
1980
1981 int btrfs_discard_extent(struct btrfs_fs_info *fs_info, u64 bytenr,
1982                          u64 num_bytes, u64 *actual_bytes)
1983 {
1984         int ret;
1985         u64 discarded_bytes = 0;
1986         struct btrfs_bio *bbio = NULL;
1987
1988
1989         /*
1990          * Avoid races with device replace and make sure our bbio has devices
1991          * associated to its stripes that don't go away while we are discarding.
1992          */
1993         btrfs_bio_counter_inc_blocked(fs_info);
1994         /* Tell the block device(s) that the sectors can be discarded */
1995         ret = btrfs_map_block(fs_info, BTRFS_MAP_DISCARD, bytenr, &num_bytes,
1996                               &bbio, 0);
1997         /* Error condition is -ENOMEM */
1998         if (!ret) {
1999                 struct btrfs_bio_stripe *stripe = bbio->stripes;
2000                 int i;
2001
2002
2003                 for (i = 0; i < bbio->num_stripes; i++, stripe++) {
2004                         u64 bytes;
2005                         struct request_queue *req_q;
2006
2007                         if (!stripe->dev->bdev) {
2008                                 ASSERT(btrfs_test_opt(fs_info, DEGRADED));
2009                                 continue;
2010                         }
2011                         req_q = bdev_get_queue(stripe->dev->bdev);
2012                         if (!blk_queue_discard(req_q))
2013                                 continue;
2014
2015                         ret = btrfs_issue_discard(stripe->dev->bdev,
2016                                                   stripe->physical,
2017                                                   stripe->length,
2018                                                   &bytes);
2019                         if (!ret)
2020                                 discarded_bytes += bytes;
2021                         else if (ret != -EOPNOTSUPP)
2022                                 break; /* Logic errors or -ENOMEM, or -EIO but I don't know how that could happen JDM */
2023
2024                         /*
2025                          * Just in case we get back EOPNOTSUPP for some reason,
2026                          * just ignore the return value so we don't screw up
2027                          * people calling discard_extent.
2028                          */
2029                         ret = 0;
2030                 }
2031                 btrfs_put_bbio(bbio);
2032         }
2033         btrfs_bio_counter_dec(fs_info);
2034
2035         if (actual_bytes)
2036                 *actual_bytes = discarded_bytes;
2037
2038
2039         if (ret == -EOPNOTSUPP)
2040                 ret = 0;
2041         return ret;
2042 }
2043
2044 /* Can return -ENOMEM */
2045 int btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
2046                          struct btrfs_root *root,
2047                          u64 bytenr, u64 num_bytes, u64 parent,
2048                          u64 root_objectid, u64 owner, u64 offset)
2049 {
2050         struct btrfs_fs_info *fs_info = root->fs_info;
2051         int old_ref_mod, new_ref_mod;
2052         int ret;
2053
2054         BUG_ON(owner < BTRFS_FIRST_FREE_OBJECTID &&
2055                root_objectid == BTRFS_TREE_LOG_OBJECTID);
2056
2057         btrfs_ref_tree_mod(root, bytenr, num_bytes, parent, root_objectid,
2058                            owner, offset, BTRFS_ADD_DELAYED_REF);
2059
2060         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
2061                 ret = btrfs_add_delayed_tree_ref(trans, bytenr,
2062                                                  num_bytes, parent,
2063                                                  root_objectid, (int)owner,
2064                                                  BTRFS_ADD_DELAYED_REF, NULL,
2065                                                  &old_ref_mod, &new_ref_mod);
2066         } else {
2067                 ret = btrfs_add_delayed_data_ref(trans, bytenr,
2068                                                  num_bytes, parent,
2069                                                  root_objectid, owner, offset,
2070                                                  0, BTRFS_ADD_DELAYED_REF,
2071                                                  &old_ref_mod, &new_ref_mod);
2072         }
2073
2074         if (ret == 0 && old_ref_mod < 0 && new_ref_mod >= 0) {
2075                 bool metadata = owner < BTRFS_FIRST_FREE_OBJECTID;
2076
2077                 add_pinned_bytes(fs_info, -num_bytes, metadata, root_objectid);
2078         }
2079
2080         return ret;
2081 }
2082
2083 /*
2084  * __btrfs_inc_extent_ref - insert backreference for a given extent
2085  *
2086  * @trans:          Handle of transaction
2087  *
2088  * @node:           The delayed ref node used to get the bytenr/length for
2089  *                  extent whose references are incremented.
2090  *
2091  * @parent:         If this is a shared extent (BTRFS_SHARED_DATA_REF_KEY/
2092  *                  BTRFS_SHARED_BLOCK_REF_KEY) then it holds the logical
2093  *                  bytenr of the parent block. Since new extents are always
2094  *                  created with indirect references, this will only be the case
2095  *                  when relocating a shared extent. In that case, root_objectid
2096  *                  will be BTRFS_TREE_RELOC_OBJECTID. Otheriwse, parent must
2097  *                  be 0
2098  *
2099  * @root_objectid:  The id of the root where this modification has originated,
2100  *                  this can be either one of the well-known metadata trees or
2101  *                  the subvolume id which references this extent.
2102  *
2103  * @owner:          For data extents it is the inode number of the owning file.
2104  *                  For metadata extents this parameter holds the level in the
2105  *                  tree of the extent.
2106  *
2107  * @offset:         For metadata extents the offset is ignored and is currently
2108  *                  always passed as 0. For data extents it is the fileoffset
2109  *                  this extent belongs to.
2110  *
2111  * @refs_to_add     Number of references to add
2112  *
2113  * @extent_op       Pointer to a structure, holding information necessary when
2114  *                  updating a tree block's flags
2115  *
2116  */
2117 static int __btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
2118                                   struct btrfs_delayed_ref_node *node,
2119                                   u64 parent, u64 root_objectid,
2120                                   u64 owner, u64 offset, int refs_to_add,
2121                                   struct btrfs_delayed_extent_op *extent_op)
2122 {
2123         struct btrfs_path *path;
2124         struct extent_buffer *leaf;
2125         struct btrfs_extent_item *item;
2126         struct btrfs_key key;
2127         u64 bytenr = node->bytenr;
2128         u64 num_bytes = node->num_bytes;
2129         u64 refs;
2130         int ret;
2131
2132         path = btrfs_alloc_path();
2133         if (!path)
2134                 return -ENOMEM;
2135
2136         path->reada = READA_FORWARD;
2137         path->leave_spinning = 1;
2138         /* this will setup the path even if it fails to insert the back ref */
2139         ret = insert_inline_extent_backref(trans, path, bytenr, num_bytes,
2140                                            parent, root_objectid, owner,
2141                                            offset, refs_to_add, extent_op);
2142         if ((ret < 0 && ret != -EAGAIN) || !ret)
2143                 goto out;
2144
2145         /*
2146          * Ok we had -EAGAIN which means we didn't have space to insert and
2147          * inline extent ref, so just update the reference count and add a
2148          * normal backref.
2149          */
2150         leaf = path->nodes[0];
2151         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
2152         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2153         refs = btrfs_extent_refs(leaf, item);
2154         btrfs_set_extent_refs(leaf, item, refs + refs_to_add);
2155         if (extent_op)
2156                 __run_delayed_extent_op(extent_op, leaf, item);
2157
2158         btrfs_mark_buffer_dirty(leaf);
2159         btrfs_release_path(path);
2160
2161         path->reada = READA_FORWARD;
2162         path->leave_spinning = 1;
2163         /* now insert the actual backref */
2164         ret = insert_extent_backref(trans, path, bytenr, parent, root_objectid,
2165                                     owner, offset, refs_to_add);
2166         if (ret)
2167                 btrfs_abort_transaction(trans, ret);
2168 out:
2169         btrfs_free_path(path);
2170         return ret;
2171 }
2172
2173 static int run_delayed_data_ref(struct btrfs_trans_handle *trans,
2174                                 struct btrfs_delayed_ref_node *node,
2175                                 struct btrfs_delayed_extent_op *extent_op,
2176                                 int insert_reserved)
2177 {
2178         int ret = 0;
2179         struct btrfs_delayed_data_ref *ref;
2180         struct btrfs_key ins;
2181         u64 parent = 0;
2182         u64 ref_root = 0;
2183         u64 flags = 0;
2184
2185         ins.objectid = node->bytenr;
2186         ins.offset = node->num_bytes;
2187         ins.type = BTRFS_EXTENT_ITEM_KEY;
2188
2189         ref = btrfs_delayed_node_to_data_ref(node);
2190         trace_run_delayed_data_ref(trans->fs_info, node, ref, node->action);
2191
2192         if (node->type == BTRFS_SHARED_DATA_REF_KEY)
2193                 parent = ref->parent;
2194         ref_root = ref->root;
2195
2196         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
2197                 if (extent_op)
2198                         flags |= extent_op->flags_to_set;
2199                 ret = alloc_reserved_file_extent(trans, parent, ref_root,
2200                                                  flags, ref->objectid,
2201                                                  ref->offset, &ins,
2202                                                  node->ref_mod);
2203         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
2204                 ret = __btrfs_inc_extent_ref(trans, node, parent, ref_root,
2205                                              ref->objectid, ref->offset,
2206                                              node->ref_mod, extent_op);
2207         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
2208                 ret = __btrfs_free_extent(trans, node, parent,
2209                                           ref_root, ref->objectid,
2210                                           ref->offset, node->ref_mod,
2211                                           extent_op);
2212         } else {
2213                 BUG();
2214         }
2215         return ret;
2216 }
2217
2218 static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op,
2219                                     struct extent_buffer *leaf,
2220                                     struct btrfs_extent_item *ei)
2221 {
2222         u64 flags = btrfs_extent_flags(leaf, ei);
2223         if (extent_op->update_flags) {
2224                 flags |= extent_op->flags_to_set;
2225                 btrfs_set_extent_flags(leaf, ei, flags);
2226         }
2227
2228         if (extent_op->update_key) {
2229                 struct btrfs_tree_block_info *bi;
2230                 BUG_ON(!(flags & BTRFS_EXTENT_FLAG_TREE_BLOCK));
2231                 bi = (struct btrfs_tree_block_info *)(ei + 1);
2232                 btrfs_set_tree_block_key(leaf, bi, &extent_op->key);
2233         }
2234 }
2235
2236 static int run_delayed_extent_op(struct btrfs_trans_handle *trans,
2237                                  struct btrfs_delayed_ref_head *head,
2238                                  struct btrfs_delayed_extent_op *extent_op)
2239 {
2240         struct btrfs_fs_info *fs_info = trans->fs_info;
2241         struct btrfs_key key;
2242         struct btrfs_path *path;
2243         struct btrfs_extent_item *ei;
2244         struct extent_buffer *leaf;
2245         u32 item_size;
2246         int ret;
2247         int err = 0;
2248         int metadata = !extent_op->is_data;
2249
2250         if (trans->aborted)
2251                 return 0;
2252
2253         if (metadata && !btrfs_fs_incompat(fs_info, SKINNY_METADATA))
2254                 metadata = 0;
2255
2256         path = btrfs_alloc_path();
2257         if (!path)
2258                 return -ENOMEM;
2259
2260         key.objectid = head->bytenr;
2261
2262         if (metadata) {
2263                 key.type = BTRFS_METADATA_ITEM_KEY;
2264                 key.offset = extent_op->level;
2265         } else {
2266                 key.type = BTRFS_EXTENT_ITEM_KEY;
2267                 key.offset = head->num_bytes;
2268         }
2269
2270 again:
2271         path->reada = READA_FORWARD;
2272         path->leave_spinning = 1;
2273         ret = btrfs_search_slot(trans, fs_info->extent_root, &key, path, 0, 1);
2274         if (ret < 0) {
2275                 err = ret;
2276                 goto out;
2277         }
2278         if (ret > 0) {
2279                 if (metadata) {
2280                         if (path->slots[0] > 0) {
2281                                 path->slots[0]--;
2282                                 btrfs_item_key_to_cpu(path->nodes[0], &key,
2283                                                       path->slots[0]);
2284                                 if (key.objectid == head->bytenr &&
2285                                     key.type == BTRFS_EXTENT_ITEM_KEY &&
2286                                     key.offset == head->num_bytes)
2287                                         ret = 0;
2288                         }
2289                         if (ret > 0) {
2290                                 btrfs_release_path(path);
2291                                 metadata = 0;
2292
2293                                 key.objectid = head->bytenr;
2294                                 key.offset = head->num_bytes;
2295                                 key.type = BTRFS_EXTENT_ITEM_KEY;
2296                                 goto again;
2297                         }
2298                 } else {
2299                         err = -EIO;
2300                         goto out;
2301                 }
2302         }
2303
2304         leaf = path->nodes[0];
2305         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2306
2307         if (unlikely(item_size < sizeof(*ei))) {
2308                 err = -EINVAL;
2309                 btrfs_print_v0_err(fs_info);
2310                 btrfs_abort_transaction(trans, err);
2311                 goto out;
2312         }
2313
2314         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2315         __run_delayed_extent_op(extent_op, leaf, ei);
2316
2317         btrfs_mark_buffer_dirty(leaf);
2318 out:
2319         btrfs_free_path(path);
2320         return err;
2321 }
2322
2323 static int run_delayed_tree_ref(struct btrfs_trans_handle *trans,
2324                                 struct btrfs_delayed_ref_node *node,
2325                                 struct btrfs_delayed_extent_op *extent_op,
2326                                 int insert_reserved)
2327 {
2328         int ret = 0;
2329         struct btrfs_delayed_tree_ref *ref;
2330         u64 parent = 0;
2331         u64 ref_root = 0;
2332
2333         ref = btrfs_delayed_node_to_tree_ref(node);
2334         trace_run_delayed_tree_ref(trans->fs_info, node, ref, node->action);
2335
2336         if (node->type == BTRFS_SHARED_BLOCK_REF_KEY)
2337                 parent = ref->parent;
2338         ref_root = ref->root;
2339
2340         if (node->ref_mod != 1) {
2341                 btrfs_err(trans->fs_info,
2342         "btree block(%llu) has %d references rather than 1: action %d ref_root %llu parent %llu",
2343                           node->bytenr, node->ref_mod, node->action, ref_root,
2344                           parent);
2345                 return -EIO;
2346         }
2347         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
2348                 BUG_ON(!extent_op || !extent_op->update_flags);
2349                 ret = alloc_reserved_tree_block(trans, node, extent_op);
2350         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
2351                 ret = __btrfs_inc_extent_ref(trans, node, parent, ref_root,
2352                                              ref->level, 0, 1, extent_op);
2353         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
2354                 ret = __btrfs_free_extent(trans, node, parent, ref_root,
2355                                           ref->level, 0, 1, extent_op);
2356         } else {
2357                 BUG();
2358         }
2359         return ret;
2360 }
2361
2362 /* helper function to actually process a single delayed ref entry */
2363 static int run_one_delayed_ref(struct btrfs_trans_handle *trans,
2364                                struct btrfs_delayed_ref_node *node,
2365                                struct btrfs_delayed_extent_op *extent_op,
2366                                int insert_reserved)
2367 {
2368         int ret = 0;
2369
2370         if (trans->aborted) {
2371                 if (insert_reserved)
2372                         btrfs_pin_extent(trans->fs_info, node->bytenr,
2373                                          node->num_bytes, 1);
2374                 return 0;
2375         }
2376
2377         if (node->type == BTRFS_TREE_BLOCK_REF_KEY ||
2378             node->type == BTRFS_SHARED_BLOCK_REF_KEY)
2379                 ret = run_delayed_tree_ref(trans, node, extent_op,
2380                                            insert_reserved);
2381         else if (node->type == BTRFS_EXTENT_DATA_REF_KEY ||
2382                  node->type == BTRFS_SHARED_DATA_REF_KEY)
2383                 ret = run_delayed_data_ref(trans, node, extent_op,
2384                                            insert_reserved);
2385         else
2386                 BUG();
2387         if (ret && insert_reserved)
2388                 btrfs_pin_extent(trans->fs_info, node->bytenr,
2389                                  node->num_bytes, 1);
2390         return ret;
2391 }
2392
2393 static inline struct btrfs_delayed_ref_node *
2394 select_delayed_ref(struct btrfs_delayed_ref_head *head)
2395 {
2396         struct btrfs_delayed_ref_node *ref;
2397
2398         if (RB_EMPTY_ROOT(&head->ref_tree.rb_root))
2399                 return NULL;
2400
2401         /*
2402          * Select a delayed ref of type BTRFS_ADD_DELAYED_REF first.
2403          * This is to prevent a ref count from going down to zero, which deletes
2404          * the extent item from the extent tree, when there still are references
2405          * to add, which would fail because they would not find the extent item.
2406          */
2407         if (!list_empty(&head->ref_add_list))
2408                 return list_first_entry(&head->ref_add_list,
2409                                 struct btrfs_delayed_ref_node, add_list);
2410
2411         ref = rb_entry(rb_first_cached(&head->ref_tree),
2412                        struct btrfs_delayed_ref_node, ref_node);
2413         ASSERT(list_empty(&ref->add_list));
2414         return ref;
2415 }
2416
2417 static void unselect_delayed_ref_head(struct btrfs_delayed_ref_root *delayed_refs,
2418                                       struct btrfs_delayed_ref_head *head)
2419 {
2420         spin_lock(&delayed_refs->lock);
2421         head->processing = 0;
2422         delayed_refs->num_heads_ready++;
2423         spin_unlock(&delayed_refs->lock);
2424         btrfs_delayed_ref_unlock(head);
2425 }
2426
2427 static struct btrfs_delayed_extent_op *cleanup_extent_op(
2428                                 struct btrfs_delayed_ref_head *head)
2429 {
2430         struct btrfs_delayed_extent_op *extent_op = head->extent_op;
2431
2432         if (!extent_op)
2433                 return NULL;
2434
2435         if (head->must_insert_reserved) {
2436                 head->extent_op = NULL;
2437                 btrfs_free_delayed_extent_op(extent_op);
2438                 return NULL;
2439         }
2440         return extent_op;
2441 }
2442
2443 static int run_and_cleanup_extent_op(struct btrfs_trans_handle *trans,
2444                                      struct btrfs_delayed_ref_head *head)
2445 {
2446         struct btrfs_delayed_extent_op *extent_op;
2447         int ret;
2448
2449         extent_op = cleanup_extent_op(head);
2450         if (!extent_op)
2451                 return 0;
2452         head->extent_op = NULL;
2453         spin_unlock(&head->lock);
2454         ret = run_delayed_extent_op(trans, head, extent_op);
2455         btrfs_free_delayed_extent_op(extent_op);
2456         return ret ? ret : 1;
2457 }
2458
2459 void btrfs_cleanup_ref_head_accounting(struct btrfs_fs_info *fs_info,
2460                                   struct btrfs_delayed_ref_root *delayed_refs,
2461                                   struct btrfs_delayed_ref_head *head)
2462 {
2463         int nr_items = 1;       /* Dropping this ref head update. */
2464
2465         if (head->total_ref_mod < 0) {
2466                 struct btrfs_space_info *space_info;
2467                 u64 flags;
2468
2469                 if (head->is_data)
2470                         flags = BTRFS_BLOCK_GROUP_DATA;
2471                 else if (head->is_system)
2472                         flags = BTRFS_BLOCK_GROUP_SYSTEM;
2473                 else
2474                         flags = BTRFS_BLOCK_GROUP_METADATA;
2475                 space_info = __find_space_info(fs_info, flags);
2476                 ASSERT(space_info);
2477                 percpu_counter_add_batch(&space_info->total_bytes_pinned,
2478                                    -head->num_bytes,
2479                                    BTRFS_TOTAL_BYTES_PINNED_BATCH);
2480
2481                 /*
2482                  * We had csum deletions accounted for in our delayed refs rsv,
2483                  * we need to drop the csum leaves for this update from our
2484                  * delayed_refs_rsv.
2485                  */
2486                 if (head->is_data) {
2487                         spin_lock(&delayed_refs->lock);
2488                         delayed_refs->pending_csums -= head->num_bytes;
2489                         spin_unlock(&delayed_refs->lock);
2490                         nr_items += btrfs_csum_bytes_to_leaves(fs_info,
2491                                 head->num_bytes);
2492                 }
2493         }
2494
2495         /* Also free its reserved qgroup space */
2496         btrfs_qgroup_free_delayed_ref(fs_info, head->qgroup_ref_root,
2497                                       head->qgroup_reserved);
2498         btrfs_delayed_refs_rsv_release(fs_info, nr_items);
2499 }
2500
2501 static int cleanup_ref_head(struct btrfs_trans_handle *trans,
2502                             struct btrfs_delayed_ref_head *head)
2503 {
2504
2505         struct btrfs_fs_info *fs_info = trans->fs_info;
2506         struct btrfs_delayed_ref_root *delayed_refs;
2507         int ret;
2508
2509         delayed_refs = &trans->transaction->delayed_refs;
2510
2511         ret = run_and_cleanup_extent_op(trans, head);
2512         if (ret < 0) {
2513                 unselect_delayed_ref_head(delayed_refs, head);
2514                 btrfs_debug(fs_info, "run_delayed_extent_op returned %d", ret);
2515                 return ret;
2516         } else if (ret) {
2517                 return ret;
2518         }
2519
2520         /*
2521          * Need to drop our head ref lock and re-acquire the delayed ref lock
2522          * and then re-check to make sure nobody got added.
2523          */
2524         spin_unlock(&head->lock);
2525         spin_lock(&delayed_refs->lock);
2526         spin_lock(&head->lock);
2527         if (!RB_EMPTY_ROOT(&head->ref_tree.rb_root) || head->extent_op) {
2528                 spin_unlock(&head->lock);
2529                 spin_unlock(&delayed_refs->lock);
2530                 return 1;
2531         }
2532         btrfs_delete_ref_head(delayed_refs, head);
2533         spin_unlock(&head->lock);
2534         spin_unlock(&delayed_refs->lock);
2535
2536         if (head->must_insert_reserved) {
2537                 btrfs_pin_extent(fs_info, head->bytenr,
2538                                  head->num_bytes, 1);
2539                 if (head->is_data) {
2540                         ret = btrfs_del_csums(trans, fs_info, head->bytenr,
2541                                               head->num_bytes);
2542                 }
2543         }
2544
2545         btrfs_cleanup_ref_head_accounting(fs_info, delayed_refs, head);
2546
2547         trace_run_delayed_ref_head(fs_info, head, 0);
2548         btrfs_delayed_ref_unlock(head);
2549         btrfs_put_delayed_ref_head(head);
2550         return 0;
2551 }
2552
2553 static struct btrfs_delayed_ref_head *btrfs_obtain_ref_head(
2554                                         struct btrfs_trans_handle *trans)
2555 {
2556         struct btrfs_delayed_ref_root *delayed_refs =
2557                 &trans->transaction->delayed_refs;
2558         struct btrfs_delayed_ref_head *head = NULL;
2559         int ret;
2560
2561         spin_lock(&delayed_refs->lock);
2562         head = btrfs_select_ref_head(delayed_refs);
2563         if (!head) {
2564                 spin_unlock(&delayed_refs->lock);
2565                 return head;
2566         }
2567
2568         /*
2569          * Grab the lock that says we are going to process all the refs for
2570          * this head
2571          */
2572         ret = btrfs_delayed_ref_lock(delayed_refs, head);
2573         spin_unlock(&delayed_refs->lock);
2574
2575         /*
2576          * We may have dropped the spin lock to get the head mutex lock, and
2577          * that might have given someone else time to free the head.  If that's
2578          * true, it has been removed from our list and we can move on.
2579          */
2580         if (ret == -EAGAIN)
2581                 head = ERR_PTR(-EAGAIN);
2582
2583         return head;
2584 }
2585
2586 static int btrfs_run_delayed_refs_for_head(struct btrfs_trans_handle *trans,
2587                                     struct btrfs_delayed_ref_head *locked_ref,
2588                                     unsigned long *run_refs)
2589 {
2590         struct btrfs_fs_info *fs_info = trans->fs_info;
2591         struct btrfs_delayed_ref_root *delayed_refs;
2592         struct btrfs_delayed_extent_op *extent_op;
2593         struct btrfs_delayed_ref_node *ref;
2594         int must_insert_reserved = 0;
2595         int ret;
2596
2597         delayed_refs = &trans->transaction->delayed_refs;
2598
2599         lockdep_assert_held(&locked_ref->mutex);
2600         lockdep_assert_held(&locked_ref->lock);
2601
2602         while ((ref = select_delayed_ref(locked_ref))) {
2603                 if (ref->seq &&
2604                     btrfs_check_delayed_seq(fs_info, ref->seq)) {
2605                         spin_unlock(&locked_ref->lock);
2606                         unselect_delayed_ref_head(delayed_refs, locked_ref);
2607                         return -EAGAIN;
2608                 }
2609
2610                 (*run_refs)++;
2611                 ref->in_tree = 0;
2612                 rb_erase_cached(&ref->ref_node, &locked_ref->ref_tree);
2613                 RB_CLEAR_NODE(&ref->ref_node);
2614                 if (!list_empty(&ref->add_list))
2615                         list_del(&ref->add_list);
2616                 /*
2617                  * When we play the delayed ref, also correct the ref_mod on
2618                  * head
2619                  */
2620                 switch (ref->action) {
2621                 case BTRFS_ADD_DELAYED_REF:
2622                 case BTRFS_ADD_DELAYED_EXTENT:
2623                         locked_ref->ref_mod -= ref->ref_mod;
2624                         break;
2625                 case BTRFS_DROP_DELAYED_REF:
2626                         locked_ref->ref_mod += ref->ref_mod;
2627                         break;
2628                 default:
2629                         WARN_ON(1);
2630                 }
2631                 atomic_dec(&delayed_refs->num_entries);
2632
2633                 /*
2634                  * Record the must_insert_reserved flag before we drop the
2635                  * spin lock.
2636                  */
2637                 must_insert_reserved = locked_ref->must_insert_reserved;
2638                 locked_ref->must_insert_reserved = 0;
2639
2640                 extent_op = locked_ref->extent_op;
2641                 locked_ref->extent_op = NULL;
2642                 spin_unlock(&locked_ref->lock);
2643
2644                 ret = run_one_delayed_ref(trans, ref, extent_op,
2645                                           must_insert_reserved);
2646
2647                 btrfs_free_delayed_extent_op(extent_op);
2648                 if (ret) {
2649                         unselect_delayed_ref_head(delayed_refs, locked_ref);
2650                         btrfs_put_delayed_ref(ref);
2651                         btrfs_debug(fs_info, "run_one_delayed_ref returned %d",
2652                                     ret);
2653                         return ret;
2654                 }
2655
2656                 btrfs_put_delayed_ref(ref);
2657                 cond_resched();
2658
2659                 spin_lock(&locked_ref->lock);
2660                 btrfs_merge_delayed_refs(trans, delayed_refs, locked_ref);
2661         }
2662
2663         return 0;
2664 }
2665
2666 /*
2667  * Returns 0 on success or if called with an already aborted transaction.
2668  * Returns -ENOMEM or -EIO on failure and will abort the transaction.
2669  */
2670 static noinline int __btrfs_run_delayed_refs(struct btrfs_trans_handle *trans,
2671                                              unsigned long nr)
2672 {
2673         struct btrfs_fs_info *fs_info = trans->fs_info;
2674         struct btrfs_delayed_ref_root *delayed_refs;
2675         struct btrfs_delayed_ref_head *locked_ref = NULL;
2676         ktime_t start = ktime_get();
2677         int ret;
2678         unsigned long count = 0;
2679         unsigned long actual_count = 0;
2680
2681         delayed_refs = &trans->transaction->delayed_refs;
2682         do {
2683                 if (!locked_ref) {
2684                         locked_ref = btrfs_obtain_ref_head(trans);
2685                         if (IS_ERR_OR_NULL(locked_ref)) {
2686                                 if (PTR_ERR(locked_ref) == -EAGAIN) {
2687                                         continue;
2688                                 } else {
2689                                         break;
2690                                 }
2691                         }
2692                         count++;
2693                 }
2694                 /*
2695                  * We need to try and merge add/drops of the same ref since we
2696                  * can run into issues with relocate dropping the implicit ref
2697                  * and then it being added back again before the drop can
2698                  * finish.  If we merged anything we need to re-loop so we can
2699                  * get a good ref.
2700                  * Or we can get node references of the same type that weren't
2701                  * merged when created due to bumps in the tree mod seq, and
2702                  * we need to merge them to prevent adding an inline extent
2703                  * backref before dropping it (triggering a BUG_ON at
2704                  * insert_inline_extent_backref()).
2705                  */
2706                 spin_lock(&locked_ref->lock);
2707                 btrfs_merge_delayed_refs(trans, delayed_refs, locked_ref);
2708
2709                 ret = btrfs_run_delayed_refs_for_head(trans, locked_ref,
2710                                                       &actual_count);
2711                 if (ret < 0 && ret != -EAGAIN) {
2712                         /*
2713                          * Error, btrfs_run_delayed_refs_for_head already
2714                          * unlocked everything so just bail out
2715                          */
2716                         return ret;
2717                 } else if (!ret) {
2718                         /*
2719                          * Success, perform the usual cleanup of a processed
2720                          * head
2721                          */
2722                         ret = cleanup_ref_head(trans, locked_ref);
2723                         if (ret > 0 ) {
2724                                 /* We dropped our lock, we need to loop. */
2725                                 ret = 0;
2726                                 continue;
2727                         } else if (ret) {
2728                                 return ret;
2729                         }
2730                 }
2731
2732                 /*
2733                  * Either success case or btrfs_run_delayed_refs_for_head
2734                  * returned -EAGAIN, meaning we need to select another head
2735                  */
2736
2737                 locked_ref = NULL;
2738                 cond_resched();
2739         } while ((nr != -1 && count < nr) || locked_ref);
2740
2741         /*
2742          * We don't want to include ref heads since we can have empty ref heads
2743          * and those will drastically skew our runtime down since we just do
2744          * accounting, no actual extent tree updates.
2745          */
2746         if (actual_count > 0) {
2747                 u64 runtime = ktime_to_ns(ktime_sub(ktime_get(), start));
2748                 u64 avg;
2749
2750                 /*
2751                  * We weigh the current average higher than our current runtime
2752                  * to avoid large swings in the average.
2753                  */
2754                 spin_lock(&delayed_refs->lock);
2755                 avg = fs_info->avg_delayed_ref_runtime * 3 + runtime;
2756                 fs_info->avg_delayed_ref_runtime = avg >> 2;    /* div by 4 */
2757                 spin_unlock(&delayed_refs->lock);
2758         }
2759         return 0;
2760 }
2761
2762 #ifdef SCRAMBLE_DELAYED_REFS
2763 /*
2764  * Normally delayed refs get processed in ascending bytenr order. This
2765  * correlates in most cases to the order added. To expose dependencies on this
2766  * order, we start to process the tree in the middle instead of the beginning
2767  */
2768 static u64 find_middle(struct rb_root *root)
2769 {
2770         struct rb_node *n = root->rb_node;
2771         struct btrfs_delayed_ref_node *entry;
2772         int alt = 1;
2773         u64 middle;
2774         u64 first = 0, last = 0;
2775
2776         n = rb_first(root);
2777         if (n) {
2778                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2779                 first = entry->bytenr;
2780         }
2781         n = rb_last(root);
2782         if (n) {
2783                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2784                 last = entry->bytenr;
2785         }
2786         n = root->rb_node;
2787
2788         while (n) {
2789                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2790                 WARN_ON(!entry->in_tree);
2791
2792                 middle = entry->bytenr;
2793
2794                 if (alt)
2795                         n = n->rb_left;
2796                 else
2797                         n = n->rb_right;
2798
2799                 alt = 1 - alt;
2800         }
2801         return middle;
2802 }
2803 #endif
2804
2805 static inline u64 heads_to_leaves(struct btrfs_fs_info *fs_info, u64 heads)
2806 {
2807         u64 num_bytes;
2808
2809         num_bytes = heads * (sizeof(struct btrfs_extent_item) +
2810                              sizeof(struct btrfs_extent_inline_ref));
2811         if (!btrfs_fs_incompat(fs_info, SKINNY_METADATA))
2812                 num_bytes += heads * sizeof(struct btrfs_tree_block_info);
2813
2814         /*
2815          * We don't ever fill up leaves all the way so multiply by 2 just to be
2816          * closer to what we're really going to want to use.
2817          */
2818         return div_u64(num_bytes, BTRFS_LEAF_DATA_SIZE(fs_info));
2819 }
2820
2821 /*
2822  * Takes the number of bytes to be csumm'ed and figures out how many leaves it
2823  * would require to store the csums for that many bytes.
2824  */
2825 u64 btrfs_csum_bytes_to_leaves(struct btrfs_fs_info *fs_info, u64 csum_bytes)
2826 {
2827         u64 csum_size;
2828         u64 num_csums_per_leaf;
2829         u64 num_csums;
2830
2831         csum_size = BTRFS_MAX_ITEM_SIZE(fs_info);
2832         num_csums_per_leaf = div64_u64(csum_size,
2833                         (u64)btrfs_super_csum_size(fs_info->super_copy));
2834         num_csums = div64_u64(csum_bytes, fs_info->sectorsize);
2835         num_csums += num_csums_per_leaf - 1;
2836         num_csums = div64_u64(num_csums, num_csums_per_leaf);
2837         return num_csums;
2838 }
2839
2840 bool btrfs_check_space_for_delayed_refs(struct btrfs_fs_info *fs_info)
2841 {
2842         struct btrfs_block_rsv *delayed_refs_rsv = &fs_info->delayed_refs_rsv;
2843         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
2844         bool ret = false;
2845         u64 reserved;
2846
2847         spin_lock(&global_rsv->lock);
2848         reserved = global_rsv->reserved;
2849         spin_unlock(&global_rsv->lock);
2850
2851         /*
2852          * Since the global reserve is just kind of magic we don't really want
2853          * to rely on it to save our bacon, so if our size is more than the
2854          * delayed_refs_rsv and the global rsv then it's time to think about
2855          * bailing.
2856          */
2857         spin_lock(&delayed_refs_rsv->lock);
2858         reserved += delayed_refs_rsv->reserved;
2859         if (delayed_refs_rsv->size >= reserved)
2860                 ret = true;
2861         spin_unlock(&delayed_refs_rsv->lock);
2862         return ret;
2863 }
2864
2865 int btrfs_should_throttle_delayed_refs(struct btrfs_trans_handle *trans)
2866 {
2867         u64 num_entries =
2868                 atomic_read(&trans->transaction->delayed_refs.num_entries);
2869         u64 avg_runtime;
2870         u64 val;
2871
2872         smp_mb();
2873         avg_runtime = trans->fs_info->avg_delayed_ref_runtime;
2874         val = num_entries * avg_runtime;
2875         if (val >= NSEC_PER_SEC)
2876                 return 1;
2877         if (val >= NSEC_PER_SEC / 2)
2878                 return 2;
2879
2880         return btrfs_check_space_for_delayed_refs(trans->fs_info);
2881 }
2882
2883 struct async_delayed_refs {
2884         struct btrfs_root *root;
2885         u64 transid;
2886         int count;
2887         int error;
2888         int sync;
2889         struct completion wait;
2890         struct btrfs_work work;
2891 };
2892
2893 static inline struct async_delayed_refs *
2894 to_async_delayed_refs(struct btrfs_work *work)
2895 {
2896         return container_of(work, struct async_delayed_refs, work);
2897 }
2898
2899 static void delayed_ref_async_start(struct btrfs_work *work)
2900 {
2901         struct async_delayed_refs *async = to_async_delayed_refs(work);
2902         struct btrfs_trans_handle *trans;
2903         struct btrfs_fs_info *fs_info = async->root->fs_info;
2904         int ret;
2905
2906         /* if the commit is already started, we don't need to wait here */
2907         if (btrfs_transaction_blocked(fs_info))
2908                 goto done;
2909
2910         trans = btrfs_join_transaction(async->root);
2911         if (IS_ERR(trans)) {
2912                 async->error = PTR_ERR(trans);
2913                 goto done;
2914         }
2915
2916         /*
2917          * trans->sync means that when we call end_transaction, we won't
2918          * wait on delayed refs
2919          */
2920         trans->sync = true;
2921
2922         /* Don't bother flushing if we got into a different transaction */
2923         if (trans->transid > async->transid)
2924                 goto end;
2925
2926         ret = btrfs_run_delayed_refs(trans, async->count);
2927         if (ret)
2928                 async->error = ret;
2929 end:
2930         ret = btrfs_end_transaction(trans);
2931         if (ret && !async->error)
2932                 async->error = ret;
2933 done:
2934         if (async->sync)
2935                 complete(&async->wait);
2936         else
2937                 kfree(async);
2938 }
2939
2940 int btrfs_async_run_delayed_refs(struct btrfs_fs_info *fs_info,
2941                                  unsigned long count, u64 transid, int wait)
2942 {
2943         struct async_delayed_refs *async;
2944         int ret;
2945
2946         async = kmalloc(sizeof(*async), GFP_NOFS);
2947         if (!async)
2948                 return -ENOMEM;
2949
2950         async->root = fs_info->tree_root;
2951         async->count = count;
2952         async->error = 0;
2953         async->transid = transid;
2954         if (wait)
2955                 async->sync = 1;
2956         else
2957                 async->sync = 0;
2958         init_completion(&async->wait);
2959
2960         btrfs_init_work(&async->work, btrfs_extent_refs_helper,
2961                         delayed_ref_async_start, NULL, NULL);
2962
2963         btrfs_queue_work(fs_info->extent_workers, &async->work);
2964
2965         if (wait) {
2966                 wait_for_completion(&async->wait);
2967                 ret = async->error;
2968                 kfree(async);
2969                 return ret;
2970         }
2971         return 0;
2972 }
2973
2974 /*
2975  * this starts processing the delayed reference count updates and
2976  * extent insertions we have queued up so far.  count can be
2977  * 0, which means to process everything in the tree at the start
2978  * of the run (but not newly added entries), or it can be some target
2979  * number you'd like to process.
2980  *
2981  * Returns 0 on success or if called with an aborted transaction
2982  * Returns <0 on error and aborts the transaction
2983  */
2984 int btrfs_run_delayed_refs(struct btrfs_trans_handle *trans,
2985                            unsigned long count)
2986 {
2987         struct btrfs_fs_info *fs_info = trans->fs_info;
2988         struct rb_node *node;
2989         struct btrfs_delayed_ref_root *delayed_refs;
2990         struct btrfs_delayed_ref_head *head;
2991         int ret;
2992         int run_all = count == (unsigned long)-1;
2993
2994         /* We'll clean this up in btrfs_cleanup_transaction */
2995         if (trans->aborted)
2996                 return 0;
2997
2998         if (test_bit(BTRFS_FS_CREATING_FREE_SPACE_TREE, &fs_info->flags))
2999                 return 0;
3000
3001         delayed_refs = &trans->transaction->delayed_refs;
3002         if (count == 0)
3003                 count = atomic_read(&delayed_refs->num_entries) * 2;
3004
3005 again:
3006 #ifdef SCRAMBLE_DELAYED_REFS
3007         delayed_refs->run_delayed_start = find_middle(&delayed_refs->root);
3008 #endif
3009         ret = __btrfs_run_delayed_refs(trans, count);
3010         if (ret < 0) {
3011                 btrfs_abort_transaction(trans, ret);
3012                 return ret;
3013         }
3014
3015         if (run_all) {
3016                 btrfs_create_pending_block_groups(trans);
3017
3018                 spin_lock(&delayed_refs->lock);
3019                 node = rb_first_cached(&delayed_refs->href_root);
3020                 if (!node) {
3021                         spin_unlock(&delayed_refs->lock);
3022                         goto out;
3023                 }
3024                 head = rb_entry(node, struct btrfs_delayed_ref_head,
3025                                 href_node);
3026                 refcount_inc(&head->refs);
3027                 spin_unlock(&delayed_refs->lock);
3028
3029                 /* Mutex was contended, block until it's released and retry. */
3030                 mutex_lock(&head->mutex);
3031                 mutex_unlock(&head->mutex);
3032
3033                 btrfs_put_delayed_ref_head(head);
3034                 cond_resched();
3035                 goto again;
3036         }
3037 out:
3038         return 0;
3039 }
3040
3041 int btrfs_set_disk_extent_flags(struct btrfs_trans_handle *trans,
3042                                 struct btrfs_fs_info *fs_info,
3043                                 u64 bytenr, u64 num_bytes, u64 flags,
3044                                 int level, int is_data)
3045 {
3046         struct btrfs_delayed_extent_op *extent_op;
3047         int ret;
3048
3049         extent_op = btrfs_alloc_delayed_extent_op();
3050         if (!extent_op)
3051                 return -ENOMEM;
3052
3053         extent_op->flags_to_set = flags;
3054         extent_op->update_flags = true;
3055         extent_op->update_key = false;
3056         extent_op->is_data = is_data ? true : false;
3057         extent_op->level = level;
3058
3059         ret = btrfs_add_delayed_extent_op(fs_info, trans, bytenr,
3060                                           num_bytes, extent_op);
3061         if (ret)
3062                 btrfs_free_delayed_extent_op(extent_op);
3063         return ret;
3064 }
3065
3066 static noinline int check_delayed_ref(struct btrfs_root *root,
3067                                       struct btrfs_path *path,
3068                                       u64 objectid, u64 offset, u64 bytenr)
3069 {
3070         struct btrfs_delayed_ref_head *head;
3071         struct btrfs_delayed_ref_node *ref;
3072         struct btrfs_delayed_data_ref *data_ref;
3073         struct btrfs_delayed_ref_root *delayed_refs;
3074         struct btrfs_transaction *cur_trans;
3075         struct rb_node *node;
3076         int ret = 0;
3077
3078         spin_lock(&root->fs_info->trans_lock);
3079         cur_trans = root->fs_info->running_transaction;
3080         if (cur_trans)
3081                 refcount_inc(&cur_trans->use_count);
3082         spin_unlock(&root->fs_info->trans_lock);
3083         if (!cur_trans)
3084                 return 0;
3085
3086         delayed_refs = &cur_trans->delayed_refs;
3087         spin_lock(&delayed_refs->lock);
3088         head = btrfs_find_delayed_ref_head(delayed_refs, bytenr);
3089         if (!head) {
3090                 spin_unlock(&delayed_refs->lock);
3091                 btrfs_put_transaction(cur_trans);
3092                 return 0;
3093         }
3094
3095         if (!mutex_trylock(&head->mutex)) {
3096                 refcount_inc(&head->refs);
3097                 spin_unlock(&delayed_refs->lock);
3098
3099                 btrfs_release_path(path);
3100
3101                 /*
3102                  * Mutex was contended, block until it's released and let
3103                  * caller try again
3104                  */
3105                 mutex_lock(&head->mutex);
3106                 mutex_unlock(&head->mutex);
3107                 btrfs_put_delayed_ref_head(head);
3108                 btrfs_put_transaction(cur_trans);
3109                 return -EAGAIN;
3110         }
3111         spin_unlock(&delayed_refs->lock);
3112
3113         spin_lock(&head->lock);
3114         /*
3115          * XXX: We should replace this with a proper search function in the
3116          * future.
3117          */
3118         for (node = rb_first_cached(&head->ref_tree); node;
3119              node = rb_next(node)) {
3120                 ref = rb_entry(node, struct btrfs_delayed_ref_node, ref_node);
3121                 /* If it's a shared ref we know a cross reference exists */
3122                 if (ref->type != BTRFS_EXTENT_DATA_REF_KEY) {
3123                         ret = 1;
3124                         break;
3125                 }
3126
3127                 data_ref = btrfs_delayed_node_to_data_ref(ref);
3128
3129                 /*
3130                  * If our ref doesn't match the one we're currently looking at
3131                  * then we have a cross reference.
