fs/btrfs/extent-tree.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Copyright (C) 2007 Oracle.  All rights reserved.
   4  */
   5
   6 #include <linux/sched.h>
   7 #include <linux/sched/signal.h>
   8 #include <linux/pagemap.h>
   9 #include <linux/writeback.h>
  10 #include <linux/blkdev.h>
  11 #include <linux/sort.h>
  12 #include <linux/rcupdate.h>
  13 #include <linux/kthread.h>
  14 #include <linux/slab.h>
  15 #include <linux/ratelimit.h>
  16 #include <linux/percpu_counter.h>
  17 #include <linux/lockdep.h>
  18 #include <linux/crc32c.h>
  19 #include "tree-log.h"
  20 #include "disk-io.h"
  21 #include "print-tree.h"
  22 #include "volumes.h"
  23 #include "raid56.h"
  24 #include "locking.h"
  25 #include "free-space-cache.h"
  26 #include "free-space-tree.h"
  27 #include "math.h"
  28 #include "sysfs.h"
  29 #include "qgroup.h"
  30 #include "ref-verify.h"
  31
  32 #undef SCRAMBLE_DELAYED_REFS
  33
  34 /*
  35  * control flags for do_chunk_alloc's force field
  36  * CHUNK_ALLOC_NO_FORCE means to only allocate a chunk
  37  * if we really need one.
  38  *
  39  * CHUNK_ALLOC_LIMITED means to only try and allocate one
  40  * if we have very few chunks already allocated.  This is
  41  * used as part of the clustering code to help make sure
  42  * we have a good pool of storage to cluster in, without
  43  * filling the FS with empty chunks
  44  *
  45  * CHUNK_ALLOC_FORCE means it must try to allocate one
  46  *
  47  */
  48 enum {
  49         CHUNK_ALLOC_NO_FORCE = 0,
  50         CHUNK_ALLOC_LIMITED = 1,
  51         CHUNK_ALLOC_FORCE = 2,
  52 };
  53
  54 /*
  55  * Declare a helper function to detect underflow of various space info members
  56  */
  57 #define DECLARE_SPACE_INFO_UPDATE(name)                                 \
  58 static inline void update_##name(struct btrfs_space_info *sinfo,        \
  59                                  s64 bytes)                             \
  60 {                                                                       \
  61         if (bytes < 0 && sinfo->name < -bytes) {                        \
  62                 WARN_ON(1);                                             \
  63                 sinfo->name = 0;                                        \
  64                 return;                                                 \
  65         }                                                               \
  66         sinfo->name += bytes;                                           \
  67 }
  68
  69 DECLARE_SPACE_INFO_UPDATE(bytes_may_use);
  70 DECLARE_SPACE_INFO_UPDATE(bytes_pinned);
  71
  72 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
  73                                struct btrfs_delayed_ref_node *node, u64 parent,
  74                                u64 root_objectid, u64 owner_objectid,
  75                                u64 owner_offset, int refs_to_drop,
  76                                struct btrfs_delayed_extent_op *extra_op);
  77 static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op,
  78                                     struct extent_buffer *leaf,
  79                                     struct btrfs_extent_item *ei);
  80 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
  81                                       u64 parent, u64 root_objectid,
  82                                       u64 flags, u64 owner, u64 offset,
  83                                       struct btrfs_key *ins, int ref_mod);
  84 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
  85                                      struct btrfs_delayed_ref_node *node,
  86                                      struct btrfs_delayed_extent_op *extent_op);
  87 static int do_chunk_alloc(struct btrfs_trans_handle *trans, u64 flags,
  88                           int force);
  89 static int find_next_key(struct btrfs_path *path, int level,
  90                          struct btrfs_key *key);
  91 static void dump_space_info(struct btrfs_fs_info *fs_info,
  92                             struct btrfs_space_info *info, u64 bytes,
  93                             int dump_block_groups);
  94 static int block_rsv_use_bytes(struct btrfs_block_rsv *block_rsv,
  95                                u64 num_bytes);
  96 static void space_info_add_new_bytes(struct btrfs_fs_info *fs_info,
  97                                      struct btrfs_space_info *space_info,
  98                                      u64 num_bytes);
  99 static void space_info_add_old_bytes(struct btrfs_fs_info *fs_info,
 100                                      struct btrfs_space_info *space_info,
 101                                      u64 num_bytes);
 102
 103 static noinline int
 104 block_group_cache_done(struct btrfs_block_group_cache *cache)
 105 {
 106         smp_mb();
 107         return cache->cached == BTRFS_CACHE_FINISHED ||
 108                 cache->cached == BTRFS_CACHE_ERROR;
 109 }
 110
 111 static int block_group_bits(struct btrfs_block_group_cache *cache, u64 bits)
 112 {
 113         return (cache->flags & bits) == bits;
 114 }
 115
 116 void btrfs_get_block_group(struct btrfs_block_group_cache *cache)
 117 {
 118         atomic_inc(&cache->count);
 119 }
 120
 121 void btrfs_put_block_group(struct btrfs_block_group_cache *cache)
 122 {
 123         if (atomic_dec_and_test(&cache->count)) {
 124                 WARN_ON(cache->pinned > 0);
 125                 WARN_ON(cache->reserved > 0);
 126
 127                 /*
 128                  * If not empty, someone is still holding mutex of
 129                  * full_stripe_lock, which can only be released by caller.
 130                  * And it will definitely cause use-after-free when caller
 131                  * tries to release full stripe lock.
 132                  *
 133                  * No better way to resolve, but only to warn.
 134                  */
 135                 WARN_ON(!RB_EMPTY_ROOT(&cache->full_stripe_locks_root.root));
 136                 kfree(cache->free_space_ctl);
 137                 kfree(cache);
 138         }
 139 }
 140
 141 /*
 142  * this adds the block group to the fs_info rb tree for the block group
 143  * cache
 144  */
 145 static int btrfs_add_block_group_cache(struct btrfs_fs_info *info,
 146                                 struct btrfs_block_group_cache *block_group)
 147 {
 148         struct rb_node **p;
 149         struct rb_node *parent = NULL;
 150         struct btrfs_block_group_cache *cache;
 151
 152         spin_lock(&info->block_group_cache_lock);
 153         p = &info->block_group_cache_tree.rb_node;
 154
 155         while (*p) {
 156                 parent = *p;
 157                 cache = rb_entry(parent, struct btrfs_block_group_cache,
 158                                  cache_node);
 159                 if (block_group->key.objectid < cache->key.objectid) {
 160                         p = &(*p)->rb_left;
 161                 } else if (block_group->key.objectid > cache->key.objectid) {
 162                         p = &(*p)->rb_right;
 163                 } else {
 164                         spin_unlock(&info->block_group_cache_lock);
 165                         return -EEXIST;
 166                 }
 167         }
 168
 169         rb_link_node(&block_group->cache_node, parent, p);
 170         rb_insert_color(&block_group->cache_node,
 171                         &info->block_group_cache_tree);
 172
 173         if (info->first_logical_byte > block_group->key.objectid)
 174                 info->first_logical_byte = block_group->key.objectid;
 175
 176         spin_unlock(&info->block_group_cache_lock);
 177
 178         return 0;
 179 }
 180
 181 /*
 182  * This will return the block group at or after bytenr if contains is 0, else
 183  * it will return the block group that contains the bytenr
 184  */
 185 static struct btrfs_block_group_cache *
 186 block_group_cache_tree_search(struct btrfs_fs_info *info, u64 bytenr,
 187                               int contains)
 188 {
 189         struct btrfs_block_group_cache *cache, *ret = NULL;
 190         struct rb_node *n;
 191         u64 end, start;
 192
 193         spin_lock(&info->block_group_cache_lock);
 194         n = info->block_group_cache_tree.rb_node;
 195
 196         while (n) {
 197                 cache = rb_entry(n, struct btrfs_block_group_cache,
 198                                  cache_node);
 199                 end = cache->key.objectid + cache->key.offset - 1;
 200                 start = cache->key.objectid;
 201
 202                 if (bytenr < start) {
 203                         if (!contains && (!ret || start < ret->key.objectid))
 204                                 ret = cache;
 205                         n = n->rb_left;
 206                 } else if (bytenr > start) {
 207                         if (contains && bytenr <= end) {
 208                                 ret = cache;
 209                                 break;
 210                         }
 211                         n = n->rb_right;
 212                 } else {
 213                         ret = cache;
 214                         break;
 215                 }
 216         }
 217         if (ret) {
 218                 btrfs_get_block_group(ret);
 219                 if (bytenr == 0 && info->first_logical_byte > ret->key.objectid)
 220                         info->first_logical_byte = ret->key.objectid;
 221         }
 222         spin_unlock(&info->block_group_cache_lock);
 223
 224         return ret;
 225 }
 226
 227 static int add_excluded_extent(struct btrfs_fs_info *fs_info,
 228                                u64 start, u64 num_bytes)
 229 {
 230         u64 end = start + num_bytes - 1;
 231         set_extent_bits(&fs_info->freed_extents[0],
 232                         start, end, EXTENT_UPTODATE);
 233         set_extent_bits(&fs_info->freed_extents[1],
 234                         start, end, EXTENT_UPTODATE);
 235         return 0;
 236 }
 237
 238 static void free_excluded_extents(struct btrfs_block_group_cache *cache)
 239 {
 240         struct btrfs_fs_info *fs_info = cache->fs_info;
 241         u64 start, end;
 242
 243         start = cache->key.objectid;
 244         end = start + cache->key.offset - 1;
 245
 246         clear_extent_bits(&fs_info->freed_extents[0],
 247                           start, end, EXTENT_UPTODATE);
 248         clear_extent_bits(&fs_info->freed_extents[1],
 249                           start, end, EXTENT_UPTODATE);
 250 }
 251
 252 static int exclude_super_stripes(struct btrfs_block_group_cache *cache)
 253 {
 254         struct btrfs_fs_info *fs_info = cache->fs_info;
 255         u64 bytenr;
 256         u64 *logical;
 257         int stripe_len;
 258         int i, nr, ret;
 259
 260         if (cache->key.objectid < BTRFS_SUPER_INFO_OFFSET) {
 261                 stripe_len = BTRFS_SUPER_INFO_OFFSET - cache->key.objectid;
 262                 cache->bytes_super += stripe_len;
 263                 ret = add_excluded_extent(fs_info, cache->key.objectid,
 264                                           stripe_len);
 265                 if (ret)
 266                         return ret;
 267         }
 268
 269         for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
 270                 bytenr = btrfs_sb_offset(i);
 271                 ret = btrfs_rmap_block(fs_info, cache->key.objectid,
 272                                        bytenr, &logical, &nr, &stripe_len);
 273                 if (ret)
 274                         return ret;
 275
 276                 while (nr--) {
 277                         u64 start, len;
 278
 279                         if (logical[nr] > cache->key.objectid +
 280                             cache->key.offset)
 281                                 continue;
 282
 283                         if (logical[nr] + stripe_len <= cache->key.objectid)
 284                                 continue;
 285
 286                         start = logical[nr];
 287                         if (start < cache->key.objectid) {
 288                                 start = cache->key.objectid;
 289                                 len = (logical[nr] + stripe_len) - start;
 290                         } else {
 291                                 len = min_t(u64, stripe_len,
 292                                             cache->key.objectid +
 293                                             cache->key.offset - start);
 294                         }
 295
 296                         cache->bytes_super += len;
 297                         ret = add_excluded_extent(fs_info, start, len);
 298                         if (ret) {
 299                                 kfree(logical);
 300                                 return ret;
 301                         }
 302                 }
 303
 304                 kfree(logical);
 305         }
 306         return 0;
 307 }
 308
 309 static struct btrfs_caching_control *
 310 get_caching_control(struct btrfs_block_group_cache *cache)
 311 {
 312         struct btrfs_caching_control *ctl;
 313
 314         spin_lock(&cache->lock);
 315         if (!cache->caching_ctl) {
 316                 spin_unlock(&cache->lock);
 317                 return NULL;
 318         }
 319
 320         ctl = cache->caching_ctl;
 321         refcount_inc(&ctl->count);
 322         spin_unlock(&cache->lock);
 323         return ctl;
 324 }
 325
 326 static void put_caching_control(struct btrfs_caching_control *ctl)
 327 {
 328         if (refcount_dec_and_test(&ctl->count))
 329                 kfree(ctl);
 330 }
 331
 332 #ifdef CONFIG_BTRFS_DEBUG
 333 static void fragment_free_space(struct btrfs_block_group_cache *block_group)
 334 {
 335         struct btrfs_fs_info *fs_info = block_group->fs_info;
 336         u64 start = block_group->key.objectid;
 337         u64 len = block_group->key.offset;
 338         u64 chunk = block_group->flags & BTRFS_BLOCK_GROUP_METADATA ?
 339                 fs_info->nodesize : fs_info->sectorsize;
 340         u64 step = chunk << 1;
 341
 342         while (len > chunk) {
 343                 btrfs_remove_free_space(block_group, start, chunk);
 344                 start += step;
 345                 if (len < step)
 346                         len = 0;
 347                 else
 348                         len -= step;
 349         }
 350 }
 351 #endif
 352
 353 /*
 354  * this is only called by cache_block_group, since we could have freed extents
 355  * we need to check the pinned_extents for any extents that can't be used yet
 356  * since their free space will be released as soon as the transaction commits.
 357  */
 358 u64 add_new_free_space(struct btrfs_block_group_cache *block_group,
 359                        u64 start, u64 end)
 360 {
 361         struct btrfs_fs_info *info = block_group->fs_info;
 362         u64 extent_start, extent_end, size, total_added = 0;
 363         int ret;
 364
 365         while (start < end) {
 366                 ret = find_first_extent_bit(info->pinned_extents, start,
 367                                             &extent_start, &extent_end,
 368                                             EXTENT_DIRTY | EXTENT_UPTODATE,
 369                                             NULL);
 370                 if (ret)
 371                         break;
 372
 373                 if (extent_start <= start) {
 374                         start = extent_end + 1;
 375                 } else if (extent_start > start && extent_start < end) {
 376                         size = extent_start - start;
 377                         total_added += size;
 378                         ret = btrfs_add_free_space(block_group, start,
 379                                                    size);
 380                         BUG_ON(ret); /* -ENOMEM or logic error */
 381                         start = extent_end + 1;
 382                 } else {
 383                         break;
 384                 }
 385         }
 386
 387         if (start < end) {
 388                 size = end - start;
 389                 total_added += size;
 390                 ret = btrfs_add_free_space(block_group, start, size);
 391                 BUG_ON(ret); /* -ENOMEM or logic error */
 392         }
 393
 394         return total_added;
 395 }
 396
 397 static int load_extent_tree_free(struct btrfs_caching_control *caching_ctl)
 398 {
 399         struct btrfs_block_group_cache *block_group = caching_ctl->block_group;
 400         struct btrfs_fs_info *fs_info = block_group->fs_info;
 401         struct btrfs_root *extent_root = fs_info->extent_root;
 402         struct btrfs_path *path;
 403         struct extent_buffer *leaf;
 404         struct btrfs_key key;
 405         u64 total_found = 0;
 406         u64 last = 0;
 407         u32 nritems;
 408         int ret;
 409         bool wakeup = true;
 410
 411         path = btrfs_alloc_path();
 412         if (!path)
 413                 return -ENOMEM;
 414
 415         last = max_t(u64, block_group->key.objectid, BTRFS_SUPER_INFO_OFFSET);
 416
 417 #ifdef CONFIG_BTRFS_DEBUG
 418         /*
 419          * If we're fragmenting we don't want to make anybody think we can
 420          * allocate from this block group until we've had a chance to fragment
 421          * the free space.
 422          */
 423         if (btrfs_should_fragment_free_space(block_group))
 424                 wakeup = false;
 425 #endif
 426         /*
 427          * We don't want to deadlock with somebody trying to allocate a new
 428          * extent for the extent root while also trying to search the extent
 429          * root to add free space.  So we skip locking and search the commit
 430          * root, since its read-only
 431          */
 432         path->skip_locking = 1;
 433         path->search_commit_root = 1;
 434         path->reada = READA_FORWARD;
 435
 436         key.objectid = last;
 437         key.offset = 0;
 438         key.type = BTRFS_EXTENT_ITEM_KEY;
 439
 440 next:
 441         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
 442         if (ret < 0)
 443                 goto out;
 444
 445         leaf = path->nodes[0];
 446         nritems = btrfs_header_nritems(leaf);
 447
 448         while (1) {
 449                 if (btrfs_fs_closing(fs_info) > 1) {
 450                         last = (u64)-1;
 451                         break;
 452                 }
 453
 454                 if (path->slots[0] < nritems) {
 455                         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
 456                 } else {
 457                         ret = find_next_key(path, 0, &key);
 458                         if (ret)
 459                                 break;
 460
 461                         if (need_resched() ||
 462                             rwsem_is_contended(&fs_info->commit_root_sem)) {
 463                                 if (wakeup)
 464                                         caching_ctl->progress = last;
 465                                 btrfs_release_path(path);
 466                                 up_read(&fs_info->commit_root_sem);
 467                                 mutex_unlock(&caching_ctl->mutex);
 468                                 cond_resched();
 469                                 mutex_lock(&caching_ctl->mutex);
 470                                 down_read(&fs_info->commit_root_sem);
 471                                 goto next;
 472                         }
 473
 474                         ret = btrfs_next_leaf(extent_root, path);
 475                         if (ret < 0)
 476                                 goto out;
 477                         if (ret)
 478                                 break;
 479                         leaf = path->nodes[0];
 480                         nritems = btrfs_header_nritems(leaf);
 481                         continue;
 482                 }
 483
 484                 if (key.objectid < last) {
 485                         key.objectid = last;
 486                         key.offset = 0;
 487                         key.type = BTRFS_EXTENT_ITEM_KEY;
 488
 489                         if (wakeup)
 490                                 caching_ctl->progress = last;
 491                         btrfs_release_path(path);
 492                         goto next;
 493                 }
 494
 495                 if (key.objectid < block_group->key.objectid) {
 496                         path->slots[0]++;
 497                         continue;
 498                 }
 499
 500                 if (key.objectid >= block_group->key.objectid +
 501                     block_group->key.offset)
 502                         break;
 503
 504                 if (key.type == BTRFS_EXTENT_ITEM_KEY ||
 505                     key.type == BTRFS_METADATA_ITEM_KEY) {
 506                         total_found += add_new_free_space(block_group, last,
 507                                                           key.objectid);
 508                         if (key.type == BTRFS_METADATA_ITEM_KEY)
 509                                 last = key.objectid +
 510                                         fs_info->nodesize;
 511                         else
 512                                 last = key.objectid + key.offset;
 513
 514                         if (total_found > CACHING_CTL_WAKE_UP) {
 515                                 total_found = 0;
 516                                 if (wakeup)
 517                                         wake_up(&caching_ctl->wait);
 518                         }
 519                 }
 520                 path->slots[0]++;
 521         }
 522         ret = 0;
 523
 524         total_found += add_new_free_space(block_group, last,
 525                                           block_group->key.objectid +
 526                                           block_group->key.offset);
 527         caching_ctl->progress = (u64)-1;
 528
 529 out:
 530         btrfs_free_path(path);
 531         return ret;
 532 }
 533
 534 static noinline void caching_thread(struct btrfs_work *work)
 535 {
 536         struct btrfs_block_group_cache *block_group;
 537         struct btrfs_fs_info *fs_info;
 538         struct btrfs_caching_control *caching_ctl;
 539         int ret;
 540
 541         caching_ctl = container_of(work, struct btrfs_caching_control, work);
 542         block_group = caching_ctl->block_group;
 543         fs_info = block_group->fs_info;
 544
 545         mutex_lock(&caching_ctl->mutex);
 546         down_read(&fs_info->commit_root_sem);
 547
 548         if (btrfs_fs_compat_ro(fs_info, FREE_SPACE_TREE))
 549                 ret = load_free_space_tree(caching_ctl);
 550         else
 551                 ret = load_extent_tree_free(caching_ctl);
 552
 553         spin_lock(&block_group->lock);
 554         block_group->caching_ctl = NULL;
 555         block_group->cached = ret ? BTRFS_CACHE_ERROR : BTRFS_CACHE_FINISHED;
 556         spin_unlock(&block_group->lock);
 557
 558 #ifdef CONFIG_BTRFS_DEBUG
 559         if (btrfs_should_fragment_free_space(block_group)) {
 560                 u64 bytes_used;
 561
 562                 spin_lock(&block_group->space_info->lock);
 563                 spin_lock(&block_group->lock);
 564                 bytes_used = block_group->key.offset -
 565                         btrfs_block_group_used(&block_group->item);
 566                 block_group->space_info->bytes_used += bytes_used >> 1;
 567                 spin_unlock(&block_group->lock);
 568                 spin_unlock(&block_group->space_info->lock);
 569                 fragment_free_space(block_group);
 570         }
 571 #endif
 572
 573         caching_ctl->progress = (u64)-1;
 574
 575         up_read(&fs_info->commit_root_sem);
 576         free_excluded_extents(block_group);
 577         mutex_unlock(&caching_ctl->mutex);
 578
 579         wake_up(&caching_ctl->wait);
 580
 581         put_caching_control(caching_ctl);
 582         btrfs_put_block_group(block_group);
 583 }
 584
 585 static int cache_block_group(struct btrfs_block_group_cache *cache,
 586                              int load_cache_only)
 587 {
 588         DEFINE_WAIT(wait);
 589         struct btrfs_fs_info *fs_info = cache->fs_info;
 590         struct btrfs_caching_control *caching_ctl;
 591         int ret = 0;
 592
 593         caching_ctl = kzalloc(sizeof(*caching_ctl), GFP_NOFS);
 594         if (!caching_ctl)
 595                 return -ENOMEM;
 596
 597         INIT_LIST_HEAD(&caching_ctl->list);
 598         mutex_init(&caching_ctl->mutex);
 599         init_waitqueue_head(&caching_ctl->wait);
 600         caching_ctl->block_group = cache;
 601         caching_ctl->progress = cache->key.objectid;
 602         refcount_set(&caching_ctl->count, 1);
 603         btrfs_init_work(&caching_ctl->work, btrfs_cache_helper,
 604                         caching_thread, NULL, NULL);
 605
 606         spin_lock(&cache->lock);
 607         /*
 608          * This should be a rare occasion, but this could happen I think in the
 609          * case where one thread starts to load the space cache info, and then
 610          * some other thread starts a transaction commit which tries to do an
 611          * allocation while the other thread is still loading the space cache
 612          * info.  The previous loop should have kept us from choosing this block
 613          * group, but if we've moved to the state where we will wait on caching
 614          * block groups we need to first check if we're doing a fast load here,
 615          * so we can wait for it to finish, otherwise we could end up allocating
 616          * from a block group who's cache gets evicted for one reason or
 617          * another.
 618          */
 619         while (cache->cached == BTRFS_CACHE_FAST) {
 620                 struct btrfs_caching_control *ctl;
 621
 622                 ctl = cache->caching_ctl;
 623                 refcount_inc(&ctl->count);
 624                 prepare_to_wait(&ctl->wait, &wait, TASK_UNINTERRUPTIBLE);
 625                 spin_unlock(&cache->lock);
 626
 627                 schedule();
 628
 629                 finish_wait(&ctl->wait, &wait);
 630                 put_caching_control(ctl);
 631                 spin_lock(&cache->lock);
 632         }
 633
 634         if (cache->cached != BTRFS_CACHE_NO) {
 635                 spin_unlock(&cache->lock);
 636                 kfree(caching_ctl);
 637                 return 0;
 638         }
 639         WARN_ON(cache->caching_ctl);
 640         cache->caching_ctl = caching_ctl;
 641         cache->cached = BTRFS_CACHE_FAST;
 642         spin_unlock(&cache->lock);
 643
 644         if (btrfs_test_opt(fs_info, SPACE_CACHE)) {
 645                 mutex_lock(&caching_ctl->mutex);
 646                 ret = load_free_space_cache(fs_info, cache);
 647
 648                 spin_lock(&cache->lock);
 649                 if (ret == 1) {
 650                         cache->caching_ctl = NULL;
 651                         cache->cached = BTRFS_CACHE_FINISHED;
 652                         cache->last_byte_to_unpin = (u64)-1;
 653                         caching_ctl->progress = (u64)-1;
 654                 } else {
 655                         if (load_cache_only) {
 656                                 cache->caching_ctl = NULL;
 657                                 cache->cached = BTRFS_CACHE_NO;
 658                         } else {
 659                                 cache->cached = BTRFS_CACHE_STARTED;
 660                                 cache->has_caching_ctl = 1;
 661                         }
 662                 }
 663                 spin_unlock(&cache->lock);
 664 #ifdef CONFIG_BTRFS_DEBUG
 665                 if (ret == 1 &&
 666                     btrfs_should_fragment_free_space(cache)) {
 667                         u64 bytes_used;
 668
 669                         spin_lock(&cache->space_info->lock);
 670                         spin_lock(&cache->lock);
 671                         bytes_used = cache->key.offset -
 672                                 btrfs_block_group_used(&cache->item);
 673                         cache->space_info->bytes_used += bytes_used >> 1;
 674                         spin_unlock(&cache->lock);
 675                         spin_unlock(&cache->space_info->lock);
 676                         fragment_free_space(cache);
 677                 }
 678 #endif
 679                 mutex_unlock(&caching_ctl->mutex);
 680
 681                 wake_up(&caching_ctl->wait);
 682                 if (ret == 1) {
 683                         put_caching_control(caching_ctl);
 684                         free_excluded_extents(cache);
 685                         return 0;
 686                 }
 687         } else {
 688                 /*
 689                  * We're either using the free space tree or no caching at all.
 690                  * Set cached to the appropriate value and wakeup any waiters.
 691                  */
 692                 spin_lock(&cache->lock);
 693                 if (load_cache_only) {
 694                         cache->caching_ctl = NULL;
 695                         cache->cached = BTRFS_CACHE_NO;
 696                 } else {
 697                         cache->cached = BTRFS_CACHE_STARTED;
 698                         cache->has_caching_ctl = 1;
 699                 }
 700                 spin_unlock(&cache->lock);
 701                 wake_up(&caching_ctl->wait);
 702         }
 703
 704         if (load_cache_only) {
 705                 put_caching_control(caching_ctl);
 706                 return 0;
 707         }
 708
 709         down_write(&fs_info->commit_root_sem);
 710         refcount_inc(&caching_ctl->count);
 711         list_add_tail(&caching_ctl->list, &fs_info->caching_block_groups);
 712         up_write(&fs_info->commit_root_sem);
 713
 714         btrfs_get_block_group(cache);
 715
 716         btrfs_queue_work(fs_info->caching_workers, &caching_ctl->work);
 717
 718         return ret;
 719 }
 720
 721 /*
 722  * return the block group that starts at or after bytenr
 723  */
 724 static struct btrfs_block_group_cache *
 725 btrfs_lookup_first_block_group(struct btrfs_fs_info *info, u64 bytenr)
 726 {
 727         return block_group_cache_tree_search(info, bytenr, 0);
 728 }
 729
 730 /*
 731  * return the block group that contains the given bytenr
 732  */
 733 struct btrfs_block_group_cache *btrfs_lookup_block_group(
 734                                                  struct btrfs_fs_info *info,
 735                                                  u64 bytenr)
 736 {
 737         return block_group_cache_tree_search(info, bytenr, 1);
 738 }
 739
 740 static struct btrfs_space_info *__find_space_info(struct btrfs_fs_info *info,
 741                                                   u64 flags)
 742 {
 743         struct list_head *head = &info->space_info;
 744         struct btrfs_space_info *found;
 745
 746         flags &= BTRFS_BLOCK_GROUP_TYPE_MASK;
 747
 748         rcu_read_lock();
 749         list_for_each_entry_rcu(found, head, list) {
 750                 if (found->flags & flags) {
 751                         rcu_read_unlock();
 752                         return found;
 753                 }
 754         }
 755         rcu_read_unlock();
 756         return NULL;
 757 }
 758
 759 static void add_pinned_bytes(struct btrfs_fs_info *fs_info, s64 num_bytes,
 760                              bool metadata, u64 root_objectid)
 761 {
 762         struct btrfs_space_info *space_info;
 763         u64 flags;
 764
 765         if (metadata) {
 766                 if (root_objectid == BTRFS_CHUNK_TREE_OBJECTID)
 767                         flags = BTRFS_BLOCK_GROUP_SYSTEM;
 768                 else
 769                         flags = BTRFS_BLOCK_GROUP_METADATA;
 770         } else {
 771                 flags = BTRFS_BLOCK_GROUP_DATA;
 772         }
 773
 774         space_info = __find_space_info(fs_info, flags);
 775         ASSERT(space_info);
 776         percpu_counter_add_batch(&space_info->total_bytes_pinned, num_bytes,
 777                     BTRFS_TOTAL_BYTES_PINNED_BATCH);
 778 }
 779
 780 /*
 781  * after adding space to the filesystem, we need to clear the full flags
 782  * on all the space infos.
 783  */
 784 void btrfs_clear_space_info_full(struct btrfs_fs_info *info)
 785 {
 786         struct list_head *head = &info->space_info;
 787         struct btrfs_space_info *found;
 788
 789         rcu_read_lock();
 790         list_for_each_entry_rcu(found, head, list)
 791                 found->full = 0;
 792         rcu_read_unlock();
 793 }
 794
 795 /* simple helper to search for an existing data extent at a given offset */
 796 int btrfs_lookup_data_extent(struct btrfs_fs_info *fs_info, u64 start, u64 len)
 797 {
 798         int ret;
 799         struct btrfs_key key;
 800         struct btrfs_path *path;
 801
 802         path = btrfs_alloc_path();
 803         if (!path)
 804                 return -ENOMEM;
 805
 806         key.objectid = start;
 807         key.offset = len;
 808         key.type = BTRFS_EXTENT_ITEM_KEY;
 809         ret = btrfs_search_slot(NULL, fs_info->extent_root, &key, path, 0, 0);
 810         btrfs_free_path(path);
 811         return ret;
 812 }
 813
 814 /*
 815  * helper function to lookup reference count and flags of a tree block.
 816  *
 817  * the head node for delayed ref is used to store the sum of all the
 818  * reference count modifications queued up in the rbtree. the head
 819  * node may also store the extent flags to set. This way you can check
 820  * to see what the reference count and extent flags would be if all of
 821  * the delayed refs are not processed.
 822  */
 823 int btrfs_lookup_extent_info(struct btrfs_trans_handle *trans,
 824                              struct btrfs_fs_info *fs_info, u64 bytenr,
 825                              u64 offset, int metadata, u64 *refs, u64 *flags)
 826 {
 827         struct btrfs_delayed_ref_head *head;
 828         struct btrfs_delayed_ref_root *delayed_refs;
 829         struct btrfs_path *path;
 830         struct btrfs_extent_item *ei;
 831         struct extent_buffer *leaf;
 832         struct btrfs_key key;
 833         u32 item_size;
 834         u64 num_refs;
 835         u64 extent_flags;
 836         int ret;
 837
 838         /*
 839          * If we don't have skinny metadata, don't bother doing anything
 840          * different
 841          */
 842         if (metadata && !btrfs_fs_incompat(fs_info, SKINNY_METADATA)) {
 843                 offset = fs_info->nodesize;
 844                 metadata = 0;
 845         }
 846
 847         path = btrfs_alloc_path();
 848         if (!path)
 849                 return -ENOMEM;
 850
 851         if (!trans) {
 852                 path->skip_locking = 1;
 853                 path->search_commit_root = 1;
 854         }
 855
 856 search_again:
 857         key.objectid = bytenr;
 858         key.offset = offset;
 859         if (metadata)
 860                 key.type = BTRFS_METADATA_ITEM_KEY;
 861         else
 862                 key.type = BTRFS_EXTENT_ITEM_KEY;
 863
 864         ret = btrfs_search_slot(trans, fs_info->extent_root, &key, path, 0, 0);
 865         if (ret < 0)
 866                 goto out_free;
 867
 868         if (ret > 0 && metadata && key.type == BTRFS_METADATA_ITEM_KEY) {
 869                 if (path->slots[0]) {
 870                         path->slots[0]--;
 871                         btrfs_item_key_to_cpu(path->nodes[0], &key,
 872                                               path->slots[0]);
 873                         if (key.objectid == bytenr &&
 874                             key.type == BTRFS_EXTENT_ITEM_KEY &&
 875                             key.offset == fs_info->nodesize)
 876                                 ret = 0;
 877                 }
 878         }
 879
 880         if (ret == 0) {
 881                 leaf = path->nodes[0];
 882                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
 883                 if (item_size >= sizeof(*ei)) {
 884                         ei = btrfs_item_ptr(leaf, path->slots[0],
 885                                             struct btrfs_extent_item);
 886                         num_refs = btrfs_extent_refs(leaf, ei);
 887                         extent_flags = btrfs_extent_flags(leaf, ei);
 888                 } else {
 889                         ret = -EINVAL;
 890                         btrfs_print_v0_err(fs_info);
 891                         if (trans)
 892                                 btrfs_abort_transaction(trans, ret);
 893                         else
 894                                 btrfs_handle_fs_error(fs_info, ret, NULL);
 895
 896                         goto out_free;
 897                 }
 898
 899                 BUG_ON(num_refs == 0);
 900         } else {
 901                 num_refs = 0;
 902                 extent_flags = 0;
 903                 ret = 0;
 904         }
 905
 906         if (!trans)
 907                 goto out;
 908
 909         delayed_refs = &trans->transaction->delayed_refs;
 910         spin_lock(&delayed_refs->lock);
 911         head = btrfs_find_delayed_ref_head(delayed_refs, bytenr);
 912         if (head) {
 913                 if (!mutex_trylock(&head->mutex)) {
 914                         refcount_inc(&head->refs);
 915                         spin_unlock(&delayed_refs->lock);
 916
 917                         btrfs_release_path(path);
 918
 919                         /*
 920                          * Mutex was contended, block until it's released and try
 921                          * again
 922                          */
 923                         mutex_lock(&head->mutex);
 924                         mutex_unlock(&head->mutex);
 925                         btrfs_put_delayed_ref_head(head);
 926                         goto search_again;
 927                 }
 928                 spin_lock(&head->lock);
 929                 if (head->extent_op && head->extent_op->update_flags)
 930                         extent_flags |= head->extent_op->flags_to_set;
 931                 else
 932                         BUG_ON(num_refs == 0);
 933
 934                 num_refs += head->ref_mod;
 935                 spin_unlock(&head->lock);
 936                 mutex_unlock(&head->mutex);
 937         }
 938         spin_unlock(&delayed_refs->lock);
 939 out:
 940         WARN_ON(num_refs == 0);
 941         if (refs)
 942                 *refs = num_refs;
 943         if (flags)
 944                 *flags = extent_flags;
 945 out_free:
 946         btrfs_free_path(path);
 947         return ret;
 948 }
 949
 950 /*
 951  * Back reference rules.  Back refs have three main goals:
 952  *
 953  * 1) differentiate between all holders of references to an extent so that
 954  *    when a reference is dropped we can make sure it was a valid reference
 955  *    before freeing the extent.
 956  *
 957  * 2) Provide enough information to quickly find the holders of an extent
 958  *    if we notice a given block is corrupted or bad.
 959  *
 960  * 3) Make it easy to migrate blocks for FS shrinking or storage pool
 961  *    maintenance.  This is actually the same as #2, but with a slightly
 962  *    different use case.
 963  *
 964  * There are two kinds of back refs. The implicit back refs is optimized
 965  * for pointers in non-shared tree blocks. For a given pointer in a block,
 966  * back refs of this kind provide information about the block's owner tree
 967  * and the pointer's key. These information allow us to find the block by
 968  * b-tree searching. The full back refs is for pointers in tree blocks not
 969  * referenced by their owner trees. The location of tree block is recorded
 970  * in the back refs. Actually the full back refs is generic, and can be
 971  * used in all cases the implicit back refs is used. The major shortcoming
 972  * of the full back refs is its overhead. Every time a tree block gets
 973  * COWed, we have to update back refs entry for all pointers in it.
 974  *
 975  * For a newly allocated tree block, we use implicit back refs for
 976  * pointers in it. This means most tree related operations only involve
 977  * implicit back refs. For a tree block created in old transaction, the
 978  * only way to drop a reference to it is COW it. So we can detect the
 979  * event that tree block loses its owner tree's reference and do the
 980  * back refs conversion.
 981  *
 982  * When a tree block is COWed through a tree, there are four cases:
 983  *
 984  * The reference count of the block is one and the tree is the block's
 985  * owner tree. Nothing to do in this case.
 986  *
 987  * The reference count of the block is one and the tree is not the
 988  * block's owner tree. In this case, full back refs is used for pointers
 989  * in the block. Remove these full back refs, add implicit back refs for
 990  * every pointers in the new block.
 991  *
 992  * The reference count of the block is greater than one and the tree is
 993  * the block's owner tree. In this case, implicit back refs is used for
 994  * pointers in the block. Add full back refs for every pointers in the
 995  * block, increase lower level extents' reference counts. The original
 996  * implicit back refs are entailed to the new block.
 997  *
 998  * The reference count of the block is greater than one and the tree is
 999  * not the block's owner tree. Add implicit back refs for every pointer in
1000  * the new block, increase lower level extents' reference count.
1001  *
1002  * Back Reference Key composing:
1003  *
1004  * The key objectid corresponds to the first byte in the extent,
1005  * The key type is used to differentiate between types of back refs.
1006  * There are different meanings of the key offset for different types
1007  * of back refs.
1008  *
1009  * File extents can be referenced by:
1010  *
1011  * - multiple snapshots, subvolumes, or different generations in one subvol
1012  * - different files inside a single subvolume
1013  * - different offsets inside a file (bookend extents in file.c)
1014  *
1015  * The extent ref structure for the implicit back refs has fields for:
1016  *
1017  * - Objectid of the subvolume root
1018  * - objectid of the file holding the reference
1019  * - original offset in the file
1020  * - how many bookend extents
1021  *
1022  * The key offset for the implicit back refs is hash of the first
1023  * three fields.
1024  *
1025  * The extent ref structure for the full back refs has field for:
1026  *
1027  * - number of pointers in the tree leaf
1028  *
1029  * The key offset for the implicit back refs is the first byte of
1030  * the tree leaf
1031  *
1032  * When a file extent is allocated, The implicit back refs is used.
1033  * the fields are filled in:
1034  *
1035  *     (root_key.objectid, inode objectid, offset in file, 1)
1036  *
1037  * When a file extent is removed file truncation, we find the
1038  * corresponding implicit back refs and check the following fields:
1039  *
1040  *     (btrfs_header_owner(leaf), inode objectid, offset in file)
1041  *
1042  * Btree extents can be referenced by:
1043  *
1044  * - Different subvolumes
1045  *
1046  * Both the implicit back refs and the full back refs for tree blocks
1047  * only consist of key. The key offset for the implicit back refs is
1048  * objectid of block's owner tree. The key offset for the full back refs
1049  * is the first byte of parent block.
1050  *
1051  * When implicit back refs is used, information about the lowest key and
1052  * level of the tree block are required. These information are stored in
1053  * tree block info structure.
1054  */
1055
1056 /*
1057  * is_data == BTRFS_REF_TYPE_BLOCK, tree block type is required,
1058  * is_data == BTRFS_REF_TYPE_DATA, data type is requiried,
1059  * is_data == BTRFS_REF_TYPE_ANY, either type is OK.
1060  */
1061 int btrfs_get_extent_inline_ref_type(const struct extent_buffer *eb,
1062                                      struct btrfs_extent_inline_ref *iref,
1063                                      enum btrfs_inline_ref_type is_data)
1064 {
1065         int type = btrfs_extent_inline_ref_type(eb, iref);
1066         u64 offset = btrfs_extent_inline_ref_offset(eb, iref);
1067
1068         if (type == BTRFS_TREE_BLOCK_REF_KEY ||
1069             type == BTRFS_SHARED_BLOCK_REF_KEY ||
1070             type == BTRFS_SHARED_DATA_REF_KEY ||
1071             type == BTRFS_EXTENT_DATA_REF_KEY) {
1072                 if (is_data == BTRFS_REF_TYPE_BLOCK) {
1073                         if (type == BTRFS_TREE_BLOCK_REF_KEY)
1074                                 return type;
1075                         if (type == BTRFS_SHARED_BLOCK_REF_KEY) {
1076                                 ASSERT(eb->fs_info);
1077                                 /*
1078                                  * Every shared one has parent tree
1079                                  * block, which must be aligned to
1080                                  * nodesize.
1081                                  */
1082                                 if (offset &&
1083                                     IS_ALIGNED(offset, eb->fs_info->nodesize))
1084                                         return type;
1085                         }
1086                 } else if (is_data == BTRFS_REF_TYPE_DATA) {
1087                         if (type == BTRFS_EXTENT_DATA_REF_KEY)
1088                                 return type;
1089                         if (type == BTRFS_SHARED_DATA_REF_KEY) {
1090                                 ASSERT(eb->fs_info);
1091                                 /*
1092                                  * Every shared one has parent tree
1093                                  * block, which must be aligned to
1094                                  * nodesize.
