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