3132                  */
3133                 if (data_ref->root != root->root_key.objectid ||
3134                     data_ref->objectid != objectid ||
3135                     data_ref->offset != offset) {
3136                         ret = 1;
3137                         break;
3138                 }
3139         }
3140         spin_unlock(&head->lock);
3141         mutex_unlock(&head->mutex);
3142         btrfs_put_transaction(cur_trans);
3143         return ret;
3144 }
3145
3146 static noinline int check_committed_ref(struct btrfs_root *root,
3147                                         struct btrfs_path *path,
3148                                         u64 objectid, u64 offset, u64 bytenr)
3149 {
3150         struct btrfs_fs_info *fs_info = root->fs_info;
3151         struct btrfs_root *extent_root = fs_info->extent_root;
3152         struct extent_buffer *leaf;
3153         struct btrfs_extent_data_ref *ref;
3154         struct btrfs_extent_inline_ref *iref;
3155         struct btrfs_extent_item *ei;
3156         struct btrfs_key key;
3157         u32 item_size;
3158         int type;
3159         int ret;
3160
3161         key.objectid = bytenr;
3162         key.offset = (u64)-1;
3163         key.type = BTRFS_EXTENT_ITEM_KEY;
3164
3165         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
3166         if (ret < 0)
3167                 goto out;
3168         BUG_ON(ret == 0); /* Corruption */
3169
3170         ret = -ENOENT;
3171         if (path->slots[0] == 0)
3172                 goto out;
3173
3174         path->slots[0]--;
3175         leaf = path->nodes[0];
3176         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
3177
3178         if (key.objectid != bytenr || key.type != BTRFS_EXTENT_ITEM_KEY)
3179                 goto out;
3180
3181         ret = 1;
3182         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
3183         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
3184
3185         if (item_size != sizeof(*ei) +
3186             btrfs_extent_inline_ref_size(BTRFS_EXTENT_DATA_REF_KEY))
3187                 goto out;
3188
3189         if (btrfs_extent_generation(leaf, ei) <=
3190             btrfs_root_last_snapshot(&root->root_item))
3191                 goto out;
3192
3193         iref = (struct btrfs_extent_inline_ref *)(ei + 1);
3194
3195         type = btrfs_get_extent_inline_ref_type(leaf, iref, BTRFS_REF_TYPE_DATA);
3196         if (type != BTRFS_EXTENT_DATA_REF_KEY)
3197                 goto out;
3198
3199         ref = (struct btrfs_extent_data_ref *)(&iref->offset);
3200         if (btrfs_extent_refs(leaf, ei) !=
3201             btrfs_extent_data_ref_count(leaf, ref) ||
3202             btrfs_extent_data_ref_root(leaf, ref) !=
3203             root->root_key.objectid ||
3204             btrfs_extent_data_ref_objectid(leaf, ref) != objectid ||
3205             btrfs_extent_data_ref_offset(leaf, ref) != offset)
3206                 goto out;
3207
3208         ret = 0;
3209 out:
3210         return ret;
3211 }
3212
3213 int btrfs_cross_ref_exist(struct btrfs_root *root, u64 objectid, u64 offset,
3214                           u64 bytenr)
3215 {
3216         struct btrfs_path *path;
3217         int ret;
3218
3219         path = btrfs_alloc_path();
3220         if (!path)
3221                 return -ENOMEM;
3222
3223         do {
3224                 ret = check_committed_ref(root, path, objectid,
3225                                           offset, bytenr);
3226                 if (ret && ret != -ENOENT)
3227                         goto out;
3228
3229                 ret = check_delayed_ref(root, path, objectid, offset, bytenr);
3230         } while (ret == -EAGAIN);
3231
3232 out:
3233         btrfs_free_path(path);
3234         if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID)
3235                 WARN_ON(ret > 0);
3236         return ret;
3237 }
3238
3239 static int __btrfs_mod_ref(struct btrfs_trans_handle *trans,
3240                            struct btrfs_root *root,
3241                            struct extent_buffer *buf,
3242                            int full_backref, int inc)
3243 {
3244         struct btrfs_fs_info *fs_info = root->fs_info;
3245         u64 bytenr;
3246         u64 num_bytes;
3247         u64 parent;
3248         u64 ref_root;
3249         u32 nritems;
3250         struct btrfs_key key;
3251         struct btrfs_file_extent_item *fi;
3252         int i;
3253         int level;
3254         int ret = 0;
3255         int (*process_func)(struct btrfs_trans_handle *,
3256                             struct btrfs_root *,
3257                             u64, u64, u64, u64, u64, u64);
3258
3259
3260         if (btrfs_is_testing(fs_info))
3261                 return 0;
3262
3263         ref_root = btrfs_header_owner(buf);
3264         nritems = btrfs_header_nritems(buf);
3265         level = btrfs_header_level(buf);
3266
3267         if (!test_bit(BTRFS_ROOT_REF_COWS, &root->state) && level == 0)
3268                 return 0;
3269
3270         if (inc)
3271                 process_func = btrfs_inc_extent_ref;
3272         else
3273                 process_func = btrfs_free_extent;
3274
3275         if (full_backref)
3276                 parent = buf->start;
3277         else
3278                 parent = 0;
3279
3280         for (i = 0; i < nritems; i++) {
3281                 if (level == 0) {
3282                         btrfs_item_key_to_cpu(buf, &key, i);
3283                         if (key.type != BTRFS_EXTENT_DATA_KEY)
3284                                 continue;
3285                         fi = btrfs_item_ptr(buf, i,
3286                                             struct btrfs_file_extent_item);
3287                         if (btrfs_file_extent_type(buf, fi) ==
3288                             BTRFS_FILE_EXTENT_INLINE)
3289                                 continue;
3290                         bytenr = btrfs_file_extent_disk_bytenr(buf, fi);
3291                         if (bytenr == 0)
3292                                 continue;
3293
3294                         num_bytes = btrfs_file_extent_disk_num_bytes(buf, fi);
3295                         key.offset -= btrfs_file_extent_offset(buf, fi);
3296                         ret = process_func(trans, root, bytenr, num_bytes,
3297                                            parent, ref_root, key.objectid,
3298                                            key.offset);
3299                         if (ret)
3300                                 goto fail;
3301                 } else {
3302                         bytenr = btrfs_node_blockptr(buf, i);
3303                         num_bytes = fs_info->nodesize;
3304                         ret = process_func(trans, root, bytenr, num_bytes,
3305                                            parent, ref_root, level - 1, 0);
3306                         if (ret)
3307                                 goto fail;
3308                 }
3309         }
3310         return 0;
3311 fail:
3312         return ret;
3313 }
3314
3315 int btrfs_inc_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
3316                   struct extent_buffer *buf, int full_backref)
3317 {
3318         return __btrfs_mod_ref(trans, root, buf, full_backref, 1);
3319 }
3320
3321 int btrfs_dec_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
3322                   struct extent_buffer *buf, int full_backref)
3323 {
3324         return __btrfs_mod_ref(trans, root, buf, full_backref, 0);
3325 }
3326
3327 static int write_one_cache_group(struct btrfs_trans_handle *trans,
3328                                  struct btrfs_fs_info *fs_info,
3329                                  struct btrfs_path *path,
3330                                  struct btrfs_block_group_cache *cache)
3331 {
3332         int ret;
3333         struct btrfs_root *extent_root = fs_info->extent_root;
3334         unsigned long bi;
3335         struct extent_buffer *leaf;
3336
3337         ret = btrfs_search_slot(trans, extent_root, &cache->key, path, 0, 1);
3338         if (ret) {
3339                 if (ret > 0)
3340                         ret = -ENOENT;
3341                 goto fail;
3342         }
3343
3344         leaf = path->nodes[0];
3345         bi = btrfs_item_ptr_offset(leaf, path->slots[0]);
3346         write_extent_buffer(leaf, &cache->item, bi, sizeof(cache->item));
3347         btrfs_mark_buffer_dirty(leaf);
3348 fail:
3349         btrfs_release_path(path);
3350         return ret;
3351
3352 }
3353
3354 static struct btrfs_block_group_cache *
3355 next_block_group(struct btrfs_fs_info *fs_info,
3356                  struct btrfs_block_group_cache *cache)
3357 {
3358         struct rb_node *node;
3359
3360         spin_lock(&fs_info->block_group_cache_lock);
3361
3362         /* If our block group was removed, we need a full search. */
3363         if (RB_EMPTY_NODE(&cache->cache_node)) {
3364                 const u64 next_bytenr = cache->key.objectid + cache->key.offset;
3365
3366                 spin_unlock(&fs_info->block_group_cache_lock);
3367                 btrfs_put_block_group(cache);
3368                 cache = btrfs_lookup_first_block_group(fs_info, next_bytenr); return cache;
3369         }
3370         node = rb_next(&cache->cache_node);
3371         btrfs_put_block_group(cache);
3372         if (node) {
3373                 cache = rb_entry(node, struct btrfs_block_group_cache,
3374                                  cache_node);
3375                 btrfs_get_block_group(cache);
3376         } else
3377                 cache = NULL;
3378         spin_unlock(&fs_info->block_group_cache_lock);
3379         return cache;
3380 }
3381
3382 static int cache_save_setup(struct btrfs_block_group_cache *block_group,
3383                             struct btrfs_trans_handle *trans,
3384                             struct btrfs_path *path)
3385 {
3386         struct btrfs_fs_info *fs_info = block_group->fs_info;
3387         struct btrfs_root *root = fs_info->tree_root;
3388         struct inode *inode = NULL;
3389         struct extent_changeset *data_reserved = NULL;
3390         u64 alloc_hint = 0;
3391         int dcs = BTRFS_DC_ERROR;
3392         u64 num_pages = 0;
3393         int retries = 0;
3394         int ret = 0;
3395
3396         /*
3397          * If this block group is smaller than 100 megs don't bother caching the
3398          * block group.
3399          */
3400         if (block_group->key.offset < (100 * SZ_1M)) {
3401                 spin_lock(&block_group->lock);
3402                 block_group->disk_cache_state = BTRFS_DC_WRITTEN;
3403                 spin_unlock(&block_group->lock);
3404                 return 0;
3405         }
3406
3407         if (trans->aborted)
3408                 return 0;
3409 again:
3410         inode = lookup_free_space_inode(fs_info, block_group, path);
3411         if (IS_ERR(inode) && PTR_ERR(inode) != -ENOENT) {
3412                 ret = PTR_ERR(inode);
3413                 btrfs_release_path(path);
3414                 goto out;
3415         }
3416
3417         if (IS_ERR(inode)) {
3418                 BUG_ON(retries);
3419                 retries++;
3420
3421                 if (block_group->ro)
3422                         goto out_free;
3423
3424                 ret = create_free_space_inode(fs_info, trans, block_group,
3425                                               path);
3426                 if (ret)
3427                         goto out_free;
3428                 goto again;
3429         }
3430
3431         /*
3432          * We want to set the generation to 0, that way if anything goes wrong
3433          * from here on out we know not to trust this cache when we load up next
3434          * time.
3435          */
3436         BTRFS_I(inode)->generation = 0;
3437         ret = btrfs_update_inode(trans, root, inode);
3438         if (ret) {
3439                 /*
3440                  * So theoretically we could recover from this, simply set the
3441                  * super cache generation to 0 so we know to invalidate the
3442                  * cache, but then we'd have to keep track of the block groups
3443                  * that fail this way so we know we _have_ to reset this cache
3444                  * before the next commit or risk reading stale cache.  So to
3445                  * limit our exposure to horrible edge cases lets just abort the
3446                  * transaction, this only happens in really bad situations
3447                  * anyway.
3448                  */
3449                 btrfs_abort_transaction(trans, ret);
3450                 goto out_put;
3451         }
3452         WARN_ON(ret);
3453
3454         /* We've already setup this transaction, go ahead and exit */
3455         if (block_group->cache_generation == trans->transid &&
3456             i_size_read(inode)) {
3457                 dcs = BTRFS_DC_SETUP;
3458                 goto out_put;
3459         }
3460
3461         if (i_size_read(inode) > 0) {
3462                 ret = btrfs_check_trunc_cache_free_space(fs_info,
3463                                         &fs_info->global_block_rsv);
3464                 if (ret)
3465                         goto out_put;
3466
3467                 ret = btrfs_truncate_free_space_cache(trans, NULL, inode);
3468                 if (ret)
3469                         goto out_put;
3470         }
3471
3472         spin_lock(&block_group->lock);
3473         if (block_group->cached != BTRFS_CACHE_FINISHED ||
3474             !btrfs_test_opt(fs_info, SPACE_CACHE)) {
3475                 /*
3476                  * don't bother trying to write stuff out _if_
3477                  * a) we're not cached,
3478                  * b) we're with nospace_cache mount option,
3479                  * c) we're with v2 space_cache (FREE_SPACE_TREE).
3480                  */
3481                 dcs = BTRFS_DC_WRITTEN;
3482                 spin_unlock(&block_group->lock);
3483                 goto out_put;
3484         }
3485         spin_unlock(&block_group->lock);
3486
3487         /*
3488          * We hit an ENOSPC when setting up the cache in this transaction, just
3489          * skip doing the setup, we've already cleared the cache so we're safe.
3490          */
3491         if (test_bit(BTRFS_TRANS_CACHE_ENOSPC, &trans->transaction->flags)) {
3492                 ret = -ENOSPC;
3493                 goto out_put;
3494         }
3495
3496         /*
3497          * Try to preallocate enough space based on how big the block group is.
3498          * Keep in mind this has to include any pinned space which could end up
3499          * taking up quite a bit since it's not folded into the other space
3500          * cache.
3501          */
3502         num_pages = div_u64(block_group->key.offset, SZ_256M);
3503         if (!num_pages)
3504                 num_pages = 1;
3505
3506         num_pages *= 16;
3507         num_pages *= PAGE_SIZE;
3508
3509         ret = btrfs_check_data_free_space(inode, &data_reserved, 0, num_pages);
3510         if (ret)
3511                 goto out_put;
3512
3513         ret = btrfs_prealloc_file_range_trans(inode, trans, 0, 0, num_pages,
3514                                               num_pages, num_pages,
3515                                               &alloc_hint);
3516         /*
3517          * Our cache requires contiguous chunks so that we don't modify a bunch
3518          * of metadata or split extents when writing the cache out, which means
3519          * we can enospc if we are heavily fragmented in addition to just normal
3520          * out of space conditions.  So if we hit this just skip setting up any
3521          * other block groups for this transaction, maybe we'll unpin enough
3522          * space the next time around.
3523          */
3524         if (!ret)
3525                 dcs = BTRFS_DC_SETUP;
3526         else if (ret == -ENOSPC)
3527                 set_bit(BTRFS_TRANS_CACHE_ENOSPC, &trans->transaction->flags);
3528
3529 out_put:
3530         iput(inode);
3531 out_free:
3532         btrfs_release_path(path);
3533 out:
3534         spin_lock(&block_group->lock);
3535         if (!ret && dcs == BTRFS_DC_SETUP)
3536                 block_group->cache_generation = trans->transid;
3537         block_group->disk_cache_state = dcs;
3538         spin_unlock(&block_group->lock);
3539
3540         extent_changeset_free(data_reserved);
3541         return ret;
3542 }
3543
3544 int btrfs_setup_space_cache(struct btrfs_trans_handle *trans,
3545                             struct btrfs_fs_info *fs_info)
3546 {
3547         struct btrfs_block_group_cache *cache, *tmp;
3548         struct btrfs_transaction *cur_trans = trans->transaction;
3549         struct btrfs_path *path;
3550
3551         if (list_empty(&cur_trans->dirty_bgs) ||
3552             !btrfs_test_opt(fs_info, SPACE_CACHE))
3553                 return 0;
3554
3555         path = btrfs_alloc_path();
3556         if (!path)
3557                 return -ENOMEM;
3558
3559         /* Could add new block groups, use _safe just in case */
3560         list_for_each_entry_safe(cache, tmp, &cur_trans->dirty_bgs,
3561                                  dirty_list) {
3562                 if (cache->disk_cache_state == BTRFS_DC_CLEAR)
3563                         cache_save_setup(cache, trans, path);
3564         }
3565
3566         btrfs_free_path(path);
3567         return 0;
3568 }
3569
3570 /*
3571  * transaction commit does final block group cache writeback during a
3572  * critical section where nothing is allowed to change the FS.  This is
3573  * required in order for the cache to actually match the block group,
3574  * but can introduce a lot of latency into the commit.
3575  *
3576  * So, btrfs_start_dirty_block_groups is here to kick off block group
3577  * cache IO.  There's a chance we'll have to redo some of it if the
3578  * block group changes again during the commit, but it greatly reduces
3579  * the commit latency by getting rid of the easy block groups while
3580  * we're still allowing others to join the commit.
3581  */
3582 int btrfs_start_dirty_block_groups(struct btrfs_trans_handle *trans)
3583 {
3584         struct btrfs_fs_info *fs_info = trans->fs_info;
3585         struct btrfs_block_group_cache *cache;
3586         struct btrfs_transaction *cur_trans = trans->transaction;
3587         int ret = 0;
3588         int should_put;
3589         struct btrfs_path *path = NULL;
3590         LIST_HEAD(dirty);
3591         struct list_head *io = &cur_trans->io_bgs;
3592         int num_started = 0;
3593         int loops = 0;
3594
3595         spin_lock(&cur_trans->dirty_bgs_lock);
3596         if (list_empty(&cur_trans->dirty_bgs)) {
3597                 spin_unlock(&cur_trans->dirty_bgs_lock);
3598                 return 0;
3599         }
3600         list_splice_init(&cur_trans->dirty_bgs, &dirty);
3601         spin_unlock(&cur_trans->dirty_bgs_lock);
3602
3603 again:
3604         /*
3605          * make sure all the block groups on our dirty list actually
3606          * exist
3607          */
3608         btrfs_create_pending_block_groups(trans);
3609
3610         if (!path) {
3611                 path = btrfs_alloc_path();
3612                 if (!path)
3613                         return -ENOMEM;
3614         }
3615
3616         /*
3617          * cache_write_mutex is here only to save us from balance or automatic
3618          * removal of empty block groups deleting this block group while we are
3619          * writing out the cache
3620          */
3621         mutex_lock(&trans->transaction->cache_write_mutex);
3622         while (!list_empty(&dirty)) {
3623                 bool drop_reserve = true;
3624
3625                 cache = list_first_entry(&dirty,
3626                                          struct btrfs_block_group_cache,
3627                                          dirty_list);
3628                 /*
3629                  * this can happen if something re-dirties a block
3630                  * group that is already under IO.  Just wait for it to
3631                  * finish and then do it all again
3632                  */
3633                 if (!list_empty(&cache->io_list)) {
3634                         list_del_init(&cache->io_list);
3635                         btrfs_wait_cache_io(trans, cache, path);
3636                         btrfs_put_block_group(cache);
3637                 }
3638
3639
3640                 /*
3641                  * btrfs_wait_cache_io uses the cache->dirty_list to decide
3642                  * if it should update the cache_state.  Don't delete
3643                  * until after we wait.
3644                  *
3645                  * Since we're not running in the commit critical section
3646                  * we need the dirty_bgs_lock to protect from update_block_group
3647                  */
3648                 spin_lock(&cur_trans->dirty_bgs_lock);
3649                 list_del_init(&cache->dirty_list);
3650                 spin_unlock(&cur_trans->dirty_bgs_lock);
3651
3652                 should_put = 1;
3653
3654                 cache_save_setup(cache, trans, path);
3655
3656                 if (cache->disk_cache_state == BTRFS_DC_SETUP) {
3657                         cache->io_ctl.inode = NULL;
3658                         ret = btrfs_write_out_cache(fs_info, trans,
3659                                                     cache, path);
3660                         if (ret == 0 && cache->io_ctl.inode) {
3661                                 num_started++;
3662                                 should_put = 0;
3663
3664                                 /*
3665                                  * The cache_write_mutex is protecting the
3666                                  * io_list, also refer to the definition of
3667                                  * btrfs_transaction::io_bgs for more details
3668                                  */
3669                                 list_add_tail(&cache->io_list, io);
3670                         } else {
3671                                 /*
3672                                  * if we failed to write the cache, the
3673                                  * generation will be bad and life goes on
3674                                  */
3675                                 ret = 0;
3676                         }
3677                 }
3678                 if (!ret) {
3679                         ret = write_one_cache_group(trans, fs_info,
3680                                                     path, cache);
3681                         /*
3682                          * Our block group might still be attached to the list
3683                          * of new block groups in the transaction handle of some
3684                          * other task (struct btrfs_trans_handle->new_bgs). This
3685                          * means its block group item isn't yet in the extent
3686                          * tree. If this happens ignore the error, as we will
3687                          * try again later in the critical section of the
3688                          * transaction commit.
3689                          */
3690                         if (ret == -ENOENT) {
3691                                 ret = 0;
3692                                 spin_lock(&cur_trans->dirty_bgs_lock);
3693                                 if (list_empty(&cache->dirty_list)) {
3694                                         list_add_tail(&cache->dirty_list,
3695                                                       &cur_trans->dirty_bgs);
3696                                         btrfs_get_block_group(cache);
3697                                         drop_reserve = false;
3698                                 }
3699                                 spin_unlock(&cur_trans->dirty_bgs_lock);
3700                         } else if (ret) {
3701                                 btrfs_abort_transaction(trans, ret);
3702                         }
3703                 }
3704
3705                 /* if it's not on the io list, we need to put the block group */
3706                 if (should_put)
3707                         btrfs_put_block_group(cache);
3708                 if (drop_reserve)
3709                         btrfs_delayed_refs_rsv_release(fs_info, 1);
3710
3711                 if (ret)
3712                         break;
3713
3714                 /*
3715                  * Avoid blocking other tasks for too long. It might even save
3716                  * us from writing caches for block groups that are going to be
3717                  * removed.
3718                  */
3719                 mutex_unlock(&trans->transaction->cache_write_mutex);
3720                 mutex_lock(&trans->transaction->cache_write_mutex);
3721         }
3722         mutex_unlock(&trans->transaction->cache_write_mutex);
3723
3724         /*
3725          * go through delayed refs for all the stuff we've just kicked off
3726          * and then loop back (just once)
3727          */
3728         ret = btrfs_run_delayed_refs(trans, 0);
3729         if (!ret && loops == 0) {
3730                 loops++;
3731                 spin_lock(&cur_trans->dirty_bgs_lock);
3732                 list_splice_init(&cur_trans->dirty_bgs, &dirty);
3733                 /*
3734                  * dirty_bgs_lock protects us from concurrent block group
3735                  * deletes too (not just cache_write_mutex).
3736                  */
3737                 if (!list_empty(&dirty)) {
3738                         spin_unlock(&cur_trans->dirty_bgs_lock);
3739                         goto again;
3740                 }
3741                 spin_unlock(&cur_trans->dirty_bgs_lock);
3742         } else if (ret < 0) {
3743                 btrfs_cleanup_dirty_bgs(cur_trans, fs_info);
3744         }
3745
3746         btrfs_free_path(path);
3747         return ret;
3748 }
3749
3750 int btrfs_write_dirty_block_groups(struct btrfs_trans_handle *trans,
3751                                    struct btrfs_fs_info *fs_info)
3752 {
3753         struct btrfs_block_group_cache *cache;
3754         struct btrfs_transaction *cur_trans = trans->transaction;
3755         int ret = 0;
3756         int should_put;
3757         struct btrfs_path *path;
3758         struct list_head *io = &cur_trans->io_bgs;
3759         int num_started = 0;
3760
3761         path = btrfs_alloc_path();
3762         if (!path)
3763                 return -ENOMEM;
3764
3765         /*
3766          * Even though we are in the critical section of the transaction commit,
3767          * we can still have concurrent tasks adding elements to this
3768          * transaction's list of dirty block groups. These tasks correspond to
3769          * endio free space workers started when writeback finishes for a
3770          * space cache, which run inode.c:btrfs_finish_ordered_io(), and can
3771          * allocate new block groups as a result of COWing nodes of the root
3772          * tree when updating the free space inode. The writeback for the space
3773          * caches is triggered by an earlier call to
3774          * btrfs_start_dirty_block_groups() and iterations of the following
3775          * loop.
3776          * Also we want to do the cache_save_setup first and then run the
3777          * delayed refs to make sure we have the best chance at doing this all
3778          * in one shot.
3779          */
3780         spin_lock(&cur_trans->dirty_bgs_lock);
3781         while (!list_empty(&cur_trans->dirty_bgs)) {
3782                 cache = list_first_entry(&cur_trans->dirty_bgs,
3783                                          struct btrfs_block_group_cache,
3784                                          dirty_list);
3785
3786                 /*
3787                  * this can happen if cache_save_setup re-dirties a block
3788                  * group that is already under IO.  Just wait for it to
3789                  * finish and then do it all again
3790                  */
3791                 if (!list_empty(&cache->io_list)) {
3792                         spin_unlock(&cur_trans->dirty_bgs_lock);
3793                         list_del_init(&cache->io_list);
3794                         btrfs_wait_cache_io(trans, cache, path);
3795                         btrfs_put_block_group(cache);
3796                         spin_lock(&cur_trans->dirty_bgs_lock);
3797                 }
3798
3799                 /*
3800                  * don't remove from the dirty list until after we've waited
3801                  * on any pending IO
3802                  */
3803                 list_del_init(&cache->dirty_list);
3804                 spin_unlock(&cur_trans->dirty_bgs_lock);
3805                 should_put = 1;
3806
3807                 cache_save_setup(cache, trans, path);
3808
3809                 if (!ret)
3810                         ret = btrfs_run_delayed_refs(trans,
3811                                                      (unsigned long) -1);
3812
3813                 if (!ret && cache->disk_cache_state == BTRFS_DC_SETUP) {
3814                         cache->io_ctl.inode = NULL;
3815                         ret = btrfs_write_out_cache(fs_info, trans,
3816                                                     cache, path);
3817                         if (ret == 0 && cache->io_ctl.inode) {
3818                                 num_started++;
3819                                 should_put = 0;
3820                                 list_add_tail(&cache->io_list, io);
3821                         } else {
3822                                 /*
3823                                  * if we failed to write the cache, the
3824                                  * generation will be bad and life goes on
3825                                  */
3826                                 ret = 0;
3827                         }
3828                 }
3829                 if (!ret) {
3830                         ret = write_one_cache_group(trans, fs_info,
3831                                                     path, cache);
3832                         /*
3833                          * One of the free space endio workers might have
3834                          * created a new block group while updating a free space
3835                          * cache's inode (at inode.c:btrfs_finish_ordered_io())
3836                          * and hasn't released its transaction handle yet, in
3837                          * which case the new block group is still attached to
3838                          * its transaction handle and its creation has not
3839                          * finished yet (no block group item in the extent tree
3840                          * yet, etc). If this is the case, wait for all free
3841                          * space endio workers to finish and retry. This is a
3842                          * a very rare case so no need for a more efficient and
3843                          * complex approach.
3844                          */
3845                         if (ret == -ENOENT) {
3846                                 wait_event(cur_trans->writer_wait,
3847                                    atomic_read(&cur_trans->num_writers) == 1);
3848                                 ret = write_one_cache_group(trans, fs_info,
3849                                                             path, cache);
3850                         }
3851                         if (ret)
3852                                 btrfs_abort_transaction(trans, ret);
3853                 }
3854
3855                 /* if its not on the io list, we need to put the block group */
3856                 if (should_put)
3857                         btrfs_put_block_group(cache);
3858                 btrfs_delayed_refs_rsv_release(fs_info, 1);
3859                 spin_lock(&cur_trans->dirty_bgs_lock);
3860         }
3861         spin_unlock(&cur_trans->dirty_bgs_lock);
3862
3863         /*
3864          * Refer to the definition of io_bgs member for details why it's safe
3865          * to use it without any locking
3866          */
3867         while (!list_empty(io)) {
3868                 cache = list_first_entry(io, struct btrfs_block_group_cache,
3869                                          io_list);
3870                 list_del_init(&cache->io_list);
3871                 btrfs_wait_cache_io(trans, cache, path);
3872                 btrfs_put_block_group(cache);
3873         }
3874
3875         btrfs_free_path(path);
3876         return ret;
3877 }
3878
3879 int btrfs_extent_readonly(struct btrfs_fs_info *fs_info, u64 bytenr)
3880 {
3881         struct btrfs_block_group_cache *block_group;
3882         int readonly = 0;
3883
3884         block_group = btrfs_lookup_block_group(fs_info, bytenr);
3885         if (!block_group || block_group->ro)
3886                 readonly = 1;
3887         if (block_group)
3888                 btrfs_put_block_group(block_group);
3889         return readonly;
3890 }
3891
3892 bool btrfs_inc_nocow_writers(struct btrfs_fs_info *fs_info, u64 bytenr)
3893 {
3894         struct btrfs_block_group_cache *bg;
3895         bool ret = true;
3896
3897         bg = btrfs_lookup_block_group(fs_info, bytenr);
3898         if (!bg)
3899                 return false;
3900
3901         spin_lock(&bg->lock);
3902         if (bg->ro)
3903                 ret = false;
3904         else
3905                 atomic_inc(&bg->nocow_writers);
3906         spin_unlock(&bg->lock);
3907
3908         /* no put on block group, done by btrfs_dec_nocow_writers */
3909         if (!ret)
3910                 btrfs_put_block_group(bg);
3911
3912         return ret;
3913
3914 }
3915
3916 void btrfs_dec_nocow_writers(struct btrfs_fs_info *fs_info, u64 bytenr)
3917 {
3918         struct btrfs_block_group_cache *bg;
3919
3920         bg = btrfs_lookup_block_group(fs_info, bytenr);
3921         ASSERT(bg);
3922         if (atomic_dec_and_test(&bg->nocow_writers))
3923                 wake_up_var(&bg->nocow_writers);
3924         /*
3925          * Once for our lookup and once for the lookup done by a previous call
3926          * to btrfs_inc_nocow_writers()
3927          */
3928         btrfs_put_block_group(bg);
3929         btrfs_put_block_group(bg);
3930 }
3931
3932 void btrfs_wait_nocow_writers(struct btrfs_block_group_cache *bg)
3933 {
3934         wait_var_event(&bg->nocow_writers, !atomic_read(&bg->nocow_writers));
3935 }
3936
3937 static const char *alloc_name(u64 flags)
3938 {
3939         switch (flags) {
3940         case BTRFS_BLOCK_GROUP_METADATA|BTRFS_BLOCK_GROUP_DATA:
3941                 return "mixed";
3942         case BTRFS_BLOCK_GROUP_METADATA:
3943                 return "metadata";
3944         case BTRFS_BLOCK_GROUP_DATA:
3945                 return "data";
3946         case BTRFS_BLOCK_GROUP_SYSTEM:
3947                 return "system";
3948         default:
3949                 WARN_ON(1);
3950                 return "invalid-combination";
3951         };
3952 }
3953
3954 static int create_space_info(struct btrfs_fs_info *info, u64 flags)
3955 {
3956
3957         struct btrfs_space_info *space_info;
3958         int i;
3959         int ret;
3960
3961         space_info = kzalloc(sizeof(*space_info), GFP_NOFS);
3962         if (!space_info)
3963                 return -ENOMEM;
3964
3965         ret = percpu_counter_init(&space_info->total_bytes_pinned, 0,
3966                                  GFP_KERNEL);
3967         if (ret) {
3968                 kfree(space_info);
3969                 return ret;
3970         }
3971
3972         for (i = 0; i < BTRFS_NR_RAID_TYPES; i++)
3973                 INIT_LIST_HEAD(&space_info->block_groups[i]);
3974         init_rwsem(&space_info->groups_sem);
3975         spin_lock_init(&space_info->lock);
3976         space_info->flags = flags & BTRFS_BLOCK_GROUP_TYPE_MASK;
3977         space_info->force_alloc = CHUNK_ALLOC_NO_FORCE;
3978         init_waitqueue_head(&space_info->wait);
3979         INIT_LIST_HEAD(&space_info->ro_bgs);
3980         INIT_LIST_HEAD(&space_info->tickets);
3981         INIT_LIST_HEAD(&space_info->priority_tickets);
3982
3983         ret = kobject_init_and_add(&space_info->kobj, &space_info_ktype,
3984                                     info->space_info_kobj, "%s",
3985                                     alloc_name(space_info->flags));
3986         if (ret) {
3987                 percpu_counter_destroy(&space_info->total_bytes_pinned);
3988                 kfree(space_info);
3989                 return ret;
3990         }
3991
3992         list_add_rcu(&space_info->list, &info->space_info);
3993         if (flags & BTRFS_BLOCK_GROUP_DATA)
3994                 info->data_sinfo = space_info;
3995
3996         return ret;
3997 }
3998
3999 static void update_space_info(struct btrfs_fs_info *info, u64 flags,
4000                              u64 total_bytes, u64 bytes_used,
4001                              u64 bytes_readonly,
4002                              struct btrfs_space_info **space_info)
4003 {
4004         struct btrfs_space_info *found;
4005         int factor;
4006
4007         factor = btrfs_bg_type_to_factor(flags);
4008
4009         found = __find_space_info(info, flags);
4010         ASSERT(found);
4011         spin_lock(&found->lock);
4012         found->total_bytes += total_bytes;
4013         found->disk_total += total_bytes * factor;
4014         found->bytes_used += bytes_used;
4015         found->disk_used += bytes_used * factor;
4016         found->bytes_readonly += bytes_readonly;
4017         if (total_bytes > 0)
4018                 found->full = 0;
4019         space_info_add_new_bytes(info, found, total_bytes -
4020                                  bytes_used - bytes_readonly);
4021         spin_unlock(&found->lock);
4022         *space_info = found;
4023 }
4024
4025 static void set_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
4026 {
4027         u64 extra_flags = chunk_to_extended(flags) &
4028                                 BTRFS_EXTENDED_PROFILE_MASK;
4029
4030         write_seqlock(&fs_info->profiles_lock);
4031         if (flags & BTRFS_BLOCK_GROUP_DATA)
4032                 fs_info->avail_data_alloc_bits |= extra_flags;
4033         if (flags & BTRFS_BLOCK_GROUP_METADATA)
4034                 fs_info->avail_metadata_alloc_bits |= extra_flags;
4035         if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
4036                 fs_info->avail_system_alloc_bits |= extra_flags;
4037         write_sequnlock(&fs_info->profiles_lock);
4038 }
4039
4040 /*
4041  * returns target flags in extended format or 0 if restripe for this
4042  * chunk_type is not in progress
4043  *
4044  * should be called with balance_lock held
4045  */
4046 static u64 get_restripe_target(struct btrfs_fs_info *fs_info, u64 flags)
4047 {
4048         struct btrfs_balance_control *bctl = fs_info->balance_ctl;
4049         u64 target = 0;
4050
4051         if (!bctl)
4052                 return 0;
4053
4054         if (flags & BTRFS_BLOCK_GROUP_DATA &&
4055             bctl->data.flags & BTRFS_BALANCE_ARGS_CONVERT) {
4056                 target = BTRFS_BLOCK_GROUP_DATA | bctl->data.target;
4057         } else if (flags & BTRFS_BLOCK_GROUP_SYSTEM &&
4058                    bctl->sys.flags & BTRFS_BALANCE_ARGS_CONVERT) {
4059                 target = BTRFS_BLOCK_GROUP_SYSTEM | bctl->sys.target;
4060         } else if (flags & BTRFS_BLOCK_GROUP_METADATA &&
4061                    bctl->meta.flags & BTRFS_BALANCE_ARGS_CONVERT) {
4062                 target = BTRFS_BLOCK_GROUP_METADATA | bctl->meta.target;
4063         }
4064
4065         return target;
4066 }
4067
4068 /*
4069  * @flags: available profiles in extended format (see ctree.h)
4070  *
4071  * Returns reduced profile in chunk format.  If profile changing is in
4072  * progress (either running or paused) picks the target profile (if it's
4073  * already available), otherwise falls back to plain reducing.
4074  */
4075 static u64 btrfs_reduce_alloc_profile(struct btrfs_fs_info *fs_info, u64 flags)
4076 {
4077         u64 num_devices = fs_info->fs_devices->rw_devices;
4078         u64 target;
4079         u64 raid_type;
4080         u64 allowed = 0;
4081
4082         /*
4083          * see if restripe for this chunk_type is in progress, if so
4084          * try to reduce to the target profile
4085          */
4086         spin_lock(&fs_info->balance_lock);
4087         target = get_restripe_target(fs_info, flags);
4088         if (target) {
4089                 /* pick target profile only if it's already available */
4090                 if ((flags & target) & BTRFS_EXTENDED_PROFILE_MASK) {
4091                         spin_unlock(&fs_info->balance_lock);
4092                         return extended_to_chunk(target);
4093                 }
4094         }
4095         spin_unlock(&fs_info->balance_lock);
4096
4097         /* First, mask out the RAID levels which aren't possible */
4098         for (raid_type = 0; raid_type < BTRFS_NR_RAID_TYPES; raid_type++) {
4099                 if (num_devices >= btrfs_raid_array[raid_type].devs_min)
4100                         allowed |= btrfs_raid_array[raid_type].bg_flag;
4101         }
4102         allowed &= flags;
4103
4104         if (allowed & BTRFS_BLOCK_GROUP_RAID6)
4105                 allowed = BTRFS_BLOCK_GROUP_RAID6;
4106         else if (allowed & BTRFS_BLOCK_GROUP_RAID5)
4107                 allowed = BTRFS_BLOCK_GROUP_RAID5;
4108         else if (allowed & BTRFS_BLOCK_GROUP_RAID10)
4109                 allowed = BTRFS_BLOCK_GROUP_RAID10;
4110         else if (allowed & BTRFS_BLOCK_GROUP_RAID1)
4111                 allowed = BTRFS_BLOCK_GROUP_RAID1;
4112         else if (allowed & BTRFS_BLOCK_GROUP_RAID0)
4113                 allowed = BTRFS_BLOCK_GROUP_RAID0;
4114
4115         flags &= ~BTRFS_BLOCK_GROUP_PROFILE_MASK;
4116
4117         return extended_to_chunk(flags | allowed);
4118 }
4119
4120 static u64 get_alloc_profile(struct btrfs_fs_info *fs_info, u64 orig_flags)
4121 {
4122         unsigned seq;
4123         u64 flags;
4124
4125         do {
4126                 flags = orig_flags;
4127                 seq = read_seqbegin(&fs_info->profiles_lock);
4128
4129                 if (flags & BTRFS_BLOCK_GROUP_DATA)
4130                         flags |= fs_info->avail_data_alloc_bits;
4131                 else if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
4132                         flags |= fs_info->avail_system_alloc_bits;
4133                 else if (flags & BTRFS_BLOCK_GROUP_METADATA)
4134                         flags |= fs_info->avail_metadata_alloc_bits;
4135         } while (read_seqretry(&fs_info->profiles_lock, seq));
4136
4137         return btrfs_reduce_alloc_profile(fs_info, flags);
4138 }
4139
4140 static u64 get_alloc_profile_by_root(struct btrfs_root *root, int data)
4141 {
4142         struct btrfs_fs_info *fs_info = root->fs_info;
4143         u64 flags;
4144         u64 ret;
4145
4146         if (data)
4147                 flags = BTRFS_BLOCK_GROUP_DATA;
4148         else if (root == fs_info->chunk_root)
4149                 flags = BTRFS_BLOCK_GROUP_SYSTEM;
4150         else
4151                 flags = BTRFS_BLOCK_GROUP_METADATA;
4152
4153         ret = get_alloc_profile(fs_info, flags);
4154         return ret;
4155 }
4156
4157 u64 btrfs_data_alloc_profile(struct btrfs_fs_info *fs_info)
4158 {
4159         return get_alloc_profile(fs_info, BTRFS_BLOCK_GROUP_DATA);
4160 }
4161
4162 u64 btrfs_metadata_alloc_profile(struct btrfs_fs_info *fs_info)
4163 {
4164         return get_alloc_profile(fs_info, BTRFS_BLOCK_GROUP_METADATA);
4165 }
4166
4167 u64 btrfs_system_alloc_profile(struct btrfs_fs_info *fs_info)
4168 {
4169         return get_alloc_profile(fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
4170 }
4171
4172 static u64 btrfs_space_info_used(struct btrfs_space_info *s_info,
4173                                  bool may_use_included)
4174 {
4175         ASSERT(s_info);
4176         return s_info->bytes_used + s_info->bytes_reserved +
4177                 s_info->bytes_pinned + s_info->bytes_readonly +
4178                 (may_use_included ? s_info->bytes_may_use : 0);
4179 }
4180
4181 int btrfs_alloc_data_chunk_ondemand(struct btrfs_inode *inode, u64 bytes)
4182 {
4183         struct btrfs_root *root = inode->root;
4184         struct btrfs_fs_info *fs_info = root->fs_info;
4185         struct btrfs_space_info *data_sinfo = fs_info->data_sinfo;
4186         u64 used;
4187         int ret = 0;
4188         int need_commit = 2;
4189         int have_pinned_space;
4190
4191         /* make sure bytes are sectorsize aligned */
4192         bytes = ALIGN(bytes, fs_info->sectorsize);
4193
4194         if (btrfs_is_free_space_inode(inode)) {
4195                 need_commit = 0;
4196                 ASSERT(current->journal_info);
4197         }
4198
4199 again:
4200         /* make sure we have enough space to handle the data first */
4201         spin_lock(&data_sinfo->lock);
4202         used = btrfs_space_info_used(data_sinfo, true);
4203
4204         if (used + bytes > data_sinfo->total_bytes) {
4205                 struct btrfs_trans_handle *trans;
4206
4207                 /*
4208                  * if we don't have enough free bytes in this space then we need
4209                  * to alloc a new chunk.
4210                  */
4211                 if (!data_sinfo->full) {
4212                         u64 alloc_target;
4213
4214                         data_sinfo->force_alloc = CHUNK_ALLOC_FORCE;
4215                         spin_unlock(&data_sinfo->lock);
4216
4217                         alloc_target = btrfs_data_alloc_profile(fs_info);
4218                         /*
4219                          * It is ugly that we don't call nolock join
4220                          * transaction for the free space inode case here.
4221                          * But it is safe because we only do the data space
4222                          * reservation for the free space cache in the
4223                          * transaction context, the common join transaction
4224                          * just increase the counter of the current transaction
4225                          * handler, doesn't try to acquire the trans_lock of
4226                          * the fs.
4227                          */
4228                         trans = btrfs_join_transaction(root);
4229                         if (IS_ERR(trans))
4230                                 return PTR_ERR(trans);
4231
4232                         ret = do_chunk_alloc(trans, alloc_target,
4233                                              CHUNK_ALLOC_NO_FORCE);
4234                         btrfs_end_transaction(trans);
4235                         if (ret < 0) {
4236                                 if (ret != -ENOSPC)
4237                                         return ret;
4238                                 else {
4239                                         have_pinned_space = 1;
4240                                         goto commit_trans;
4241                                 }
4242                         }
4243
4244                         goto again;
4245                 }
4246
4247                 /*
4248                  * If we don't have enough pinned space to deal with this
4249                  * allocation, and no removed chunk in current transaction,
4250                  * don't bother committing the transaction.
4251                  */
4252                 have_pinned_space = __percpu_counter_compare(
4253                         &data_sinfo->total_bytes_pinned,
4254                         used + bytes - data_sinfo->total_bytes,
4255                         BTRFS_TOTAL_BYTES_PINNED_BATCH);
4256                 spin_unlock(&data_sinfo->lock);
4257
4258                 /* commit the current transaction and try again */
4259 commit_trans:
4260                 if (need_commit) {
4261                         need_commit--;
4262
4263                         if (need_commit > 0) {
4264                                 btrfs_start_delalloc_roots(fs_info, -1);
4265                                 btrfs_wait_ordered_roots(fs_info, U64_MAX, 0,
4266                                                          (u64)-1);
4267                         }
4268
4269                         trans = btrfs_join_transaction(root);
4270                         if (IS_ERR(trans))
4271                                 return PTR_ERR(trans);
4272                         if (have_pinned_space >= 0 ||
4273                             test_bit(BTRFS_TRANS_HAVE_FREE_BGS,
4274                                      &trans->transaction->flags) ||
4275                             need_commit > 0) {
4276                                 ret = btrfs_commit_transaction(trans);
4277                                 if (ret)
4278                                         return ret;
4279                                 /*
4280                                  * The cleaner kthread might still be doing iput
4281                                  * operations. Wait for it to finish so that
4282                                  * more space is released.  We don't need to
4283                                  * explicitly run the delayed iputs here because
4284                                  * the commit_transaction would have woken up
4285                                  * the cleaner.
4286                                  */
4287                                 ret = btrfs_wait_on_delayed_iputs(fs_info);
4288                                 if (ret)
4289                                         return ret;
4290                                 goto again;
4291                         } else {
4292                                 btrfs_end_transaction(trans);
4293                         }
4294                 }
4295
4296                 trace_btrfs_space_reservation(fs_info,
4297                                               "space_info:enospc",
4298                                               data_sinfo->flags, bytes, 1);
4299                 return -ENOSPC;
4300         }
4301         update_bytes_may_use(data_sinfo, bytes);
4302         trace_btrfs_space_reservation(fs_info, "space_info",
4303                                       data_sinfo->flags, bytes, 1);
4304         spin_unlock(&data_sinfo->lock);
4305
4306         return 0;
4307 }
4308
4309 int btrfs_check_data_free_space(struct inode *inode,
4310                         struct extent_changeset **reserved, u64 start, u64 len)
4311 {
4312         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
4313         int ret;
4314
4315         /* align the range */
4316         len = round_up(start + len, fs_info->sectorsize) -
4317               round_down(start, fs_info->sectorsize);
4318         start = round_down(start, fs_info->sectorsize);
4319
4320         ret = btrfs_alloc_data_chunk_ondemand(BTRFS_I(inode), len);
4321         if (ret < 0)
4322                 return ret;
4323
4324         /* Use new btrfs_qgroup_reserve_data to reserve precious data space. */
4325         ret = btrfs_qgroup_reserve_data(inode, reserved, start, len);
4326         if (ret < 0)
4327                 btrfs_free_reserved_data_space_noquota(inode, start, len);
4328         else
4329                 ret = 0;
4330         return ret;
4331 }
4332
4333 /*
4334  * Called if we need to clear a data reservation for this inode
4335  * Normally in a error case.
4336  *
4337  * This one will *NOT* use accurate qgroup reserved space API, just for case
4338  * which we can't sleep and is sure it won't affect qgroup reserved space.
4339  * Like clear_bit_hook().
4340  */
4341 void btrfs_free_reserved_data_space_noquota(struct inode *inode, u64 start,
4342                                             u64 len)
4343 {
4344         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
4345         struct btrfs_space_info *data_sinfo;
4346
4347         /* Make sure the range is aligned to sectorsize */
4348         len = round_up(start + len, fs_info->sectorsize) -
4349               round_down(start, fs_info->sectorsize);
4350         start = round_down(start, fs_info->sectorsize);
4351
4352         data_sinfo = fs_info->data_sinfo;
4353         spin_lock(&data_sinfo->lock);
4354         update_bytes_may_use(data_sinfo, -len);
4355         trace_btrfs_space_reservation(fs_info, "space_info",
4356                                       data_sinfo->flags, len, 0);
4357         spin_unlock(&data_sinfo->lock);
4358 }
4359
4360 /*
4361  * Called if we need to clear a data reservation for this inode
4362  * Normally in a error case.
4363  *
4364  * This one will handle the per-inode data rsv map for accurate reserved
4365  * space framework.
4366  */
4367 void btrfs_free_reserved_data_space(struct inode *inode,
4368                         struct extent_changeset *reserved, u64 start, u64 len)
4369 {
4370         struct btrfs_root *root = BTRFS_I(inode)->root;
4371
4372         /* Make sure the range is aligned to sectorsize */
4373         len = round_up(start + len, root->fs_info->sectorsize) -
4374               round_down(start, root->fs_info->sectorsize);
4375         start = round_down(start, root->fs_info->sectorsize);
4376
4377         btrfs_free_reserved_data_space_noquota(inode, start, len);
4378         btrfs_qgroup_free_data(inode, reserved, start, len);
4379 }
4380
4381 static void force_metadata_allocation(struct btrfs_fs_info *info)
4382 {
4383         struct list_head *head = &info->space_info;
4384         struct btrfs_space_info *found;
4385
4386         rcu_read_lock();
4387         list_for_each_entry_rcu(found, head, list) {
4388                 if (found->flags & BTRFS_BLOCK_GROUP_METADATA)
4389                         found->force_alloc = CHUNK_ALLOC_FORCE;
4390         }
4391         rcu_read_unlock();
4392 }
4393
4394 static inline u64 calc_global_rsv_need_space(struct btrfs_block_rsv *global)
4395 {
4396         return (global->size << 1);
4397 }
4398
4399 static int should_alloc_chunk(struct btrfs_fs_info *fs_info,
4400                               struct btrfs_space_info *sinfo, int force)
4401 {
4402         u64 bytes_used = btrfs_space_info_used(sinfo, false);
4403         u64 thresh;
4404
4405         if (force == CHUNK_ALLOC_FORCE)
4406                 return 1;
4407
4408         /*
4409          * in limited mode, we want to have some free space up to
4410          * about 1% of the FS size.