1095                                  */
1096                                 if (offset &&
1097                                     IS_ALIGNED(offset, eb->fs_info->nodesize))
1098                                         return type;
1099                         }
1100                 } else {
1101                         ASSERT(is_data == BTRFS_REF_TYPE_ANY);
1102                         return type;
1103                 }
1104         }
1105
1106         btrfs_print_leaf((struct extent_buffer *)eb);
1107         btrfs_err(eb->fs_info, "eb %llu invalid extent inline ref type %d",
1108                   eb->start, type);
1109         WARN_ON(1);
1110
1111         return BTRFS_REF_TYPE_INVALID;
1112 }
1113
1114 static u64 hash_extent_data_ref(u64 root_objectid, u64 owner, u64 offset)
1115 {
1116         u32 high_crc = ~(u32)0;
1117         u32 low_crc = ~(u32)0;
1118         __le64 lenum;
1119
1120         lenum = cpu_to_le64(root_objectid);
1121         high_crc = crc32c(high_crc, &lenum, sizeof(lenum));
1122         lenum = cpu_to_le64(owner);
1123         low_crc = crc32c(low_crc, &lenum, sizeof(lenum));
1124         lenum = cpu_to_le64(offset);
1125         low_crc = crc32c(low_crc, &lenum, sizeof(lenum));
1126
1127         return ((u64)high_crc << 31) ^ (u64)low_crc;
1128 }
1129
1130 static u64 hash_extent_data_ref_item(struct extent_buffer *leaf,
1131                                      struct btrfs_extent_data_ref *ref)
1132 {
1133         return hash_extent_data_ref(btrfs_extent_data_ref_root(leaf, ref),
1134                                     btrfs_extent_data_ref_objectid(leaf, ref),
1135                                     btrfs_extent_data_ref_offset(leaf, ref));
1136 }
1137
1138 static int match_extent_data_ref(struct extent_buffer *leaf,
1139                                  struct btrfs_extent_data_ref *ref,
1140                                  u64 root_objectid, u64 owner, u64 offset)
1141 {
1142         if (btrfs_extent_data_ref_root(leaf, ref) != root_objectid ||
1143             btrfs_extent_data_ref_objectid(leaf, ref) != owner ||
1144             btrfs_extent_data_ref_offset(leaf, ref) != offset)
1145                 return 0;
1146         return 1;
1147 }
1148
1149 static noinline int lookup_extent_data_ref(struct btrfs_trans_handle *trans,
1150                                            struct btrfs_path *path,
1151                                            u64 bytenr, u64 parent,
1152                                            u64 root_objectid,
1153                                            u64 owner, u64 offset)
1154 {
1155         struct btrfs_root *root = trans->fs_info->extent_root;
1156         struct btrfs_key key;
1157         struct btrfs_extent_data_ref *ref;
1158         struct extent_buffer *leaf;
1159         u32 nritems;
1160         int ret;
1161         int recow;
1162         int err = -ENOENT;
1163
1164         key.objectid = bytenr;
1165         if (parent) {
1166                 key.type = BTRFS_SHARED_DATA_REF_KEY;
1167                 key.offset = parent;
1168         } else {
1169                 key.type = BTRFS_EXTENT_DATA_REF_KEY;
1170                 key.offset = hash_extent_data_ref(root_objectid,
1171                                                   owner, offset);
1172         }
1173 again:
1174         recow = 0;
1175         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1176         if (ret < 0) {
1177                 err = ret;
1178                 goto fail;
1179         }
1180
1181         if (parent) {
1182                 if (!ret)
1183                         return 0;
1184                 goto fail;
1185         }
1186
1187         leaf = path->nodes[0];
1188         nritems = btrfs_header_nritems(leaf);
1189         while (1) {
1190                 if (path->slots[0] >= nritems) {
1191                         ret = btrfs_next_leaf(root, path);
1192                         if (ret < 0)
1193                                 err = ret;
1194                         if (ret)
1195                                 goto fail;
1196
1197                         leaf = path->nodes[0];
1198                         nritems = btrfs_header_nritems(leaf);
1199                         recow = 1;
1200                 }
1201
1202                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1203                 if (key.objectid != bytenr ||
1204                     key.type != BTRFS_EXTENT_DATA_REF_KEY)
1205                         goto fail;
1206
1207                 ref = btrfs_item_ptr(leaf, path->slots[0],
1208                                      struct btrfs_extent_data_ref);
1209
1210                 if (match_extent_data_ref(leaf, ref, root_objectid,
1211                                           owner, offset)) {
1212                         if (recow) {
1213                                 btrfs_release_path(path);
1214                                 goto again;
1215                         }
1216                         err = 0;
1217                         break;
1218                 }
1219                 path->slots[0]++;
1220         }
1221 fail:
1222         return err;
1223 }
1224
1225 static noinline int insert_extent_data_ref(struct btrfs_trans_handle *trans,
1226                                            struct btrfs_path *path,
1227                                            u64 bytenr, u64 parent,
1228                                            u64 root_objectid, u64 owner,
1229                                            u64 offset, int refs_to_add)
1230 {
1231         struct btrfs_root *root = trans->fs_info->extent_root;
1232         struct btrfs_key key;
1233         struct extent_buffer *leaf;
1234         u32 size;
1235         u32 num_refs;
1236         int ret;
1237
1238         key.objectid = bytenr;
1239         if (parent) {
1240                 key.type = BTRFS_SHARED_DATA_REF_KEY;
1241                 key.offset = parent;
1242                 size = sizeof(struct btrfs_shared_data_ref);
1243         } else {
1244                 key.type = BTRFS_EXTENT_DATA_REF_KEY;
1245                 key.offset = hash_extent_data_ref(root_objectid,
1246                                                   owner, offset);
1247                 size = sizeof(struct btrfs_extent_data_ref);
1248         }
1249
1250         ret = btrfs_insert_empty_item(trans, root, path, &key, size);
1251         if (ret && ret != -EEXIST)
1252                 goto fail;
1253
1254         leaf = path->nodes[0];
1255         if (parent) {
1256                 struct btrfs_shared_data_ref *ref;
1257                 ref = btrfs_item_ptr(leaf, path->slots[0],
1258                                      struct btrfs_shared_data_ref);
1259                 if (ret == 0) {
1260                         btrfs_set_shared_data_ref_count(leaf, ref, refs_to_add);
1261                 } else {
1262                         num_refs = btrfs_shared_data_ref_count(leaf, ref);
1263                         num_refs += refs_to_add;
1264                         btrfs_set_shared_data_ref_count(leaf, ref, num_refs);
1265                 }
1266         } else {
1267                 struct btrfs_extent_data_ref *ref;
1268                 while (ret == -EEXIST) {
1269                         ref = btrfs_item_ptr(leaf, path->slots[0],
1270                                              struct btrfs_extent_data_ref);
1271                         if (match_extent_data_ref(leaf, ref, root_objectid,
1272                                                   owner, offset))
1273                                 break;
1274                         btrfs_release_path(path);
1275                         key.offset++;
1276                         ret = btrfs_insert_empty_item(trans, root, path, &key,
1277                                                       size);
1278                         if (ret && ret != -EEXIST)
1279                                 goto fail;
1280
1281                         leaf = path->nodes[0];
1282                 }
1283                 ref = btrfs_item_ptr(leaf, path->slots[0],
1284                                      struct btrfs_extent_data_ref);
1285                 if (ret == 0) {
1286                         btrfs_set_extent_data_ref_root(leaf, ref,
1287                                                        root_objectid);
1288                         btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
1289                         btrfs_set_extent_data_ref_offset(leaf, ref, offset);
1290                         btrfs_set_extent_data_ref_count(leaf, ref, refs_to_add);
1291                 } else {
1292                         num_refs = btrfs_extent_data_ref_count(leaf, ref);
1293                         num_refs += refs_to_add;
1294                         btrfs_set_extent_data_ref_count(leaf, ref, num_refs);
1295                 }
1296         }
1297         btrfs_mark_buffer_dirty(leaf);
1298         ret = 0;
1299 fail:
1300         btrfs_release_path(path);
1301         return ret;
1302 }
1303
1304 static noinline int remove_extent_data_ref(struct btrfs_trans_handle *trans,
1305                                            struct btrfs_path *path,
1306                                            int refs_to_drop, int *last_ref)
1307 {
1308         struct btrfs_key key;
1309         struct btrfs_extent_data_ref *ref1 = NULL;
1310         struct btrfs_shared_data_ref *ref2 = NULL;
1311         struct extent_buffer *leaf;
1312         u32 num_refs = 0;
1313         int ret = 0;
1314
1315         leaf = path->nodes[0];
1316         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1317
1318         if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
1319                 ref1 = btrfs_item_ptr(leaf, path->slots[0],
1320                                       struct btrfs_extent_data_ref);
1321                 num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1322         } else if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
1323                 ref2 = btrfs_item_ptr(leaf, path->slots[0],
1324                                       struct btrfs_shared_data_ref);
1325                 num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1326         } else if (unlikely(key.type == BTRFS_EXTENT_REF_V0_KEY)) {
1327                 btrfs_print_v0_err(trans->fs_info);
1328                 btrfs_abort_transaction(trans, -EINVAL);
1329                 return -EINVAL;
1330         } else {
1331                 BUG();
1332         }
1333
1334         BUG_ON(num_refs < refs_to_drop);
1335         num_refs -= refs_to_drop;
1336
1337         if (num_refs == 0) {
1338                 ret = btrfs_del_item(trans, trans->fs_info->extent_root, path);
1339                 *last_ref = 1;
1340         } else {
1341                 if (key.type == BTRFS_EXTENT_DATA_REF_KEY)
1342                         btrfs_set_extent_data_ref_count(leaf, ref1, num_refs);
1343                 else if (key.type == BTRFS_SHARED_DATA_REF_KEY)
1344                         btrfs_set_shared_data_ref_count(leaf, ref2, num_refs);
1345                 btrfs_mark_buffer_dirty(leaf);
1346         }
1347         return ret;
1348 }
1349
1350 static noinline u32 extent_data_ref_count(struct btrfs_path *path,
1351                                           struct btrfs_extent_inline_ref *iref)
1352 {
1353         struct btrfs_key key;
1354         struct extent_buffer *leaf;
1355         struct btrfs_extent_data_ref *ref1;
1356         struct btrfs_shared_data_ref *ref2;
1357         u32 num_refs = 0;
1358         int type;
1359
1360         leaf = path->nodes[0];
1361         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1362
1363         BUG_ON(key.type == BTRFS_EXTENT_REF_V0_KEY);
1364         if (iref) {
1365                 /*
1366                  * If type is invalid, we should have bailed out earlier than
1367                  * this call.
1368                  */
1369                 type = btrfs_get_extent_inline_ref_type(leaf, iref, BTRFS_REF_TYPE_DATA);
1370                 ASSERT(type != BTRFS_REF_TYPE_INVALID);
1371                 if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1372                         ref1 = (struct btrfs_extent_data_ref *)(&iref->offset);
1373                         num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1374                 } else {
1375                         ref2 = (struct btrfs_shared_data_ref *)(iref + 1);
1376                         num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1377                 }
1378         } else if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
1379                 ref1 = btrfs_item_ptr(leaf, path->slots[0],
1380                                       struct btrfs_extent_data_ref);
1381                 num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1382         } else if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
1383                 ref2 = btrfs_item_ptr(leaf, path->slots[0],
1384                                       struct btrfs_shared_data_ref);
1385                 num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1386         } else {
1387                 WARN_ON(1);
1388         }
1389         return num_refs;
1390 }
1391
1392 static noinline int lookup_tree_block_ref(struct btrfs_trans_handle *trans,
1393                                           struct btrfs_path *path,
1394                                           u64 bytenr, u64 parent,
1395                                           u64 root_objectid)
1396 {
1397         struct btrfs_root *root = trans->fs_info->extent_root;
1398         struct btrfs_key key;
1399         int ret;
1400
1401         key.objectid = bytenr;
1402         if (parent) {
1403                 key.type = BTRFS_SHARED_BLOCK_REF_KEY;
1404                 key.offset = parent;
1405         } else {
1406                 key.type = BTRFS_TREE_BLOCK_REF_KEY;
1407                 key.offset = root_objectid;
1408         }
1409
1410         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1411         if (ret > 0)
1412                 ret = -ENOENT;
1413         return ret;
1414 }
1415
1416 static noinline int insert_tree_block_ref(struct btrfs_trans_handle *trans,
1417                                           struct btrfs_path *path,
1418                                           u64 bytenr, u64 parent,
1419                                           u64 root_objectid)
1420 {
1421         struct btrfs_key key;
1422         int ret;
1423
1424         key.objectid = bytenr;
1425         if (parent) {
1426                 key.type = BTRFS_SHARED_BLOCK_REF_KEY;
1427                 key.offset = parent;
1428         } else {
1429                 key.type = BTRFS_TREE_BLOCK_REF_KEY;
1430                 key.offset = root_objectid;
1431         }
1432
1433         ret = btrfs_insert_empty_item(trans, trans->fs_info->extent_root,
1434                                       path, &key, 0);
1435         btrfs_release_path(path);
1436         return ret;
1437 }
1438
1439 static inline int extent_ref_type(u64 parent, u64 owner)
1440 {
1441         int type;
1442         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1443                 if (parent > 0)
1444                         type = BTRFS_SHARED_BLOCK_REF_KEY;
1445                 else
1446                         type = BTRFS_TREE_BLOCK_REF_KEY;
1447         } else {
1448                 if (parent > 0)
1449                         type = BTRFS_SHARED_DATA_REF_KEY;
1450                 else
1451                         type = BTRFS_EXTENT_DATA_REF_KEY;
1452         }
1453         return type;
1454 }
1455
1456 static int find_next_key(struct btrfs_path *path, int level,
1457                          struct btrfs_key *key)
1458
1459 {
1460         for (; level < BTRFS_MAX_LEVEL; level++) {
1461                 if (!path->nodes[level])
1462                         break;
1463                 if (path->slots[level] + 1 >=
1464                     btrfs_header_nritems(path->nodes[level]))
1465                         continue;
1466                 if (level == 0)
1467                         btrfs_item_key_to_cpu(path->nodes[level], key,
1468                                               path->slots[level] + 1);
1469                 else
1470                         btrfs_node_key_to_cpu(path->nodes[level], key,
1471                                               path->slots[level] + 1);
1472                 return 0;
1473         }
1474         return 1;
1475 }
1476
1477 /*
1478  * look for inline back ref. if back ref is found, *ref_ret is set
1479  * to the address of inline back ref, and 0 is returned.
1480  *
1481  * if back ref isn't found, *ref_ret is set to the address where it
1482  * should be inserted, and -ENOENT is returned.
1483  *
1484  * if insert is true and there are too many inline back refs, the path
1485  * points to the extent item, and -EAGAIN is returned.
1486  *
1487  * NOTE: inline back refs are ordered in the same way that back ref
1488  *       items in the tree are ordered.
1489  */
1490 static noinline_for_stack
1491 int lookup_inline_extent_backref(struct btrfs_trans_handle *trans,
1492                                  struct btrfs_path *path,
1493                                  struct btrfs_extent_inline_ref **ref_ret,
1494                                  u64 bytenr, u64 num_bytes,
1495                                  u64 parent, u64 root_objectid,
1496                                  u64 owner, u64 offset, int insert)
1497 {
1498         struct btrfs_fs_info *fs_info = trans->fs_info;
1499         struct btrfs_root *root = fs_info->extent_root;
1500         struct btrfs_key key;
1501         struct extent_buffer *leaf;
1502         struct btrfs_extent_item *ei;
1503         struct btrfs_extent_inline_ref *iref;
1504         u64 flags;
1505         u64 item_size;
1506         unsigned long ptr;
1507         unsigned long end;
1508         int extra_size;
1509         int type;
1510         int want;
1511         int ret;
1512         int err = 0;
1513         bool skinny_metadata = btrfs_fs_incompat(fs_info, SKINNY_METADATA);
1514         int needed;
1515
1516         key.objectid = bytenr;
1517         key.type = BTRFS_EXTENT_ITEM_KEY;
1518         key.offset = num_bytes;
1519
1520         want = extent_ref_type(parent, owner);
1521         if (insert) {
1522                 extra_size = btrfs_extent_inline_ref_size(want);
1523                 path->keep_locks = 1;
1524         } else
1525                 extra_size = -1;
1526
1527         /*
1528          * Owner is our level, so we can just add one to get the level for the
1529          * block we are interested in.
1530          */
1531         if (skinny_metadata && owner < BTRFS_FIRST_FREE_OBJECTID) {
1532                 key.type = BTRFS_METADATA_ITEM_KEY;
1533                 key.offset = owner;
1534         }
1535
1536 again:
1537         ret = btrfs_search_slot(trans, root, &key, path, extra_size, 1);
1538         if (ret < 0) {
1539                 err = ret;
1540                 goto out;
1541         }
1542
1543         /*
1544          * We may be a newly converted file system which still has the old fat
1545          * extent entries for metadata, so try and see if we have one of those.
1546          */
1547         if (ret > 0 && skinny_metadata) {
1548                 skinny_metadata = false;
1549                 if (path->slots[0]) {
1550                         path->slots[0]--;
1551                         btrfs_item_key_to_cpu(path->nodes[0], &key,
1552                                               path->slots[0]);
1553                         if (key.objectid == bytenr &&
1554                             key.type == BTRFS_EXTENT_ITEM_KEY &&
1555                             key.offset == num_bytes)
1556                                 ret = 0;
1557                 }
1558                 if (ret) {
1559                         key.objectid = bytenr;
1560                         key.type = BTRFS_EXTENT_ITEM_KEY;
1561                         key.offset = num_bytes;
1562                         btrfs_release_path(path);
1563                         goto again;
1564                 }
1565         }
1566
1567         if (ret && !insert) {
1568                 err = -ENOENT;
1569                 goto out;
1570         } else if (WARN_ON(ret)) {
1571                 err = -EIO;
1572                 goto out;
1573         }
1574
1575         leaf = path->nodes[0];
1576         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1577         if (unlikely(item_size < sizeof(*ei))) {
1578                 err = -EINVAL;
1579                 btrfs_print_v0_err(fs_info);
1580                 btrfs_abort_transaction(trans, err);
1581                 goto out;
1582         }
1583
1584         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1585         flags = btrfs_extent_flags(leaf, ei);
1586
1587         ptr = (unsigned long)(ei + 1);
1588         end = (unsigned long)ei + item_size;
1589
1590         if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK && !skinny_metadata) {
1591                 ptr += sizeof(struct btrfs_tree_block_info);
1592                 BUG_ON(ptr > end);
1593         }
1594
1595         if (owner >= BTRFS_FIRST_FREE_OBJECTID)
1596                 needed = BTRFS_REF_TYPE_DATA;
1597         else
1598                 needed = BTRFS_REF_TYPE_BLOCK;
1599
1600         err = -ENOENT;
1601         while (1) {
1602                 if (ptr >= end) {
1603                         WARN_ON(ptr > end);
1604                         break;
1605                 }
1606                 iref = (struct btrfs_extent_inline_ref *)ptr;
1607                 type = btrfs_get_extent_inline_ref_type(leaf, iref, needed);
1608                 if (type == BTRFS_REF_TYPE_INVALID) {
1609                         err = -EUCLEAN;
1610                         goto out;
1611                 }
1612
1613                 if (want < type)
1614                         break;
1615                 if (want > type) {
1616                         ptr += btrfs_extent_inline_ref_size(type);
1617                         continue;
1618                 }
1619
1620                 if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1621                         struct btrfs_extent_data_ref *dref;
1622                         dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1623                         if (match_extent_data_ref(leaf, dref, root_objectid,
1624                                                   owner, offset)) {
1625                                 err = 0;
1626                                 break;
1627                         }
1628                         if (hash_extent_data_ref_item(leaf, dref) <
1629                             hash_extent_data_ref(root_objectid, owner, offset))
1630                                 break;
1631                 } else {
1632                         u64 ref_offset;
1633                         ref_offset = btrfs_extent_inline_ref_offset(leaf, iref);
1634                         if (parent > 0) {
1635                                 if (parent == ref_offset) {
1636                                         err = 0;
1637                                         break;
1638                                 }
1639                                 if (ref_offset < parent)
1640                                         break;
1641                         } else {
1642                                 if (root_objectid == ref_offset) {
1643                                         err = 0;
1644                                         break;
1645                                 }
1646                                 if (ref_offset < root_objectid)
1647                                         break;
1648                         }
1649                 }
1650                 ptr += btrfs_extent_inline_ref_size(type);
1651         }
1652         if (err == -ENOENT && insert) {
1653                 if (item_size + extra_size >=
1654                     BTRFS_MAX_EXTENT_ITEM_SIZE(root)) {
1655                         err = -EAGAIN;
1656                         goto out;
1657                 }
1658                 /*
1659                  * To add new inline back ref, we have to make sure
1660                  * there is no corresponding back ref item.
1661                  * For simplicity, we just do not add new inline back
1662                  * ref if there is any kind of item for this block
1663                  */
1664                 if (find_next_key(path, 0, &key) == 0 &&
1665                     key.objectid == bytenr &&
1666                     key.type < BTRFS_BLOCK_GROUP_ITEM_KEY) {
1667                         err = -EAGAIN;
1668                         goto out;
1669                 }
1670         }
1671         *ref_ret = (struct btrfs_extent_inline_ref *)ptr;
1672 out:
1673         if (insert) {
1674                 path->keep_locks = 0;
1675                 btrfs_unlock_up_safe(path, 1);
1676         }
1677         return err;
1678 }
1679
1680 /*
1681  * helper to add new inline back ref
1682  */
1683 static noinline_for_stack
1684 void setup_inline_extent_backref(struct btrfs_fs_info *fs_info,
1685                                  struct btrfs_path *path,
1686                                  struct btrfs_extent_inline_ref *iref,
1687                                  u64 parent, u64 root_objectid,
1688                                  u64 owner, u64 offset, int refs_to_add,
1689                                  struct btrfs_delayed_extent_op *extent_op)
1690 {
1691         struct extent_buffer *leaf;
1692         struct btrfs_extent_item *ei;
1693         unsigned long ptr;
1694         unsigned long end;
1695         unsigned long item_offset;
1696         u64 refs;
1697         int size;
1698         int type;
1699
1700         leaf = path->nodes[0];
1701         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1702         item_offset = (unsigned long)iref - (unsigned long)ei;
1703
1704         type = extent_ref_type(parent, owner);
1705         size = btrfs_extent_inline_ref_size(type);
1706
1707         btrfs_extend_item(fs_info, path, size);
1708
1709         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1710         refs = btrfs_extent_refs(leaf, ei);
1711         refs += refs_to_add;
1712         btrfs_set_extent_refs(leaf, ei, refs);
1713         if (extent_op)
1714                 __run_delayed_extent_op(extent_op, leaf, ei);
1715
1716         ptr = (unsigned long)ei + item_offset;
1717         end = (unsigned long)ei + btrfs_item_size_nr(leaf, path->slots[0]);
1718         if (ptr < end - size)
1719                 memmove_extent_buffer(leaf, ptr + size, ptr,
1720                                       end - size - ptr);
1721
1722         iref = (struct btrfs_extent_inline_ref *)ptr;
1723         btrfs_set_extent_inline_ref_type(leaf, iref, type);
1724         if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1725                 struct btrfs_extent_data_ref *dref;
1726                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1727                 btrfs_set_extent_data_ref_root(leaf, dref, root_objectid);
1728                 btrfs_set_extent_data_ref_objectid(leaf, dref, owner);
1729                 btrfs_set_extent_data_ref_offset(leaf, dref, offset);
1730                 btrfs_set_extent_data_ref_count(leaf, dref, refs_to_add);
1731         } else if (type == BTRFS_SHARED_DATA_REF_KEY) {
1732                 struct btrfs_shared_data_ref *sref;
1733                 sref = (struct btrfs_shared_data_ref *)(iref + 1);
1734                 btrfs_set_shared_data_ref_count(leaf, sref, refs_to_add);
1735                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
1736         } else if (type == BTRFS_SHARED_BLOCK_REF_KEY) {
1737                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
1738         } else {
1739                 btrfs_set_extent_inline_ref_offset(leaf, iref, root_objectid);
1740         }
1741         btrfs_mark_buffer_dirty(leaf);
1742 }
1743
1744 static int lookup_extent_backref(struct btrfs_trans_handle *trans,
1745                                  struct btrfs_path *path,
1746                                  struct btrfs_extent_inline_ref **ref_ret,
1747                                  u64 bytenr, u64 num_bytes, u64 parent,
1748                                  u64 root_objectid, u64 owner, u64 offset)
1749 {
1750         int ret;
1751
1752         ret = lookup_inline_extent_backref(trans, path, ref_ret, bytenr,
1753                                            num_bytes, parent, root_objectid,
1754                                            owner, offset, 0);
1755         if (ret != -ENOENT)
1756                 return ret;
1757
1758         btrfs_release_path(path);
1759         *ref_ret = NULL;
1760
1761         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1762                 ret = lookup_tree_block_ref(trans, path, bytenr, parent,
1763                                             root_objectid);
1764         } else {
1765                 ret = lookup_extent_data_ref(trans, path, bytenr, parent,
1766                                              root_objectid, owner, offset);
1767         }
1768         return ret;
1769 }
1770
1771 /*
1772  * helper to update/remove inline back ref
1773  */
1774 static noinline_for_stack
1775 void update_inline_extent_backref(struct btrfs_path *path,
1776                                   struct btrfs_extent_inline_ref *iref,
1777                                   int refs_to_mod,
1778                                   struct btrfs_delayed_extent_op *extent_op,
1779                                   int *last_ref)
1780 {
1781         struct extent_buffer *leaf = path->nodes[0];
1782         struct btrfs_fs_info *fs_info = leaf->fs_info;
1783         struct btrfs_extent_item *ei;
1784         struct btrfs_extent_data_ref *dref = NULL;
1785         struct btrfs_shared_data_ref *sref = NULL;
1786         unsigned long ptr;
1787         unsigned long end;
1788         u32 item_size;
1789         int size;
1790         int type;
1791         u64 refs;
1792
1793         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1794         refs = btrfs_extent_refs(leaf, ei);
1795         WARN_ON(refs_to_mod < 0 && refs + refs_to_mod <= 0);
1796         refs += refs_to_mod;
1797         btrfs_set_extent_refs(leaf, ei, refs);
1798         if (extent_op)
1799                 __run_delayed_extent_op(extent_op, leaf, ei);
1800
1801         /*
1802          * If type is invalid, we should have bailed out after
1803          * lookup_inline_extent_backref().
1804          */
1805         type = btrfs_get_extent_inline_ref_type(leaf, iref, BTRFS_REF_TYPE_ANY);
1806         ASSERT(type != BTRFS_REF_TYPE_INVALID);
1807
1808         if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1809                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1810                 refs = btrfs_extent_data_ref_count(leaf, dref);
1811         } else if (type == BTRFS_SHARED_DATA_REF_KEY) {
1812                 sref = (struct btrfs_shared_data_ref *)(iref + 1);
1813                 refs = btrfs_shared_data_ref_count(leaf, sref);
1814         } else {
1815                 refs = 1;
1816                 BUG_ON(refs_to_mod != -1);
1817         }
1818
1819         BUG_ON(refs_to_mod < 0 && refs < -refs_to_mod);
1820         refs += refs_to_mod;
1821
1822         if (refs > 0) {
1823                 if (type == BTRFS_EXTENT_DATA_REF_KEY)
1824                         btrfs_set_extent_data_ref_count(leaf, dref, refs);
1825                 else
1826                         btrfs_set_shared_data_ref_count(leaf, sref, refs);
1827         } else {
1828                 *last_ref = 1;
1829                 size =  btrfs_extent_inline_ref_size(type);
1830                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1831                 ptr = (unsigned long)iref;
1832                 end = (unsigned long)ei + item_size;
1833                 if (ptr + size < end)
1834                         memmove_extent_buffer(leaf, ptr, ptr + size,
1835                                               end - ptr - size);
1836                 item_size -= size;
1837                 btrfs_truncate_item(fs_info, path, item_size, 1);
1838         }
1839         btrfs_mark_buffer_dirty(leaf);
1840 }
1841
1842 static noinline_for_stack
1843 int insert_inline_extent_backref(struct btrfs_trans_handle *trans,
1844                                  struct btrfs_path *path,
1845                                  u64 bytenr, u64 num_bytes, u64 parent,
1846                                  u64 root_objectid, u64 owner,
1847                                  u64 offset, int refs_to_add,
1848                                  struct btrfs_delayed_extent_op *extent_op)
1849 {
1850         struct btrfs_extent_inline_ref *iref;
1851         int ret;
1852
1853         ret = lookup_inline_extent_backref(trans, path, &iref, bytenr,
1854                                            num_bytes, parent, root_objectid,
1855                                            owner, offset, 1);
1856         if (ret == 0) {
1857                 BUG_ON(owner < BTRFS_FIRST_FREE_OBJECTID);
1858                 update_inline_extent_backref(path, iref, refs_to_add,
1859                                              extent_op, NULL);
1860         } else if (ret == -ENOENT) {
1861                 setup_inline_extent_backref(trans->fs_info, path, iref, parent,
1862                                             root_objectid, owner, offset,
1863                                             refs_to_add, extent_op);
1864                 ret = 0;
1865         }
1866         return ret;
1867 }
1868
1869 static int insert_extent_backref(struct btrfs_trans_handle *trans,
1870                                  struct btrfs_path *path,
1871                                  u64 bytenr, u64 parent, u64 root_objectid,
1872                                  u64 owner, u64 offset, int refs_to_add)
1873 {
1874         int ret;
1875         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1876                 BUG_ON(refs_to_add != 1);
1877                 ret = insert_tree_block_ref(trans, path, bytenr, parent,
1878                                             root_objectid);
1879         } else {
1880                 ret = insert_extent_data_ref(trans, path, bytenr, parent,
1881                                              root_objectid, owner, offset,
1882                                              refs_to_add);
1883         }
1884         return ret;
1885 }
1886
1887 static int remove_extent_backref(struct btrfs_trans_handle *trans,
1888                                  struct btrfs_path *path,
1889                                  struct btrfs_extent_inline_ref *iref,
1890                                  int refs_to_drop, int is_data, int *last_ref)
1891 {
1892         int ret = 0;
1893
1894         BUG_ON(!is_data && refs_to_drop != 1);
1895         if (iref) {
1896                 update_inline_extent_backref(path, iref, -refs_to_drop, NULL,
1897                                              last_ref);
1898         } else if (is_data) {
1899                 ret = remove_extent_data_ref(trans, path, refs_to_drop,
1900                                              last_ref);
1901         } else {
1902                 *last_ref = 1;
1903                 ret = btrfs_del_item(trans, trans->fs_info->extent_root, path);
1904         }
1905         return ret;
1906 }
1907
1908 #define in_range(b, first, len)        ((b) >= (first) && (b) < (first) + (len))
1909 static int btrfs_issue_discard(struct block_device *bdev, u64 start, u64 len,
1910                                u64 *discarded_bytes)
1911 {
1912         int j, ret = 0;
1913         u64 bytes_left, end;
1914         u64 aligned_start = ALIGN(start, 1 << 9);
1915
1916         if (WARN_ON(start != aligned_start)) {
1917                 len -= aligned_start - start;
1918                 len = round_down(len, 1 << 9);
1919                 start = aligned_start;
1920         }
1921
1922         *discarded_bytes = 0;
1923
1924         if (!len)
1925                 return 0;
1926
1927         end = start + len;
1928         bytes_left = len;
1929
1930         /* Skip any superblocks on this device. */
1931         for (j = 0; j < BTRFS_SUPER_MIRROR_MAX; j++) {
1932                 u64 sb_start = btrfs_sb_offset(j);
1933                 u64 sb_end = sb_start + BTRFS_SUPER_INFO_SIZE;
1934                 u64 size = sb_start - start;
1935
1936                 if (!in_range(sb_start, start, bytes_left) &&
1937                     !in_range(sb_end, start, bytes_left) &&
1938                     !in_range(start, sb_start, BTRFS_SUPER_INFO_SIZE))
1939                         continue;
1940
1941                 /*
1942                  * Superblock spans beginning of range.  Adjust start and
1943                  * try again.
1944                  */
1945                 if (sb_start <= start) {
1946                         start += sb_end - start;
1947                         if (start > end) {
1948                                 bytes_left = 0;
1949                                 break;
1950                         }
1951                         bytes_left = end - start;
1952                         continue;
1953                 }
1954
1955                 if (size) {
1956                         ret = blkdev_issue_discard(bdev, start >> 9, size >> 9,
1957                                                    GFP_NOFS, 0);
1958                         if (!ret)
1959                                 *discarded_bytes += size;
1960                         else if (ret != -EOPNOTSUPP)
1961                                 return ret;
1962                 }
1963
1964                 start = sb_end;
1965                 if (start > end) {
1966                         bytes_left = 0;
1967                         break;
1968                 }
1969                 bytes_left = end - start;
1970         }
1971
1972         if (bytes_left) {
1973                 ret = blkdev_issue_discard(bdev, start >> 9, bytes_left >> 9,
1974                                            GFP_NOFS, 0);
1975                 if (!ret)
1976                         *discarded_bytes += bytes_left;
1977         }
1978         return ret;
1979 }
1980
1981 int btrfs_discard_extent(struct btrfs_fs_info *fs_info, u64 bytenr,
1982                          u64 num_bytes, u64 *actual_bytes)
1983 {
1984         int ret;
1985         u64 discarded_bytes = 0;
1986         struct btrfs_bio *bbio = NULL;
1987
1988
1989         /*
1990          * Avoid races with device replace and make sure our bbio has devices
1991          * associated to its stripes that don't go away while we are discarding.
1992          */
1993         btrfs_bio_counter_inc_blocked(fs_info);
1994         /* Tell the block device(s) that the sectors can be discarded */
1995         ret = btrfs_map_block(fs_info, BTRFS_MAP_DISCARD, bytenr, &num_bytes,
1996                               &bbio, 0);
1997         /* Error condition is -ENOMEM */
1998         if (!ret) {
1999                 struct btrfs_bio_stripe *stripe = bbio->stripes;
2000                 int i;
2001
2002
2003                 for (i = 0; i < bbio->num_stripes; i++, stripe++) {
2004                         u64 bytes;
2005                         struct request_queue *req_q;
2006
2007                         if (!stripe->dev->bdev) {
2008                                 ASSERT(btrfs_test_opt(fs_info, DEGRADED));
2009                                 continue;
2010                         }
2011                         req_q = bdev_get_queue(stripe->dev->bdev);
2012                         if (!blk_queue_discard(req_q))
2013                                 continue;
2014
2015                         ret = btrfs_issue_discard(stripe->dev->bdev,
2016                                                   stripe->physical,
2017                                                   stripe->length,
2018                                                   &bytes);
2019                         if (!ret)
2020                                 discarded_bytes += bytes;
2021                         else if (ret != -EOPNOTSUPP)
2022                                 break; /* Logic errors or -ENOMEM, or -EIO but I don't know how that could happen JDM */
2023
2024                         /*
2025                          * Just in case we get back EOPNOTSUPP for some reason,
2026                          * just ignore the return value so we don't screw up
2027                          * people calling discard_extent.
2028                          */
2029                         ret = 0;
2030                 }
2031                 btrfs_put_bbio(bbio);
2032         }
2033         btrfs_bio_counter_dec(fs_info);
2034
2035         if (actual_bytes)
2036                 *actual_bytes = discarded_bytes;
2037
2038
2039         if (ret == -EOPNOTSUPP)
2040                 ret = 0;
2041         return ret;
2042 }
2043
2044 /* Can return -ENOMEM */
2045 int btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
2046                          struct btrfs_root *root,
2047                          u64 bytenr, u64 num_bytes, u64 parent,
2048                          u64 root_objectid, u64 owner, u64 offset)
2049 {
2050         struct btrfs_fs_info *fs_info = root->fs_info;
2051         int old_ref_mod, new_ref_mod;
2052         int ret;
2053
2054         BUG_ON(owner < BTRFS_FIRST_FREE_OBJECTID &&
2055                root_objectid == BTRFS_TREE_LOG_OBJECTID);
2056
2057         btrfs_ref_tree_mod(root, bytenr, num_bytes, parent, root_objectid,
2058                            owner, offset, BTRFS_ADD_DELAYED_REF);
2059
2060         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
2061                 ret = btrfs_add_delayed_tree_ref(trans, bytenr,
2062                                                  num_bytes, parent,
2063                                                  root_objectid, (int)owner,
2064                                                  BTRFS_ADD_DELAYED_REF, NULL,
2065                                                  &old_ref_mod, &new_ref_mod);
2066         } else {
2067                 ret = btrfs_add_delayed_data_ref(trans, bytenr,
2068                                                  num_bytes, parent,
2069                                                  root_objectid, owner, offset,
2070                                                  0, BTRFS_ADD_DELAYED_REF,
2071                                                  &old_ref_mod, &new_ref_mod);
2072         }
2073
2074         if (ret == 0 && old_ref_mod < 0 && new_ref_mod >= 0) {
2075                 bool metadata = owner < BTRFS_FIRST_FREE_OBJECTID;
2076
2077                 add_pinned_bytes(fs_info, -num_bytes, metadata, root_objectid);
2078         }
2079
2080         return ret;
2081 }
2082
2083 /*
2084  * __btrfs_inc_extent_ref - insert backreference for a given extent
2085  *
2086  * @trans:          Handle of transaction
2087  *
2088  * @node:           The delayed ref node used to get the bytenr/length for
2089  *                  extent whose references are incremented.
2090  *
2091  * @parent:         If this is a shared extent (BTRFS_SHARED_DATA_REF_KEY/
2092  *                  BTRFS_SHARED_BLOCK_REF_KEY) then it holds the logical
2093  *                  bytenr of the parent block. Since new extents are always
2094  *                  created with indirect references, this will only be the case
2095  *                  when relocating a shared extent. In that case, root_objectid
2096  *                  will be BTRFS_TREE_RELOC_OBJECTID. Otheriwse, parent must
2097  *                  be 0
2098  *
2099  * @root_objectid:  The id of the root where this modification has originated,
2100  *                  this can be either one of the well-known metadata trees or
2101  *                  the subvolume id which references this extent.
2102  *
2103  * @owner:          For data extents it is the inode number of the owning file.
2104  *                  For metadata extents this parameter holds the level in the
2105  *                  tree of the extent.
2106  *
2107  * @offset:         For metadata extents the offset is ignored and is currently
2108  *                  always passed as 0. For data extents it is the fileoffset
2109  *                  this extent belongs to.
2110  *
2111  * @refs_to_add     Number of references to add
2112  *
2113  * @extent_op       Pointer to a structure, holding information necessary when
2114  *                  updating a tree block's flags
2115  *
2116  */
2117 static int __btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
2118                                   struct btrfs_delayed_ref_node *node,
2119                                   u64 parent, u64 root_objectid,
2120                                   u64 owner, u64 offset, int refs_to_add,
2121                                   struct btrfs_delayed_extent_op *extent_op)
2122 {
2123         struct btrfs_path *path;
2124         struct extent_buffer *leaf;
2125         struct btrfs_extent_item *item;
2126         struct btrfs_key key;
2127         u64 bytenr = node->bytenr;
2128         u64 num_bytes = node->num_bytes;
2129         u64 refs;
2130         int ret;
2131
2132         path = btrfs_alloc_path();
2133         if (!path)
2134                 return -ENOMEM;
2135
2136         path->reada = READA_FORWARD;
2137         path->leave_spinning = 1;
2138         /* this will setup the path even if it fails to insert the back ref */
2139         ret = insert_inline_extent_backref(trans, path, bytenr, num_bytes,
2140                                            parent, root_objectid, owner,
2141                                            offset, refs_to_add, extent_op);
2142         if ((ret < 0 && ret != -EAGAIN) || !ret)
2143                 goto out;
2144
2145         /*
2146          * Ok we had -EAGAIN which means we didn't have space to insert and
2147          * inline extent ref, so just update the reference count and add a
2148          * normal backref.
2149          */
2150         leaf = path->nodes[0];
2151         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
2152         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2153         refs = btrfs_extent_refs(leaf, item);
2154         btrfs_set_extent_refs(leaf, item, refs + refs_to_add);
2155         if (extent_op)
2156                 __run_delayed_extent_op(extent_op, leaf, item);
2157
2158         btrfs_mark_buffer_dirty(leaf);
2159         btrfs_release_path(path);
2160
2161         path->reada = READA_FORWARD;
2162         path->leave_spinning = 1;
2163         /* now insert the actual backref */
2164         ret = insert_extent_backref(trans, path, bytenr, parent, root_objectid,
2165                                     owner, offset, refs_to_add);
2166         if (ret)
2167                 btrfs_abort_transaction(trans, ret);
2168 out:
2169         btrfs_free_path(path);
2170         return ret;
2171 }
2172
2173 static int run_delayed_data_ref(struct btrfs_trans_handle *trans,
2174                                 struct btrfs_delayed_ref_node *node,
2175                                 struct btrfs_delayed_extent_op *extent_op,
2176                                 int insert_reserved)
2177 {
2178         int ret = 0;
2179         struct btrfs_delayed_data_ref *ref;
2180         struct btrfs_key ins;
2181         u64 parent = 0;
2182         u64 ref_root = 0;
2183         u64 flags = 0;
2184
2185         ins.objectid = node->bytenr;
2186         ins.offset = node->num_bytes;
2187         ins.type = BTRFS_EXTENT_ITEM_KEY;
2188
2189         ref = btrfs_delayed_node_to_data_ref(node);
2190         trace_run_delayed_data_ref(trans->fs_info, node, ref, node->action);
2191
2192         if (node->type == BTRFS_SHARED_DATA_REF_KEY)
2193                 parent = ref->parent;
2194         ref_root = ref->root;
2195
2196         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
2197                 if (extent_op)
2198                         flags |= extent_op->flags_to_set;
2199                 ret = alloc_reserved_file_extent(trans, parent, ref_root,
2200                                                  flags, ref->objectid,
2201                                                  ref->offset, &ins,
2202                                                  node->ref_mod);
2203         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
2204                 ret = __btrfs_inc_extent_ref(trans, node, parent, ref_root,
2205                                              ref->objectid, ref->offset,
2206                                              node->ref_mod, extent_op);
2207         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
2208                 ret = __btrfs_free_extent(trans, node, parent,
2209                                           ref_root, ref->objectid,
2210                                           ref->offset, node->ref_mod,
2211                                           extent_op);
2212         } else {
2213                 BUG();
2214         }
2215         return ret;
2216 }
2217
2218 static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op,
2219                                     struct extent_buffer *leaf,
2220                                     struct btrfs_extent_item *ei)
2221 {
2222         u64 flags = btrfs_extent_flags(leaf, ei);
2223         if (extent_op->update_flags) {
2224                 flags |= extent_op->flags_to_set;
2225                 btrfs_set_extent_flags(leaf, ei, flags);
2226         }
2227
2228         if (extent_op->update_key) {
2229                 struct btrfs_tree_block_info *bi;
2230                 BUG_ON(!(flags & BTRFS_EXTENT_FLAG_TREE_BLOCK));
2231                 bi = (struct btrfs_tree_block_info *)(ei + 1);
2232                 btrfs_set_tree_block_key(leaf, bi, &extent_op->key);
2233         }
2234 }
2235
2236 static int run_delayed_extent_op(struct btrfs_trans_handle *trans,
2237                                  struct btrfs_delayed_ref_head *head,
2238                                  struct btrfs_delayed_extent_op *extent_op)
2239 {
2240         struct btrfs_fs_info *fs_info = trans->fs_info;
2241         struct btrfs_key key;
2242         struct btrfs_path *path;
2243         struct btrfs_extent_item *ei;
2244         struct extent_buffer *leaf;
2245         u32 item_size;
2246         int ret;
2247         int err = 0;
2248         int metadata = !extent_op->is_data;
2249
2250         if (trans->aborted)
2251                 return 0;
2252
2253         if (metadata && !btrfs_fs_incompat(fs_info, SKINNY_METADATA))
2254                 metadata = 0;
2255
2256         path = btrfs_alloc_path();
2257         if (!path)
2258                 return -ENOMEM;
2259
2260         key.objectid = head->bytenr;
2261
2262         if (metadata) {
2263                 key.type = BTRFS_METADATA_ITEM_KEY;
2264                 key.offset = extent_op->level;
2265         } else {
2266                 key.type = BTRFS_EXTENT_ITEM_KEY;
2267                 key.offset = head->num_bytes;
2268         }
2269
2270 again:
2271         path->reada = READA_FORWARD;
2272         path->leave_spinning = 1;
2273         ret = btrfs_search_slot(trans, fs_info->extent_root, &key, path, 0, 1);
2274         if (ret < 0) {
2275                 err = ret;
2276                 goto out;
2277         }
2278         if (ret > 0) {
2279                 if (metadata) {
2280                         if (path->slots[0] > 0) {
2281                                 path->slots[0]--;
2282                                 btrfs_item_key_to_cpu(path->nodes[0], &key,
2283                                                       path->slots[0]);
2284                                 if (key.objectid == head->bytenr &&
2285                                     key.type == BTRFS_EXTENT_ITEM_KEY &&
2286                                     key.offset == head->num_bytes)
2287                                         ret = 0;
2288                         }
2289                         if (ret > 0) {
2290                                 btrfs_release_path(path);
2291                                 metadata = 0;
2292
2293                                 key.objectid = head->bytenr;
2294                                 key.offset = head->num_bytes;
2295                                 key.type = BTRFS_EXTENT_ITEM_KEY;
2296                                 goto again;
2297                         }
2298                 } else {
2299                         err = -EIO;
2300                         goto out;
2301                 }
2302         }
2303
2304         leaf = path->nodes[0];
2305         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2306
2307         if (unlikely(item_size < sizeof(*ei))) {
2308                 err = -EINVAL;
2309                 btrfs_print_v0_err(fs_info);
2310                 btrfs_abort_transaction(trans, err);
2311                 goto out;
2312         }
2313
2314         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2315         __run_delayed_extent_op(extent_op, leaf, ei);
2316
2317         btrfs_mark_buffer_dirty(leaf);
2318 out:
2319         btrfs_free_path(path);
2320         return err;
2321 }
2322
2323 static int run_delayed_tree_ref(struct btrfs_trans_handle *trans,
2324                                 struct btrfs_delayed_ref_node *node,
2325                                 struct btrfs_delayed_extent_op *extent_op,
2326                                 int insert_reserved)
2327 {
2328         int ret = 0;
2329         struct btrfs_delayed_tree_ref *ref;
2330         u64 parent = 0;
2331         u64 ref_root = 0;
2332
2333         ref = btrfs_delayed_node_to_tree_ref(node);
2334         trace_run_delayed_tree_ref(trans->fs_info, node, ref, node->action);
2335
2336         if (node->type == BTRFS_SHARED_BLOCK_REF_KEY)
2337                 parent = ref->parent;
2338         ref_root = ref->root;
2339
2340         if (node->ref_mod != 1) {
2341                 btrfs_err(trans->fs_info,
2342         "btree block(%llu) has %d references rather than 1: action %d ref_root %llu parent %llu",
2343                           node->bytenr, node->ref_mod, node->action, ref_root,
2344                           parent);
2345                 return -EIO;
2346         }
2347         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
2348                 BUG_ON(!extent_op || !extent_op->update_flags);
2349                 ret = alloc_reserved_tree_block(trans, node, extent_op);
2350         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
2351                 ret = __btrfs_inc_extent_ref(trans, node, parent, ref_root,
2352                                              ref->level, 0, 1, extent_op);
2353         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
2354                 ret = __btrfs_free_extent(trans, node, parent, ref_root,
2355                                           ref->level, 0, 1, extent_op);
2356         } else {
2357                 BUG();
2358         }
2359         return ret;
2360 }
2361
2362 /* helper function to actually process a single delayed ref entry */
2363 static int run_one_delayed_ref(struct btrfs_trans_handle *trans,
2364                                struct btrfs_delayed_ref_node *node,
2365                                struct btrfs_delayed_extent_op *extent_op,
2366                                int insert_reserved)
2367 {
2368         int ret = 0;
2369
2370         if (trans->aborted) {
2371                 if (insert_reserved)
2372                         btrfs_pin_extent(trans->fs_info, node->bytenr,
2373                                          node->num_bytes, 1);
2374                 return 0;
2375         }
2376
2377         if (node->type == BTRFS_TREE_BLOCK_REF_KEY ||
2378             node->type == BTRFS_SHARED_BLOCK_REF_KEY)
2379                 ret = run_delayed_tree_ref(trans, node, extent_op,
2380                                            insert_reserved);
2381         else if (node->type == BTRFS_EXTENT_DATA_REF_KEY ||
2382                  node->type == BTRFS_SHARED_DATA_REF_KEY)
2383                 ret = run_delayed_data_ref(trans, node, extent_op,
2384                                            insert_reserved);
2385         else
2386                 BUG();
2387         if (ret && insert_reserved)
2388                 btrfs_pin_extent(trans->fs_info, node->bytenr,
2389                                  node->num_bytes, 1);
2390         return ret;
2391 }
2392
2393 static inline struct btrfs_delayed_ref_node *
2394 select_delayed_ref(struct btrfs_delayed_ref_head *head)
2395 {
2396         struct btrfs_delayed_ref_node *ref;
2397
2398         if (RB_EMPTY_ROOT(&head->ref_tree.rb_root))
2399                 return NULL;
2400
2401         /*
2402          * Select a delayed ref of type BTRFS_ADD_DELAYED_REF first.