4411          */
4412         if (force == CHUNK_ALLOC_LIMITED) {
4413                 thresh = btrfs_super_total_bytes(fs_info->super_copy);
4414                 thresh = max_t(u64, SZ_64M, div_factor_fine(thresh, 1));
4415
4416                 if (sinfo->total_bytes - bytes_used < thresh)
4417                         return 1;
4418         }
4419
4420         if (bytes_used + SZ_2M < div_factor(sinfo->total_bytes, 8))
4421                 return 0;
4422         return 1;
4423 }
4424
4425 static u64 get_profile_num_devs(struct btrfs_fs_info *fs_info, u64 type)
4426 {
4427         u64 num_dev;
4428
4429         if (type & (BTRFS_BLOCK_GROUP_RAID10 |
4430                     BTRFS_BLOCK_GROUP_RAID0 |
4431                     BTRFS_BLOCK_GROUP_RAID5 |
4432                     BTRFS_BLOCK_GROUP_RAID6))
4433                 num_dev = fs_info->fs_devices->rw_devices;
4434         else if (type & BTRFS_BLOCK_GROUP_RAID1)
4435                 num_dev = 2;
4436         else
4437                 num_dev = 1;    /* DUP or single */
4438
4439         return num_dev;
4440 }
4441
4442 /*
4443  * If @is_allocation is true, reserve space in the system space info necessary
4444  * for allocating a chunk, otherwise if it's false, reserve space necessary for
4445  * removing a chunk.
4446  */
4447 void check_system_chunk(struct btrfs_trans_handle *trans, u64 type)
4448 {
4449         struct btrfs_fs_info *fs_info = trans->fs_info;
4450         struct btrfs_space_info *info;
4451         u64 left;
4452         u64 thresh;
4453         int ret = 0;
4454         u64 num_devs;
4455
4456         /*
4457          * Needed because we can end up allocating a system chunk and for an
4458          * atomic and race free space reservation in the chunk block reserve.
4459          */
4460         lockdep_assert_held(&fs_info->chunk_mutex);
4461
4462         info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
4463         spin_lock(&info->lock);
4464         left = info->total_bytes - btrfs_space_info_used(info, true);
4465         spin_unlock(&info->lock);
4466
4467         num_devs = get_profile_num_devs(fs_info, type);
4468
4469         /* num_devs device items to update and 1 chunk item to add or remove */
4470         thresh = btrfs_calc_trunc_metadata_size(fs_info, num_devs) +
4471                 btrfs_calc_trans_metadata_size(fs_info, 1);
4472
4473         if (left < thresh && btrfs_test_opt(fs_info, ENOSPC_DEBUG)) {
4474                 btrfs_info(fs_info, "left=%llu, need=%llu, flags=%llu",
4475                            left, thresh, type);
4476                 dump_space_info(fs_info, info, 0, 0);
4477         }
4478
4479         if (left < thresh) {
4480                 u64 flags = btrfs_system_alloc_profile(fs_info);
4481
4482                 /*
4483                  * Ignore failure to create system chunk. We might end up not
4484                  * needing it, as we might not need to COW all nodes/leafs from
4485                  * the paths we visit in the chunk tree (they were already COWed
4486                  * or created in the current transaction for example).
4487                  */
4488                 ret = btrfs_alloc_chunk(trans, flags);
4489         }
4490
4491         if (!ret) {
4492                 ret = btrfs_block_rsv_add(fs_info->chunk_root,
4493                                           &fs_info->chunk_block_rsv,
4494                                           thresh, BTRFS_RESERVE_NO_FLUSH);
4495                 if (!ret)
4496                         trans->chunk_bytes_reserved += thresh;
4497         }
4498 }
4499
4500 /*
4501  * If force is CHUNK_ALLOC_FORCE:
4502  *    - return 1 if it successfully allocates a chunk,
4503  *    - return errors including -ENOSPC otherwise.
4504  * If force is NOT CHUNK_ALLOC_FORCE:
4505  *    - return 0 if it doesn't need to allocate a new chunk,
4506  *    - return 1 if it successfully allocates a chunk,
4507  *    - return errors including -ENOSPC otherwise.
4508  */
4509 static int do_chunk_alloc(struct btrfs_trans_handle *trans, u64 flags,
4510                           int force)
4511 {
4512         struct btrfs_fs_info *fs_info = trans->fs_info;
4513         struct btrfs_space_info *space_info;
4514         bool wait_for_alloc = false;
4515         bool should_alloc = false;
4516         int ret = 0;
4517
4518         /* Don't re-enter if we're already allocating a chunk */
4519         if (trans->allocating_chunk)
4520                 return -ENOSPC;
4521
4522         space_info = __find_space_info(fs_info, flags);
4523         ASSERT(space_info);
4524
4525         do {
4526                 spin_lock(&space_info->lock);
4527                 if (force < space_info->force_alloc)
4528                         force = space_info->force_alloc;
4529                 should_alloc = should_alloc_chunk(fs_info, space_info, force);
4530                 if (space_info->full) {
4531                         /* No more free physical space */
4532                         if (should_alloc)
4533                                 ret = -ENOSPC;
4534                         else
4535                                 ret = 0;
4536                         spin_unlock(&space_info->lock);
4537                         return ret;
4538                 } else if (!should_alloc) {
4539                         spin_unlock(&space_info->lock);
4540                         return 0;
4541                 } else if (space_info->chunk_alloc) {
4542                         /*
4543                          * Someone is already allocating, so we need to block
4544                          * until this someone is finished and then loop to
4545                          * recheck if we should continue with our allocation
4546                          * attempt.
4547                          */
4548                         wait_for_alloc = true;
4549                         spin_unlock(&space_info->lock);
4550                         mutex_lock(&fs_info->chunk_mutex);
4551                         mutex_unlock(&fs_info->chunk_mutex);
4552                 } else {
4553                         /* Proceed with allocation */
4554                         space_info->chunk_alloc = 1;
4555                         wait_for_alloc = false;
4556                         spin_unlock(&space_info->lock);
4557                 }
4558
4559                 cond_resched();
4560         } while (wait_for_alloc);
4561
4562         mutex_lock(&fs_info->chunk_mutex);
4563         trans->allocating_chunk = true;
4564
4565         /*
4566          * If we have mixed data/metadata chunks we want to make sure we keep
4567          * allocating mixed chunks instead of individual chunks.
4568          */
4569         if (btrfs_mixed_space_info(space_info))
4570                 flags |= (BTRFS_BLOCK_GROUP_DATA | BTRFS_BLOCK_GROUP_METADATA);
4571
4572         /*
4573          * if we're doing a data chunk, go ahead and make sure that
4574          * we keep a reasonable number of metadata chunks allocated in the
4575          * FS as well.
4576          */
4577         if (flags & BTRFS_BLOCK_GROUP_DATA && fs_info->metadata_ratio) {
4578                 fs_info->data_chunk_allocations++;
4579                 if (!(fs_info->data_chunk_allocations %
4580                       fs_info->metadata_ratio))
4581                         force_metadata_allocation(fs_info);
4582         }
4583
4584         /*
4585          * Check if we have enough space in SYSTEM chunk because we may need
4586          * to update devices.
4587          */
4588         check_system_chunk(trans, flags);
4589
4590         ret = btrfs_alloc_chunk(trans, flags);
4591         trans->allocating_chunk = false;
4592
4593         spin_lock(&space_info->lock);
4594         if (ret < 0) {
4595                 if (ret == -ENOSPC)
4596                         space_info->full = 1;
4597                 else
4598                         goto out;
4599         } else {
4600                 ret = 1;
4601                 space_info->max_extent_size = 0;
4602         }
4603
4604         space_info->force_alloc = CHUNK_ALLOC_NO_FORCE;
4605 out:
4606         space_info->chunk_alloc = 0;
4607         spin_unlock(&space_info->lock);
4608         mutex_unlock(&fs_info->chunk_mutex);
4609         /*
4610          * When we allocate a new chunk we reserve space in the chunk block
4611          * reserve to make sure we can COW nodes/leafs in the chunk tree or
4612          * add new nodes/leafs to it if we end up needing to do it when
4613          * inserting the chunk item and updating device items as part of the
4614          * second phase of chunk allocation, performed by
4615          * btrfs_finish_chunk_alloc(). So make sure we don't accumulate a
4616          * large number of new block groups to create in our transaction
4617          * handle's new_bgs list to avoid exhausting the chunk block reserve
4618          * in extreme cases - like having a single transaction create many new
4619          * block groups when starting to write out the free space caches of all
4620          * the block groups that were made dirty during the lifetime of the
4621          * transaction.
4622          */
4623         if (trans->chunk_bytes_reserved >= (u64)SZ_2M)
4624                 btrfs_create_pending_block_groups(trans);
4625
4626         return ret;
4627 }
4628
4629 static int can_overcommit(struct btrfs_fs_info *fs_info,
4630                           struct btrfs_space_info *space_info, u64 bytes,
4631                           enum btrfs_reserve_flush_enum flush,
4632                           bool system_chunk)
4633 {
4634         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
4635         u64 profile;
4636         u64 space_size;
4637         u64 avail;
4638         u64 used;
4639         int factor;
4640
4641         /* Don't overcommit when in mixed mode. */
4642         if (space_info->flags & BTRFS_BLOCK_GROUP_DATA)
4643                 return 0;
4644
4645         if (system_chunk)
4646                 profile = btrfs_system_alloc_profile(fs_info);
4647         else
4648                 profile = btrfs_metadata_alloc_profile(fs_info);
4649
4650         used = btrfs_space_info_used(space_info, false);
4651
4652         /*
4653          * We only want to allow over committing if we have lots of actual space
4654          * free, but if we don't have enough space to handle the global reserve
4655          * space then we could end up having a real enospc problem when trying
4656          * to allocate a chunk or some other such important allocation.
4657          */
4658         spin_lock(&global_rsv->lock);
4659         space_size = calc_global_rsv_need_space(global_rsv);
4660         spin_unlock(&global_rsv->lock);
4661         if (used + space_size >= space_info->total_bytes)
4662                 return 0;
4663
4664         used += space_info->bytes_may_use;
4665
4666         avail = atomic64_read(&fs_info->free_chunk_space);
4667
4668         /*
4669          * If we have dup, raid1 or raid10 then only half of the free
4670          * space is actually usable.  For raid56, the space info used
4671          * doesn't include the parity drive, so we don't have to
4672          * change the math
4673          */
4674         factor = btrfs_bg_type_to_factor(profile);
4675         avail = div_u64(avail, factor);
4676
4677         /*
4678          * If we aren't flushing all things, let us overcommit up to
4679          * 1/2th of the space. If we can flush, don't let us overcommit
4680          * too much, let it overcommit up to 1/8 of the space.
4681          */
4682         if (flush == BTRFS_RESERVE_FLUSH_ALL)
4683                 avail >>= 3;
4684         else
4685                 avail >>= 1;
4686
4687         if (used + bytes < space_info->total_bytes + avail)
4688                 return 1;
4689         return 0;
4690 }
4691
4692 static void btrfs_writeback_inodes_sb_nr(struct btrfs_fs_info *fs_info,
4693                                          unsigned long nr_pages, int nr_items)
4694 {
4695         struct super_block *sb = fs_info->sb;
4696
4697         if (down_read_trylock(&sb->s_umount)) {
4698                 writeback_inodes_sb_nr(sb, nr_pages, WB_REASON_FS_FREE_SPACE);
4699                 up_read(&sb->s_umount);
4700         } else {
4701                 /*
4702                  * We needn't worry the filesystem going from r/w to r/o though
4703                  * we don't acquire ->s_umount mutex, because the filesystem
4704                  * should guarantee the delalloc inodes list be empty after
4705                  * the filesystem is readonly(all dirty pages are written to
4706                  * the disk).
4707                  */
4708                 btrfs_start_delalloc_roots(fs_info, nr_items);
4709                 if (!current->journal_info)
4710                         btrfs_wait_ordered_roots(fs_info, nr_items, 0, (u64)-1);
4711         }
4712 }
4713
4714 static inline u64 calc_reclaim_items_nr(struct btrfs_fs_info *fs_info,
4715                                         u64 to_reclaim)
4716 {
4717         u64 bytes;
4718         u64 nr;
4719
4720         bytes = btrfs_calc_trans_metadata_size(fs_info, 1);
4721         nr = div64_u64(to_reclaim, bytes);
4722         if (!nr)
4723                 nr = 1;
4724         return nr;
4725 }
4726
4727 #define EXTENT_SIZE_PER_ITEM    SZ_256K
4728
4729 /*
4730  * shrink metadata reservation for delalloc
4731  */
4732 static void shrink_delalloc(struct btrfs_fs_info *fs_info, u64 to_reclaim,
4733                             u64 orig, bool wait_ordered)
4734 {
4735         struct btrfs_space_info *space_info;
4736         struct btrfs_trans_handle *trans;
4737         u64 delalloc_bytes;
4738         u64 async_pages;
4739         u64 items;
4740         long time_left;
4741         unsigned long nr_pages;
4742         int loops;
4743
4744         /* Calc the number of the pages we need flush for space reservation */
4745         items = calc_reclaim_items_nr(fs_info, to_reclaim);
4746         to_reclaim = items * EXTENT_SIZE_PER_ITEM;
4747
4748         trans = (struct btrfs_trans_handle *)current->journal_info;
4749         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
4750
4751         delalloc_bytes = percpu_counter_sum_positive(
4752                                                 &fs_info->delalloc_bytes);
4753         if (delalloc_bytes == 0) {
4754                 if (trans)
4755                         return;
4756                 if (wait_ordered)
4757                         btrfs_wait_ordered_roots(fs_info, items, 0, (u64)-1);
4758                 return;
4759         }
4760
4761         loops = 0;
4762         while (delalloc_bytes && loops < 3) {
4763                 nr_pages = min(delalloc_bytes, to_reclaim) >> PAGE_SHIFT;
4764
4765                 /*
4766                  * Triggers inode writeback for up to nr_pages. This will invoke
4767                  * ->writepages callback and trigger delalloc filling
4768                  *  (btrfs_run_delalloc_range()).
4769                  */
4770                 btrfs_writeback_inodes_sb_nr(fs_info, nr_pages, items);
4771
4772                 /*
4773                  * We need to wait for the compressed pages to start before
4774                  * we continue.
4775                  */
4776                 async_pages = atomic_read(&fs_info->async_delalloc_pages);
4777                 if (!async_pages)
4778                         goto skip_async;
4779
4780                 /*
4781                  * Calculate how many compressed pages we want to be written
4782                  * before we continue. I.e if there are more async pages than we
4783                  * require wait_event will wait until nr_pages are written.
4784                  */
4785                 if (async_pages <= nr_pages)
4786                         async_pages = 0;
4787                 else
4788                         async_pages -= nr_pages;
4789
4790                 wait_event(fs_info->async_submit_wait,
4791                            atomic_read(&fs_info->async_delalloc_pages) <=
4792                            (int)async_pages);
4793 skip_async:
4794                 spin_lock(&space_info->lock);
4795                 if (list_empty(&space_info->tickets) &&
4796                     list_empty(&space_info->priority_tickets)) {
4797                         spin_unlock(&space_info->lock);
4798                         break;
4799                 }
4800                 spin_unlock(&space_info->lock);
4801
4802                 loops++;
4803                 if (wait_ordered && !trans) {
4804                         btrfs_wait_ordered_roots(fs_info, items, 0, (u64)-1);
4805                 } else {
4806                         time_left = schedule_timeout_killable(1);
4807                         if (time_left)
4808                                 break;
4809                 }
4810                 delalloc_bytes = percpu_counter_sum_positive(
4811                                                 &fs_info->delalloc_bytes);
4812         }
4813 }
4814
4815 struct reserve_ticket {
4816         u64 orig_bytes;
4817         u64 bytes;
4818         int error;
4819         struct list_head list;
4820         wait_queue_head_t wait;
4821 };
4822
4823 /**
4824  * maybe_commit_transaction - possibly commit the transaction if its ok to
4825  * @root - the root we're allocating for
4826  * @bytes - the number of bytes we want to reserve
4827  * @force - force the commit
4828  *
4829  * This will check to make sure that committing the transaction will actually
4830  * get us somewhere and then commit the transaction if it does.  Otherwise it
4831  * will return -ENOSPC.
4832  */
4833 static int may_commit_transaction(struct btrfs_fs_info *fs_info,
4834                                   struct btrfs_space_info *space_info)
4835 {
4836         struct reserve_ticket *ticket = NULL;
4837         struct btrfs_block_rsv *delayed_rsv = &fs_info->delayed_block_rsv;
4838         struct btrfs_block_rsv *delayed_refs_rsv = &fs_info->delayed_refs_rsv;
4839         struct btrfs_trans_handle *trans;
4840         u64 bytes_needed;
4841         u64 reclaim_bytes = 0;
4842
4843         trans = (struct btrfs_trans_handle *)current->journal_info;
4844         if (trans)
4845                 return -EAGAIN;
4846
4847         spin_lock(&space_info->lock);
4848         if (!list_empty(&space_info->priority_tickets))
4849                 ticket = list_first_entry(&space_info->priority_tickets,
4850                                           struct reserve_ticket, list);
4851         else if (!list_empty(&space_info->tickets))
4852                 ticket = list_first_entry(&space_info->tickets,
4853                                           struct reserve_ticket, list);
4854         bytes_needed = (ticket) ? ticket->bytes : 0;
4855         spin_unlock(&space_info->lock);
4856
4857         if (!bytes_needed)
4858                 return 0;
4859
4860         trans = btrfs_join_transaction(fs_info->extent_root);
4861         if (IS_ERR(trans))
4862                 return PTR_ERR(trans);
4863
4864         /*
4865          * See if there is enough pinned space to make this reservation, or if
4866          * we have block groups that are going to be freed, allowing us to
4867          * possibly do a chunk allocation the next loop through.
4868          */
4869         if (test_bit(BTRFS_TRANS_HAVE_FREE_BGS, &trans->transaction->flags) ||
4870             __percpu_counter_compare(&space_info->total_bytes_pinned,
4871                                      bytes_needed,
4872                                      BTRFS_TOTAL_BYTES_PINNED_BATCH) >= 0)
4873                 goto commit;
4874
4875         /*
4876          * See if there is some space in the delayed insertion reservation for
4877          * this reservation.
4878          */
4879         if (space_info != delayed_rsv->space_info)
4880                 goto enospc;
4881
4882         spin_lock(&delayed_rsv->lock);
4883         reclaim_bytes += delayed_rsv->reserved;
4884         spin_unlock(&delayed_rsv->lock);
4885
4886         spin_lock(&delayed_refs_rsv->lock);
4887         reclaim_bytes += delayed_refs_rsv->reserved;
4888         spin_unlock(&delayed_refs_rsv->lock);
4889         if (reclaim_bytes >= bytes_needed)
4890                 goto commit;
4891         bytes_needed -= reclaim_bytes;
4892
4893         if (__percpu_counter_compare(&space_info->total_bytes_pinned,
4894                                    bytes_needed,
4895                                    BTRFS_TOTAL_BYTES_PINNED_BATCH) < 0)
4896                 goto enospc;
4897
4898 commit:
4899         return btrfs_commit_transaction(trans);
4900 enospc:
4901         btrfs_end_transaction(trans);
4902         return -ENOSPC;
4903 }
4904
4905 /*
4906  * Try to flush some data based on policy set by @state. This is only advisory
4907  * and may fail for various reasons. The caller is supposed to examine the
4908  * state of @space_info to detect the outcome.
4909  */
4910 static void flush_space(struct btrfs_fs_info *fs_info,
4911                        struct btrfs_space_info *space_info, u64 num_bytes,
4912                        int state)
4913 {
4914         struct btrfs_root *root = fs_info->extent_root;
4915         struct btrfs_trans_handle *trans;
4916         int nr;
4917         int ret = 0;
4918
4919         switch (state) {
4920         case FLUSH_DELAYED_ITEMS_NR:
4921         case FLUSH_DELAYED_ITEMS:
4922                 if (state == FLUSH_DELAYED_ITEMS_NR)
4923                         nr = calc_reclaim_items_nr(fs_info, num_bytes) * 2;
4924                 else
4925                         nr = -1;
4926
4927                 trans = btrfs_join_transaction(root);
4928                 if (IS_ERR(trans)) {
4929                         ret = PTR_ERR(trans);
4930                         break;
4931                 }
4932                 ret = btrfs_run_delayed_items_nr(trans, nr);
4933                 btrfs_end_transaction(trans);
4934                 break;
4935         case FLUSH_DELALLOC:
4936         case FLUSH_DELALLOC_WAIT:
4937                 shrink_delalloc(fs_info, num_bytes * 2, num_bytes,
4938                                 state == FLUSH_DELALLOC_WAIT);
4939                 break;
4940         case FLUSH_DELAYED_REFS_NR:
4941         case FLUSH_DELAYED_REFS:
4942                 trans = btrfs_join_transaction(root);
4943                 if (IS_ERR(trans)) {
4944                         ret = PTR_ERR(trans);
4945                         break;
4946                 }
4947                 if (state == FLUSH_DELAYED_REFS_NR)
4948                         nr = calc_reclaim_items_nr(fs_info, num_bytes);
4949                 else
4950                         nr = 0;
4951                 btrfs_run_delayed_refs(trans, nr);
4952                 btrfs_end_transaction(trans);
4953                 break;
4954         case ALLOC_CHUNK:
4955         case ALLOC_CHUNK_FORCE:
4956                 trans = btrfs_join_transaction(root);
4957                 if (IS_ERR(trans)) {
4958                         ret = PTR_ERR(trans);
4959                         break;
4960                 }
4961                 ret = do_chunk_alloc(trans,
4962                                      btrfs_metadata_alloc_profile(fs_info),
4963                                      (state == ALLOC_CHUNK) ?
4964                                       CHUNK_ALLOC_NO_FORCE : CHUNK_ALLOC_FORCE);
4965                 btrfs_end_transaction(trans);
4966                 if (ret > 0 || ret == -ENOSPC)
4967                         ret = 0;
4968                 break;
4969         case COMMIT_TRANS:
4970                 /*
4971                  * If we have pending delayed iputs then we could free up a
4972                  * bunch of pinned space, so make sure we run the iputs before
4973                  * we do our pinned bytes check below.
4974                  */
4975                 btrfs_run_delayed_iputs(fs_info);
4976                 btrfs_wait_on_delayed_iputs(fs_info);
4977
4978                 ret = may_commit_transaction(fs_info, space_info);
4979                 break;
4980         default:
4981                 ret = -ENOSPC;
4982                 break;
4983         }
4984
4985         trace_btrfs_flush_space(fs_info, space_info->flags, num_bytes, state,
4986                                 ret);
4987         return;
4988 }
4989
4990 static inline u64
4991 btrfs_calc_reclaim_metadata_size(struct btrfs_fs_info *fs_info,
4992                                  struct btrfs_space_info *space_info,
4993                                  bool system_chunk)
4994 {
4995         struct reserve_ticket *ticket;
4996         u64 used;
4997         u64 expected;
4998         u64 to_reclaim = 0;
4999
5000         list_for_each_entry(ticket, &space_info->tickets, list)
5001                 to_reclaim += ticket->bytes;
5002         list_for_each_entry(ticket, &space_info->priority_tickets, list)
5003                 to_reclaim += ticket->bytes;
5004         if (to_reclaim)
5005                 return to_reclaim;
5006
5007         to_reclaim = min_t(u64, num_online_cpus() * SZ_1M, SZ_16M);
5008         if (can_overcommit(fs_info, space_info, to_reclaim,
5009                            BTRFS_RESERVE_FLUSH_ALL, system_chunk))
5010                 return 0;
5011
5012         used = btrfs_space_info_used(space_info, true);
5013
5014         if (can_overcommit(fs_info, space_info, SZ_1M,
5015                            BTRFS_RESERVE_FLUSH_ALL, system_chunk))
5016                 expected = div_factor_fine(space_info->total_bytes, 95);
5017         else
5018                 expected = div_factor_fine(space_info->total_bytes, 90);
5019
5020         if (used > expected)
5021                 to_reclaim = used - expected;
5022         else
5023                 to_reclaim = 0;
5024         to_reclaim = min(to_reclaim, space_info->bytes_may_use +
5025                                      space_info->bytes_reserved);
5026         return to_reclaim;
5027 }
5028
5029 static inline int need_do_async_reclaim(struct btrfs_fs_info *fs_info,
5030                                         struct btrfs_space_info *space_info,
5031                                         u64 used, bool system_chunk)
5032 {
5033         u64 thresh = div_factor_fine(space_info->total_bytes, 98);
5034
5035         /* If we're just plain full then async reclaim just slows us down. */
5036         if ((space_info->bytes_used + space_info->bytes_reserved) >= thresh)
5037                 return 0;
5038
5039         if (!btrfs_calc_reclaim_metadata_size(fs_info, space_info,
5040                                               system_chunk))
5041                 return 0;
5042
5043         return (used >= thresh && !btrfs_fs_closing(fs_info) &&
5044                 !test_bit(BTRFS_FS_STATE_REMOUNTING, &fs_info->fs_state));
5045 }
5046
5047 static bool wake_all_tickets(struct list_head *head)
5048 {
5049         struct reserve_ticket *ticket;
5050
5051         while (!list_empty(head)) {
5052                 ticket = list_first_entry(head, struct reserve_ticket, list);
5053                 list_del_init(&ticket->list);
5054                 ticket->error = -ENOSPC;
5055                 wake_up(&ticket->wait);
5056                 if (ticket->bytes != ticket->orig_bytes)
5057                         return true;
5058         }
5059         return false;
5060 }
5061
5062 /*
5063  * This is for normal flushers, we can wait all goddamned day if we want to.  We
5064  * will loop and continuously try to flush as long as we are making progress.
5065  * We count progress as clearing off tickets each time we have to loop.
5066  */
5067 static void btrfs_async_reclaim_metadata_space(struct work_struct *work)
5068 {
5069         struct btrfs_fs_info *fs_info;
5070         struct btrfs_space_info *space_info;
5071         u64 to_reclaim;
5072         int flush_state;
5073         int commit_cycles = 0;
5074         u64 last_tickets_id;
5075
5076         fs_info = container_of(work, struct btrfs_fs_info, async_reclaim_work);
5077         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
5078
5079         spin_lock(&space_info->lock);
5080         to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info, space_info,
5081                                                       false);
5082         if (!to_reclaim) {
5083                 space_info->flush = 0;
5084                 spin_unlock(&space_info->lock);
5085                 return;
5086         }
5087         last_tickets_id = space_info->tickets_id;
5088         spin_unlock(&space_info->lock);
5089
5090         flush_state = FLUSH_DELAYED_ITEMS_NR;
5091         do {
5092                 flush_space(fs_info, space_info, to_reclaim, flush_state);
5093                 spin_lock(&space_info->lock);
5094                 if (list_empty(&space_info->tickets)) {
5095                         space_info->flush = 0;
5096                         spin_unlock(&space_info->lock);
5097                         return;
5098                 }
5099                 to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info,
5100                                                               space_info,
5101                                                               false);
5102                 if (last_tickets_id == space_info->tickets_id) {
5103                         flush_state++;
5104                 } else {
5105                         last_tickets_id = space_info->tickets_id;
5106                         flush_state = FLUSH_DELAYED_ITEMS_NR;
5107                         if (commit_cycles)
5108                                 commit_cycles--;
5109                 }
5110
5111                 /*
5112                  * We don't want to force a chunk allocation until we've tried
5113                  * pretty hard to reclaim space.  Think of the case where we
5114                  * freed up a bunch of space and so have a lot of pinned space
5115                  * to reclaim.  We would rather use that than possibly create a
5116                  * underutilized metadata chunk.  So if this is our first run
5117                  * through the flushing state machine skip ALLOC_CHUNK_FORCE and
5118                  * commit the transaction.  If nothing has changed the next go
5119                  * around then we can force a chunk allocation.
5120                  */
5121                 if (flush_state == ALLOC_CHUNK_FORCE && !commit_cycles)
5122                         flush_state++;
5123
5124                 if (flush_state > COMMIT_TRANS) {
5125                         commit_cycles++;
5126                         if (commit_cycles > 2) {
5127                                 if (wake_all_tickets(&space_info->tickets)) {
5128                                         flush_state = FLUSH_DELAYED_ITEMS_NR;
5129                                         commit_cycles--;
5130                                 } else {
5131                                         space_info->flush = 0;
5132                                 }
5133                         } else {
5134                                 flush_state = FLUSH_DELAYED_ITEMS_NR;
5135                         }
5136                 }
5137                 spin_unlock(&space_info->lock);
5138         } while (flush_state <= COMMIT_TRANS);
5139 }
5140
5141 void btrfs_init_async_reclaim_work(struct work_struct *work)
5142 {
5143         INIT_WORK(work, btrfs_async_reclaim_metadata_space);
5144 }
5145
5146 static void priority_reclaim_metadata_space(struct btrfs_fs_info *fs_info,
5147                                             struct btrfs_space_info *space_info,
5148                                             struct reserve_ticket *ticket)
5149 {
5150         u64 to_reclaim;
5151         int flush_state = FLUSH_DELAYED_ITEMS_NR;
5152
5153         spin_lock(&space_info->lock);
5154         to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info, space_info,
5155                                                       false);
5156         if (!to_reclaim) {
5157                 spin_unlock(&space_info->lock);
5158                 return;
5159         }
5160         spin_unlock(&space_info->lock);
5161
5162         do {
5163                 flush_space(fs_info, space_info, to_reclaim, flush_state);
5164                 flush_state++;
5165                 spin_lock(&space_info->lock);
5166                 if (ticket->bytes == 0) {
5167                         spin_unlock(&space_info->lock);
5168                         return;
5169                 }
5170                 spin_unlock(&space_info->lock);
5171
5172                 /*
5173                  * Priority flushers can't wait on delalloc without
5174                  * deadlocking.
5175                  */
5176                 if (flush_state == FLUSH_DELALLOC ||
5177                     flush_state == FLUSH_DELALLOC_WAIT)
5178                         flush_state = ALLOC_CHUNK;
5179         } while (flush_state < COMMIT_TRANS);
5180 }
5181
5182 static int wait_reserve_ticket(struct btrfs_fs_info *fs_info,
5183                                struct btrfs_space_info *space_info,
5184                                struct reserve_ticket *ticket)
5185
5186 {
5187         DEFINE_WAIT(wait);
5188         u64 reclaim_bytes = 0;
5189         int ret = 0;
5190
5191         spin_lock(&space_info->lock);
5192         while (ticket->bytes > 0 && ticket->error == 0) {
5193                 ret = prepare_to_wait_event(&ticket->wait, &wait, TASK_KILLABLE);
5194                 if (ret) {
5195                         ret = -EINTR;
5196                         break;
5197                 }
5198                 spin_unlock(&space_info->lock);
5199
5200                 schedule();
5201
5202                 finish_wait(&ticket->wait, &wait);
5203                 spin_lock(&space_info->lock);
5204         }
5205         if (!ret)
5206                 ret = ticket->error;
5207         if (!list_empty(&ticket->list))
5208                 list_del_init(&ticket->list);
5209         if (ticket->bytes && ticket->bytes < ticket->orig_bytes)
5210                 reclaim_bytes = ticket->orig_bytes - ticket->bytes;
5211         spin_unlock(&space_info->lock);
5212
5213         if (reclaim_bytes)
5214                 space_info_add_old_bytes(fs_info, space_info, reclaim_bytes);
5215         return ret;
5216 }
5217
5218 /**
5219  * reserve_metadata_bytes - try to reserve bytes from the block_rsv's space
5220  * @root - the root we're allocating for
5221  * @space_info - the space info we want to allocate from
5222  * @orig_bytes - the number of bytes we want
5223  * @flush - whether or not we can flush to make our reservation
5224  *
5225  * This will reserve orig_bytes number of bytes from the space info associated
5226  * with the block_rsv.  If there is not enough space it will make an attempt to
5227  * flush out space to make room.  It will do this by flushing delalloc if
5228  * possible or committing the transaction.  If flush is 0 then no attempts to
5229  * regain reservations will be made and this will fail if there is not enough
5230  * space already.
5231  */
5232 static int __reserve_metadata_bytes(struct btrfs_fs_info *fs_info,
5233                                     struct btrfs_space_info *space_info,
5234                                     u64 orig_bytes,
5235                                     enum btrfs_reserve_flush_enum flush,
5236                                     bool system_chunk)
5237 {
5238         struct reserve_ticket ticket;
5239         u64 used;
5240         u64 reclaim_bytes = 0;
5241         int ret = 0;
5242
5243         ASSERT(orig_bytes);
5244         ASSERT(!current->journal_info || flush != BTRFS_RESERVE_FLUSH_ALL);
5245
5246         spin_lock(&space_info->lock);
5247         ret = -ENOSPC;
5248         used = btrfs_space_info_used(space_info, true);
5249
5250         /*
5251          * If we have enough space then hooray, make our reservation and carry
5252          * on.  If not see if we can overcommit, and if we can, hooray carry on.
5253          * If not things get more complicated.
5254          */
5255         if (used + orig_bytes <= space_info->total_bytes) {
5256                 update_bytes_may_use(space_info, orig_bytes);
5257                 trace_btrfs_space_reservation(fs_info, "space_info",
5258                                               space_info->flags, orig_bytes, 1);
5259                 ret = 0;
5260         } else if (can_overcommit(fs_info, space_info, orig_bytes, flush,
5261                                   system_chunk)) {
5262                 update_bytes_may_use(space_info, orig_bytes);
5263                 trace_btrfs_space_reservation(fs_info, "space_info",
5264                                               space_info->flags, orig_bytes, 1);
5265                 ret = 0;
5266         }
5267
5268         /*
5269          * If we couldn't make a reservation then setup our reservation ticket
5270          * and kick the async worker if it's not already running.
5271          *
5272          * If we are a priority flusher then we just need to add our ticket to
5273          * the list and we will do our own flushing further down.
5274          */
5275         if (ret && flush != BTRFS_RESERVE_NO_FLUSH) {
5276                 ticket.orig_bytes = orig_bytes;
5277                 ticket.bytes = orig_bytes;
5278                 ticket.error = 0;
5279                 init_waitqueue_head(&ticket.wait);
5280                 if (flush == BTRFS_RESERVE_FLUSH_ALL) {
5281                         list_add_tail(&ticket.list, &space_info->tickets);
5282                         if (!space_info->flush) {
5283                                 space_info->flush = 1;
5284                                 trace_btrfs_trigger_flush(fs_info,
5285                                                           space_info->flags,
5286                                                           orig_bytes, flush,
5287                                                           "enospc");
5288                                 queue_work(system_unbound_wq,
5289                                            &fs_info->async_reclaim_work);
5290                         }
5291                 } else {
5292                         list_add_tail(&ticket.list,
5293                                       &space_info->priority_tickets);
5294                 }
5295         } else if (!ret && space_info->flags & BTRFS_BLOCK_GROUP_METADATA) {
5296                 used += orig_bytes;
5297                 /*
5298                  * We will do the space reservation dance during log replay,
5299                  * which means we won't have fs_info->fs_root set, so don't do
5300                  * the async reclaim as we will panic.
5301                  */
5302                 if (!test_bit(BTRFS_FS_LOG_RECOVERING, &fs_info->flags) &&
5303                     need_do_async_reclaim(fs_info, space_info,
5304                                           used, system_chunk) &&
5305                     !work_busy(&fs_info->async_reclaim_work)) {
5306                         trace_btrfs_trigger_flush(fs_info, space_info->flags,
5307                                                   orig_bytes, flush, "preempt");
5308                         queue_work(system_unbound_wq,
5309                                    &fs_info->async_reclaim_work);
5310                 }
5311         }
5312         spin_unlock(&space_info->lock);
5313         if (!ret || flush == BTRFS_RESERVE_NO_FLUSH)
5314                 return ret;
5315
5316         if (flush == BTRFS_RESERVE_FLUSH_ALL)
5317                 return wait_reserve_ticket(fs_info, space_info, &ticket);
5318
5319         ret = 0;
5320         priority_reclaim_metadata_space(fs_info, space_info, &ticket);
5321         spin_lock(&space_info->lock);
5322         if (ticket.bytes) {
5323                 if (ticket.bytes < orig_bytes)
5324                         reclaim_bytes = orig_bytes - ticket.bytes;
5325                 list_del_init(&ticket.list);
5326                 ret = -ENOSPC;
5327         }
5328         spin_unlock(&space_info->lock);
5329
5330         if (reclaim_bytes)
5331                 space_info_add_old_bytes(fs_info, space_info, reclaim_bytes);
5332         ASSERT(list_empty(&ticket.list));
5333         return ret;
5334 }
5335
5336 /**
5337  * reserve_metadata_bytes - try to reserve bytes from the block_rsv's space
5338  * @root - the root we're allocating for
5339  * @block_rsv - the block_rsv we're allocating for
5340  * @orig_bytes - the number of bytes we want
5341  * @flush - whether or not we can flush to make our reservation
5342  *
5343  * This will reserve orig_bytes number of bytes from the space info associated
5344  * with the block_rsv.  If there is not enough space it will make an attempt to
5345  * flush out space to make room.  It will do this by flushing delalloc if
5346  * possible or committing the transaction.  If flush is 0 then no attempts to
5347  * regain reservations will be made and this will fail if there is not enough
5348  * space already.
5349  */
5350 static int reserve_metadata_bytes(struct btrfs_root *root,
5351                                   struct btrfs_block_rsv *block_rsv,
5352                                   u64 orig_bytes,
5353                                   enum btrfs_reserve_flush_enum flush)
5354 {
5355         struct btrfs_fs_info *fs_info = root->fs_info;
5356         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
5357         int ret;
5358         bool system_chunk = (root == fs_info->chunk_root);
5359
5360         ret = __reserve_metadata_bytes(fs_info, block_rsv->space_info,
5361                                        orig_bytes, flush, system_chunk);
5362         if (ret == -ENOSPC &&
5363             unlikely(root->orphan_cleanup_state == ORPHAN_CLEANUP_STARTED)) {
5364                 if (block_rsv != global_rsv &&
5365                     !block_rsv_use_bytes(global_rsv, orig_bytes))
5366                         ret = 0;
5367         }
5368         if (ret == -ENOSPC) {
5369                 trace_btrfs_space_reservation(fs_info, "space_info:enospc",
5370                                               block_rsv->space_info->flags,
5371                                               orig_bytes, 1);
5372
5373                 if (btrfs_test_opt(fs_info, ENOSPC_DEBUG))
5374                         dump_space_info(fs_info, block_rsv->space_info,
5375                                         orig_bytes, 0);
5376         }
5377         return ret;
5378 }
5379
5380 static struct btrfs_block_rsv *get_block_rsv(
5381                                         const struct btrfs_trans_handle *trans,
5382                                         const struct btrfs_root *root)
5383 {
5384         struct btrfs_fs_info *fs_info = root->fs_info;
5385         struct btrfs_block_rsv *block_rsv = NULL;
5386
5387         if (test_bit(BTRFS_ROOT_REF_COWS, &root->state) ||
5388             (root == fs_info->csum_root && trans->adding_csums) ||
5389             (root == fs_info->uuid_root))
5390                 block_rsv = trans->block_rsv;
5391
5392         if (!block_rsv)
5393                 block_rsv = root->block_rsv;
5394
5395         if (!block_rsv)
5396                 block_rsv = &fs_info->empty_block_rsv;
5397
5398         return block_rsv;
5399 }
5400
5401 static int block_rsv_use_bytes(struct btrfs_block_rsv *block_rsv,
5402                                u64 num_bytes)
5403 {
5404         int ret = -ENOSPC;
5405         spin_lock(&block_rsv->lock);
5406         if (block_rsv->reserved >= num_bytes) {
5407                 block_rsv->reserved -= num_bytes;
5408                 if (block_rsv->reserved < block_rsv->size)
5409                         block_rsv->full = 0;
5410                 ret = 0;
5411         }
5412         spin_unlock(&block_rsv->lock);
5413         return ret;
5414 }
5415
5416 static void block_rsv_add_bytes(struct btrfs_block_rsv *block_rsv,
5417                                 u64 num_bytes, bool update_size)
5418 {
5419         spin_lock(&block_rsv->lock);
5420         block_rsv->reserved += num_bytes;
5421         if (update_size)
5422                 block_rsv->size += num_bytes;
5423         else if (block_rsv->reserved >= block_rsv->size)
5424                 block_rsv->full = 1;
5425         spin_unlock(&block_rsv->lock);
5426 }
5427
5428 int btrfs_cond_migrate_bytes(struct btrfs_fs_info *fs_info,
5429                              struct btrfs_block_rsv *dest, u64 num_bytes,
5430                              int min_factor)
5431 {
5432         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
5433         u64 min_bytes;
5434
5435         if (global_rsv->space_info != dest->space_info)
5436                 return -ENOSPC;
5437
5438         spin_lock(&global_rsv->lock);
5439         min_bytes = div_factor(global_rsv->size, min_factor);
5440         if (global_rsv->reserved < min_bytes + num_bytes) {
5441                 spin_unlock(&global_rsv->lock);
5442                 return -ENOSPC;
5443         }
5444         global_rsv->reserved -= num_bytes;
5445         if (global_rsv->reserved < global_rsv->size)
5446                 global_rsv->full = 0;
5447         spin_unlock(&global_rsv->lock);
5448
5449         block_rsv_add_bytes(dest, num_bytes, true);
5450         return 0;
5451 }
5452
5453 /**
5454  * btrfs_migrate_to_delayed_refs_rsv - transfer bytes to our delayed refs rsv.
5455  * @fs_info - the fs info for our fs.
5456  * @src - the source block rsv to transfer from.
5457  * @num_bytes - the number of bytes to transfer.
5458  *
5459  * This transfers up to the num_bytes amount from the src rsv to the
5460  * delayed_refs_rsv.  Any extra bytes are returned to the space info.
5461  */
5462 void btrfs_migrate_to_delayed_refs_rsv(struct btrfs_fs_info *fs_info,
5463                                        struct btrfs_block_rsv *src,
5464                                        u64 num_bytes)
5465 {
5466         struct btrfs_block_rsv *delayed_refs_rsv = &fs_info->delayed_refs_rsv;
5467         u64 to_free = 0;
5468
5469         spin_lock(&src->lock);
5470         src->reserved -= num_bytes;
5471         src->size -= num_bytes;
5472         spin_unlock(&src->lock);
5473
5474         spin_lock(&delayed_refs_rsv->lock);
5475         if (delayed_refs_rsv->size > delayed_refs_rsv->reserved) {
5476                 u64 delta = delayed_refs_rsv->size -
5477                         delayed_refs_rsv->reserved;
5478                 if (num_bytes > delta) {
5479                         to_free = num_bytes - delta;
5480                         num_bytes = delta;
5481                 }
5482         } else {
5483                 to_free = num_bytes;
5484                 num_bytes = 0;
5485         }
5486
5487         if (num_bytes)
5488                 delayed_refs_rsv->reserved += num_bytes;
5489         if (delayed_refs_rsv->reserved >= delayed_refs_rsv->size)
5490                 delayed_refs_rsv->full = 1;
5491         spin_unlock(&delayed_refs_rsv->lock);
5492
5493         if (num_bytes)
5494                 trace_btrfs_space_reservation(fs_info, "delayed_refs_rsv",
5495                                               0, num_bytes, 1);
5496         if (to_free)
5497                 space_info_add_old_bytes(fs_info, delayed_refs_rsv->space_info,
5498                                          to_free);
5499 }
5500
5501 /**
5502  * btrfs_delayed_refs_rsv_refill - refill based on our delayed refs usage.
5503  * @fs_info - the fs_info for our fs.
5504  * @flush - control how we can flush for this reservation.
5505  *
5506  * This will refill the delayed block_rsv up to 1 items size worth of space and
5507  * will return -ENOSPC if we can't make the reservation.
5508  */
5509 int btrfs_delayed_refs_rsv_refill(struct btrfs_fs_info *fs_info,
5510                                   enum btrfs_reserve_flush_enum flush)
5511 {
5512         struct btrfs_block_rsv *block_rsv = &fs_info->delayed_refs_rsv;
5513         u64 limit = btrfs_calc_trans_metadata_size(fs_info, 1);
5514         u64 num_bytes = 0;
5515         int ret = -ENOSPC;
5516
5517         spin_lock(&block_rsv->lock);
5518         if (block_rsv->reserved < block_rsv->size) {
5519                 num_bytes = block_rsv->size - block_rsv->reserved;
5520                 num_bytes = min(num_bytes, limit);
5521         }
5522         spin_unlock(&block_rsv->lock);
5523
5524         if (!num_bytes)
5525                 return 0;
5526
5527         ret = reserve_metadata_bytes(fs_info->extent_root, block_rsv,
5528                                      num_bytes, flush);
5529         if (ret)
5530                 return ret;
5531         block_rsv_add_bytes(block_rsv, num_bytes, 0);
5532         trace_btrfs_space_reservation(fs_info, "delayed_refs_rsv",
5533                                       0, num_bytes, 1);
5534         return 0;
5535 }
5536
5537 /*
5538  * This is for space we already have accounted in space_info->bytes_may_use, so
5539  * basically when we're returning space from block_rsv's.
5540  */
5541 static void space_info_add_old_bytes(struct btrfs_fs_info *fs_info,
5542                                      struct btrfs_space_info *space_info,
5543                                      u64 num_bytes)
5544 {
5545         struct reserve_ticket *ticket;
5546         struct list_head *head;
5547         u64 used;
5548         enum btrfs_reserve_flush_enum flush = BTRFS_RESERVE_NO_FLUSH;
5549         bool check_overcommit = false;
5550
5551         spin_lock(&space_info->lock);
5552         head = &space_info->priority_tickets;
5553
5554         /*
5555          * If we are over our limit then we need to check and see if we can
5556          * overcommit, and if we can't then we just need to free up our space
5557          * and not satisfy any requests.
5558          */
5559         used = btrfs_space_info_used(space_info, true);
5560         if (used - num_bytes >= space_info->total_bytes)
5561                 check_overcommit = true;
5562 again:
5563         while (!list_empty(head) && num_bytes) {
5564                 ticket = list_first_entry(head, struct reserve_ticket,
5565                                           list);
5566                 /*
5567                  * We use 0 bytes because this space is already reserved, so
5568                  * adding the ticket space would be a double count.
5569                  */
5570                 if (check_overcommit &&
5571                     !can_overcommit(fs_info, space_info, 0, flush, false))
5572                         break;
5573                 if (num_bytes >= ticket->bytes) {
5574                         list_del_init(&ticket->list);
5575                         num_bytes -= ticket->bytes;
5576                         ticket->bytes = 0;
5577                         space_info->tickets_id++;
5578                         wake_up(&ticket->wait);
5579                 } else {
5580                         ticket->bytes -= num_bytes;
5581                         num_bytes = 0;
5582                 }
5583         }
5584
5585         if (num_bytes && head == &space_info->priority_tickets) {
5586                 head = &space_info->tickets;
5587                 flush = BTRFS_RESERVE_FLUSH_ALL;
5588                 goto again;
5589         }
5590         update_bytes_may_use(space_info, -num_bytes);
5591         trace_btrfs_space_reservation(fs_info, "space_info",
5592                                       space_info->flags, num_bytes, 0);
5593         spin_unlock(&space_info->lock);
5594 }
5595
5596 /*
5597  * This is for newly allocated space that isn't accounted in
5598  * space_info->bytes_may_use yet.  So if we allocate a chunk or unpin an extent
5599  * we use this helper.
5600  */
5601 static void space_info_add_new_bytes(struct btrfs_fs_info *fs_info,
5602                                      struct btrfs_space_info *space_info,
5603                                      u64 num_bytes)
5604 {
5605         struct reserve_ticket *ticket;
5606         struct list_head *head = &space_info->priority_tickets;
5607
5608 again:
5609         while (!list_empty(head) && num_bytes) {
5610                 ticket = list_first_entry(head, struct reserve_ticket,
5611                                           list);
5612                 if (num_bytes >= ticket->bytes) {
5613                         trace_btrfs_space_reservation(fs_info, "space_info",
5614                                                       space_info->flags,
5615                                                       ticket->bytes, 1);
5616                         list_del_init(&ticket->list);
5617                         num_bytes -= ticket->bytes;
5618                         update_bytes_may_use(space_info, ticket->bytes);
5619                         ticket->bytes = 0;
5620                         space_info->tickets_id++;
5621                         wake_up(&ticket->wait);
5622                 } else {
5623                         trace_btrfs_space_reservation(fs_info, "space_info",
5624                                                       space_info->flags,
5625                                                       num_bytes, 1);
5626                         update_bytes_may_use(space_info, num_bytes);
5627                         ticket->bytes -= num_bytes;
5628                         num_bytes = 0;
5629                 }
5630         }
5631
5632         if (num_bytes && head == &space_info->priority_tickets) {
5633                 head = &space_info->tickets;
5634                 goto again;
5635         }
5636 }
5637
5638 static u64 block_rsv_release_bytes(struct btrfs_fs_info *fs_info,
5639                                     struct btrfs_block_rsv *block_rsv,
5640                                     struct btrfs_block_rsv *dest, u64 num_bytes,
5641                                     u64 *qgroup_to_release_ret)
5642 {
5643         struct btrfs_space_info *space_info = block_rsv->space_info;
5644         u64 qgroup_to_release = 0;
5645         u64 ret;
5646
5647         spin_lock(&block_rsv->lock);
5648         if (num_bytes == (u64)-1) {
5649                 num_bytes = block_rsv->size;
5650                 qgroup_to_release = block_rsv->qgroup_rsv_size;
5651         }
5652         block_rsv->size -= num_bytes;
5653         if (block_rsv->reserved >= block_rsv->size) {
5654                 num_bytes = block_rsv->reserved - block_rsv->size;
5655                 block_rsv->reserved = block_rsv->size;
5656                 block_rsv->full = 1;
5657         } else {
5658                 num_bytes = 0;
5659         }
5660         if (block_rsv->qgroup_rsv_reserved >= block_rsv->qgroup_rsv_size) {
5661                 qgroup_to_release = block_rsv->qgroup_rsv_reserved -
5662                                     block_rsv->qgroup_rsv_size;
5663                 block_rsv->qgroup_rsv_reserved = block_rsv->qgroup_rsv_size;
5664         } else {
5665                 qgroup_to_release = 0;
5666         }
5667         spin_unlock(&block_rsv->lock);
5668
5669         ret = num_bytes;
5670         if (num_bytes > 0) {
5671                 if (dest) {
5672                         spin_lock(&dest->lock);
5673                         if (!dest->full) {
5674                                 u64 bytes_to_add;
5675
5676                                 bytes_to_add = dest->size - dest->reserved;
5677                                 bytes_to_add = min(num_bytes, bytes_to_add);
5678                                 dest->reserved += bytes_to_add;
5679                                 if (dest->reserved >= dest->size)
5680                                         dest->full = 1;
5681                                 num_bytes -= bytes_to_add;
5682                         }
5683                         spin_unlock(&dest->lock);
5684                 }
5685                 if (num_bytes)
5686                         space_info_add_old_bytes(fs_info, space_info,
5687                                                  num_bytes);
5688         }
5689         if (qgroup_to_release_ret)
5690                 *qgroup_to_release_ret = qgroup_to_release;
5691         return ret;
5692 }
5693
5694 int btrfs_block_rsv_migrate(struct btrfs_block_rsv *src,
5695                             struct btrfs_block_rsv *dst, u64 num_bytes,
5696                             bool update_size)
5697 {
5698         int ret;
5699
5700         ret = block_rsv_use_bytes(src, num_bytes);
5701         if (ret)
5702                 return ret;
5703
5704         block_rsv_add_bytes(dst, num_bytes, update_size);
5705         return 0;
5706 }
5707
5708 void btrfs_init_block_rsv(struct btrfs_block_rsv *rsv, unsigned short type)
5709 {
5710         memset(rsv, 0, sizeof(*rsv));
5711         spin_lock_init(&rsv->lock);
5712         rsv->type = type;
5713 }
5714
5715 void btrfs_init_metadata_block_rsv(struct btrfs_fs_info *fs_info,
5716                                    struct btrfs_block_rsv *rsv,
5717                                    unsigned short type)
5718 {
5719         btrfs_init_block_rsv(rsv, type);
5720         rsv->space_info = __find_space_info(fs_info,
5721                                             BTRFS_BLOCK_GROUP_METADATA);
5722 }
5723
5724 struct btrfs_block_rsv *btrfs_alloc_block_rsv(struct btrfs_fs_info *fs_info,
5725                                               unsigned short type)
5726 {
5727         struct btrfs_block_rsv *block_rsv;
5728
5729         block_rsv = kmalloc(sizeof(*block_rsv), GFP_NOFS);
5730         if (!block_rsv)
5731                 return NULL;
5732
5733         btrfs_init_metadata_block_rsv(fs_info, block_rsv, type);
5734         return block_rsv;
5735 }
5736
5737 void btrfs_free_block_rsv(struct btrfs_fs_info *fs_info,
5738                           struct btrfs_block_rsv *rsv)
5739 {
5740         if (!rsv)
5741                 return;
5742         btrfs_block_rsv_release(fs_info, rsv, (u64)-1);
5743         kfree(rsv);
5744 }
5745
5746 int btrfs_block_rsv_add(struct btrfs_root *root,
5747                         struct btrfs_block_rsv *block_rsv, u64 num_bytes,
5748                         enum btrfs_reserve_flush_enum flush)
5749 {
5750         int ret;
5751
5752         if (num_bytes == 0)
5753                 return 0;
5754
5755         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
5756         if (!ret)
5757                 block_rsv_add_bytes(block_rsv, num_bytes, true);
5758
5759         return ret;
5760 }
5761
5762 int btrfs_block_rsv_check(struct btrfs_block_rsv *block_rsv, int min_factor)
5763 {
5764         u64 num_bytes = 0;
5765         int ret = -ENOSPC;
5766
5767         if (!block_rsv)
5768                 return 0;
5769
5770         spin_lock(&block_rsv->lock);
5771         num_bytes = div_factor(block_rsv->size, min_factor);
5772         if (block_rsv->reserved >= num_bytes)
5773                 ret = 0;
5774         spin_unlock(&block_rsv->lock);
5775
5776         return ret;
5777 }
5778
5779 int btrfs_block_rsv_refill(struct btrfs_root *root,
5780                            struct btrfs_block_rsv *block_rsv, u64 min_reserved,
5781                            enum btrfs_reserve_flush_enum flush)
5782 {
5783         u64 num_bytes = 0;
5784         int ret = -ENOSPC;
5785
5786         if (!block_rsv)
5787                 return 0;
5788
5789         spin_lock(&block_rsv->lock);
5790         num_bytes = min_reserved;
5791         if (block_rsv->reserved >= num_bytes)
5792                 ret = 0;
5793         else
5794                 num_bytes -= block_rsv->reserved;
5795         spin_unlock(&block_rsv->lock);
5796
5797         if (!ret)
5798                 return 0;
5799
5800         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
5801         if (!ret) {
5802                 block_rsv_add_bytes(block_rsv, num_bytes, false);
5803                 return 0;
5804         }
5805
5806         return ret;
5807 }
5808
5809 /**
5810  * btrfs_inode_rsv_refill - refill the inode block rsv.
5811  * @inode - the inode we are refilling.
5812  * @flush - the flushing restriction.
5813  *
5814  * Essentially the same as btrfs_block_rsv_refill, except it uses the
5815  * block_rsv->size as the minimum size.  We'll either refill the missing amount
5816  * or return if we already have enough space.  This will also handle the reserve
5817  * tracepoint for the reserved amount.
5818  */
5819 static int btrfs_inode_rsv_refill(struct btrfs_inode *inode,
5820                                   enum btrfs_reserve_flush_enum flush)
5821 {
5822         struct btrfs_root *root = inode->root;
5823         struct btrfs_block_rsv *block_rsv = &inode->block_rsv;
5824         u64 num_bytes = 0;
5825         u64 qgroup_num_bytes = 0;
5826         int ret = -ENOSPC;
5827
5828         spin_lock(&block_rsv->lock);
5829         if (block_rsv->reserved < block_rsv->size)
5830                 num_bytes = block_rsv->size - block_rsv->reserved;
5831         if (block_rsv->qgroup_rsv_reserved < block_rsv->qgroup_rsv_size)
5832                 qgroup_num_bytes = block_rsv->qgroup_rsv_size -
5833                                    block_rsv->qgroup_rsv_reserved;
5834         spin_unlock(&block_rsv->lock);
5835
5836         if (num_bytes == 0)
5837                 return 0;
5838
5839         ret = btrfs_qgroup_reserve_meta_prealloc(root, qgroup_num_bytes, true);
5840         if (ret)
5841                 return ret;
5842         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
5843         if (!ret) {
5844                 block_rsv_add_bytes(block_rsv, num_bytes, false);
5845                 trace_btrfs_space_reservation(root->fs_info, "delalloc",
5846                                               btrfs_ino(inode), num_bytes, 1);
5847
5848                 /* Don't forget to increase qgroup_rsv_reserved */
5849                 spin_lock(&block_rsv->lock);
5850                 block_rsv->qgroup_rsv_reserved += qgroup_num_bytes;
5851                 spin_unlock(&block_rsv->lock);
5852         } else
5853                 btrfs_qgroup_free_meta_prealloc(root, qgroup_num_bytes);
5854         return ret;
5855 }
5856
5857 static u64 __btrfs_block_rsv_release(struct btrfs_fs_info *fs_info,
5858                                      struct btrfs_block_rsv *block_rsv,
5859                                      u64 num_bytes, u64 *qgroup_to_release)
5860 {
5861         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
5862         struct btrfs_block_rsv *delayed_rsv = &fs_info->delayed_refs_rsv;
5863         struct btrfs_block_rsv *target = delayed_rsv;
5864
5865         if (target->full || target == block_rsv)
5866                 target = global_rsv;
5867
5868         if (block_rsv->space_info != target->space_info)
5869                 target = NULL;
5870
5871         return block_rsv_release_bytes(fs_info, block_rsv, target, num_bytes,
5872                                        qgroup_to_release);
5873 }
5874
5875 void btrfs_block_rsv_release(struct btrfs_fs_info *fs_info,
5876                              struct btrfs_block_rsv *block_rsv,
5877                              u64 num_bytes)
5878 {
5879         __btrfs_block_rsv_release(fs_info, block_rsv, num_bytes, NULL);
5880 }
5881
5882 /**
5883  * btrfs_inode_rsv_release - release any excessive reservation.
5884  * @inode - the inode we need to release from.
5885  * @qgroup_free - free or convert qgroup meta.
5886  *   Unlike normal operation, qgroup meta reservation needs to know if we are
5887  *   freeing qgroup reservation or just converting it into per-trans.  Normally
5888  *   @qgroup_free is true for error handling, and false for normal release.
5889  *
5890  * This is the same as btrfs_block_rsv_release, except that it handles the
5891  * tracepoint for the reservation.
5892  */
5893 static void btrfs_inode_rsv_release(struct btrfs_inode *inode, bool qgroup_free)
5894 {
5895         struct btrfs_fs_info *fs_info = inode->root->fs_info;
5896         struct btrfs_block_rsv *block_rsv = &inode->block_rsv;
5897         u64 released = 0;
5898         u64 qgroup_to_release = 0;
5899
5900         /*
5901          * Since we statically set the block_rsv->size we just want to say we
5902          * are releasing 0 bytes, and then we'll just get the reservation over
5903          * the size free'd.
5904          */
5905         released = __btrfs_block_rsv_release(fs_info, block_rsv, 0,
5906                                              &qgroup_to_release);
5907         if (released > 0)
5908                 trace_btrfs_space_reservation(fs_info, "delalloc",
5909                                               btrfs_ino(inode), released, 0);
5910         if (qgroup_free)
5911                 btrfs_qgroup_free_meta_prealloc(inode->root, qgroup_to_release);
5912         else
5913                 btrfs_qgroup_convert_reserved_meta(inode->root,
5914                                                    qgroup_to_release);
5915 }
5916
5917 /**
5918  * btrfs_delayed_refs_rsv_release - release a ref head's reservation.
5919  * @fs_info - the fs_info for our fs.
5920  * @nr - the number of items to drop.
5921  *
5922  * This drops the delayed ref head's count from the delayed refs rsv and frees
5923  * any excess reservation we had.
5924  */
5925 void btrfs_delayed_refs_rsv_release(struct btrfs_fs_info *fs_info, int nr)
5926 {
5927         struct btrfs_block_rsv *block_rsv = &fs_info->delayed_refs_rsv;
5928         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
5929         u64 num_bytes = btrfs_calc_trans_metadata_size(fs_info, nr);
5930         u64 released = 0;
5931
5932         released = block_rsv_release_bytes(fs_info, block_rsv, global_rsv,
5933                                            num_bytes, NULL);
5934         if (released)
5935                 trace_btrfs_space_reservation(fs_info, "delayed_refs_rsv",
5936                                               0, released, 0);
5937 }
5938
5939 static void update_global_block_rsv(struct btrfs_fs_info *fs_info)
5940 {
5941         struct btrfs_block_rsv *block_rsv = &fs_info->global_block_rsv;
5942         struct btrfs_space_info *sinfo = block_rsv->space_info;
5943         u64 num_bytes;
5944
5945         /*
5946          * The global block rsv is based on the size of the extent tree, the
5947          * checksum tree and the root tree.  If the fs is empty we want to set
5948          * it to a minimal amount for safety.
5949          */
5950         num_bytes = btrfs_root_used(&fs_info->extent_root->root_item) +
5951                 btrfs_root_used(&fs_info->csum_root->root_item) +
5952                 btrfs_root_used(&fs_info->tree_root->root_item);
5953         num_bytes = max_t(u64, num_bytes, SZ_16M);
5954
5955         spin_lock(&sinfo->lock);
5956         spin_lock(&block_rsv->lock);
5957
5958         block_rsv->size = min_t(u64, num_bytes, SZ_512M);
5959
5960         if (block_rsv->reserved < block_rsv->size) {
5961                 num_bytes = btrfs_space_info_used(sinfo, true);
5962                 if (sinfo->total_bytes > num_bytes) {
5963                         num_bytes = sinfo->total_bytes - num_bytes;
5964                         num_bytes = min(num_bytes,
5965                                         block_rsv->size - block_rsv->reserved);
5966                         block_rsv->reserved += num_bytes;
5967                         update_bytes_may_use(sinfo, num_bytes);
5968                         trace_btrfs_space_reservation(fs_info, "space_info",
5969                                                       sinfo->flags, num_bytes,
5970                                                       1);
5971                 }
5972         } else if (block_rsv->reserved > block_rsv->size) {
5973                 num_bytes = block_rsv->reserved - block_rsv->size;
5974                 update_bytes_may_use(sinfo, -num_bytes);
5975                 trace_btrfs_space_reservation(fs_info, "space_info",
5976                                       sinfo->flags, num_bytes, 0);
5977                 block_rsv->reserved = block_rsv->size;
5978         }
5979
5980         if (block_rsv->reserved == block_rsv->size)
5981                 block_rsv->full = 1;
5982         else
5983                 block_rsv->full = 0;
5984
5985         spin_unlock(&block_rsv->lock);
5986         spin_unlock(&sinfo->lock);
5987 }
5988
5989 static void init_global_block_rsv(struct btrfs_fs_info *fs_info)
5990 {
5991         struct btrfs_space_info *space_info;
5992
5993         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
5994         fs_info->chunk_block_rsv.space_info = space_info;
5995
5996         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
5997         fs_info->global_block_rsv.space_info = space_info;
5998         fs_info->trans_block_rsv.space_info = space_info;
5999         fs_info->empty_block_rsv.space_info = space_info;
6000         fs_info->delayed_block_rsv.space_info = space_info;
6001         fs_info->delayed_refs_rsv.space_info = space_info;
6002
6003         fs_info->extent_root->block_rsv = &fs_info->delayed_refs_rsv;
6004         fs_info->csum_root->block_rsv = &fs_info->delayed_refs_rsv;
6005         fs_info->dev_root->block_rsv = &fs_info->global_block_rsv;
6006         fs_info->tree_root->block_rsv = &fs_info->global_block_rsv;
6007         if (fs_info->quota_root)
6008                 fs_info->quota_root->block_rsv = &fs_info->global_block_rsv;
6009         fs_info->chunk_root->block_rsv = &fs_info->chunk_block_rsv;
6010
6011         update_global_block_rsv(fs_info);
6012 }
6013
6014 static void release_global_block_rsv(struct btrfs_fs_info *fs_info)
6015 {
6016         block_rsv_release_bytes(fs_info, &fs_info->global_block_rsv, NULL,
6017                                 (u64)-1, NULL);
6018         WARN_ON(fs_info->trans_block_rsv.size > 0);
6019         WARN_ON(fs_info->trans_block_rsv.reserved > 0);
6020         WARN_ON(fs_info->chunk_block_rsv.size > 0);
6021         WARN_ON(fs_info->chunk_block_rsv.reserved > 0);
6022         WARN_ON(fs_info->delayed_block_rsv.size > 0);
6023         WARN_ON(fs_info->delayed_block_rsv.reserved > 0);
6024         WARN_ON(fs_info->delayed_refs_rsv.reserved > 0);
6025         WARN_ON(fs_info->delayed_refs_rsv.size > 0);
6026 }
6027
6028 /*
6029  * btrfs_update_delayed_refs_rsv - adjust the size of the delayed refs rsv
6030  * @trans - the trans that may have generated delayed refs
6031  *
6032  * This is to be called anytime we may have adjusted trans->delayed_ref_updates,
6033  * it'll calculate the additional size and add it to the delayed_refs_rsv.
6034  */
6035 void btrfs_update_delayed_refs_rsv(struct btrfs_trans_handle *trans)
6036 {
6037         struct btrfs_fs_info *fs_info = trans->fs_info;
6038         struct btrfs_block_rsv *delayed_rsv = &fs_info->delayed_refs_rsv;
6039         u64 num_bytes;
6040
6041         if (!trans->delayed_ref_updates)
6042                 return;
6043
6044         num_bytes = btrfs_calc_trans_metadata_size(fs_info,
6045                                                    trans->delayed_ref_updates);
6046         spin_lock(&delayed_rsv->lock);
6047         delayed_rsv->size += num_bytes;
6048         delayed_rsv->full = 0;
6049         spin_unlock(&delayed_rsv->lock);
6050         trans->delayed_ref_updates = 0;
6051 }
6052
6053 /*
6054  * To be called after all the new block groups attached to the transaction
6055  * handle have been created (btrfs_create_pending_block_groups()).
6056  */
6057 void btrfs_trans_release_chunk_metadata(struct btrfs_trans_handle *trans)
6058 {
6059         struct btrfs_fs_info *fs_info = trans->fs_info;
6060
6061         if (!trans->chunk_bytes_reserved)
6062                 return;
6063
6064         WARN_ON_ONCE(!list_empty(&trans->new_bgs));
6065
6066         block_rsv_release_bytes(fs_info, &fs_info->chunk_block_rsv, NULL,
6067                                 trans->chunk_bytes_reserved, NULL);
6068         trans->chunk_bytes_reserved = 0;
6069 }
6070
6071 /*
6072  * btrfs_subvolume_reserve_metadata() - reserve space for subvolume operation
6073  * root: the root of the parent directory
6074  * rsv: block reservation
6075  * items: the number of items that we need do reservation
6076  * use_global_rsv: allow fallback to the global block reservation
6077  *
6078  * This function is used to reserve the space for snapshot/subvolume
6079  * creation and deletion. Those operations are different with the
6080  * common file/directory operations, they change two fs/file trees
6081  * and root tree, the number of items that the qgroup reserves is
6082  * different with the free space reservation. So we can not use
6083  * the space reservation mechanism in start_transaction().
6084  */
6085 int btrfs_subvolume_reserve_metadata(struct btrfs_root *root,
6086                                      struct btrfs_block_rsv *rsv, int items,
6087                                      bool use_global_rsv)
6088 {
6089         u64 qgroup_num_bytes = 0;
6090         u64 num_bytes;
6091         int ret;
6092         struct btrfs_fs_info *fs_info = root->fs_info;
6093         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
6094
6095         if (test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags)) {
6096                 /* One for parent inode, two for dir entries */
6097                 qgroup_num_bytes = 3 * fs_info->nodesize;
6098                 ret = btrfs_qgroup_reserve_meta_prealloc(root,
6099                                 qgroup_num_bytes, true);
6100                 if (ret)
6101                         return ret;
6102         }
6103
6104         num_bytes = btrfs_calc_trans_metadata_size(fs_info, items);
6105         rsv->space_info = __find_space_info(fs_info,
6106                                             BTRFS_BLOCK_GROUP_METADATA);
6107         ret = btrfs_block_rsv_add(root, rsv, num_bytes,
6108                                   BTRFS_RESERVE_FLUSH_ALL);
6109
6110         if (ret == -ENOSPC && use_global_rsv)
6111                 ret = btrfs_block_rsv_migrate(global_rsv, rsv, num_bytes, true);
6112
6113         if (ret && qgroup_num_bytes)
6114                 btrfs_qgroup_free_meta_prealloc(root, qgroup_num_bytes);
6115
6116         return ret;
6117 }
6118
6119 void btrfs_subvolume_release_metadata(struct btrfs_fs_info *fs_info,
6120                                       struct btrfs_block_rsv *rsv)
6121 {
6122         btrfs_block_rsv_release(fs_info, rsv, (u64)-1);
6123 }
6124
6125 static void btrfs_calculate_inode_block_rsv_size(struct btrfs_fs_info *fs_info,
6126                                                  struct btrfs_inode *inode)
6127 {
6128         struct btrfs_block_rsv *block_rsv = &inode->block_rsv;
6129         u64 reserve_size = 0;
6130         u64 qgroup_rsv_size = 0;
6131         u64 csum_leaves;
6132         unsigned outstanding_extents;
6133
6134         lockdep_assert_held(&inode->lock);
6135         outstanding_extents = inode->outstanding_extents;
6136         if (outstanding_extents)
6137                 reserve_size = btrfs_calc_trans_metadata_size(fs_info,
6138                                                 outstanding_extents + 1);
6139         csum_leaves = btrfs_csum_bytes_to_leaves(fs_info,
6140                                                  inode->csum_bytes);
6141         reserve_size += btrfs_calc_trans_metadata_size(fs_info,
6142                                                        csum_leaves);
6143         /*
6144          * For qgroup rsv, the calculation is very simple:
6145          * account one nodesize for each outstanding extent
6146          *
6147          * This is overestimating in most cases.
6148          */
6149         qgroup_rsv_size = outstanding_extents * fs_info->nodesize;
6150
6151         spin_lock(&block_rsv->lock);
6152         block_rsv->size = reserve_size;
6153         block_rsv->qgroup_rsv_size = qgroup_rsv_size;
6154         spin_unlock(&block_rsv->lock);
6155 }
6156
6157 int btrfs_delalloc_reserve_metadata(struct btrfs_inode *inode, u64 num_bytes)
6158 {
6159         struct btrfs_fs_info *fs_info = inode->root->fs_info;
6160         unsigned nr_extents;
6161         enum btrfs_reserve_flush_enum flush = BTRFS_RESERVE_FLUSH_ALL;
6162         int ret = 0;
6163         bool delalloc_lock = true;
6164
6165         /* If we are a free space inode we need to not flush since we will be in
6166          * the middle of a transaction commit.  We also don't need the delalloc
6167          * mutex since we won't race with anybody.  We need this mostly to make
6168          * lockdep shut its filthy mouth.
6169          *
6170          * If we have a transaction open (can happen if we call truncate_block
6171          * from truncate), then we need FLUSH_LIMIT so we don't deadlock.
6172          */
6173         if (btrfs_is_free_space_inode(inode)) {
6174                 flush = BTRFS_RESERVE_NO_FLUSH;
6175                 delalloc_lock = false;
6176         } else {
6177                 if (current->journal_info)
6178                         flush = BTRFS_RESERVE_FLUSH_LIMIT;
6179
6180                 if (btrfs_transaction_in_commit(fs_info))
6181                         schedule_timeout(1);
6182         }
6183
6184         if (delalloc_lock)
6185                 mutex_lock(&inode->delalloc_mutex);
6186
6187         num_bytes = ALIGN(num_bytes, fs_info->sectorsize);
6188
6189         /* Add our new extents and calculate the new rsv size. */
6190         spin_lock(&inode->lock);
6191         nr_extents = count_max_extents(num_bytes);
6192         btrfs_mod_outstanding_extents(inode, nr_extents);
6193         inode->csum_bytes += num_bytes;
6194         btrfs_calculate_inode_block_rsv_size(fs_info, inode);
6195         spin_unlock(&inode->lock);
6196
6197         ret = btrfs_inode_rsv_refill(inode, flush);
6198         if (unlikely(ret))
6199                 goto out_fail;
6200
6201         if (delalloc_lock)
6202                 mutex_unlock(&inode->delalloc_mutex);
6203         return 0;
6204
6205 out_fail:
6206         spin_lock(&inode->lock);
6207         nr_extents = count_max_extents(num_bytes);
6208         btrfs_mod_outstanding_extents(inode, -nr_extents);
6209         inode->csum_bytes -= num_bytes;
6210         btrfs_calculate_inode_block_rsv_size(fs_info, inode);
6211         spin_unlock(&inode->lock);
6212
6213         btrfs_inode_rsv_release(inode, true);
6214         if (delalloc_lock)
6215                 mutex_unlock(&inode->delalloc_mutex);
6216         return ret;
6217 }
6218
6219 /**
6220  * btrfs_delalloc_release_metadata - release a metadata reservation for an inode
6221  * @inode: the inode to release the reservation for.
6222  * @num_bytes: the number of bytes we are releasing.
6223  * @qgroup_free: free qgroup reservation or convert it to per-trans reservation
6224  *
6225  * This will release the metadata reservation for an inode.  This can be called
6226  * once we complete IO for a given set of bytes to release their metadata
6227  * reservations, or on error for the same reason.
6228  */
6229 void btrfs_delalloc_release_metadata(struct btrfs_inode *inode, u64 num_bytes,
6230                                      bool qgroup_free)
6231 {
6232         struct btrfs_fs_info *fs_info = inode->root->fs_info;
6233
6234         num_bytes = ALIGN(num_bytes, fs_info->sectorsize);
6235         spin_lock(&inode->lock);
6236         inode->csum_bytes -= num_bytes;
6237         btrfs_calculate_inode_block_rsv_size(fs_info, inode);
6238         spin_unlock(&inode->lock);
6239
6240         if (btrfs_is_testing(fs_info))
6241                 return;
6242
6243         btrfs_inode_rsv_release(inode, qgroup_free);
6244 }
6245
6246 /**
6247  * btrfs_delalloc_release_extents - release our outstanding_extents
6248  * @inode: the inode to balance the reservation for.
6249  * @num_bytes: the number of bytes we originally reserved with
6250  * @qgroup_free: do we need to free qgroup meta reservation or convert them.
6251  *
6252  * When we reserve space we increase outstanding_extents for the extents we may
6253  * add.  Once we've set the range as delalloc or created our ordered extents we
6254  * have outstanding_extents to track the real usage, so we use this to free our
6255  * temporarily tracked outstanding_extents.  This _must_ be used in conjunction
6256  * with btrfs_delalloc_reserve_metadata.
6257  */
6258 void btrfs_delalloc_release_extents(struct btrfs_inode *inode, u64 num_bytes,
6259                                     bool qgroup_free)
6260 {
6261         struct btrfs_fs_info *fs_info = inode->root->fs_info;
6262         unsigned num_extents;
6263
6264         spin_lock(&inode->lock);
6265         num_extents = count_max_extents(num_bytes);
6266         btrfs_mod_outstanding_extents(inode, -num_extents);
6267         btrfs_calculate_inode_block_rsv_size(fs_info, inode);
6268         spin_unlock(&inode->lock);
6269
6270         if (btrfs_is_testing(fs_info))
6271                 return;
6272
6273         btrfs_inode_rsv_release(inode, qgroup_free);
6274 }
6275
6276 /**
6277  * btrfs_delalloc_reserve_space - reserve data and metadata space for
6278  * delalloc
6279  * @inode: inode we're writing to
6280  * @start: start range we are writing to
6281  * @len: how long the range we are writing to
6282  * @reserved: mandatory parameter, record actually reserved qgroup ranges of
6283  *            current reservation.
6284  *
6285  * This will do the following things
6286  *
6287  * o reserve space in data space info for num bytes
6288  *   and reserve precious corresponding qgroup space
6289  *   (Done in check_data_free_space)
6290  *
6291  * o reserve space for metadata space, based on the number of outstanding
6292  *   extents and how much csums will be needed
6293  *   also reserve metadata space in a per root over-reserve method.
6294  * o add to the inodes->delalloc_bytes
6295  * o add it to the fs_info's delalloc inodes list.
6296  *   (Above 3 all done in delalloc_reserve_metadata)
6297  *
6298  * Return 0 for success
6299  * Return <0 for error(-ENOSPC or -EQUOT)
6300  */
6301 int btrfs_delalloc_reserve_space(struct inode *inode,
6302                         struct extent_changeset **reserved, u64 start, u64 len)
6303 {
6304         int ret;
6305
6306         ret = btrfs_check_data_free_space(inode, reserved, start, len);
6307         if (ret < 0)
6308                 return ret;
6309         ret = btrfs_delalloc_reserve_metadata(BTRFS_I(inode), len);
6310         if (ret < 0)
6311                 btrfs_free_reserved_data_space(inode, *reserved, start, len);
6312         return ret;
6313 }
6314
6315 /**
6316  * btrfs_delalloc_release_space - release data and metadata space for delalloc
6317  * @inode: inode we're releasing space for
6318  * @start: start position of the space already reserved
6319  * @len: the len of the space already reserved
6320  * @release_bytes: the len of the space we consumed or didn't use
6321  *
6322  * This function will release the metadata space that was not used and will
6323  * decrement ->delalloc_bytes and remove it from the fs_info delalloc_inodes
6324  * list if there are no delalloc bytes left.
6325  * Also it will handle the qgroup reserved space.
6326  */
6327 void btrfs_delalloc_release_space(struct inode *inode,
6328                                   struct extent_changeset *reserved,
6329                                   u64 start, u64 len, bool qgroup_free)
6330 {
6331         btrfs_delalloc_release_metadata(BTRFS_I(inode), len, qgroup_free);
6332         btrfs_free_reserved_data_space(inode, reserved, start, len);
6333 }
6334
6335 static int update_block_group(struct btrfs_trans_handle *trans,
6336                               struct btrfs_fs_info *info, u64 bytenr,
6337                               u64 num_bytes, int alloc)
6338 {
6339         struct btrfs_block_group_cache *cache = NULL;
6340         u64 total = num_bytes;
6341         u64 old_val;
6342         u64 byte_in_group;
6343         int factor;
6344         int ret = 0;
6345
6346         /* block accounting for super block */
6347         spin_lock(&info->delalloc_root_lock);
6348         old_val = btrfs_super_bytes_used(info->super_copy);
6349         if (alloc)
6350                 old_val += num_bytes;
6351         else
6352                 old_val -= num_bytes;
6353         btrfs_set_super_bytes_used(info->super_copy, old_val);
6354         spin_unlock(&info->delalloc_root_lock);
6355
6356         while (total) {
6357                 cache = btrfs_lookup_block_group(info, bytenr);
6358                 if (!cache) {
6359                         ret = -ENOENT;
6360                         break;
6361                 }
6362                 factor = btrfs_bg_type_to_factor(cache->flags);
6363
6364                 /*
6365                  * If this block group has free space cache written out, we
6366                  * need to make sure to load it if we are removing space.  This
6367                  * is because we need the unpinning stage to actually add the
6368                  * space back to the block group, otherwise we will leak space.
6369                  */
6370                 if (!alloc && cache->cached == BTRFS_CACHE_NO)
6371                         cache_block_group(cache, 1);
6372
6373                 byte_in_group = bytenr - cache->key.objectid;
6374                 WARN_ON(byte_in_group > cache->key.offset);
6375
6376                 spin_lock(&cache->space_info->lock);
6377                 spin_lock(&cache->lock);
6378
6379                 if (btrfs_test_opt(info, SPACE_CACHE) &&
6380                     cache->disk_cache_state < BTRFS_DC_CLEAR)
6381                         cache->disk_cache_state = BTRFS_DC_CLEAR;
6382
6383                 old_val = btrfs_block_group_used(&cache->item);
6384                 num_bytes = min(total, cache->key.offset - byte_in_group);
6385                 if (alloc) {
6386                         old_val += num_bytes;
6387                         btrfs_set_block_group_used(&cache->item, old_val);
6388                         cache->reserved -= num_bytes;
6389                         cache->space_info->bytes_reserved -= num_bytes;
6390                         cache->space_info->bytes_used += num_bytes;
6391                         cache->space_info->disk_used += num_bytes * factor;
6392                         spin_unlock(&cache->lock);
6393                         spin_unlock(&cache->space_info->lock);
6394                 } else {
6395                         old_val -= num_bytes;
6396                         btrfs_set_block_group_used(&cache->item, old_val);
6397                         cache->pinned += num_bytes;
6398                         update_bytes_pinned(cache->space_info, num_bytes);
6399                         cache->space_info->bytes_used -= num_bytes;
6400                         cache->space_info->disk_used -= num_bytes * factor;
6401                         spin_unlock(&cache->lock);
6402                         spin_unlock(&cache->space_info->lock);
6403
6404                         trace_btrfs_space_reservation(info, "pinned",
6405                                                       cache->space_info->flags,
6406                                                       num_bytes, 1);
6407                         percpu_counter_add_batch(&cache->space_info->total_bytes_pinned,
6408                                            num_bytes,
6409                                            BTRFS_TOTAL_BYTES_PINNED_BATCH);
6410                         set_extent_dirty(info->pinned_extents,
6411                                          bytenr, bytenr + num_bytes - 1,
6412                                          GFP_NOFS | __GFP_NOFAIL);
6413                 }
6414
6415                 spin_lock(&trans->transaction->dirty_bgs_lock);
6416                 if (list_empty(&cache->dirty_list)) {
6417                         list_add_tail(&cache->dirty_list,
6418                                       &trans->transaction->dirty_bgs);
6419                         trans->transaction->num_dirty_bgs++;
6420                         trans->delayed_ref_updates++;
6421                         btrfs_get_block_group(cache);
6422                 }
6423                 spin_unlock(&trans->transaction->dirty_bgs_lock);
6424
6425                 /*
6426                  * No longer have used bytes in this block group, queue it for
6427                  * deletion. We do this after adding the block group to the
6428                  * dirty list to avoid races between cleaner kthread and space
6429                  * cache writeout.
6430                  */
6431                 if (!alloc && old_val == 0)
6432                         btrfs_mark_bg_unused(cache);
6433
6434                 btrfs_put_block_group(cache);
6435                 total -= num_bytes;
6436                 bytenr += num_bytes;
6437         }
6438
6439         /* Modified block groups are accounted for in the delayed_refs_rsv. */
6440         btrfs_update_delayed_refs_rsv(trans);
6441         return ret;
6442 }
6443
6444 static u64 first_logical_byte(struct btrfs_fs_info *fs_info, u64 search_start)
6445 {
6446         struct btrfs_block_group_cache *cache;
6447         u64 bytenr;
6448
6449         spin_lock(&fs_info->block_group_cache_lock);
6450         bytenr = fs_info->first_logical_byte;
6451         spin_unlock(&fs_info->block_group_cache_lock);
6452
6453         if (bytenr < (u64)-1)
6454                 return bytenr;
6455
6456         cache = btrfs_lookup_first_block_group(fs_info, search_start);
6457         if (!cache)
6458                 return 0;
6459
6460         bytenr = cache->key.objectid;
6461         btrfs_put_block_group(cache);
6462
6463         return bytenr;
6464 }
6465
6466 static int pin_down_extent(struct btrfs_fs_info *fs_info,
6467                            struct btrfs_block_group_cache *cache,
6468                            u64 bytenr, u64 num_bytes, int reserved)
6469 {
6470         spin_lock(&cache->space_info->lock);
6471         spin_lock(&cache->lock);
6472         cache->pinned += num_bytes;
6473         update_bytes_pinned(cache->space_info, num_bytes);
6474         if (reserved) {
6475                 cache->reserved -= num_bytes;
6476                 cache->space_info->bytes_reserved -= num_bytes;
6477         }
6478         spin_unlock(&cache->lock);
6479         spin_unlock(&cache->space_info->lock);
6480
6481         trace_btrfs_space_reservation(fs_info, "pinned",
6482                                       cache->space_info->flags, num_bytes, 1);
6483         percpu_counter_add_batch(&cache->space_info->total_bytes_pinned,
6484                     num_bytes, BTRFS_TOTAL_BYTES_PINNED_BATCH);
6485         set_extent_dirty(fs_info->pinned_extents, bytenr,
6486                          bytenr + num_bytes - 1, GFP_NOFS | __GFP_NOFAIL);
6487         return 0;
6488 }
6489
6490 /*
6491  * this function must be called within transaction
6492  */
6493 int btrfs_pin_extent(struct btrfs_fs_info *fs_info,
6494                      u64 bytenr, u64 num_bytes, int reserved)
6495 {
6496         struct btrfs_block_group_cache *cache;
6497
6498         cache = btrfs_lookup_block_group(fs_info, bytenr);
6499         BUG_ON(!cache); /* Logic error */
6500
6501         pin_down_extent(fs_info, cache, bytenr, num_bytes, reserved);
6502
6503         btrfs_put_block_group(cache);
6504         return 0;
6505 }
6506
6507 /*
6508  * this function must be called within transaction
6509  */
6510 int btrfs_pin_extent_for_log_replay(struct btrfs_fs_info *fs_info,
6511                                     u64 bytenr, u64 num_bytes)
6512 {
6513         struct btrfs_block_group_cache *cache;
6514         int ret;
6515
6516         cache = btrfs_lookup_block_group(fs_info, bytenr);
6517         if (!cache)
6518                 return -EINVAL;
6519
6520         /*
6521          * pull in the free space cache (if any) so that our pin
6522          * removes the free space from the cache.  We have load_only set
6523          * to one because the slow code to read in the free extents does check
6524          * the pinned extents.
6525          */
6526         cache_block_group(cache, 1);
6527
6528         pin_down_extent(fs_info, cache, bytenr, num_bytes, 0);
6529
6530         /* remove us from the free space cache (if we're there at all) */
6531         ret = btrfs_remove_free_space(cache, bytenr, num_bytes);
6532         btrfs_put_block_group(cache);
6533         return ret;
6534 }
6535
6536 static int __exclude_logged_extent(struct btrfs_fs_info *fs_info,
6537                                    u64 start, u64 num_bytes)
6538 {
6539         int ret;
6540         struct btrfs_block_group_cache *block_group;
6541         struct btrfs_caching_control *caching_ctl;
6542
6543         block_group = btrfs_lookup_block_group(fs_info, start);
6544         if (!block_group)
6545                 return -EINVAL;
6546
6547         cache_block_group(block_group, 0);
6548         caching_ctl = get_caching_control(block_group);
6549
6550         if (!caching_ctl) {
6551                 /* Logic error */
6552                 BUG_ON(!block_group_cache_done(block_group));
6553                 ret = btrfs_remove_free_space(block_group, start, num_bytes);
6554         } else {
6555                 mutex_lock(&caching_ctl->mutex);
6556
6557                 if (start >= caching_ctl->progress) {
6558                         ret = add_excluded_extent(fs_info, start, num_bytes);
6559                 } else if (start + num_bytes <= caching_ctl->progress) {
6560                         ret = btrfs_remove_free_space(block_group,
6561                                                       start, num_bytes);
6562                 } else {
6563                         num_bytes = caching_ctl->progress - start;
6564                         ret = btrfs_remove_free_space(block_group,
6565                                                       start, num_bytes);
6566                         if (ret)
6567                                 goto out_lock;
6568
6569                         num_bytes = (start + num_bytes) -
6570                                 caching_ctl->progress;
6571                         start = caching_ctl->progress;
6572                         ret = add_excluded_extent(fs_info, start, num_bytes);
6573                 }
6574 out_lock:
6575                 mutex_unlock(&caching_ctl->mutex);
6576                 put_caching_control(caching_ctl);
6577         }
6578         btrfs_put_block_group(block_group);
6579         return ret;
6580 }
6581
6582 int btrfs_exclude_logged_extents(struct btrfs_fs_info *fs_info,
6583                                  struct extent_buffer *eb)
6584 {
6585         struct btrfs_file_extent_item *item;
6586         struct btrfs_key key;
6587         int found_type;
6588         int i;
6589         int ret = 0;
6590
6591         if (!btrfs_fs_incompat(fs_info, MIXED_GROUPS))
6592                 return 0;
6593
6594         for (i = 0; i < btrfs_header_nritems(eb); i++) {
6595                 btrfs_item_key_to_cpu(eb, &key, i);
6596                 if (key.type != BTRFS_EXTENT_DATA_KEY)
6597                         continue;
6598                 item = btrfs_item_ptr(eb, i, struct btrfs_file_extent_item);
6599                 found_type = btrfs_file_extent_type(eb, item);
6600                 if (found_type == BTRFS_FILE_EXTENT_INLINE)
6601                         continue;
6602                 if (btrfs_file_extent_disk_bytenr(eb, item) == 0)
6603                         continue;
6604                 key.objectid = btrfs_file_extent_disk_bytenr(eb, item);
6605                 key.offset = btrfs_file_extent_disk_num_bytes(eb, item);
6606                 ret = __exclude_logged_extent(fs_info, key.objectid, key.offset);
6607                 if (ret)
6608                         break;
6609         }
6610
6611         return ret;
6612 }
6613
6614 static void
6615 btrfs_inc_block_group_reservations(struct btrfs_block_group_cache *bg)
6616 {
6617         atomic_inc(&bg->reservations);
6618 }
6619
6620 void btrfs_dec_block_group_reservations(struct btrfs_fs_info *fs_info,
6621                                         const u64 start)
6622 {
6623         struct btrfs_block_group_cache *bg;
6624
6625         bg = btrfs_lookup_block_group(fs_info, start);
6626         ASSERT(bg);
6627         if (atomic_dec_and_test(&bg->reservations))
6628                 wake_up_var(&bg->reservations);
6629         btrfs_put_block_group(bg);
6630 }
6631
6632 void btrfs_wait_block_group_reservations(struct btrfs_block_group_cache *bg)
6633 {
6634         struct btrfs_space_info *space_info = bg->space_info;
6635
6636         ASSERT(bg->ro);
6637
6638         if (!(bg->flags & BTRFS_BLOCK_GROUP_DATA))
6639                 return;
6640
6641         /*
6642          * Our block group is read only but before we set it to read only,
6643          * some task might have had allocated an extent from it already, but it
6644          * has not yet created a respective ordered extent (and added it to a
6645          * root's list of ordered extents).
6646          * Therefore wait for any task currently allocating extents, since the
6647          * block group's reservations counter is incremented while a read lock
6648          * on the groups' semaphore is held and decremented after releasing
6649          * the read access on that semaphore and creating the ordered extent.
6650          */
6651         down_write(&space_info->groups_sem);
6652         up_write(&space_info->groups_sem);
6653
6654         wait_var_event(&bg->reservations, !atomic_read(&bg->reservations));
6655 }
6656
6657 /**
6658  * btrfs_add_reserved_bytes - update the block_group and space info counters
6659  * @cache:      The cache we are manipulating
6660  * @ram_bytes:  The number of bytes of file content, and will be same to
6661  *              @num_bytes except for the compress path.
6662  * @num_bytes:  The number of bytes in question
6663  * @delalloc:   The blocks are allocated for the delalloc write
6664  *
6665  * This is called by the allocator when it reserves space. If this is a
6666  * reservation and the block group has become read only we cannot make the
6667  * reservation and return -EAGAIN, otherwise this function always succeeds.
6668  */
6669 static int btrfs_add_reserved_bytes(struct btrfs_block_group_cache *cache,
6670                                     u64 ram_bytes, u64 num_bytes, int delalloc)
6671 {
6672         struct btrfs_space_info *space_info = cache->space_info;
6673         int ret = 0;
6674
6675         spin_lock(&space_info->lock);
6676         spin_lock(&cache->lock);
6677         if (cache->ro) {
6678                 ret = -EAGAIN;
6679         } else {
6680                 cache->reserved += num_bytes;
6681                 space_info->bytes_reserved += num_bytes;
6682                 update_bytes_may_use(space_info, -ram_bytes);
6683                 if (delalloc)
6684                         cache->delalloc_bytes += num_bytes;
6685         }
6686         spin_unlock(&cache->lock);
6687         spin_unlock(&space_info->lock);
6688         return ret;
6689 }
6690
6691 /**
6692  * btrfs_free_reserved_bytes - update the block_group and space info counters
6693  * @cache:      The cache we are manipulating
6694  * @num_bytes:  The number of bytes in question
6695  * @delalloc:   The blocks are allocated for the delalloc write
6696  *
6697  * This is called by somebody who is freeing space that was never actually used
6698  * on disk.  For example if you reserve some space for a new leaf in transaction
6699  * A and before transaction A commits you free that leaf, you call this with
6700  * reserve set to 0 in order to clear the reservation.