2403          * This is to prevent a ref count from going down to zero, which deletes
2404          * the extent item from the extent tree, when there still are references
2405          * to add, which would fail because they would not find the extent item.
2406          */
2407         if (!list_empty(&head->ref_add_list))
2408                 return list_first_entry(&head->ref_add_list,
2409                                 struct btrfs_delayed_ref_node, add_list);
2410
2411         ref = rb_entry(rb_first_cached(&head->ref_tree),
2412                        struct btrfs_delayed_ref_node, ref_node);
2413         ASSERT(list_empty(&ref->add_list));
2414         return ref;
2415 }
2416
2417 static void unselect_delayed_ref_head(struct btrfs_delayed_ref_root *delayed_refs,
2418                                       struct btrfs_delayed_ref_head *head)
2419 {
2420         spin_lock(&delayed_refs->lock);
2421         head->processing = 0;
2422         delayed_refs->num_heads_ready++;
2423         spin_unlock(&delayed_refs->lock);
2424         btrfs_delayed_ref_unlock(head);
2425 }
2426
2427 static struct btrfs_delayed_extent_op *cleanup_extent_op(
2428                                 struct btrfs_delayed_ref_head *head)
2429 {
2430         struct btrfs_delayed_extent_op *extent_op = head->extent_op;
2431
2432         if (!extent_op)
2433                 return NULL;
2434
2435         if (head->must_insert_reserved) {
2436                 head->extent_op = NULL;
2437                 btrfs_free_delayed_extent_op(extent_op);
2438                 return NULL;
2439         }
2440         return extent_op;
2441 }
2442
2443 static int run_and_cleanup_extent_op(struct btrfs_trans_handle *trans,
2444                                      struct btrfs_delayed_ref_head *head)
2445 {
2446         struct btrfs_delayed_extent_op *extent_op;
2447         int ret;
2448
2449         extent_op = cleanup_extent_op(head);
2450         if (!extent_op)
2451                 return 0;
2452         head->extent_op = NULL;
2453         spin_unlock(&head->lock);
2454         ret = run_delayed_extent_op(trans, head, extent_op);
2455         btrfs_free_delayed_extent_op(extent_op);
2456         return ret ? ret : 1;
2457 }
2458
2459 void btrfs_cleanup_ref_head_accounting(struct btrfs_fs_info *fs_info,
2460                                   struct btrfs_delayed_ref_root *delayed_refs,
2461                                   struct btrfs_delayed_ref_head *head)
2462 {
2463         int nr_items = 1;       /* Dropping this ref head update. */
2464
2465         if (head->total_ref_mod < 0) {
2466                 struct btrfs_space_info *space_info;
2467                 u64 flags;
2468
2469                 if (head->is_data)
2470                         flags = BTRFS_BLOCK_GROUP_DATA;
2471                 else if (head->is_system)
2472                         flags = BTRFS_BLOCK_GROUP_SYSTEM;
2473                 else
2474                         flags = BTRFS_BLOCK_GROUP_METADATA;
2475                 space_info = __find_space_info(fs_info, flags);
2476                 ASSERT(space_info);
2477                 percpu_counter_add_batch(&space_info->total_bytes_pinned,
2478                                    -head->num_bytes,
2479                                    BTRFS_TOTAL_BYTES_PINNED_BATCH);
2480
2481                 /*
2482                  * We had csum deletions accounted for in our delayed refs rsv,
2483                  * we need to drop the csum leaves for this update from our
2484                  * delayed_refs_rsv.
2485                  */
2486                 if (head->is_data) {
2487                         spin_lock(&delayed_refs->lock);
2488                         delayed_refs->pending_csums -= head->num_bytes;
2489                         spin_unlock(&delayed_refs->lock);
2490                         nr_items += btrfs_csum_bytes_to_leaves(fs_info,
2491                                 head->num_bytes);
2492                 }
2493         }
2494
2495         btrfs_delayed_refs_rsv_release(fs_info, nr_items);
2496 }
2497
2498 static int cleanup_ref_head(struct btrfs_trans_handle *trans,
2499                             struct btrfs_delayed_ref_head *head)
2500 {
2501
2502         struct btrfs_fs_info *fs_info = trans->fs_info;
2503         struct btrfs_delayed_ref_root *delayed_refs;
2504         int ret;
2505
2506         delayed_refs = &trans->transaction->delayed_refs;
2507
2508         ret = run_and_cleanup_extent_op(trans, head);
2509         if (ret < 0) {
2510                 unselect_delayed_ref_head(delayed_refs, head);
2511                 btrfs_debug(fs_info, "run_delayed_extent_op returned %d", ret);
2512                 return ret;
2513         } else if (ret) {
2514                 return ret;
2515         }
2516
2517         /*
2518          * Need to drop our head ref lock and re-acquire the delayed ref lock
2519          * and then re-check to make sure nobody got added.
2520          */
2521         spin_unlock(&head->lock);
2522         spin_lock(&delayed_refs->lock);
2523         spin_lock(&head->lock);
2524         if (!RB_EMPTY_ROOT(&head->ref_tree.rb_root) || head->extent_op) {
2525                 spin_unlock(&head->lock);
2526                 spin_unlock(&delayed_refs->lock);
2527                 return 1;
2528         }
2529         btrfs_delete_ref_head(delayed_refs, head);
2530         spin_unlock(&head->lock);
2531         spin_unlock(&delayed_refs->lock);
2532
2533         if (head->must_insert_reserved) {
2534                 btrfs_pin_extent(fs_info, head->bytenr,
2535                                  head->num_bytes, 1);
2536                 if (head->is_data) {
2537                         ret = btrfs_del_csums(trans, fs_info, head->bytenr,
2538                                               head->num_bytes);
2539                 }
2540         }
2541
2542         btrfs_cleanup_ref_head_accounting(fs_info, delayed_refs, head);
2543
2544         trace_run_delayed_ref_head(fs_info, head, 0);
2545         btrfs_delayed_ref_unlock(head);
2546         btrfs_put_delayed_ref_head(head);
2547         return 0;
2548 }
2549
2550 static struct btrfs_delayed_ref_head *btrfs_obtain_ref_head(
2551                                         struct btrfs_trans_handle *trans)
2552 {
2553         struct btrfs_delayed_ref_root *delayed_refs =
2554                 &trans->transaction->delayed_refs;
2555         struct btrfs_delayed_ref_head *head = NULL;
2556         int ret;
2557
2558         spin_lock(&delayed_refs->lock);
2559         head = btrfs_select_ref_head(delayed_refs);
2560         if (!head) {
2561                 spin_unlock(&delayed_refs->lock);
2562                 return head;
2563         }
2564
2565         /*
2566          * Grab the lock that says we are going to process all the refs for
2567          * this head
2568          */
2569         ret = btrfs_delayed_ref_lock(delayed_refs, head);
2570         spin_unlock(&delayed_refs->lock);
2571
2572         /*
2573          * We may have dropped the spin lock to get the head mutex lock, and
2574          * that might have given someone else time to free the head.  If that's
2575          * true, it has been removed from our list and we can move on.
2576          */
2577         if (ret == -EAGAIN)
2578                 head = ERR_PTR(-EAGAIN);
2579
2580         return head;
2581 }
2582
2583 static int btrfs_run_delayed_refs_for_head(struct btrfs_trans_handle *trans,
2584                                     struct btrfs_delayed_ref_head *locked_ref,
2585                                     unsigned long *run_refs)
2586 {
2587         struct btrfs_fs_info *fs_info = trans->fs_info;
2588         struct btrfs_delayed_ref_root *delayed_refs;
2589         struct btrfs_delayed_extent_op *extent_op;
2590         struct btrfs_delayed_ref_node *ref;
2591         int must_insert_reserved = 0;
2592         int ret;
2593
2594         delayed_refs = &trans->transaction->delayed_refs;
2595
2596         lockdep_assert_held(&locked_ref->mutex);
2597         lockdep_assert_held(&locked_ref->lock);
2598
2599         while ((ref = select_delayed_ref(locked_ref))) {
2600                 if (ref->seq &&
2601                     btrfs_check_delayed_seq(fs_info, ref->seq)) {
2602                         spin_unlock(&locked_ref->lock);
2603                         unselect_delayed_ref_head(delayed_refs, locked_ref);
2604                         return -EAGAIN;
2605                 }
2606
2607                 (*run_refs)++;
2608                 ref->in_tree = 0;
2609                 rb_erase_cached(&ref->ref_node, &locked_ref->ref_tree);
2610                 RB_CLEAR_NODE(&ref->ref_node);
2611                 if (!list_empty(&ref->add_list))
2612                         list_del(&ref->add_list);
2613                 /*
2614                  * When we play the delayed ref, also correct the ref_mod on
2615                  * head
2616                  */
2617                 switch (ref->action) {
2618                 case BTRFS_ADD_DELAYED_REF:
2619                 case BTRFS_ADD_DELAYED_EXTENT:
2620                         locked_ref->ref_mod -= ref->ref_mod;
2621                         break;
2622                 case BTRFS_DROP_DELAYED_REF:
2623                         locked_ref->ref_mod += ref->ref_mod;
2624                         break;
2625                 default:
2626                         WARN_ON(1);
2627                 }
2628                 atomic_dec(&delayed_refs->num_entries);
2629
2630                 /*
2631                  * Record the must_insert_reserved flag before we drop the
2632                  * spin lock.
2633                  */
2634                 must_insert_reserved = locked_ref->must_insert_reserved;
2635                 locked_ref->must_insert_reserved = 0;
2636
2637                 extent_op = locked_ref->extent_op;
2638                 locked_ref->extent_op = NULL;
2639                 spin_unlock(&locked_ref->lock);
2640
2641                 ret = run_one_delayed_ref(trans, ref, extent_op,
2642                                           must_insert_reserved);
2643
2644                 btrfs_free_delayed_extent_op(extent_op);
2645                 if (ret) {
2646                         unselect_delayed_ref_head(delayed_refs, locked_ref);
2647                         btrfs_put_delayed_ref(ref);
2648                         btrfs_debug(fs_info, "run_one_delayed_ref returned %d",
2649                                     ret);
2650                         return ret;
2651                 }
2652
2653                 btrfs_put_delayed_ref(ref);
2654                 cond_resched();
2655
2656                 spin_lock(&locked_ref->lock);
2657                 btrfs_merge_delayed_refs(trans, delayed_refs, locked_ref);
2658         }
2659
2660         return 0;
2661 }
2662
2663 /*
2664  * Returns 0 on success or if called with an already aborted transaction.
2665  * Returns -ENOMEM or -EIO on failure and will abort the transaction.
2666  */
2667 static noinline int __btrfs_run_delayed_refs(struct btrfs_trans_handle *trans,
2668                                              unsigned long nr)
2669 {
2670         struct btrfs_fs_info *fs_info = trans->fs_info;
2671         struct btrfs_delayed_ref_root *delayed_refs;
2672         struct btrfs_delayed_ref_head *locked_ref = NULL;
2673         ktime_t start = ktime_get();
2674         int ret;
2675         unsigned long count = 0;
2676         unsigned long actual_count = 0;
2677
2678         delayed_refs = &trans->transaction->delayed_refs;
2679         do {
2680                 if (!locked_ref) {
2681                         locked_ref = btrfs_obtain_ref_head(trans);
2682                         if (IS_ERR_OR_NULL(locked_ref)) {
2683                                 if (PTR_ERR(locked_ref) == -EAGAIN) {
2684                                         continue;
2685                                 } else {
2686                                         break;
2687                                 }
2688                         }
2689                         count++;
2690                 }
2691                 /*
2692                  * We need to try and merge add/drops of the same ref since we
2693                  * can run into issues with relocate dropping the implicit ref
2694                  * and then it being added back again before the drop can
2695                  * finish.  If we merged anything we need to re-loop so we can
2696                  * get a good ref.
2697                  * Or we can get node references of the same type that weren't
2698                  * merged when created due to bumps in the tree mod seq, and
2699                  * we need to merge them to prevent adding an inline extent
2700                  * backref before dropping it (triggering a BUG_ON at
2701                  * insert_inline_extent_backref()).
2702                  */
2703                 spin_lock(&locked_ref->lock);
2704                 btrfs_merge_delayed_refs(trans, delayed_refs, locked_ref);
2705
2706                 ret = btrfs_run_delayed_refs_for_head(trans, locked_ref,
2707                                                       &actual_count);
2708                 if (ret < 0 && ret != -EAGAIN) {
2709                         /*
2710                          * Error, btrfs_run_delayed_refs_for_head already
2711                          * unlocked everything so just bail out
2712                          */
2713                         return ret;
2714                 } else if (!ret) {
2715                         /*
2716                          * Success, perform the usual cleanup of a processed
2717                          * head
2718                          */
2719                         ret = cleanup_ref_head(trans, locked_ref);
2720                         if (ret > 0 ) {
2721                                 /* We dropped our lock, we need to loop. */
2722                                 ret = 0;
2723                                 continue;
2724                         } else if (ret) {
2725                                 return ret;
2726                         }
2727                 }
2728
2729                 /*
2730                  * Either success case or btrfs_run_delayed_refs_for_head
2731                  * returned -EAGAIN, meaning we need to select another head
2732                  */
2733
2734                 locked_ref = NULL;
2735                 cond_resched();
2736         } while ((nr != -1 && count < nr) || locked_ref);
2737
2738         /*
2739          * We don't want to include ref heads since we can have empty ref heads
2740          * and those will drastically skew our runtime down since we just do
2741          * accounting, no actual extent tree updates.
2742          */
2743         if (actual_count > 0) {
2744                 u64 runtime = ktime_to_ns(ktime_sub(ktime_get(), start));
2745                 u64 avg;
2746
2747                 /*
2748                  * We weigh the current average higher than our current runtime
2749                  * to avoid large swings in the average.
2750                  */
2751                 spin_lock(&delayed_refs->lock);
2752                 avg = fs_info->avg_delayed_ref_runtime * 3 + runtime;
2753                 fs_info->avg_delayed_ref_runtime = avg >> 2;    /* div by 4 */
2754                 spin_unlock(&delayed_refs->lock);
2755         }
2756         return 0;
2757 }
2758
2759 #ifdef SCRAMBLE_DELAYED_REFS
2760 /*
2761  * Normally delayed refs get processed in ascending bytenr order. This
2762  * correlates in most cases to the order added. To expose dependencies on this
2763  * order, we start to process the tree in the middle instead of the beginning
2764  */
2765 static u64 find_middle(struct rb_root *root)
2766 {
2767         struct rb_node *n = root->rb_node;
2768         struct btrfs_delayed_ref_node *entry;
2769         int alt = 1;
2770         u64 middle;
2771         u64 first = 0, last = 0;
2772
2773         n = rb_first(root);
2774         if (n) {
2775                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2776                 first = entry->bytenr;
2777         }
2778         n = rb_last(root);
2779         if (n) {
2780                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2781                 last = entry->bytenr;
2782         }
2783         n = root->rb_node;
2784
2785         while (n) {
2786                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2787                 WARN_ON(!entry->in_tree);
2788
2789                 middle = entry->bytenr;
2790
2791                 if (alt)
2792                         n = n->rb_left;
2793                 else
2794                         n = n->rb_right;
2795
2796                 alt = 1 - alt;
2797         }
2798         return middle;
2799 }
2800 #endif
2801
2802 static inline u64 heads_to_leaves(struct btrfs_fs_info *fs_info, u64 heads)
2803 {
2804         u64 num_bytes;
2805
2806         num_bytes = heads * (sizeof(struct btrfs_extent_item) +
2807                              sizeof(struct btrfs_extent_inline_ref));
2808         if (!btrfs_fs_incompat(fs_info, SKINNY_METADATA))
2809                 num_bytes += heads * sizeof(struct btrfs_tree_block_info);
2810
2811         /*
2812          * We don't ever fill up leaves all the way so multiply by 2 just to be
2813          * closer to what we're really going to want to use.
2814          */
2815         return div_u64(num_bytes, BTRFS_LEAF_DATA_SIZE(fs_info));
2816 }
2817
2818 /*
2819  * Takes the number of bytes to be csumm'ed and figures out how many leaves it
2820  * would require to store the csums for that many bytes.
2821  */
2822 u64 btrfs_csum_bytes_to_leaves(struct btrfs_fs_info *fs_info, u64 csum_bytes)
2823 {
2824         u64 csum_size;
2825         u64 num_csums_per_leaf;
2826         u64 num_csums;
2827
2828         csum_size = BTRFS_MAX_ITEM_SIZE(fs_info);
2829         num_csums_per_leaf = div64_u64(csum_size,
2830                         (u64)btrfs_super_csum_size(fs_info->super_copy));
2831         num_csums = div64_u64(csum_bytes, fs_info->sectorsize);
2832         num_csums += num_csums_per_leaf - 1;
2833         num_csums = div64_u64(num_csums, num_csums_per_leaf);
2834         return num_csums;
2835 }
2836
2837 bool btrfs_check_space_for_delayed_refs(struct btrfs_fs_info *fs_info)
2838 {
2839         struct btrfs_block_rsv *delayed_refs_rsv = &fs_info->delayed_refs_rsv;
2840         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
2841         bool ret = false;
2842         u64 reserved;
2843
2844         spin_lock(&global_rsv->lock);
2845         reserved = global_rsv->reserved;
2846         spin_unlock(&global_rsv->lock);
2847
2848         /*
2849          * Since the global reserve is just kind of magic we don't really want
2850          * to rely on it to save our bacon, so if our size is more than the
2851          * delayed_refs_rsv and the global rsv then it's time to think about
2852          * bailing.
2853          */
2854         spin_lock(&delayed_refs_rsv->lock);
2855         reserved += delayed_refs_rsv->reserved;
2856         if (delayed_refs_rsv->size >= reserved)
2857                 ret = true;
2858         spin_unlock(&delayed_refs_rsv->lock);
2859         return ret;
2860 }
2861
2862 int btrfs_should_throttle_delayed_refs(struct btrfs_trans_handle *trans)
2863 {
2864         u64 num_entries =
2865                 atomic_read(&trans->transaction->delayed_refs.num_entries);
2866         u64 avg_runtime;
2867         u64 val;
2868
2869         smp_mb();
2870         avg_runtime = trans->fs_info->avg_delayed_ref_runtime;
2871         val = num_entries * avg_runtime;
2872         if (val >= NSEC_PER_SEC)
2873                 return 1;
2874         if (val >= NSEC_PER_SEC / 2)
2875                 return 2;
2876
2877         return btrfs_check_space_for_delayed_refs(trans->fs_info);
2878 }
2879
2880 struct async_delayed_refs {
2881         struct btrfs_root *root;
2882         u64 transid;
2883         int count;
2884         int error;
2885         int sync;
2886         struct completion wait;
2887         struct btrfs_work work;
2888 };
2889
2890 static inline struct async_delayed_refs *
2891 to_async_delayed_refs(struct btrfs_work *work)
2892 {
2893         return container_of(work, struct async_delayed_refs, work);
2894 }
2895
2896 static void delayed_ref_async_start(struct btrfs_work *work)
2897 {
2898         struct async_delayed_refs *async = to_async_delayed_refs(work);
2899         struct btrfs_trans_handle *trans;
2900         struct btrfs_fs_info *fs_info = async->root->fs_info;
2901         int ret;
2902
2903         /* if the commit is already started, we don't need to wait here */
2904         if (btrfs_transaction_blocked(fs_info))
2905                 goto done;
2906
2907         trans = btrfs_join_transaction(async->root);
2908         if (IS_ERR(trans)) {
2909                 async->error = PTR_ERR(trans);
2910                 goto done;
2911         }
2912
2913         /*
2914          * trans->sync means that when we call end_transaction, we won't
2915          * wait on delayed refs
2916          */
2917         trans->sync = true;
2918
2919         /* Don't bother flushing if we got into a different transaction */
2920         if (trans->transid > async->transid)
2921                 goto end;
2922
2923         ret = btrfs_run_delayed_refs(trans, async->count);
2924         if (ret)
2925                 async->error = ret;
2926 end:
2927         ret = btrfs_end_transaction(trans);
2928         if (ret && !async->error)
2929                 async->error = ret;
2930 done:
2931         if (async->sync)
2932                 complete(&async->wait);
2933         else
2934                 kfree(async);
2935 }
2936
2937 int btrfs_async_run_delayed_refs(struct btrfs_fs_info *fs_info,
2938                                  unsigned long count, u64 transid, int wait)
2939 {
2940         struct async_delayed_refs *async;
2941         int ret;
2942
2943         async = kmalloc(sizeof(*async), GFP_NOFS);
2944         if (!async)
2945                 return -ENOMEM;
2946
2947         async->root = fs_info->tree_root;
2948         async->count = count;
2949         async->error = 0;
2950         async->transid = transid;
2951         if (wait)
2952                 async->sync = 1;
2953         else
2954                 async->sync = 0;
2955         init_completion(&async->wait);
2956
2957         btrfs_init_work(&async->work, btrfs_extent_refs_helper,
2958                         delayed_ref_async_start, NULL, NULL);
2959
2960         btrfs_queue_work(fs_info->extent_workers, &async->work);
2961
2962         if (wait) {
2963                 wait_for_completion(&async->wait);
2964                 ret = async->error;
2965                 kfree(async);
2966                 return ret;
2967         }
2968         return 0;
2969 }
2970
2971 /*
2972  * this starts processing the delayed reference count updates and
2973  * extent insertions we have queued up so far.  count can be
2974  * 0, which means to process everything in the tree at the start
2975  * of the run (but not newly added entries), or it can be some target
2976  * number you'd like to process.
2977  *
2978  * Returns 0 on success or if called with an aborted transaction
2979  * Returns <0 on error and aborts the transaction
2980  */
2981 int btrfs_run_delayed_refs(struct btrfs_trans_handle *trans,
2982                            unsigned long count)
2983 {
2984         struct btrfs_fs_info *fs_info = trans->fs_info;
2985         struct rb_node *node;
2986         struct btrfs_delayed_ref_root *delayed_refs;
2987         struct btrfs_delayed_ref_head *head;
2988         int ret;
2989         int run_all = count == (unsigned long)-1;
2990
2991         /* We'll clean this up in btrfs_cleanup_transaction */
2992         if (trans->aborted)
2993                 return 0;
2994
2995         if (test_bit(BTRFS_FS_CREATING_FREE_SPACE_TREE, &fs_info->flags))
2996                 return 0;
2997
2998         delayed_refs = &trans->transaction->delayed_refs;
2999         if (count == 0)
3000                 count = atomic_read(&delayed_refs->num_entries) * 2;
3001
3002 again:
3003 #ifdef SCRAMBLE_DELAYED_REFS
3004         delayed_refs->run_delayed_start = find_middle(&delayed_refs->root);
3005 #endif
3006         ret = __btrfs_run_delayed_refs(trans, count);
3007         if (ret < 0) {
3008                 btrfs_abort_transaction(trans, ret);
3009                 return ret;
3010         }
3011
3012         if (run_all) {
3013                 btrfs_create_pending_block_groups(trans);
3014
3015                 spin_lock(&delayed_refs->lock);
3016                 node = rb_first_cached(&delayed_refs->href_root);
3017                 if (!node) {
3018                         spin_unlock(&delayed_refs->lock);
3019                         goto out;
3020                 }
3021                 head = rb_entry(node, struct btrfs_delayed_ref_head,
3022                                 href_node);
3023                 refcount_inc(&head->refs);
3024                 spin_unlock(&delayed_refs->lock);
3025
3026                 /* Mutex was contended, block until it's released and retry. */
3027                 mutex_lock(&head->mutex);
3028                 mutex_unlock(&head->mutex);
3029
3030                 btrfs_put_delayed_ref_head(head);
3031                 cond_resched();
3032                 goto again;
3033         }
3034 out:
3035         return 0;
3036 }
3037
3038 int btrfs_set_disk_extent_flags(struct btrfs_trans_handle *trans,
3039                                 struct btrfs_fs_info *fs_info,
3040                                 u64 bytenr, u64 num_bytes, u64 flags,
3041                                 int level, int is_data)
3042 {
3043         struct btrfs_delayed_extent_op *extent_op;
3044         int ret;
3045
3046         extent_op = btrfs_alloc_delayed_extent_op();
3047         if (!extent_op)
3048                 return -ENOMEM;
3049
3050         extent_op->flags_to_set = flags;
3051         extent_op->update_flags = true;
3052         extent_op->update_key = false;
3053         extent_op->is_data = is_data ? true : false;
3054         extent_op->level = level;
3055
3056         ret = btrfs_add_delayed_extent_op(fs_info, trans, bytenr,
3057                                           num_bytes, extent_op);
3058         if (ret)
3059                 btrfs_free_delayed_extent_op(extent_op);
3060         return ret;
3061 }
3062
3063 static noinline int check_delayed_ref(struct btrfs_root *root,
3064                                       struct btrfs_path *path,
3065                                       u64 objectid, u64 offset, u64 bytenr)
3066 {
3067         struct btrfs_delayed_ref_head *head;
3068         struct btrfs_delayed_ref_node *ref;
3069         struct btrfs_delayed_data_ref *data_ref;
3070         struct btrfs_delayed_ref_root *delayed_refs;
3071         struct btrfs_transaction *cur_trans;
3072         struct rb_node *node;
3073         int ret = 0;
3074
3075         spin_lock(&root->fs_info->trans_lock);
3076         cur_trans = root->fs_info->running_transaction;
3077         if (cur_trans)
3078                 refcount_inc(&cur_trans->use_count);
3079         spin_unlock(&root->fs_info->trans_lock);
3080         if (!cur_trans)
3081                 return 0;
3082
3083         delayed_refs = &cur_trans->delayed_refs;
3084         spin_lock(&delayed_refs->lock);
3085         head = btrfs_find_delayed_ref_head(delayed_refs, bytenr);
3086         if (!head) {
3087                 spin_unlock(&delayed_refs->lock);
3088                 btrfs_put_transaction(cur_trans);
3089                 return 0;
3090         }
3091
3092         if (!mutex_trylock(&head->mutex)) {
3093                 refcount_inc(&head->refs);
3094                 spin_unlock(&delayed_refs->lock);
3095
3096                 btrfs_release_path(path);
3097
3098                 /*
3099                  * Mutex was contended, block until it's released and let
3100                  * caller try again
3101                  */
3102                 mutex_lock(&head->mutex);
3103                 mutex_unlock(&head->mutex);
3104                 btrfs_put_delayed_ref_head(head);
3105                 btrfs_put_transaction(cur_trans);
3106                 return -EAGAIN;
3107         }
3108         spin_unlock(&delayed_refs->lock);
3109
3110         spin_lock(&head->lock);
3111         /*
3112          * XXX: We should replace this with a proper search function in the
3113          * future.
3114          */
3115         for (node = rb_first_cached(&head->ref_tree); node;
3116              node = rb_next(node)) {
3117                 ref = rb_entry(node, struct btrfs_delayed_ref_node, ref_node);
3118                 /* If it's a shared ref we know a cross reference exists */
3119                 if (ref->type != BTRFS_EXTENT_DATA_REF_KEY) {
3120                         ret = 1;
3121                         break;
3122                 }
3123
3124                 data_ref = btrfs_delayed_node_to_data_ref(ref);
3125
3126                 /*
3127                  * If our ref doesn't match the one we're currently looking at
3128                  * then we have a cross reference.
3129                  */
3130                 if (data_ref->root != root->root_key.objectid ||
3131                     data_ref->objectid != objectid ||
3132                     data_ref->offset != offset) {
3133                         ret = 1;
3134                         break;
3135                 }
3136         }
3137         spin_unlock(&head->lock);
3138         mutex_unlock(&head->mutex);
3139         btrfs_put_transaction(cur_trans);
3140         return ret;
3141 }
3142
3143 static noinline int check_committed_ref(struct btrfs_root *root,
3144                                         struct btrfs_path *path,
3145                                         u64 objectid, u64 offset, u64 bytenr)
3146 {
3147         struct btrfs_fs_info *fs_info = root->fs_info;
3148         struct btrfs_root *extent_root = fs_info->extent_root;
3149         struct extent_buffer *leaf;
3150         struct btrfs_extent_data_ref *ref;
3151         struct btrfs_extent_inline_ref *iref;
3152         struct btrfs_extent_item *ei;
3153         struct btrfs_key key;
3154         u32 item_size;
3155         int type;
3156         int ret;
3157
3158         key.objectid = bytenr;
3159         key.offset = (u64)-1;
3160         key.type = BTRFS_EXTENT_ITEM_KEY;
3161
3162         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
3163         if (ret < 0)
3164                 goto out;
3165         BUG_ON(ret == 0); /* Corruption */
3166
3167         ret = -ENOENT;
3168         if (path->slots[0] == 0)
3169                 goto out;
3170
3171         path->slots[0]--;
3172         leaf = path->nodes[0];
3173         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
3174
3175         if (key.objectid != bytenr || key.type != BTRFS_EXTENT_ITEM_KEY)
3176                 goto out;
3177
3178         ret = 1;
3179         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
3180         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
3181
3182         if (item_size != sizeof(*ei) +
3183             btrfs_extent_inline_ref_size(BTRFS_EXTENT_DATA_REF_KEY))
3184                 goto out;
3185
3186         if (btrfs_extent_generation(leaf, ei) <=
3187             btrfs_root_last_snapshot(&root->root_item))
3188                 goto out;
3189
3190         iref = (struct btrfs_extent_inline_ref *)(ei + 1);
3191
3192         type = btrfs_get_extent_inline_ref_type(leaf, iref, BTRFS_REF_TYPE_DATA);
3193         if (type != BTRFS_EXTENT_DATA_REF_KEY)
3194                 goto out;
3195
3196         ref = (struct btrfs_extent_data_ref *)(&iref->offset);
3197         if (btrfs_extent_refs(leaf, ei) !=
3198             btrfs_extent_data_ref_count(leaf, ref) ||
3199             btrfs_extent_data_ref_root(leaf, ref) !=
3200             root->root_key.objectid ||
3201             btrfs_extent_data_ref_objectid(leaf, ref) != objectid ||
3202             btrfs_extent_data_ref_offset(leaf, ref) != offset)
3203                 goto out;
3204
3205         ret = 0;
3206 out:
3207         return ret;
3208 }
3209
3210 int btrfs_cross_ref_exist(struct btrfs_root *root, u64 objectid, u64 offset,
3211                           u64 bytenr)
3212 {
3213         struct btrfs_path *path;
3214         int ret;
3215
3216         path = btrfs_alloc_path();
3217         if (!path)
3218                 return -ENOMEM;
3219
3220         do {
3221                 ret = check_committed_ref(root, path, objectid,
3222                                           offset, bytenr);
3223                 if (ret && ret != -ENOENT)
3224                         goto out;
3225
3226                 ret = check_delayed_ref(root, path, objectid, offset, bytenr);
3227         } while (ret == -EAGAIN);
3228
3229 out:
3230         btrfs_free_path(path);
3231         if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID)
3232                 WARN_ON(ret > 0);
3233         return ret;
3234 }
3235
3236 static int __btrfs_mod_ref(struct btrfs_trans_handle *trans,
3237                            struct btrfs_root *root,
3238                            struct extent_buffer *buf,
3239                            int full_backref, int inc)
3240 {
3241         struct btrfs_fs_info *fs_info = root->fs_info;
3242         u64 bytenr;
3243         u64 num_bytes;
3244         u64 parent;
3245         u64 ref_root;
3246         u32 nritems;
3247         struct btrfs_key key;
3248         struct btrfs_file_extent_item *fi;
3249         int i;
3250         int level;
3251         int ret = 0;
3252         int (*process_func)(struct btrfs_trans_handle *,
3253                             struct btrfs_root *,
3254                             u64, u64, u64, u64, u64, u64);
3255
3256
3257         if (btrfs_is_testing(fs_info))
3258                 return 0;
3259
3260         ref_root = btrfs_header_owner(buf);
3261         nritems = btrfs_header_nritems(buf);
3262         level = btrfs_header_level(buf);
3263
3264         if (!test_bit(BTRFS_ROOT_REF_COWS, &root->state) && level == 0)
3265                 return 0;
3266
3267         if (inc)
3268                 process_func = btrfs_inc_extent_ref;
3269         else
3270                 process_func = btrfs_free_extent;
3271
3272         if (full_backref)
3273                 parent = buf->start;
3274         else
3275                 parent = 0;
3276
3277         for (i = 0; i < nritems; i++) {
3278                 if (level == 0) {
3279                         btrfs_item_key_to_cpu(buf, &key, i);
3280                         if (key.type != BTRFS_EXTENT_DATA_KEY)
3281                                 continue;
3282                         fi = btrfs_item_ptr(buf, i,
3283                                             struct btrfs_file_extent_item);
3284                         if (btrfs_file_extent_type(buf, fi) ==
3285                             BTRFS_FILE_EXTENT_INLINE)
3286                                 continue;
3287                         bytenr = btrfs_file_extent_disk_bytenr(buf, fi);
3288                         if (bytenr == 0)
3289                                 continue;
3290
3291                         num_bytes = btrfs_file_extent_disk_num_bytes(buf, fi);
3292                         key.offset -= btrfs_file_extent_offset(buf, fi);
3293                         ret = process_func(trans, root, bytenr, num_bytes,
3294                                            parent, ref_root, key.objectid,
3295                                            key.offset);
3296                         if (ret)
3297                                 goto fail;
3298                 } else {
3299                         bytenr = btrfs_node_blockptr(buf, i);
3300                         num_bytes = fs_info->nodesize;
3301                         ret = process_func(trans, root, bytenr, num_bytes,
3302                                            parent, ref_root, level - 1, 0);
3303                         if (ret)
3304                                 goto fail;
3305                 }
3306         }
3307         return 0;
3308 fail:
3309         return ret;
3310 }
3311
3312 int btrfs_inc_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
3313                   struct extent_buffer *buf, int full_backref)
3314 {
3315         return __btrfs_mod_ref(trans, root, buf, full_backref, 1);
3316 }
3317
3318 int btrfs_dec_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
3319                   struct extent_buffer *buf, int full_backref)
3320 {
3321         return __btrfs_mod_ref(trans, root, buf, full_backref, 0);
3322 }
3323
3324 static int write_one_cache_group(struct btrfs_trans_handle *trans,
3325                                  struct btrfs_fs_info *fs_info,
3326                                  struct btrfs_path *path,
3327                                  struct btrfs_block_group_cache *cache)
3328 {
3329         int ret;
3330         struct btrfs_root *extent_root = fs_info->extent_root;
3331         unsigned long bi;
3332         struct extent_buffer *leaf;
3333
3334         ret = btrfs_search_slot(trans, extent_root, &cache->key, path, 0, 1);
3335         if (ret) {
3336                 if (ret > 0)
3337                         ret = -ENOENT;
3338                 goto fail;
3339         }
3340
3341         leaf = path->nodes[0];
3342         bi = btrfs_item_ptr_offset(leaf, path->slots[0]);
3343         write_extent_buffer(leaf, &cache->item, bi, sizeof(cache->item));
3344         btrfs_mark_buffer_dirty(leaf);
3345 fail:
3346         btrfs_release_path(path);
3347         return ret;
3348
3349 }
3350
3351 static struct btrfs_block_group_cache *
3352 next_block_group(struct btrfs_fs_info *fs_info,
3353                  struct btrfs_block_group_cache *cache)
3354 {
3355         struct rb_node *node;
3356
3357         spin_lock(&fs_info->block_group_cache_lock);
3358
3359         /* If our block group was removed, we need a full search. */
3360         if (RB_EMPTY_NODE(&cache->cache_node)) {
3361                 const u64 next_bytenr = cache->key.objectid + cache->key.offset;
3362
3363                 spin_unlock(&fs_info->block_group_cache_lock);
3364                 btrfs_put_block_group(cache);
3365                 cache = btrfs_lookup_first_block_group(fs_info, next_bytenr); return cache;
3366         }
3367         node = rb_next(&cache->cache_node);
3368         btrfs_put_block_group(cache);
3369         if (node) {
3370                 cache = rb_entry(node, struct btrfs_block_group_cache,
3371                                  cache_node);
3372                 btrfs_get_block_group(cache);
3373         } else
3374                 cache = NULL;
3375         spin_unlock(&fs_info->block_group_cache_lock);
3376         return cache;
3377 }
3378
3379 static int cache_save_setup(struct btrfs_block_group_cache *block_group,
3380                             struct btrfs_trans_handle *trans,
3381                             struct btrfs_path *path)
3382 {
3383         struct btrfs_fs_info *fs_info = block_group->fs_info;
3384         struct btrfs_root *root = fs_info->tree_root;
3385         struct inode *inode = NULL;
3386         struct extent_changeset *data_reserved = NULL;
3387         u64 alloc_hint = 0;
3388         int dcs = BTRFS_DC_ERROR;
3389         u64 num_pages = 0;
3390         int retries = 0;
3391         int ret = 0;
3392
3393         /*
3394          * If this block group is smaller than 100 megs don't bother caching the
3395          * block group.
3396          */
3397         if (block_group->key.offset < (100 * SZ_1M)) {
3398                 spin_lock(&block_group->lock);
3399                 block_group->disk_cache_state = BTRFS_DC_WRITTEN;
3400                 spin_unlock(&block_group->lock);
3401                 return 0;
3402         }
3403
3404         if (trans->aborted)
3405                 return 0;
3406 again:
3407         inode = lookup_free_space_inode(fs_info, block_group, path);
3408         if (IS_ERR(inode) && PTR_ERR(inode) != -ENOENT) {
3409                 ret = PTR_ERR(inode);
3410                 btrfs_release_path(path);
3411                 goto out;
3412         }
3413
3414         if (IS_ERR(inode)) {
3415                 BUG_ON(retries);
3416                 retries++;
3417
3418                 if (block_group->ro)
3419                         goto out_free;
3420
3421                 ret = create_free_space_inode(fs_info, trans, block_group,
3422                                               path);
3423                 if (ret)
3424                         goto out_free;
3425                 goto again;
3426         }
3427
3428         /*
3429          * We want to set the generation to 0, that way if anything goes wrong
3430          * from here on out we know not to trust this cache when we load up next
3431          * time.
3432          */
3433         BTRFS_I(inode)->generation = 0;
3434         ret = btrfs_update_inode(trans, root, inode);
3435         if (ret) {
3436                 /*
3437                  * So theoretically we could recover from this, simply set the
3438                  * super cache generation to 0 so we know to invalidate the
3439                  * cache, but then we'd have to keep track of the block groups
3440                  * that fail this way so we know we _have_ to reset this cache
3441                  * before the next commit or risk reading stale cache.  So to
3442                  * limit our exposure to horrible edge cases lets just abort the
3443                  * transaction, this only happens in really bad situations
3444                  * anyway.
3445                  */
3446                 btrfs_abort_transaction(trans, ret);
3447                 goto out_put;
3448         }
3449         WARN_ON(ret);
3450
3451         /* We've already setup this transaction, go ahead and exit */
3452         if (block_group->cache_generation == trans->transid &&
3453             i_size_read(inode)) {
3454                 dcs = BTRFS_DC_SETUP;
3455                 goto out_put;
3456         }
3457
3458         if (i_size_read(inode) > 0) {
3459                 ret = btrfs_check_trunc_cache_free_space(fs_info,
3460                                         &fs_info->global_block_rsv);
3461                 if (ret)
3462                         goto out_put;
3463
3464                 ret = btrfs_truncate_free_space_cache(trans, NULL, inode);
3465                 if (ret)
3466                         goto out_put;
3467         }
3468
3469         spin_lock(&block_group->lock);
3470         if (block_group->cached != BTRFS_CACHE_FINISHED ||
3471             !btrfs_test_opt(fs_info, SPACE_CACHE)) {
3472                 /*
3473                  * don't bother trying to write stuff out _if_
3474                  * a) we're not cached,
3475                  * b) we're with nospace_cache mount option,
3476                  * c) we're with v2 space_cache (FREE_SPACE_TREE).
3477                  */
3478                 dcs = BTRFS_DC_WRITTEN;
3479                 spin_unlock(&block_group->lock);
3480                 goto out_put;
3481         }
3482         spin_unlock(&block_group->lock);
3483
3484         /*
3485          * We hit an ENOSPC when setting up the cache in this transaction, just
3486          * skip doing the setup, we've already cleared the cache so we're safe.
3487          */
3488         if (test_bit(BTRFS_TRANS_CACHE_ENOSPC, &trans->transaction->flags)) {
3489                 ret = -ENOSPC;
3490                 goto out_put;
3491         }
3492
3493         /*
3494          * Try to preallocate enough space based on how big the block group is.
3495          * Keep in mind this has to include any pinned space which could end up
3496          * taking up quite a bit since it's not folded into the other space
3497          * cache.
3498          */
3499         num_pages = div_u64(block_group->key.offset, SZ_256M);
3500         if (!num_pages)
3501                 num_pages = 1;
3502
3503         num_pages *= 16;
3504         num_pages *= PAGE_SIZE;
3505
3506         ret = btrfs_check_data_free_space(inode, &data_reserved, 0, num_pages);
3507         if (ret)
3508                 goto out_put;
3509
3510         ret = btrfs_prealloc_file_range_trans(inode, trans, 0, 0, num_pages,
3511                                               num_pages, num_pages,
3512                                               &alloc_hint);
3513         /*
3514          * Our cache requires contiguous chunks so that we don't modify a bunch
3515          * of metadata or split extents when writing the cache out, which means
3516          * we can enospc if we are heavily fragmented in addition to just normal
3517          * out of space conditions.  So if we hit this just skip setting up any
3518          * other block groups for this transaction, maybe we'll unpin enough
3519          * space the next time around.
3520          */
3521         if (!ret)
3522                 dcs = BTRFS_DC_SETUP;
3523         else if (ret == -ENOSPC)
3524                 set_bit(BTRFS_TRANS_CACHE_ENOSPC, &trans->transaction->flags);
3525
3526 out_put:
3527         iput(inode);
3528 out_free:
3529         btrfs_release_path(path);
3530 out:
3531         spin_lock(&block_group->lock);
3532         if (!ret && dcs == BTRFS_DC_SETUP)
3533                 block_group->cache_generation = trans->transid;
3534         block_group->disk_cache_state = dcs;
3535         spin_unlock(&block_group->lock);
3536
3537         extent_changeset_free(data_reserved);
3538         return ret;
3539 }
3540
3541 int btrfs_setup_space_cache(struct btrfs_trans_handle *trans,
3542                             struct btrfs_fs_info *fs_info)
3543 {
3544         struct btrfs_block_group_cache *cache, *tmp;
3545         struct btrfs_transaction *cur_trans = trans->transaction;
3546         struct btrfs_path *path;
3547
3548         if (list_empty(&cur_trans->dirty_bgs) ||
3549             !btrfs_test_opt(fs_info, SPACE_CACHE))
3550                 return 0;
3551
3552         path = btrfs_alloc_path();
3553         if (!path)
3554                 return -ENOMEM;
3555
3556         /* Could add new block groups, use _safe just in case */
3557         list_for_each_entry_safe(cache, tmp, &cur_trans->dirty_bgs,
3558                                  dirty_list) {
3559                 if (cache->disk_cache_state == BTRFS_DC_CLEAR)
3560                         cache_save_setup(cache, trans, path);
3561         }
3562
3563         btrfs_free_path(path);
3564         return 0;
3565 }
3566
3567 /*
3568  * transaction commit does final block group cache writeback during a
3569  * critical section where nothing is allowed to change the FS.  This is
3570  * required in order for the cache to actually match the block group,
3571  * but can introduce a lot of latency into the commit.
3572  *
3573  * So, btrfs_start_dirty_block_groups is here to kick off block group
3574  * cache IO.  There's a chance we'll have to redo some of it if the
3575  * block group changes again during the commit, but it greatly reduces
3576  * the commit latency by getting rid of the easy block groups while
3577  * we're still allowing others to join the commit.