6701  */
6702
6703 static void btrfs_free_reserved_bytes(struct btrfs_block_group_cache *cache,
6704                                       u64 num_bytes, int delalloc)
6705 {
6706         struct btrfs_space_info *space_info = cache->space_info;
6707
6708         spin_lock(&space_info->lock);
6709         spin_lock(&cache->lock);
6710         if (cache->ro)
6711                 space_info->bytes_readonly += num_bytes;
6712         cache->reserved -= num_bytes;
6713         space_info->bytes_reserved -= num_bytes;
6714         space_info->max_extent_size = 0;
6715
6716         if (delalloc)
6717                 cache->delalloc_bytes -= num_bytes;
6718         spin_unlock(&cache->lock);
6719         spin_unlock(&space_info->lock);
6720 }
6721 void btrfs_prepare_extent_commit(struct btrfs_fs_info *fs_info)
6722 {
6723         struct btrfs_caching_control *next;
6724         struct btrfs_caching_control *caching_ctl;
6725         struct btrfs_block_group_cache *cache;
6726
6727         down_write(&fs_info->commit_root_sem);
6728
6729         list_for_each_entry_safe(caching_ctl, next,
6730                                  &fs_info->caching_block_groups, list) {
6731                 cache = caching_ctl->block_group;
6732                 if (block_group_cache_done(cache)) {
6733                         cache->last_byte_to_unpin = (u64)-1;
6734                         list_del_init(&caching_ctl->list);
6735                         put_caching_control(caching_ctl);
6736                 } else {
6737                         cache->last_byte_to_unpin = caching_ctl->progress;
6738                 }
6739         }
6740
6741         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
6742                 fs_info->pinned_extents = &fs_info->freed_extents[1];
6743         else
6744                 fs_info->pinned_extents = &fs_info->freed_extents[0];
6745
6746         up_write(&fs_info->commit_root_sem);
6747
6748         update_global_block_rsv(fs_info);
6749 }
6750
6751 /*
6752  * Returns the free cluster for the given space info and sets empty_cluster to
6753  * what it should be based on the mount options.
6754  */
6755 static struct btrfs_free_cluster *
6756 fetch_cluster_info(struct btrfs_fs_info *fs_info,
6757                    struct btrfs_space_info *space_info, u64 *empty_cluster)
6758 {
6759         struct btrfs_free_cluster *ret = NULL;
6760
6761         *empty_cluster = 0;
6762         if (btrfs_mixed_space_info(space_info))
6763                 return ret;
6764
6765         if (space_info->flags & BTRFS_BLOCK_GROUP_METADATA) {
6766                 ret = &fs_info->meta_alloc_cluster;
6767                 if (btrfs_test_opt(fs_info, SSD))
6768                         *empty_cluster = SZ_2M;
6769                 else
6770                         *empty_cluster = SZ_64K;
6771         } else if ((space_info->flags & BTRFS_BLOCK_GROUP_DATA) &&
6772                    btrfs_test_opt(fs_info, SSD_SPREAD)) {
6773                 *empty_cluster = SZ_2M;
6774                 ret = &fs_info->data_alloc_cluster;
6775         }
6776
6777         return ret;
6778 }
6779
6780 static int unpin_extent_range(struct btrfs_fs_info *fs_info,
6781                               u64 start, u64 end,
6782                               const bool return_free_space)
6783 {
6784         struct btrfs_block_group_cache *cache = NULL;
6785         struct btrfs_space_info *space_info;
6786         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
6787         struct btrfs_free_cluster *cluster = NULL;
6788         u64 len;
6789         u64 total_unpinned = 0;
6790         u64 empty_cluster = 0;
6791         bool readonly;
6792
6793         while (start <= end) {
6794                 readonly = false;
6795                 if (!cache ||
6796                     start >= cache->key.objectid + cache->key.offset) {
6797                         if (cache)
6798                                 btrfs_put_block_group(cache);
6799                         total_unpinned = 0;
6800                         cache = btrfs_lookup_block_group(fs_info, start);
6801                         BUG_ON(!cache); /* Logic error */
6802
6803                         cluster = fetch_cluster_info(fs_info,
6804                                                      cache->space_info,
6805                                                      &empty_cluster);
6806                         empty_cluster <<= 1;
6807                 }
6808
6809                 len = cache->key.objectid + cache->key.offset - start;
6810                 len = min(len, end + 1 - start);
6811
6812                 if (start < cache->last_byte_to_unpin) {
6813                         len = min(len, cache->last_byte_to_unpin - start);
6814                         if (return_free_space)
6815                                 btrfs_add_free_space(cache, start, len);
6816                 }
6817
6818                 start += len;
6819                 total_unpinned += len;
6820                 space_info = cache->space_info;
6821
6822                 /*
6823                  * If this space cluster has been marked as fragmented and we've
6824                  * unpinned enough in this block group to potentially allow a
6825                  * cluster to be created inside of it go ahead and clear the
6826                  * fragmented check.
6827                  */
6828                 if (cluster && cluster->fragmented &&
6829                     total_unpinned > empty_cluster) {
6830                         spin_lock(&cluster->lock);
6831                         cluster->fragmented = 0;
6832                         spin_unlock(&cluster->lock);
6833                 }
6834
6835                 spin_lock(&space_info->lock);
6836                 spin_lock(&cache->lock);
6837                 cache->pinned -= len;
6838                 update_bytes_pinned(space_info, -len);
6839
6840                 trace_btrfs_space_reservation(fs_info, "pinned",
6841                                               space_info->flags, len, 0);
6842                 space_info->max_extent_size = 0;
6843                 percpu_counter_add_batch(&space_info->total_bytes_pinned,
6844                             -len, BTRFS_TOTAL_BYTES_PINNED_BATCH);
6845                 if (cache->ro) {
6846                         space_info->bytes_readonly += len;
6847                         readonly = true;
6848                 }
6849                 spin_unlock(&cache->lock);
6850                 if (!readonly && return_free_space &&
6851                     global_rsv->space_info == space_info) {
6852                         u64 to_add = len;
6853
6854                         spin_lock(&global_rsv->lock);
6855                         if (!global_rsv->full) {
6856                                 to_add = min(len, global_rsv->size -
6857                                              global_rsv->reserved);
6858                                 global_rsv->reserved += to_add;
6859                                 update_bytes_may_use(space_info, to_add);
6860                                 if (global_rsv->reserved >= global_rsv->size)
6861                                         global_rsv->full = 1;
6862                                 trace_btrfs_space_reservation(fs_info,
6863                                                               "space_info",
6864                                                               space_info->flags,
6865                                                               to_add, 1);
6866                                 len -= to_add;
6867                         }
6868                         spin_unlock(&global_rsv->lock);
6869                         /* Add to any tickets we may have */
6870                         if (len)
6871                                 space_info_add_new_bytes(fs_info, space_info,
6872                                                          len);
6873                 }
6874                 spin_unlock(&space_info->lock);
6875         }
6876
6877         if (cache)
6878                 btrfs_put_block_group(cache);
6879         return 0;
6880 }
6881
6882 int btrfs_finish_extent_commit(struct btrfs_trans_handle *trans)
6883 {
6884         struct btrfs_fs_info *fs_info = trans->fs_info;
6885         struct btrfs_block_group_cache *block_group, *tmp;
6886         struct list_head *deleted_bgs;
6887         struct extent_io_tree *unpin;
6888         u64 start;
6889         u64 end;
6890         int ret;
6891
6892         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
6893                 unpin = &fs_info->freed_extents[1];
6894         else
6895                 unpin = &fs_info->freed_extents[0];
6896
6897         while (!trans->aborted) {
6898                 struct extent_state *cached_state = NULL;
6899
6900                 mutex_lock(&fs_info->unused_bg_unpin_mutex);
6901                 ret = find_first_extent_bit(unpin, 0, &start, &end,
6902                                             EXTENT_DIRTY, &cached_state);
6903                 if (ret) {
6904                         mutex_unlock(&fs_info->unused_bg_unpin_mutex);
6905                         break;
6906                 }
6907
6908                 if (btrfs_test_opt(fs_info, DISCARD))
6909                         ret = btrfs_discard_extent(fs_info, start,
6910                                                    end + 1 - start, NULL);
6911
6912                 clear_extent_dirty(unpin, start, end, &cached_state);
6913                 unpin_extent_range(fs_info, start, end, true);
6914                 mutex_unlock(&fs_info->unused_bg_unpin_mutex);
6915                 free_extent_state(cached_state);
6916                 cond_resched();
6917         }
6918
6919         /*
6920          * Transaction is finished.  We don't need the lock anymore.  We
6921          * do need to clean up the block groups in case of a transaction
6922          * abort.
6923          */
6924         deleted_bgs = &trans->transaction->deleted_bgs;
6925         list_for_each_entry_safe(block_group, tmp, deleted_bgs, bg_list) {
6926                 u64 trimmed = 0;
6927
6928                 ret = -EROFS;
6929                 if (!trans->aborted)
6930                         ret = btrfs_discard_extent(fs_info,
6931                                                    block_group->key.objectid,
6932                                                    block_group->key.offset,
6933                                                    &trimmed);
6934
6935                 list_del_init(&block_group->bg_list);
6936                 btrfs_put_block_group_trimming(block_group);
6937                 btrfs_put_block_group(block_group);
6938
6939                 if (ret) {
6940                         const char *errstr = btrfs_decode_error(ret);
6941                         btrfs_warn(fs_info,
6942                            "discard failed while removing blockgroup: errno=%d %s",
6943                                    ret, errstr);
6944                 }
6945         }
6946
6947         return 0;
6948 }
6949
6950 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
6951                                struct btrfs_delayed_ref_node *node, u64 parent,
6952                                u64 root_objectid, u64 owner_objectid,
6953                                u64 owner_offset, int refs_to_drop,
6954                                struct btrfs_delayed_extent_op *extent_op)
6955 {
6956         struct btrfs_fs_info *info = trans->fs_info;
6957         struct btrfs_key key;
6958         struct btrfs_path *path;
6959         struct btrfs_root *extent_root = info->extent_root;
6960         struct extent_buffer *leaf;
6961         struct btrfs_extent_item *ei;
6962         struct btrfs_extent_inline_ref *iref;
6963         int ret;
6964         int is_data;
6965         int extent_slot = 0;
6966         int found_extent = 0;
6967         int num_to_del = 1;
6968         u32 item_size;
6969         u64 refs;
6970         u64 bytenr = node->bytenr;
6971         u64 num_bytes = node->num_bytes;
6972         int last_ref = 0;
6973         bool skinny_metadata = btrfs_fs_incompat(info, SKINNY_METADATA);
6974
6975         path = btrfs_alloc_path();
6976         if (!path)
6977                 return -ENOMEM;
6978
6979         path->reada = READA_FORWARD;
6980         path->leave_spinning = 1;
6981
6982         is_data = owner_objectid >= BTRFS_FIRST_FREE_OBJECTID;
6983         BUG_ON(!is_data && refs_to_drop != 1);
6984
6985         if (is_data)
6986                 skinny_metadata = false;
6987
6988         ret = lookup_extent_backref(trans, path, &iref, bytenr, num_bytes,
6989                                     parent, root_objectid, owner_objectid,
6990                                     owner_offset);
6991         if (ret == 0) {
6992                 extent_slot = path->slots[0];
6993                 while (extent_slot >= 0) {
6994                         btrfs_item_key_to_cpu(path->nodes[0], &key,
6995                                               extent_slot);
6996                         if (key.objectid != bytenr)
6997                                 break;
6998                         if (key.type == BTRFS_EXTENT_ITEM_KEY &&
6999                             key.offset == num_bytes) {
7000                                 found_extent = 1;
7001                                 break;
7002                         }
7003                         if (key.type == BTRFS_METADATA_ITEM_KEY &&
7004                             key.offset == owner_objectid) {
7005                                 found_extent = 1;
7006                                 break;
7007                         }
7008                         if (path->slots[0] - extent_slot > 5)
7009                                 break;
7010                         extent_slot--;
7011                 }
7012
7013                 if (!found_extent) {
7014                         BUG_ON(iref);
7015                         ret = remove_extent_backref(trans, path, NULL,
7016                                                     refs_to_drop,
7017                                                     is_data, &last_ref);
7018                         if (ret) {
7019                                 btrfs_abort_transaction(trans, ret);
7020                                 goto out;
7021                         }
7022                         btrfs_release_path(path);
7023                         path->leave_spinning = 1;
7024
7025                         key.objectid = bytenr;
7026                         key.type = BTRFS_EXTENT_ITEM_KEY;
7027                         key.offset = num_bytes;
7028
7029                         if (!is_data && skinny_metadata) {
7030                                 key.type = BTRFS_METADATA_ITEM_KEY;
7031                                 key.offset = owner_objectid;
7032                         }
7033
7034                         ret = btrfs_search_slot(trans, extent_root,
7035                                                 &key, path, -1, 1);
7036                         if (ret > 0 && skinny_metadata && path->slots[0]) {
7037                                 /*
7038                                  * Couldn't find our skinny metadata item,
7039                                  * see if we have ye olde extent item.
7040                                  */
7041                                 path->slots[0]--;
7042                                 btrfs_item_key_to_cpu(path->nodes[0], &key,
7043                                                       path->slots[0]);
7044                                 if (key.objectid == bytenr &&
7045                                     key.type == BTRFS_EXTENT_ITEM_KEY &&
7046                                     key.offset == num_bytes)
7047                                         ret = 0;
7048                         }
7049
7050                         if (ret > 0 && skinny_metadata) {
7051                                 skinny_metadata = false;
7052                                 key.objectid = bytenr;
7053                                 key.type = BTRFS_EXTENT_ITEM_KEY;
7054                                 key.offset = num_bytes;
7055                                 btrfs_release_path(path);
7056                                 ret = btrfs_search_slot(trans, extent_root,
7057                                                         &key, path, -1, 1);
7058                         }
7059
7060                         if (ret) {
7061                                 btrfs_err(info,
7062                                           "umm, got %d back from search, was looking for %llu",
7063                                           ret, bytenr);
7064                                 if (ret > 0)
7065                                         btrfs_print_leaf(path->nodes[0]);
7066                         }
7067                         if (ret < 0) {
7068                                 btrfs_abort_transaction(trans, ret);
7069                                 goto out;
7070                         }
7071                         extent_slot = path->slots[0];
7072                 }
7073         } else if (WARN_ON(ret == -ENOENT)) {
7074                 btrfs_print_leaf(path->nodes[0]);
7075                 btrfs_err(info,
7076                         "unable to find ref byte nr %llu parent %llu root %llu  owner %llu offset %llu",
7077                         bytenr, parent, root_objectid, owner_objectid,
7078                         owner_offset);
7079                 btrfs_abort_transaction(trans, ret);
7080                 goto out;
7081         } else {
7082                 btrfs_abort_transaction(trans, ret);
7083                 goto out;
7084         }
7085
7086         leaf = path->nodes[0];
7087         item_size = btrfs_item_size_nr(leaf, extent_slot);
7088         if (unlikely(item_size < sizeof(*ei))) {
7089                 ret = -EINVAL;
7090                 btrfs_print_v0_err(info);
7091                 btrfs_abort_transaction(trans, ret);
7092                 goto out;
7093         }
7094         ei = btrfs_item_ptr(leaf, extent_slot,
7095                             struct btrfs_extent_item);
7096         if (owner_objectid < BTRFS_FIRST_FREE_OBJECTID &&
7097             key.type == BTRFS_EXTENT_ITEM_KEY) {
7098                 struct btrfs_tree_block_info *bi;
7099                 BUG_ON(item_size < sizeof(*ei) + sizeof(*bi));
7100                 bi = (struct btrfs_tree_block_info *)(ei + 1);
7101                 WARN_ON(owner_objectid != btrfs_tree_block_level(leaf, bi));
7102         }
7103
7104         refs = btrfs_extent_refs(leaf, ei);
7105         if (refs < refs_to_drop) {
7106                 btrfs_err(info,
7107                           "trying to drop %d refs but we only have %Lu for bytenr %Lu",
7108                           refs_to_drop, refs, bytenr);
7109                 ret = -EINVAL;
7110                 btrfs_abort_transaction(trans, ret);
7111                 goto out;
7112         }
7113         refs -= refs_to_drop;
7114
7115         if (refs > 0) {
7116                 if (extent_op)
7117                         __run_delayed_extent_op(extent_op, leaf, ei);
7118                 /*
7119                  * In the case of inline back ref, reference count will
7120                  * be updated by remove_extent_backref
7121                  */
7122                 if (iref) {
7123                         BUG_ON(!found_extent);
7124                 } else {
7125                         btrfs_set_extent_refs(leaf, ei, refs);
7126                         btrfs_mark_buffer_dirty(leaf);
7127                 }
7128                 if (found_extent) {
7129                         ret = remove_extent_backref(trans, path, iref,
7130                                                     refs_to_drop, is_data,
7131                                                     &last_ref);
7132                         if (ret) {
7133                                 btrfs_abort_transaction(trans, ret);
7134                                 goto out;
7135                         }
7136                 }
7137         } else {
7138                 if (found_extent) {
7139                         BUG_ON(is_data && refs_to_drop !=
7140                                extent_data_ref_count(path, iref));
7141                         if (iref) {
7142                                 BUG_ON(path->slots[0] != extent_slot);
7143                         } else {
7144                                 BUG_ON(path->slots[0] != extent_slot + 1);
7145                                 path->slots[0] = extent_slot;
7146                                 num_to_del = 2;
7147                         }
7148                 }
7149
7150                 last_ref = 1;
7151                 ret = btrfs_del_items(trans, extent_root, path, path->slots[0],
7152                                       num_to_del);
7153                 if (ret) {
7154                         btrfs_abort_transaction(trans, ret);
7155                         goto out;
7156                 }
7157                 btrfs_release_path(path);
7158
7159                 if (is_data) {
7160                         ret = btrfs_del_csums(trans, info, bytenr, num_bytes);
7161                         if (ret) {
7162                                 btrfs_abort_transaction(trans, ret);
7163                                 goto out;
7164                         }
7165                 }
7166
7167                 ret = add_to_free_space_tree(trans, bytenr, num_bytes);
7168                 if (ret) {
7169                         btrfs_abort_transaction(trans, ret);
7170                         goto out;
7171                 }
7172
7173                 ret = update_block_group(trans, info, bytenr, num_bytes, 0);
7174                 if (ret) {
7175                         btrfs_abort_transaction(trans, ret);
7176                         goto out;
7177                 }
7178         }
7179         btrfs_release_path(path);
7180
7181 out:
7182         btrfs_free_path(path);
7183         return ret;
7184 }
7185
7186 /*
7187  * when we free an block, it is possible (and likely) that we free the last
7188  * delayed ref for that extent as well.  This searches the delayed ref tree for
7189  * a given extent, and if there are no other delayed refs to be processed, it
7190  * removes it from the tree.
7191  */
7192 static noinline int check_ref_cleanup(struct btrfs_trans_handle *trans,
7193                                       u64 bytenr)
7194 {
7195         struct btrfs_delayed_ref_head *head;
7196         struct btrfs_delayed_ref_root *delayed_refs;
7197         int ret = 0;
7198
7199         delayed_refs = &trans->transaction->delayed_refs;
7200         spin_lock(&delayed_refs->lock);
7201         head = btrfs_find_delayed_ref_head(delayed_refs, bytenr);
7202         if (!head)
7203                 goto out_delayed_unlock;
7204
7205         spin_lock(&head->lock);
7206         if (!RB_EMPTY_ROOT(&head->ref_tree.rb_root))
7207                 goto out;
7208
7209         if (cleanup_extent_op(head) != NULL)
7210                 goto out;
7211
7212         /*
7213          * waiting for the lock here would deadlock.  If someone else has it
7214          * locked they are already in the process of dropping it anyway
7215          */
7216         if (!mutex_trylock(&head->mutex))
7217                 goto out;
7218
7219         btrfs_delete_ref_head(delayed_refs, head);
7220         head->processing = 0;
7221
7222         spin_unlock(&head->lock);
7223         spin_unlock(&delayed_refs->lock);
7224
7225         BUG_ON(head->extent_op);
7226         if (head->must_insert_reserved)
7227                 ret = 1;
7228
7229         btrfs_cleanup_ref_head_accounting(trans->fs_info, delayed_refs, head);
7230         mutex_unlock(&head->mutex);
7231         btrfs_put_delayed_ref_head(head);
7232         return ret;
7233 out:
7234         spin_unlock(&head->lock);
7235
7236 out_delayed_unlock:
7237         spin_unlock(&delayed_refs->lock);
7238         return 0;
7239 }
7240
7241 void btrfs_free_tree_block(struct btrfs_trans_handle *trans,
7242                            struct btrfs_root *root,
7243                            struct extent_buffer *buf,
7244                            u64 parent, int last_ref)
7245 {
7246         struct btrfs_fs_info *fs_info = root->fs_info;
7247         int pin = 1;
7248         int ret;
7249
7250         if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
7251                 int old_ref_mod, new_ref_mod;
7252
7253                 btrfs_ref_tree_mod(root, buf->start, buf->len, parent,
7254                                    root->root_key.objectid,
7255                                    btrfs_header_level(buf), 0,
7256                                    BTRFS_DROP_DELAYED_REF);
7257                 ret = btrfs_add_delayed_tree_ref(trans, buf->start,
7258                                                  buf->len, parent,
7259                                                  root->root_key.objectid,
7260                                                  btrfs_header_level(buf),
7261                                                  BTRFS_DROP_DELAYED_REF, NULL,
7262                                                  &old_ref_mod, &new_ref_mod);
7263                 BUG_ON(ret); /* -ENOMEM */
7264                 pin = old_ref_mod >= 0 && new_ref_mod < 0;
7265         }
7266
7267         if (last_ref && btrfs_header_generation(buf) == trans->transid) {
7268                 struct btrfs_block_group_cache *cache;
7269
7270                 if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
7271                         ret = check_ref_cleanup(trans, buf->start);
7272                         if (!ret)
7273                                 goto out;
7274                 }
7275
7276                 pin = 0;
7277                 cache = btrfs_lookup_block_group(fs_info, buf->start);
7278
7279                 if (btrfs_header_flag(buf, BTRFS_HEADER_FLAG_WRITTEN)) {
7280                         pin_down_extent(fs_info, cache, buf->start,
7281                                         buf->len, 1);
7282                         btrfs_put_block_group(cache);
7283                         goto out;
7284                 }
7285
7286                 WARN_ON(test_bit(EXTENT_BUFFER_DIRTY, &buf->bflags));
7287
7288                 btrfs_add_free_space(cache, buf->start, buf->len);
7289                 btrfs_free_reserved_bytes(cache, buf->len, 0);
7290                 btrfs_put_block_group(cache);
7291                 trace_btrfs_reserved_extent_free(fs_info, buf->start, buf->len);
7292         }
7293 out:
7294         if (pin)
7295                 add_pinned_bytes(fs_info, buf->len, true,
7296                                  root->root_key.objectid);
7297
7298         if (last_ref) {
7299                 /*
7300                  * Deleting the buffer, clear the corrupt flag since it doesn't
7301                  * matter anymore.
7302                  */
7303                 clear_bit(EXTENT_BUFFER_CORRUPT, &buf->bflags);
7304         }
7305 }
7306
7307 /* Can return -ENOMEM */
7308 int btrfs_free_extent(struct btrfs_trans_handle *trans,
7309                       struct btrfs_root *root,
7310                       u64 bytenr, u64 num_bytes, u64 parent, u64 root_objectid,
7311                       u64 owner, u64 offset)
7312 {
7313         struct btrfs_fs_info *fs_info = root->fs_info;
7314         int old_ref_mod, new_ref_mod;
7315         int ret;
7316
7317         if (btrfs_is_testing(fs_info))
7318                 return 0;
7319
7320         if (root_objectid != BTRFS_TREE_LOG_OBJECTID)
7321                 btrfs_ref_tree_mod(root, bytenr, num_bytes, parent,
7322                                    root_objectid, owner, offset,
7323                                    BTRFS_DROP_DELAYED_REF);
7324
7325         /*
7326          * tree log blocks never actually go into the extent allocation
7327          * tree, just update pinning info and exit early.
7328          */
7329         if (root_objectid == BTRFS_TREE_LOG_OBJECTID) {
7330                 WARN_ON(owner >= BTRFS_FIRST_FREE_OBJECTID);
7331                 /* unlocks the pinned mutex */
7332                 btrfs_pin_extent(fs_info, bytenr, num_bytes, 1);
7333                 old_ref_mod = new_ref_mod = 0;
7334                 ret = 0;
7335         } else if (owner < BTRFS_FIRST_FREE_OBJECTID) {
7336                 ret = btrfs_add_delayed_tree_ref(trans, bytenr,
7337                                                  num_bytes, parent,
7338                                                  root_objectid, (int)owner,
7339                                                  BTRFS_DROP_DELAYED_REF, NULL,
7340                                                  &old_ref_mod, &new_ref_mod);
7341         } else {
7342                 ret = btrfs_add_delayed_data_ref(trans, bytenr,
7343                                                  num_bytes, parent,
7344                                                  root_objectid, owner, offset,
7345                                                  0, BTRFS_DROP_DELAYED_REF,
7346                                                  &old_ref_mod, &new_ref_mod);
7347         }
7348
7349         if (ret == 0 && old_ref_mod >= 0 && new_ref_mod < 0) {
7350                 bool metadata = owner < BTRFS_FIRST_FREE_OBJECTID;
7351
7352                 add_pinned_bytes(fs_info, num_bytes, metadata, root_objectid);
7353         }
7354
7355         return ret;
7356 }
7357
7358 /*
7359  * when we wait for progress in the block group caching, its because
7360  * our allocation attempt failed at least once.  So, we must sleep
7361  * and let some progress happen before we try again.
7362  *
7363  * This function will sleep at least once waiting for new free space to
7364  * show up, and then it will check the block group free space numbers
7365  * for our min num_bytes.  Another option is to have it go ahead
7366  * and look in the rbtree for a free extent of a given size, but this
7367  * is a good start.
7368  *
7369  * Callers of this must check if cache->cached == BTRFS_CACHE_ERROR before using
7370  * any of the information in this block group.
7371  */
7372 static noinline void
7373 wait_block_group_cache_progress(struct btrfs_block_group_cache *cache,
7374                                 u64 num_bytes)
7375 {
7376         struct btrfs_caching_control *caching_ctl;
7377
7378         caching_ctl = get_caching_control(cache);
7379         if (!caching_ctl)
7380                 return;
7381
7382         wait_event(caching_ctl->wait, block_group_cache_done(cache) ||
7383                    (cache->free_space_ctl->free_space >= num_bytes));
7384
7385         put_caching_control(caching_ctl);
7386 }
7387
7388 static noinline int
7389 wait_block_group_cache_done(struct btrfs_block_group_cache *cache)
7390 {
7391         struct btrfs_caching_control *caching_ctl;
7392         int ret = 0;
7393
7394         caching_ctl = get_caching_control(cache);
7395         if (!caching_ctl)
7396                 return (cache->cached == BTRFS_CACHE_ERROR) ? -EIO : 0;
7397
7398         wait_event(caching_ctl->wait, block_group_cache_done(cache));
7399         if (cache->cached == BTRFS_CACHE_ERROR)
7400                 ret = -EIO;
7401         put_caching_control(caching_ctl);
7402         return ret;
7403 }
7404
7405 enum btrfs_loop_type {
7406         LOOP_CACHING_NOWAIT = 0,
7407         LOOP_CACHING_WAIT = 1,
7408         LOOP_ALLOC_CHUNK = 2,
7409         LOOP_NO_EMPTY_SIZE = 3,
7410 };
7411
7412 static inline void
7413 btrfs_lock_block_group(struct btrfs_block_group_cache *cache,
7414                        int delalloc)
7415 {
7416         if (delalloc)
7417                 down_read(&cache->data_rwsem);
7418 }
7419
7420 static inline void
7421 btrfs_grab_block_group(struct btrfs_block_group_cache *cache,
7422                        int delalloc)
7423 {
7424         btrfs_get_block_group(cache);
7425         if (delalloc)
7426                 down_read(&cache->data_rwsem);
7427 }
7428
7429 static struct btrfs_block_group_cache *
7430 btrfs_lock_cluster(struct btrfs_block_group_cache *block_group,
7431                    struct btrfs_free_cluster *cluster,
7432                    int delalloc)
7433 {
7434         struct btrfs_block_group_cache *used_bg = NULL;
7435
7436         spin_lock(&cluster->refill_lock);
7437         while (1) {
7438                 used_bg = cluster->block_group;
7439                 if (!used_bg)
7440                         return NULL;
7441
7442                 if (used_bg == block_group)
7443                         return used_bg;
7444
7445                 btrfs_get_block_group(used_bg);
7446
7447                 if (!delalloc)
7448                         return used_bg;
7449
7450                 if (down_read_trylock(&used_bg->data_rwsem))
7451                         return used_bg;
7452
7453                 spin_unlock(&cluster->refill_lock);
7454
7455                 /* We should only have one-level nested. */
7456                 down_read_nested(&used_bg->data_rwsem, SINGLE_DEPTH_NESTING);
7457
7458                 spin_lock(&cluster->refill_lock);
7459                 if (used_bg == cluster->block_group)
7460                         return used_bg;
7461
7462                 up_read(&used_bg->data_rwsem);
7463                 btrfs_put_block_group(used_bg);
7464         }
7465 }
7466
7467 static inline void
7468 btrfs_release_block_group(struct btrfs_block_group_cache *cache,
7469                          int delalloc)
7470 {
7471         if (delalloc)
7472                 up_read(&cache->data_rwsem);
7473         btrfs_put_block_group(cache);
7474 }
7475
7476 /*
7477  * Structure used internally for find_free_extent() function.  Wraps needed
7478  * parameters.
7479  */
7480 struct find_free_extent_ctl {
7481         /* Basic allocation info */
7482         u64 ram_bytes;
7483         u64 num_bytes;
7484         u64 empty_size;
7485         u64 flags;
7486         int delalloc;
7487
7488         /* Where to start the search inside the bg */
7489         u64 search_start;
7490
7491         /* For clustered allocation */
7492         u64 empty_cluster;
7493
7494         bool have_caching_bg;
7495         bool orig_have_caching_bg;
7496
7497         /* RAID index, converted from flags */
7498         int index;
7499
7500         /*
7501          * Current loop number, check find_free_extent_update_loop() for details
7502          */
7503         int loop;
7504
7505         /*
7506          * Whether we're refilling a cluster, if true we need to re-search
7507          * current block group but don't try to refill the cluster again.
7508          */
7509         bool retry_clustered;
7510
7511         /*
7512          * Whether we're updating free space cache, if true we need to re-search
7513          * current block group but don't try updating free space cache again.
7514          */
7515         bool retry_unclustered;
7516
7517         /* If current block group is cached */
7518         int cached;
7519
7520         /* Max contiguous hole found */
7521         u64 max_extent_size;
7522
7523         /* Total free space from free space cache, not always contiguous */
7524         u64 total_free_space;
7525
7526         /* Found result */
7527         u64 found_offset;
7528 };
7529
7530
7531 /*
7532  * Helper function for find_free_extent().
7533  *
7534  * Return -ENOENT to inform caller that we need fallback to unclustered mode.
7535  * Return -EAGAIN to inform caller that we need to re-search this block group
7536  * Return >0 to inform caller that we find nothing
7537  * Return 0 means we have found a location and set ffe_ctl->found_offset.
7538  */
7539 static int find_free_extent_clustered(struct btrfs_block_group_cache *bg,
7540                 struct btrfs_free_cluster *last_ptr,
7541                 struct find_free_extent_ctl *ffe_ctl,
7542                 struct btrfs_block_group_cache **cluster_bg_ret)
7543 {
7544         struct btrfs_fs_info *fs_info = bg->fs_info;
7545         struct btrfs_block_group_cache *cluster_bg;
7546         u64 aligned_cluster;
7547         u64 offset;
7548         int ret;
7549
7550         cluster_bg = btrfs_lock_cluster(bg, last_ptr, ffe_ctl->delalloc);
7551         if (!cluster_bg)
7552                 goto refill_cluster;
7553         if (cluster_bg != bg && (cluster_bg->ro ||
7554             !block_group_bits(cluster_bg, ffe_ctl->flags)))
7555                 goto release_cluster;
7556
7557         offset = btrfs_alloc_from_cluster(cluster_bg, last_ptr,
7558                         ffe_ctl->num_bytes, cluster_bg->key.objectid,
7559                         &ffe_ctl->max_extent_size);
7560         if (offset) {
7561                 /* We have a block, we're done */
7562                 spin_unlock(&last_ptr->refill_lock);
7563                 trace_btrfs_reserve_extent_cluster(cluster_bg,
7564                                 ffe_ctl->search_start, ffe_ctl->num_bytes);
7565                 *cluster_bg_ret = cluster_bg;
7566                 ffe_ctl->found_offset = offset;
7567                 return 0;
7568         }
7569         WARN_ON(last_ptr->block_group != cluster_bg);
7570
7571 release_cluster:
7572         /*
7573          * If we are on LOOP_NO_EMPTY_SIZE, we can't set up a new clusters, so
7574          * lets just skip it and let the allocator find whatever block it can
7575          * find. If we reach this point, we will have tried the cluster
7576          * allocator plenty of times and not have found anything, so we are
7577          * likely way too fragmented for the clustering stuff to find anything.
7578          *
7579          * However, if the cluster is taken from the current block group,
7580          * release the cluster first, so that we stand a better chance of
7581          * succeeding in the unclustered allocation.
7582          */
7583         if (ffe_ctl->loop >= LOOP_NO_EMPTY_SIZE && cluster_bg != bg) {
7584                 spin_unlock(&last_ptr->refill_lock);
7585                 btrfs_release_block_group(cluster_bg, ffe_ctl->delalloc);
7586                 return -ENOENT;
7587         }
7588
7589         /* This cluster didn't work out, free it and start over */
7590         btrfs_return_cluster_to_free_space(NULL, last_ptr);
7591
7592         if (cluster_bg != bg)
7593                 btrfs_release_block_group(cluster_bg, ffe_ctl->delalloc);
7594
7595 refill_cluster:
7596         if (ffe_ctl->loop >= LOOP_NO_EMPTY_SIZE) {
7597                 spin_unlock(&last_ptr->refill_lock);
7598                 return -ENOENT;
7599         }
7600
7601         aligned_cluster = max_t(u64,
7602                         ffe_ctl->empty_cluster + ffe_ctl->empty_size,
7603                         bg->full_stripe_len);
7604         ret = btrfs_find_space_cluster(fs_info, bg, last_ptr,
7605                         ffe_ctl->search_start, ffe_ctl->num_bytes,
7606                         aligned_cluster);
7607         if (ret == 0) {
7608                 /* Now pull our allocation out of this cluster */
7609                 offset = btrfs_alloc_from_cluster(bg, last_ptr,
7610                                 ffe_ctl->num_bytes, ffe_ctl->search_start,
7611                                 &ffe_ctl->max_extent_size);
7612                 if (offset) {
7613                         /* We found one, proceed */
7614                         spin_unlock(&last_ptr->refill_lock);
7615                         trace_btrfs_reserve_extent_cluster(bg,
7616                                         ffe_ctl->search_start,
7617                                         ffe_ctl->num_bytes);
7618                         ffe_ctl->found_offset = offset;
7619                         return 0;
7620                 }
7621         } else if (!ffe_ctl->cached && ffe_ctl->loop > LOOP_CACHING_NOWAIT &&
7622                    !ffe_ctl->retry_clustered) {
7623                 spin_unlock(&last_ptr->refill_lock);
7624
7625                 ffe_ctl->retry_clustered = true;
7626                 wait_block_group_cache_progress(bg, ffe_ctl->num_bytes +
7627                                 ffe_ctl->empty_cluster + ffe_ctl->empty_size);
7628                 return -EAGAIN;
7629         }
7630         /*
7631          * At this point we either didn't find a cluster or we weren't able to
7632          * allocate a block from our cluster.  Free the cluster we've been
7633          * trying to use, and go to the next block group.
7634          */
7635         btrfs_return_cluster_to_free_space(NULL, last_ptr);
7636         spin_unlock(&last_ptr->refill_lock);
7637         return 1;
7638 }
7639
7640 /*
7641  * Return >0 to inform caller that we find nothing
7642  * Return 0 when we found an free extent and set ffe_ctrl->found_offset
7643  * Return -EAGAIN to inform caller that we need to re-search this block group
7644  */
7645 static int find_free_extent_unclustered(struct btrfs_block_group_cache *bg,
7646                 struct btrfs_free_cluster *last_ptr,
7647                 struct find_free_extent_ctl *ffe_ctl)
7648 {
7649         u64 offset;
7650
7651         /*
7652          * We are doing an unclustered allocation, set the fragmented flag so
7653          * we don't bother trying to setup a cluster again until we get more
7654          * space.
7655          */
7656         if (unlikely(last_ptr)) {
7657                 spin_lock(&last_ptr->lock);
7658                 last_ptr->fragmented = 1;
7659                 spin_unlock(&last_ptr->lock);
7660         }
7661         if (ffe_ctl->cached) {
7662                 struct btrfs_free_space_ctl *free_space_ctl;
7663
7664                 free_space_ctl = bg->free_space_ctl;
7665                 spin_lock(&free_space_ctl->tree_lock);
7666                 if (free_space_ctl->free_space <
7667                     ffe_ctl->num_bytes + ffe_ctl->empty_cluster +
7668                     ffe_ctl->empty_size) {
7669                         ffe_ctl->total_free_space = max_t(u64,
7670                                         ffe_ctl->total_free_space,
7671                                         free_space_ctl->free_space);
7672                         spin_unlock(&free_space_ctl->tree_lock);
7673                         return 1;
7674                 }
7675                 spin_unlock(&free_space_ctl->tree_lock);
7676         }
7677
7678         offset = btrfs_find_space_for_alloc(bg, ffe_ctl->search_start,
7679                         ffe_ctl->num_bytes, ffe_ctl->empty_size,
7680                         &ffe_ctl->max_extent_size);
7681
7682         /*
7683          * If we didn't find a chunk, and we haven't failed on this block group
7684          * before, and this block group is in the middle of caching and we are
7685          * ok with waiting, then go ahead and wait for progress to be made, and
7686          * set @retry_unclustered to true.
7687          *
7688          * If @retry_unclustered is true then we've already waited on this
7689          * block group once and should move on to the next block group.
7690          */
7691         if (!offset && !ffe_ctl->retry_unclustered && !ffe_ctl->cached &&
7692             ffe_ctl->loop > LOOP_CACHING_NOWAIT) {
7693                 wait_block_group_cache_progress(bg, ffe_ctl->num_bytes +
7694                                                 ffe_ctl->empty_size);
7695                 ffe_ctl->retry_unclustered = true;
7696                 return -EAGAIN;
7697         } else if (!offset) {
7698                 return 1;
7699         }
7700         ffe_ctl->found_offset = offset;
7701         return 0;
7702 }
7703
7704 /*
7705  * Return >0 means caller needs to re-search for free extent
7706  * Return 0 means we have the needed free extent.
7707  * Return <0 means we failed to locate any free extent.
7708  */
7709 static int find_free_extent_update_loop(struct btrfs_fs_info *fs_info,
7710                                         struct btrfs_free_cluster *last_ptr,
7711                                         struct btrfs_key *ins,
7712                                         struct find_free_extent_ctl *ffe_ctl,
7713                                         int full_search, bool use_cluster)
7714 {
7715         struct btrfs_root *root = fs_info->extent_root;
7716         int ret;
7717
7718         if ((ffe_ctl->loop == LOOP_CACHING_NOWAIT) &&
7719             ffe_ctl->have_caching_bg && !ffe_ctl->orig_have_caching_bg)
7720                 ffe_ctl->orig_have_caching_bg = true;
7721
7722         if (!ins->objectid && ffe_ctl->loop >= LOOP_CACHING_WAIT &&
7723             ffe_ctl->have_caching_bg)
7724                 return 1;
7725
7726         if (!ins->objectid && ++(ffe_ctl->index) < BTRFS_NR_RAID_TYPES)
7727                 return 1;
7728
7729         if (ins->objectid) {
7730                 if (!use_cluster && last_ptr) {
7731                         spin_lock(&last_ptr->lock);
7732                         last_ptr->window_start = ins->objectid;
7733                         spin_unlock(&last_ptr->lock);
7734                 }
7735                 return 0;
7736         }
7737
7738         /*
7739          * LOOP_CACHING_NOWAIT, search partially cached block groups, kicking
7740          *                      caching kthreads as we move along
7741          * LOOP_CACHING_WAIT, search everything, and wait if our bg is caching
7742          * LOOP_ALLOC_CHUNK, force a chunk allocation and try again
7743          * LOOP_NO_EMPTY_SIZE, set empty_size and empty_cluster to 0 and try
7744          *                     again
7745          */
7746         if (ffe_ctl->loop < LOOP_NO_EMPTY_SIZE) {
7747                 ffe_ctl->index = 0;
7748                 if (ffe_ctl->loop == LOOP_CACHING_NOWAIT) {
7749                         /*
7750                          * We want to skip the LOOP_CACHING_WAIT step if we
7751                          * don't have any uncached bgs and we've already done a
7752                          * full search through.
7753                          */
7754                         if (ffe_ctl->orig_have_caching_bg || !full_search)
7755                                 ffe_ctl->loop = LOOP_CACHING_WAIT;
7756                         else
7757                                 ffe_ctl->loop = LOOP_ALLOC_CHUNK;
7758                 } else {
7759                         ffe_ctl->loop++;
7760                 }
7761
7762                 if (ffe_ctl->loop == LOOP_ALLOC_CHUNK) {
7763                         struct btrfs_trans_handle *trans;
7764                         int exist = 0;
7765
7766                         trans = current->journal_info;
7767                         if (trans)
7768                                 exist = 1;
7769                         else
7770                                 trans = btrfs_join_transaction(root);
7771
7772                         if (IS_ERR(trans)) {
7773                                 ret = PTR_ERR(trans);
7774                                 return ret;
7775                         }
7776
7777                         ret = do_chunk_alloc(trans, ffe_ctl->flags,
7778                                              CHUNK_ALLOC_FORCE);
7779
7780                         /*
7781                          * If we can't allocate a new chunk we've already looped
7782                          * through at least once, move on to the NO_EMPTY_SIZE
7783                          * case.
7784                          */
7785                         if (ret == -ENOSPC)
7786                                 ffe_ctl->loop = LOOP_NO_EMPTY_SIZE;
7787
7788                         /* Do not bail out on ENOSPC since we can do more. */
7789                         if (ret < 0 && ret != -ENOSPC)
7790                                 btrfs_abort_transaction(trans, ret);
7791                         else
7792                                 ret = 0;
7793                         if (!exist)
7794                                 btrfs_end_transaction(trans);
7795                         if (ret)
7796                                 return ret;
7797                 }
7798
7799                 if (ffe_ctl->loop == LOOP_NO_EMPTY_SIZE) {
7800                         /*
7801                          * Don't loop again if we already have no empty_size and
7802                          * no empty_cluster.
7803                          */
7804                         if (ffe_ctl->empty_size == 0 &&
7805                             ffe_ctl->empty_cluster == 0)
7806                                 return -ENOSPC;
7807                         ffe_ctl->empty_size = 0;
7808                         ffe_ctl->empty_cluster = 0;
7809                 }
7810                 return 1;
7811         }
7812         return -ENOSPC;
7813 }
7814
7815 /*
7816  * walks the btree of allocated extents and find a hole of a given size.
7817  * The key ins is changed to record the hole:
7818  * ins->objectid == start position
7819  * ins->flags = BTRFS_EXTENT_ITEM_KEY
7820  * ins->offset == the size of the hole.
7821  * Any available blocks before search_start are skipped.
7822  *
7823  * If there is no suitable free space, we will record the max size of
7824  * the free space extent currently.
7825  *
7826  * The overall logic and call chain:
7827  *
7828  * find_free_extent()
7829  * |- Iterate through all block groups
7830  * |  |- Get a valid block group
7831  * |  |- Try to do clustered allocation in that block group
7832  * |  |- Try to do unclustered allocation in that block group
7833  * |  |- Check if the result is valid
7834  * |  |  |- If valid, then exit
7835  * |  |- Jump to next block group
7836  * |
7837  * |- Push harder to find free extents
7838  *    |- If not found, re-iterate all block groups
7839  */
7840 static noinline int find_free_extent(struct btrfs_fs_info *fs_info,
7841                                 u64 ram_bytes, u64 num_bytes, u64 empty_size,
7842                                 u64 hint_byte, struct btrfs_key *ins,
7843                                 u64 flags, int delalloc)
7844 {
7845         int ret = 0;
7846         struct btrfs_free_cluster *last_ptr = NULL;
7847         struct btrfs_block_group_cache *block_group = NULL;
7848         struct find_free_extent_ctl ffe_ctl = {0};
7849         struct btrfs_space_info *space_info;
7850         bool use_cluster = true;
7851         bool full_search = false;
7852
7853         WARN_ON(num_bytes < fs_info->sectorsize);
7854
7855         ffe_ctl.ram_bytes = ram_bytes;
7856         ffe_ctl.num_bytes = num_bytes;
7857         ffe_ctl.empty_size = empty_size;
7858         ffe_ctl.flags = flags;
7859         ffe_ctl.search_start = 0;
7860         ffe_ctl.retry_clustered = false;
7861         ffe_ctl.retry_unclustered = false;
7862         ffe_ctl.delalloc = delalloc;
7863         ffe_ctl.index = btrfs_bg_flags_to_raid_index(flags);
7864         ffe_ctl.have_caching_bg = false;
7865         ffe_ctl.orig_have_caching_bg = false;
7866         ffe_ctl.found_offset = 0;
7867
7868         ins->type = BTRFS_EXTENT_ITEM_KEY;
7869         ins->objectid = 0;
7870         ins->offset = 0;
7871
7872         trace_find_free_extent(fs_info, num_bytes, empty_size, flags);
7873
7874         space_info = __find_space_info(fs_info, flags);
7875         if (!space_info) {
7876                 btrfs_err(fs_info, "No space info for %llu", flags);
7877                 return -ENOSPC;
7878         }
7879
7880         /*
7881          * If our free space is heavily fragmented we may not be able to make
7882          * big contiguous allocations, so instead of doing the expensive search
7883          * for free space, simply return ENOSPC with our max_extent_size so we
7884          * can go ahead and search for a more manageable chunk.
7885          *
7886          * If our max_extent_size is large enough for our allocation simply
7887          * disable clustering since we will likely not be able to find enough
7888          * space to create a cluster and induce latency trying.
7889          */
7890         if (unlikely(space_info->max_extent_size)) {
7891                 spin_lock(&space_info->lock);
7892                 if (space_info->max_extent_size &&
7893                     num_bytes > space_info->max_extent_size) {
7894                         ins->offset = space_info->max_extent_size;
7895                         spin_unlock(&space_info->lock);
7896                         return -ENOSPC;
7897                 } else if (space_info->max_extent_size) {
7898                         use_cluster = false;
7899                 }
7900                 spin_unlock(&space_info->lock);
7901         }
7902
7903         last_ptr = fetch_cluster_info(fs_info, space_info,
7904                                       &ffe_ctl.empty_cluster);
7905         if (last_ptr) {
7906                 spin_lock(&last_ptr->lock);
7907                 if (last_ptr->block_group)
7908                         hint_byte = last_ptr->window_start;
7909                 if (last_ptr->fragmented) {
7910                         /*
7911                          * We still set window_start so we can keep track of the
7912                          * last place we found an allocation to try and save
7913                          * some time.
7914                          */
7915                         hint_byte = last_ptr->window_start;
7916                         use_cluster = false;
7917                 }
7918                 spin_unlock(&last_ptr->lock);
7919         }
7920
7921         ffe_ctl.search_start = max(ffe_ctl.search_start,
7922                                    first_logical_byte(fs_info, 0));
7923         ffe_ctl.search_start = max(ffe_ctl.search_start, hint_byte);
7924         if (ffe_ctl.search_start == hint_byte) {
7925                 block_group = btrfs_lookup_block_group(fs_info,
7926                                                        ffe_ctl.search_start);
7927                 /*
7928                  * we don't want to use the block group if it doesn't match our
7929                  * allocation bits, or if its not cached.
7930                  *
7931                  * However if we are re-searching with an ideal block group
7932                  * picked out then we don't care that the block group is cached.
7933                  */
7934                 if (block_group && block_group_bits(block_group, flags) &&
7935                     block_group->cached != BTRFS_CACHE_NO) {
7936                         down_read(&space_info->groups_sem);
7937                         if (list_empty(&block_group->list) ||
7938                             block_group->ro) {
7939                                 /*
7940                                  * someone is removing this block group,
7941                                  * we can't jump into the have_block_group
7942                                  * target because our list pointers are not
7943                                  * valid
7944                                  */
7945                                 btrfs_put_block_group(block_group);
7946                                 up_read(&space_info->groups_sem);
7947                         } else {
7948                                 ffe_ctl.index = btrfs_bg_flags_to_raid_index(
7949                                                 block_group->flags);
7950                                 btrfs_lock_block_group(block_group, delalloc);
7951                                 goto have_block_group;
7952                         }
7953                 } else if (block_group) {
7954                         btrfs_put_block_group(block_group);
7955                 }
7956         }
7957 search:
7958         ffe_ctl.have_caching_bg = false;
7959         if (ffe_ctl.index == btrfs_bg_flags_to_raid_index(flags) ||
7960             ffe_ctl.index == 0)
7961                 full_search = true;
7962         down_read(&space_info->groups_sem);
7963         list_for_each_entry(block_group,
7964                             &space_info->block_groups[ffe_ctl.index], list) {
7965                 /* If the block group is read-only, we can skip it entirely. */
7966                 if (unlikely(block_group->ro))
7967                         continue;
7968
7969                 btrfs_grab_block_group(block_group, delalloc);
7970                 ffe_ctl.search_start = block_group->key.objectid;
7971
7972                 /*
7973                  * this can happen if we end up cycling through all the
7974                  * raid types, but we want to make sure we only allocate
7975                  * for the proper type.
7976                  */
7977                 if (!block_group_bits(block_group, flags)) {
7978                         u64 extra = BTRFS_BLOCK_GROUP_DUP |
7979                                 BTRFS_BLOCK_GROUP_RAID1 |
7980                                 BTRFS_BLOCK_GROUP_RAID5 |
7981                                 BTRFS_BLOCK_GROUP_RAID6 |
7982                                 BTRFS_BLOCK_GROUP_RAID10;
7983
7984                         /*
7985                          * if they asked for extra copies and this block group
7986                          * doesn't provide them, bail.  This does allow us to
7987                          * fill raid0 from raid1.
7988                          */
7989                         if ((flags & extra) && !(block_group->flags & extra))
7990                                 goto loop;
7991                 }
7992
7993 have_block_group:
7994                 ffe_ctl.cached = block_group_cache_done(block_group);
7995                 if (unlikely(!ffe_ctl.cached)) {
7996                         ffe_ctl.have_caching_bg = true;
7997                         ret = cache_block_group(block_group, 0);
7998                         BUG_ON(ret < 0);
7999                         ret = 0;
8000                 }
8001
8002                 if (unlikely(block_group->cached == BTRFS_CACHE_ERROR))
8003                         goto loop;
8004
8005                 /*
8006                  * Ok we want to try and use the cluster allocator, so
8007                  * lets look there
8008                  */
8009                 if (last_ptr && use_cluster) {
8010                         struct btrfs_block_group_cache *cluster_bg = NULL;
8011
8012                         ret = find_free_extent_clustered(block_group, last_ptr,
8013                                                          &ffe_ctl, &cluster_bg);
8014
8015                         if (ret == 0) {
8016                                 if (cluster_bg && cluster_bg != block_group) {
8017                                         btrfs_release_block_group(block_group,
8018                                                                   delalloc);
8019                                         block_group = cluster_bg;
8020                                 }
8021                                 goto checks;
8022                         } else if (ret == -EAGAIN) {
8023                                 goto have_block_group;
8024                         } else if (ret > 0) {
8025                                 goto loop;
8026                         }
8027                         /* ret == -ENOENT case falls through */
8028                 }
8029
8030                 ret = find_free_extent_unclustered(block_group, last_ptr,
8031                                                    &ffe_ctl);
8032                 if (ret == -EAGAIN)
8033                         goto have_block_group;
8034                 else if (ret > 0)
8035                         goto loop;
8036                 /* ret == 0 case falls through */
8037 checks:
8038                 ffe_ctl.search_start = round_up(ffe_ctl.found_offset,
8039                                              fs_info->stripesize);
8040
8041                 /* move on to the next group */
8042                 if (ffe_ctl.search_start + num_bytes >
8043                     block_group->key.objectid + block_group->key.offset) {
8044                         btrfs_add_free_space(block_group, ffe_ctl.found_offset,
8045                                              num_bytes);
8046                         goto loop;
8047                 }
8048
8049                 if (ffe_ctl.found_offset < ffe_ctl.search_start)
8050                         btrfs_add_free_space(block_group, ffe_ctl.found_offset,
8051                                 ffe_ctl.search_start - ffe_ctl.found_offset);
8052
8053                 ret = btrfs_add_reserved_bytes(block_group, ram_bytes,
8054                                 num_bytes, delalloc);
8055                 if (ret == -EAGAIN) {
8056                         btrfs_add_free_space(block_group, ffe_ctl.found_offset,
8057                                              num_bytes);
8058                         goto loop;
8059                 }
8060                 btrfs_inc_block_group_reservations(block_group);
8061
8062                 /* we are all good, lets return */
8063                 ins->objectid = ffe_ctl.search_start;
8064                 ins->offset = num_bytes;
8065
8066                 trace_btrfs_reserve_extent(block_group, ffe_ctl.search_start,
8067                                            num_bytes);
8068                 btrfs_release_block_group(block_group, delalloc);
8069                 break;
8070 loop:
8071                 ffe_ctl.retry_clustered = false;
8072                 ffe_ctl.retry_unclustered = false;
8073                 BUG_ON(btrfs_bg_flags_to_raid_index(block_group->flags) !=
8074                        ffe_ctl.index);
8075                 btrfs_release_block_group(block_group, delalloc);
8076                 cond_resched();
8077         }
8078         up_read(&space_info->groups_sem);
8079
8080         ret = find_free_extent_update_loop(fs_info, last_ptr, ins, &ffe_ctl,
8081                                            full_search, use_cluster);
8082         if (ret > 0)
8083                 goto search;
8084
8085         if (ret == -ENOSPC) {
8086                 /*
8087                  * Use ffe_ctl->total_free_space as fallback if we can't find
8088                  * any contiguous hole.
8089                  */
8090                 if (!ffe_ctl.max_extent_size)
8091                         ffe_ctl.max_extent_size = ffe_ctl.total_free_space;
8092                 spin_lock(&space_info->lock);
8093                 space_info->max_extent_size = ffe_ctl.max_extent_size;
8094                 spin_unlock(&space_info->lock);
8095                 ins->offset = ffe_ctl.max_extent_size;
8096         }
8097         return ret;
8098 }
8099
8100 #define DUMP_BLOCK_RSV(fs_info, rsv_name)                               \
8101 do {                                                                    \
8102         struct btrfs_block_rsv *__rsv = &(fs_info)->rsv_name;           \
8103         spin_lock(&__rsv->lock);                                        \
8104         btrfs_info(fs_info, #rsv_name ": size %llu reserved %llu",      \
8105                    __rsv->size, __rsv->reserved);                       \
8106         spin_unlock(&__rsv->lock);                                      \
8107 } while (0)
8108
8109 static void dump_space_info(struct btrfs_fs_info *fs_info,
8110                             struct btrfs_space_info *info, u64 bytes,
8111                             int dump_block_groups)
8112 {
8113         struct btrfs_block_group_cache *cache;
8114         int index = 0;
8115
8116         spin_lock(&info->lock);
8117         btrfs_info(fs_info, "space_info %llu has %llu free, is %sfull",
8118                    info->flags,
8119                    info->total_bytes - btrfs_space_info_used(info, true),
8120                    info->full ? "" : "not ");
8121         btrfs_info(fs_info,
8122                 "space_info total=%llu, used=%llu, pinned=%llu, reserved=%llu, may_use=%llu, readonly=%llu",
8123                 info->total_bytes, info->bytes_used, info->bytes_pinned,
8124                 info->bytes_reserved, info->bytes_may_use,
8125                 info->bytes_readonly);
8126         spin_unlock(&info->lock);
8127
8128         DUMP_BLOCK_RSV(fs_info, global_block_rsv);
8129         DUMP_BLOCK_RSV(fs_info, trans_block_rsv);
8130         DUMP_BLOCK_RSV(fs_info, chunk_block_rsv);
8131         DUMP_BLOCK_RSV(fs_info, delayed_block_rsv);
8132         DUMP_BLOCK_RSV(fs_info, delayed_refs_rsv);
8133
8134         if (!dump_block_groups)
8135                 return;
8136
8137         down_read(&info->groups_sem);
8138 again:
8139         list_for_each_entry(cache, &info->block_groups[index], list) {
8140                 spin_lock(&cache->lock);
8141                 btrfs_info(fs_info,
8142                         "block group %llu has %llu bytes, %llu used %llu pinned %llu reserved %s",
8143                         cache->key.objectid, cache->key.offset,
8144                         btrfs_block_group_used(&cache->item), cache->pinned,
8145                         cache->reserved, cache->ro ? "[readonly]" : "");
8146                 btrfs_dump_free_space(cache, bytes);
8147                 spin_unlock(&cache->lock);
8148         }
8149         if (++index < BTRFS_NR_RAID_TYPES)
8150                 goto again;
8151         up_read(&info->groups_sem);
8152 }
8153
8154 /*
8155  * btrfs_reserve_extent - entry point to the extent allocator. Tries to find a
8156  *                        hole that is at least as big as @num_bytes.
8157  *
8158  * @root           -    The root that will contain this extent
8159  *
8160  * @ram_bytes      -    The amount of space in ram that @num_bytes take. This
8161  *                      is used for accounting purposes. This value differs
8162  *                      from @num_bytes only in the case of compressed extents.
8163  *
8164  * @num_bytes      -    Number of bytes to allocate on-disk.
8165  *
8166  * @min_alloc_size -    Indicates the minimum amount of space that the
8167  *                      allocator should try to satisfy. In some cases
8168  *                      @num_bytes may be larger than what is required and if
8169  *                      the filesystem is fragmented then allocation fails.
8170  *                      However, the presence of @min_alloc_size gives a
8171  *                      chance to try and satisfy the smaller allocation.
8172  *
8173  * @empty_size     -    A hint that you plan on doing more COW. This is the
8174  *                      size in bytes the allocator should try to find free
8175  *                      next to the block it returns.  This is just a hint and
8176  *                      may be ignored by the allocator.
8177  *
8178  * @hint_byte      -    Hint to the allocator to start searching above the byte
8179  *                      address passed. It might be ignored.
8180  *
8181  * @ins            -    This key is modified to record the found hole. It will
8182  *                      have the following values:
8183  *                      ins->objectid == start position
8184  *                      ins->flags = BTRFS_EXTENT_ITEM_KEY
8185  *                      ins->offset == the size of the hole.
8186  *
8187  * @is_data        -    Boolean flag indicating whether an extent is
8188  *                      allocated for data (true) or metadata (false)
8189  *
8190  * @delalloc       -    Boolean flag indicating whether this allocation is for
8191  *                      delalloc or not. If 'true' data_rwsem of block groups
8192  *                      is going to be acquired.
8193  *
8194  *
8195  * Returns 0 when an allocation succeeded or < 0 when an error occurred. In
8196  * case -ENOSPC is returned then @ins->offset will contain the size of the
8197  * largest available hole the allocator managed to find.
8198  */
8199 int btrfs_reserve_extent(struct btrfs_root *root, u64 ram_bytes,
8200                          u64 num_bytes, u64 min_alloc_size,
8201                          u64 empty_size, u64 hint_byte,
8202                          struct btrfs_key *ins, int is_data, int delalloc)
8203 {
8204         struct btrfs_fs_info *fs_info = root->fs_info;
8205         bool final_tried = num_bytes == min_alloc_size;
8206         u64 flags;
8207         int ret;
8208
8209         flags = get_alloc_profile_by_root(root, is_data);
8210 again:
8211         WARN_ON(num_bytes < fs_info->sectorsize);
8212         ret = find_free_extent(fs_info, ram_bytes, num_bytes, empty_size,
8213                                hint_byte, ins, flags, delalloc);
8214         if (!ret && !is_data) {
8215                 btrfs_dec_block_group_reservations(fs_info, ins->objectid);
8216         } else if (ret == -ENOSPC) {
8217                 if (!final_tried && ins->offset) {
8218                         num_bytes = min(num_bytes >> 1, ins->offset);
8219                         num_bytes = round_down(num_bytes,
8220                                                fs_info->sectorsize);
8221                         num_bytes = max(num_bytes, min_alloc_size);
8222                         ram_bytes = num_bytes;
8223                         if (num_bytes == min_alloc_size)
8224                                 final_tried = true;
8225                         goto again;
8226                 } else if (btrfs_test_opt(fs_info, ENOSPC_DEBUG)) {
8227                         struct btrfs_space_info *sinfo;
8228
8229                         sinfo = __find_space_info(fs_info, flags);
8230                         btrfs_err(fs_info,
8231                                   "allocation failed flags %llu, wanted %llu",
8232                                   flags, num_bytes);
8233                         if (sinfo)
8234                                 dump_space_info(fs_info, sinfo, num_bytes, 1);
8235                 }
8236         }
8237
8238         return ret;
8239 }
8240
8241 static int __btrfs_free_reserved_extent(struct btrfs_fs_info *fs_info,
8242                                         u64 start, u64 len,
8243                                         int pin, int delalloc)
8244 {
8245         struct btrfs_block_group_cache *cache;
8246         int ret = 0;
8247
8248         cache = btrfs_lookup_block_group(fs_info, start);
8249         if (!cache) {
8250                 btrfs_err(fs_info, "Unable to find block group for %llu",
8251                           start);
8252                 return -ENOSPC;
8253         }
8254
8255         if (pin)
8256                 pin_down_extent(fs_info, cache, start, len, 1);
8257         else {
8258                 if (btrfs_test_opt(fs_info, DISCARD))
8259                         ret = btrfs_discard_extent(fs_info, start, len, NULL);
8260                 btrfs_add_free_space(cache, start, len);
8261                 btrfs_free_reserved_bytes(cache, len, delalloc);
8262                 trace_btrfs_reserved_extent_free(fs_info, start, len);
8263         }
8264
8265         btrfs_put_block_group(cache);
8266         return ret;
8267 }
8268
8269 int btrfs_free_reserved_extent(struct btrfs_fs_info *fs_info,
8270                                u64 start, u64 len, int delalloc)
8271 {
8272         return __btrfs_free_reserved_extent(fs_info, start, len, 0, delalloc);
8273 }
8274
8275 int btrfs_free_and_pin_reserved_extent(struct btrfs_fs_info *fs_info,
8276                                        u64 start, u64 len)
8277 {
8278         return __btrfs_free_reserved_extent(fs_info, start, len, 1, 0);
8279 }
8280
8281 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
8282                                       u64 parent, u64 root_objectid,
8283                                       u64 flags, u64 owner, u64 offset,
8284                                       struct btrfs_key *ins, int ref_mod)
8285 {
8286         struct btrfs_fs_info *fs_info = trans->fs_info;
8287         int ret;
8288         struct btrfs_extent_item *extent_item;
8289         struct btrfs_extent_inline_ref *iref;
8290         struct btrfs_path *path;
8291         struct extent_buffer *leaf;
8292         int type;
8293         u32 size;
8294
8295         if (parent > 0)
8296                 type = BTRFS_SHARED_DATA_REF_KEY;
8297         else
8298                 type = BTRFS_EXTENT_DATA_REF_KEY;
8299
8300         size = sizeof(*extent_item) + btrfs_extent_inline_ref_size(type);
8301
8302         path = btrfs_alloc_path();
8303         if (!path)
8304                 return -ENOMEM;
8305
8306         path->leave_spinning = 1;
8307         ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path,
8308                                       ins, size);
8309         if (ret) {
8310                 btrfs_free_path(path);
8311                 return ret;
8312         }
8313
8314         leaf = path->nodes[0];
8315         extent_item = btrfs_item_ptr(leaf, path->slots[0],
8316                                      struct btrfs_extent_item);
8317         btrfs_set_extent_refs(leaf, extent_item, ref_mod);
8318         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
8319         btrfs_set_extent_flags(leaf, extent_item,
8320                                flags | BTRFS_EXTENT_FLAG_DATA);
8321
8322         iref = (struct btrfs_extent_inline_ref *)(extent_item + 1);
8323         btrfs_set_extent_inline_ref_type(leaf, iref, type);
8324         if (parent > 0) {
8325                 struct btrfs_shared_data_ref *ref;
8326                 ref = (struct btrfs_shared_data_ref *)(iref + 1);
8327                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
8328                 btrfs_set_shared_data_ref_count(leaf, ref, ref_mod);
8329         } else {
8330                 struct btrfs_extent_data_ref *ref;
8331                 ref = (struct btrfs_extent_data_ref *)(&iref->offset);
8332                 btrfs_set_extent_data_ref_root(leaf, ref, root_objectid);
8333                 btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
8334                 btrfs_set_extent_data_ref_offset(leaf, ref, offset);
8335                 btrfs_set_extent_data_ref_count(leaf, ref, ref_mod);
8336         }
8337
8338         btrfs_mark_buffer_dirty(path->nodes[0]);
8339         btrfs_free_path(path);
8340
8341         ret = remove_from_free_space_tree(trans, ins->objectid, ins->offset);
8342         if (ret)
8343                 return ret;
8344
8345         ret = update_block_group(trans, fs_info, ins->objectid, ins->offset, 1);
8346         if (ret) { /* -ENOENT, logic error */
8347                 btrfs_err(fs_info, "update block group failed for %llu %llu",
8348                         ins->objectid, ins->offset);
8349                 BUG();
8350         }
8351         trace_btrfs_reserved_extent_alloc(fs_info, ins->objectid, ins->offset);
8352         return ret;
8353 }
8354
8355 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
8356                                      struct btrfs_delayed_ref_node *node,
8357                                      struct btrfs_delayed_extent_op *extent_op)
8358 {
8359         struct btrfs_fs_info *fs_info = trans->fs_info;
8360         int ret;
8361         struct btrfs_extent_item *extent_item;
8362         struct btrfs_key extent_key;
8363         struct btrfs_tree_block_info *block_info;
8364         struct btrfs_extent_inline_ref *iref;
8365         struct btrfs_path *path;
8366         struct extent_buffer *leaf;
8367         struct btrfs_delayed_tree_ref *ref;
8368         u32 size = sizeof(*extent_item) + sizeof(*iref);
8369         u64 num_bytes;
8370         u64 flags = extent_op->flags_to_set;
8371         bool skinny_metadata = btrfs_fs_incompat(fs_info, SKINNY_METADATA);
8372
8373         ref = btrfs_delayed_node_to_tree_ref(node);
8374
8375         extent_key.objectid = node->bytenr;
8376         if (skinny_metadata) {
8377                 extent_key.offset = ref->level;
8378                 extent_key.type = BTRFS_METADATA_ITEM_KEY;
8379                 num_bytes = fs_info->nodesize;
8380         } else {
8381                 extent_key.offset = node->num_bytes;
8382                 extent_key.type = BTRFS_EXTENT_ITEM_KEY;
8383                 size += sizeof(*block_info);
8384                 num_bytes = node->num_bytes;
8385         }
8386
8387         path = btrfs_alloc_path();
8388         if (!path)
8389                 return -ENOMEM;
8390
8391         path->leave_spinning = 1;
8392         ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path,
8393                                       &extent_key, size);
8394         if (ret) {
8395                 btrfs_free_path(path);
8396                 return ret;
8397         }
8398
8399         leaf = path->nodes[0];
8400         extent_item = btrfs_item_ptr(leaf, path->slots[0],
8401                                      struct btrfs_extent_item);
8402         btrfs_set_extent_refs(leaf, extent_item, 1);
8403         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
8404         btrfs_set_extent_flags(leaf, extent_item,
8405                                flags | BTRFS_EXTENT_FLAG_TREE_BLOCK);
8406
8407         if (skinny_metadata) {
8408                 iref = (struct btrfs_extent_inline_ref *)(extent_item + 1);
8409         } else {
8410                 block_info = (struct btrfs_tree_block_info *)(extent_item + 1);
8411                 btrfs_set_tree_block_key(leaf, block_info, &extent_op->key);
8412                 btrfs_set_tree_block_level(leaf, block_info, ref->level);
8413                 iref = (struct btrfs_extent_inline_ref *)(block_info + 1);
8414         }
8415
8416         if (node->type == BTRFS_SHARED_BLOCK_REF_KEY) {
8417                 BUG_ON(!(flags & BTRFS_BLOCK_FLAG_FULL_BACKREF));
8418                 btrfs_set_extent_inline_ref_type(leaf, iref,
8419                                                  BTRFS_SHARED_BLOCK_REF_KEY);
8420                 btrfs_set_extent_inline_ref_offset(leaf, iref, ref->parent);
8421         } else {
8422                 btrfs_set_extent_inline_ref_type(leaf, iref,
8423                                                  BTRFS_TREE_BLOCK_REF_KEY);
8424                 btrfs_set_extent_inline_ref_offset(leaf, iref, ref->root);
8425         }
8426
8427         btrfs_mark_buffer_dirty(leaf);
8428         btrfs_free_path(path);
8429
8430         ret = remove_from_free_space_tree(trans, extent_key.objectid,
8431                                           num_bytes);
8432         if (ret)
8433                 return ret;
8434
8435         ret = update_block_group(trans, fs_info, extent_key.objectid,
8436                                  fs_info->nodesize, 1);
8437         if (ret) { /* -ENOENT, logic error */
8438                 btrfs_err(fs_info, "update block group failed for %llu %llu",
8439                         extent_key.objectid, extent_key.offset);
8440                 BUG();
8441         }
8442
8443         trace_btrfs_reserved_extent_alloc(fs_info, extent_key.objectid,
8444                                           fs_info->nodesize);
8445         return ret;
8446 }
8447
8448 int btrfs_alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
8449                                      struct btrfs_root *root, u64 owner,
8450                                      u64 offset, u64 ram_bytes,
8451                                      struct btrfs_key *ins)
8452 {
8453         int ret;
8454
8455         BUG_ON(root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID);
8456
8457         btrfs_ref_tree_mod(root, ins->objectid, ins->offset, 0,
8458                            root->root_key.objectid, owner, offset,
8459                            BTRFS_ADD_DELAYED_EXTENT);
8460
8461         ret = btrfs_add_delayed_data_ref(trans, ins->objectid,
8462                                          ins->offset, 0,
8463                                          root->root_key.objectid, owner,
8464                                          offset, ram_bytes,
8465                                          BTRFS_ADD_DELAYED_EXTENT, NULL, NULL);
8466         return ret;
8467 }
8468
8469 /*
8470  * this is used by the tree logging recovery code.  It records that
8471  * an extent has been allocated and makes sure to clear the free
8472  * space cache bits as well
8473  */
8474 int btrfs_alloc_logged_file_extent(struct btrfs_trans_handle *trans,
8475                                    u64 root_objectid, u64 owner, u64 offset,
8476                                    struct btrfs_key *ins)
8477 {
8478         struct btrfs_fs_info *fs_info = trans->fs_info;
8479         int ret;
8480         struct btrfs_block_group_cache *block_group;
8481         struct btrfs_space_info *space_info;
8482
8483         /*
8484          * Mixed block groups will exclude before processing the log so we only
8485          * need to do the exclude dance if this fs isn't mixed.
8486          */
8487         if (!btrfs_fs_incompat(fs_info, MIXED_GROUPS)) {
8488                 ret = __exclude_logged_extent(fs_info, ins->objectid,
8489                                               ins->offset);
8490                 if (ret)
8491                         return ret;
8492         }
8493
8494         block_group = btrfs_lookup_block_group(fs_info, ins->objectid);
8495         if (!block_group)
8496                 return -EINVAL;
8497
8498         space_info = block_group->space_info;
8499         spin_lock(&space_info->lock);
8500         spin_lock(&block_group->lock);
8501         space_info->bytes_reserved += ins->offset;
8502         block_group->reserved += ins->offset;
8503         spin_unlock(&block_group->lock);
8504         spin_unlock(&space_info->lock);
8505
8506         ret = alloc_reserved_file_extent(trans, 0, root_objectid, 0, owner,
8507                                          offset, ins, 1);
8508         btrfs_put_block_group(block_group);
8509         return ret;
8510 }
8511
8512 static struct extent_buffer *
8513 btrfs_init_new_buffer(struct btrfs_trans_handle *trans, struct btrfs_root *root,
8514                       u64 bytenr, int level, u64 owner)
8515 {
8516         struct btrfs_fs_info *fs_info = root->fs_info;
8517         struct extent_buffer *buf;
8518
8519         buf = btrfs_find_create_tree_block(fs_info, bytenr);
8520         if (IS_ERR(buf))
8521                 return buf;
8522
8523         /*
8524          * Extra safety check in case the extent tree is corrupted and extent
8525          * allocator chooses to use a tree block which is already used and
8526          * locked.
8527          */
8528         if (buf->lock_owner == current->pid) {
8529                 btrfs_err_rl(fs_info,
8530 "tree block %llu owner %llu already locked by pid=%d, extent tree corruption detected",
8531                         buf->start, btrfs_header_owner(buf), current->pid);
8532                 free_extent_buffer(buf);
8533                 return ERR_PTR(-EUCLEAN);
8534         }
8535
8536         btrfs_set_buffer_lockdep_class(root->root_key.objectid, buf, level);
8537         btrfs_tree_lock(buf);
8538         clean_tree_block(fs_info, buf);
8539         clear_bit(EXTENT_BUFFER_STALE, &buf->bflags);
8540
8541         btrfs_set_lock_blocking_write(buf);
8542         set_extent_buffer_uptodate(buf);
8543
8544         memzero_extent_buffer(buf, 0, sizeof(struct btrfs_header));
8545         btrfs_set_header_level(buf, level);
8546         btrfs_set_header_bytenr(buf, buf->start);
8547         btrfs_set_header_generation(buf, trans->transid);
8548         btrfs_set_header_backref_rev(buf, BTRFS_MIXED_BACKREF_REV);
8549         btrfs_set_header_owner(buf, owner);
8550         write_extent_buffer_fsid(buf, fs_info->fs_devices->metadata_uuid);
8551         write_extent_buffer_chunk_tree_uuid(buf, fs_info->chunk_tree_uuid);
8552         if (root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID) {
8553                 buf->log_index = root->log_transid % 2;
8554                 /*
8555                  * we allow two log transactions at a time, use different
8556                  * EXTENT bit to differentiate dirty pages.
8557                  */
8558                 if (buf->log_index == 0)
8559                         set_extent_dirty(&root->dirty_log_pages, buf->start,
8560                                         buf->start + buf->len - 1, GFP_NOFS);
8561                 else
8562                         set_extent_new(&root->dirty_log_pages, buf->start,
8563                                         buf->start + buf->len - 1);
8564         } else {
8565                 buf->log_index = -1;
8566                 set_extent_dirty(&trans->transaction->dirty_pages, buf->start,
8567                          buf->start + buf->len - 1, GFP_NOFS);
8568         }
8569         trans->dirty = true;
8570         /* this returns a buffer locked for blocking */
8571         return buf;
8572 }
8573
8574 static struct btrfs_block_rsv *
8575 use_block_rsv(struct btrfs_trans_handle *trans,
8576               struct btrfs_root *root, u32 blocksize)
8577 {
8578         struct btrfs_fs_info *fs_info = root->fs_info;
8579         struct btrfs_block_rsv *block_rsv;
8580         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
8581         int ret;
8582         bool global_updated = false;
8583
8584         block_rsv = get_block_rsv(trans, root);
8585
8586         if (unlikely(block_rsv->size == 0))
8587                 goto try_reserve;
8588 again:
8589         ret = block_rsv_use_bytes(block_rsv, blocksize);
8590         if (!ret)
8591                 return block_rsv;
8592
8593         if (block_rsv->failfast)
8594                 return ERR_PTR(ret);
8595
8596         if (block_rsv->type == BTRFS_BLOCK_RSV_GLOBAL && !global_updated) {
8597                 global_updated = true;
8598                 update_global_block_rsv(fs_info);
8599                 goto again;
8600         }
8601
8602         /*
8603          * The global reserve still exists to save us from ourselves, so don't
8604          * warn_on if we are short on our delayed refs reserve.
8605          */
8606         if (block_rsv->type != BTRFS_BLOCK_RSV_DELREFS &&
8607             btrfs_test_opt(fs_info, ENOSPC_DEBUG)) {
8608                 static DEFINE_RATELIMIT_STATE(_rs,
8609                                 DEFAULT_RATELIMIT_INTERVAL * 10,
8610                                 /*DEFAULT_RATELIMIT_BURST*/ 1);
8611                 if (__ratelimit(&_rs))
8612                         WARN(1, KERN_DEBUG
8613                                 "BTRFS: block rsv returned %d\n", ret);
8614         }
8615 try_reserve:
8616         ret = reserve_metadata_bytes(root, block_rsv, blocksize,
8617                                      BTRFS_RESERVE_NO_FLUSH);
8618         if (!ret)
8619                 return block_rsv;
8620         /*
8621          * If we couldn't reserve metadata bytes try and use some from
8622          * the global reserve if its space type is the same as the global
8623          * reservation.
8624          */
8625         if (block_rsv->type != BTRFS_BLOCK_RSV_GLOBAL &&
8626             block_rsv->space_info == global_rsv->space_info) {
8627                 ret = block_rsv_use_bytes(global_rsv, blocksize);
8628                 if (!ret)
8629                         return global_rsv;
8630         }
8631         return ERR_PTR(ret);
8632 }
8633
8634 static void unuse_block_rsv(struct btrfs_fs_info *fs_info,
8635                             struct btrfs_block_rsv *block_rsv, u32 blocksize)
8636 {
8637         block_rsv_add_bytes(block_rsv, blocksize, false);
8638         block_rsv_release_bytes(fs_info, block_rsv, NULL, 0, NULL);
8639 }
8640
8641 /*
8642  * finds a free extent and does all the dirty work required for allocation
8643  * returns the tree buffer or an ERR_PTR on error.
8644  */
8645 struct extent_buffer *btrfs_alloc_tree_block(struct btrfs_trans_handle *trans,
8646                                              struct btrfs_root *root,
8647                                              u64 parent, u64 root_objectid,
8648                                              const struct btrfs_disk_key *key,
8649                                              int level, u64 hint,
8650                                              u64 empty_size)
8651 {
8652         struct btrfs_fs_info *fs_info = root->fs_info;
8653         struct btrfs_key ins;
8654         struct btrfs_block_rsv *block_rsv;
8655         struct extent_buffer *buf;
8656         struct btrfs_delayed_extent_op *extent_op;
8657         u64 flags = 0;
8658         int ret;
8659         u32 blocksize = fs_info->nodesize;
8660         bool skinny_metadata = btrfs_fs_incompat(fs_info, SKINNY_METADATA);
8661
8662 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
8663         if (btrfs_is_testing(fs_info)) {
8664                 buf = btrfs_init_new_buffer(trans, root, root->alloc_bytenr,
8665                                             level, root_objectid);
8666                 if (!IS_ERR(buf))
8667                         root->alloc_bytenr += blocksize;
8668                 return buf;
8669         }
8670 #endif
8671
8672         block_rsv = use_block_rsv(trans, root, blocksize);
8673         if (IS_ERR(block_rsv))
8674                 return ERR_CAST(block_rsv);
8675
8676         ret = btrfs_reserve_extent(root, blocksize, blocksize, blocksize,
8677                                    empty_size, hint, &ins, 0, 0);
8678         if (ret)
8679                 goto out_unuse;
8680
8681         buf = btrfs_init_new_buffer(trans, root, ins.objectid, level,
8682                                     root_objectid);
8683         if (IS_ERR(buf)) {
8684                 ret = PTR_ERR(buf);
8685                 goto out_free_reserved;
8686         }
8687
8688         if (root_objectid == BTRFS_TREE_RELOC_OBJECTID) {
8689                 if (parent == 0)
8690                         parent = ins.objectid;
8691                 flags |= BTRFS_BLOCK_FLAG_FULL_BACKREF;
8692         } else
8693                 BUG_ON(parent > 0);
8694
8695         if (root_objectid != BTRFS_TREE_LOG_OBJECTID) {
8696                 extent_op = btrfs_alloc_delayed_extent_op();
8697                 if (!extent_op) {
8698                         ret = -ENOMEM;
8699                         goto out_free_buf;
8700                 }
8701                 if (key)
8702                         memcpy(&extent_op->key, key, sizeof(extent_op->key));
8703                 else
8704                         memset(&extent_op->key, 0, sizeof(extent_op->key));
8705                 extent_op->flags_to_set = flags;
8706                 extent_op->update_key = skinny_metadata ? false : true;
8707                 extent_op->update_flags = true;
8708                 extent_op->is_data = false;
8709                 extent_op->level = level;
8710
8711                 btrfs_ref_tree_mod(root, ins.objectid, ins.offset, parent,
8712                                    root_objectid, level, 0,
8713                                    BTRFS_ADD_DELAYED_EXTENT);
8714                 ret = btrfs_add_delayed_tree_ref(trans, ins.objectid,
8715                                                  ins.offset, parent,
8716                                                  root_objectid, level,
8717                                                  BTRFS_ADD_DELAYED_EXTENT,
8718                                                  extent_op, NULL, NULL);
8719                 if (ret)
8720                         goto out_free_delayed;
8721         }
8722         return buf;
8723
8724 out_free_delayed:
8725         btrfs_free_delayed_extent_op(extent_op);
8726 out_free_buf:
8727         free_extent_buffer(buf);
8728 out_free_reserved:
8729         btrfs_free_reserved_extent(fs_info, ins.objectid, ins.offset, 0);
8730 out_unuse:
8731         unuse_block_rsv(fs_info, block_rsv, blocksize);
8732         return ERR_PTR(ret);
8733 }
8734
8735 struct walk_control {
8736         u64 refs[BTRFS_MAX_LEVEL];
8737         u64 flags[BTRFS_MAX_LEVEL];
8738         struct btrfs_key update_progress;
8739         int stage;
8740         int level;
8741         int shared_level;
8742         int update_ref;
8743         int keep_locks;
8744         int reada_slot;
8745         int reada_count;
8746 };
8747
8748 #define DROP_REFERENCE  1
8749 #define UPDATE_BACKREF  2
8750
8751 static noinline void reada_walk_down(struct btrfs_trans_handle *trans,
8752                                      struct btrfs_root *root,
8753                                      struct walk_control *wc,
8754                                      struct btrfs_path *path)
8755 {
8756         struct btrfs_fs_info *fs_info = root->fs_info;
8757         u64 bytenr;
8758         u64 generation;
8759         u64 refs;
8760         u64 flags;
8761         u32 nritems;
8762         struct btrfs_key key;
8763         struct extent_buffer *eb;
8764         int ret;
8765         int slot;
8766         int nread = 0;
8767
8768         if (path->slots[wc->level] < wc->reada_slot) {
8769                 wc->reada_count = wc->reada_count * 2 / 3;
8770                 wc->reada_count = max(wc->reada_count, 2);
8771         } else {
8772                 wc->reada_count = wc->reada_count * 3 / 2;
8773                 wc->reada_count = min_t(int, wc->reada_count,
8774                                         BTRFS_NODEPTRS_PER_BLOCK(fs_info));
8775         }
8776
8777         eb = path->nodes[wc->level];
8778         nritems = btrfs_header_nritems(eb);
8779
8780         for (slot = path->slots[wc->level]; slot < nritems; slot++) {
8781                 if (nread >= wc->reada_count)
8782                         break;
8783
8784                 cond_resched();
8785                 bytenr = btrfs_node_blockptr(eb, slot);
8786                 generation = btrfs_node_ptr_generation(eb, slot);
8787
8788                 if (slot == path->slots[wc->level])
8789                         goto reada;
8790
8791                 if (wc->stage == UPDATE_BACKREF &&
8792                     generation <= root->root_key.offset)
8793                         continue;
8794
8795                 /* We don't lock the tree block, it's OK to be racy here */
8796                 ret = btrfs_lookup_extent_info(trans, fs_info, bytenr,
8797                                                wc->level - 1, 1, &refs,
8798                                                &flags);
8799                 /* We don't care about errors in readahead. */
8800                 if (ret < 0)
8801                         continue;
8802                 BUG_ON(refs == 0);
8803
8804                 if (wc->stage == DROP_REFERENCE) {
8805                         if (refs == 1)
8806                                 goto reada;
8807
8808                         if (wc->level == 1 &&
8809                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
8810                                 continue;
8811                         if (!wc->update_ref ||
8812                             generation <= root->root_key.offset)
8813                                 continue;
8814                         btrfs_node_key_to_cpu(eb, &key, slot);
8815                         ret = btrfs_comp_cpu_keys(&key,
8816                                                   &wc->update_progress);
8817                         if (ret < 0)
8818                                 continue;
8819                 } else {
8820                         if (wc->level == 1 &&
8821                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
8822                                 continue;
8823                 }
8824 reada:
8825                 readahead_tree_block(fs_info, bytenr);
8826                 nread++;
8827         }
8828         wc->reada_slot = slot;
8829 }
8830
8831 /*
8832  * helper to process tree block while walking down the tree.
8833  *
8834  * when wc->stage == UPDATE_BACKREF, this function updates
8835  * back refs for pointers in the block.
8836  *
8837  * NOTE: return value 1 means we should stop walking down.
8838  */
8839 static noinline int walk_down_proc(struct btrfs_trans_handle *trans,
8840                                    struct btrfs_root *root,
8841                                    struct btrfs_path *path,
8842                                    struct walk_control *wc, int lookup_info)
8843 {
8844         struct btrfs_fs_info *fs_info = root->fs_info;
8845         int level = wc->level;
8846         struct extent_buffer *eb = path->nodes[level];
8847         u64 flag = BTRFS_BLOCK_FLAG_FULL_BACKREF;
8848         int ret;
8849
8850         if (wc->stage == UPDATE_BACKREF &&
8851             btrfs_header_owner(eb) != root->root_key.objectid)
8852                 return 1;
8853
8854         /*
8855          * when reference count of tree block is 1, it won't increase
8856          * again. once full backref flag is set, we never clear it.
8857          */
8858         if (lookup_info &&
8859             ((wc->stage == DROP_REFERENCE && wc->refs[level] != 1) ||
8860              (wc->stage == UPDATE_BACKREF && !(wc->flags[level] & flag)))) {
8861                 BUG_ON(!path->locks[level]);
8862                 ret = btrfs_lookup_extent_info(trans, fs_info,
8863                                                eb->start, level, 1,
8864                                                &wc->refs[level],
8865                                                &wc->flags[level]);
8866                 BUG_ON(ret == -ENOMEM);
8867                 if (ret)
8868                         return ret;
8869                 BUG_ON(wc->refs[level] == 0);
8870         }
8871
8872         if (wc->stage == DROP_REFERENCE) {
8873                 if (wc->refs[level] > 1)
8874                         return 1;
8875
8876                 if (path->locks[level] && !wc->keep_locks) {
8877                         btrfs_tree_unlock_rw(eb, path->locks[level]);
8878                         path->locks[level] = 0;
8879                 }
8880                 return 0;
8881         }
8882
8883         /* wc->stage == UPDATE_BACKREF */
8884         if (!(wc->flags[level] & flag)) {
8885                 BUG_ON(!path->locks[level]);
8886                 ret = btrfs_inc_ref(trans, root, eb, 1);
8887                 BUG_ON(ret); /* -ENOMEM */
8888                 ret = btrfs_dec_ref(trans, root, eb, 0);
8889                 BUG_ON(ret); /* -ENOMEM */
8890                 ret = btrfs_set_disk_extent_flags(trans, fs_info, eb->start,
8891                                                   eb->len, flag,
8892                                                   btrfs_header_level(eb), 0);
8893                 BUG_ON(ret); /* -ENOMEM */
8894                 wc->flags[level] |= flag;
8895         }
8896
8897         /*
8898          * the block is shared by multiple trees, so it's not good to
8899          * keep the tree lock
8900          */
8901         if (path->locks[level] && level > 0) {
8902                 btrfs_tree_unlock_rw(eb, path->locks[level]);
8903                 path->locks[level] = 0;
8904         }
8905         return 0;
8906 }
8907
8908 /*
8909  * helper to process tree block pointer.
8910  *
8911  * when wc->stage == DROP_REFERENCE, this function checks
8912  * reference count of the block pointed to. if the block
8913  * is shared and we need update back refs for the subtree
8914  * rooted at the block, this function changes wc->stage to
8915  * UPDATE_BACKREF. if the block is shared and there is no
8916  * need to update back, this function drops the reference
8917  * to the block.
8918  *
8919  * NOTE: return value 1 means we should stop walking down.
8920  */
8921 static noinline int do_walk_down(struct btrfs_trans_handle *trans,
8922                                  struct btrfs_root *root,
8923                                  struct btrfs_path *path,
8924                                  struct walk_control *wc, int *lookup_info)
8925 {
8926         struct btrfs_fs_info *fs_info = root->fs_info;
8927         u64 bytenr;
8928         u64 generation;
8929         u64 parent;
8930         struct btrfs_key key;
8931         struct btrfs_key first_key;
8932         struct extent_buffer *next;
8933         int level = wc->level;
8934         int reada = 0;
8935         int ret = 0;
8936         bool need_account = false;
8937
8938         generation = btrfs_node_ptr_generation(path->nodes[level],
8939                                                path->slots[level]);
8940         /*
8941          * if the lower level block was created before the snapshot
8942          * was created, we know there is no need to update back refs
8943          * for the subtree
8944          */
8945         if (wc->stage == UPDATE_BACKREF &&
8946             generation <= root->root_key.offset) {
8947                 *lookup_info = 1;
8948                 return 1;
8949         }
8950
8951         bytenr = btrfs_node_blockptr(path->nodes[level], path->slots[level]);
8952         btrfs_node_key_to_cpu(path->nodes[level], &first_key,
8953                               path->slots[level]);
8954
8955         next = find_extent_buffer(fs_info, bytenr);
8956         if (!next) {
8957                 next = btrfs_find_create_tree_block(fs_info, bytenr);
8958                 if (IS_ERR(next))
8959                         return PTR_ERR(next);
8960
8961                 btrfs_set_buffer_lockdep_class(root->root_key.objectid, next,
8962                                                level - 1);
8963                 reada = 1;
8964         }
8965         btrfs_tree_lock(next);
8966         btrfs_set_lock_blocking_write(next);
8967
8968         ret = btrfs_lookup_extent_info(trans, fs_info, bytenr, level - 1, 1,
8969                                        &wc->refs[level - 1],
8970                                        &wc->flags[level - 1]);
8971         if (ret < 0)
8972                 goto out_unlock;
8973
8974         if (unlikely(wc->refs[level - 1] == 0)) {
8975                 btrfs_err(fs_info, "Missing references.");
8976                 ret = -EIO;
8977                 goto out_unlock;
8978         }
8979         *lookup_info = 0;
8980
8981         if (wc->stage == DROP_REFERENCE) {
8982                 if (wc->refs[level - 1] > 1) {
8983                         need_account = true;
8984                         if (level == 1 &&
8985                             (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
8986                                 goto skip;
8987
8988                         if (!wc->update_ref ||
8989                             generation <= root->root_key.offset)
8990                                 goto skip;
8991
8992                         btrfs_node_key_to_cpu(path->nodes[level], &key,
8993                                               path->slots[level]);
8994                         ret = btrfs_comp_cpu_keys(&key, &wc->update_progress);
8995                         if (ret < 0)
8996                                 goto skip;
8997
8998                         wc->stage = UPDATE_BACKREF;
8999                         wc->shared_level = level - 1;
9000                 }
9001         } else {
9002                 if (level == 1 &&
9003                     (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
9004                         goto skip;
9005         }
9006
9007         if (!btrfs_buffer_uptodate(next, generation, 0)) {
9008                 btrfs_tree_unlock(next);
9009                 free_extent_buffer(next);
9010                 next = NULL;
9011                 *lookup_info = 1;
9012         }
9013
9014         if (!next) {
9015                 if (reada && level == 1)
9016                         reada_walk_down(trans, root, wc, path);
9017                 next = read_tree_block(fs_info, bytenr, generation, level - 1,
9018                                        &first_key);
9019                 if (IS_ERR(next)) {
9020                         return PTR_ERR(next);
9021                 } else if (!extent_buffer_uptodate(next)) {
9022                         free_extent_buffer(next);
9023                         return -EIO;
9024                 }
9025                 btrfs_tree_lock(next);
9026                 btrfs_set_lock_blocking_write(next);
9027         }
9028
9029         level--;
9030         ASSERT(level == btrfs_header_level(next));
9031         if (level != btrfs_header_level(next)) {
9032                 btrfs_err(root->fs_info, "mismatched level");
9033                 ret = -EIO;
9034                 goto out_unlock;
9035         }
9036         path->nodes[level] = next;
9037         path->slots[level] = 0;
9038         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
9039         wc->level = level;
9040         if (wc->level == 1)
9041                 wc->reada_slot = 0;
9042         return 0;
9043 skip:
9044         wc->refs[level - 1] = 0;
9045         wc->flags[level - 1] = 0;
9046         if (wc->stage == DROP_REFERENCE) {
9047                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF) {
9048                         parent = path->nodes[level]->start;
9049                 } else {
9050                         ASSERT(root->root_key.objectid ==
9051                                btrfs_header_owner(path->nodes[level]));
9052                         if (root->root_key.objectid !=
9053                             btrfs_header_owner(path->nodes[level])) {
9054                                 btrfs_err(root->fs_info,
9055                                                 "mismatched block owner");
9056                                 ret = -EIO;
9057                                 goto out_unlock;
9058                         }
9059                         parent = 0;
9060                 }
9061
9062                 /*
9063                  * Reloc tree doesn't contribute to qgroup numbers, and we have
9064                  * already accounted them at merge time (replace_path),
9065                  * thus we could skip expensive subtree trace here.