3578  */
3579 int btrfs_start_dirty_block_groups(struct btrfs_trans_handle *trans)
3580 {
3581         struct btrfs_fs_info *fs_info = trans->fs_info;
3582         struct btrfs_block_group_cache *cache;
3583         struct btrfs_transaction *cur_trans = trans->transaction;
3584         int ret = 0;
3585         int should_put;
3586         struct btrfs_path *path = NULL;
3587         LIST_HEAD(dirty);
3588         struct list_head *io = &cur_trans->io_bgs;
3589         int num_started = 0;
3590         int loops = 0;
3591
3592         spin_lock(&cur_trans->dirty_bgs_lock);
3593         if (list_empty(&cur_trans->dirty_bgs)) {
3594                 spin_unlock(&cur_trans->dirty_bgs_lock);
3595                 return 0;
3596         }
3597         list_splice_init(&cur_trans->dirty_bgs, &dirty);
3598         spin_unlock(&cur_trans->dirty_bgs_lock);
3599
3600 again:
3601         /*
3602          * make sure all the block groups on our dirty list actually
3603          * exist
3604          */
3605         btrfs_create_pending_block_groups(trans);
3606
3607         if (!path) {
3608                 path = btrfs_alloc_path();
3609                 if (!path)
3610                         return -ENOMEM;
3611         }
3612
3613         /*
3614          * cache_write_mutex is here only to save us from balance or automatic
3615          * removal of empty block groups deleting this block group while we are
3616          * writing out the cache
3617          */
3618         mutex_lock(&trans->transaction->cache_write_mutex);
3619         while (!list_empty(&dirty)) {
3620                 bool drop_reserve = true;
3621
3622                 cache = list_first_entry(&dirty,
3623                                          struct btrfs_block_group_cache,
3624                                          dirty_list);
3625                 /*
3626                  * this can happen if something re-dirties a block
3627                  * group that is already under IO.  Just wait for it to
3628                  * finish and then do it all again
3629                  */
3630                 if (!list_empty(&cache->io_list)) {
3631                         list_del_init(&cache->io_list);
3632                         btrfs_wait_cache_io(trans, cache, path);
3633                         btrfs_put_block_group(cache);
3634                 }
3635
3636
3637                 /*
3638                  * btrfs_wait_cache_io uses the cache->dirty_list to decide
3639                  * if it should update the cache_state.  Don't delete
3640                  * until after we wait.
3641                  *
3642                  * Since we're not running in the commit critical section
3643                  * we need the dirty_bgs_lock to protect from update_block_group
3644                  */
3645                 spin_lock(&cur_trans->dirty_bgs_lock);
3646                 list_del_init(&cache->dirty_list);
3647                 spin_unlock(&cur_trans->dirty_bgs_lock);
3648
3649                 should_put = 1;
3650
3651                 cache_save_setup(cache, trans, path);
3652
3653                 if (cache->disk_cache_state == BTRFS_DC_SETUP) {
3654                         cache->io_ctl.inode = NULL;
3655                         ret = btrfs_write_out_cache(fs_info, trans,
3656                                                     cache, path);
3657                         if (ret == 0 && cache->io_ctl.inode) {
3658                                 num_started++;
3659                                 should_put = 0;
3660
3661                                 /*
3662                                  * The cache_write_mutex is protecting the
3663                                  * io_list, also refer to the definition of
3664                                  * btrfs_transaction::io_bgs for more details
3665                                  */
3666                                 list_add_tail(&cache->io_list, io);
3667                         } else {
3668                                 /*
3669                                  * if we failed to write the cache, the
3670                                  * generation will be bad and life goes on
3671                                  */
3672                                 ret = 0;
3673                         }
3674                 }
3675                 if (!ret) {
3676                         ret = write_one_cache_group(trans, fs_info,
3677                                                     path, cache);
3678                         /*
3679                          * Our block group might still be attached to the list
3680                          * of new block groups in the transaction handle of some
3681                          * other task (struct btrfs_trans_handle->new_bgs). This
3682                          * means its block group item isn't yet in the extent
3683                          * tree. If this happens ignore the error, as we will
3684                          * try again later in the critical section of the
3685                          * transaction commit.
3686                          */
3687                         if (ret == -ENOENT) {
3688                                 ret = 0;
3689                                 spin_lock(&cur_trans->dirty_bgs_lock);
3690                                 if (list_empty(&cache->dirty_list)) {
3691                                         list_add_tail(&cache->dirty_list,
3692                                                       &cur_trans->dirty_bgs);
3693                                         btrfs_get_block_group(cache);
3694                                         drop_reserve = false;
3695                                 }
3696                                 spin_unlock(&cur_trans->dirty_bgs_lock);
3697                         } else if (ret) {
3698                                 btrfs_abort_transaction(trans, ret);
3699                         }
3700                 }
3701
3702                 /* if it's not on the io list, we need to put the block group */
3703                 if (should_put)
3704                         btrfs_put_block_group(cache);
3705                 if (drop_reserve)
3706                         btrfs_delayed_refs_rsv_release(fs_info, 1);
3707
3708                 if (ret)
3709                         break;
3710
3711                 /*
3712                  * Avoid blocking other tasks for too long. It might even save
3713                  * us from writing caches for block groups that are going to be
3714                  * removed.
3715                  */
3716                 mutex_unlock(&trans->transaction->cache_write_mutex);
3717                 mutex_lock(&trans->transaction->cache_write_mutex);
3718         }
3719         mutex_unlock(&trans->transaction->cache_write_mutex);
3720
3721         /*
3722          * go through delayed refs for all the stuff we've just kicked off
3723          * and then loop back (just once)
3724          */
3725         ret = btrfs_run_delayed_refs(trans, 0);
3726         if (!ret && loops == 0) {
3727                 loops++;
3728                 spin_lock(&cur_trans->dirty_bgs_lock);
3729                 list_splice_init(&cur_trans->dirty_bgs, &dirty);
3730                 /*
3731                  * dirty_bgs_lock protects us from concurrent block group
3732                  * deletes too (not just cache_write_mutex).
3733                  */
3734                 if (!list_empty(&dirty)) {
3735                         spin_unlock(&cur_trans->dirty_bgs_lock);
3736                         goto again;
3737                 }
3738                 spin_unlock(&cur_trans->dirty_bgs_lock);
3739         } else if (ret < 0) {
3740                 btrfs_cleanup_dirty_bgs(cur_trans, fs_info);
3741         }
3742
3743         btrfs_free_path(path);
3744         return ret;
3745 }
3746
3747 int btrfs_write_dirty_block_groups(struct btrfs_trans_handle *trans,
3748                                    struct btrfs_fs_info *fs_info)
3749 {
3750         struct btrfs_block_group_cache *cache;
3751         struct btrfs_transaction *cur_trans = trans->transaction;
3752         int ret = 0;
3753         int should_put;
3754         struct btrfs_path *path;
3755         struct list_head *io = &cur_trans->io_bgs;
3756         int num_started = 0;
3757
3758         path = btrfs_alloc_path();
3759         if (!path)
3760                 return -ENOMEM;
3761
3762         /*
3763          * Even though we are in the critical section of the transaction commit,
3764          * we can still have concurrent tasks adding elements to this
3765          * transaction's list of dirty block groups. These tasks correspond to
3766          * endio free space workers started when writeback finishes for a
3767          * space cache, which run inode.c:btrfs_finish_ordered_io(), and can
3768          * allocate new block groups as a result of COWing nodes of the root
3769          * tree when updating the free space inode. The writeback for the space
3770          * caches is triggered by an earlier call to
3771          * btrfs_start_dirty_block_groups() and iterations of the following
3772          * loop.
3773          * Also we want to do the cache_save_setup first and then run the
3774          * delayed refs to make sure we have the best chance at doing this all
3775          * in one shot.
3776          */
3777         spin_lock(&cur_trans->dirty_bgs_lock);
3778         while (!list_empty(&cur_trans->dirty_bgs)) {
3779                 cache = list_first_entry(&cur_trans->dirty_bgs,
3780                                          struct btrfs_block_group_cache,
3781                                          dirty_list);
3782
3783                 /*
3784                  * this can happen if cache_save_setup re-dirties a block
3785                  * group that is already under IO.  Just wait for it to
3786                  * finish and then do it all again
3787                  */
3788                 if (!list_empty(&cache->io_list)) {
3789                         spin_unlock(&cur_trans->dirty_bgs_lock);
3790                         list_del_init(&cache->io_list);
3791                         btrfs_wait_cache_io(trans, cache, path);
3792                         btrfs_put_block_group(cache);
3793                         spin_lock(&cur_trans->dirty_bgs_lock);
3794                 }
3795
3796                 /*
3797                  * don't remove from the dirty list until after we've waited
3798                  * on any pending IO
3799                  */
3800                 list_del_init(&cache->dirty_list);
3801                 spin_unlock(&cur_trans->dirty_bgs_lock);
3802                 should_put = 1;
3803
3804                 cache_save_setup(cache, trans, path);
3805
3806                 if (!ret)
3807                         ret = btrfs_run_delayed_refs(trans,
3808                                                      (unsigned long) -1);
3809
3810                 if (!ret && cache->disk_cache_state == BTRFS_DC_SETUP) {
3811                         cache->io_ctl.inode = NULL;
3812                         ret = btrfs_write_out_cache(fs_info, trans,
3813                                                     cache, path);
3814                         if (ret == 0 && cache->io_ctl.inode) {
3815                                 num_started++;
3816                                 should_put = 0;
3817                                 list_add_tail(&cache->io_list, io);
3818                         } else {
3819                                 /*
3820                                  * if we failed to write the cache, the
3821                                  * generation will be bad and life goes on
3822                                  */
3823                                 ret = 0;
3824                         }
3825                 }
3826                 if (!ret) {
3827                         ret = write_one_cache_group(trans, fs_info,
3828                                                     path, cache);
3829                         /*
3830                          * One of the free space endio workers might have
3831                          * created a new block group while updating a free space
3832                          * cache's inode (at inode.c:btrfs_finish_ordered_io())
3833                          * and hasn't released its transaction handle yet, in
3834                          * which case the new block group is still attached to
3835                          * its transaction handle and its creation has not
3836                          * finished yet (no block group item in the extent tree
3837                          * yet, etc). If this is the case, wait for all free
3838                          * space endio workers to finish and retry. This is a
3839                          * a very rare case so no need for a more efficient and
3840                          * complex approach.
3841                          */
3842                         if (ret == -ENOENT) {
3843                                 wait_event(cur_trans->writer_wait,
3844                                    atomic_read(&cur_trans->num_writers) == 1);
3845                                 ret = write_one_cache_group(trans, fs_info,
3846                                                             path, cache);
3847                         }
3848                         if (ret)
3849                                 btrfs_abort_transaction(trans, ret);
3850                 }
3851
3852                 /* if its not on the io list, we need to put the block group */
3853                 if (should_put)
3854                         btrfs_put_block_group(cache);
3855                 btrfs_delayed_refs_rsv_release(fs_info, 1);
3856                 spin_lock(&cur_trans->dirty_bgs_lock);
3857         }
3858         spin_unlock(&cur_trans->dirty_bgs_lock);
3859
3860         /*
3861          * Refer to the definition of io_bgs member for details why it's safe
3862          * to use it without any locking
3863          */
3864         while (!list_empty(io)) {
3865                 cache = list_first_entry(io, struct btrfs_block_group_cache,
3866                                          io_list);
3867                 list_del_init(&cache->io_list);
3868                 btrfs_wait_cache_io(trans, cache, path);
3869                 btrfs_put_block_group(cache);
3870         }
3871
3872         btrfs_free_path(path);
3873         return ret;
3874 }
3875
3876 int btrfs_extent_readonly(struct btrfs_fs_info *fs_info, u64 bytenr)
3877 {
3878         struct btrfs_block_group_cache *block_group;
3879         int readonly = 0;
3880
3881         block_group = btrfs_lookup_block_group(fs_info, bytenr);
3882         if (!block_group || block_group->ro)
3883                 readonly = 1;
3884         if (block_group)
3885                 btrfs_put_block_group(block_group);
3886         return readonly;
3887 }
3888
3889 bool btrfs_inc_nocow_writers(struct btrfs_fs_info *fs_info, u64 bytenr)
3890 {
3891         struct btrfs_block_group_cache *bg;
3892         bool ret = true;
3893
3894         bg = btrfs_lookup_block_group(fs_info, bytenr);
3895         if (!bg)
3896                 return false;
3897
3898         spin_lock(&bg->lock);
3899         if (bg->ro)
3900                 ret = false;
3901         else
3902                 atomic_inc(&bg->nocow_writers);
3903         spin_unlock(&bg->lock);
3904
3905         /* no put on block group, done by btrfs_dec_nocow_writers */
3906         if (!ret)
3907                 btrfs_put_block_group(bg);
3908
3909         return ret;
3910
3911 }
3912
3913 void btrfs_dec_nocow_writers(struct btrfs_fs_info *fs_info, u64 bytenr)
3914 {
3915         struct btrfs_block_group_cache *bg;
3916
3917         bg = btrfs_lookup_block_group(fs_info, bytenr);
3918         ASSERT(bg);
3919         if (atomic_dec_and_test(&bg->nocow_writers))
3920                 wake_up_var(&bg->nocow_writers);
3921         /*
3922          * Once for our lookup and once for the lookup done by a previous call
3923          * to btrfs_inc_nocow_writers()
3924          */
3925         btrfs_put_block_group(bg);
3926         btrfs_put_block_group(bg);
3927 }
3928
3929 void btrfs_wait_nocow_writers(struct btrfs_block_group_cache *bg)
3930 {
3931         wait_var_event(&bg->nocow_writers, !atomic_read(&bg->nocow_writers));
3932 }
3933
3934 static const char *alloc_name(u64 flags)
3935 {
3936         switch (flags) {
3937         case BTRFS_BLOCK_GROUP_METADATA|BTRFS_BLOCK_GROUP_DATA:
3938                 return "mixed";
3939         case BTRFS_BLOCK_GROUP_METADATA:
3940                 return "metadata";
3941         case BTRFS_BLOCK_GROUP_DATA:
3942                 return "data";
3943         case BTRFS_BLOCK_GROUP_SYSTEM:
3944                 return "system";
3945         default:
3946                 WARN_ON(1);
3947                 return "invalid-combination";
3948         };
3949 }
3950
3951 static int create_space_info(struct btrfs_fs_info *info, u64 flags)
3952 {
3953
3954         struct btrfs_space_info *space_info;
3955         int i;
3956         int ret;
3957
3958         space_info = kzalloc(sizeof(*space_info), GFP_NOFS);
3959         if (!space_info)
3960                 return -ENOMEM;
3961
3962         ret = percpu_counter_init(&space_info->total_bytes_pinned, 0,
3963                                  GFP_KERNEL);
3964         if (ret) {
3965                 kfree(space_info);
3966                 return ret;
3967         }
3968
3969         for (i = 0; i < BTRFS_NR_RAID_TYPES; i++)
3970                 INIT_LIST_HEAD(&space_info->block_groups[i]);
3971         init_rwsem(&space_info->groups_sem);
3972         spin_lock_init(&space_info->lock);
3973         space_info->flags = flags & BTRFS_BLOCK_GROUP_TYPE_MASK;
3974         space_info->force_alloc = CHUNK_ALLOC_NO_FORCE;
3975         init_waitqueue_head(&space_info->wait);
3976         INIT_LIST_HEAD(&space_info->ro_bgs);
3977         INIT_LIST_HEAD(&space_info->tickets);
3978         INIT_LIST_HEAD(&space_info->priority_tickets);
3979
3980         ret = kobject_init_and_add(&space_info->kobj, &space_info_ktype,
3981                                     info->space_info_kobj, "%s",
3982                                     alloc_name(space_info->flags));
3983         if (ret) {
3984                 percpu_counter_destroy(&space_info->total_bytes_pinned);
3985                 kfree(space_info);
3986                 return ret;
3987         }
3988
3989         list_add_rcu(&space_info->list, &info->space_info);
3990         if (flags & BTRFS_BLOCK_GROUP_DATA)
3991                 info->data_sinfo = space_info;
3992
3993         return ret;
3994 }
3995
3996 static void update_space_info(struct btrfs_fs_info *info, u64 flags,
3997                              u64 total_bytes, u64 bytes_used,
3998                              u64 bytes_readonly,
3999                              struct btrfs_space_info **space_info)
4000 {
4001         struct btrfs_space_info *found;
4002         int factor;
4003
4004         factor = btrfs_bg_type_to_factor(flags);
4005
4006         found = __find_space_info(info, flags);
4007         ASSERT(found);
4008         spin_lock(&found->lock);
4009         found->total_bytes += total_bytes;
4010         found->disk_total += total_bytes * factor;
4011         found->bytes_used += bytes_used;
4012         found->disk_used += bytes_used * factor;
4013         found->bytes_readonly += bytes_readonly;
4014         if (total_bytes > 0)
4015                 found->full = 0;
4016         space_info_add_new_bytes(info, found, total_bytes -
4017                                  bytes_used - bytes_readonly);
4018         spin_unlock(&found->lock);
4019         *space_info = found;
4020 }
4021
4022 static void set_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
4023 {
4024         u64 extra_flags = chunk_to_extended(flags) &
4025                                 BTRFS_EXTENDED_PROFILE_MASK;
4026
4027         write_seqlock(&fs_info->profiles_lock);
4028         if (flags & BTRFS_BLOCK_GROUP_DATA)
4029                 fs_info->avail_data_alloc_bits |= extra_flags;
4030         if (flags & BTRFS_BLOCK_GROUP_METADATA)
4031                 fs_info->avail_metadata_alloc_bits |= extra_flags;
4032         if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
4033                 fs_info->avail_system_alloc_bits |= extra_flags;
4034         write_sequnlock(&fs_info->profiles_lock);
4035 }
4036
4037 /*
4038  * returns target flags in extended format or 0 if restripe for this
4039  * chunk_type is not in progress
4040  *
4041  * should be called with balance_lock held
4042  */
4043 static u64 get_restripe_target(struct btrfs_fs_info *fs_info, u64 flags)
4044 {
4045         struct btrfs_balance_control *bctl = fs_info->balance_ctl;
4046         u64 target = 0;
4047
4048         if (!bctl)
4049                 return 0;
4050
4051         if (flags & BTRFS_BLOCK_GROUP_DATA &&
4052             bctl->data.flags & BTRFS_BALANCE_ARGS_CONVERT) {
4053                 target = BTRFS_BLOCK_GROUP_DATA | bctl->data.target;
4054         } else if (flags & BTRFS_BLOCK_GROUP_SYSTEM &&
4055                    bctl->sys.flags & BTRFS_BALANCE_ARGS_CONVERT) {
4056                 target = BTRFS_BLOCK_GROUP_SYSTEM | bctl->sys.target;
4057         } else if (flags & BTRFS_BLOCK_GROUP_METADATA &&
4058                    bctl->meta.flags & BTRFS_BALANCE_ARGS_CONVERT) {
4059                 target = BTRFS_BLOCK_GROUP_METADATA | bctl->meta.target;
4060         }
4061
4062         return target;
4063 }
4064
4065 /*
4066  * @flags: available profiles in extended format (see ctree.h)
4067  *
4068  * Returns reduced profile in chunk format.  If profile changing is in
4069  * progress (either running or paused) picks the target profile (if it's
4070  * already available), otherwise falls back to plain reducing.
4071  */
4072 static u64 btrfs_reduce_alloc_profile(struct btrfs_fs_info *fs_info, u64 flags)
4073 {
4074         u64 num_devices = fs_info->fs_devices->rw_devices;
4075         u64 target;
4076         u64 raid_type;
4077         u64 allowed = 0;
4078
4079         /*
4080          * see if restripe for this chunk_type is in progress, if so
4081          * try to reduce to the target profile
4082          */
4083         spin_lock(&fs_info->balance_lock);
4084         target = get_restripe_target(fs_info, flags);
4085         if (target) {
4086                 /* pick target profile only if it's already available */
4087                 if ((flags & target) & BTRFS_EXTENDED_PROFILE_MASK) {
4088                         spin_unlock(&fs_info->balance_lock);
4089                         return extended_to_chunk(target);
4090                 }
4091         }
4092         spin_unlock(&fs_info->balance_lock);
4093
4094         /* First, mask out the RAID levels which aren't possible */
4095         for (raid_type = 0; raid_type < BTRFS_NR_RAID_TYPES; raid_type++) {
4096                 if (num_devices >= btrfs_raid_array[raid_type].devs_min)
4097                         allowed |= btrfs_raid_array[raid_type].bg_flag;
4098         }
4099         allowed &= flags;
4100
4101         if (allowed & BTRFS_BLOCK_GROUP_RAID6)
4102                 allowed = BTRFS_BLOCK_GROUP_RAID6;
4103         else if (allowed & BTRFS_BLOCK_GROUP_RAID5)
4104                 allowed = BTRFS_BLOCK_GROUP_RAID5;
4105         else if (allowed & BTRFS_BLOCK_GROUP_RAID10)
4106                 allowed = BTRFS_BLOCK_GROUP_RAID10;
4107         else if (allowed & BTRFS_BLOCK_GROUP_RAID1)
4108                 allowed = BTRFS_BLOCK_GROUP_RAID1;
4109         else if (allowed & BTRFS_BLOCK_GROUP_RAID0)
4110                 allowed = BTRFS_BLOCK_GROUP_RAID0;
4111
4112         flags &= ~BTRFS_BLOCK_GROUP_PROFILE_MASK;
4113
4114         return extended_to_chunk(flags | allowed);
4115 }
4116
4117 static u64 get_alloc_profile(struct btrfs_fs_info *fs_info, u64 orig_flags)
4118 {
4119         unsigned seq;
4120         u64 flags;
4121
4122         do {
4123                 flags = orig_flags;
4124                 seq = read_seqbegin(&fs_info->profiles_lock);
4125
4126                 if (flags & BTRFS_BLOCK_GROUP_DATA)
4127                         flags |= fs_info->avail_data_alloc_bits;
4128                 else if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
4129                         flags |= fs_info->avail_system_alloc_bits;
4130                 else if (flags & BTRFS_BLOCK_GROUP_METADATA)
4131                         flags |= fs_info->avail_metadata_alloc_bits;
4132         } while (read_seqretry(&fs_info->profiles_lock, seq));
4133
4134         return btrfs_reduce_alloc_profile(fs_info, flags);
4135 }
4136
4137 static u64 get_alloc_profile_by_root(struct btrfs_root *root, int data)
4138 {
4139         struct btrfs_fs_info *fs_info = root->fs_info;
4140         u64 flags;
4141         u64 ret;
4142
4143         if (data)
4144                 flags = BTRFS_BLOCK_GROUP_DATA;
4145         else if (root == fs_info->chunk_root)
4146                 flags = BTRFS_BLOCK_GROUP_SYSTEM;
4147         else
4148                 flags = BTRFS_BLOCK_GROUP_METADATA;
4149
4150         ret = get_alloc_profile(fs_info, flags);
4151         return ret;
4152 }
4153
4154 u64 btrfs_data_alloc_profile(struct btrfs_fs_info *fs_info)
4155 {
4156         return get_alloc_profile(fs_info, BTRFS_BLOCK_GROUP_DATA);
4157 }
4158
4159 u64 btrfs_metadata_alloc_profile(struct btrfs_fs_info *fs_info)
4160 {
4161         return get_alloc_profile(fs_info, BTRFS_BLOCK_GROUP_METADATA);
4162 }
4163
4164 u64 btrfs_system_alloc_profile(struct btrfs_fs_info *fs_info)
4165 {
4166         return get_alloc_profile(fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
4167 }
4168
4169 static u64 btrfs_space_info_used(struct btrfs_space_info *s_info,
4170                                  bool may_use_included)
4171 {
4172         ASSERT(s_info);
4173         return s_info->bytes_used + s_info->bytes_reserved +
4174                 s_info->bytes_pinned + s_info->bytes_readonly +
4175                 (may_use_included ? s_info->bytes_may_use : 0);
4176 }
4177
4178 int btrfs_alloc_data_chunk_ondemand(struct btrfs_inode *inode, u64 bytes)
4179 {
4180         struct btrfs_root *root = inode->root;
4181         struct btrfs_fs_info *fs_info = root->fs_info;
4182         struct btrfs_space_info *data_sinfo = fs_info->data_sinfo;
4183         u64 used;
4184         int ret = 0;
4185         int need_commit = 2;
4186         int have_pinned_space;
4187
4188         /* make sure bytes are sectorsize aligned */
4189         bytes = ALIGN(bytes, fs_info->sectorsize);
4190
4191         if (btrfs_is_free_space_inode(inode)) {
4192                 need_commit = 0;
4193                 ASSERT(current->journal_info);
4194         }
4195
4196 again:
4197         /* make sure we have enough space to handle the data first */
4198         spin_lock(&data_sinfo->lock);
4199         used = btrfs_space_info_used(data_sinfo, true);
4200
4201         if (used + bytes > data_sinfo->total_bytes) {
4202                 struct btrfs_trans_handle *trans;
4203
4204                 /*
4205                  * if we don't have enough free bytes in this space then we need
4206                  * to alloc a new chunk.
4207                  */
4208                 if (!data_sinfo->full) {
4209                         u64 alloc_target;
4210
4211                         data_sinfo->force_alloc = CHUNK_ALLOC_FORCE;
4212                         spin_unlock(&data_sinfo->lock);
4213
4214                         alloc_target = btrfs_data_alloc_profile(fs_info);
4215                         /*
4216                          * It is ugly that we don't call nolock join
4217                          * transaction for the free space inode case here.
4218                          * But it is safe because we only do the data space
4219                          * reservation for the free space cache in the
4220                          * transaction context, the common join transaction
4221                          * just increase the counter of the current transaction
4222                          * handler, doesn't try to acquire the trans_lock of
4223                          * the fs.
4224                          */
4225                         trans = btrfs_join_transaction(root);
4226                         if (IS_ERR(trans))
4227                                 return PTR_ERR(trans);
4228
4229                         ret = do_chunk_alloc(trans, alloc_target,
4230                                              CHUNK_ALLOC_NO_FORCE);
4231                         btrfs_end_transaction(trans);
4232                         if (ret < 0) {
4233                                 if (ret != -ENOSPC)
4234                                         return ret;
4235                                 else {
4236                                         have_pinned_space = 1;
4237                                         goto commit_trans;
4238                                 }
4239                         }
4240
4241                         goto again;
4242                 }
4243
4244                 /*
4245                  * If we don't have enough pinned space to deal with this
4246                  * allocation, and no removed chunk in current transaction,
4247                  * don't bother committing the transaction.
4248                  */
4249                 have_pinned_space = __percpu_counter_compare(
4250                         &data_sinfo->total_bytes_pinned,
4251                         used + bytes - data_sinfo->total_bytes,
4252                         BTRFS_TOTAL_BYTES_PINNED_BATCH);
4253                 spin_unlock(&data_sinfo->lock);
4254
4255                 /* commit the current transaction and try again */
4256 commit_trans:
4257                 if (need_commit) {
4258                         need_commit--;
4259
4260                         if (need_commit > 0) {
4261                                 btrfs_start_delalloc_roots(fs_info, -1);
4262                                 btrfs_wait_ordered_roots(fs_info, U64_MAX, 0,
4263                                                          (u64)-1);
4264                         }
4265
4266                         trans = btrfs_join_transaction(root);
4267                         if (IS_ERR(trans))
4268                                 return PTR_ERR(trans);
4269                         if (have_pinned_space >= 0 ||
4270                             test_bit(BTRFS_TRANS_HAVE_FREE_BGS,
4271                                      &trans->transaction->flags) ||
4272                             need_commit > 0) {
4273                                 ret = btrfs_commit_transaction(trans);
4274                                 if (ret)
4275                                         return ret;
4276                                 /*
4277                                  * The cleaner kthread might still be doing iput
4278                                  * operations. Wait for it to finish so that
4279                                  * more space is released.  We don't need to
4280                                  * explicitly run the delayed iputs here because
4281                                  * the commit_transaction would have woken up
4282                                  * the cleaner.
4283                                  */
4284                                 ret = btrfs_wait_on_delayed_iputs(fs_info);
4285                                 if (ret)
4286                                         return ret;
4287                                 goto again;
4288                         } else {
4289                                 btrfs_end_transaction(trans);
4290                         }
4291                 }
4292
4293                 trace_btrfs_space_reservation(fs_info,
4294                                               "space_info:enospc",
4295                                               data_sinfo->flags, bytes, 1);
4296                 return -ENOSPC;
4297         }
4298         update_bytes_may_use(data_sinfo, bytes);
4299         trace_btrfs_space_reservation(fs_info, "space_info",
4300                                       data_sinfo->flags, bytes, 1);
4301         spin_unlock(&data_sinfo->lock);
4302
4303         return 0;
4304 }
4305
4306 int btrfs_check_data_free_space(struct inode *inode,
4307                         struct extent_changeset **reserved, u64 start, u64 len)
4308 {
4309         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
4310         int ret;
4311
4312         /* align the range */
4313         len = round_up(start + len, fs_info->sectorsize) -
4314               round_down(start, fs_info->sectorsize);
4315         start = round_down(start, fs_info->sectorsize);
4316
4317         ret = btrfs_alloc_data_chunk_ondemand(BTRFS_I(inode), len);
4318         if (ret < 0)
4319                 return ret;
4320
4321         /* Use new btrfs_qgroup_reserve_data to reserve precious data space. */
4322         ret = btrfs_qgroup_reserve_data(inode, reserved, start, len);
4323         if (ret < 0)
4324                 btrfs_free_reserved_data_space_noquota(inode, start, len);
4325         else
4326                 ret = 0;
4327         return ret;
4328 }
4329
4330 /*
4331  * Called if we need to clear a data reservation for this inode
4332  * Normally in a error case.
4333  *
4334  * This one will *NOT* use accurate qgroup reserved space API, just for case
4335  * which we can't sleep and is sure it won't affect qgroup reserved space.
4336  * Like clear_bit_hook().
4337  */
4338 void btrfs_free_reserved_data_space_noquota(struct inode *inode, u64 start,
4339                                             u64 len)
4340 {
4341         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
4342         struct btrfs_space_info *data_sinfo;
4343
4344         /* Make sure the range is aligned to sectorsize */
4345         len = round_up(start + len, fs_info->sectorsize) -
4346               round_down(start, fs_info->sectorsize);
4347         start = round_down(start, fs_info->sectorsize);
4348
4349         data_sinfo = fs_info->data_sinfo;
4350         spin_lock(&data_sinfo->lock);
4351         update_bytes_may_use(data_sinfo, -len);
4352         trace_btrfs_space_reservation(fs_info, "space_info",
4353                                       data_sinfo->flags, len, 0);
4354         spin_unlock(&data_sinfo->lock);
4355 }
4356
4357 /*
4358  * Called if we need to clear a data reservation for this inode
4359  * Normally in a error case.
4360  *
4361  * This one will handle the per-inode data rsv map for accurate reserved
4362  * space framework.
4363  */
4364 void btrfs_free_reserved_data_space(struct inode *inode,
4365                         struct extent_changeset *reserved, u64 start, u64 len)
4366 {
4367         struct btrfs_root *root = BTRFS_I(inode)->root;
4368
4369         /* Make sure the range is aligned to sectorsize */
4370         len = round_up(start + len, root->fs_info->sectorsize) -
4371               round_down(start, root->fs_info->sectorsize);
4372         start = round_down(start, root->fs_info->sectorsize);
4373
4374         btrfs_free_reserved_data_space_noquota(inode, start, len);
4375         btrfs_qgroup_free_data(inode, reserved, start, len);
4376 }
4377
4378 static void force_metadata_allocation(struct btrfs_fs_info *info)
4379 {
4380         struct list_head *head = &info->space_info;
4381         struct btrfs_space_info *found;
4382
4383         rcu_read_lock();
4384         list_for_each_entry_rcu(found, head, list) {
4385                 if (found->flags & BTRFS_BLOCK_GROUP_METADATA)
4386                         found->force_alloc = CHUNK_ALLOC_FORCE;
4387         }
4388         rcu_read_unlock();
4389 }
4390
4391 static inline u64 calc_global_rsv_need_space(struct btrfs_block_rsv *global)
4392 {
4393         return (global->size << 1);
4394 }
4395
4396 static int should_alloc_chunk(struct btrfs_fs_info *fs_info,
4397                               struct btrfs_space_info *sinfo, int force)
4398 {
4399         u64 bytes_used = btrfs_space_info_used(sinfo, false);
4400         u64 thresh;
4401
4402         if (force == CHUNK_ALLOC_FORCE)
4403                 return 1;
4404
4405         /*
4406          * in limited mode, we want to have some free space up to
4407          * about 1% of the FS size.
4408          */
4409         if (force == CHUNK_ALLOC_LIMITED) {
4410                 thresh = btrfs_super_total_bytes(fs_info->super_copy);
4411                 thresh = max_t(u64, SZ_64M, div_factor_fine(thresh, 1));
4412
4413                 if (sinfo->total_bytes - bytes_used < thresh)
4414                         return 1;
4415         }
4416
4417         if (bytes_used + SZ_2M < div_factor(sinfo->total_bytes, 8))
4418                 return 0;
4419         return 1;
4420 }
4421
4422 static u64 get_profile_num_devs(struct btrfs_fs_info *fs_info, u64 type)
4423 {
4424         u64 num_dev;
4425
4426         if (type & (BTRFS_BLOCK_GROUP_RAID10 |
4427                     BTRFS_BLOCK_GROUP_RAID0 |
4428                     BTRFS_BLOCK_GROUP_RAID5 |
4429                     BTRFS_BLOCK_GROUP_RAID6))
4430                 num_dev = fs_info->fs_devices->rw_devices;
4431         else if (type & BTRFS_BLOCK_GROUP_RAID1)
4432                 num_dev = 2;
4433         else
4434                 num_dev = 1;    /* DUP or single */
4435
4436         return num_dev;
4437 }
4438
4439 /*
4440  * If @is_allocation is true, reserve space in the system space info necessary
4441  * for allocating a chunk, otherwise if it's false, reserve space necessary for
4442  * removing a chunk.
4443  */
4444 void check_system_chunk(struct btrfs_trans_handle *trans, u64 type)
4445 {
4446         struct btrfs_fs_info *fs_info = trans->fs_info;
4447         struct btrfs_space_info *info;
4448         u64 left;
4449         u64 thresh;
4450         int ret = 0;
4451         u64 num_devs;
4452
4453         /*
4454          * Needed because we can end up allocating a system chunk and for an
4455          * atomic and race free space reservation in the chunk block reserve.
4456          */
4457         lockdep_assert_held(&fs_info->chunk_mutex);
4458
4459         info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
4460         spin_lock(&info->lock);
4461         left = info->total_bytes - btrfs_space_info_used(info, true);
4462         spin_unlock(&info->lock);
4463
4464         num_devs = get_profile_num_devs(fs_info, type);
4465
4466         /* num_devs device items to update and 1 chunk item to add or remove */
4467         thresh = btrfs_calc_trunc_metadata_size(fs_info, num_devs) +
4468                 btrfs_calc_trans_metadata_size(fs_info, 1);
4469
4470         if (left < thresh && btrfs_test_opt(fs_info, ENOSPC_DEBUG)) {
4471                 btrfs_info(fs_info, "left=%llu, need=%llu, flags=%llu",
4472                            left, thresh, type);
4473                 dump_space_info(fs_info, info, 0, 0);
4474         }
4475
4476         if (left < thresh) {
4477                 u64 flags = btrfs_system_alloc_profile(fs_info);
4478
4479                 /*
4480                  * Ignore failure to create system chunk. We might end up not
4481                  * needing it, as we might not need to COW all nodes/leafs from
4482                  * the paths we visit in the chunk tree (they were already COWed
4483                  * or created in the current transaction for example).
4484                  */
4485                 ret = btrfs_alloc_chunk(trans, flags);
4486         }
4487
4488         if (!ret) {
4489                 ret = btrfs_block_rsv_add(fs_info->chunk_root,
4490                                           &fs_info->chunk_block_rsv,
4491                                           thresh, BTRFS_RESERVE_NO_FLUSH);
4492                 if (!ret)
4493                         trans->chunk_bytes_reserved += thresh;
4494         }
4495 }
4496
4497 /*
4498  * If force is CHUNK_ALLOC_FORCE:
4499  *    - return 1 if it successfully allocates a chunk,
4500  *    - return errors including -ENOSPC otherwise.
4501  * If force is NOT CHUNK_ALLOC_FORCE:
4502  *    - return 0 if it doesn't need to allocate a new chunk,
4503  *    - return 1 if it successfully allocates a chunk,
4504  *    - return errors including -ENOSPC otherwise.
4505  */
4506 static int do_chunk_alloc(struct btrfs_trans_handle *trans, u64 flags,
4507                           int force)
4508 {
4509         struct btrfs_fs_info *fs_info = trans->fs_info;
4510         struct btrfs_space_info *space_info;
4511         bool wait_for_alloc = false;
4512         bool should_alloc = false;
4513         int ret = 0;
4514
4515         /* Don't re-enter if we're already allocating a chunk */
4516         if (trans->allocating_chunk)
4517                 return -ENOSPC;
4518
4519         space_info = __find_space_info(fs_info, flags);
4520         ASSERT(space_info);
4521
4522         do {
4523                 spin_lock(&space_info->lock);
4524                 if (force < space_info->force_alloc)
4525                         force = space_info->force_alloc;
4526                 should_alloc = should_alloc_chunk(fs_info, space_info, force);
4527                 if (space_info->full) {
4528                         /* No more free physical space */
4529                         if (should_alloc)
4530                                 ret = -ENOSPC;
4531                         else
4532                                 ret = 0;
4533                         spin_unlock(&space_info->lock);
4534                         return ret;
4535                 } else if (!should_alloc) {
4536                         spin_unlock(&space_info->lock);
4537                         return 0;
4538                 } else if (space_info->chunk_alloc) {
4539                         /*
4540                          * Someone is already allocating, so we need to block
4541                          * until this someone is finished and then loop to
4542                          * recheck if we should continue with our allocation
4543                          * attempt.
4544                          */
4545                         wait_for_alloc = true;
4546                         spin_unlock(&space_info->lock);
4547                         mutex_lock(&fs_info->chunk_mutex);
4548                         mutex_unlock(&fs_info->chunk_mutex);
4549                 } else {
4550                         /* Proceed with allocation */
4551                         space_info->chunk_alloc = 1;
4552                         wait_for_alloc = false;
4553                         spin_unlock(&space_info->lock);
4554                 }
4555
4556                 cond_resched();
4557         } while (wait_for_alloc);
4558
4559         mutex_lock(&fs_info->chunk_mutex);
4560         trans->allocating_chunk = true;
4561
4562         /*
4563          * If we have mixed data/metadata chunks we want to make sure we keep
4564          * allocating mixed chunks instead of individual chunks.
4565          */
4566         if (btrfs_mixed_space_info(space_info))
4567                 flags |= (BTRFS_BLOCK_GROUP_DATA | BTRFS_BLOCK_GROUP_METADATA);
4568
4569         /*
4570          * if we're doing a data chunk, go ahead and make sure that
4571          * we keep a reasonable number of metadata chunks allocated in the
4572          * FS as well.
4573          */
4574         if (flags & BTRFS_BLOCK_GROUP_DATA && fs_info->metadata_ratio) {
4575                 fs_info->data_chunk_allocations++;
4576                 if (!(fs_info->data_chunk_allocations %
4577                       fs_info->metadata_ratio))
4578                         force_metadata_allocation(fs_info);
4579         }
4580
4581         /*
4582          * Check if we have enough space in SYSTEM chunk because we may need
4583          * to update devices.
4584          */
4585         check_system_chunk(trans, flags);
4586
4587         ret = btrfs_alloc_chunk(trans, flags);
4588         trans->allocating_chunk = false;
4589
4590         spin_lock(&space_info->lock);
4591         if (ret < 0) {
4592                 if (ret == -ENOSPC)
4593                         space_info->full = 1;
4594                 else
4595                         goto out;
4596         } else {
4597                 ret = 1;
4598                 space_info->max_extent_size = 0;
4599         }
4600
4601         space_info->force_alloc = CHUNK_ALLOC_NO_FORCE;
4602 out:
4603         space_info->chunk_alloc = 0;
4604         spin_unlock(&space_info->lock);
4605         mutex_unlock(&fs_info->chunk_mutex);
4606         /*
4607          * When we allocate a new chunk we reserve space in the chunk block
4608          * reserve to make sure we can COW nodes/leafs in the chunk tree or
4609          * add new nodes/leafs to it if we end up needing to do it when
4610          * inserting the chunk item and updating device items as part of the
4611          * second phase of chunk allocation, performed by
4612          * btrfs_finish_chunk_alloc(). So make sure we don't accumulate a
4613          * large number of new block groups to create in our transaction
4614          * handle's new_bgs list to avoid exhausting the chunk block reserve
4615          * in extreme cases - like having a single transaction create many new
4616          * block groups when starting to write out the free space caches of all
4617          * the block groups that were made dirty during the lifetime of the
4618          * transaction.
4619          */
4620         if (trans->chunk_bytes_reserved >= (u64)SZ_2M)
4621                 btrfs_create_pending_block_groups(trans);
4622
4623         return ret;
4624 }
4625
4626 static int can_overcommit(struct btrfs_fs_info *fs_info,
4627                           struct btrfs_space_info *space_info, u64 bytes,
4628                           enum btrfs_reserve_flush_enum flush,
4629                           bool system_chunk)
4630 {
4631         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
4632         u64 profile;
4633         u64 space_size;
4634         u64 avail;
4635         u64 used;
4636         int factor;
4637
4638         /* Don't overcommit when in mixed mode. */
4639         if (space_info->flags & BTRFS_BLOCK_GROUP_DATA)
4640                 return 0;
4641
4642         if (system_chunk)
4643                 profile = btrfs_system_alloc_profile(fs_info);
4644         else
4645                 profile = btrfs_metadata_alloc_profile(fs_info);
4646
4647         used = btrfs_space_info_used(space_info, false);
4648
4649         /*
4650          * We only want to allow over committing if we have lots of actual space
4651          * free, but if we don't have enough space to handle the global reserve
4652          * space then we could end up having a real enospc problem when trying
4653          * to allocate a chunk or some other such important allocation.
4654          */
4655         spin_lock(&global_rsv->lock);
4656         space_size = calc_global_rsv_need_space(global_rsv);
4657         spin_unlock(&global_rsv->lock);
4658         if (used + space_size >= space_info->total_bytes)
4659                 return 0;
4660
4661         used += space_info->bytes_may_use;
4662
4663         avail = atomic64_read(&fs_info->free_chunk_space);
4664
4665         /*
4666          * If we have dup, raid1 or raid10 then only half of the free
4667          * space is actually usable.  For raid56, the space info used
4668          * doesn't include the parity drive, so we don't have to
4669          * change the math
4670          */
4671         factor = btrfs_bg_type_to_factor(profile);
4672         avail = div_u64(avail, factor);
4673
4674         /*
4675          * If we aren't flushing all things, let us overcommit up to
4676          * 1/2th of the space. If we can flush, don't let us overcommit
4677          * too much, let it overcommit up to 1/8 of the space.
4678          */
4679         if (flush == BTRFS_RESERVE_FLUSH_ALL)
4680                 avail >>= 3;
4681         else
4682                 avail >>= 1;
4683
4684         if (used + bytes < space_info->total_bytes + avail)
4685                 return 1;
4686         return 0;
4687 }
4688
4689 static void btrfs_writeback_inodes_sb_nr(struct btrfs_fs_info *fs_info,
4690                                          unsigned long nr_pages, int nr_items)
4691 {
4692         struct super_block *sb = fs_info->sb;
4693
4694         if (down_read_trylock(&sb->s_umount)) {
4695                 writeback_inodes_sb_nr(sb, nr_pages, WB_REASON_FS_FREE_SPACE);
4696                 up_read(&sb->s_umount);
4697         } else {
4698                 /*
4699                  * We needn't worry the filesystem going from r/w to r/o though
4700                  * we don't acquire ->s_umount mutex, because the filesystem
4701                  * should guarantee the delalloc inodes list be empty after
4702                  * the filesystem is readonly(all dirty pages are written to
4703                  * the disk).