9066                  */
9067                 if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID &&
9068                     need_account) {
9069                         ret = btrfs_qgroup_trace_subtree(trans, next,
9070                                                          generation, level - 1);
9071                         if (ret) {
9072                                 btrfs_err_rl(fs_info,
9073                                              "Error %d accounting shared subtree. Quota is out of sync, rescan required.",
9074                                              ret);
9075                         }
9076                 }
9077                 ret = btrfs_free_extent(trans, root, bytenr, fs_info->nodesize,
9078                                         parent, root->root_key.objectid,
9079                                         level - 1, 0);
9080                 if (ret)
9081                         goto out_unlock;
9082         }
9083
9084         *lookup_info = 1;
9085         ret = 1;
9086
9087 out_unlock:
9088         btrfs_tree_unlock(next);
9089         free_extent_buffer(next);
9090
9091         return ret;
9092 }
9093
9094 /*
9095  * helper to process tree block while walking up the tree.
9096  *
9097  * when wc->stage == DROP_REFERENCE, this function drops
9098  * reference count on the block.
9099  *
9100  * when wc->stage == UPDATE_BACKREF, this function changes
9101  * wc->stage back to DROP_REFERENCE if we changed wc->stage
9102  * to UPDATE_BACKREF previously while processing the block.
9103  *
9104  * NOTE: return value 1 means we should stop walking up.
9105  */
9106 static noinline int walk_up_proc(struct btrfs_trans_handle *trans,
9107                                  struct btrfs_root *root,
9108                                  struct btrfs_path *path,
9109                                  struct walk_control *wc)
9110 {
9111         struct btrfs_fs_info *fs_info = root->fs_info;
9112         int ret;
9113         int level = wc->level;
9114         struct extent_buffer *eb = path->nodes[level];
9115         u64 parent = 0;
9116
9117         if (wc->stage == UPDATE_BACKREF) {
9118                 BUG_ON(wc->shared_level < level);
9119                 if (level < wc->shared_level)
9120                         goto out;
9121
9122                 ret = find_next_key(path, level + 1, &wc->update_progress);
9123                 if (ret > 0)
9124                         wc->update_ref = 0;
9125
9126                 wc->stage = DROP_REFERENCE;
9127                 wc->shared_level = -1;
9128                 path->slots[level] = 0;
9129
9130                 /*
9131                  * check reference count again if the block isn't locked.
9132                  * we should start walking down the tree again if reference
9133                  * count is one.
9134                  */
9135                 if (!path->locks[level]) {
9136                         BUG_ON(level == 0);
9137                         btrfs_tree_lock(eb);
9138                         btrfs_set_lock_blocking_write(eb);
9139                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
9140
9141                         ret = btrfs_lookup_extent_info(trans, fs_info,
9142                                                        eb->start, level, 1,
9143                                                        &wc->refs[level],
9144                                                        &wc->flags[level]);
9145                         if (ret < 0) {
9146                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
9147                                 path->locks[level] = 0;
9148                                 return ret;
9149                         }
9150                         BUG_ON(wc->refs[level] == 0);
9151                         if (wc->refs[level] == 1) {
9152                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
9153                                 path->locks[level] = 0;
9154                                 return 1;
9155                         }
9156                 }
9157         }
9158
9159         /* wc->stage == DROP_REFERENCE */
9160         BUG_ON(wc->refs[level] > 1 && !path->locks[level]);
9161
9162         if (wc->refs[level] == 1) {
9163                 if (level == 0) {
9164                         if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
9165                                 ret = btrfs_dec_ref(trans, root, eb, 1);
9166                         else
9167                                 ret = btrfs_dec_ref(trans, root, eb, 0);
9168                         BUG_ON(ret); /* -ENOMEM */
9169                         ret = btrfs_qgroup_trace_leaf_items(trans, eb);
9170                         if (ret) {
9171                                 btrfs_err_rl(fs_info,
9172                                              "error %d accounting leaf items. Quota is out of sync, rescan required.",
9173                                              ret);
9174                         }
9175                 }
9176                 /* make block locked assertion in clean_tree_block happy */
9177                 if (!path->locks[level] &&
9178                     btrfs_header_generation(eb) == trans->transid) {
9179                         btrfs_tree_lock(eb);
9180                         btrfs_set_lock_blocking_write(eb);
9181                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
9182                 }
9183                 clean_tree_block(fs_info, eb);
9184         }
9185
9186         if (eb == root->node) {
9187                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
9188                         parent = eb->start;
9189                 else if (root->root_key.objectid != btrfs_header_owner(eb))
9190                         goto owner_mismatch;
9191         } else {
9192                 if (wc->flags[level + 1] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
9193                         parent = path->nodes[level + 1]->start;
9194                 else if (root->root_key.objectid !=
9195                          btrfs_header_owner(path->nodes[level + 1]))
9196                         goto owner_mismatch;
9197         }
9198
9199         btrfs_free_tree_block(trans, root, eb, parent, wc->refs[level] == 1);
9200 out:
9201         wc->refs[level] = 0;
9202         wc->flags[level] = 0;
9203         return 0;
9204
9205 owner_mismatch:
9206         btrfs_err_rl(fs_info, "unexpected tree owner, have %llu expect %llu",
9207                      btrfs_header_owner(eb), root->root_key.objectid);
9208         return -EUCLEAN;
9209 }
9210
9211 static noinline int walk_down_tree(struct btrfs_trans_handle *trans,
9212                                    struct btrfs_root *root,
9213                                    struct btrfs_path *path,
9214                                    struct walk_control *wc)
9215 {
9216         int level = wc->level;
9217         int lookup_info = 1;
9218         int ret;
9219
9220         while (level >= 0) {
9221                 ret = walk_down_proc(trans, root, path, wc, lookup_info);
9222                 if (ret > 0)
9223                         break;
9224
9225                 if (level == 0)
9226                         break;
9227
9228                 if (path->slots[level] >=
9229                     btrfs_header_nritems(path->nodes[level]))
9230                         break;
9231
9232                 ret = do_walk_down(trans, root, path, wc, &lookup_info);
9233                 if (ret > 0) {
9234                         path->slots[level]++;
9235                         continue;
9236                 } else if (ret < 0)
9237                         return ret;
9238                 level = wc->level;
9239         }
9240         return 0;
9241 }
9242
9243 static noinline int walk_up_tree(struct btrfs_trans_handle *trans,
9244                                  struct btrfs_root *root,
9245                                  struct btrfs_path *path,
9246                                  struct walk_control *wc, int max_level)
9247 {
9248         int level = wc->level;
9249         int ret;
9250
9251         path->slots[level] = btrfs_header_nritems(path->nodes[level]);
9252         while (level < max_level && path->nodes[level]) {
9253                 wc->level = level;
9254                 if (path->slots[level] + 1 <
9255                     btrfs_header_nritems(path->nodes[level])) {
9256                         path->slots[level]++;
9257                         return 0;
9258                 } else {
9259                         ret = walk_up_proc(trans, root, path, wc);
9260                         if (ret > 0)
9261                                 return 0;
9262                         if (ret < 0)
9263                                 return ret;
9264
9265                         if (path->locks[level]) {
9266                                 btrfs_tree_unlock_rw(path->nodes[level],
9267                                                      path->locks[level]);
9268                                 path->locks[level] = 0;
9269                         }
9270                         free_extent_buffer(path->nodes[level]);
9271                         path->nodes[level] = NULL;
9272                         level++;
9273                 }
9274         }
9275         return 1;
9276 }
9277
9278 /*
9279  * drop a subvolume tree.
9280  *
9281  * this function traverses the tree freeing any blocks that only
9282  * referenced by the tree.
9283  *
9284  * when a shared tree block is found. this function decreases its
9285  * reference count by one. if update_ref is true, this function
9286  * also make sure backrefs for the shared block and all lower level
9287  * blocks are properly updated.
9288  *
9289  * If called with for_reloc == 0, may exit early with -EAGAIN
9290  */
9291 int btrfs_drop_snapshot(struct btrfs_root *root,
9292                          struct btrfs_block_rsv *block_rsv, int update_ref,
9293                          int for_reloc)
9294 {
9295         struct btrfs_fs_info *fs_info = root->fs_info;
9296         struct btrfs_path *path;
9297         struct btrfs_trans_handle *trans;
9298         struct btrfs_root *tree_root = fs_info->tree_root;
9299         struct btrfs_root_item *root_item = &root->root_item;
9300         struct walk_control *wc;
9301         struct btrfs_key key;
9302         int err = 0;
9303         int ret;
9304         int level;
9305         bool root_dropped = false;
9306
9307         btrfs_debug(fs_info, "Drop subvolume %llu", root->root_key.objectid);
9308
9309         path = btrfs_alloc_path();
9310         if (!path) {
9311                 err = -ENOMEM;
9312                 goto out;
9313         }
9314
9315         wc = kzalloc(sizeof(*wc), GFP_NOFS);
9316         if (!wc) {
9317                 btrfs_free_path(path);
9318                 err = -ENOMEM;
9319                 goto out;
9320         }
9321
9322         trans = btrfs_start_transaction(tree_root, 0);
9323         if (IS_ERR(trans)) {
9324                 err = PTR_ERR(trans);
9325                 goto out_free;
9326         }
9327
9328         err = btrfs_run_delayed_items(trans);
9329         if (err)
9330                 goto out_end_trans;
9331
9332         if (block_rsv)
9333                 trans->block_rsv = block_rsv;
9334
9335         /*
9336          * This will help us catch people modifying the fs tree while we're
9337          * dropping it.  It is unsafe to mess with the fs tree while it's being
9338          * dropped as we unlock the root node and parent nodes as we walk down
9339          * the tree, assuming nothing will change.  If something does change
9340          * then we'll have stale information and drop references to blocks we've
9341          * already dropped.
9342          */
9343         set_bit(BTRFS_ROOT_DELETING, &root->state);
9344         if (btrfs_disk_key_objectid(&root_item->drop_progress) == 0) {
9345                 level = btrfs_header_level(root->node);
9346                 path->nodes[level] = btrfs_lock_root_node(root);
9347                 btrfs_set_lock_blocking_write(path->nodes[level]);
9348                 path->slots[level] = 0;
9349                 path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
9350                 memset(&wc->update_progress, 0,
9351                        sizeof(wc->update_progress));
9352         } else {
9353                 btrfs_disk_key_to_cpu(&key, &root_item->drop_progress);
9354                 memcpy(&wc->update_progress, &key,
9355                        sizeof(wc->update_progress));
9356
9357                 level = root_item->drop_level;
9358                 BUG_ON(level == 0);
9359                 path->lowest_level = level;
9360                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
9361                 path->lowest_level = 0;
9362                 if (ret < 0) {
9363                         err = ret;
9364                         goto out_end_trans;
9365                 }
9366                 WARN_ON(ret > 0);
9367
9368                 /*
9369                  * unlock our path, this is safe because only this
9370                  * function is allowed to delete this snapshot
9371                  */
9372                 btrfs_unlock_up_safe(path, 0);
9373
9374                 level = btrfs_header_level(root->node);
9375                 while (1) {
9376                         btrfs_tree_lock(path->nodes[level]);
9377                         btrfs_set_lock_blocking_write(path->nodes[level]);
9378                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
9379
9380                         ret = btrfs_lookup_extent_info(trans, fs_info,
9381                                                 path->nodes[level]->start,
9382                                                 level, 1, &wc->refs[level],
9383                                                 &wc->flags[level]);
9384                         if (ret < 0) {
9385                                 err = ret;
9386                                 goto out_end_trans;
9387                         }
9388                         BUG_ON(wc->refs[level] == 0);
9389
9390                         if (level == root_item->drop_level)
9391                                 break;
9392
9393                         btrfs_tree_unlock(path->nodes[level]);
9394                         path->locks[level] = 0;
9395                         WARN_ON(wc->refs[level] != 1);
9396                         level--;
9397                 }
9398         }
9399
9400         wc->level = level;
9401         wc->shared_level = -1;
9402         wc->stage = DROP_REFERENCE;
9403         wc->update_ref = update_ref;
9404         wc->keep_locks = 0;
9405         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(fs_info);
9406
9407         while (1) {
9408
9409                 ret = walk_down_tree(trans, root, path, wc);
9410                 if (ret < 0) {
9411                         err = ret;
9412                         break;
9413                 }
9414
9415                 ret = walk_up_tree(trans, root, path, wc, BTRFS_MAX_LEVEL);
9416                 if (ret < 0) {
9417                         err = ret;
9418                         break;
9419                 }
9420
9421                 if (ret > 0) {
9422                         BUG_ON(wc->stage != DROP_REFERENCE);
9423                         break;
9424                 }
9425
9426                 if (wc->stage == DROP_REFERENCE) {
9427                         level = wc->level;
9428                         btrfs_node_key(path->nodes[level],
9429                                        &root_item->drop_progress,
9430                                        path->slots[level]);
9431                         root_item->drop_level = level;
9432                 }
9433
9434                 BUG_ON(wc->level == 0);
9435                 if (btrfs_should_end_transaction(trans) ||
9436                     (!for_reloc && btrfs_need_cleaner_sleep(fs_info))) {
9437                         ret = btrfs_update_root(trans, tree_root,
9438                                                 &root->root_key,
9439                                                 root_item);
9440                         if (ret) {
9441                                 btrfs_abort_transaction(trans, ret);
9442                                 err = ret;
9443                                 goto out_end_trans;
9444                         }
9445
9446                         btrfs_end_transaction_throttle(trans);
9447                         if (!for_reloc && btrfs_need_cleaner_sleep(fs_info)) {
9448                                 btrfs_debug(fs_info,
9449                                             "drop snapshot early exit");
9450                                 err = -EAGAIN;
9451                                 goto out_free;
9452                         }
9453
9454                         trans = btrfs_start_transaction(tree_root, 0);
9455                         if (IS_ERR(trans)) {
9456                                 err = PTR_ERR(trans);
9457                                 goto out_free;
9458                         }
9459                         if (block_rsv)
9460                                 trans->block_rsv = block_rsv;
9461                 }
9462         }
9463         btrfs_release_path(path);
9464         if (err)
9465                 goto out_end_trans;
9466
9467         ret = btrfs_del_root(trans, &root->root_key);
9468         if (ret) {
9469                 btrfs_abort_transaction(trans, ret);
9470                 err = ret;
9471                 goto out_end_trans;
9472         }
9473
9474         if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) {
9475                 ret = btrfs_find_root(tree_root, &root->root_key, path,
9476                                       NULL, NULL);
9477                 if (ret < 0) {
9478                         btrfs_abort_transaction(trans, ret);
9479                         err = ret;
9480                         goto out_end_trans;
9481                 } else if (ret > 0) {
9482                         /* if we fail to delete the orphan item this time
9483                          * around, it'll get picked up the next time.
9484                          *
9485                          * The most common failure here is just -ENOENT.
9486                          */
9487                         btrfs_del_orphan_item(trans, tree_root,
9488                                               root->root_key.objectid);
9489                 }
9490         }
9491
9492         if (test_bit(BTRFS_ROOT_IN_RADIX, &root->state)) {
9493                 btrfs_add_dropped_root(trans, root);
9494         } else {
9495                 free_extent_buffer(root->node);
9496                 free_extent_buffer(root->commit_root);
9497                 btrfs_put_fs_root(root);
9498         }
9499         root_dropped = true;
9500 out_end_trans:
9501         btrfs_end_transaction_throttle(trans);
9502 out_free:
9503         kfree(wc);
9504         btrfs_free_path(path);
9505 out:
9506         /*
9507          * So if we need to stop dropping the snapshot for whatever reason we
9508          * need to make sure to add it back to the dead root list so that we
9509          * keep trying to do the work later.  This also cleans up roots if we
9510          * don't have it in the radix (like when we recover after a power fail
9511          * or unmount) so we don't leak memory.
9512          */
9513         if (!for_reloc && !root_dropped)
9514                 btrfs_add_dead_root(root);
9515         if (err && err != -EAGAIN)
9516                 btrfs_handle_fs_error(fs_info, err, NULL);
9517         return err;
9518 }
9519
9520 /*
9521  * drop subtree rooted at tree block 'node'.
9522  *
9523  * NOTE: this function will unlock and release tree block 'node'
9524  * only used by relocation code
9525  */
9526 int btrfs_drop_subtree(struct btrfs_trans_handle *trans,
9527                         struct btrfs_root *root,
9528                         struct extent_buffer *node,
9529                         struct extent_buffer *parent)
9530 {
9531         struct btrfs_fs_info *fs_info = root->fs_info;
9532         struct btrfs_path *path;
9533         struct walk_control *wc;
9534         int level;
9535         int parent_level;
9536         int ret = 0;
9537         int wret;
9538
9539         BUG_ON(root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID);
9540
9541         path = btrfs_alloc_path();
9542         if (!path)
9543                 return -ENOMEM;
9544
9545         wc = kzalloc(sizeof(*wc), GFP_NOFS);
9546         if (!wc) {
9547                 btrfs_free_path(path);
9548                 return -ENOMEM;
9549         }
9550
9551         btrfs_assert_tree_locked(parent);
9552         parent_level = btrfs_header_level(parent);
9553         extent_buffer_get(parent);
9554         path->nodes[parent_level] = parent;
9555         path->slots[parent_level] = btrfs_header_nritems(parent);
9556
9557         btrfs_assert_tree_locked(node);
9558         level = btrfs_header_level(node);
9559         path->nodes[level] = node;
9560         path->slots[level] = 0;
9561         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
9562
9563         wc->refs[parent_level] = 1;
9564         wc->flags[parent_level] = BTRFS_BLOCK_FLAG_FULL_BACKREF;
9565         wc->level = level;
9566         wc->shared_level = -1;
9567         wc->stage = DROP_REFERENCE;
9568         wc->update_ref = 0;
9569         wc->keep_locks = 1;
9570         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(fs_info);
9571
9572         while (1) {
9573                 wret = walk_down_tree(trans, root, path, wc);
9574                 if (wret < 0) {
9575                         ret = wret;
9576                         break;
9577                 }
9578
9579                 wret = walk_up_tree(trans, root, path, wc, parent_level);
9580                 if (wret < 0)
9581                         ret = wret;
9582                 if (wret != 0)
9583                         break;
9584         }
9585
9586         kfree(wc);
9587         btrfs_free_path(path);
9588         return ret;
9589 }
9590
9591 static u64 update_block_group_flags(struct btrfs_fs_info *fs_info, u64 flags)
9592 {
9593         u64 num_devices;
9594         u64 stripped;
9595
9596         /*
9597          * if restripe for this chunk_type is on pick target profile and
9598          * return, otherwise do the usual balance
9599          */
9600         stripped = get_restripe_target(fs_info, flags);
9601         if (stripped)
9602                 return extended_to_chunk(stripped);
9603
9604         num_devices = fs_info->fs_devices->rw_devices;
9605
9606         stripped = BTRFS_BLOCK_GROUP_RAID0 |
9607                 BTRFS_BLOCK_GROUP_RAID5 | BTRFS_BLOCK_GROUP_RAID6 |
9608                 BTRFS_BLOCK_GROUP_RAID1 | BTRFS_BLOCK_GROUP_RAID10;
9609
9610         if (num_devices == 1) {
9611                 stripped |= BTRFS_BLOCK_GROUP_DUP;
9612                 stripped = flags & ~stripped;
9613
9614                 /* turn raid0 into single device chunks */
9615                 if (flags & BTRFS_BLOCK_GROUP_RAID0)
9616                         return stripped;
9617
9618                 /* turn mirroring into duplication */
9619                 if (flags & (BTRFS_BLOCK_GROUP_RAID1 |
9620                              BTRFS_BLOCK_GROUP_RAID10))
9621                         return stripped | BTRFS_BLOCK_GROUP_DUP;
9622         } else {
9623                 /* they already had raid on here, just return */
9624                 if (flags & stripped)
9625                         return flags;
9626
9627                 stripped |= BTRFS_BLOCK_GROUP_DUP;
9628                 stripped = flags & ~stripped;
9629
9630                 /* switch duplicated blocks with raid1 */
9631                 if (flags & BTRFS_BLOCK_GROUP_DUP)
9632                         return stripped | BTRFS_BLOCK_GROUP_RAID1;
9633
9634                 /* this is drive concat, leave it alone */
9635         }
9636
9637         return flags;
9638 }
9639
9640 static int inc_block_group_ro(struct btrfs_block_group_cache *cache, int force)
9641 {
9642         struct btrfs_space_info *sinfo = cache->space_info;
9643         u64 num_bytes;
9644         u64 sinfo_used;
9645         u64 min_allocable_bytes;
9646         int ret = -ENOSPC;
9647
9648         /*
9649          * We need some metadata space and system metadata space for
9650          * allocating chunks in some corner cases until we force to set
9651          * it to be readonly.
9652          */
9653         if ((sinfo->flags &
9654              (BTRFS_BLOCK_GROUP_SYSTEM | BTRFS_BLOCK_GROUP_METADATA)) &&
9655             !force)
9656                 min_allocable_bytes = SZ_1M;
9657         else
9658                 min_allocable_bytes = 0;
9659
9660         spin_lock(&sinfo->lock);
9661         spin_lock(&cache->lock);
9662
9663         if (cache->ro) {
9664                 cache->ro++;
9665                 ret = 0;
9666                 goto out;
9667         }
9668
9669         num_bytes = cache->key.offset - cache->reserved - cache->pinned -
9670                     cache->bytes_super - btrfs_block_group_used(&cache->item);
9671         sinfo_used = btrfs_space_info_used(sinfo, true);
9672
9673         if (sinfo_used + num_bytes + min_allocable_bytes <=
9674             sinfo->total_bytes) {
9675                 sinfo->bytes_readonly += num_bytes;
9676                 cache->ro++;
9677                 list_add_tail(&cache->ro_list, &sinfo->ro_bgs);
9678                 ret = 0;
9679         }
9680 out:
9681         spin_unlock(&cache->lock);
9682         spin_unlock(&sinfo->lock);
9683         if (ret == -ENOSPC && btrfs_test_opt(cache->fs_info, ENOSPC_DEBUG)) {
9684                 btrfs_info(cache->fs_info,
9685                         "unable to make block group %llu ro",
9686                         cache->key.objectid);
9687                 btrfs_info(cache->fs_info,
9688                         "sinfo_used=%llu bg_num_bytes=%llu min_allocable=%llu",
9689                         sinfo_used, num_bytes, min_allocable_bytes);
9690                 dump_space_info(cache->fs_info, cache->space_info, 0, 0);
9691         }
9692         return ret;
9693 }
9694
9695 int btrfs_inc_block_group_ro(struct btrfs_block_group_cache *cache)
9696
9697 {
9698         struct btrfs_fs_info *fs_info = cache->fs_info;
9699         struct btrfs_trans_handle *trans;
9700         u64 alloc_flags;
9701         int ret;
9702
9703 again:
9704         trans = btrfs_join_transaction(fs_info->extent_root);
9705         if (IS_ERR(trans))
9706                 return PTR_ERR(trans);
9707
9708         /*
9709          * we're not allowed to set block groups readonly after the dirty
9710          * block groups cache has started writing.  If it already started,
9711          * back off and let this transaction commit
9712          */
9713         mutex_lock(&fs_info->ro_block_group_mutex);
9714         if (test_bit(BTRFS_TRANS_DIRTY_BG_RUN, &trans->transaction->flags)) {
9715                 u64 transid = trans->transid;
9716
9717                 mutex_unlock(&fs_info->ro_block_group_mutex);
9718                 btrfs_end_transaction(trans);
9719
9720                 ret = btrfs_wait_for_commit(fs_info, transid);
9721                 if (ret)
9722                         return ret;
9723                 goto again;
9724         }
9725
9726         /*
9727          * if we are changing raid levels, try to allocate a corresponding
9728          * block group with the new raid level.
9729          */
9730         alloc_flags = update_block_group_flags(fs_info, cache->flags);
9731         if (alloc_flags != cache->flags) {
9732                 ret = do_chunk_alloc(trans, alloc_flags,
9733                                      CHUNK_ALLOC_FORCE);
9734                 /*
9735                  * ENOSPC is allowed here, we may have enough space
9736                  * already allocated at the new raid level to
9737                  * carry on
9738                  */
9739                 if (ret == -ENOSPC)
9740                         ret = 0;
9741                 if (ret < 0)
9742                         goto out;
9743         }
9744
9745         ret = inc_block_group_ro(cache, 0);
9746         if (!ret)
9747                 goto out;
9748         alloc_flags = get_alloc_profile(fs_info, cache->space_info->flags);
9749         ret = do_chunk_alloc(trans, alloc_flags, CHUNK_ALLOC_FORCE);
9750         if (ret < 0)
9751                 goto out;
9752         ret = inc_block_group_ro(cache, 0);
9753 out:
9754         if (cache->flags & BTRFS_BLOCK_GROUP_SYSTEM) {
9755                 alloc_flags = update_block_group_flags(fs_info, cache->flags);
9756                 mutex_lock(&fs_info->chunk_mutex);
9757                 check_system_chunk(trans, alloc_flags);
9758                 mutex_unlock(&fs_info->chunk_mutex);
9759         }
9760         mutex_unlock(&fs_info->ro_block_group_mutex);
9761
9762         btrfs_end_transaction(trans);
9763         return ret;
9764 }
9765
9766 int btrfs_force_chunk_alloc(struct btrfs_trans_handle *trans, u64 type)
9767 {
9768         u64 alloc_flags = get_alloc_profile(trans->fs_info, type);
9769
9770         return do_chunk_alloc(trans, alloc_flags, CHUNK_ALLOC_FORCE);
9771 }
9772
9773 /*
9774  * helper to account the unused space of all the readonly block group in the
9775  * space_info. takes mirrors into account.
9776  */
9777 u64 btrfs_account_ro_block_groups_free_space(struct btrfs_space_info *sinfo)
9778 {
9779         struct btrfs_block_group_cache *block_group;
9780         u64 free_bytes = 0;
9781         int factor;
9782
9783         /* It's df, we don't care if it's racy */
9784         if (list_empty(&sinfo->ro_bgs))
9785                 return 0;
9786
9787         spin_lock(&sinfo->lock);
9788         list_for_each_entry(block_group, &sinfo->ro_bgs, ro_list) {
9789                 spin_lock(&block_group->lock);
9790
9791                 if (!block_group->ro) {
9792                         spin_unlock(&block_group->lock);
9793                         continue;
9794                 }
9795
9796                 factor = btrfs_bg_type_to_factor(block_group->flags);
9797                 free_bytes += (block_group->key.offset -
9798                                btrfs_block_group_used(&block_group->item)) *
9799                                factor;
9800
9801                 spin_unlock(&block_group->lock);
9802         }
9803         spin_unlock(&sinfo->lock);
9804
9805         return free_bytes;
9806 }
9807
9808 void btrfs_dec_block_group_ro(struct btrfs_block_group_cache *cache)
9809 {
9810         struct btrfs_space_info *sinfo = cache->space_info;
9811         u64 num_bytes;
9812
9813         BUG_ON(!cache->ro);
9814
9815         spin_lock(&sinfo->lock);
9816         spin_lock(&cache->lock);
9817         if (!--cache->ro) {
9818                 num_bytes = cache->key.offset - cache->reserved -
9819                             cache->pinned - cache->bytes_super -
9820                             btrfs_block_group_used(&cache->item);
9821                 sinfo->bytes_readonly -= num_bytes;
9822                 list_del_init(&cache->ro_list);
9823         }
9824         spin_unlock(&cache->lock);
9825         spin_unlock(&sinfo->lock);
9826 }
9827
9828 /*
9829  * Checks to see if it's even possible to relocate this block group.
9830  *
9831  * @return - -1 if it's not a good idea to relocate this block group, 0 if its
9832  * ok to go ahead and try.
9833  */
9834 int btrfs_can_relocate(struct btrfs_fs_info *fs_info, u64 bytenr)
9835 {
9836         struct btrfs_root *root = fs_info->extent_root;
9837         struct btrfs_block_group_cache *block_group;
9838         struct btrfs_space_info *space_info;
9839         struct btrfs_fs_devices *fs_devices = fs_info->fs_devices;
9840         struct btrfs_device *device;
9841         struct btrfs_trans_handle *trans;
9842         u64 min_free;
9843         u64 dev_min = 1;
9844         u64 dev_nr = 0;
9845         u64 target;
9846         int debug;
9847         int index;
9848         int full = 0;
9849         int ret = 0;
9850
9851         debug = btrfs_test_opt(fs_info, ENOSPC_DEBUG);
9852
9853         block_group = btrfs_lookup_block_group(fs_info, bytenr);
9854
9855         /* odd, couldn't find the block group, leave it alone */
9856         if (!block_group) {
9857                 if (debug)
9858                         btrfs_warn(fs_info,
9859                                    "can't find block group for bytenr %llu",
9860                                    bytenr);
9861                 return -1;
9862         }
9863
9864         min_free = btrfs_block_group_used(&block_group->item);
9865
9866         /* no bytes used, we're good */
9867         if (!min_free)
9868                 goto out;
9869
9870         space_info = block_group->space_info;
9871         spin_lock(&space_info->lock);
9872
9873         full = space_info->full;
9874
9875         /*
9876          * if this is the last block group we have in this space, we can't
9877          * relocate it unless we're able to allocate a new chunk below.
9878          *
9879          * Otherwise, we need to make sure we have room in the space to handle
9880          * all of the extents from this block group.  If we can, we're good
9881          */
9882         if ((space_info->total_bytes != block_group->key.offset) &&
9883             (btrfs_space_info_used(space_info, false) + min_free <
9884              space_info->total_bytes)) {
9885                 spin_unlock(&space_info->lock);
9886                 goto out;
9887         }
9888         spin_unlock(&space_info->lock);
9889
9890         /*
9891          * ok we don't have enough space, but maybe we have free space on our
9892          * devices to allocate new chunks for relocation, so loop through our
9893          * alloc devices and guess if we have enough space.  if this block
9894          * group is going to be restriped, run checks against the target
9895          * profile instead of the current one.
9896          */
9897         ret = -1;
9898
9899         /*
9900          * index:
9901          *      0: raid10
9902          *      1: raid1
9903          *      2: dup
9904          *      3: raid0
9905          *      4: single
9906          */
9907         target = get_restripe_target(fs_info, block_group->flags);
9908         if (target) {
9909                 index = btrfs_bg_flags_to_raid_index(extended_to_chunk(target));
9910         } else {
9911                 /*
9912                  * this is just a balance, so if we were marked as full
9913                  * we know there is no space for a new chunk
9914                  */
9915                 if (full) {
9916                         if (debug)
9917                                 btrfs_warn(fs_info,
9918                                            "no space to alloc new chunk for block group %llu",
9919                                            block_group->key.objectid);
9920                         goto out;
9921                 }
9922
9923                 index = btrfs_bg_flags_to_raid_index(block_group->flags);
9924         }
9925
9926         if (index == BTRFS_RAID_RAID10) {
9927                 dev_min = 4;
9928                 /* Divide by 2 */
9929                 min_free >>= 1;
9930         } else if (index == BTRFS_RAID_RAID1) {
9931                 dev_min = 2;
9932         } else if (index == BTRFS_RAID_DUP) {
9933                 /* Multiply by 2 */
9934                 min_free <<= 1;
9935         } else if (index == BTRFS_RAID_RAID0) {
9936                 dev_min = fs_devices->rw_devices;
9937                 min_free = div64_u64(min_free, dev_min);
9938         }
9939
9940         /* We need to do this so that we can look at pending chunks */
9941         trans = btrfs_join_transaction(root);
9942         if (IS_ERR(trans)) {
9943                 ret = PTR_ERR(trans);
9944                 goto out;
9945         }
9946
9947         mutex_lock(&fs_info->chunk_mutex);
9948         list_for_each_entry(device, &fs_devices->alloc_list, dev_alloc_list) {
9949                 u64 dev_offset;
9950
9951                 /*
9952                  * check to make sure we can actually find a chunk with enough
9953                  * space to fit our block group in.
9954                  */
9955                 if (device->total_bytes > device->bytes_used + min_free &&
9956                     !test_bit(BTRFS_DEV_STATE_REPLACE_TGT, &device->dev_state)) {
9957                         ret = find_free_dev_extent(trans, device, min_free,
9958                                                    &dev_offset, NULL);
9959                         if (!ret)
9960                                 dev_nr++;
9961
9962                         if (dev_nr >= dev_min)
9963                                 break;
9964
9965                         ret = -1;
9966                 }
9967         }
9968         if (debug && ret == -1)
9969                 btrfs_warn(fs_info,
9970                            "no space to allocate a new chunk for block group %llu",
9971                            block_group->key.objectid);
9972         mutex_unlock(&fs_info->chunk_mutex);
9973         btrfs_end_transaction(trans);
9974 out:
9975         btrfs_put_block_group(block_group);
9976         return ret;
9977 }
9978
9979 static int find_first_block_group(struct btrfs_fs_info *fs_info,
9980                                   struct btrfs_path *path,
9981                                   struct btrfs_key *key)
9982 {
9983         struct btrfs_root *root = fs_info->extent_root;
9984         int ret = 0;
9985         struct btrfs_key found_key;
9986         struct extent_buffer *leaf;
9987         struct btrfs_block_group_item bg;
9988         u64 flags;
9989         int slot;
9990
9991         ret = btrfs_search_slot(NULL, root, key, path, 0, 0);
9992         if (ret < 0)
9993                 goto out;
9994
9995         while (1) {
9996                 slot = path->slots[0];
9997                 leaf = path->nodes[0];
9998                 if (slot >= btrfs_header_nritems(leaf)) {
9999                         ret = btrfs_next_leaf(root, path);
10000                         if (ret == 0)
10001                                 continue;
10002                         if (ret < 0)
10003                                 goto out;
10004                         break;
10005                 }
10006                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
10007
10008                 if (found_key.objectid >= key->objectid &&
10009                     found_key.type == BTRFS_BLOCK_GROUP_ITEM_KEY) {
10010                         struct extent_map_tree *em_tree;
10011                         struct extent_map *em;
10012
10013                         em_tree = &root->fs_info->mapping_tree.map_tree;
10014                         read_lock(&em_tree->lock);
10015                         em = lookup_extent_mapping(em_tree, found_key.objectid,
10016                                                    found_key.offset);
10017                         read_unlock(&em_tree->lock);
10018                         if (!em) {
10019                                 btrfs_err(fs_info,
10020                         "logical %llu len %llu found bg but no related chunk",
10021                                           found_key.objectid, found_key.offset);
10022                                 ret = -ENOENT;
10023                         } else if (em->start != found_key.objectid ||
10024                                    em->len != found_key.offset) {
10025                                 btrfs_err(fs_info,
10026                 "block group %llu len %llu mismatch with chunk %llu len %llu",
10027                                           found_key.objectid, found_key.offset,
10028                                           em->start, em->len);
10029                                 ret = -EUCLEAN;
10030                         } else {
10031                                 read_extent_buffer(leaf, &bg,
10032                                         btrfs_item_ptr_offset(leaf, slot),
10033                                         sizeof(bg));
10034                                 flags = btrfs_block_group_flags(&bg) &
10035                                         BTRFS_BLOCK_GROUP_TYPE_MASK;
10036
10037                                 if (flags != (em->map_lookup->type &
10038                                               BTRFS_BLOCK_GROUP_TYPE_MASK)) {
10039                                         btrfs_err(fs_info,
10040 "block group %llu len %llu type flags 0x%llx mismatch with chunk type flags 0x%llx",
10041                                                 found_key.objectid,
10042                                                 found_key.offset, flags,
10043                                                 (BTRFS_BLOCK_GROUP_TYPE_MASK &
10044                                                  em->map_lookup->type));
10045                                         ret = -EUCLEAN;
10046                                 } else {
10047                                         ret = 0;
10048                                 }
10049                         }
10050                         free_extent_map(em);
10051                         goto out;
10052                 }
10053                 path->slots[0]++;
10054         }
10055 out:
10056         return ret;
10057 }
10058
10059 void btrfs_put_block_group_cache(struct btrfs_fs_info *info)
10060 {
10061         struct btrfs_block_group_cache *block_group;
10062         u64 last = 0;
10063
10064         while (1) {
10065                 struct inode *inode;
10066
10067                 block_group = btrfs_lookup_first_block_group(info, last);
10068                 while (block_group) {
10069                         wait_block_group_cache_done(block_group);
10070                         spin_lock(&block_group->lock);
10071                         if (block_group->iref)
10072                                 break;
10073                         spin_unlock(&block_group->lock);
10074                         block_group = next_block_group(info, block_group);
10075                 }
10076                 if (!block_group) {
10077                         if (last == 0)
10078                                 break;
10079                         last = 0;
10080                         continue;
10081                 }
10082
10083                 inode = block_group->inode;
10084                 block_group->iref = 0;
10085                 block_group->inode = NULL;
10086                 spin_unlock(&block_group->lock);
10087                 ASSERT(block_group->io_ctl.inode == NULL);
10088                 iput(inode);
10089                 last = block_group->key.objectid + block_group->key.offset;
10090                 btrfs_put_block_group(block_group);
10091         }
10092 }
10093
10094 /*
10095  * Must be called only after stopping all workers, since we could have block
10096  * group caching kthreads running, and therefore they could race with us if we
10097  * freed the block groups before stopping them.
10098  */
10099 int btrfs_free_block_groups(struct btrfs_fs_info *info)
10100 {
10101         struct btrfs_block_group_cache *block_group;
10102         struct btrfs_space_info *space_info;
10103         struct btrfs_caching_control *caching_ctl;
10104         struct rb_node *n;
10105
10106         down_write(&info->commit_root_sem);
10107         while (!list_empty(&info->caching_block_groups)) {
10108                 caching_ctl = list_entry(info->caching_block_groups.next,
10109                                          struct btrfs_caching_control, list);
10110                 list_del(&caching_ctl->list);
10111                 put_caching_control(caching_ctl);
10112         }
10113         up_write(&info->commit_root_sem);
10114
10115         spin_lock(&info->unused_bgs_lock);
10116         while (!list_empty(&info->unused_bgs)) {
10117                 block_group = list_first_entry(&info->unused_bgs,
10118                                                struct btrfs_block_group_cache,
10119                                                bg_list);
10120                 list_del_init(&block_group->bg_list);
10121                 btrfs_put_block_group(block_group);
10122         }
10123         spin_unlock(&info->unused_bgs_lock);
10124
10125         spin_lock(&info->block_group_cache_lock);
10126         while ((n = rb_last(&info->block_group_cache_tree)) != NULL) {
10127                 block_group = rb_entry(n, struct btrfs_block_group_cache,
10128                                        cache_node);
10129                 rb_erase(&block_group->cache_node,
10130                          &info->block_group_cache_tree);
10131                 RB_CLEAR_NODE(&block_group->cache_node);
10132                 spin_unlock(&info->block_group_cache_lock);
10133
10134                 down_write(&block_group->space_info->groups_sem);
10135                 list_del(&block_group->list);
10136                 up_write(&block_group->space_info->groups_sem);
10137
10138                 /*
10139                  * We haven't cached this block group, which means we could
10140                  * possibly have excluded extents on this block group.
10141                  */
10142                 if (block_group->cached == BTRFS_CACHE_NO ||
10143                     block_group->cached == BTRFS_CACHE_ERROR)
10144                         free_excluded_extents(block_group);
10145
10146                 btrfs_remove_free_space_cache(block_group);
10147                 ASSERT(block_group->cached != BTRFS_CACHE_STARTED);
10148                 ASSERT(list_empty(&block_group->dirty_list));
10149                 ASSERT(list_empty(&block_group->io_list));
10150                 ASSERT(list_empty(&block_group->bg_list));
10151                 ASSERT(atomic_read(&block_group->count) == 1);
10152                 btrfs_put_block_group(block_group);
10153
10154                 spin_lock(&info->block_group_cache_lock);
10155         }
10156         spin_unlock(&info->block_group_cache_lock);
10157
10158         /* now that all the block groups are freed, go through and
10159          * free all the space_info structs.  This is only called during
10160          * the final stages of unmount, and so we know nobody is
10161          * using them.  We call synchronize_rcu() once before we start,
10162          * just to be on the safe side.
10163          */
10164         synchronize_rcu();
10165
10166         release_global_block_rsv(info);
10167
10168         while (!list_empty(&info->space_info)) {
10169                 int i;
10170
10171                 space_info = list_entry(info->space_info.next,
10172                                         struct btrfs_space_info,
10173                                         list);
10174
10175                 /*
10176                  * Do not hide this behind enospc_debug, this is actually
10177                  * important and indicates a real bug if this happens.
10178                  */
10179                 if (WARN_ON(space_info->bytes_pinned > 0 ||
10180                             space_info->bytes_reserved > 0 ||
10181                             space_info->bytes_may_use > 0))
10182                         dump_space_info(info, space_info, 0, 0);
10183                 list_del(&space_info->list);
10184                 for (i = 0; i < BTRFS_NR_RAID_TYPES; i++) {
10185                         struct kobject *kobj;
10186                         kobj = space_info->block_group_kobjs[i];
10187                         space_info->block_group_kobjs[i] = NULL;
10188                         if (kobj) {
10189                                 kobject_del(kobj);
10190                                 kobject_put(kobj);
10191                         }
10192                 }
10193                 kobject_del(&space_info->kobj);
10194                 kobject_put(&space_info->kobj);
10195         }
10196         return 0;
10197 }
10198
10199 /* link_block_group will queue up kobjects to add when we're reclaim-safe */
10200 void btrfs_add_raid_kobjects(struct btrfs_fs_info *fs_info)
10201 {
10202         struct btrfs_space_info *space_info;
10203         struct raid_kobject *rkobj;
10204         LIST_HEAD(list);
10205         int index;
10206         int ret = 0;
10207
10208         spin_lock(&fs_info->pending_raid_kobjs_lock);
10209         list_splice_init(&fs_info->pending_raid_kobjs, &list);
10210         spin_unlock(&fs_info->pending_raid_kobjs_lock);
10211
10212         list_for_each_entry(rkobj, &list, list) {
10213                 space_info = __find_space_info(fs_info, rkobj->flags);
10214                 index = btrfs_bg_flags_to_raid_index(rkobj->flags);
10215
10216                 ret = kobject_add(&rkobj->kobj, &space_info->kobj,
10217                                   "%s", get_raid_name(index));
10218                 if (ret) {
10219                         kobject_put(&rkobj->kobj);
10220                         break;
10221                 }
10222         }
10223         if (ret)
10224                 btrfs_warn(fs_info,
10225                            "failed to add kobject for block cache, ignoring");
10226 }
10227
10228 static void link_block_group(struct btrfs_block_group_cache *cache)
10229 {
10230         struct btrfs_space_info *space_info = cache->space_info;
10231         struct btrfs_fs_info *fs_info = cache->fs_info;
10232         int index = btrfs_bg_flags_to_raid_index(cache->flags);
10233         bool first = false;
10234
10235         down_write(&space_info->groups_sem);
10236         if (list_empty(&space_info->block_groups[index]))
10237                 first = true;
10238         list_add_tail(&cache->list, &space_info->block_groups[index]);
10239         up_write(&space_info->groups_sem);
10240
10241         if (first) {
10242                 struct raid_kobject *rkobj = kzalloc(sizeof(*rkobj), GFP_NOFS);
10243                 if (!rkobj) {
10244                         btrfs_warn(cache->fs_info,
10245                                 "couldn't alloc memory for raid level kobject");
10246                         return;
10247                 }
10248                 rkobj->flags = cache->flags;
10249                 kobject_init(&rkobj->kobj, &btrfs_raid_ktype);
10250
10251                 spin_lock(&fs_info->pending_raid_kobjs_lock);
10252                 list_add_tail(&rkobj->list, &fs_info->pending_raid_kobjs);
10253                 spin_unlock(&fs_info->pending_raid_kobjs_lock);
10254                 space_info->block_group_kobjs[index] = &rkobj->kobj;
10255         }
10256 }
10257
10258 static struct btrfs_block_group_cache *
10259 btrfs_create_block_group_cache(struct btrfs_fs_info *fs_info,
10260                                u64 start, u64 size)
10261 {
10262         struct btrfs_block_group_cache *cache;
10263
10264         cache = kzalloc(sizeof(*cache), GFP_NOFS);
10265         if (!cache)
10266                 return NULL;
10267
10268         cache->free_space_ctl = kzalloc(sizeof(*cache->free_space_ctl),
10269                                         GFP_NOFS);
10270         if (!cache->free_space_ctl) {
10271                 kfree(cache);
10272                 return NULL;
10273         }
10274
10275         cache->key.objectid = start;
10276         cache->key.offset = size;
10277         cache->key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
10278
10279         cache->fs_info = fs_info;
10280         cache->full_stripe_len = btrfs_full_stripe_len(fs_info, start);
10281         set_free_space_tree_thresholds(cache);
10282
10283         atomic_set(&cache->count, 1);
10284         spin_lock_init(&cache->lock);
10285         init_rwsem(&cache->data_rwsem);
10286         INIT_LIST_HEAD(&cache->list);
10287         INIT_LIST_HEAD(&cache->cluster_list);
10288         INIT_LIST_HEAD(&cache->bg_list);
10289         INIT_LIST_HEAD(&cache->ro_list);
10290         INIT_LIST_HEAD(&cache->dirty_list);
10291         INIT_LIST_HEAD(&cache->io_list);
10292         btrfs_init_free_space_ctl(cache);
10293         atomic_set(&cache->trimming, 0);
10294         mutex_init(&cache->free_space_lock);
10295         btrfs_init_full_stripe_locks_tree(&cache->full_stripe_locks_root);
10296
10297         return cache;
10298 }
10299
10300
10301 /*
10302  * Iterate all chunks and verify that each of them has the corresponding block
10303  * group
10304  */
10305 static int check_chunk_block_group_mappings(struct btrfs_fs_info *fs_info)
10306 {
10307         struct btrfs_mapping_tree *map_tree = &fs_info->mapping_tree;
10308         struct extent_map *em;
10309         struct btrfs_block_group_cache *bg;
10310         u64 start = 0;
10311         int ret = 0;
10312
10313         while (1) {
10314                 read_lock(&map_tree->map_tree.lock);
10315                 /*
10316                  * lookup_extent_mapping will return the first extent map
10317                  * intersecting the range, so setting @len to 1 is enough to
10318                  * get the first chunk.