4704                  */
4705                 btrfs_start_delalloc_roots(fs_info, nr_items);
4706                 if (!current->journal_info)
4707                         btrfs_wait_ordered_roots(fs_info, nr_items, 0, (u64)-1);
4708         }
4709 }
4710
4711 static inline u64 calc_reclaim_items_nr(struct btrfs_fs_info *fs_info,
4712                                         u64 to_reclaim)
4713 {
4714         u64 bytes;
4715         u64 nr;
4716
4717         bytes = btrfs_calc_trans_metadata_size(fs_info, 1);
4718         nr = div64_u64(to_reclaim, bytes);
4719         if (!nr)
4720                 nr = 1;
4721         return nr;
4722 }
4723
4724 #define EXTENT_SIZE_PER_ITEM    SZ_256K
4725
4726 /*
4727  * shrink metadata reservation for delalloc
4728  */
4729 static void shrink_delalloc(struct btrfs_fs_info *fs_info, u64 to_reclaim,
4730                             u64 orig, bool wait_ordered)
4731 {
4732         struct btrfs_space_info *space_info;
4733         struct btrfs_trans_handle *trans;
4734         u64 delalloc_bytes;
4735         u64 async_pages;
4736         u64 items;
4737         long time_left;
4738         unsigned long nr_pages;
4739         int loops;
4740
4741         /* Calc the number of the pages we need flush for space reservation */
4742         items = calc_reclaim_items_nr(fs_info, to_reclaim);
4743         to_reclaim = items * EXTENT_SIZE_PER_ITEM;
4744
4745         trans = (struct btrfs_trans_handle *)current->journal_info;
4746         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
4747
4748         delalloc_bytes = percpu_counter_sum_positive(
4749                                                 &fs_info->delalloc_bytes);
4750         if (delalloc_bytes == 0) {
4751                 if (trans)
4752                         return;
4753                 if (wait_ordered)
4754                         btrfs_wait_ordered_roots(fs_info, items, 0, (u64)-1);
4755                 return;
4756         }
4757
4758         loops = 0;
4759         while (delalloc_bytes && loops < 3) {
4760                 nr_pages = min(delalloc_bytes, to_reclaim) >> PAGE_SHIFT;
4761
4762                 /*
4763                  * Triggers inode writeback for up to nr_pages. This will invoke
4764                  * ->writepages callback and trigger delalloc filling
4765                  *  (btrfs_run_delalloc_range()).
4766                  */
4767                 btrfs_writeback_inodes_sb_nr(fs_info, nr_pages, items);
4768
4769                 /*
4770                  * We need to wait for the compressed pages to start before
4771                  * we continue.
4772                  */
4773                 async_pages = atomic_read(&fs_info->async_delalloc_pages);
4774                 if (!async_pages)
4775                         goto skip_async;
4776
4777                 /*
4778                  * Calculate how many compressed pages we want to be written
4779                  * before we continue. I.e if there are more async pages than we
4780                  * require wait_event will wait until nr_pages are written.
4781                  */
4782                 if (async_pages <= nr_pages)
4783                         async_pages = 0;
4784                 else
4785                         async_pages -= nr_pages;
4786
4787                 wait_event(fs_info->async_submit_wait,
4788                            atomic_read(&fs_info->async_delalloc_pages) <=
4789                            (int)async_pages);
4790 skip_async:
4791                 spin_lock(&space_info->lock);
4792                 if (list_empty(&space_info->tickets) &&
4793                     list_empty(&space_info->priority_tickets)) {
4794                         spin_unlock(&space_info->lock);
4795                         break;
4796                 }
4797                 spin_unlock(&space_info->lock);
4798
4799                 loops++;
4800                 if (wait_ordered && !trans) {
4801                         btrfs_wait_ordered_roots(fs_info, items, 0, (u64)-1);
4802                 } else {
4803                         time_left = schedule_timeout_killable(1);
4804                         if (time_left)
4805                                 break;
4806                 }
4807                 delalloc_bytes = percpu_counter_sum_positive(
4808                                                 &fs_info->delalloc_bytes);
4809         }
4810 }
4811
4812 struct reserve_ticket {
4813         u64 orig_bytes;
4814         u64 bytes;
4815         int error;
4816         struct list_head list;
4817         wait_queue_head_t wait;
4818 };
4819
4820 /**
4821  * maybe_commit_transaction - possibly commit the transaction if its ok to
4822  * @root - the root we're allocating for
4823  * @bytes - the number of bytes we want to reserve
4824  * @force - force the commit
4825  *
4826  * This will check to make sure that committing the transaction will actually
4827  * get us somewhere and then commit the transaction if it does.  Otherwise it
4828  * will return -ENOSPC.
4829  */
4830 static int may_commit_transaction(struct btrfs_fs_info *fs_info,
4831                                   struct btrfs_space_info *space_info)
4832 {
4833         struct reserve_ticket *ticket = NULL;
4834         struct btrfs_block_rsv *delayed_rsv = &fs_info->delayed_block_rsv;
4835         struct btrfs_block_rsv *delayed_refs_rsv = &fs_info->delayed_refs_rsv;
4836         struct btrfs_trans_handle *trans;
4837         u64 bytes_needed;
4838         u64 reclaim_bytes = 0;
4839
4840         trans = (struct btrfs_trans_handle *)current->journal_info;
4841         if (trans)
4842                 return -EAGAIN;
4843
4844         spin_lock(&space_info->lock);
4845         if (!list_empty(&space_info->priority_tickets))
4846                 ticket = list_first_entry(&space_info->priority_tickets,
4847                                           struct reserve_ticket, list);
4848         else if (!list_empty(&space_info->tickets))
4849                 ticket = list_first_entry(&space_info->tickets,
4850                                           struct reserve_ticket, list);
4851         bytes_needed = (ticket) ? ticket->bytes : 0;
4852         spin_unlock(&space_info->lock);
4853
4854         if (!bytes_needed)
4855                 return 0;
4856
4857         trans = btrfs_join_transaction(fs_info->extent_root);
4858         if (IS_ERR(trans))
4859                 return PTR_ERR(trans);
4860
4861         /*
4862          * See if there is enough pinned space to make this reservation, or if
4863          * we have block groups that are going to be freed, allowing us to
4864          * possibly do a chunk allocation the next loop through.
4865          */
4866         if (test_bit(BTRFS_TRANS_HAVE_FREE_BGS, &trans->transaction->flags) ||
4867             __percpu_counter_compare(&space_info->total_bytes_pinned,
4868                                      bytes_needed,
4869                                      BTRFS_TOTAL_BYTES_PINNED_BATCH) >= 0)
4870                 goto commit;
4871
4872         /*
4873          * See if there is some space in the delayed insertion reservation for
4874          * this reservation.
4875          */
4876         if (space_info != delayed_rsv->space_info)
4877                 goto enospc;
4878
4879         spin_lock(&delayed_rsv->lock);
4880         reclaim_bytes += delayed_rsv->reserved;
4881         spin_unlock(&delayed_rsv->lock);
4882
4883         spin_lock(&delayed_refs_rsv->lock);
4884         reclaim_bytes += delayed_refs_rsv->reserved;
4885         spin_unlock(&delayed_refs_rsv->lock);
4886         if (reclaim_bytes >= bytes_needed)
4887                 goto commit;
4888         bytes_needed -= reclaim_bytes;
4889
4890         if (__percpu_counter_compare(&space_info->total_bytes_pinned,
4891                                    bytes_needed,
4892                                    BTRFS_TOTAL_BYTES_PINNED_BATCH) < 0)
4893                 goto enospc;
4894
4895 commit:
4896         return btrfs_commit_transaction(trans);
4897 enospc:
4898         btrfs_end_transaction(trans);
4899         return -ENOSPC;
4900 }
4901
4902 /*
4903  * Try to flush some data based on policy set by @state. This is only advisory
4904  * and may fail for various reasons. The caller is supposed to examine the
4905  * state of @space_info to detect the outcome.
4906  */
4907 static void flush_space(struct btrfs_fs_info *fs_info,
4908                        struct btrfs_space_info *space_info, u64 num_bytes,
4909                        int state)
4910 {
4911         struct btrfs_root *root = fs_info->extent_root;
4912         struct btrfs_trans_handle *trans;
4913         int nr;
4914         int ret = 0;
4915
4916         switch (state) {
4917         case FLUSH_DELAYED_ITEMS_NR:
4918         case FLUSH_DELAYED_ITEMS:
4919                 if (state == FLUSH_DELAYED_ITEMS_NR)
4920                         nr = calc_reclaim_items_nr(fs_info, num_bytes) * 2;
4921                 else
4922                         nr = -1;
4923
4924                 trans = btrfs_join_transaction(root);
4925                 if (IS_ERR(trans)) {
4926                         ret = PTR_ERR(trans);
4927                         break;
4928                 }
4929                 ret = btrfs_run_delayed_items_nr(trans, nr);
4930                 btrfs_end_transaction(trans);
4931                 break;
4932         case FLUSH_DELALLOC:
4933         case FLUSH_DELALLOC_WAIT:
4934                 shrink_delalloc(fs_info, num_bytes * 2, num_bytes,
4935                                 state == FLUSH_DELALLOC_WAIT);
4936                 break;
4937         case FLUSH_DELAYED_REFS_NR:
4938         case FLUSH_DELAYED_REFS:
4939                 trans = btrfs_join_transaction(root);
4940                 if (IS_ERR(trans)) {
4941                         ret = PTR_ERR(trans);
4942                         break;
4943                 }
4944                 if (state == FLUSH_DELAYED_REFS_NR)
4945                         nr = calc_reclaim_items_nr(fs_info, num_bytes);
4946                 else
4947                         nr = 0;
4948                 btrfs_run_delayed_refs(trans, nr);
4949                 btrfs_end_transaction(trans);
4950                 break;
4951         case ALLOC_CHUNK:
4952         case ALLOC_CHUNK_FORCE:
4953                 trans = btrfs_join_transaction(root);
4954                 if (IS_ERR(trans)) {
4955                         ret = PTR_ERR(trans);
4956                         break;
4957                 }
4958                 ret = do_chunk_alloc(trans,
4959                                      btrfs_metadata_alloc_profile(fs_info),
4960                                      (state == ALLOC_CHUNK) ?
4961                                       CHUNK_ALLOC_NO_FORCE : CHUNK_ALLOC_FORCE);
4962                 btrfs_end_transaction(trans);
4963                 if (ret > 0 || ret == -ENOSPC)
4964                         ret = 0;
4965                 break;
4966         case COMMIT_TRANS:
4967                 /*
4968                  * If we have pending delayed iputs then we could free up a
4969                  * bunch of pinned space, so make sure we run the iputs before
4970                  * we do our pinned bytes check below.
4971                  */
4972                 btrfs_run_delayed_iputs(fs_info);
4973                 btrfs_wait_on_delayed_iputs(fs_info);
4974
4975                 ret = may_commit_transaction(fs_info, space_info);
4976                 break;
4977         default:
4978                 ret = -ENOSPC;
4979                 break;
4980         }
4981
4982         trace_btrfs_flush_space(fs_info, space_info->flags, num_bytes, state,
4983                                 ret);
4984         return;
4985 }
4986
4987 static inline u64
4988 btrfs_calc_reclaim_metadata_size(struct btrfs_fs_info *fs_info,
4989                                  struct btrfs_space_info *space_info,
4990                                  bool system_chunk)
4991 {
4992         struct reserve_ticket *ticket;
4993         u64 used;
4994         u64 expected;
4995         u64 to_reclaim = 0;
4996
4997         list_for_each_entry(ticket, &space_info->tickets, list)
4998                 to_reclaim += ticket->bytes;
4999         list_for_each_entry(ticket, &space_info->priority_tickets, list)
5000                 to_reclaim += ticket->bytes;
5001         if (to_reclaim)
5002                 return to_reclaim;
5003
5004         to_reclaim = min_t(u64, num_online_cpus() * SZ_1M, SZ_16M);
5005         if (can_overcommit(fs_info, space_info, to_reclaim,
5006                            BTRFS_RESERVE_FLUSH_ALL, system_chunk))
5007                 return 0;
5008
5009         used = btrfs_space_info_used(space_info, true);
5010
5011         if (can_overcommit(fs_info, space_info, SZ_1M,
5012                            BTRFS_RESERVE_FLUSH_ALL, system_chunk))
5013                 expected = div_factor_fine(space_info->total_bytes, 95);
5014         else
5015                 expected = div_factor_fine(space_info->total_bytes, 90);
5016
5017         if (used > expected)
5018                 to_reclaim = used - expected;
5019         else
5020                 to_reclaim = 0;
5021         to_reclaim = min(to_reclaim, space_info->bytes_may_use +
5022                                      space_info->bytes_reserved);
5023         return to_reclaim;
5024 }
5025
5026 static inline int need_do_async_reclaim(struct btrfs_fs_info *fs_info,
5027                                         struct btrfs_space_info *space_info,
5028                                         u64 used, bool system_chunk)
5029 {
5030         u64 thresh = div_factor_fine(space_info->total_bytes, 98);
5031
5032         /* If we're just plain full then async reclaim just slows us down. */
5033         if ((space_info->bytes_used + space_info->bytes_reserved) >= thresh)
5034                 return 0;
5035
5036         if (!btrfs_calc_reclaim_metadata_size(fs_info, space_info,
5037                                               system_chunk))
5038                 return 0;
5039
5040         return (used >= thresh && !btrfs_fs_closing(fs_info) &&
5041                 !test_bit(BTRFS_FS_STATE_REMOUNTING, &fs_info->fs_state));
5042 }
5043
5044 static bool wake_all_tickets(struct list_head *head)
5045 {
5046         struct reserve_ticket *ticket;
5047
5048         while (!list_empty(head)) {
5049                 ticket = list_first_entry(head, struct reserve_ticket, list);
5050                 list_del_init(&ticket->list);
5051                 ticket->error = -ENOSPC;
5052                 wake_up(&ticket->wait);
5053                 if (ticket->bytes != ticket->orig_bytes)
5054                         return true;
5055         }
5056         return false;
5057 }
5058
5059 /*
5060  * This is for normal flushers, we can wait all goddamned day if we want to.  We
5061  * will loop and continuously try to flush as long as we are making progress.
5062  * We count progress as clearing off tickets each time we have to loop.
5063  */
5064 static void btrfs_async_reclaim_metadata_space(struct work_struct *work)
5065 {
5066         struct btrfs_fs_info *fs_info;
5067         struct btrfs_space_info *space_info;
5068         u64 to_reclaim;
5069         int flush_state;
5070         int commit_cycles = 0;
5071         u64 last_tickets_id;
5072
5073         fs_info = container_of(work, struct btrfs_fs_info, async_reclaim_work);
5074         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
5075
5076         spin_lock(&space_info->lock);
5077         to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info, space_info,
5078                                                       false);
5079         if (!to_reclaim) {
5080                 space_info->flush = 0;
5081                 spin_unlock(&space_info->lock);
5082                 return;
5083         }
5084         last_tickets_id = space_info->tickets_id;
5085         spin_unlock(&space_info->lock);
5086
5087         flush_state = FLUSH_DELAYED_ITEMS_NR;
5088         do {
5089                 flush_space(fs_info, space_info, to_reclaim, flush_state);
5090                 spin_lock(&space_info->lock);
5091                 if (list_empty(&space_info->tickets)) {
5092                         space_info->flush = 0;
5093                         spin_unlock(&space_info->lock);
5094                         return;
5095                 }
5096                 to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info,
5097                                                               space_info,
5098                                                               false);
5099                 if (last_tickets_id == space_info->tickets_id) {
5100                         flush_state++;
5101                 } else {
5102                         last_tickets_id = space_info->tickets_id;
5103                         flush_state = FLUSH_DELAYED_ITEMS_NR;
5104                         if (commit_cycles)
5105                                 commit_cycles--;
5106                 }
5107
5108                 /*
5109                  * We don't want to force a chunk allocation until we've tried
5110                  * pretty hard to reclaim space.  Think of the case where we
5111                  * freed up a bunch of space and so have a lot of pinned space
5112                  * to reclaim.  We would rather use that than possibly create a
5113                  * underutilized metadata chunk.  So if this is our first run
5114                  * through the flushing state machine skip ALLOC_CHUNK_FORCE and
5115                  * commit the transaction.  If nothing has changed the next go
5116                  * around then we can force a chunk allocation.
5117                  */
5118                 if (flush_state == ALLOC_CHUNK_FORCE && !commit_cycles)
5119                         flush_state++;
5120
5121                 if (flush_state > COMMIT_TRANS) {
5122                         commit_cycles++;
5123                         if (commit_cycles > 2) {
5124                                 if (wake_all_tickets(&space_info->tickets)) {
5125                                         flush_state = FLUSH_DELAYED_ITEMS_NR;
5126                                         commit_cycles--;
5127                                 } else {
5128                                         space_info->flush = 0;
5129                                 }
5130                         } else {
5131                                 flush_state = FLUSH_DELAYED_ITEMS_NR;
5132                         }
5133                 }
5134                 spin_unlock(&space_info->lock);
5135         } while (flush_state <= COMMIT_TRANS);
5136 }
5137
5138 void btrfs_init_async_reclaim_work(struct work_struct *work)
5139 {
5140         INIT_WORK(work, btrfs_async_reclaim_metadata_space);
5141 }
5142
5143 static const enum btrfs_flush_state priority_flush_states[] = {
5144         FLUSH_DELAYED_ITEMS_NR,
5145         FLUSH_DELAYED_ITEMS,
5146         ALLOC_CHUNK,
5147 };
5148
5149 static void priority_reclaim_metadata_space(struct btrfs_fs_info *fs_info,
5150                                             struct btrfs_space_info *space_info,
5151                                             struct reserve_ticket *ticket)
5152 {
5153         u64 to_reclaim;
5154         int flush_state;
5155
5156         spin_lock(&space_info->lock);
5157         to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info, space_info,
5158                                                       false);
5159         if (!to_reclaim) {
5160                 spin_unlock(&space_info->lock);
5161                 return;
5162         }
5163         spin_unlock(&space_info->lock);
5164
5165         flush_state = 0;
5166         do {
5167                 flush_space(fs_info, space_info, to_reclaim,
5168                             priority_flush_states[flush_state]);
5169                 flush_state++;
5170                 spin_lock(&space_info->lock);
5171                 if (ticket->bytes == 0) {
5172                         spin_unlock(&space_info->lock);
5173                         return;
5174                 }
5175                 spin_unlock(&space_info->lock);
5176         } while (flush_state < ARRAY_SIZE(priority_flush_states));
5177 }
5178
5179 static int wait_reserve_ticket(struct btrfs_fs_info *fs_info,
5180                                struct btrfs_space_info *space_info,
5181                                struct reserve_ticket *ticket)
5182
5183 {
5184         DEFINE_WAIT(wait);
5185         u64 reclaim_bytes = 0;
5186         int ret = 0;
5187
5188         spin_lock(&space_info->lock);
5189         while (ticket->bytes > 0 && ticket->error == 0) {
5190                 ret = prepare_to_wait_event(&ticket->wait, &wait, TASK_KILLABLE);
5191                 if (ret) {
5192                         ret = -EINTR;
5193                         break;
5194                 }
5195                 spin_unlock(&space_info->lock);
5196
5197                 schedule();
5198
5199                 finish_wait(&ticket->wait, &wait);
5200                 spin_lock(&space_info->lock);
5201         }
5202         if (!ret)
5203                 ret = ticket->error;
5204         if (!list_empty(&ticket->list))
5205                 list_del_init(&ticket->list);
5206         if (ticket->bytes && ticket->bytes < ticket->orig_bytes)
5207                 reclaim_bytes = ticket->orig_bytes - ticket->bytes;
5208         spin_unlock(&space_info->lock);
5209
5210         if (reclaim_bytes)
5211                 space_info_add_old_bytes(fs_info, space_info, reclaim_bytes);
5212         return ret;
5213 }
5214
5215 /**
5216  * reserve_metadata_bytes - try to reserve bytes from the block_rsv's space
5217  * @root - the root we're allocating for
5218  * @space_info - the space info we want to allocate from
5219  * @orig_bytes - the number of bytes we want
5220  * @flush - whether or not we can flush to make our reservation
5221  *
5222  * This will reserve orig_bytes number of bytes from the space info associated
5223  * with the block_rsv.  If there is not enough space it will make an attempt to
5224  * flush out space to make room.  It will do this by flushing delalloc if
5225  * possible or committing the transaction.  If flush is 0 then no attempts to
5226  * regain reservations will be made and this will fail if there is not enough
5227  * space already.
5228  */
5229 static int __reserve_metadata_bytes(struct btrfs_fs_info *fs_info,
5230                                     struct btrfs_space_info *space_info,
5231                                     u64 orig_bytes,
5232                                     enum btrfs_reserve_flush_enum flush,
5233                                     bool system_chunk)
5234 {
5235         struct reserve_ticket ticket;
5236         u64 used;
5237         u64 reclaim_bytes = 0;
5238         int ret = 0;
5239
5240         ASSERT(orig_bytes);
5241         ASSERT(!current->journal_info || flush != BTRFS_RESERVE_FLUSH_ALL);
5242
5243         spin_lock(&space_info->lock);
5244         ret = -ENOSPC;
5245         used = btrfs_space_info_used(space_info, true);
5246
5247         /*
5248          * If we have enough space then hooray, make our reservation and carry
5249          * on.  If not see if we can overcommit, and if we can, hooray carry on.
5250          * If not things get more complicated.
5251          */
5252         if (used + orig_bytes <= space_info->total_bytes) {
5253                 update_bytes_may_use(space_info, orig_bytes);
5254                 trace_btrfs_space_reservation(fs_info, "space_info",
5255                                               space_info->flags, orig_bytes, 1);
5256                 ret = 0;
5257         } else if (can_overcommit(fs_info, space_info, orig_bytes, flush,
5258                                   system_chunk)) {
5259                 update_bytes_may_use(space_info, orig_bytes);
5260                 trace_btrfs_space_reservation(fs_info, "space_info",
5261                                               space_info->flags, orig_bytes, 1);
5262                 ret = 0;
5263         }
5264
5265         /*
5266          * If we couldn't make a reservation then setup our reservation ticket
5267          * and kick the async worker if it's not already running.
5268          *
5269          * If we are a priority flusher then we just need to add our ticket to
5270          * the list and we will do our own flushing further down.
5271          */
5272         if (ret && flush != BTRFS_RESERVE_NO_FLUSH) {
5273                 ticket.orig_bytes = orig_bytes;
5274                 ticket.bytes = orig_bytes;
5275                 ticket.error = 0;
5276                 init_waitqueue_head(&ticket.wait);
5277                 if (flush == BTRFS_RESERVE_FLUSH_ALL) {
5278                         list_add_tail(&ticket.list, &space_info->tickets);
5279                         if (!space_info->flush) {
5280                                 space_info->flush = 1;
5281                                 trace_btrfs_trigger_flush(fs_info,
5282                                                           space_info->flags,
5283                                                           orig_bytes, flush,
5284                                                           "enospc");
5285                                 queue_work(system_unbound_wq,
5286                                            &fs_info->async_reclaim_work);
5287                         }
5288                 } else {
5289                         list_add_tail(&ticket.list,
5290                                       &space_info->priority_tickets);
5291                 }
5292         } else if (!ret && space_info->flags & BTRFS_BLOCK_GROUP_METADATA) {
5293                 used += orig_bytes;
5294                 /*
5295                  * We will do the space reservation dance during log replay,
5296                  * which means we won't have fs_info->fs_root set, so don't do
5297                  * the async reclaim as we will panic.
5298                  */
5299                 if (!test_bit(BTRFS_FS_LOG_RECOVERING, &fs_info->flags) &&
5300                     need_do_async_reclaim(fs_info, space_info,
5301                                           used, system_chunk) &&
5302                     !work_busy(&fs_info->async_reclaim_work)) {
5303                         trace_btrfs_trigger_flush(fs_info, space_info->flags,
5304                                                   orig_bytes, flush, "preempt");
5305                         queue_work(system_unbound_wq,
5306                                    &fs_info->async_reclaim_work);
5307                 }
5308         }
5309         spin_unlock(&space_info->lock);
5310         if (!ret || flush == BTRFS_RESERVE_NO_FLUSH)
5311                 return ret;
5312
5313         if (flush == BTRFS_RESERVE_FLUSH_ALL)
5314                 return wait_reserve_ticket(fs_info, space_info, &ticket);
5315
5316         ret = 0;
5317         priority_reclaim_metadata_space(fs_info, space_info, &ticket);
5318         spin_lock(&space_info->lock);
5319         if (ticket.bytes) {
5320                 if (ticket.bytes < orig_bytes)
5321                         reclaim_bytes = orig_bytes - ticket.bytes;
5322                 list_del_init(&ticket.list);
5323                 ret = -ENOSPC;
5324         }
5325         spin_unlock(&space_info->lock);
5326
5327         if (reclaim_bytes)
5328                 space_info_add_old_bytes(fs_info, space_info, reclaim_bytes);
5329         ASSERT(list_empty(&ticket.list));
5330         return ret;
5331 }
5332
5333 /**
5334  * reserve_metadata_bytes - try to reserve bytes from the block_rsv's space
5335  * @root - the root we're allocating for
5336  * @block_rsv - the block_rsv we're allocating for
5337  * @orig_bytes - the number of bytes we want
5338  * @flush - whether or not we can flush to make our reservation
5339  *
5340  * This will reserve orig_bytes number of bytes from the space info associated
5341  * with the block_rsv.  If there is not enough space it will make an attempt to
5342  * flush out space to make room.  It will do this by flushing delalloc if
5343  * possible or committing the transaction.  If flush is 0 then no attempts to
5344  * regain reservations will be made and this will fail if there is not enough
5345  * space already.
5346  */
5347 static int reserve_metadata_bytes(struct btrfs_root *root,
5348                                   struct btrfs_block_rsv *block_rsv,
5349                                   u64 orig_bytes,
5350                                   enum btrfs_reserve_flush_enum flush)
5351 {
5352         struct btrfs_fs_info *fs_info = root->fs_info;
5353         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
5354         int ret;
5355         bool system_chunk = (root == fs_info->chunk_root);
5356
5357         ret = __reserve_metadata_bytes(fs_info, block_rsv->space_info,
5358                                        orig_bytes, flush, system_chunk);
5359         if (ret == -ENOSPC &&
5360             unlikely(root->orphan_cleanup_state == ORPHAN_CLEANUP_STARTED)) {
5361                 if (block_rsv != global_rsv &&
5362                     !block_rsv_use_bytes(global_rsv, orig_bytes))
5363                         ret = 0;
5364         }
5365         if (ret == -ENOSPC) {
5366                 trace_btrfs_space_reservation(fs_info, "space_info:enospc",
5367                                               block_rsv->space_info->flags,
5368                                               orig_bytes, 1);
5369
5370                 if (btrfs_test_opt(fs_info, ENOSPC_DEBUG))
5371                         dump_space_info(fs_info, block_rsv->space_info,
5372                                         orig_bytes, 0);
5373         }
5374         return ret;
5375 }
5376
5377 static struct btrfs_block_rsv *get_block_rsv(
5378                                         const struct btrfs_trans_handle *trans,
5379                                         const struct btrfs_root *root)
5380 {
5381         struct btrfs_fs_info *fs_info = root->fs_info;
5382         struct btrfs_block_rsv *block_rsv = NULL;
5383
5384         if (test_bit(BTRFS_ROOT_REF_COWS, &root->state) ||
5385             (root == fs_info->csum_root && trans->adding_csums) ||
5386             (root == fs_info->uuid_root))
5387                 block_rsv = trans->block_rsv;
5388
5389         if (!block_rsv)
5390                 block_rsv = root->block_rsv;
5391
5392         if (!block_rsv)
5393                 block_rsv = &fs_info->empty_block_rsv;
5394
5395         return block_rsv;
5396 }
5397
5398 static int block_rsv_use_bytes(struct btrfs_block_rsv *block_rsv,
5399                                u64 num_bytes)
5400 {
5401         int ret = -ENOSPC;
5402         spin_lock(&block_rsv->lock);
5403         if (block_rsv->reserved >= num_bytes) {
5404                 block_rsv->reserved -= num_bytes;
5405                 if (block_rsv->reserved < block_rsv->size)
5406                         block_rsv->full = 0;
5407                 ret = 0;
5408         }
5409         spin_unlock(&block_rsv->lock);
5410         return ret;
5411 }
5412
5413 static void block_rsv_add_bytes(struct btrfs_block_rsv *block_rsv,
5414                                 u64 num_bytes, bool update_size)
5415 {
5416         spin_lock(&block_rsv->lock);
5417         block_rsv->reserved += num_bytes;
5418         if (update_size)
5419                 block_rsv->size += num_bytes;
5420         else if (block_rsv->reserved >= block_rsv->size)
5421                 block_rsv->full = 1;
5422         spin_unlock(&block_rsv->lock);
5423 }
5424
5425 int btrfs_cond_migrate_bytes(struct btrfs_fs_info *fs_info,
5426                              struct btrfs_block_rsv *dest, u64 num_bytes,
5427                              int min_factor)
5428 {
5429         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
5430         u64 min_bytes;
5431
5432         if (global_rsv->space_info != dest->space_info)
5433                 return -ENOSPC;
5434
5435         spin_lock(&global_rsv->lock);
5436         min_bytes = div_factor(global_rsv->size, min_factor);
5437         if (global_rsv->reserved < min_bytes + num_bytes) {
5438                 spin_unlock(&global_rsv->lock);
5439                 return -ENOSPC;
5440         }
5441         global_rsv->reserved -= num_bytes;
5442         if (global_rsv->reserved < global_rsv->size)
5443                 global_rsv->full = 0;
5444         spin_unlock(&global_rsv->lock);
5445
5446         block_rsv_add_bytes(dest, num_bytes, true);
5447         return 0;
5448 }
5449
5450 /**
5451  * btrfs_migrate_to_delayed_refs_rsv - transfer bytes to our delayed refs rsv.
5452  * @fs_info - the fs info for our fs.
5453  * @src - the source block rsv to transfer from.
5454  * @num_bytes - the number of bytes to transfer.
5455  *
5456  * This transfers up to the num_bytes amount from the src rsv to the
5457  * delayed_refs_rsv.  Any extra bytes are returned to the space info.
5458  */
5459 void btrfs_migrate_to_delayed_refs_rsv(struct btrfs_fs_info *fs_info,
5460                                        struct btrfs_block_rsv *src,
5461                                        u64 num_bytes)
5462 {
5463         struct btrfs_block_rsv *delayed_refs_rsv = &fs_info->delayed_refs_rsv;
5464         u64 to_free = 0;
5465
5466         spin_lock(&src->lock);
5467         src->reserved -= num_bytes;
5468         src->size -= num_bytes;
5469         spin_unlock(&src->lock);
5470
5471         spin_lock(&delayed_refs_rsv->lock);
5472         if (delayed_refs_rsv->size > delayed_refs_rsv->reserved) {
5473                 u64 delta = delayed_refs_rsv->size -
5474                         delayed_refs_rsv->reserved;
5475                 if (num_bytes > delta) {
5476                         to_free = num_bytes - delta;
5477                         num_bytes = delta;
5478                 }
5479         } else {
5480                 to_free = num_bytes;
5481                 num_bytes = 0;
5482         }
5483
5484         if (num_bytes)
5485                 delayed_refs_rsv->reserved += num_bytes;
5486         if (delayed_refs_rsv->reserved >= delayed_refs_rsv->size)
5487                 delayed_refs_rsv->full = 1;
5488         spin_unlock(&delayed_refs_rsv->lock);
5489
5490         if (num_bytes)
5491                 trace_btrfs_space_reservation(fs_info, "delayed_refs_rsv",
5492                                               0, num_bytes, 1);
5493         if (to_free)
5494                 space_info_add_old_bytes(fs_info, delayed_refs_rsv->space_info,
5495                                          to_free);
5496 }
5497
5498 /**
5499  * btrfs_delayed_refs_rsv_refill - refill based on our delayed refs usage.
5500  * @fs_info - the fs_info for our fs.
5501  * @flush - control how we can flush for this reservation.
5502  *
5503  * This will refill the delayed block_rsv up to 1 items size worth of space and
5504  * will return -ENOSPC if we can't make the reservation.
5505  */
5506 int btrfs_delayed_refs_rsv_refill(struct btrfs_fs_info *fs_info,
5507                                   enum btrfs_reserve_flush_enum flush)
5508 {
5509         struct btrfs_block_rsv *block_rsv = &fs_info->delayed_refs_rsv;
5510         u64 limit = btrfs_calc_trans_metadata_size(fs_info, 1);
5511         u64 num_bytes = 0;
5512         int ret = -ENOSPC;
5513
5514         spin_lock(&block_rsv->lock);
5515         if (block_rsv->reserved < block_rsv->size) {
5516                 num_bytes = block_rsv->size - block_rsv->reserved;
5517                 num_bytes = min(num_bytes, limit);
5518         }
5519         spin_unlock(&block_rsv->lock);
5520
5521         if (!num_bytes)
5522                 return 0;
5523
5524         ret = reserve_metadata_bytes(fs_info->extent_root, block_rsv,
5525                                      num_bytes, flush);
5526         if (ret)
5527                 return ret;
5528         block_rsv_add_bytes(block_rsv, num_bytes, 0);
5529         trace_btrfs_space_reservation(fs_info, "delayed_refs_rsv",
5530                                       0, num_bytes, 1);
5531         return 0;
5532 }
5533
5534 /*
5535  * This is for space we already have accounted in space_info->bytes_may_use, so
5536  * basically when we're returning space from block_rsv's.
5537  */
5538 static void space_info_add_old_bytes(struct btrfs_fs_info *fs_info,
5539                                      struct btrfs_space_info *space_info,
5540                                      u64 num_bytes)
5541 {
5542         struct reserve_ticket *ticket;
5543         struct list_head *head;
5544         u64 used;
5545         enum btrfs_reserve_flush_enum flush = BTRFS_RESERVE_NO_FLUSH;
5546         bool check_overcommit = false;
5547
5548         spin_lock(&space_info->lock);
5549         head = &space_info->priority_tickets;
5550
5551         /*
5552          * If we are over our limit then we need to check and see if we can
5553          * overcommit, and if we can't then we just need to free up our space
5554          * and not satisfy any requests.
5555          */
5556         used = btrfs_space_info_used(space_info, true);
5557         if (used - num_bytes >= space_info->total_bytes)
5558                 check_overcommit = true;
5559 again:
5560         while (!list_empty(head) && num_bytes) {
5561                 ticket = list_first_entry(head, struct reserve_ticket,
5562                                           list);
5563                 /*
5564                  * We use 0 bytes because this space is already reserved, so
5565                  * adding the ticket space would be a double count.
5566                  */
5567                 if (check_overcommit &&
5568                     !can_overcommit(fs_info, space_info, 0, flush, false))
5569                         break;
5570                 if (num_bytes >= ticket->bytes) {
5571                         list_del_init(&ticket->list);
5572                         num_bytes -= ticket->bytes;
5573                         ticket->bytes = 0;
5574                         space_info->tickets_id++;
5575                         wake_up(&ticket->wait);
5576                 } else {
5577                         ticket->bytes -= num_bytes;
5578                         num_bytes = 0;
5579                 }
5580         }
5581
5582         if (num_bytes && head == &space_info->priority_tickets) {
5583                 head = &space_info->tickets;
5584                 flush = BTRFS_RESERVE_FLUSH_ALL;
5585                 goto again;
5586         }
5587         update_bytes_may_use(space_info, -num_bytes);
5588         trace_btrfs_space_reservation(fs_info, "space_info",
5589                                       space_info->flags, num_bytes, 0);
5590         spin_unlock(&space_info->lock);
5591 }
5592
5593 /*
5594  * This is for newly allocated space that isn't accounted in
5595  * space_info->bytes_may_use yet.  So if we allocate a chunk or unpin an extent
5596  * we use this helper.
5597  */
5598 static void space_info_add_new_bytes(struct btrfs_fs_info *fs_info,
5599                                      struct btrfs_space_info *space_info,
5600                                      u64 num_bytes)
5601 {
5602         struct reserve_ticket *ticket;
5603         struct list_head *head = &space_info->priority_tickets;
5604
5605 again:
5606         while (!list_empty(head) && num_bytes) {
5607                 ticket = list_first_entry(head, struct reserve_ticket,
5608                                           list);
5609                 if (num_bytes >= ticket->bytes) {
5610                         trace_btrfs_space_reservation(fs_info, "space_info",
5611                                                       space_info->flags,
5612                                                       ticket->bytes, 1);
5613                         list_del_init(&ticket->list);
5614                         num_bytes -= ticket->bytes;
5615                         update_bytes_may_use(space_info, ticket->bytes);
5616                         ticket->bytes = 0;
5617                         space_info->tickets_id++;
5618                         wake_up(&ticket->wait);
5619                 } else {
5620                         trace_btrfs_space_reservation(fs_info, "space_info",
5621                                                       space_info->flags,
5622                                                       num_bytes, 1);
5623                         update_bytes_may_use(space_info, num_bytes);
5624                         ticket->bytes -= num_bytes;
5625                         num_bytes = 0;
5626                 }
5627         }
5628
5629         if (num_bytes && head == &space_info->priority_tickets) {
5630                 head = &space_info->tickets;
5631                 goto again;
5632         }
5633 }
5634
5635 static u64 block_rsv_release_bytes(struct btrfs_fs_info *fs_info,
5636                                     struct btrfs_block_rsv *block_rsv,
5637                                     struct btrfs_block_rsv *dest, u64 num_bytes,
5638                                     u64 *qgroup_to_release_ret)
5639 {
5640         struct btrfs_space_info *space_info = block_rsv->space_info;
5641         u64 qgroup_to_release = 0;
5642         u64 ret;
5643
5644         spin_lock(&block_rsv->lock);
5645         if (num_bytes == (u64)-1) {
5646                 num_bytes = block_rsv->size;
5647                 qgroup_to_release = block_rsv->qgroup_rsv_size;
5648         }
5649         block_rsv->size -= num_bytes;
5650         if (block_rsv->reserved >= block_rsv->size) {
5651                 num_bytes = block_rsv->reserved - block_rsv->size;
5652                 block_rsv->reserved = block_rsv->size;
5653                 block_rsv->full = 1;
5654         } else {
5655                 num_bytes = 0;
5656         }
5657         if (block_rsv->qgroup_rsv_reserved >= block_rsv->qgroup_rsv_size) {
5658                 qgroup_to_release = block_rsv->qgroup_rsv_reserved -
5659                                     block_rsv->qgroup_rsv_size;
5660                 block_rsv->qgroup_rsv_reserved = block_rsv->qgroup_rsv_size;
5661         } else {
5662                 qgroup_to_release = 0;
5663         }
5664         spin_unlock(&block_rsv->lock);
5665
5666         ret = num_bytes;
5667         if (num_bytes > 0) {
5668                 if (dest) {
5669                         spin_lock(&dest->lock);
5670                         if (!dest->full) {
5671                                 u64 bytes_to_add;
5672
5673                                 bytes_to_add = dest->size - dest->reserved;
5674                                 bytes_to_add = min(num_bytes, bytes_to_add);
5675                                 dest->reserved += bytes_to_add;
5676                                 if (dest->reserved >= dest->size)
5677                                         dest->full = 1;
5678                                 num_bytes -= bytes_to_add;
5679                         }
5680                         spin_unlock(&dest->lock);
5681                 }
5682                 if (num_bytes)
5683                         space_info_add_old_bytes(fs_info, space_info,
5684                                                  num_bytes);
5685         }
5686         if (qgroup_to_release_ret)
5687                 *qgroup_to_release_ret = qgroup_to_release;
5688         return ret;
5689 }
5690
5691 int btrfs_block_rsv_migrate(struct btrfs_block_rsv *src,
5692                             struct btrfs_block_rsv *dst, u64 num_bytes,
5693                             bool update_size)
5694 {
5695         int ret;
5696
5697         ret = block_rsv_use_bytes(src, num_bytes);
5698         if (ret)
5699                 return ret;
5700
5701         block_rsv_add_bytes(dst, num_bytes, update_size);
5702         return 0;
5703 }
5704
5705 void btrfs_init_block_rsv(struct btrfs_block_rsv *rsv, unsigned short type)
5706 {
5707         memset(rsv, 0, sizeof(*rsv));
5708         spin_lock_init(&rsv->lock);
5709         rsv->type = type;
5710 }
5711
5712 void btrfs_init_metadata_block_rsv(struct btrfs_fs_info *fs_info,
5713                                    struct btrfs_block_rsv *rsv,
5714                                    unsigned short type)
5715 {
5716         btrfs_init_block_rsv(rsv, type);
5717         rsv->space_info = __find_space_info(fs_info,
5718                                             BTRFS_BLOCK_GROUP_METADATA);
5719 }
5720
5721 struct btrfs_block_rsv *btrfs_alloc_block_rsv(struct btrfs_fs_info *fs_info,
5722                                               unsigned short type)
5723 {
5724         struct btrfs_block_rsv *block_rsv;
5725
5726         block_rsv = kmalloc(sizeof(*block_rsv), GFP_NOFS);
5727         if (!block_rsv)
5728                 return NULL;
5729
5730         btrfs_init_metadata_block_rsv(fs_info, block_rsv, type);
5731         return block_rsv;
5732 }
5733
5734 void btrfs_free_block_rsv(struct btrfs_fs_info *fs_info,
5735                           struct btrfs_block_rsv *rsv)
5736 {
5737         if (!rsv)
5738                 return;
5739         btrfs_block_rsv_release(fs_info, rsv, (u64)-1);
5740         kfree(rsv);
5741 }
5742
5743 int btrfs_block_rsv_add(struct btrfs_root *root,
5744                         struct btrfs_block_rsv *block_rsv, u64 num_bytes,
5745                         enum btrfs_reserve_flush_enum flush)
5746 {
5747         int ret;
5748
5749         if (num_bytes == 0)
5750                 return 0;
5751
5752         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
5753         if (!ret)
5754                 block_rsv_add_bytes(block_rsv, num_bytes, true);
5755
5756         return ret;
5757 }
5758
5759 int btrfs_block_rsv_check(struct btrfs_block_rsv *block_rsv, int min_factor)
5760 {
5761         u64 num_bytes = 0;
5762         int ret = -ENOSPC;
5763
5764         if (!block_rsv)
5765                 return 0;
5766
5767         spin_lock(&block_rsv->lock);
5768         num_bytes = div_factor(block_rsv->size, min_factor);
5769         if (block_rsv->reserved >= num_bytes)
5770                 ret = 0;
5771         spin_unlock(&block_rsv->lock);
5772
5773         return ret;
5774 }
5775
5776 int btrfs_block_rsv_refill(struct btrfs_root *root,
5777                            struct btrfs_block_rsv *block_rsv, u64 min_reserved,
5778                            enum btrfs_reserve_flush_enum flush)
5779 {
5780         u64 num_bytes = 0;
5781         int ret = -ENOSPC;
5782
5783         if (!block_rsv)
5784                 return 0;
5785
5786         spin_lock(&block_rsv->lock);
5787         num_bytes = min_reserved;
5788         if (block_rsv->reserved >= num_bytes)
5789                 ret = 0;
5790         else
5791                 num_bytes -= block_rsv->reserved;
5792         spin_unlock(&block_rsv->lock);
5793
5794         if (!ret)
5795                 return 0;
5796
5797         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
5798         if (!ret) {
5799                 block_rsv_add_bytes(block_rsv, num_bytes, false);
5800                 return 0;
5801         }
5802
5803         return ret;
5804 }
5805
5806 static void calc_refill_bytes(struct btrfs_block_rsv *block_rsv,
5807                                 u64 *metadata_bytes, u64 *qgroup_bytes)
5808 {
5809         *metadata_bytes = 0;
5810         *qgroup_bytes = 0;
5811
5812         spin_lock(&block_rsv->lock);
5813         if (block_rsv->reserved < block_rsv->size)
5814                 *metadata_bytes = block_rsv->size - block_rsv->reserved;
5815         if (block_rsv->qgroup_rsv_reserved < block_rsv->qgroup_rsv_size)
5816                 *qgroup_bytes = block_rsv->qgroup_rsv_size -
5817                         block_rsv->qgroup_rsv_reserved;
5818         spin_unlock(&block_rsv->lock);
5819 }
5820
5821 /**
5822  * btrfs_inode_rsv_refill - refill the inode block rsv.
5823  * @inode - the inode we are refilling.
5824  * @flush - the flushing restriction.
5825  *
5826  * Essentially the same as btrfs_block_rsv_refill, except it uses the
5827  * block_rsv->size as the minimum size.  We'll either refill the missing amount
5828  * or return if we already have enough space.  This will also handle the reserve
5829  * tracepoint for the reserved amount.
5830  */
5831 static int btrfs_inode_rsv_refill(struct btrfs_inode *inode,
5832                                   enum btrfs_reserve_flush_enum flush)
5833 {
5834         struct btrfs_root *root = inode->root;
5835         struct btrfs_block_rsv *block_rsv = &inode->block_rsv;
5836         u64 num_bytes, last = 0;
5837         u64 qgroup_num_bytes;
5838         int ret = -ENOSPC;
5839
5840         calc_refill_bytes(block_rsv, &num_bytes, &qgroup_num_bytes);
5841         if (num_bytes == 0)
5842                 return 0;
5843
5844         do {
5845                 ret = btrfs_qgroup_reserve_meta_prealloc(root, qgroup_num_bytes,
5846                                                          true);
5847                 if (ret)
5848                         return ret;
5849                 ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
5850                 if (ret) {
5851                         btrfs_qgroup_free_meta_prealloc(root, qgroup_num_bytes);
5852                         last = num_bytes;
5853                         /*
5854                          * If we are fragmented we can end up with a lot of
5855                          * outstanding extents which will make our size be much
5856                          * larger than our reserved amount.