10319                  */
10320                 em = lookup_extent_mapping(&map_tree->map_tree, start, 1);
10321                 read_unlock(&map_tree->map_tree.lock);
10322                 if (!em)
10323                         break;
10324
10325                 bg = btrfs_lookup_block_group(fs_info, em->start);
10326                 if (!bg) {
10327                         btrfs_err(fs_info,
10328         "chunk start=%llu len=%llu doesn't have corresponding block group",
10329                                      em->start, em->len);
10330                         ret = -EUCLEAN;
10331                         free_extent_map(em);
10332                         break;
10333                 }
10334                 if (bg->key.objectid != em->start ||
10335                     bg->key.offset != em->len ||
10336                     (bg->flags & BTRFS_BLOCK_GROUP_TYPE_MASK) !=
10337                     (em->map_lookup->type & BTRFS_BLOCK_GROUP_TYPE_MASK)) {
10338                         btrfs_err(fs_info,
10339 "chunk start=%llu len=%llu flags=0x%llx doesn't match block group start=%llu len=%llu flags=0x%llx",
10340                                 em->start, em->len,
10341                                 em->map_lookup->type & BTRFS_BLOCK_GROUP_TYPE_MASK,
10342                                 bg->key.objectid, bg->key.offset,
10343                                 bg->flags & BTRFS_BLOCK_GROUP_TYPE_MASK);
10344                         ret = -EUCLEAN;
10345                         free_extent_map(em);
10346                         btrfs_put_block_group(bg);
10347                         break;
10348                 }
10349                 start = em->start + em->len;
10350                 free_extent_map(em);
10351                 btrfs_put_block_group(bg);
10352         }
10353         return ret;
10354 }
10355
10356 int btrfs_read_block_groups(struct btrfs_fs_info *info)
10357 {
10358         struct btrfs_path *path;
10359         int ret;
10360         struct btrfs_block_group_cache *cache;
10361         struct btrfs_space_info *space_info;
10362         struct btrfs_key key;
10363         struct btrfs_key found_key;
10364         struct extent_buffer *leaf;
10365         int need_clear = 0;
10366         u64 cache_gen;
10367         u64 feature;
10368         int mixed;
10369
10370         feature = btrfs_super_incompat_flags(info->super_copy);
10371         mixed = !!(feature & BTRFS_FEATURE_INCOMPAT_MIXED_GROUPS);
10372
10373         key.objectid = 0;
10374         key.offset = 0;
10375         key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
10376         path = btrfs_alloc_path();
10377         if (!path)
10378                 return -ENOMEM;
10379         path->reada = READA_FORWARD;
10380
10381         cache_gen = btrfs_super_cache_generation(info->super_copy);
10382         if (btrfs_test_opt(info, SPACE_CACHE) &&
10383             btrfs_super_generation(info->super_copy) != cache_gen)
10384                 need_clear = 1;
10385         if (btrfs_test_opt(info, CLEAR_CACHE))
10386                 need_clear = 1;
10387
10388         while (1) {
10389                 ret = find_first_block_group(info, path, &key);
10390                 if (ret > 0)
10391                         break;
10392                 if (ret != 0)
10393                         goto error;
10394
10395                 leaf = path->nodes[0];
10396                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
10397
10398                 cache = btrfs_create_block_group_cache(info, found_key.objectid,
10399                                                        found_key.offset);
10400                 if (!cache) {
10401                         ret = -ENOMEM;
10402                         goto error;
10403                 }
10404
10405                 if (need_clear) {
10406                         /*
10407                          * When we mount with old space cache, we need to
10408                          * set BTRFS_DC_CLEAR and set dirty flag.
10409                          *
10410                          * a) Setting 'BTRFS_DC_CLEAR' makes sure that we
10411                          *    truncate the old free space cache inode and
10412                          *    setup a new one.
10413                          * b) Setting 'dirty flag' makes sure that we flush
10414                          *    the new space cache info onto disk.
10415                          */
10416                         if (btrfs_test_opt(info, SPACE_CACHE))
10417                                 cache->disk_cache_state = BTRFS_DC_CLEAR;
10418                 }
10419
10420                 read_extent_buffer(leaf, &cache->item,
10421                                    btrfs_item_ptr_offset(leaf, path->slots[0]),
10422                                    sizeof(cache->item));
10423                 cache->flags = btrfs_block_group_flags(&cache->item);
10424                 if (!mixed &&
10425                     ((cache->flags & BTRFS_BLOCK_GROUP_METADATA) &&
10426                     (cache->flags & BTRFS_BLOCK_GROUP_DATA))) {
10427                         btrfs_err(info,
10428 "bg %llu is a mixed block group but filesystem hasn't enabled mixed block groups",
10429                                   cache->key.objectid);
10430                         ret = -EINVAL;
10431                         goto error;
10432                 }
10433
10434                 key.objectid = found_key.objectid + found_key.offset;
10435                 btrfs_release_path(path);
10436
10437                 /*
10438                  * We need to exclude the super stripes now so that the space
10439                  * info has super bytes accounted for, otherwise we'll think
10440                  * we have more space than we actually do.
10441                  */
10442                 ret = exclude_super_stripes(cache);
10443                 if (ret) {
10444                         /*
10445                          * We may have excluded something, so call this just in
10446                          * case.
10447                          */
10448                         free_excluded_extents(cache);
10449                         btrfs_put_block_group(cache);
10450                         goto error;
10451                 }
10452
10453                 /*
10454                  * check for two cases, either we are full, and therefore
10455                  * don't need to bother with the caching work since we won't
10456                  * find any space, or we are empty, and we can just add all
10457                  * the space in and be done with it.  This saves us _a_lot_ of
10458                  * time, particularly in the full case.
10459                  */
10460                 if (found_key.offset == btrfs_block_group_used(&cache->item)) {
10461                         cache->last_byte_to_unpin = (u64)-1;
10462                         cache->cached = BTRFS_CACHE_FINISHED;
10463                         free_excluded_extents(cache);
10464                 } else if (btrfs_block_group_used(&cache->item) == 0) {
10465                         cache->last_byte_to_unpin = (u64)-1;
10466                         cache->cached = BTRFS_CACHE_FINISHED;
10467                         add_new_free_space(cache, found_key.objectid,
10468                                            found_key.objectid +
10469                                            found_key.offset);
10470                         free_excluded_extents(cache);
10471                 }
10472
10473                 ret = btrfs_add_block_group_cache(info, cache);
10474                 if (ret) {
10475                         btrfs_remove_free_space_cache(cache);
10476                         btrfs_put_block_group(cache);
10477                         goto error;
10478                 }
10479
10480                 trace_btrfs_add_block_group(info, cache, 0);
10481                 update_space_info(info, cache->flags, found_key.offset,
10482                                   btrfs_block_group_used(&cache->item),
10483                                   cache->bytes_super, &space_info);
10484
10485                 cache->space_info = space_info;
10486
10487                 link_block_group(cache);
10488
10489                 set_avail_alloc_bits(info, cache->flags);
10490                 if (btrfs_chunk_readonly(info, cache->key.objectid)) {
10491                         inc_block_group_ro(cache, 1);
10492                 } else if (btrfs_block_group_used(&cache->item) == 0) {
10493                         ASSERT(list_empty(&cache->bg_list));
10494                         btrfs_mark_bg_unused(cache);
10495                 }
10496         }
10497
10498         list_for_each_entry_rcu(space_info, &info->space_info, list) {
10499                 if (!(get_alloc_profile(info, space_info->flags) &
10500                       (BTRFS_BLOCK_GROUP_RAID10 |
10501                        BTRFS_BLOCK_GROUP_RAID1 |
10502                        BTRFS_BLOCK_GROUP_RAID5 |
10503                        BTRFS_BLOCK_GROUP_RAID6 |
10504                        BTRFS_BLOCK_GROUP_DUP)))
10505                         continue;
10506                 /*
10507                  * avoid allocating from un-mirrored block group if there are
10508                  * mirrored block groups.
10509                  */
10510                 list_for_each_entry(cache,
10511                                 &space_info->block_groups[BTRFS_RAID_RAID0],
10512                                 list)
10513                         inc_block_group_ro(cache, 1);
10514                 list_for_each_entry(cache,
10515                                 &space_info->block_groups[BTRFS_RAID_SINGLE],
10516                                 list)
10517                         inc_block_group_ro(cache, 1);
10518         }
10519
10520         btrfs_add_raid_kobjects(info);
10521         init_global_block_rsv(info);
10522         ret = check_chunk_block_group_mappings(info);
10523 error:
10524         btrfs_free_path(path);
10525         return ret;
10526 }
10527
10528 void btrfs_create_pending_block_groups(struct btrfs_trans_handle *trans)
10529 {
10530         struct btrfs_fs_info *fs_info = trans->fs_info;
10531         struct btrfs_block_group_cache *block_group;
10532         struct btrfs_root *extent_root = fs_info->extent_root;
10533         struct btrfs_block_group_item item;
10534         struct btrfs_key key;
10535         int ret = 0;
10536
10537         if (!trans->can_flush_pending_bgs)
10538                 return;
10539
10540         while (!list_empty(&trans->new_bgs)) {
10541                 block_group = list_first_entry(&trans->new_bgs,
10542                                                struct btrfs_block_group_cache,
10543                                                bg_list);
10544                 if (ret)
10545                         goto next;
10546
10547                 spin_lock(&block_group->lock);
10548                 memcpy(&item, &block_group->item, sizeof(item));
10549                 memcpy(&key, &block_group->key, sizeof(key));
10550                 spin_unlock(&block_group->lock);
10551
10552                 ret = btrfs_insert_item(trans, extent_root, &key, &item,
10553                                         sizeof(item));
10554                 if (ret)
10555                         btrfs_abort_transaction(trans, ret);
10556                 ret = btrfs_finish_chunk_alloc(trans, key.objectid, key.offset);
10557                 if (ret)
10558                         btrfs_abort_transaction(trans, ret);
10559                 add_block_group_free_space(trans, block_group);
10560                 /* already aborted the transaction if it failed. */
10561 next:
10562                 btrfs_delayed_refs_rsv_release(fs_info, 1);
10563                 list_del_init(&block_group->bg_list);
10564         }
10565         btrfs_trans_release_chunk_metadata(trans);
10566 }
10567
10568 int btrfs_make_block_group(struct btrfs_trans_handle *trans, u64 bytes_used,
10569                            u64 type, u64 chunk_offset, u64 size)
10570 {
10571         struct btrfs_fs_info *fs_info = trans->fs_info;
10572         struct btrfs_block_group_cache *cache;
10573         int ret;
10574
10575         btrfs_set_log_full_commit(fs_info, trans);
10576
10577         cache = btrfs_create_block_group_cache(fs_info, chunk_offset, size);
10578         if (!cache)
10579                 return -ENOMEM;
10580
10581         btrfs_set_block_group_used(&cache->item, bytes_used);
10582         btrfs_set_block_group_chunk_objectid(&cache->item,
10583                                              BTRFS_FIRST_CHUNK_TREE_OBJECTID);
10584         btrfs_set_block_group_flags(&cache->item, type);
10585
10586         cache->flags = type;
10587         cache->last_byte_to_unpin = (u64)-1;
10588         cache->cached = BTRFS_CACHE_FINISHED;
10589         cache->needs_free_space = 1;
10590         ret = exclude_super_stripes(cache);
10591         if (ret) {
10592                 /*
10593                  * We may have excluded something, so call this just in
10594                  * case.
10595                  */
10596                 free_excluded_extents(cache);
10597                 btrfs_put_block_group(cache);
10598                 return ret;
10599         }
10600
10601         add_new_free_space(cache, chunk_offset, chunk_offset + size);
10602
10603         free_excluded_extents(cache);
10604
10605 #ifdef CONFIG_BTRFS_DEBUG
10606         if (btrfs_should_fragment_free_space(cache)) {
10607                 u64 new_bytes_used = size - bytes_used;
10608
10609                 bytes_used += new_bytes_used >> 1;
10610                 fragment_free_space(cache);
10611         }
10612 #endif
10613         /*
10614          * Ensure the corresponding space_info object is created and
10615          * assigned to our block group. We want our bg to be added to the rbtree
10616          * with its ->space_info set.
10617          */
10618         cache->space_info = __find_space_info(fs_info, cache->flags);
10619         ASSERT(cache->space_info);
10620
10621         ret = btrfs_add_block_group_cache(fs_info, cache);
10622         if (ret) {
10623                 btrfs_remove_free_space_cache(cache);
10624                 btrfs_put_block_group(cache);
10625                 return ret;
10626         }
10627
10628         /*
10629          * Now that our block group has its ->space_info set and is inserted in
10630          * the rbtree, update the space info's counters.
10631          */
10632         trace_btrfs_add_block_group(fs_info, cache, 1);
10633         update_space_info(fs_info, cache->flags, size, bytes_used,
10634                                 cache->bytes_super, &cache->space_info);
10635         update_global_block_rsv(fs_info);
10636
10637         link_block_group(cache);
10638
10639         list_add_tail(&cache->bg_list, &trans->new_bgs);
10640         trans->delayed_ref_updates++;
10641         btrfs_update_delayed_refs_rsv(trans);
10642
10643         set_avail_alloc_bits(fs_info, type);
10644         return 0;
10645 }
10646
10647 static void clear_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
10648 {
10649         u64 extra_flags = chunk_to_extended(flags) &
10650                                 BTRFS_EXTENDED_PROFILE_MASK;
10651
10652         write_seqlock(&fs_info->profiles_lock);
10653         if (flags & BTRFS_BLOCK_GROUP_DATA)
10654                 fs_info->avail_data_alloc_bits &= ~extra_flags;
10655         if (flags & BTRFS_BLOCK_GROUP_METADATA)
10656                 fs_info->avail_metadata_alloc_bits &= ~extra_flags;
10657         if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
10658                 fs_info->avail_system_alloc_bits &= ~extra_flags;
10659         write_sequnlock(&fs_info->profiles_lock);
10660 }
10661
10662 int btrfs_remove_block_group(struct btrfs_trans_handle *trans,
10663                              u64 group_start, struct extent_map *em)
10664 {
10665         struct btrfs_fs_info *fs_info = trans->fs_info;
10666         struct btrfs_root *root = fs_info->extent_root;
10667         struct btrfs_path *path;
10668         struct btrfs_block_group_cache *block_group;
10669         struct btrfs_free_cluster *cluster;
10670         struct btrfs_root *tree_root = fs_info->tree_root;
10671         struct btrfs_key key;
10672         struct inode *inode;
10673         struct kobject *kobj = NULL;
10674         int ret;
10675         int index;
10676         int factor;
10677         struct btrfs_caching_control *caching_ctl = NULL;
10678         bool remove_em;
10679         bool remove_rsv = false;
10680
10681         block_group = btrfs_lookup_block_group(fs_info, group_start);
10682         BUG_ON(!block_group);
10683         BUG_ON(!block_group->ro);
10684
10685         trace_btrfs_remove_block_group(block_group);
10686         /*
10687          * Free the reserved super bytes from this block group before
10688          * remove it.
10689          */
10690         free_excluded_extents(block_group);
10691         btrfs_free_ref_tree_range(fs_info, block_group->key.objectid,
10692                                   block_group->key.offset);
10693
10694         memcpy(&key, &block_group->key, sizeof(key));
10695         index = btrfs_bg_flags_to_raid_index(block_group->flags);
10696         factor = btrfs_bg_type_to_factor(block_group->flags);
10697
10698         /* make sure this block group isn't part of an allocation cluster */
10699         cluster = &fs_info->data_alloc_cluster;
10700         spin_lock(&cluster->refill_lock);
10701         btrfs_return_cluster_to_free_space(block_group, cluster);
10702         spin_unlock(&cluster->refill_lock);
10703
10704         /*
10705          * make sure this block group isn't part of a metadata
10706          * allocation cluster
10707          */
10708         cluster = &fs_info->meta_alloc_cluster;
10709         spin_lock(&cluster->refill_lock);
10710         btrfs_return_cluster_to_free_space(block_group, cluster);
10711         spin_unlock(&cluster->refill_lock);
10712
10713         path = btrfs_alloc_path();
10714         if (!path) {
10715                 ret = -ENOMEM;
10716                 goto out;
10717         }
10718
10719         /*
10720          * get the inode first so any iput calls done for the io_list
10721          * aren't the final iput (no unlinks allowed now)
10722          */
10723         inode = lookup_free_space_inode(fs_info, block_group, path);
10724
10725         mutex_lock(&trans->transaction->cache_write_mutex);
10726         /*
10727          * Make sure our free space cache IO is done before removing the
10728          * free space inode
10729          */
10730         spin_lock(&trans->transaction->dirty_bgs_lock);
10731         if (!list_empty(&block_group->io_list)) {
10732                 list_del_init(&block_group->io_list);
10733
10734                 WARN_ON(!IS_ERR(inode) && inode != block_group->io_ctl.inode);
10735
10736                 spin_unlock(&trans->transaction->dirty_bgs_lock);
10737                 btrfs_wait_cache_io(trans, block_group, path);
10738                 btrfs_put_block_group(block_group);
10739                 spin_lock(&trans->transaction->dirty_bgs_lock);
10740         }
10741
10742         if (!list_empty(&block_group->dirty_list)) {
10743                 list_del_init(&block_group->dirty_list);
10744                 remove_rsv = true;
10745                 btrfs_put_block_group(block_group);
10746         }
10747         spin_unlock(&trans->transaction->dirty_bgs_lock);
10748         mutex_unlock(&trans->transaction->cache_write_mutex);
10749
10750         if (!IS_ERR(inode)) {
10751                 ret = btrfs_orphan_add(trans, BTRFS_I(inode));
10752                 if (ret) {
10753                         btrfs_add_delayed_iput(inode);
10754                         goto out;
10755                 }
10756                 clear_nlink(inode);
10757                 /* One for the block groups ref */
10758                 spin_lock(&block_group->lock);
10759                 if (block_group->iref) {
10760                         block_group->iref = 0;
10761                         block_group->inode = NULL;
10762                         spin_unlock(&block_group->lock);
10763                         iput(inode);
10764                 } else {
10765                         spin_unlock(&block_group->lock);
10766                 }
10767                 /* One for our lookup ref */
10768                 btrfs_add_delayed_iput(inode);
10769         }
10770
10771         key.objectid = BTRFS_FREE_SPACE_OBJECTID;
10772         key.offset = block_group->key.objectid;
10773         key.type = 0;
10774
10775         ret = btrfs_search_slot(trans, tree_root, &key, path, -1, 1);
10776         if (ret < 0)
10777                 goto out;
10778         if (ret > 0)
10779                 btrfs_release_path(path);
10780         if (ret == 0) {
10781                 ret = btrfs_del_item(trans, tree_root, path);
10782                 if (ret)
10783                         goto out;
10784                 btrfs_release_path(path);
10785         }
10786
10787         spin_lock(&fs_info->block_group_cache_lock);
10788         rb_erase(&block_group->cache_node,
10789                  &fs_info->block_group_cache_tree);
10790         RB_CLEAR_NODE(&block_group->cache_node);
10791
10792         if (fs_info->first_logical_byte == block_group->key.objectid)
10793                 fs_info->first_logical_byte = (u64)-1;
10794         spin_unlock(&fs_info->block_group_cache_lock);
10795
10796         down_write(&block_group->space_info->groups_sem);
10797         /*
10798          * we must use list_del_init so people can check to see if they
10799          * are still on the list after taking the semaphore
10800          */
10801         list_del_init(&block_group->list);
10802         if (list_empty(&block_group->space_info->block_groups[index])) {
10803                 kobj = block_group->space_info->block_group_kobjs[index];
10804                 block_group->space_info->block_group_kobjs[index] = NULL;
10805                 clear_avail_alloc_bits(fs_info, block_group->flags);
10806         }
10807         up_write(&block_group->space_info->groups_sem);
10808         if (kobj) {
10809                 kobject_del(kobj);
10810                 kobject_put(kobj);
10811         }
10812
10813         if (block_group->has_caching_ctl)
10814                 caching_ctl = get_caching_control(block_group);
10815         if (block_group->cached == BTRFS_CACHE_STARTED)
10816                 wait_block_group_cache_done(block_group);
10817         if (block_group->has_caching_ctl) {
10818                 down_write(&fs_info->commit_root_sem);
10819                 if (!caching_ctl) {
10820                         struct btrfs_caching_control *ctl;
10821
10822                         list_for_each_entry(ctl,
10823                                     &fs_info->caching_block_groups, list)
10824                                 if (ctl->block_group == block_group) {
10825                                         caching_ctl = ctl;
10826                                         refcount_inc(&caching_ctl->count);
10827                                         break;
10828                                 }
10829                 }
10830                 if (caching_ctl)
10831                         list_del_init(&caching_ctl->list);
10832                 up_write(&fs_info->commit_root_sem);
10833                 if (caching_ctl) {
10834                         /* Once for the caching bgs list and once for us. */
10835                         put_caching_control(caching_ctl);
10836                         put_caching_control(caching_ctl);
10837                 }
10838         }
10839
10840         spin_lock(&trans->transaction->dirty_bgs_lock);
10841         if (!list_empty(&block_group->dirty_list)) {
10842                 WARN_ON(1);
10843         }
10844         if (!list_empty(&block_group->io_list)) {
10845                 WARN_ON(1);
10846         }
10847         spin_unlock(&trans->transaction->dirty_bgs_lock);
10848         btrfs_remove_free_space_cache(block_group);
10849
10850         spin_lock(&block_group->space_info->lock);
10851         list_del_init(&block_group->ro_list);
10852
10853         if (btrfs_test_opt(fs_info, ENOSPC_DEBUG)) {
10854                 WARN_ON(block_group->space_info->total_bytes
10855                         < block_group->key.offset);
10856                 WARN_ON(block_group->space_info->bytes_readonly
10857                         < block_group->key.offset);
10858                 WARN_ON(block_group->space_info->disk_total
10859                         < block_group->key.offset * factor);
10860         }
10861         block_group->space_info->total_bytes -= block_group->key.offset;
10862         block_group->space_info->bytes_readonly -= block_group->key.offset;
10863         block_group->space_info->disk_total -= block_group->key.offset * factor;
10864
10865         spin_unlock(&block_group->space_info->lock);
10866
10867         memcpy(&key, &block_group->key, sizeof(key));
10868
10869         mutex_lock(&fs_info->chunk_mutex);
10870         if (!list_empty(&em->list)) {
10871                 /* We're in the transaction->pending_chunks list. */
10872                 free_extent_map(em);
10873         }
10874         spin_lock(&block_group->lock);
10875         block_group->removed = 1;
10876         /*
10877          * At this point trimming can't start on this block group, because we
10878          * removed the block group from the tree fs_info->block_group_cache_tree
10879          * so no one can't find it anymore and even if someone already got this
10880          * block group before we removed it from the rbtree, they have already
10881          * incremented block_group->trimming - if they didn't, they won't find
10882          * any free space entries because we already removed them all when we
10883          * called btrfs_remove_free_space_cache().
10884          *
10885          * And we must not remove the extent map from the fs_info->mapping_tree
10886          * to prevent the same logical address range and physical device space
10887          * ranges from being reused for a new block group. This is because our
10888          * fs trim operation (btrfs_trim_fs() / btrfs_ioctl_fitrim()) is
10889          * completely transactionless, so while it is trimming a range the
10890          * currently running transaction might finish and a new one start,
10891          * allowing for new block groups to be created that can reuse the same
10892          * physical device locations unless we take this special care.
10893          *
10894          * There may also be an implicit trim operation if the file system
10895          * is mounted with -odiscard. The same protections must remain
10896          * in place until the extents have been discarded completely when
10897          * the transaction commit has completed.
10898          */
10899         remove_em = (atomic_read(&block_group->trimming) == 0);
10900         /*
10901          * Make sure a trimmer task always sees the em in the pinned_chunks list
10902          * if it sees block_group->removed == 1 (needs to lock block_group->lock
10903          * before checking block_group->removed).
10904          */
10905         if (!remove_em) {
10906                 /*
10907                  * Our em might be in trans->transaction->pending_chunks which
10908                  * is protected by fs_info->chunk_mutex ([lock|unlock]_chunks),
10909                  * and so is the fs_info->pinned_chunks list.
10910                  *
10911                  * So at this point we must be holding the chunk_mutex to avoid
10912                  * any races with chunk allocation (more specifically at
10913                  * volumes.c:contains_pending_extent()), to ensure it always
10914                  * sees the em, either in the pending_chunks list or in the
10915                  * pinned_chunks list.
10916                  */
10917                 list_move_tail(&em->list, &fs_info->pinned_chunks);
10918         }
10919         spin_unlock(&block_group->lock);
10920
10921         if (remove_em) {
10922                 struct extent_map_tree *em_tree;
10923
10924                 em_tree = &fs_info->mapping_tree.map_tree;
10925                 write_lock(&em_tree->lock);
10926                 /*
10927                  * The em might be in the pending_chunks list, so make sure the
10928                  * chunk mutex is locked, since remove_extent_mapping() will
10929                  * delete us from that list.
10930                  */
10931                 remove_extent_mapping(em_tree, em);
10932                 write_unlock(&em_tree->lock);
10933                 /* once for the tree */
10934                 free_extent_map(em);
10935         }
10936
10937         mutex_unlock(&fs_info->chunk_mutex);
10938
10939         ret = remove_block_group_free_space(trans, block_group);
10940         if (ret)
10941                 goto out;
10942
10943         btrfs_put_block_group(block_group);
10944         btrfs_put_block_group(block_group);
10945
10946         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
10947         if (ret > 0)
10948                 ret = -EIO;
10949         if (ret < 0)
10950                 goto out;
10951
10952         ret = btrfs_del_item(trans, root, path);
10953 out:
10954         if (remove_rsv)
10955                 btrfs_delayed_refs_rsv_release(fs_info, 1);
10956         btrfs_free_path(path);
10957         return ret;
10958 }
10959
10960 struct btrfs_trans_handle *
10961 btrfs_start_trans_remove_block_group(struct btrfs_fs_info *fs_info,
10962                                      const u64 chunk_offset)
10963 {
10964         struct extent_map_tree *em_tree = &fs_info->mapping_tree.map_tree;
10965         struct extent_map *em;
10966         struct map_lookup *map;
10967         unsigned int num_items;
10968
10969         read_lock(&em_tree->lock);
10970         em = lookup_extent_mapping(em_tree, chunk_offset, 1);
10971         read_unlock(&em_tree->lock);
10972         ASSERT(em && em->start == chunk_offset);
10973
10974         /*
10975          * We need to reserve 3 + N units from the metadata space info in order
10976          * to remove a block group (done at btrfs_remove_chunk() and at
10977          * btrfs_remove_block_group()), which are used for:
10978          *
10979          * 1 unit for adding the free space inode's orphan (located in the tree
10980          * of tree roots).
10981          * 1 unit for deleting the block group item (located in the extent
10982          * tree).
10983          * 1 unit for deleting the free space item (located in tree of tree
10984          * roots).
10985          * N units for deleting N device extent items corresponding to each
10986          * stripe (located in the device tree).
10987          *
10988          * In order to remove a block group we also need to reserve units in the
10989          * system space info in order to update the chunk tree (update one or
10990          * more device items and remove one chunk item), but this is done at
10991          * btrfs_remove_chunk() through a call to check_system_chunk().
10992          */
10993         map = em->map_lookup;
10994         num_items = 3 + map->num_stripes;
10995         free_extent_map(em);
10996
10997         return btrfs_start_transaction_fallback_global_rsv(fs_info->extent_root,
10998                                                            num_items, 1);
10999 }
11000
11001 /*
11002  * Process the unused_bgs list and remove any that don't have any allocated
11003  * space inside of them.
11004  */
11005 void btrfs_delete_unused_bgs(struct btrfs_fs_info *fs_info)
11006 {
11007         struct btrfs_block_group_cache *block_group;
11008         struct btrfs_space_info *space_info;
11009         struct btrfs_trans_handle *trans;
11010         int ret = 0;
11011
11012         if (!test_bit(BTRFS_FS_OPEN, &fs_info->flags))
11013                 return;
11014
11015         spin_lock(&fs_info->unused_bgs_lock);
11016         while (!list_empty(&fs_info->unused_bgs)) {
11017                 u64 start, end;
11018                 int trimming;
11019
11020                 block_group = list_first_entry(&fs_info->unused_bgs,
11021                                                struct btrfs_block_group_cache,
11022                                                bg_list);
11023                 list_del_init(&block_group->bg_list);
11024
11025                 space_info = block_group->space_info;
11026
11027                 if (ret || btrfs_mixed_space_info(space_info)) {
11028                         btrfs_put_block_group(block_group);
11029                         continue;
11030                 }
11031                 spin_unlock(&fs_info->unused_bgs_lock);
11032
11033                 mutex_lock(&fs_info->delete_unused_bgs_mutex);
11034
11035                 /* Don't want to race with allocators so take the groups_sem */
11036                 down_write(&space_info->groups_sem);
11037                 spin_lock(&block_group->lock);
11038                 if (block_group->reserved || block_group->pinned ||
11039                     btrfs_block_group_used(&block_group->item) ||
11040                     block_group->ro ||
11041                     list_is_singular(&block_group->list)) {
11042                         /*
11043                          * We want to bail if we made new allocations or have
11044                          * outstanding allocations in this block group.  We do
11045                          * the ro check in case balance is currently acting on
11046                          * this block group.
11047                          */
11048                         trace_btrfs_skip_unused_block_group(block_group);
11049                         spin_unlock(&block_group->lock);
11050                         up_write(&space_info->groups_sem);
11051                         goto next;
11052                 }
11053                 spin_unlock(&block_group->lock);
11054
11055                 /* We don't want to force the issue, only flip if it's ok. */
11056                 ret = inc_block_group_ro(block_group, 0);
11057                 up_write(&space_info->groups_sem);
11058                 if (ret < 0) {
11059                         ret = 0;
11060                         goto next;
11061                 }
11062
11063                 /*
11064                  * Want to do this before we do anything else so we can recover
11065                  * properly if we fail to join the transaction.
11066                  */
11067                 trans = btrfs_start_trans_remove_block_group(fs_info,
11068                                                      block_group->key.objectid);
11069                 if (IS_ERR(trans)) {
11070                         btrfs_dec_block_group_ro(block_group);
11071                         ret = PTR_ERR(trans);
11072                         goto next;
11073                 }
11074
11075                 /*
11076                  * We could have pending pinned extents for this block group,
11077                  * just delete them, we don't care about them anymore.
11078                  */
11079                 start = block_group->key.objectid;
11080                 end = start + block_group->key.offset - 1;
11081                 /*
11082                  * Hold the unused_bg_unpin_mutex lock to avoid racing with
11083                  * btrfs_finish_extent_commit(). If we are at transaction N,
11084                  * another task might be running finish_extent_commit() for the
11085                  * previous transaction N - 1, and have seen a range belonging
11086                  * to the block group in freed_extents[] before we were able to
11087                  * clear the whole block group range from freed_extents[]. This
11088                  * means that task can lookup for the block group after we
11089                  * unpinned it from freed_extents[] and removed it, leading to
11090                  * a BUG_ON() at btrfs_unpin_extent_range().
11091                  */
11092                 mutex_lock(&fs_info->unused_bg_unpin_mutex);
11093                 ret = clear_extent_bits(&fs_info->freed_extents[0], start, end,
11094                                   EXTENT_DIRTY);
11095                 if (ret) {
11096                         mutex_unlock(&fs_info->unused_bg_unpin_mutex);
11097                         btrfs_dec_block_group_ro(block_group);
11098                         goto end_trans;
11099                 }
11100                 ret = clear_extent_bits(&fs_info->freed_extents[1], start, end,
11101                                   EXTENT_DIRTY);
11102                 if (ret) {
11103                         mutex_unlock(&fs_info->unused_bg_unpin_mutex);
11104                         btrfs_dec_block_group_ro(block_group);
11105                         goto end_trans;
11106                 }
11107                 mutex_unlock(&fs_info->unused_bg_unpin_mutex);
11108
11109                 /* Reset pinned so btrfs_put_block_group doesn't complain */
11110                 spin_lock(&space_info->lock);
11111                 spin_lock(&block_group->lock);
11112
11113                 update_bytes_pinned(space_info, -block_group->pinned);
11114                 space_info->bytes_readonly += block_group->pinned;
11115                 percpu_counter_add_batch(&space_info->total_bytes_pinned,
11116                                    -block_group->pinned,
11117                                    BTRFS_TOTAL_BYTES_PINNED_BATCH);
11118                 block_group->pinned = 0;
11119
11120                 spin_unlock(&block_group->lock);
11121                 spin_unlock(&space_info->lock);
11122
11123                 /* DISCARD can flip during remount */
11124                 trimming = btrfs_test_opt(fs_info, DISCARD);
11125
11126                 /* Implicit trim during transaction commit. */
11127                 if (trimming)
11128                         btrfs_get_block_group_trimming(block_group);
11129
11130                 /*
11131                  * Btrfs_remove_chunk will abort the transaction if things go
11132                  * horribly wrong.
11133                  */
11134                 ret = btrfs_remove_chunk(trans, block_group->key.objectid);
11135
11136                 if (ret) {
11137                         if (trimming)
11138                                 btrfs_put_block_group_trimming(block_group);
11139                         goto end_trans;
11140                 }
11141
11142                 /*
11143                  * If we're not mounted with -odiscard, we can just forget
11144                  * about this block group. Otherwise we'll need to wait
11145                  * until transaction commit to do the actual discard.
11146                  */
11147                 if (trimming) {
11148                         spin_lock(&fs_info->unused_bgs_lock);
11149                         /*
11150                          * A concurrent scrub might have added us to the list
11151                          * fs_info->unused_bgs, so use a list_move operation
11152                          * to add the block group to the deleted_bgs list.
11153                          */
11154                         list_move(&block_group->bg_list,
11155                                   &trans->transaction->deleted_bgs);
11156                         spin_unlock(&fs_info->unused_bgs_lock);
11157                         btrfs_get_block_group(block_group);
11158                 }
11159 end_trans:
11160                 btrfs_end_transaction(trans);
11161 next:
11162                 mutex_unlock(&fs_info->delete_unused_bgs_mutex);
11163                 btrfs_put_block_group(block_group);
11164                 spin_lock(&fs_info->unused_bgs_lock);
11165         }
11166         spin_unlock(&fs_info->unused_bgs_lock);
11167 }
11168
11169 int btrfs_init_space_info(struct btrfs_fs_info *fs_info)
11170 {
11171         struct btrfs_super_block *disk_super;
11172         u64 features;
11173         u64 flags;
11174         int mixed = 0;
11175         int ret;
11176
11177         disk_super = fs_info->super_copy;
11178         if (!btrfs_super_root(disk_super))
11179                 return -EINVAL;
11180
11181         features = btrfs_super_incompat_flags(disk_super);
11182         if (features & BTRFS_FEATURE_INCOMPAT_MIXED_GROUPS)
11183                 mixed = 1;
11184
11185         flags = BTRFS_BLOCK_GROUP_SYSTEM;
11186         ret = create_space_info(fs_info, flags);
11187         if (ret)
11188                 goto out;
11189
11190         if (mixed) {
11191                 flags = BTRFS_BLOCK_GROUP_METADATA | BTRFS_BLOCK_GROUP_DATA;
11192                 ret = create_space_info(fs_info, flags);
11193         } else {
11194                 flags = BTRFS_BLOCK_GROUP_METADATA;
11195                 ret = create_space_info(fs_info, flags);
11196                 if (ret)
11197                         goto out;
11198
11199                 flags = BTRFS_BLOCK_GROUP_DATA;
11200                 ret = create_space_info(fs_info, flags);
11201         }
11202 out:
11203         return ret;
11204 }
11205
11206 int btrfs_error_unpin_extent_range(struct btrfs_fs_info *fs_info,
11207                                    u64 start, u64 end)
11208 {
11209         return unpin_extent_range(fs_info, start, end, false);
11210 }
11211
11212 /*
11213  * It used to be that old block groups would be left around forever.
11214  * Iterating over them would be enough to trim unused space.  Since we
11215  * now automatically remove them, we also need to iterate over unallocated
11216  * space.
11217  *
11218  * We don't want a transaction for this since the discard may take a
11219  * substantial amount of time.  We don't require that a transaction be
11220  * running, but we do need to take a running transaction into account
11221  * to ensure that we're not discarding chunks that were released or
11222  * allocated in the current transaction.
11223  *
11224  * Holding the chunks lock will prevent other threads from allocating
11225  * or releasing chunks, but it won't prevent a running transaction
11226  * from committing and releasing the memory that the pending chunks
11227  * list head uses.  For that, we need to take a reference to the
11228  * transaction and hold the commit root sem.  We only need to hold
11229  * it while performing the free space search since we have already
11230  * held back allocations.
11231  */
11232 static int btrfs_trim_free_extents(struct btrfs_device *device,
11233                                    u64 minlen, u64 *trimmed)
11234 {
11235         u64 start = 0, len = 0;
11236         int ret;
11237
11238         *trimmed = 0;
11239
11240         /* Discard not supported = nothing to do. */
11241         if (!blk_queue_discard(bdev_get_queue(device->bdev)))
11242                 return 0;
11243
11244         /* Not writable = nothing to do. */
11245         if (!test_bit(BTRFS_DEV_STATE_WRITEABLE, &device->dev_state))
11246                 return 0;
11247
11248         /* No free space = nothing to do. */
11249         if (device->total_bytes <= device->bytes_used)
11250                 return 0;
11251
11252         ret = 0;
11253
11254         while (1) {
11255                 struct btrfs_fs_info *fs_info = device->fs_info;
11256                 struct btrfs_transaction *trans;
11257                 u64 bytes;
11258
11259                 ret = mutex_lock_interruptible(&fs_info->chunk_mutex);
11260                 if (ret)
11261                         break;
11262
11263                 ret = down_read_killable(&fs_info->commit_root_sem);
11264                 if (ret) {
11265                         mutex_unlock(&fs_info->chunk_mutex);
11266                         break;
11267                 }
11268
11269                 spin_lock(&fs_info->trans_lock);
11270                 trans = fs_info->running_transaction;
11271                 if (trans)
11272                         refcount_inc(&trans->use_count);
11273                 spin_unlock(&fs_info->trans_lock);
11274
11275                 if (!trans)
11276                         up_read(&fs_info->commit_root_sem);
11277
11278                 ret = find_free_dev_extent_start(trans, device, minlen, start,
11279                                                  &start, &len);
11280                 if (trans) {
11281                         up_read(&fs_info->commit_root_sem);
11282                         btrfs_put_transaction(trans);
11283                 }
11284
11285                 if (ret) {
11286                         mutex_unlock(&fs_info->chunk_mutex);
11287                         if (ret == -ENOSPC)
11288                                 ret = 0;
11289                         break;
11290                 }
11291
11292                 ret = btrfs_issue_discard(device->bdev, start, len, &bytes);
11293                 mutex_unlock(&fs_info->chunk_mutex);
11294
11295                 if (ret)
11296                         break;
11297
11298                 start += len;
11299                 *trimmed += bytes;
11300
11301                 if (fatal_signal_pending(current)) {
11302                         ret = -ERESTARTSYS;
11303                         break;
11304                 }
11305
11306                 cond_resched();
11307         }
11308
11309         return ret;
11310 }
11311
11312 /*
11313  * Trim the whole filesystem by:
11314  * 1) trimming the free space in each block group
11315  * 2) trimming the unallocated space on each device
11316  *
11317  * This will also continue trimming even if a block group or device encounters
11318  * an error.  The return value will be the last error, or 0 if nothing bad
11319  * happens.
11320  */
11321 int btrfs_trim_fs(struct btrfs_fs_info *fs_info, struct fstrim_range *range)
11322 {
11323         struct btrfs_block_group_cache *cache = NULL;
11324         struct btrfs_device *device;
11325         struct list_head *devices;
11326         u64 group_trimmed;
11327         u64 start;
11328         u64 end;
11329         u64 trimmed = 0;
11330         u64 bg_failed = 0;
11331         u64 dev_failed = 0;
11332         int bg_ret = 0;
11333         int dev_ret = 0;
11334         int ret = 0;
11335
11336         cache = btrfs_lookup_first_block_group(fs_info, range->start);
11337         for (; cache; cache = next_block_group(fs_info, cache)) {
11338                 if (cache->key.objectid >= (range->start + range->len)) {
11339                         btrfs_put_block_group(cache);
11340                         break;
11341                 }
11342
11343                 start = max(range->start, cache->key.objectid);
11344                 end = min(range->start + range->len,
11345                                 cache->key.objectid + cache->key.offset);
11346
11347                 if (end - start >= range->minlen) {
11348                         if (!block_group_cache_done(cache)) {
11349                                 ret = cache_block_group(cache, 0);
11350                                 if (ret) {
11351                                         bg_failed++;
11352                                         bg_ret = ret;
11353                                         continue;
11354                                 }
11355                                 ret = wait_block_group_cache_done(cache);
11356                                 if (ret) {
11357                                         bg_failed++;
11358                                         bg_ret = ret;
11359                                         continue;
11360                                 }
11361                         }
11362                         ret = btrfs_trim_block_group(cache,
11363                                                      &group_trimmed,
11364                                                      start,
11365                                                      end,
11366                                                      range->minlen);
11367
11368                         trimmed += group_trimmed;
11369                         if (ret) {
11370                                 bg_failed++;
11371                                 bg_ret = ret;
11372                                 continue;
11373                         }
11374                 }
11375         }
11376
11377         if (bg_failed)
11378                 btrfs_warn(fs_info,
11379                         "failed to trim %llu block group(s), last error %d",
11380                         bg_failed, bg_ret);
11381         mutex_lock(&fs_info->fs_devices->device_list_mutex);
11382         devices = &fs_info->fs_devices->devices;
11383         list_for_each_entry(device, devices, dev_list) {
11384                 ret = btrfs_trim_free_extents(device, range->minlen,
11385                                               &group_trimmed);
11386                 if (ret) {
11387                         dev_failed++;
11388                         dev_ret = ret;
11389                         break;
11390                 }
11391
11392                 trimmed += group_trimmed;
11393         }
11394         mutex_unlock(&fs_info->fs_devices->device_list_mutex);
11395
11396         if (dev_failed)
11397                 btrfs_warn(fs_info,
11398                         "failed to trim %llu device(s), last error %d",
11399                         dev_failed, dev_ret);
11400         range->len = trimmed;
11401         if (bg_ret)
11402                 return bg_ret;
11403         return dev_ret;
11404 }
11405
11406 /*
11407  * btrfs_{start,end}_write_no_snapshotting() are similar to
11408  * mnt_{want,drop}_write(), they are used to prevent some tasks from writing
11409  * data into the page cache through nocow before the subvolume is snapshoted,
11410  * but flush the data into disk after the snapshot creation, or to prevent
11411  * operations while snapshotting is ongoing and that cause the snapshot to be
11412  * inconsistent (writes followed by expanding truncates for example).
11413  */
11414 void btrfs_end_write_no_snapshotting(struct btrfs_root *root)
11415 {
11416         percpu_counter_dec(&root->subv_writers->counter);
11417         cond_wake_up(&root->subv_writers->wait);
11418 }
11419
11420 int btrfs_start_write_no_snapshotting(struct btrfs_root *root)
11421 {
11422         if (atomic_read(&root->will_be_snapshotted))
11423                 return 0;
11424
11425         percpu_counter_inc(&root->subv_writers->counter);
11426         /*
11427          * Make sure counter is updated before we check for snapshot creation.
11428          */
11429         smp_mb();
11430         if (atomic_read(&root->will_be_snapshotted)) {
11431                 btrfs_end_write_no_snapshotting(root);
11432                 return 0;
11433         }
11434         return 1;
11435 }
11436
11437 void btrfs_wait_for_snapshot_creation(struct btrfs_root *root)
11438 {
11439         while (true) {
11440                 int ret;
11441
11442                 ret = btrfs_start_write_no_snapshotting(root);
11443                 if (ret)
11444                         break;
11445                 wait_var_event(&root->will_be_snapshotted,
11446                                !atomic_read(&root->will_be_snapshotted));
11447         }
11448 }
11449
11450 void btrfs_mark_bg_unused(struct btrfs_block_group_cache *bg)
11451 {
11452         struct btrfs_fs_info *fs_info = bg->fs_info;
11453
11454         spin_lock(&fs_info->unused_bgs_lock);
11455         if (list_empty(&bg->bg_list)) {
11456                 btrfs_get_block_group(bg);
11457                 trace_btrfs_add_unused_block_group(bg);
11458                 list_add_tail(&bg->bg_list, &fs_info->unused_bgs);
11459         }
11460         spin_unlock(&fs_info->unused_bgs_lock);
11461 }