5857                          *
5858                          * If the reservation happens here, it might be very
5859                          * big though not needed in the end, if the delalloc
5860                          * flushing happens.
5861                          *
5862                          * If this is the case try and do the reserve again.
5863                          */
5864                         if (flush == BTRFS_RESERVE_FLUSH_ALL)
5865                                 calc_refill_bytes(block_rsv, &num_bytes,
5866                                                    &qgroup_num_bytes);
5867                         if (num_bytes == 0)
5868                                 return 0;
5869                 }
5870         } while (ret && last != num_bytes);
5871
5872         if (!ret) {
5873                 block_rsv_add_bytes(block_rsv, num_bytes, false);
5874                 trace_btrfs_space_reservation(root->fs_info, "delalloc",
5875                                               btrfs_ino(inode), num_bytes, 1);
5876
5877                 /* Don't forget to increase qgroup_rsv_reserved */
5878                 spin_lock(&block_rsv->lock);
5879                 block_rsv->qgroup_rsv_reserved += qgroup_num_bytes;
5880                 spin_unlock(&block_rsv->lock);
5881         }
5882         return ret;
5883 }
5884
5885 static u64 __btrfs_block_rsv_release(struct btrfs_fs_info *fs_info,
5886                                      struct btrfs_block_rsv *block_rsv,
5887                                      u64 num_bytes, u64 *qgroup_to_release)
5888 {
5889         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
5890         struct btrfs_block_rsv *delayed_rsv = &fs_info->delayed_refs_rsv;
5891         struct btrfs_block_rsv *target = delayed_rsv;
5892
5893         if (target->full || target == block_rsv)
5894                 target = global_rsv;
5895
5896         if (block_rsv->space_info != target->space_info)
5897                 target = NULL;
5898
5899         return block_rsv_release_bytes(fs_info, block_rsv, target, num_bytes,
5900                                        qgroup_to_release);
5901 }
5902
5903 void btrfs_block_rsv_release(struct btrfs_fs_info *fs_info,
5904                              struct btrfs_block_rsv *block_rsv,
5905                              u64 num_bytes)
5906 {
5907         __btrfs_block_rsv_release(fs_info, block_rsv, num_bytes, NULL);
5908 }
5909
5910 /**
5911  * btrfs_inode_rsv_release - release any excessive reservation.
5912  * @inode - the inode we need to release from.
5913  * @qgroup_free - free or convert qgroup meta.
5914  *   Unlike normal operation, qgroup meta reservation needs to know if we are
5915  *   freeing qgroup reservation or just converting it into per-trans.  Normally
5916  *   @qgroup_free is true for error handling, and false for normal release.
5917  *
5918  * This is the same as btrfs_block_rsv_release, except that it handles the
5919  * tracepoint for the reservation.
5920  */
5921 static void btrfs_inode_rsv_release(struct btrfs_inode *inode, bool qgroup_free)
5922 {
5923         struct btrfs_fs_info *fs_info = inode->root->fs_info;
5924         struct btrfs_block_rsv *block_rsv = &inode->block_rsv;
5925         u64 released = 0;
5926         u64 qgroup_to_release = 0;
5927
5928         /*
5929          * Since we statically set the block_rsv->size we just want to say we
5930          * are releasing 0 bytes, and then we'll just get the reservation over
5931          * the size free'd.
5932          */
5933         released = __btrfs_block_rsv_release(fs_info, block_rsv, 0,
5934                                              &qgroup_to_release);
5935         if (released > 0)
5936                 trace_btrfs_space_reservation(fs_info, "delalloc",
5937                                               btrfs_ino(inode), released, 0);
5938         if (qgroup_free)
5939                 btrfs_qgroup_free_meta_prealloc(inode->root, qgroup_to_release);
5940         else
5941                 btrfs_qgroup_convert_reserved_meta(inode->root,
5942                                                    qgroup_to_release);
5943 }
5944
5945 /**
5946  * btrfs_delayed_refs_rsv_release - release a ref head's reservation.
5947  * @fs_info - the fs_info for our fs.
5948  * @nr - the number of items to drop.
5949  *
5950  * This drops the delayed ref head's count from the delayed refs rsv and frees
5951  * any excess reservation we had.
5952  */
5953 void btrfs_delayed_refs_rsv_release(struct btrfs_fs_info *fs_info, int nr)
5954 {
5955         struct btrfs_block_rsv *block_rsv = &fs_info->delayed_refs_rsv;
5956         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
5957         u64 num_bytes = btrfs_calc_trans_metadata_size(fs_info, nr);
5958         u64 released = 0;
5959
5960         released = block_rsv_release_bytes(fs_info, block_rsv, global_rsv,
5961                                            num_bytes, NULL);
5962         if (released)
5963                 trace_btrfs_space_reservation(fs_info, "delayed_refs_rsv",
5964                                               0, released, 0);
5965 }
5966
5967 static void update_global_block_rsv(struct btrfs_fs_info *fs_info)
5968 {
5969         struct btrfs_block_rsv *block_rsv = &fs_info->global_block_rsv;
5970         struct btrfs_space_info *sinfo = block_rsv->space_info;
5971         u64 num_bytes;
5972
5973         /*
5974          * The global block rsv is based on the size of the extent tree, the
5975          * checksum tree and the root tree.  If the fs is empty we want to set
5976          * it to a minimal amount for safety.
5977          */
5978         num_bytes = btrfs_root_used(&fs_info->extent_root->root_item) +
5979                 btrfs_root_used(&fs_info->csum_root->root_item) +
5980                 btrfs_root_used(&fs_info->tree_root->root_item);
5981         num_bytes = max_t(u64, num_bytes, SZ_16M);
5982
5983         spin_lock(&sinfo->lock);
5984         spin_lock(&block_rsv->lock);
5985
5986         block_rsv->size = min_t(u64, num_bytes, SZ_512M);
5987
5988         if (block_rsv->reserved < block_rsv->size) {
5989                 num_bytes = btrfs_space_info_used(sinfo, true);
5990                 if (sinfo->total_bytes > num_bytes) {
5991                         num_bytes = sinfo->total_bytes - num_bytes;
5992                         num_bytes = min(num_bytes,
5993                                         block_rsv->size - block_rsv->reserved);
5994                         block_rsv->reserved += num_bytes;
5995                         update_bytes_may_use(sinfo, num_bytes);
5996                         trace_btrfs_space_reservation(fs_info, "space_info",
5997                                                       sinfo->flags, num_bytes,
5998                                                       1);
5999                 }
6000         } else if (block_rsv->reserved > block_rsv->size) {
6001                 num_bytes = block_rsv->reserved - block_rsv->size;
6002                 update_bytes_may_use(sinfo, -num_bytes);
6003                 trace_btrfs_space_reservation(fs_info, "space_info",
6004                                       sinfo->flags, num_bytes, 0);
6005                 block_rsv->reserved = block_rsv->size;
6006         }
6007
6008         if (block_rsv->reserved == block_rsv->size)
6009                 block_rsv->full = 1;
6010         else
6011                 block_rsv->full = 0;
6012
6013         spin_unlock(&block_rsv->lock);
6014         spin_unlock(&sinfo->lock);
6015 }
6016
6017 static void init_global_block_rsv(struct btrfs_fs_info *fs_info)
6018 {
6019         struct btrfs_space_info *space_info;
6020
6021         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
6022         fs_info->chunk_block_rsv.space_info = space_info;
6023
6024         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
6025         fs_info->global_block_rsv.space_info = space_info;
6026         fs_info->trans_block_rsv.space_info = space_info;
6027         fs_info->empty_block_rsv.space_info = space_info;
6028         fs_info->delayed_block_rsv.space_info = space_info;
6029         fs_info->delayed_refs_rsv.space_info = space_info;
6030
6031         fs_info->extent_root->block_rsv = &fs_info->delayed_refs_rsv;
6032         fs_info->csum_root->block_rsv = &fs_info->delayed_refs_rsv;
6033         fs_info->dev_root->block_rsv = &fs_info->global_block_rsv;
6034         fs_info->tree_root->block_rsv = &fs_info->global_block_rsv;
6035         if (fs_info->quota_root)
6036                 fs_info->quota_root->block_rsv = &fs_info->global_block_rsv;
6037         fs_info->chunk_root->block_rsv = &fs_info->chunk_block_rsv;
6038
6039         update_global_block_rsv(fs_info);
6040 }
6041
6042 static void release_global_block_rsv(struct btrfs_fs_info *fs_info)
6043 {
6044         block_rsv_release_bytes(fs_info, &fs_info->global_block_rsv, NULL,
6045                                 (u64)-1, NULL);
6046         WARN_ON(fs_info->trans_block_rsv.size > 0);
6047         WARN_ON(fs_info->trans_block_rsv.reserved > 0);
6048         WARN_ON(fs_info->chunk_block_rsv.size > 0);
6049         WARN_ON(fs_info->chunk_block_rsv.reserved > 0);
6050         WARN_ON(fs_info->delayed_block_rsv.size > 0);
6051         WARN_ON(fs_info->delayed_block_rsv.reserved > 0);
6052         WARN_ON(fs_info->delayed_refs_rsv.reserved > 0);
6053         WARN_ON(fs_info->delayed_refs_rsv.size > 0);
6054 }
6055
6056 /*
6057  * btrfs_update_delayed_refs_rsv - adjust the size of the delayed refs rsv
6058  * @trans - the trans that may have generated delayed refs
6059  *
6060  * This is to be called anytime we may have adjusted trans->delayed_ref_updates,
6061  * it'll calculate the additional size and add it to the delayed_refs_rsv.
6062  */
6063 void btrfs_update_delayed_refs_rsv(struct btrfs_trans_handle *trans)
6064 {
6065         struct btrfs_fs_info *fs_info = trans->fs_info;
6066         struct btrfs_block_rsv *delayed_rsv = &fs_info->delayed_refs_rsv;
6067         u64 num_bytes;
6068
6069         if (!trans->delayed_ref_updates)
6070                 return;
6071
6072         num_bytes = btrfs_calc_trans_metadata_size(fs_info,
6073                                                    trans->delayed_ref_updates);
6074         spin_lock(&delayed_rsv->lock);
6075         delayed_rsv->size += num_bytes;
6076         delayed_rsv->full = 0;
6077         spin_unlock(&delayed_rsv->lock);
6078         trans->delayed_ref_updates = 0;
6079 }
6080
6081 /*
6082  * To be called after all the new block groups attached to the transaction
6083  * handle have been created (btrfs_create_pending_block_groups()).
6084  */
6085 void btrfs_trans_release_chunk_metadata(struct btrfs_trans_handle *trans)
6086 {
6087         struct btrfs_fs_info *fs_info = trans->fs_info;
6088
6089         if (!trans->chunk_bytes_reserved)
6090                 return;
6091
6092         WARN_ON_ONCE(!list_empty(&trans->new_bgs));
6093
6094         block_rsv_release_bytes(fs_info, &fs_info->chunk_block_rsv, NULL,
6095                                 trans->chunk_bytes_reserved, NULL);
6096         trans->chunk_bytes_reserved = 0;
6097 }
6098
6099 /*
6100  * btrfs_subvolume_reserve_metadata() - reserve space for subvolume operation
6101  * root: the root of the parent directory
6102  * rsv: block reservation
6103  * items: the number of items that we need do reservation
6104  * use_global_rsv: allow fallback to the global block reservation
6105  *
6106  * This function is used to reserve the space for snapshot/subvolume
6107  * creation and deletion. Those operations are different with the
6108  * common file/directory operations, they change two fs/file trees
6109  * and root tree, the number of items that the qgroup reserves is
6110  * different with the free space reservation. So we can not use
6111  * the space reservation mechanism in start_transaction().
6112  */
6113 int btrfs_subvolume_reserve_metadata(struct btrfs_root *root,
6114                                      struct btrfs_block_rsv *rsv, int items,
6115                                      bool use_global_rsv)
6116 {
6117         u64 qgroup_num_bytes = 0;
6118         u64 num_bytes;
6119         int ret;
6120         struct btrfs_fs_info *fs_info = root->fs_info;
6121         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
6122
6123         if (test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags)) {
6124                 /* One for parent inode, two for dir entries */
6125                 qgroup_num_bytes = 3 * fs_info->nodesize;
6126                 ret = btrfs_qgroup_reserve_meta_prealloc(root,
6127                                 qgroup_num_bytes, true);
6128                 if (ret)
6129                         return ret;
6130         }
6131
6132         num_bytes = btrfs_calc_trans_metadata_size(fs_info, items);
6133         rsv->space_info = __find_space_info(fs_info,
6134                                             BTRFS_BLOCK_GROUP_METADATA);
6135         ret = btrfs_block_rsv_add(root, rsv, num_bytes,
6136                                   BTRFS_RESERVE_FLUSH_ALL);
6137
6138         if (ret == -ENOSPC && use_global_rsv)
6139                 ret = btrfs_block_rsv_migrate(global_rsv, rsv, num_bytes, true);
6140
6141         if (ret && qgroup_num_bytes)
6142                 btrfs_qgroup_free_meta_prealloc(root, qgroup_num_bytes);
6143
6144         return ret;
6145 }
6146
6147 void btrfs_subvolume_release_metadata(struct btrfs_fs_info *fs_info,
6148                                       struct btrfs_block_rsv *rsv)
6149 {
6150         btrfs_block_rsv_release(fs_info, rsv, (u64)-1);
6151 }
6152
6153 static void btrfs_calculate_inode_block_rsv_size(struct btrfs_fs_info *fs_info,
6154                                                  struct btrfs_inode *inode)
6155 {
6156         struct btrfs_block_rsv *block_rsv = &inode->block_rsv;
6157         u64 reserve_size = 0;
6158         u64 qgroup_rsv_size = 0;
6159         u64 csum_leaves;
6160         unsigned outstanding_extents;
6161
6162         lockdep_assert_held(&inode->lock);
6163         outstanding_extents = inode->outstanding_extents;
6164         if (outstanding_extents)
6165                 reserve_size = btrfs_calc_trans_metadata_size(fs_info,
6166                                                 outstanding_extents + 1);
6167         csum_leaves = btrfs_csum_bytes_to_leaves(fs_info,
6168                                                  inode->csum_bytes);
6169         reserve_size += btrfs_calc_trans_metadata_size(fs_info,
6170                                                        csum_leaves);
6171         /*
6172          * For qgroup rsv, the calculation is very simple:
6173          * account one nodesize for each outstanding extent
6174          *
6175          * This is overestimating in most cases.
6176          */
6177         qgroup_rsv_size = (u64)outstanding_extents * fs_info->nodesize;
6178
6179         spin_lock(&block_rsv->lock);
6180         block_rsv->size = reserve_size;
6181         block_rsv->qgroup_rsv_size = qgroup_rsv_size;
6182         spin_unlock(&block_rsv->lock);
6183 }
6184
6185 int btrfs_delalloc_reserve_metadata(struct btrfs_inode *inode, u64 num_bytes)
6186 {
6187         struct btrfs_fs_info *fs_info = inode->root->fs_info;
6188         unsigned nr_extents;
6189         enum btrfs_reserve_flush_enum flush = BTRFS_RESERVE_FLUSH_ALL;
6190         int ret = 0;
6191         bool delalloc_lock = true;
6192
6193         /* If we are a free space inode we need to not flush since we will be in
6194          * the middle of a transaction commit.  We also don't need the delalloc
6195          * mutex since we won't race with anybody.  We need this mostly to make
6196          * lockdep shut its filthy mouth.
6197          *
6198          * If we have a transaction open (can happen if we call truncate_block
6199          * from truncate), then we need FLUSH_LIMIT so we don't deadlock.
6200          */
6201         if (btrfs_is_free_space_inode(inode)) {
6202                 flush = BTRFS_RESERVE_NO_FLUSH;
6203                 delalloc_lock = false;
6204         } else {
6205                 if (current->journal_info)
6206                         flush = BTRFS_RESERVE_FLUSH_LIMIT;
6207
6208                 if (btrfs_transaction_in_commit(fs_info))
6209                         schedule_timeout(1);
6210         }
6211
6212         if (delalloc_lock)
6213                 mutex_lock(&inode->delalloc_mutex);
6214
6215         num_bytes = ALIGN(num_bytes, fs_info->sectorsize);
6216
6217         /* Add our new extents and calculate the new rsv size. */
6218         spin_lock(&inode->lock);
6219         nr_extents = count_max_extents(num_bytes);
6220         btrfs_mod_outstanding_extents(inode, nr_extents);
6221         inode->csum_bytes += num_bytes;
6222         btrfs_calculate_inode_block_rsv_size(fs_info, inode);
6223         spin_unlock(&inode->lock);
6224
6225         ret = btrfs_inode_rsv_refill(inode, flush);
6226         if (unlikely(ret))
6227                 goto out_fail;
6228
6229         if (delalloc_lock)
6230                 mutex_unlock(&inode->delalloc_mutex);
6231         return 0;
6232
6233 out_fail:
6234         spin_lock(&inode->lock);
6235         nr_extents = count_max_extents(num_bytes);
6236         btrfs_mod_outstanding_extents(inode, -nr_extents);
6237         inode->csum_bytes -= num_bytes;
6238         btrfs_calculate_inode_block_rsv_size(fs_info, inode);
6239         spin_unlock(&inode->lock);
6240
6241         btrfs_inode_rsv_release(inode, true);
6242         if (delalloc_lock)
6243                 mutex_unlock(&inode->delalloc_mutex);
6244         return ret;
6245 }
6246
6247 /**
6248  * btrfs_delalloc_release_metadata - release a metadata reservation for an inode
6249  * @inode: the inode to release the reservation for.
6250  * @num_bytes: the number of bytes we are releasing.
6251  * @qgroup_free: free qgroup reservation or convert it to per-trans reservation
6252  *
6253  * This will release the metadata reservation for an inode.  This can be called
6254  * once we complete IO for a given set of bytes to release their metadata
6255  * reservations, or on error for the same reason.
6256  */
6257 void btrfs_delalloc_release_metadata(struct btrfs_inode *inode, u64 num_bytes,
6258                                      bool qgroup_free)
6259 {
6260         struct btrfs_fs_info *fs_info = inode->root->fs_info;
6261
6262         num_bytes = ALIGN(num_bytes, fs_info->sectorsize);
6263         spin_lock(&inode->lock);
6264         inode->csum_bytes -= num_bytes;
6265         btrfs_calculate_inode_block_rsv_size(fs_info, inode);
6266         spin_unlock(&inode->lock);
6267
6268         if (btrfs_is_testing(fs_info))
6269                 return;
6270
6271         btrfs_inode_rsv_release(inode, qgroup_free);
6272 }
6273
6274 /**
6275  * btrfs_delalloc_release_extents - release our outstanding_extents
6276  * @inode: the inode to balance the reservation for.
6277  * @num_bytes: the number of bytes we originally reserved with
6278  * @qgroup_free: do we need to free qgroup meta reservation or convert them.
6279  *
6280  * When we reserve space we increase outstanding_extents for the extents we may
6281  * add.  Once we've set the range as delalloc or created our ordered extents we
6282  * have outstanding_extents to track the real usage, so we use this to free our
6283  * temporarily tracked outstanding_extents.  This _must_ be used in conjunction
6284  * with btrfs_delalloc_reserve_metadata.
6285  */
6286 void btrfs_delalloc_release_extents(struct btrfs_inode *inode, u64 num_bytes,
6287                                     bool qgroup_free)
6288 {
6289         struct btrfs_fs_info *fs_info = inode->root->fs_info;
6290         unsigned num_extents;
6291
6292         spin_lock(&inode->lock);
6293         num_extents = count_max_extents(num_bytes);
6294         btrfs_mod_outstanding_extents(inode, -num_extents);
6295         btrfs_calculate_inode_block_rsv_size(fs_info, inode);
6296         spin_unlock(&inode->lock);
6297
6298         if (btrfs_is_testing(fs_info))
6299                 return;
6300
6301         btrfs_inode_rsv_release(inode, qgroup_free);
6302 }
6303
6304 /**
6305  * btrfs_delalloc_reserve_space - reserve data and metadata space for
6306  * delalloc
6307  * @inode: inode we're writing to
6308  * @start: start range we are writing to
6309  * @len: how long the range we are writing to
6310  * @reserved: mandatory parameter, record actually reserved qgroup ranges of
6311  *            current reservation.
6312  *
6313  * This will do the following things
6314  *
6315  * o reserve space in data space info for num bytes
6316  *   and reserve precious corresponding qgroup space
6317  *   (Done in check_data_free_space)
6318  *
6319  * o reserve space for metadata space, based on the number of outstanding
6320  *   extents and how much csums will be needed
6321  *   also reserve metadata space in a per root over-reserve method.
6322  * o add to the inodes->delalloc_bytes
6323  * o add it to the fs_info's delalloc inodes list.
6324  *   (Above 3 all done in delalloc_reserve_metadata)
6325  *
6326  * Return 0 for success
6327  * Return <0 for error(-ENOSPC or -EQUOT)
6328  */
6329 int btrfs_delalloc_reserve_space(struct inode *inode,
6330                         struct extent_changeset **reserved, u64 start, u64 len)
6331 {
6332         int ret;
6333
6334         ret = btrfs_check_data_free_space(inode, reserved, start, len);
6335         if (ret < 0)
6336                 return ret;
6337         ret = btrfs_delalloc_reserve_metadata(BTRFS_I(inode), len);
6338         if (ret < 0)
6339                 btrfs_free_reserved_data_space(inode, *reserved, start, len);
6340         return ret;
6341 }
6342
6343 /**
6344  * btrfs_delalloc_release_space - release data and metadata space for delalloc
6345  * @inode: inode we're releasing space for
6346  * @start: start position of the space already reserved
6347  * @len: the len of the space already reserved
6348  * @release_bytes: the len of the space we consumed or didn't use
6349  *
6350  * This function will release the metadata space that was not used and will
6351  * decrement ->delalloc_bytes and remove it from the fs_info delalloc_inodes
6352  * list if there are no delalloc bytes left.
6353  * Also it will handle the qgroup reserved space.
6354  */
6355 void btrfs_delalloc_release_space(struct inode *inode,
6356                                   struct extent_changeset *reserved,
6357                                   u64 start, u64 len, bool qgroup_free)
6358 {
6359         btrfs_delalloc_release_metadata(BTRFS_I(inode), len, qgroup_free);
6360         btrfs_free_reserved_data_space(inode, reserved, start, len);
6361 }
6362
6363 static int update_block_group(struct btrfs_trans_handle *trans,
6364                               struct btrfs_fs_info *info, u64 bytenr,
6365                               u64 num_bytes, int alloc)
6366 {
6367         struct btrfs_block_group_cache *cache = NULL;
6368         u64 total = num_bytes;
6369         u64 old_val;
6370         u64 byte_in_group;
6371         int factor;
6372         int ret = 0;
6373
6374         /* block accounting for super block */
6375         spin_lock(&info->delalloc_root_lock);
6376         old_val = btrfs_super_bytes_used(info->super_copy);
6377         if (alloc)
6378                 old_val += num_bytes;
6379         else
6380                 old_val -= num_bytes;
6381         btrfs_set_super_bytes_used(info->super_copy, old_val);
6382         spin_unlock(&info->delalloc_root_lock);
6383
6384         while (total) {
6385                 cache = btrfs_lookup_block_group(info, bytenr);
6386                 if (!cache) {
6387                         ret = -ENOENT;
6388                         break;
6389                 }
6390                 factor = btrfs_bg_type_to_factor(cache->flags);
6391
6392                 /*
6393                  * If this block group has free space cache written out, we
6394                  * need to make sure to load it if we are removing space.  This
6395                  * is because we need the unpinning stage to actually add the
6396                  * space back to the block group, otherwise we will leak space.
6397                  */
6398                 if (!alloc && cache->cached == BTRFS_CACHE_NO)
6399                         cache_block_group(cache, 1);
6400
6401                 byte_in_group = bytenr - cache->key.objectid;
6402                 WARN_ON(byte_in_group > cache->key.offset);
6403
6404                 spin_lock(&cache->space_info->lock);
6405                 spin_lock(&cache->lock);
6406
6407                 if (btrfs_test_opt(info, SPACE_CACHE) &&
6408                     cache->disk_cache_state < BTRFS_DC_CLEAR)
6409                         cache->disk_cache_state = BTRFS_DC_CLEAR;
6410
6411                 old_val = btrfs_block_group_used(&cache->item);
6412                 num_bytes = min(total, cache->key.offset - byte_in_group);
6413                 if (alloc) {
6414                         old_val += num_bytes;
6415                         btrfs_set_block_group_used(&cache->item, old_val);
6416                         cache->reserved -= num_bytes;
6417                         cache->space_info->bytes_reserved -= num_bytes;
6418                         cache->space_info->bytes_used += num_bytes;
6419                         cache->space_info->disk_used += num_bytes * factor;
6420                         spin_unlock(&cache->lock);
6421                         spin_unlock(&cache->space_info->lock);
6422                 } else {
6423                         old_val -= num_bytes;
6424                         btrfs_set_block_group_used(&cache->item, old_val);
6425                         cache->pinned += num_bytes;
6426                         update_bytes_pinned(cache->space_info, num_bytes);
6427                         cache->space_info->bytes_used -= num_bytes;
6428                         cache->space_info->disk_used -= num_bytes * factor;
6429                         spin_unlock(&cache->lock);
6430                         spin_unlock(&cache->space_info->lock);
6431
6432                         trace_btrfs_space_reservation(info, "pinned",
6433                                                       cache->space_info->flags,
6434                                                       num_bytes, 1);
6435                         percpu_counter_add_batch(&cache->space_info->total_bytes_pinned,
6436                                            num_bytes,
6437                                            BTRFS_TOTAL_BYTES_PINNED_BATCH);
6438                         set_extent_dirty(info->pinned_extents,
6439                                          bytenr, bytenr + num_bytes - 1,
6440                                          GFP_NOFS | __GFP_NOFAIL);
6441                 }
6442
6443                 spin_lock(&trans->transaction->dirty_bgs_lock);
6444                 if (list_empty(&cache->dirty_list)) {
6445                         list_add_tail(&cache->dirty_list,
6446                                       &trans->transaction->dirty_bgs);
6447                         trans->transaction->num_dirty_bgs++;
6448                         trans->delayed_ref_updates++;
6449                         btrfs_get_block_group(cache);
6450                 }
6451                 spin_unlock(&trans->transaction->dirty_bgs_lock);
6452
6453                 /*
6454                  * No longer have used bytes in this block group, queue it for
6455                  * deletion. We do this after adding the block group to the
6456                  * dirty list to avoid races between cleaner kthread and space
6457                  * cache writeout.
6458                  */
6459                 if (!alloc && old_val == 0)
6460                         btrfs_mark_bg_unused(cache);
6461
6462                 btrfs_put_block_group(cache);
6463                 total -= num_bytes;
6464                 bytenr += num_bytes;
6465         }
6466
6467         /* Modified block groups are accounted for in the delayed_refs_rsv. */
6468         btrfs_update_delayed_refs_rsv(trans);
6469         return ret;
6470 }
6471
6472 static u64 first_logical_byte(struct btrfs_fs_info *fs_info, u64 search_start)
6473 {
6474         struct btrfs_block_group_cache *cache;
6475         u64 bytenr;
6476
6477         spin_lock(&fs_info->block_group_cache_lock);
6478         bytenr = fs_info->first_logical_byte;
6479         spin_unlock(&fs_info->block_group_cache_lock);
6480
6481         if (bytenr < (u64)-1)
6482                 return bytenr;
6483
6484         cache = btrfs_lookup_first_block_group(fs_info, search_start);
6485         if (!cache)
6486                 return 0;
6487
6488         bytenr = cache->key.objectid;
6489         btrfs_put_block_group(cache);
6490
6491         return bytenr;
6492 }
6493
6494 static int pin_down_extent(struct btrfs_fs_info *fs_info,
6495                            struct btrfs_block_group_cache *cache,
6496                            u64 bytenr, u64 num_bytes, int reserved)
6497 {
6498         spin_lock(&cache->space_info->lock);
6499         spin_lock(&cache->lock);
6500         cache->pinned += num_bytes;
6501         update_bytes_pinned(cache->space_info, num_bytes);
6502         if (reserved) {
6503                 cache->reserved -= num_bytes;
6504                 cache->space_info->bytes_reserved -= num_bytes;
6505         }
6506         spin_unlock(&cache->lock);
6507         spin_unlock(&cache->space_info->lock);
6508
6509         trace_btrfs_space_reservation(fs_info, "pinned",
6510                                       cache->space_info->flags, num_bytes, 1);
6511         percpu_counter_add_batch(&cache->space_info->total_bytes_pinned,
6512                     num_bytes, BTRFS_TOTAL_BYTES_PINNED_BATCH);
6513         set_extent_dirty(fs_info->pinned_extents, bytenr,
6514                          bytenr + num_bytes - 1, GFP_NOFS | __GFP_NOFAIL);
6515         return 0;
6516 }
6517
6518 /*
6519  * this function must be called within transaction
6520  */
6521 int btrfs_pin_extent(struct btrfs_fs_info *fs_info,
6522                      u64 bytenr, u64 num_bytes, int reserved)
6523 {
6524         struct btrfs_block_group_cache *cache;
6525
6526         cache = btrfs_lookup_block_group(fs_info, bytenr);
6527         BUG_ON(!cache); /* Logic error */
6528
6529         pin_down_extent(fs_info, cache, bytenr, num_bytes, reserved);
6530
6531         btrfs_put_block_group(cache);
6532         return 0;
6533 }
6534
6535 /*
6536  * this function must be called within transaction
6537  */
6538 int btrfs_pin_extent_for_log_replay(struct btrfs_fs_info *fs_info,
6539                                     u64 bytenr, u64 num_bytes)
6540 {
6541         struct btrfs_block_group_cache *cache;
6542         int ret;
6543
6544         cache = btrfs_lookup_block_group(fs_info, bytenr);
6545         if (!cache)
6546                 return -EINVAL;
6547
6548         /*
6549          * pull in the free space cache (if any) so that our pin
6550          * removes the free space from the cache.  We have load_only set
6551          * to one because the slow code to read in the free extents does check
6552          * the pinned extents.
6553          */
6554         cache_block_group(cache, 1);
6555
6556         pin_down_extent(fs_info, cache, bytenr, num_bytes, 0);
6557
6558         /* remove us from the free space cache (if we're there at all) */
6559         ret = btrfs_remove_free_space(cache, bytenr, num_bytes);
6560         btrfs_put_block_group(cache);
6561         return ret;
6562 }
6563
6564 static int __exclude_logged_extent(struct btrfs_fs_info *fs_info,
6565                                    u64 start, u64 num_bytes)
6566 {
6567         int ret;
6568         struct btrfs_block_group_cache *block_group;
6569         struct btrfs_caching_control *caching_ctl;
6570
6571         block_group = btrfs_lookup_block_group(fs_info, start);
6572         if (!block_group)
6573                 return -EINVAL;
6574
6575         cache_block_group(block_group, 0);
6576         caching_ctl = get_caching_control(block_group);
6577
6578         if (!caching_ctl) {
6579                 /* Logic error */
6580                 BUG_ON(!block_group_cache_done(block_group));
6581                 ret = btrfs_remove_free_space(block_group, start, num_bytes);
6582         } else {
6583                 mutex_lock(&caching_ctl->mutex);
6584
6585                 if (start >= caching_ctl->progress) {
6586                         ret = add_excluded_extent(fs_info, start, num_bytes);
6587                 } else if (start + num_bytes <= caching_ctl->progress) {
6588                         ret = btrfs_remove_free_space(block_group,
6589                                                       start, num_bytes);
6590                 } else {
6591                         num_bytes = caching_ctl->progress - start;
6592                         ret = btrfs_remove_free_space(block_group,
6593                                                       start, num_bytes);
6594                         if (ret)
6595                                 goto out_lock;
6596
6597                         num_bytes = (start + num_bytes) -
6598                                 caching_ctl->progress;
6599                         start = caching_ctl->progress;
6600                         ret = add_excluded_extent(fs_info, start, num_bytes);
6601                 }
6602 out_lock:
6603                 mutex_unlock(&caching_ctl->mutex);
6604                 put_caching_control(caching_ctl);
6605         }
6606         btrfs_put_block_group(block_group);
6607         return ret;
6608 }
6609
6610 int btrfs_exclude_logged_extents(struct btrfs_fs_info *fs_info,
6611                                  struct extent_buffer *eb)
6612 {
6613         struct btrfs_file_extent_item *item;
6614         struct btrfs_key key;
6615         int found_type;
6616         int i;
6617         int ret = 0;
6618
6619         if (!btrfs_fs_incompat(fs_info, MIXED_GROUPS))
6620                 return 0;
6621
6622         for (i = 0; i < btrfs_header_nritems(eb); i++) {
6623                 btrfs_item_key_to_cpu(eb, &key, i);
6624                 if (key.type != BTRFS_EXTENT_DATA_KEY)
6625                         continue;
6626                 item = btrfs_item_ptr(eb, i, struct btrfs_file_extent_item);
6627                 found_type = btrfs_file_extent_type(eb, item);
6628                 if (found_type == BTRFS_FILE_EXTENT_INLINE)
6629                         continue;
6630                 if (btrfs_file_extent_disk_bytenr(eb, item) == 0)
6631                         continue;
6632                 key.objectid = btrfs_file_extent_disk_bytenr(eb, item);
6633                 key.offset = btrfs_file_extent_disk_num_bytes(eb, item);
6634                 ret = __exclude_logged_extent(fs_info, key.objectid, key.offset);
6635                 if (ret)
6636                         break;
6637         }
6638
6639         return ret;
6640 }
6641
6642 static void
6643 btrfs_inc_block_group_reservations(struct btrfs_block_group_cache *bg)
6644 {
6645         atomic_inc(&bg->reservations);
6646 }
6647
6648 void btrfs_dec_block_group_reservations(struct btrfs_fs_info *fs_info,
6649                                         const u64 start)
6650 {
6651         struct btrfs_block_group_cache *bg;
6652
6653         bg = btrfs_lookup_block_group(fs_info, start);
6654         ASSERT(bg);
6655         if (atomic_dec_and_test(&bg->reservations))
6656                 wake_up_var(&bg->reservations);
6657         btrfs_put_block_group(bg);
6658 }
6659
6660 void btrfs_wait_block_group_reservations(struct btrfs_block_group_cache *bg)
6661 {
6662         struct btrfs_space_info *space_info = bg->space_info;
6663
6664         ASSERT(bg->ro);
6665
6666         if (!(bg->flags & BTRFS_BLOCK_GROUP_DATA))
6667                 return;
6668
6669         /*
6670          * Our block group is read only but before we set it to read only,
6671          * some task might have had allocated an extent from it already, but it
6672          * has not yet created a respective ordered extent (and added it to a
6673          * root's list of ordered extents).
6674          * Therefore wait for any task currently allocating extents, since the
6675          * block group's reservations counter is incremented while a read lock
6676          * on the groups' semaphore is held and decremented after releasing
6677          * the read access on that semaphore and creating the ordered extent.
6678          */
6679         down_write(&space_info->groups_sem);
6680         up_write(&space_info->groups_sem);
6681
6682         wait_var_event(&bg->reservations, !atomic_read(&bg->reservations));
6683 }
6684
6685 /**
6686  * btrfs_add_reserved_bytes - update the block_group and space info counters
6687  * @cache:      The cache we are manipulating
6688  * @ram_bytes:  The number of bytes of file content, and will be same to
6689  *              @num_bytes except for the compress path.
6690  * @num_bytes:  The number of bytes in question
6691  * @delalloc:   The blocks are allocated for the delalloc write
6692  *
6693  * This is called by the allocator when it reserves space. If this is a
6694  * reservation and the block group has become read only we cannot make the
6695  * reservation and return -EAGAIN, otherwise this function always succeeds.
6696  */
6697 static int btrfs_add_reserved_bytes(struct btrfs_block_group_cache *cache,
6698                                     u64 ram_bytes, u64 num_bytes, int delalloc)
6699 {
6700         struct btrfs_space_info *space_info = cache->space_info;
6701         int ret = 0;
6702
6703         spin_lock(&space_info->lock);
6704         spin_lock(&cache->lock);
6705         if (cache->ro) {
6706                 ret = -EAGAIN;
6707         } else {
6708                 cache->reserved += num_bytes;
6709                 space_info->bytes_reserved += num_bytes;
6710                 update_bytes_may_use(space_info, -ram_bytes);
6711                 if (delalloc)
6712                         cache->delalloc_bytes += num_bytes;
6713         }
6714         spin_unlock(&cache->lock);
6715         spin_unlock(&space_info->lock);
6716         return ret;
6717 }
6718
6719 /**
6720  * btrfs_free_reserved_bytes - update the block_group and space info counters
6721  * @cache:      The cache we are manipulating
6722  * @num_bytes:  The number of bytes in question
6723  * @delalloc:   The blocks are allocated for the delalloc write
6724  *
6725  * This is called by somebody who is freeing space that was never actually used
6726  * on disk.  For example if you reserve some space for a new leaf in transaction
6727  * A and before transaction A commits you free that leaf, you call this with
6728  * reserve set to 0 in order to clear the reservation.
6729  */
6730
6731 static void btrfs_free_reserved_bytes(struct btrfs_block_group_cache *cache,
6732                                       u64 num_bytes, int delalloc)
6733 {
6734         struct btrfs_space_info *space_info = cache->space_info;
6735
6736         spin_lock(&space_info->lock);
6737         spin_lock(&cache->lock);
6738         if (cache->ro)
6739                 space_info->bytes_readonly += num_bytes;
6740         cache->reserved -= num_bytes;
6741         space_info->bytes_reserved -= num_bytes;
6742         space_info->max_extent_size = 0;
6743
6744         if (delalloc)
6745                 cache->delalloc_bytes -= num_bytes;
6746         spin_unlock(&cache->lock);
6747         spin_unlock(&space_info->lock);
6748 }
6749 void btrfs_prepare_extent_commit(struct btrfs_fs_info *fs_info)
6750 {
6751         struct btrfs_caching_control *next;
6752         struct btrfs_caching_control *caching_ctl;
6753         struct btrfs_block_group_cache *cache;
6754
6755         down_write(&fs_info->commit_root_sem);
6756
6757         list_for_each_entry_safe(caching_ctl, next,
6758                                  &fs_info->caching_block_groups, list) {
6759                 cache = caching_ctl->block_group;
6760                 if (block_group_cache_done(cache)) {
6761                         cache->last_byte_to_unpin = (u64)-1;
6762                         list_del_init(&caching_ctl->list);
6763                         put_caching_control(caching_ctl);
6764                 } else {
6765                         cache->last_byte_to_unpin = caching_ctl->progress;
6766                 }
6767         }
6768
6769         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
6770                 fs_info->pinned_extents = &fs_info->freed_extents[1];
6771         else
6772                 fs_info->pinned_extents = &fs_info->freed_extents[0];
6773
6774         up_write(&fs_info->commit_root_sem);
6775
6776         update_global_block_rsv(fs_info);
6777 }
6778
6779 /*
6780  * Returns the free cluster for the given space info and sets empty_cluster to
6781  * what it should be based on the mount options.
6782  */
6783 static struct btrfs_free_cluster *
6784 fetch_cluster_info(struct btrfs_fs_info *fs_info,
6785                    struct btrfs_space_info *space_info, u64 *empty_cluster)
6786 {
6787         struct btrfs_free_cluster *ret = NULL;
6788
6789         *empty_cluster = 0;
6790         if (btrfs_mixed_space_info(space_info))
6791                 return ret;
6792
6793         if (space_info->flags & BTRFS_BLOCK_GROUP_METADATA) {
6794                 ret = &fs_info->meta_alloc_cluster;
6795                 if (btrfs_test_opt(fs_info, SSD))
6796                         *empty_cluster = SZ_2M;
6797                 else
6798                         *empty_cluster = SZ_64K;
6799         } else if ((space_info->flags & BTRFS_BLOCK_GROUP_DATA) &&
6800                    btrfs_test_opt(fs_info, SSD_SPREAD)) {
6801                 *empty_cluster = SZ_2M;
6802                 ret = &fs_info->data_alloc_cluster;
6803         }
6804
6805         return ret;
6806 }
6807
6808 static int unpin_extent_range(struct btrfs_fs_info *fs_info,
6809                               u64 start, u64 end,
6810                               const bool return_free_space)
6811 {
6812         struct btrfs_block_group_cache *cache = NULL;
6813         struct btrfs_space_info *space_info;
6814         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
6815         struct btrfs_free_cluster *cluster = NULL;
6816         u64 len;
6817         u64 total_unpinned = 0;
6818         u64 empty_cluster = 0;
6819         bool readonly;
6820
6821         while (start <= end) {
6822                 readonly = false;
6823                 if (!cache ||
6824                     start >= cache->key.objectid + cache->key.offset) {
6825                         if (cache)
6826                                 btrfs_put_block_group(cache);
6827                         total_unpinned = 0;
6828                         cache = btrfs_lookup_block_group(fs_info, start);
6829                         BUG_ON(!cache); /* Logic error */
6830
6831                         cluster = fetch_cluster_info(fs_info,
6832                                                      cache->space_info,
6833                                                      &empty_cluster);
6834                         empty_cluster <<= 1;
6835                 }
6836
6837                 len = cache->key.objectid + cache->key.offset - start;
6838                 len = min(len, end + 1 - start);
6839
6840                 if (start < cache->last_byte_to_unpin) {
6841                         len = min(len, cache->last_byte_to_unpin - start);
6842                         if (return_free_space)
6843                                 btrfs_add_free_space(cache, start, len);
6844                 }
6845
6846                 start += len;
6847                 total_unpinned += len;
6848                 space_info = cache->space_info;
6849
6850                 /*
6851                  * If this space cluster has been marked as fragmented and we've
6852                  * unpinned enough in this block group to potentially allow a
6853                  * cluster to be created inside of it go ahead and clear the
6854                  * fragmented check.
6855                  */
6856                 if (cluster && cluster->fragmented &&
6857                     total_unpinned > empty_cluster) {
6858                         spin_lock(&cluster->lock);
6859                         cluster->fragmented = 0;
6860                         spin_unlock(&cluster->lock);
6861                 }
6862
6863                 spin_lock(&space_info->lock);
6864                 spin_lock(&cache->lock);
6865                 cache->pinned -= len;
6866                 update_bytes_pinned(space_info, -len);
6867
6868                 trace_btrfs_space_reservation(fs_info, "pinned",
6869                                               space_info->flags, len, 0);
6870                 space_info->max_extent_size = 0;
6871                 percpu_counter_add_batch(&space_info->total_bytes_pinned,
6872                             -len, BTRFS_TOTAL_BYTES_PINNED_BATCH);
6873                 if (cache->ro) {
6874                         space_info->bytes_readonly += len;
6875                         readonly = true;
6876                 }
6877                 spin_unlock(&cache->lock);
6878                 if (!readonly && return_free_space &&
6879                     global_rsv->space_info == space_info) {
6880                         u64 to_add = len;
6881
6882                         spin_lock(&global_rsv->lock);
6883                         if (!global_rsv->full) {
6884                                 to_add = min(len, global_rsv->size -
6885                                              global_rsv->reserved);
6886                                 global_rsv->reserved += to_add;
6887                                 update_bytes_may_use(space_info, to_add);
6888                                 if (global_rsv->reserved >= global_rsv->size)
6889                                         global_rsv->full = 1;
6890                                 trace_btrfs_space_reservation(fs_info,
6891                                                               "space_info",
6892                                                               space_info->flags,
6893                                                               to_add, 1);
6894                                 len -= to_add;
6895                         }
6896                         spin_unlock(&global_rsv->lock);
6897                         /* Add to any tickets we may have */
6898                         if (len)
6899                                 space_info_add_new_bytes(fs_info, space_info,
6900                                                          len);
6901                 }
6902                 spin_unlock(&space_info->lock);
6903         }
6904
6905         if (cache)
6906                 btrfs_put_block_group(cache);
6907         return 0;
6908 }
6909
6910 int btrfs_finish_extent_commit(struct btrfs_trans_handle *trans)
6911 {
6912         struct btrfs_fs_info *fs_info = trans->fs_info;
6913         struct btrfs_block_group_cache *block_group, *tmp;
6914         struct list_head *deleted_bgs;
6915         struct extent_io_tree *unpin;
6916         u64 start;
6917         u64 end;
6918         int ret;
6919
6920         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
6921                 unpin = &fs_info->freed_extents[1];
6922         else
6923                 unpin = &fs_info->freed_extents[0];
6924
6925         while (!trans->aborted) {
6926                 struct extent_state *cached_state = NULL;
6927
6928                 mutex_lock(&fs_info->unused_bg_unpin_mutex);
6929                 ret = find_first_extent_bit(unpin, 0, &start, &end,
6930                                             EXTENT_DIRTY, &cached_state);
6931                 if (ret) {
6932                         mutex_unlock(&fs_info->unused_bg_unpin_mutex);
6933                         break;
6934                 }
6935
6936                 if (btrfs_test_opt(fs_info, DISCARD))
6937                         ret = btrfs_discard_extent(fs_info, start,
6938                                                    end + 1 - start, NULL);
6939
6940                 clear_extent_dirty(unpin, start, end, &cached_state);
6941                 unpin_extent_range(fs_info, start, end, true);
6942                 mutex_unlock(&fs_info->unused_bg_unpin_mutex);
6943                 free_extent_state(cached_state);
6944                 cond_resched();
6945         }
6946
6947         /*
6948          * Transaction is finished.  We don't need the lock anymore.  We
6949          * do need to clean up the block groups in case of a transaction
6950          * abort.
6951          */
6952         deleted_bgs = &trans->transaction->deleted_bgs;
6953         list_for_each_entry_safe(block_group, tmp, deleted_bgs, bg_list) {
6954                 u64 trimmed = 0;
6955
6956                 ret = -EROFS;
6957                 if (!trans->aborted)
6958                         ret = btrfs_discard_extent(fs_info,
6959                                                    block_group->key.objectid,
6960                                                    block_group->key.offset,
6961                                                    &trimmed);
6962
6963                 list_del_init(&block_group->bg_list);
6964                 btrfs_put_block_group_trimming(block_group);
6965                 btrfs_put_block_group(block_group);
6966
6967                 if (ret) {
6968                         const char *errstr = btrfs_decode_error(ret);
6969                         btrfs_warn(fs_info,
6970                            "discard failed while removing blockgroup: errno=%d %s",
6971                                    ret, errstr);
6972                 }
6973         }
6974
6975         return 0;
6976 }
6977
6978 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
6979                                struct btrfs_delayed_ref_node *node, u64 parent,
6980                                u64 root_objectid, u64 owner_objectid,
6981                                u64 owner_offset, int refs_to_drop,
6982                                struct btrfs_delayed_extent_op *extent_op)
6983 {
6984         struct btrfs_fs_info *info = trans->fs_info;
6985         struct btrfs_key key;
6986         struct btrfs_path *path;
6987         struct btrfs_root *extent_root = info->extent_root;
6988         struct extent_buffer *leaf;
6989         struct btrfs_extent_item *ei;
6990         struct btrfs_extent_inline_ref *iref;
6991         int ret;
6992         int is_data;
6993         int extent_slot = 0;
6994         int found_extent = 0;
6995         int num_to_del = 1;
6996         u32 item_size;
6997         u64 refs;
6998         u64 bytenr = node->bytenr;
6999         u64 num_bytes = node->num_bytes;
7000         int last_ref = 0;
7001         bool skinny_metadata = btrfs_fs_incompat(info, SKINNY_METADATA);
7002
7003         path = btrfs_alloc_path();
7004         if (!path)
7005                 return -ENOMEM;
7006
7007         path->reada = READA_FORWARD;
7008         path->leave_spinning = 1;
7009
7010         is_data = owner_objectid >= BTRFS_FIRST_FREE_OBJECTID;
7011         BUG_ON(!is_data && refs_to_drop != 1);
7012
7013         if (is_data)
7014                 skinny_metadata = false;
7015
7016         ret = lookup_extent_backref(trans, path, &iref, bytenr, num_bytes,
7017                                     parent, root_objectid, owner_objectid,
7018                                     owner_offset);
7019         if (ret == 0) {
7020                 extent_slot = path->slots[0];
7021                 while (extent_slot >= 0) {
7022                         btrfs_item_key_to_cpu(path->nodes[0], &key,
7023                                               extent_slot);
7024                         if (key.objectid != bytenr)
7025                                 break;
7026                         if (key.type == BTRFS_EXTENT_ITEM_KEY &&
7027                             key.offset == num_bytes) {
7028                                 found_extent = 1;
7029                                 break;
7030                         }
7031                         if (key.type == BTRFS_METADATA_ITEM_KEY &&
7032                             key.offset == owner_objectid) {
7033                                 found_extent = 1;
7034                                 break;
7035                         }
7036                         if (path->slots[0] - extent_slot > 5)
7037                                 break;
7038                         extent_slot--;
7039                 }
7040
7041                 if (!found_extent) {
7042                         BUG_ON(iref);
7043                         ret = remove_extent_backref(trans, path, NULL,
7044                                                     refs_to_drop,
7045                                                     is_data, &last_ref);
7046                         if (ret) {
7047                                 btrfs_abort_transaction(trans, ret);
7048                                 goto out;
7049                         }
7050                         btrfs_release_path(path);
7051                         path->leave_spinning = 1;
7052
7053                         key.objectid = bytenr;
7054                         key.type = BTRFS_EXTENT_ITEM_KEY;
7055                         key.offset = num_bytes;
7056
7057                         if (!is_data && skinny_metadata) {
7058                                 key.type = BTRFS_METADATA_ITEM_KEY;
7059                                 key.offset = owner_objectid;
7060                         }
7061
7062                         ret = btrfs_search_slot(trans, extent_root,
7063                                                 &key, path, -1, 1);
7064                         if (ret > 0 && skinny_metadata && path->slots[0]) {
7065                                 /*
7066                                  * Couldn't find our skinny metadata item,
7067                                  * see if we have ye olde extent item.
7068                                  */
7069                                 path->slots[0]--;
7070                                 btrfs_item_key_to_cpu(path->nodes[0], &key,
7071                                                       path->slots[0]);
7072                                 if (key.objectid == bytenr &&
7073                                     key.type == BTRFS_EXTENT_ITEM_KEY &&
7074                                     key.offset == num_bytes)
7075                                         ret = 0;
7076                         }
7077
7078                         if (ret > 0 && skinny_metadata) {
7079                                 skinny_metadata = false;
7080                                 key.objectid = bytenr;
7081                                 key.type = BTRFS_EXTENT_ITEM_KEY;
7082                                 key.offset = num_bytes;
7083                                 btrfs_release_path(path);
7084                                 ret = btrfs_search_slot(trans, extent_root,
7085                                                         &key, path, -1, 1);
7086                         }
7087
7088                         if (ret) {
7089                                 btrfs_err(info,
7090                                           "umm, got %d back from search, was looking for %llu",
7091                                           ret, bytenr);
7092                                 if (ret > 0)
7093                                         btrfs_print_leaf(path->nodes[0]);
7094                         }
7095                         if (ret < 0) {
7096                                 btrfs_abort_transaction(trans, ret);
7097                                 goto out;
7098                         }
7099                         extent_slot = path->slots[0];
7100                 }
7101         } else if (WARN_ON(ret == -ENOENT)) {
7102                 btrfs_print_leaf(path->nodes[0]);
7103                 btrfs_err(info,
7104                         "unable to find ref byte nr %llu parent %llu root %llu  owner %llu offset %llu",
7105                         bytenr, parent, root_objectid, owner_objectid,
7106                         owner_offset);
7107                 btrfs_abort_transaction(trans, ret);
7108                 goto out;
7109         } else {
7110                 btrfs_abort_transaction(trans, ret);
7111                 goto out;
7112         }
7113
7114         leaf = path->nodes[0];
7115         item_size = btrfs_item_size_nr(leaf, extent_slot);
7116         if (unlikely(item_size < sizeof(*ei))) {
7117                 ret = -EINVAL;
7118                 btrfs_print_v0_err(info);
7119                 btrfs_abort_transaction(trans, ret);
7120                 goto out;
7121         }
7122         ei = btrfs_item_ptr(leaf, extent_slot,
7123                             struct btrfs_extent_item);
7124         if (owner_objectid < BTRFS_FIRST_FREE_OBJECTID &&
7125             key.type == BTRFS_EXTENT_ITEM_KEY) {
7126                 struct btrfs_tree_block_info *bi;
7127                 BUG_ON(item_size < sizeof(*ei) + sizeof(*bi));
7128                 bi = (struct btrfs_tree_block_info *)(ei + 1);
7129                 WARN_ON(owner_objectid != btrfs_tree_block_level(leaf, bi));
7130         }
7131
7132         refs = btrfs_extent_refs(leaf, ei);
7133         if (refs < refs_to_drop) {
7134                 btrfs_err(info,
7135                           "trying to drop %d refs but we only have %Lu for bytenr %Lu",
7136                           refs_to_drop, refs, bytenr);
7137                 ret = -EINVAL;
7138                 btrfs_abort_transaction(trans, ret);
7139                 goto out;
7140         }
7141         refs -= refs_to_drop;
7142
7143         if (refs > 0) {
7144                 if (extent_op)
7145                         __run_delayed_extent_op(extent_op, leaf, ei);
7146                 /*
7147                  * In the case of inline back ref, reference count will
7148                  * be updated by remove_extent_backref
7149                  */
7150                 if (iref) {
7151                         BUG_ON(!found_extent);
7152                 } else {
7153                         btrfs_set_extent_refs(leaf, ei, refs);
7154                         btrfs_mark_buffer_dirty(leaf);
7155                 }
7156                 if (found_extent) {
7157                         ret = remove_extent_backref(trans, path, iref,
7158                                                     refs_to_drop, is_data,
7159                                                     &last_ref);
7160                         if (ret) {
7161                                 btrfs_abort_transaction(trans, ret);
7162                                 goto out;
7163                         }
7164                 }
7165         } else {
7166                 if (found_extent) {
7167                         BUG_ON(is_data && refs_to_drop !=
7168                                extent_data_ref_count(path, iref));
7169                         if (iref) {
7170                                 BUG_ON(path->slots[0] != extent_slot);
7171                         } else {
7172                                 BUG_ON(path->slots[0] != extent_slot + 1);
7173                                 path->slots[0] = extent_slot;
7174                                 num_to_del = 2;
7175                         }
7176                 }
7177
7178                 last_ref = 1;
7179                 ret = btrfs_del_items(trans, extent_root, path, path->slots[0],
7180                                       num_to_del);
7181                 if (ret) {
7182                         btrfs_abort_transaction(trans, ret);
7183                         goto out;
7184                 }
7185                 btrfs_release_path(path);
7186
7187                 if (is_data) {
7188                         ret = btrfs_del_csums(trans, info, bytenr, num_bytes);
7189                         if (ret) {
7190                                 btrfs_abort_transaction(trans, ret);
7191                                 goto out;
7192                         }
7193                 }
7194
7195                 ret = add_to_free_space_tree(trans, bytenr, num_bytes);
7196                 if (ret) {
7197                         btrfs_abort_transaction(trans, ret);
7198                         goto out;
7199                 }
7200
7201                 ret = update_block_group(trans, info, bytenr, num_bytes, 0);
7202                 if (ret) {
7203                         btrfs_abort_transaction(trans, ret);
7204                         goto out;
7205                 }
7206         }
7207         btrfs_release_path(path);
7208
7209 out:
7210         btrfs_free_path(path);
7211         return ret;
7212 }
7213
7214 /*
7215  * when we free an block, it is possible (and likely) that we free the last
7216  * delayed ref for that extent as well.  This searches the delayed ref tree for
7217  * a given extent, and if there are no other delayed refs to be processed, it
7218  * removes it from the tree.
7219  */
7220 static noinline int check_ref_cleanup(struct btrfs_trans_handle *trans,
7221                                       u64 bytenr)
7222 {
7223         struct btrfs_delayed_ref_head *head;
7224         struct btrfs_delayed_ref_root *delayed_refs;
7225         int ret = 0;
7226
7227         delayed_refs = &trans->transaction->delayed_refs;
7228         spin_lock(&delayed_refs->lock);
7229         head = btrfs_find_delayed_ref_head(delayed_refs, bytenr);
7230         if (!head)
7231                 goto out_delayed_unlock;
7232
7233         spin_lock(&head->lock);
7234         if (!RB_EMPTY_ROOT(&head->ref_tree.rb_root))
7235                 goto out;
7236
7237         if (cleanup_extent_op(head) != NULL)
7238                 goto out;
7239
7240         /*
7241          * waiting for the lock here would deadlock.  If someone else has it
7242          * locked they are already in the process of dropping it anyway
7243          */
7244         if (!mutex_trylock(&head->mutex))
7245                 goto out;
7246
7247         btrfs_delete_ref_head(delayed_refs, head);
7248         head->processing = 0;
7249
7250         spin_unlock(&head->lock);
7251         spin_unlock(&delayed_refs->lock);
7252
7253         BUG_ON(head->extent_op);
7254         if (head->must_insert_reserved)
7255                 ret = 1;
7256
7257         btrfs_cleanup_ref_head_accounting(trans->fs_info, delayed_refs, head);
7258         mutex_unlock(&head->mutex);
7259         btrfs_put_delayed_ref_head(head);
7260         return ret;
7261 out:
7262         spin_unlock(&head->lock);
7263
7264 out_delayed_unlock:
7265         spin_unlock(&delayed_refs->lock);
7266         return 0;
7267 }
7268
7269 void btrfs_free_tree_block(struct btrfs_trans_handle *trans,
7270                            struct btrfs_root *root,
7271                            struct extent_buffer *buf,
7272                            u64 parent, int last_ref)
7273 {
7274         struct btrfs_fs_info *fs_info = root->fs_info;
7275         int pin = 1;
7276         int ret;
7277
7278         if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
7279                 int old_ref_mod, new_ref_mod;
7280
7281                 btrfs_ref_tree_mod(root, buf->start, buf->len, parent,
7282                                    root->root_key.objectid,
7283                                    btrfs_header_level(buf), 0,
7284                                    BTRFS_DROP_DELAYED_REF);
7285                 ret = btrfs_add_delayed_tree_ref(trans, buf->start,
7286                                                  buf->len, parent,
7287                                                  root->root_key.objectid,
7288                                                  btrfs_header_level(buf),
7289                                                  BTRFS_DROP_DELAYED_REF, NULL,
7290                                                  &old_ref_mod, &new_ref_mod);
7291                 BUG_ON(ret); /* -ENOMEM */
7292                 pin = old_ref_mod >= 0 && new_ref_mod < 0;
7293         }
7294
7295         if (last_ref && btrfs_header_generation(buf) == trans->transid) {
7296                 struct btrfs_block_group_cache *cache;
7297
7298                 if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
7299                         ret = check_ref_cleanup(trans, buf->start);
7300                         if (!ret)
7301                                 goto out;
7302                 }
7303
7304                 pin = 0;
7305                 cache = btrfs_lookup_block_group(fs_info, buf->start);
7306
7307                 if (btrfs_header_flag(buf, BTRFS_HEADER_FLAG_WRITTEN)) {
7308                         pin_down_extent(fs_info, cache, buf->start,
7309                                         buf->len, 1);
7310                         btrfs_put_block_group(cache);
7311                         goto out;
7312                 }
7313
7314                 WARN_ON(test_bit(EXTENT_BUFFER_DIRTY, &buf->bflags));
7315
7316                 btrfs_add_free_space(cache, buf->start, buf->len);
7317                 btrfs_free_reserved_bytes(cache, buf->len, 0);
7318                 btrfs_put_block_group(cache);
7319                 trace_btrfs_reserved_extent_free(fs_info, buf->start, buf->len);
7320         }
7321 out:
7322         if (pin)
7323                 add_pinned_bytes(fs_info, buf->len, true,
7324                                  root->root_key.objectid);
7325
7326         if (last_ref) {
7327                 /*
7328                  * Deleting the buffer, clear the corrupt flag since it doesn't
7329                  * matter anymore.
7330                  */
7331                 clear_bit(EXTENT_BUFFER_CORRUPT, &buf->bflags);
7332         }
7333 }
7334
7335 /* Can return -ENOMEM */
7336 int btrfs_free_extent(struct btrfs_trans_handle *trans,
7337                       struct btrfs_root *root,
7338                       u64 bytenr, u64 num_bytes, u64 parent, u64 root_objectid,
7339                       u64 owner, u64 offset)
7340 {
7341         struct btrfs_fs_info *fs_info = root->fs_info;
7342         int old_ref_mod, new_ref_mod;
7343         int ret;
7344
7345         if (btrfs_is_testing(fs_info))
7346                 return 0;
7347
7348         if (root_objectid != BTRFS_TREE_LOG_OBJECTID)
7349                 btrfs_ref_tree_mod(root, bytenr, num_bytes, parent,
7350                                    root_objectid, owner, offset,
7351                                    BTRFS_DROP_DELAYED_REF);
7352
7353         /*
7354          * tree log blocks never actually go into the extent allocation
7355          * tree, just update pinning info and exit early.
7356          */
7357         if (root_objectid == BTRFS_TREE_LOG_OBJECTID) {
7358                 WARN_ON(owner >= BTRFS_FIRST_FREE_OBJECTID);
7359                 /* unlocks the pinned mutex */
7360                 btrfs_pin_extent(fs_info, bytenr, num_bytes, 1);
7361                 old_ref_mod = new_ref_mod = 0;
7362                 ret = 0;
7363         } else if (owner < BTRFS_FIRST_FREE_OBJECTID) {
7364                 ret = btrfs_add_delayed_tree_ref(trans, bytenr,
7365                                                  num_bytes, parent,
7366                                                  root_objectid, (int)owner,
7367                                                  BTRFS_DROP_DELAYED_REF, NULL,
7368                                                  &old_ref_mod, &new_ref_mod);
7369         } else {
7370                 ret = btrfs_add_delayed_data_ref(trans, bytenr,
7371                                                  num_bytes, parent,
7372                                                  root_objectid, owner, offset,
7373                                                  0, BTRFS_DROP_DELAYED_REF,
7374                                                  &old_ref_mod, &new_ref_mod);
7375         }
7376
7377         if (ret == 0 && old_ref_mod >= 0 && new_ref_mod < 0) {
7378                 bool metadata = owner < BTRFS_FIRST_FREE_OBJECTID;
7379
7380                 add_pinned_bytes(fs_info, num_bytes, metadata, root_objectid);
7381         }
7382
7383         return ret;
7384 }
7385
7386 /*
7387  * when we wait for progress in the block group caching, its because
7388  * our allocation attempt failed at least once.  So, we must sleep
7389  * and let some progress happen before we try again.
7390  *
7391  * This function will sleep at least once waiting for new free space to
7392  * show up, and then it will check the block group free space numbers
7393  * for our min num_bytes.  Another option is to have it go ahead
7394  * and look in the rbtree for a free extent of a given size, but this
7395  * is a good start.
7396  *
7397  * Callers of this must check if cache->cached == BTRFS_CACHE_ERROR before using
7398  * any of the information in this block group.
7399  */
7400 static noinline void
7401 wait_block_group_cache_progress(struct btrfs_block_group_cache *cache,
7402                                 u64 num_bytes)
7403 {
7404         struct btrfs_caching_control *caching_ctl;
7405
7406         caching_ctl = get_caching_control(cache);
7407         if (!caching_ctl)
7408                 return;
7409
7410         wait_event(caching_ctl->wait, block_group_cache_done(cache) ||
7411                    (cache->free_space_ctl->free_space >= num_bytes));
7412
7413         put_caching_control(caching_ctl);
7414 }
7415
7416 static noinline int
7417 wait_block_group_cache_done(struct btrfs_block_group_cache *cache)
7418 {
7419         struct btrfs_caching_control *caching_ctl;
7420         int ret = 0;
7421
7422         caching_ctl = get_caching_control(cache);
7423         if (!caching_ctl)
7424                 return (cache->cached == BTRFS_CACHE_ERROR) ? -EIO : 0;
7425
7426         wait_event(caching_ctl->wait, block_group_cache_done(cache));
7427         if (cache->cached == BTRFS_CACHE_ERROR)
7428                 ret = -EIO;
7429         put_caching_control(caching_ctl);
7430         return ret;
7431 }
7432
7433 enum btrfs_loop_type {
7434         LOOP_CACHING_NOWAIT = 0,
7435         LOOP_CACHING_WAIT = 1,
7436         LOOP_ALLOC_CHUNK = 2,
7437         LOOP_NO_EMPTY_SIZE = 3,
7438 };
7439
7440 static inline void
7441 btrfs_lock_block_group(struct btrfs_block_group_cache *cache,
7442                        int delalloc)
7443 {
7444         if (delalloc)
7445                 down_read(&cache->data_rwsem);
7446 }
7447
7448 static inline void
7449 btrfs_grab_block_group(struct btrfs_block_group_cache *cache,
7450                        int delalloc)
7451 {
7452         btrfs_get_block_group(cache);
7453         if (delalloc)
7454                 down_read(&cache->data_rwsem);
7455 }
7456
7457 static struct btrfs_block_group_cache *
7458 btrfs_lock_cluster(struct btrfs_block_group_cache *block_group,
7459                    struct btrfs_free_cluster *cluster,
7460                    int delalloc)
7461 {
7462         struct btrfs_block_group_cache *used_bg = NULL;
7463
7464         spin_lock(&cluster->refill_lock);
7465         while (1) {
7466                 used_bg = cluster->block_group;
7467                 if (!used_bg)
7468                         return NULL;
7469
7470                 if (used_bg == block_group)
7471                         return used_bg;
7472
7473                 btrfs_get_block_group(used_bg);
7474
7475                 if (!delalloc)
7476                         return used_bg;
7477
7478                 if (down_read_trylock(&used_bg->data_rwsem))
7479                         return used_bg;
7480
7481                 spin_unlock(&cluster->refill_lock);
7482
7483                 /* We should only have one-level nested. */
7484                 down_read_nested(&used_bg->data_rwsem, SINGLE_DEPTH_NESTING);
7485
7486                 spin_lock(&cluster->refill_lock);
7487                 if (used_bg == cluster->block_group)
7488                         return used_bg;
7489
7490                 up_read(&used_bg->data_rwsem);
7491                 btrfs_put_block_group(used_bg);
7492         }
7493 }
7494
7495 static inline void
7496 btrfs_release_block_group(struct btrfs_block_group_cache *cache,
7497                          int delalloc)
7498 {
7499         if (delalloc)
7500                 up_read(&cache->data_rwsem);
7501         btrfs_put_block_group(cache);
7502 }
7503
7504 /*
7505  * Structure used internally for find_free_extent() function.  Wraps needed
7506  * parameters.
7507  */
7508 struct find_free_extent_ctl {
7509         /* Basic allocation info */
7510         u64 ram_bytes;
7511         u64 num_bytes;
7512         u64 empty_size;
7513         u64 flags;
7514         int delalloc;
7515
7516         /* Where to start the search inside the bg */
7517         u64 search_start;
7518
7519         /* For clustered allocation */
7520         u64 empty_cluster;
7521
7522         bool have_caching_bg;
7523         bool orig_have_caching_bg;
7524
7525         /* RAID index, converted from flags */
7526         int index;
7527
7528         /*
7529          * Current loop number, check find_free_extent_update_loop() for details
7530          */
7531         int loop;
7532
7533         /*
7534          * Whether we're refilling a cluster, if true we need to re-search
7535          * current block group but don't try to refill the cluster again.
7536          */
7537         bool retry_clustered;
7538
7539         /*
7540          * Whether we're updating free space cache, if true we need to re-search
7541          * current block group but don't try updating free space cache again.
7542          */
7543         bool retry_unclustered;
7544
7545         /* If current block group is cached */
7546         int cached;
7547
7548         /* Max contiguous hole found */
7549         u64 max_extent_size;
7550
7551         /* Total free space from free space cache, not always contiguous */
7552         u64 total_free_space;
7553
7554         /* Found result */
7555         u64 found_offset;
7556 };
7557
7558
7559 /*
7560  * Helper function for find_free_extent().
7561  *
7562  * Return -ENOENT to inform caller that we need fallback to unclustered mode.
7563  * Return -EAGAIN to inform caller that we need to re-search this block group
7564  * Return >0 to inform caller that we find nothing
7565  * Return 0 means we have found a location and set ffe_ctl->found_offset.
7566  */
7567 static int find_free_extent_clustered(struct btrfs_block_group_cache *bg,
7568                 struct btrfs_free_cluster *last_ptr,
7569                 struct find_free_extent_ctl *ffe_ctl,
7570                 struct btrfs_block_group_cache **cluster_bg_ret)
7571 {
7572         struct btrfs_fs_info *fs_info = bg->fs_info;
7573         struct btrfs_block_group_cache *cluster_bg;
7574         u64 aligned_cluster;
7575         u64 offset;
7576         int ret;
7577
7578         cluster_bg = btrfs_lock_cluster(bg, last_ptr, ffe_ctl->delalloc);
7579         if (!cluster_bg)
7580                 goto refill_cluster;
7581         if (cluster_bg != bg && (cluster_bg->ro ||
7582             !block_group_bits(cluster_bg, ffe_ctl->flags)))
7583                 goto release_cluster;
7584
7585         offset = btrfs_alloc_from_cluster(cluster_bg, last_ptr,
7586                         ffe_ctl->num_bytes, cluster_bg->key.objectid,
7587                         &ffe_ctl->max_extent_size);
7588         if (offset) {
7589                 /* We have a block, we're done */
7590                 spin_unlock(&last_ptr->refill_lock);
7591                 trace_btrfs_reserve_extent_cluster(cluster_bg,
7592                                 ffe_ctl->search_start, ffe_ctl->num_bytes);
7593                 *cluster_bg_ret = cluster_bg;
7594                 ffe_ctl->found_offset = offset;
7595                 return 0;
7596         }
7597         WARN_ON(last_ptr->block_group != cluster_bg);
7598
7599 release_cluster:
7600         /*
7601          * If we are on LOOP_NO_EMPTY_SIZE, we can't set up a new clusters, so
7602          * lets just skip it and let the allocator find whatever block it can
7603          * find. If we reach this point, we will have tried the cluster
7604          * allocator plenty of times and not have found anything, so we are
7605          * likely way too fragmented for the clustering stuff to find anything.
7606          *
7607          * However, if the cluster is taken from the current block group,
7608          * release the cluster first, so that we stand a better chance of
7609          * succeeding in the unclustered allocation.
7610          */
7611         if (ffe_ctl->loop >= LOOP_NO_EMPTY_SIZE && cluster_bg != bg) {
7612                 spin_unlock(&last_ptr->refill_lock);
7613                 btrfs_release_block_group(cluster_bg, ffe_ctl->delalloc);
7614                 return -ENOENT;
7615         }
7616
7617         /* This cluster didn't work out, free it and start over */
7618         btrfs_return_cluster_to_free_space(NULL, last_ptr);
7619
7620         if (cluster_bg != bg)
7621                 btrfs_release_block_group(cluster_bg, ffe_ctl->delalloc);
7622
7623 refill_cluster:
7624         if (ffe_ctl->loop >= LOOP_NO_EMPTY_SIZE) {
7625                 spin_unlock(&last_ptr->refill_lock);
7626                 return -ENOENT;
7627         }
7628
7629         aligned_cluster = max_t(u64,
7630                         ffe_ctl->empty_cluster + ffe_ctl->empty_size,
7631                         bg->full_stripe_len);
7632         ret = btrfs_find_space_cluster(fs_info, bg, last_ptr,
7633                         ffe_ctl->search_start, ffe_ctl->num_bytes,
7634                         aligned_cluster);
7635         if (ret == 0) {
7636                 /* Now pull our allocation out of this cluster */
7637                 offset = btrfs_alloc_from_cluster(bg, last_ptr,
7638                                 ffe_ctl->num_bytes, ffe_ctl->search_start,
7639                                 &ffe_ctl->max_extent_size);
7640                 if (offset) {
7641                         /* We found one, proceed */
7642                         spin_unlock(&last_ptr->refill_lock);
7643                         trace_btrfs_reserve_extent_cluster(bg,
7644                                         ffe_ctl->search_start,
7645                                         ffe_ctl->num_bytes);
7646                         ffe_ctl->found_offset = offset;
7647                         return 0;
7648                 }
7649         } else if (!ffe_ctl->cached && ffe_ctl->loop > LOOP_CACHING_NOWAIT &&
7650                    !ffe_ctl->retry_clustered) {
7651                 spin_unlock(&last_ptr->refill_lock);
7652
7653                 ffe_ctl->retry_clustered = true;
7654                 wait_block_group_cache_progress(bg, ffe_ctl->num_bytes +
7655                                 ffe_ctl->empty_cluster + ffe_ctl->empty_size);
7656                 return -EAGAIN;
7657         }
7658         /*
7659          * At this point we either didn't find a cluster or we weren't able to
7660          * allocate a block from our cluster.  Free the cluster we've been
7661          * trying to use, and go to the next block group.
7662          */
7663         btrfs_return_cluster_to_free_space(NULL, last_ptr);
7664         spin_unlock(&last_ptr->refill_lock);
7665         return 1;
7666 }
7667
7668 /*
7669  * Return >0 to inform caller that we find nothing
7670  * Return 0 when we found an free extent and set ffe_ctrl->found_offset
7671  * Return -EAGAIN to inform caller that we need to re-search this block group
7672  */
7673 static int find_free_extent_unclustered(struct btrfs_block_group_cache *bg,
7674                 struct btrfs_free_cluster *last_ptr,
7675                 struct find_free_extent_ctl *ffe_ctl)
7676 {
7677         u64 offset;
7678
7679         /*
7680          * We are doing an unclustered allocation, set the fragmented flag so
7681          * we don't bother trying to setup a cluster again until we get more
7682          * space.
7683          */
7684         if (unlikely(last_ptr)) {
7685                 spin_lock(&last_ptr->lock);
7686                 last_ptr->fragmented = 1;
7687                 spin_unlock(&last_ptr->lock);
7688         }
7689         if (ffe_ctl->cached) {
7690                 struct btrfs_free_space_ctl *free_space_ctl;
7691
7692                 free_space_ctl = bg->free_space_ctl;
7693                 spin_lock(&free_space_ctl->tree_lock);
7694                 if (free_space_ctl->free_space <
7695                     ffe_ctl->num_bytes + ffe_ctl->empty_cluster +
7696                     ffe_ctl->empty_size) {
7697                         ffe_ctl->total_free_space = max_t(u64,
7698                                         ffe_ctl->total_free_space,
7699                                         free_space_ctl->free_space);
7700                         spin_unlock(&free_space_ctl->tree_lock);
7701                         return 1;
7702                 }
7703                 spin_unlock(&free_space_ctl->tree_lock);
7704         }
7705
7706         offset = btrfs_find_space_for_alloc(bg, ffe_ctl->search_start,
7707                         ffe_ctl->num_bytes, ffe_ctl->empty_size,
7708                         &ffe_ctl->max_extent_size);
7709
7710         /*
7711          * If we didn't find a chunk, and we haven't failed on this block group
7712          * before, and this block group is in the middle of caching and we are
7713          * ok with waiting, then go ahead and wait for progress to be made, and
7714          * set @retry_unclustered to true.
7715          *
7716          * If @retry_unclustered is true then we've already waited on this
7717          * block group once and should move on to the next block group.
7718          */
7719         if (!offset && !ffe_ctl->retry_unclustered && !ffe_ctl->cached &&
7720             ffe_ctl->loop > LOOP_CACHING_NOWAIT) {
7721                 wait_block_group_cache_progress(bg, ffe_ctl->num_bytes +
7722                                                 ffe_ctl->empty_size);
7723                 ffe_ctl->retry_unclustered = true;
7724                 return -EAGAIN;
7725         } else if (!offset) {
7726                 return 1;
7727         }
7728         ffe_ctl->found_offset = offset;
7729         return 0;
7730 }
7731
7732 /*
7733  * Return >0 means caller needs to re-search for free extent
7734  * Return 0 means we have the needed free extent.
7735  * Return <0 means we failed to locate any free extent.
7736  */
7737 static int find_free_extent_update_loop(struct btrfs_fs_info *fs_info,
7738                                         struct btrfs_free_cluster *last_ptr,
7739                                         struct btrfs_key *ins,
7740                                         struct find_free_extent_ctl *ffe_ctl,
7741                                         int full_search, bool use_cluster)
7742 {
7743         struct btrfs_root *root = fs_info->extent_root;
7744         int ret;
7745
7746         if ((ffe_ctl->loop == LOOP_CACHING_NOWAIT) &&
7747             ffe_ctl->have_caching_bg && !ffe_ctl->orig_have_caching_bg)
7748                 ffe_ctl->orig_have_caching_bg = true;
7749
7750         if (!ins->objectid && ffe_ctl->loop >= LOOP_CACHING_WAIT &&
7751             ffe_ctl->have_caching_bg)
7752                 return 1;
7753
7754         if (!ins->objectid && ++(ffe_ctl->index) < BTRFS_NR_RAID_TYPES)
7755                 return 1;
7756
7757         if (ins->objectid) {
7758                 if (!use_cluster && last_ptr) {
7759                         spin_lock(&last_ptr->lock);
7760                         last_ptr->window_start = ins->objectid;
7761                         spin_unlock(&last_ptr->lock);
7762                 }
7763                 return 0;
7764         }
7765
7766         /*
7767          * LOOP_CACHING_NOWAIT, search partially cached block groups, kicking
7768          *                      caching kthreads as we move along
7769          * LOOP_CACHING_WAIT, search everything, and wait if our bg is caching
7770          * LOOP_ALLOC_CHUNK, force a chunk allocation and try again
7771          * LOOP_NO_EMPTY_SIZE, set empty_size and empty_cluster to 0 and try
7772          *                     again
7773          */
7774         if (ffe_ctl->loop < LOOP_NO_EMPTY_SIZE) {
7775                 ffe_ctl->index = 0;
7776                 if (ffe_ctl->loop == LOOP_CACHING_NOWAIT) {
7777                         /*
7778                          * We want to skip the LOOP_CACHING_WAIT step if we
7779                          * don't have any uncached bgs and we've already done a
7780                          * full search through.
7781                          */
7782                         if (ffe_ctl->orig_have_caching_bg || !full_search)
7783                                 ffe_ctl->loop = LOOP_CACHING_WAIT;
7784                         else
7785                                 ffe_ctl->loop = LOOP_ALLOC_CHUNK;
7786                 } else {
7787                         ffe_ctl->loop++;
7788                 }
7789
7790                 if (ffe_ctl->loop == LOOP_ALLOC_CHUNK) {
7791                         struct btrfs_trans_handle *trans;
7792                         int exist = 0;
7793
7794                         trans = current->journal_info;
7795                         if (trans)
7796                                 exist = 1;
7797                         else
7798                                 trans = btrfs_join_transaction(root);
7799
7800                         if (IS_ERR(trans)) {
7801                                 ret = PTR_ERR(trans);
7802                                 return ret;
7803                         }
7804
7805                         ret = do_chunk_alloc(trans, ffe_ctl->flags,
7806                                              CHUNK_ALLOC_FORCE);
7807
7808                         /*
7809                          * If we can't allocate a new chunk we've already looped
7810                          * through at least once, move on to the NO_EMPTY_SIZE
7811                          * case.
7812                          */
7813                         if (ret == -ENOSPC)
7814                                 ffe_ctl->loop = LOOP_NO_EMPTY_SIZE;
7815
7816                         /* Do not bail out on ENOSPC since we can do more. */
7817                         if (ret < 0 && ret != -ENOSPC)
7818                                 btrfs_abort_transaction(trans, ret);
7819                         else
7820                                 ret = 0;
7821                         if (!exist)
7822                                 btrfs_end_transaction(trans);
7823                         if (ret)
7824                                 return ret;
7825                 }
7826
7827                 if (ffe_ctl->loop == LOOP_NO_EMPTY_SIZE) {
7828                         /*
7829                          * Don't loop again if we already have no empty_size and
7830                          * no empty_cluster.
7831                          */
7832                         if (ffe_ctl->empty_size == 0 &&
7833                             ffe_ctl->empty_cluster == 0)
7834                                 return -ENOSPC;
7835                         ffe_ctl->empty_size = 0;
7836                         ffe_ctl->empty_cluster = 0;
7837                 }
7838                 return 1;
7839         }
7840         return -ENOSPC;
7841 }
7842
7843 /*
7844  * walks the btree of allocated extents and find a hole of a given size.
7845  * The key ins is changed to record the hole:
7846  * ins->objectid == start position
7847  * ins->flags = BTRFS_EXTENT_ITEM_KEY
7848  * ins->offset == the size of the hole.
7849  * Any available blocks before search_start are skipped.
7850  *
7851  * If there is no suitable free space, we will record the max size of
7852  * the free space extent currently.
7853  *
7854  * The overall logic and call chain:
7855  *
7856  * find_free_extent()
7857  * |- Iterate through all block groups
7858  * |  |- Get a valid block group
7859  * |  |- Try to do clustered allocation in that block group
7860  * |  |- Try to do unclustered allocation in that block group
7861  * |  |- Check if the result is valid
7862  * |  |  |- If valid, then exit
7863  * |  |- Jump to next block group
7864  * |
7865  * |- Push harder to find free extents
7866  *    |- If not found, re-iterate all block groups
7867  */
7868 static noinline int find_free_extent(struct btrfs_fs_info *fs_info,
7869                                 u64 ram_bytes, u64 num_bytes, u64 empty_size,
7870                                 u64 hint_byte, struct btrfs_key *ins,
7871                                 u64 flags, int delalloc)
7872 {
7873         int ret = 0;
7874         struct btrfs_free_cluster *last_ptr = NULL;
7875         struct btrfs_block_group_cache *block_group = NULL;
7876         struct find_free_extent_ctl ffe_ctl = {0};
7877         struct btrfs_space_info *space_info;
7878         bool use_cluster = true;
7879         bool full_search = false;
7880
7881         WARN_ON(num_bytes < fs_info->sectorsize);
7882
7883         ffe_ctl.ram_bytes = ram_bytes;
7884         ffe_ctl.num_bytes = num_bytes;
7885         ffe_ctl.empty_size = empty_size;
7886         ffe_ctl.flags = flags;
7887         ffe_ctl.search_start = 0;
7888         ffe_ctl.retry_clustered = false;
7889         ffe_ctl.retry_unclustered = false;
7890         ffe_ctl.delalloc = delalloc;
7891         ffe_ctl.index = btrfs_bg_flags_to_raid_index(flags);
7892         ffe_ctl.have_caching_bg = false;
7893         ffe_ctl.orig_have_caching_bg = false;
7894         ffe_ctl.found_offset = 0;
7895
7896         ins->type = BTRFS_EXTENT_ITEM_KEY;
7897         ins->objectid = 0;
7898         ins->offset = 0;
7899
7900         trace_find_free_extent(fs_info, num_bytes, empty_size, flags);
7901
7902         space_info = __find_space_info(fs_info, flags);
7903         if (!space_info) {
7904                 btrfs_err(fs_info, "No space info for %llu", flags);
7905                 return -ENOSPC;
7906         }
7907
7908         /*
7909          * If our free space is heavily fragmented we may not be able to make
7910          * big contiguous allocations, so instead of doing the expensive search
7911          * for free space, simply return ENOSPC with our max_extent_size so we
7912          * can go ahead and search for a more manageable chunk.
7913          *
7914          * If our max_extent_size is large enough for our allocation simply
7915          * disable clustering since we will likely not be able to find enough
7916          * space to create a cluster and induce latency trying.
7917          */
7918         if (unlikely(space_info->max_extent_size)) {
7919                 spin_lock(&space_info->lock);
7920                 if (space_info->max_extent_size &&
7921                     num_bytes > space_info->max_extent_size) {
7922                         ins->offset = space_info->max_extent_size;
7923                         spin_unlock(&space_info->lock);
7924                         return -ENOSPC;
7925                 } else if (space_info->max_extent_size) {
7926                         use_cluster = false;
7927                 }
7928                 spin_unlock(&space_info->lock);
7929         }
7930
7931         last_ptr = fetch_cluster_info(fs_info, space_info,
7932                                       &ffe_ctl.empty_cluster);
7933         if (last_ptr) {
7934                 spin_lock(&last_ptr->lock);
7935                 if (last_ptr->block_group)
7936                         hint_byte = last_ptr->window_start;
7937                 if (last_ptr->fragmented) {
7938                         /*
7939                          * We still set window_start so we can keep track of the
7940                          * last place we found an allocation to try and save
7941                          * some time.
7942                          */
7943                         hint_byte = last_ptr->window_start;
7944                         use_cluster = false;
7945                 }
7946                 spin_unlock(&last_ptr->lock);
7947         }
7948
7949         ffe_ctl.search_start = max(ffe_ctl.search_start,
7950                                    first_logical_byte(fs_info, 0));
7951         ffe_ctl.search_start = max(ffe_ctl.search_start, hint_byte);
7952         if (ffe_ctl.search_start == hint_byte) {
7953                 block_group = btrfs_lookup_block_group(fs_info,
7954                                                        ffe_ctl.search_start);
7955                 /*
7956                  * we don't want to use the block group if it doesn't match our
7957                  * allocation bits, or if its not cached.
7958                  *
7959                  * However if we are re-searching with an ideal block group
7960                  * picked out then we don't care that the block group is cached.
7961                  */
7962                 if (block_group && block_group_bits(block_group, flags) &&
7963                     block_group->cached != BTRFS_CACHE_NO) {
7964                         down_read(&space_info->groups_sem);
7965                         if (list_empty(&block_group->list) ||
7966                             block_group->ro) {
7967                                 /*
7968                                  * someone is removing this block group,
7969                                  * we can't jump into the have_block_group
7970                                  * target because our list pointers are not
7971                                  * valid
7972                                  */
7973                                 btrfs_put_block_group(block_group);
7974                                 up_read(&space_info->groups_sem);
7975                         } else {
7976                                 ffe_ctl.index = btrfs_bg_flags_to_raid_index(
7977                                                 block_group->flags);
7978                                 btrfs_lock_block_group(block_group, delalloc);
7979                                 goto have_block_group;
7980                         }
7981                 } else if (block_group) {
7982                         btrfs_put_block_group(block_group);
7983                 }
7984         }
7985 search:
7986         ffe_ctl.have_caching_bg = false;
7987         if (ffe_ctl.index == btrfs_bg_flags_to_raid_index(flags) ||
7988             ffe_ctl.index == 0)
7989                 full_search = true;
7990         down_read(&space_info->groups_sem);
7991         list_for_each_entry(block_group,
7992                             &space_info->block_groups[ffe_ctl.index], list) {
7993                 /* If the block group is read-only, we can skip it entirely. */
7994                 if (unlikely(block_group->ro))
7995                         continue;
7996
7997                 btrfs_grab_block_group(block_group, delalloc);
7998                 ffe_ctl.search_start = block_group->key.objectid;
7999
8000                 /*
8001                  * this can happen if we end up cycling through all the
8002                  * raid types, but we want to make sure we only allocate
8003                  * for the proper type.
8004                  */
8005                 if (!block_group_bits(block_group, flags)) {
8006                         u64 extra = BTRFS_BLOCK_GROUP_DUP |
8007                                 BTRFS_BLOCK_GROUP_RAID1 |
8008                                 BTRFS_BLOCK_GROUP_RAID5 |
8009                                 BTRFS_BLOCK_GROUP_RAID6 |
8010                                 BTRFS_BLOCK_GROUP_RAID10;
8011
8012                         /*
8013                          * if they asked for extra copies and this block group
8014                          * doesn't provide them, bail.  This does allow us to
8015                          * fill raid0 from raid1.
8016                          */
8017                         if ((flags & extra) && !(block_group->flags & extra))
8018                                 goto loop;
8019                 }
8020
8021 have_block_group:
8022                 ffe_ctl.cached = block_group_cache_done(block_group);
8023                 if (unlikely(!ffe_ctl.cached)) {
8024                         ffe_ctl.have_caching_bg = true;
8025                         ret = cache_block_group(block_group, 0);
8026                         BUG_ON(ret < 0);
8027                         ret = 0;
8028                 }
8029
8030                 if (unlikely(block_group->cached == BTRFS_CACHE_ERROR))
8031                         goto loop;
8032
8033                 /*
8034                  * Ok we want to try and use the cluster allocator, so
8035                  * lets look there
8036                  */
8037                 if (last_ptr && use_cluster) {
8038                         struct btrfs_block_group_cache *cluster_bg = NULL;
8039
8040                         ret = find_free_extent_clustered(block_group, last_ptr,
8041                                                          &ffe_ctl, &cluster_bg);
8042
8043                         if (ret == 0) {
8044                                 if (cluster_bg && cluster_bg != block_group) {
8045                                         btrfs_release_block_group(block_group,
8046                                                                   delalloc);
8047                                         block_group = cluster_bg;
8048                                 }
8049                                 goto checks;
8050                         } else if (ret == -EAGAIN) {
8051                                 goto have_block_group;
8052                         } else if (ret > 0) {
8053                                 goto loop;
8054                         }
8055                         /* ret == -ENOENT case falls through */
8056                 }
8057
8058                 ret = find_free_extent_unclustered(block_group, last_ptr,
8059                                                    &ffe_ctl);
8060                 if (ret == -EAGAIN)
8061                         goto have_block_group;
8062                 else if (ret > 0)
8063                         goto loop;
8064                 /* ret == 0 case falls through */
8065 checks:
8066                 ffe_ctl.search_start = round_up(ffe_ctl.found_offset,
8067                                              fs_info->stripesize);
8068
8069                 /* move on to the next group */
8070                 if (ffe_ctl.search_start + num_bytes >
8071                     block_group->key.objectid + block_group->key.offset) {
8072                         btrfs_add_free_space(block_group, ffe_ctl.found_offset,
8073                                              num_bytes);
8074                         goto loop;
8075                 }
8076
8077                 if (ffe_ctl.found_offset < ffe_ctl.search_start)
8078                         btrfs_add_free_space(block_group, ffe_ctl.found_offset,
8079                                 ffe_ctl.search_start - ffe_ctl.found_offset);
8080
8081                 ret = btrfs_add_reserved_bytes(block_group, ram_bytes,
8082                                 num_bytes, delalloc);
8083                 if (ret == -EAGAIN) {
8084                         btrfs_add_free_space(block_group, ffe_ctl.found_offset,
8085                                              num_bytes);
8086                         goto loop;
8087                 }
8088                 btrfs_inc_block_group_reservations(block_group);
8089
8090                 /* we are all good, lets return */
8091                 ins->objectid = ffe_ctl.search_start;
8092                 ins->offset = num_bytes;
8093
8094                 trace_btrfs_reserve_extent(block_group, ffe_ctl.search_start,
8095                                            num_bytes);
8096                 btrfs_release_block_group(block_group, delalloc);
8097                 break;
8098 loop:
8099                 ffe_ctl.retry_clustered = false;
8100                 ffe_ctl.retry_unclustered = false;
8101                 BUG_ON(btrfs_bg_flags_to_raid_index(block_group->flags) !=
8102                        ffe_ctl.index);
8103                 btrfs_release_block_group(block_group, delalloc);
8104                 cond_resched();
8105         }
8106         up_read(&space_info->groups_sem);
8107
8108         ret = find_free_extent_update_loop(fs_info, last_ptr, ins, &ffe_ctl,
8109                                            full_search, use_cluster);
8110         if (ret > 0)
8111                 goto search;
8112
8113         if (ret == -ENOSPC) {
8114                 /*
8115                  * Use ffe_ctl->total_free_space as fallback if we can't find
8116                  * any contiguous hole.
8117                  */
8118                 if (!ffe_ctl.max_extent_size)
8119                         ffe_ctl.max_extent_size = ffe_ctl.total_free_space;
8120                 spin_lock(&space_info->lock);
8121                 space_info->max_extent_size = ffe_ctl.max_extent_size;
8122                 spin_unlock(&space_info->lock);
8123                 ins->offset = ffe_ctl.max_extent_size;
8124         }
8125         return ret;
8126 }
8127
8128 #define DUMP_BLOCK_RSV(fs_info, rsv_name)                               \
8129 do {                                                                    \
8130         struct btrfs_block_rsv *__rsv = &(fs_info)->rsv_name;           \
8131         spin_lock(&__rsv->lock);                                        \
8132         btrfs_info(fs_info, #rsv_name ": size %llu reserved %llu",      \
8133                    __rsv->size, __rsv->reserved);                       \
8134         spin_unlock(&__rsv->lock);                                      \
8135 } while (0)
8136
8137 static void dump_space_info(struct btrfs_fs_info *fs_info,
8138                             struct btrfs_space_info *info, u64 bytes,
8139                             int dump_block_groups)
8140 {
8141         struct btrfs_block_group_cache *cache;
8142         int index = 0;
8143
8144         spin_lock(&info->lock);
8145         btrfs_info(fs_info, "space_info %llu has %llu free, is %sfull",
8146                    info->flags,
8147                    info->total_bytes - btrfs_space_info_used(info, true),
8148                    info->full ? "" : "not ");
8149         btrfs_info(fs_info,
8150                 "space_info total=%llu, used=%llu, pinned=%llu, reserved=%llu, may_use=%llu, readonly=%llu",
8151                 info->total_bytes, info->bytes_used, info->bytes_pinned,
8152                 info->bytes_reserved, info->bytes_may_use,
8153                 info->bytes_readonly);
8154         spin_unlock(&info->lock);
8155
8156         DUMP_BLOCK_RSV(fs_info, global_block_rsv);
8157         DUMP_BLOCK_RSV(fs_info, trans_block_rsv);
8158         DUMP_BLOCK_RSV(fs_info, chunk_block_rsv);
8159         DUMP_BLOCK_RSV(fs_info, delayed_block_rsv);
8160         DUMP_BLOCK_RSV(fs_info, delayed_refs_rsv);
8161
8162         if (!dump_block_groups)
8163                 return;
8164
8165         down_read(&info->groups_sem);
8166 again:
8167         list_for_each_entry(cache, &info->block_groups[index], list) {
8168                 spin_lock(&cache->lock);
8169                 btrfs_info(fs_info,
8170                         "block group %llu has %llu bytes, %llu used %llu pinned %llu reserved %s",
8171                         cache->key.objectid, cache->key.offset,
8172                         btrfs_block_group_used(&cache->item), cache->pinned,
8173                         cache->reserved, cache->ro ? "[readonly]" : "");
8174                 btrfs_dump_free_space(cache, bytes);
8175                 spin_unlock(&cache->lock);
8176         }
8177         if (++index < BTRFS_NR_RAID_TYPES)
8178                 goto again;
8179         up_read(&info->groups_sem);
8180 }
8181
8182 /*
8183  * btrfs_reserve_extent - entry point to the extent allocator. Tries to find a
8184  *                        hole that is at least as big as @num_bytes.
8185  *
8186  * @root           -    The root that will contain this extent
8187  *
8188  * @ram_bytes      -    The amount of space in ram that @num_bytes take. This
8189  *                      is used for accounting purposes. This value differs
8190  *                      from @num_bytes only in the case of compressed extents.
8191  *
8192  * @num_bytes      -    Number of bytes to allocate on-disk.
8193  *
8194  * @min_alloc_size -    Indicates the minimum amount of space that the
8195  *                      allocator should try to satisfy. In some cases
8196  *                      @num_bytes may be larger than what is required and if
8197  *                      the filesystem is fragmented then allocation fails.
8198  *                      However, the presence of @min_alloc_size gives a
8199  *                      chance to try and satisfy the smaller allocation.
8200  *
8201  * @empty_size     -    A hint that you plan on doing more COW. This is the
8202  *                      size in bytes the allocator should try to find free
8203  *                      next to the block it returns.  This is just a hint and
8204  *                      may be ignored by the allocator.
8205  *
8206  * @hint_byte      -    Hint to the allocator to start searching above the byte
8207  *                      address passed. It might be ignored.
8208  *
8209  * @ins            -    This key is modified to record the found hole. It will
8210  *                      have the following values:
8211  *                      ins->objectid == start position
8212  *                      ins->flags = BTRFS_EXTENT_ITEM_KEY
8213  *                      ins->offset == the size of the hole.
8214  *
8215  * @is_data        -    Boolean flag indicating whether an extent is
8216  *                      allocated for data (true) or metadata (false)
8217  *
8218  * @delalloc       -    Boolean flag indicating whether this allocation is for
8219  *                      delalloc or not. If 'true' data_rwsem of block groups
8220  *                      is going to be acquired.
8221  *
8222  *
8223  * Returns 0 when an allocation succeeded or < 0 when an error occurred. In
8224  * case -ENOSPC is returned then @ins->offset will contain the size of the
8225  * largest available hole the allocator managed to find.
8226  */
8227 int btrfs_reserve_extent(struct btrfs_root *root, u64 ram_bytes,
8228                          u64 num_bytes, u64 min_alloc_size,
8229                          u64 empty_size, u64 hint_byte,
8230                          struct btrfs_key *ins, int is_data, int delalloc)
8231 {
8232         struct btrfs_fs_info *fs_info = root->fs_info;
8233         bool final_tried = num_bytes == min_alloc_size;
8234         u64 flags;
8235         int ret;
8236
8237         flags = get_alloc_profile_by_root(root, is_data);
8238 again:
8239         WARN_ON(num_bytes < fs_info->sectorsize);
8240         ret = find_free_extent(fs_info, ram_bytes, num_bytes, empty_size,
8241                                hint_byte, ins, flags, delalloc);
8242         if (!ret && !is_data) {
8243                 btrfs_dec_block_group_reservations(fs_info, ins->objectid);
8244         } else if (ret == -ENOSPC) {
8245                 if (!final_tried && ins->offset) {
8246                         num_bytes = min(num_bytes >> 1, ins->offset);
8247                         num_bytes = round_down(num_bytes,
8248                                                fs_info->sectorsize);
8249                         num_bytes = max(num_bytes, min_alloc_size);
8250                         ram_bytes = num_bytes;
8251                         if (num_bytes == min_alloc_size)
8252                                 final_tried = true;
8253                         goto again;
8254                 } else if (btrfs_test_opt(fs_info, ENOSPC_DEBUG)) {
8255                         struct btrfs_space_info *sinfo;
8256
8257                         sinfo = __find_space_info(fs_info, flags);
8258                         btrfs_err(fs_info,
8259                                   "allocation failed flags %llu, wanted %llu",
8260                                   flags, num_bytes);
8261                         if (sinfo)
8262                                 dump_space_info(fs_info, sinfo, num_bytes, 1);
8263                 }
8264         }
8265
8266         return ret;
8267 }
8268
8269 static int __btrfs_free_reserved_extent(struct btrfs_fs_info *fs_info,
8270                                         u64 start, u64 len,
8271                                         int pin, int delalloc)
8272 {
8273         struct btrfs_block_group_cache *cache;
8274         int ret = 0;
8275
8276         cache = btrfs_lookup_block_group(fs_info, start);
8277         if (!cache) {
8278                 btrfs_err(fs_info, "Unable to find block group for %llu",
8279                           start);
8280                 return -ENOSPC;
8281         }
8282
8283         if (pin)
8284                 pin_down_extent(fs_info, cache, start, len, 1);
8285         else {
8286                 if (btrfs_test_opt(fs_info, DISCARD))
8287                         ret = btrfs_discard_extent(fs_info, start, len, NULL);
8288                 btrfs_add_free_space(cache, start, len);
8289                 btrfs_free_reserved_bytes(cache, len, delalloc);
8290                 trace_btrfs_reserved_extent_free(fs_info, start, len);
8291         }
8292
8293         btrfs_put_block_group(cache);
8294         return ret;
8295 }
8296
8297 int btrfs_free_reserved_extent(struct btrfs_fs_info *fs_info,
8298                                u64 start, u64 len, int delalloc)
8299 {
8300         return __btrfs_free_reserved_extent(fs_info, start, len, 0, delalloc);
8301 }
8302
8303 int btrfs_free_and_pin_reserved_extent(struct btrfs_fs_info *fs_info,
8304                                        u64 start, u64 len)
8305 {
8306         return __btrfs_free_reserved_extent(fs_info, start, len, 1, 0);
8307 }
8308
8309 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
8310                                       u64 parent, u64 root_objectid,
8311                                       u64 flags, u64 owner, u64 offset,
8312                                       struct btrfs_key *ins, int ref_mod)
8313 {
8314         struct btrfs_fs_info *fs_info = trans->fs_info;
8315         int ret;
8316         struct btrfs_extent_item *extent_item;
8317         struct btrfs_extent_inline_ref *iref;
8318         struct btrfs_path *path;
8319         struct extent_buffer *leaf;
8320         int type;
8321         u32 size;
8322
8323         if (parent > 0)
8324                 type = BTRFS_SHARED_DATA_REF_KEY;
8325         else
8326                 type = BTRFS_EXTENT_DATA_REF_KEY;
8327
8328         size = sizeof(*extent_item) + btrfs_extent_inline_ref_size(type);
8329
8330         path = btrfs_alloc_path();
8331         if (!path)
8332                 return -ENOMEM;
8333
8334         path->leave_spinning = 1;
8335         ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path,
8336                                       ins, size);
8337         if (ret) {
8338                 btrfs_free_path(path);
8339                 return ret;
8340         }
8341
8342         leaf = path->nodes[0];
8343         extent_item = btrfs_item_ptr(leaf, path->slots[0],
8344                                      struct btrfs_extent_item);
8345         btrfs_set_extent_refs(leaf, extent_item, ref_mod);
8346         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
8347         btrfs_set_extent_flags(leaf, extent_item,
8348                                flags | BTRFS_EXTENT_FLAG_DATA);
8349
8350         iref = (struct btrfs_extent_inline_ref *)(extent_item + 1);
8351         btrfs_set_extent_inline_ref_type(leaf, iref, type);
8352         if (parent > 0) {
8353                 struct btrfs_shared_data_ref *ref;
8354                 ref = (struct btrfs_shared_data_ref *)(iref + 1);
8355                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
8356                 btrfs_set_shared_data_ref_count(leaf, ref, ref_mod);
8357         } else {
8358                 struct btrfs_extent_data_ref *ref;
8359                 ref = (struct btrfs_extent_data_ref *)(&iref->offset);
8360                 btrfs_set_extent_data_ref_root(leaf, ref, root_objectid);
8361                 btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
8362                 btrfs_set_extent_data_ref_offset(leaf, ref, offset);
8363                 btrfs_set_extent_data_ref_count(leaf, ref, ref_mod);
8364         }
8365
8366         btrfs_mark_buffer_dirty(path->nodes[0]);
8367         btrfs_free_path(path);
8368
8369         ret = remove_from_free_space_tree(trans, ins->objectid, ins->offset);
8370         if (ret)
8371                 return ret;
8372
8373         ret = update_block_group(trans, fs_info, ins->objectid, ins->offset, 1);
8374         if (ret) { /* -ENOENT, logic error */
8375                 btrfs_err(fs_info, "update block group failed for %llu %llu",
8376                         ins->objectid, ins->offset);
8377                 BUG();
8378         }
8379         trace_btrfs_reserved_extent_alloc(fs_info, ins->objectid, ins->offset);
8380         return ret;
8381 }
8382
8383 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
8384                                      struct btrfs_delayed_ref_node *node,
8385                                      struct btrfs_delayed_extent_op *extent_op)
8386 {
8387         struct btrfs_fs_info *fs_info = trans->fs_info;
8388         int ret;
8389         struct btrfs_extent_item *extent_item;
8390         struct btrfs_key extent_key;
8391         struct btrfs_tree_block_info *block_info;
8392         struct btrfs_extent_inline_ref *iref;
8393         struct btrfs_path *path;
8394         struct extent_buffer *leaf;
8395         struct btrfs_delayed_tree_ref *ref;
8396         u32 size = sizeof(*extent_item) + sizeof(*iref);
8397         u64 num_bytes;
8398         u64 flags = extent_op->flags_to_set;
8399         bool skinny_metadata = btrfs_fs_incompat(fs_info, SKINNY_METADATA);
8400
8401         ref = btrfs_delayed_node_to_tree_ref(node);
8402
8403         extent_key.objectid = node->bytenr;
8404         if (skinny_metadata) {
8405                 extent_key.offset = ref->level;
8406                 extent_key.type = BTRFS_METADATA_ITEM_KEY;
8407                 num_bytes = fs_info->nodesize;
8408         } else {
8409                 extent_key.offset = node->num_bytes;
8410                 extent_key.type = BTRFS_EXTENT_ITEM_KEY;
8411                 size += sizeof(*block_info);
8412                 num_bytes = node->num_bytes;
8413         }
8414
8415         path = btrfs_alloc_path();
8416         if (!path)
8417                 return -ENOMEM;
8418
8419         path->leave_spinning = 1;
8420         ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path,
8421                                       &extent_key, size);
8422         if (ret) {
8423                 btrfs_free_path(path);
8424                 return ret;
8425         }
8426
8427         leaf = path->nodes[0];
8428         extent_item = btrfs_item_ptr(leaf, path->slots[0],
8429                                      struct btrfs_extent_item);
8430         btrfs_set_extent_refs(leaf, extent_item, 1);
8431         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
8432         btrfs_set_extent_flags(leaf, extent_item,
8433                                flags | BTRFS_EXTENT_FLAG_TREE_BLOCK);
8434
8435         if (skinny_metadata) {
8436                 iref = (struct btrfs_extent_inline_ref *)(extent_item + 1);
8437         } else {
8438                 block_info = (struct btrfs_tree_block_info *)(extent_item + 1);
8439                 btrfs_set_tree_block_key(leaf, block_info, &extent_op->key);
8440                 btrfs_set_tree_block_level(leaf, block_info, ref->level);
8441                 iref = (struct btrfs_extent_inline_ref *)(block_info + 1);
8442         }
8443
8444         if (node->type == BTRFS_SHARED_BLOCK_REF_KEY) {
8445                 BUG_ON(!(flags & BTRFS_BLOCK_FLAG_FULL_BACKREF));
8446                 btrfs_set_extent_inline_ref_type(leaf, iref,
8447                                                  BTRFS_SHARED_BLOCK_REF_KEY);
8448                 btrfs_set_extent_inline_ref_offset(leaf, iref, ref->parent);
8449         } else {
8450                 btrfs_set_extent_inline_ref_type(leaf, iref,
8451                                                  BTRFS_TREE_BLOCK_REF_KEY);
8452                 btrfs_set_extent_inline_ref_offset(leaf, iref, ref->root);
8453         }
8454
8455         btrfs_mark_buffer_dirty(leaf);
8456         btrfs_free_path(path);
8457
8458         ret = remove_from_free_space_tree(trans, extent_key.objectid,
8459                                           num_bytes);
8460         if (ret)
8461                 return ret;
8462
8463         ret = update_block_group(trans, fs_info, extent_key.objectid,
8464                                  fs_info->nodesize, 1);
8465         if (ret) { /* -ENOENT, logic error */
8466                 btrfs_err(fs_info, "update block group failed for %llu %llu",
8467                         extent_key.objectid, extent_key.offset);
8468                 BUG();
8469         }
8470
8471         trace_btrfs_reserved_extent_alloc(fs_info, extent_key.objectid,
8472                                           fs_info->nodesize);
8473         return ret;
8474 }
8475
8476 int btrfs_alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
8477                                      struct btrfs_root *root, u64 owner,
8478                                      u64 offset, u64 ram_bytes,
8479                                      struct btrfs_key *ins)
8480 {
8481         int ret;
8482
8483         BUG_ON(root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID);
8484
8485         btrfs_ref_tree_mod(root, ins->objectid, ins->offset, 0,
8486                            root->root_key.objectid, owner, offset,
8487                            BTRFS_ADD_DELAYED_EXTENT);
8488
8489         ret = btrfs_add_delayed_data_ref(trans, ins->objectid,
8490                                          ins->offset, 0,
8491                                          root->root_key.objectid, owner,
8492                                          offset, ram_bytes,
8493                                          BTRFS_ADD_DELAYED_EXTENT, NULL, NULL);
8494         return ret;
8495 }
8496
8497 /*
8498  * this is used by the tree logging recovery code.  It records that
8499  * an extent has been allocated and makes sure to clear the free
8500  * space cache bits as well
8501  */
8502 int btrfs_alloc_logged_file_extent(struct btrfs_trans_handle *trans,
8503                                    u64 root_objectid, u64 owner, u64 offset,
8504                                    struct btrfs_key *ins)
8505 {
8506         struct btrfs_fs_info *fs_info = trans->fs_info;
8507         int ret;
8508         struct btrfs_block_group_cache *block_group;
8509         struct btrfs_space_info *space_info;
8510
8511         /*
8512          * Mixed block groups will exclude before processing the log so we only
8513          * need to do the exclude dance if this fs isn't mixed.
8514          */
8515         if (!btrfs_fs_incompat(fs_info, MIXED_GROUPS)) {
8516                 ret = __exclude_logged_extent(fs_info, ins->objectid,
8517                                               ins->offset);
8518                 if (ret)
8519                         return ret;
8520         }
8521
8522         block_group = btrfs_lookup_block_group(fs_info, ins->objectid);
8523         if (!block_group)
8524                 return -EINVAL;
8525
8526         space_info = block_group->space_info;
8527         spin_lock(&space_info->lock);
8528         spin_lock(&block_group->lock);
8529         space_info->bytes_reserved += ins->offset;
8530         block_group->reserved += ins->offset;
8531         spin_unlock(&block_group->lock);
8532         spin_unlock(&space_info->lock);
8533
8534         ret = alloc_reserved_file_extent(trans, 0, root_objectid, 0, owner,
8535                                          offset, ins, 1);
8536         btrfs_put_block_group(block_group);
8537         return ret;
8538 }
8539
8540 static struct extent_buffer *
8541 btrfs_init_new_buffer(struct btrfs_trans_handle *trans, struct btrfs_root *root,
8542                       u64 bytenr, int level, u64 owner)
8543 {
8544         struct btrfs_fs_info *fs_info = root->fs_info;
8545         struct extent_buffer *buf;
8546
8547         buf = btrfs_find_create_tree_block(fs_info, bytenr);
8548         if (IS_ERR(buf))
8549                 return buf;
8550
8551         /*
8552          * Extra safety check in case the extent tree is corrupted and extent
8553          * allocator chooses to use a tree block which is already used and
8554          * locked.
8555          */
8556         if (buf->lock_owner == current->pid) {
8557                 btrfs_err_rl(fs_info,
8558 "tree block %llu owner %llu already locked by pid=%d, extent tree corruption detected",
8559                         buf->start, btrfs_header_owner(buf), current->pid);
8560                 free_extent_buffer(buf);
8561                 return ERR_PTR(-EUCLEAN);
8562         }
8563
8564         btrfs_set_buffer_lockdep_class(root->root_key.objectid, buf, level);
8565         btrfs_tree_lock(buf);
8566         clean_tree_block(fs_info, buf);
8567         clear_bit(EXTENT_BUFFER_STALE, &buf->bflags);
8568
8569         btrfs_set_lock_blocking_write(buf);
8570         set_extent_buffer_uptodate(buf);
8571
8572         memzero_extent_buffer(buf, 0, sizeof(struct btrfs_header));
8573         btrfs_set_header_level(buf, level);
8574         btrfs_set_header_bytenr(buf, buf->start);
8575         btrfs_set_header_generation(buf, trans->transid);
8576         btrfs_set_header_backref_rev(buf, BTRFS_MIXED_BACKREF_REV);
8577         btrfs_set_header_owner(buf, owner);
8578         write_extent_buffer_fsid(buf, fs_info->fs_devices->metadata_uuid);
8579         write_extent_buffer_chunk_tree_uuid(buf, fs_info->chunk_tree_uuid);
8580         if (root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID) {
8581                 buf->log_index = root->log_transid % 2;
8582                 /*
8583                  * we allow two log transactions at a time, use different
8584                  * EXTENT bit to differentiate dirty pages.
8585                  */
8586                 if (buf->log_index == 0)
8587                         set_extent_dirty(&root->dirty_log_pages, buf->start,
8588                                         buf->start + buf->len - 1, GFP_NOFS);
8589                 else
8590                         set_extent_new(&root->dirty_log_pages, buf->start,
8591                                         buf->start + buf->len - 1);
8592         } else {
8593                 buf->log_index = -1;
8594                 set_extent_dirty(&trans->transaction->dirty_pages, buf->start,
8595                          buf->start + buf->len - 1, GFP_NOFS);
8596         }
8597         trans->dirty = true;
8598         /* this returns a buffer locked for blocking */
8599         return buf;
8600 }
8601
8602 static struct btrfs_block_rsv *
8603 use_block_rsv(struct btrfs_trans_handle *trans,
8604               struct btrfs_root *root, u32 blocksize)
8605 {
8606         struct btrfs_fs_info *fs_info = root->fs_info;
8607         struct btrfs_block_rsv *block_rsv;
8608         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
8609         int ret;
8610         bool global_updated = false;
8611
8612         block_rsv = get_block_rsv(trans, root);
8613
8614         if (unlikely(block_rsv->size == 0))
8615                 goto try_reserve;
8616 again:
8617         ret = block_rsv_use_bytes(block_rsv, blocksize);
8618         if (!ret)
8619                 return block_rsv;
8620
8621         if (block_rsv->failfast)
8622                 return ERR_PTR(ret);
8623
8624         if (block_rsv->type == BTRFS_BLOCK_RSV_GLOBAL && !global_updated) {
8625                 global_updated = true;
8626                 update_global_block_rsv(fs_info);
8627                 goto again;
8628         }
8629
8630         /*
8631          * The global reserve still exists to save us from ourselves, so don't
8632          * warn_on if we are short on our delayed refs reserve.
8633          */
8634         if (block_rsv->type != BTRFS_BLOCK_RSV_DELREFS &&
8635             btrfs_test_opt(fs_info, ENOSPC_DEBUG)) {
8636                 static DEFINE_RATELIMIT_STATE(_rs,
8637                                 DEFAULT_RATELIMIT_INTERVAL * 10,
8638                                 /*DEFAULT_RATELIMIT_BURST*/ 1);
8639                 if (__ratelimit(&_rs))
8640                         WARN(1, KERN_DEBUG
8641                                 "BTRFS: block rsv returned %d\n", ret);
8642         }
8643 try_reserve:
8644         ret = reserve_metadata_bytes(root, block_rsv, blocksize,
8645                                      BTRFS_RESERVE_NO_FLUSH);
8646         if (!ret)
8647                 return block_rsv;
8648         /*
8649          * If we couldn't reserve metadata bytes try and use some from
8650          * the global reserve if its space type is the same as the global
8651          * reservation.
8652          */
8653         if (block_rsv->type != BTRFS_BLOCK_RSV_GLOBAL &&
8654             block_rsv->space_info == global_rsv->space_info) {
8655                 ret = block_rsv_use_bytes(global_rsv, blocksize);
8656                 if (!ret)
8657                         return global_rsv;
8658         }
8659         return ERR_PTR(ret);
8660 }
8661
8662 static void unuse_block_rsv(struct btrfs_fs_info *fs_info,
8663                             struct btrfs_block_rsv *block_rsv, u32 blocksize)
8664 {
8665         block_rsv_add_bytes(block_rsv, blocksize, false);
8666         block_rsv_release_bytes(fs_info, block_rsv, NULL, 0, NULL);
8667 }
8668
8669 /*
8670  * finds a free extent and does all the dirty work required for allocation
8671  * returns the tree buffer or an ERR_PTR on error.
8672  */
8673 struct extent_buffer *btrfs_alloc_tree_block(struct btrfs_trans_handle *trans,
8674                                              struct btrfs_root *root,
8675                                              u64 parent, u64 root_objectid,
8676                                              const struct btrfs_disk_key *key,
8677                                              int level, u64 hint,
8678                                              u64 empty_size)
8679 {
8680         struct btrfs_fs_info *fs_info = root->fs_info;
8681         struct btrfs_key ins;
8682         struct btrfs_block_rsv *block_rsv;
8683         struct extent_buffer *buf;
8684         struct btrfs_delayed_extent_op *extent_op;
8685         u64 flags = 0;
8686         int ret;
8687         u32 blocksize = fs_info->nodesize;
8688         bool skinny_metadata = btrfs_fs_incompat(fs_info, SKINNY_METADATA);
8689
8690 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
8691         if (btrfs_is_testing(fs_info)) {
8692                 buf = btrfs_init_new_buffer(trans, root, root->alloc_bytenr,
8693                                             level, root_objectid);
8694                 if (!IS_ERR(buf))
8695                         root->alloc_bytenr += blocksize;
8696                 return buf;
8697         }
8698 #endif
8699
8700         block_rsv = use_block_rsv(trans, root, blocksize);
8701         if (IS_ERR(block_rsv))
8702                 return ERR_CAST(block_rsv);
8703
8704         ret = btrfs_reserve_extent(root, blocksize, blocksize, blocksize,
8705                                    empty_size, hint, &ins, 0, 0);
8706         if (ret)
8707                 goto out_unuse;
8708
8709         buf = btrfs_init_new_buffer(trans, root, ins.objectid, level,
8710                                     root_objectid);
8711         if (IS_ERR(buf)) {
8712                 ret = PTR_ERR(buf);
8713                 goto out_free_reserved;
8714         }
8715
8716         if (root_objectid == BTRFS_TREE_RELOC_OBJECTID) {
8717                 if (parent == 0)
8718                         parent = ins.objectid;
8719                 flags |= BTRFS_BLOCK_FLAG_FULL_BACKREF;
8720         } else
8721                 BUG_ON(parent > 0);
8722
8723         if (root_objectid != BTRFS_TREE_LOG_OBJECTID) {
8724                 extent_op = btrfs_alloc_delayed_extent_op();
8725                 if (!extent_op) {
8726                         ret = -ENOMEM;
8727                         goto out_free_buf;
8728                 }
8729                 if (key)
8730                         memcpy(&extent_op->key, key, sizeof(extent_op->key));
8731                 else
8732                         memset(&extent_op->key, 0, sizeof(extent_op->key));
8733                 extent_op->flags_to_set = flags;
8734                 extent_op->update_key = skinny_metadata ? false : true;
8735                 extent_op->update_flags = true;
8736                 extent_op->is_data = false;
8737                 extent_op->level = level;
8738
8739                 btrfs_ref_tree_mod(root, ins.objectid, ins.offset, parent,
8740                                    root_objectid, level, 0,
8741                                    BTRFS_ADD_DELAYED_EXTENT);
8742                 ret = btrfs_add_delayed_tree_ref(trans, ins.objectid,
8743                                                  ins.offset, parent,
8744                                                  root_objectid, level,
8745                                                  BTRFS_ADD_DELAYED_EXTENT,
8746                                                  extent_op, NULL, NULL);
8747                 if (ret)
8748                         goto out_free_delayed;
8749         }
8750         return buf;
8751
8752 out_free_delayed:
8753         btrfs_free_delayed_extent_op(extent_op);
8754 out_free_buf:
8755         free_extent_buffer(buf);
8756 out_free_reserved:
8757         btrfs_free_reserved_extent(fs_info, ins.objectid, ins.offset, 0);
8758 out_unuse:
8759         unuse_block_rsv(fs_info, block_rsv, blocksize);
8760         return ERR_PTR(ret);
8761 }
8762
8763 struct walk_control {
8764         u64 refs[BTRFS_MAX_LEVEL];
8765         u64 flags[BTRFS_MAX_LEVEL];
8766         struct btrfs_key update_progress;
8767         struct btrfs_key drop_progress;
8768         int drop_level;
8769         int stage;
8770         int level;
8771         int shared_level;
8772         int update_ref;
8773         int keep_locks;
8774         int reada_slot;
8775         int reada_count;
8776         int restarted;
8777 };
8778
8779 #define DROP_REFERENCE  1
8780 #define UPDATE_BACKREF  2
8781
8782 static noinline void reada_walk_down(struct btrfs_trans_handle *trans,
8783                                      struct btrfs_root *root,
8784                                      struct walk_control *wc,
8785                                      struct btrfs_path *path)
8786 {
8787         struct btrfs_fs_info *fs_info = root->fs_info;
8788         u64 bytenr;
8789         u64 generation;
8790         u64 refs;
8791         u64 flags;
8792         u32 nritems;
8793         struct btrfs_key key;
8794         struct extent_buffer *eb;
8795         int ret;
8796         int slot;
8797         int nread = 0;
8798
8799         if (path->slots[wc->level] < wc->reada_slot) {
8800                 wc->reada_count = wc->reada_count * 2 / 3;
8801                 wc->reada_count = max(wc->reada_count, 2);
8802         } else {
8803                 wc->reada_count = wc->reada_count * 3 / 2;
8804                 wc->reada_count = min_t(int, wc->reada_count,
8805                                         BTRFS_NODEPTRS_PER_BLOCK(fs_info));
8806         }
8807
8808         eb = path->nodes[wc->level];
8809         nritems = btrfs_header_nritems(eb);
8810
8811         for (slot = path->slots[wc->level]; slot < nritems; slot++) {
8812                 if (nread >= wc->reada_count)
8813                         break;
8814
8815                 cond_resched();
8816                 bytenr = btrfs_node_blockptr(eb, slot);
8817                 generation = btrfs_node_ptr_generation(eb, slot);
8818
8819                 if (slot == path->slots[wc->level])
8820                         goto reada;
8821
8822                 if (wc->stage == UPDATE_BACKREF &&
8823                     generation <= root->root_key.offset)
8824                         continue;
8825
8826                 /* We don't lock the tree block, it's OK to be racy here */
8827                 ret = btrfs_lookup_extent_info(trans, fs_info, bytenr,
8828                                                wc->level - 1, 1, &refs,
8829                                                &flags);
8830                 /* We don't care about errors in readahead. */
8831                 if (ret < 0)
8832                         continue;
8833                 BUG_ON(refs == 0);
8834
8835                 if (wc->stage == DROP_REFERENCE) {
8836                         if (refs == 1)
8837                                 goto reada;
8838
8839                         if (wc->level == 1 &&
8840                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
8841                                 continue;
8842                         if (!wc->update_ref ||
8843                             generation <= root->root_key.offset)
8844                                 continue;
8845                         btrfs_node_key_to_cpu(eb, &key, slot);
8846                         ret = btrfs_comp_cpu_keys(&key,
8847                                                   &wc->update_progress);
8848                         if (ret < 0)
8849                                 continue;
8850                 } else {
8851                         if (wc->level == 1 &&
8852                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
8853                                 continue;
8854                 }
8855 reada:
8856                 readahead_tree_block(fs_info, bytenr);
8857                 nread++;
8858         }
8859         wc->reada_slot = slot;
8860 }
8861
8862 /*
8863  * helper to process tree block while walking down the tree.
8864  *
8865  * when wc->stage == UPDATE_BACKREF, this function updates
8866  * back refs for pointers in the block.
8867  *
8868  * NOTE: return value 1 means we should stop walking down.
8869  */
8870 static noinline int walk_down_proc(struct btrfs_trans_handle *trans,
8871                                    struct btrfs_root *root,
8872                                    struct btrfs_path *path,
8873                                    struct walk_control *wc, int lookup_info)
8874 {
8875         struct btrfs_fs_info *fs_info = root->fs_info;
8876         int level = wc->level;
8877         struct extent_buffer *eb = path->nodes[level];
8878         u64 flag = BTRFS_BLOCK_FLAG_FULL_BACKREF;
8879         int ret;
8880
8881         if (wc->stage == UPDATE_BACKREF &&
8882             btrfs_header_owner(eb) != root->root_key.objectid)
8883                 return 1;
8884
8885         /*
8886          * when reference count of tree block is 1, it won't increase
8887          * again. once full backref flag is set, we never clear it.
8888          */
8889         if (lookup_info &&
8890             ((wc->stage == DROP_REFERENCE && wc->refs[level] != 1) ||
8891              (wc->stage == UPDATE_BACKREF && !(wc->flags[level] & flag)))) {
8892                 BUG_ON(!path->locks[level]);
8893                 ret = btrfs_lookup_extent_info(trans, fs_info,
8894                                                eb->start, level, 1,
8895                                                &wc->refs[level],
8896                                                &wc->flags[level]);
8897                 BUG_ON(ret == -ENOMEM);
8898                 if (ret)
8899                         return ret;
8900                 BUG_ON(wc->refs[level] == 0);
8901         }
8902
8903         if (wc->stage == DROP_REFERENCE) {
8904                 if (wc->refs[level] > 1)
8905                         return 1;
8906
8907                 if (path->locks[level] && !wc->keep_locks) {
8908                         btrfs_tree_unlock_rw(eb, path->locks[level]);
8909                         path->locks[level] = 0;
8910                 }
8911                 return 0;
8912         }
8913
8914         /* wc->stage == UPDATE_BACKREF */
8915         if (!(wc->flags[level] & flag)) {
8916                 BUG_ON(!path->locks[level]);
8917                 ret = btrfs_inc_ref(trans, root, eb, 1);
8918                 BUG_ON(ret); /* -ENOMEM */
8919                 ret = btrfs_dec_ref(trans, root, eb, 0);
8920                 BUG_ON(ret); /* -ENOMEM */
8921                 ret = btrfs_set_disk_extent_flags(trans, fs_info, eb->start,
8922                                                   eb->len, flag,
8923                                                   btrfs_header_level(eb), 0);
8924                 BUG_ON(ret); /* -ENOMEM */
8925                 wc->flags[level] |= flag;
8926         }
8927
8928         /*
8929          * the block is shared by multiple trees, so it's not good to
8930          * keep the tree lock
8931          */
8932         if (path->locks[level] && level > 0) {
8933                 btrfs_tree_unlock_rw(eb, path->locks[level]);
8934                 path->locks[level] = 0;
8935         }
8936         return 0;
8937 }
8938
8939 /*
8940  * This is used to verify a ref exists for this root to deal with a bug where we
8941  * would have a drop_progress key that hadn't been updated properly.
8942  */
8943 static int check_ref_exists(struct btrfs_trans_handle *trans,
8944                             struct btrfs_root *root, u64 bytenr, u64 parent,
8945                             int level)
8946 {
8947         struct btrfs_path *path;
8948         struct btrfs_extent_inline_ref *iref;
8949         int ret;
8950
8951         path = btrfs_alloc_path();
8952         if (!path)
8953                 return -ENOMEM;
8954
8955         ret = lookup_extent_backref(trans, path, &iref, bytenr,
8956                                     root->fs_info->nodesize, parent,
8957                                     root->root_key.objectid, level, 0);
8958         btrfs_free_path(path);
8959         if (ret == -ENOENT)
8960                 return 0;
8961         if (ret < 0)
8962                 return ret;
8963         return 1;
8964 }
8965
8966 /*
8967  * helper to process tree block pointer.
8968  *
8969  * when wc->stage == DROP_REFERENCE, this function checks
8970  * reference count of the block pointed to. if the block
8971  * is shared and we need update back refs for the subtree
8972  * rooted at the block, this function changes wc->stage to
8973  * UPDATE_BACKREF. if the block is shared and there is no
8974  * need to update back, this function drops the reference
8975  * to the block.
8976  *
8977  * NOTE: return value 1 means we should stop walking down.
8978  */
8979 static noinline int do_walk_down(struct btrfs_trans_handle *trans,
8980                                  struct btrfs_root *root,
8981                                  struct btrfs_path *path,
8982                                  struct walk_control *wc, int *lookup_info)
8983 {
8984         struct btrfs_fs_info *fs_info = root->fs_info;
8985         u64 bytenr;
8986         u64 generation;
8987         u64 parent;
8988         struct btrfs_key key;
8989         struct btrfs_key first_key;
8990         struct extent_buffer *next;
8991         int level = wc->level;
8992         int reada = 0;
8993         int ret = 0;
8994         bool need_account = false;
8995
8996         generation = btrfs_node_ptr_generation(path->nodes[level],
8997                                                path->slots[level]);
8998         /*
8999          * if the lower level block was created before the snapshot
9000          * was created, we know there is no need to update back refs
9001          * for the subtree
9002          */
9003         if (wc->stage == UPDATE_BACKREF &&
9004             generation <= root->root_key.offset) {
9005                 *lookup_info = 1;
9006                 return 1;
9007         }
9008
9009         bytenr = btrfs_node_blockptr(path->nodes[level], path->slots[level]);
9010         btrfs_node_key_to_cpu(path->nodes[level], &first_key,
9011                               path->slots[level]);
9012
9013         next = find_extent_buffer(fs_info, bytenr);
9014         if (!next) {
9015                 next = btrfs_find_create_tree_block(fs_info, bytenr);
9016                 if (IS_ERR(next))
9017                         return PTR_ERR(next);
9018
9019                 btrfs_set_buffer_lockdep_class(root->root_key.objectid, next,
9020                                                level - 1);
9021                 reada = 1;
9022         }
9023         btrfs_tree_lock(next);
9024         btrfs_set_lock_blocking_write(next);
9025
9026         ret = btrfs_lookup_extent_info(trans, fs_info, bytenr, level - 1, 1,
9027                                        &wc->refs[level - 1],
9028                                        &wc->flags[level - 1]);
9029         if (ret < 0)
9030                 goto out_unlock;
9031
9032         if (unlikely(wc->refs[level - 1] == 0)) {
9033                 btrfs_err(fs_info, "Missing references.");
9034                 ret = -EIO;
9035                 goto out_unlock;
9036         }
9037         *lookup_info = 0;
9038
9039         if (wc->stage == DROP_REFERENCE) {
9040                 if (wc->refs[level - 1] > 1) {
9041                         need_account = true;
9042                         if (level == 1 &&
9043                             (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
9044                                 goto skip;
9045
9046                         if (!wc->update_ref ||
9047                             generation <= root->root_key.offset)
9048                                 goto skip;
9049
9050                         btrfs_node_key_to_cpu(path->nodes[level], &key,
9051                                               path->slots[level]);
9052                         ret = btrfs_comp_cpu_keys(&key, &wc->update_progress);
9053                         if (ret < 0)
9054                                 goto skip;
9055
9056                         wc->stage = UPDATE_BACKREF;
9057                         wc->shared_level = level - 1;
9058                 }
9059         } else {
9060                 if (level == 1 &&
9061                     (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
9062                         goto skip;
9063         }
9064
9065         if (!btrfs_buffer_uptodate(next, generation, 0)) {
9066                 btrfs_tree_unlock(next);
9067                 free_extent_buffer(next);
9068                 next = NULL;
9069                 *lookup_info = 1;
9070         }
9071
9072         if (!next) {
9073                 if (reada && level == 1)
9074                         reada_walk_down(trans, root, wc, path);
9075                 next = read_tree_block(fs_info, bytenr, generation, level - 1,
9076                                        &first_key);
9077                 if (IS_ERR(next)) {
9078                         return PTR_ERR(next);
9079                 } else if (!extent_buffer_uptodate(next)) {
9080                         free_extent_buffer(next);
9081                         return -EIO;
9082                 }
9083                 btrfs_tree_lock(next);
9084                 btrfs_set_lock_blocking_write(next);
9085         }
9086
9087         level--;
9088         ASSERT(level == btrfs_header_level(next));
9089         if (level != btrfs_header_level(next)) {
9090                 btrfs_err(root->fs_info, "mismatched level");
9091                 ret = -EIO;
9092                 goto out_unlock;
9093         }
9094         path->nodes[level] = next;
9095         path->slots[level] = 0;
9096         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
9097         wc->level = level;
9098         if (wc->level == 1)
9099                 wc->reada_slot = 0;
9100         return 0;
9101 skip:
9102         wc->refs[level - 1] = 0;
9103         wc->flags[level - 1] = 0;
9104         if (wc->stage == DROP_REFERENCE) {
9105                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF) {
9106                         parent = path->nodes[level]->start;
9107                 } else {
9108                         ASSERT(root->root_key.objectid ==
9109                                btrfs_header_owner(path->nodes[level]));
9110                         if (root->root_key.objectid !=
9111                             btrfs_header_owner(path->nodes[level])) {
9112                                 btrfs_err(root->fs_info,
9113                                                 "mismatched block owner");
9114                                 ret = -EIO;
9115                                 goto out_unlock;
9116                         }
9117                         parent = 0;
9118                 }
9119
9120                 /*
9121                  * If we had a drop_progress we need to verify the refs are set
9122                  * as expected.  If we find our ref then we know that from here
9123                  * on out everything should be correct, and we can clear the
9124                  * ->restarted flag.
9125                  */
9126                 if (wc->restarted) {
9127                         ret = check_ref_exists(trans, root, bytenr, parent,
9128                                                level - 1);
9129                         if (ret < 0)
9130                                 goto out_unlock;
9131                         if (ret == 0)
9132                                 goto no_delete;
9133                         ret = 0;
9134                         wc->restarted = 0;
9135                 }
9136
9137                 /*
9138                  * Reloc tree doesn't contribute to qgroup numbers, and we have
9139                  * already accounted them at merge time (replace_path),
9140                  * thus we could skip expensive subtree trace here.
9141                  */
9142                 if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID &&
9143                     need_account) {
9144                         ret = btrfs_qgroup_trace_subtree(trans, next,
9145                                                          generation, level - 1);
9146                         if (ret) {
9147                                 btrfs_err_rl(fs_info,
9148                                              "Error %d accounting shared subtree. Quota is out of sync, rescan required.",
9149                                              ret);
9150                         }
9151                 }
9152
9153                 /*
9154                  * We need to update the next key in our walk control so we can
9155                  * update the drop_progress key accordingly.  We don't care if
9156                  * find_next_key doesn't find a key because that means we're at
9157                  * the end and are going to clean up now.
9158                  */
9159                 wc->drop_level = level;
9160                 find_next_key(path, level, &wc->drop_progress);
9161
9162                 ret = btrfs_free_extent(trans, root, bytenr, fs_info->nodesize,
9163                                         parent, root->root_key.objectid,
9164                                         level - 1, 0);
9165                 if (ret)
9166                         goto out_unlock;
9167         }
9168 no_delete:
9169         *lookup_info = 1;
9170         ret = 1;
9171
9172 out_unlock:
9173         btrfs_tree_unlock(next);
9174         free_extent_buffer(next);
9175
9176         return ret;
9177 }
9178
9179 /*
9180  * helper to process tree block while walking up the tree.
9181  *
9182  * when wc->stage == DROP_REFERENCE, this function drops
9183  * reference count on the block.
9184  *
9185  * when wc->stage == UPDATE_BACKREF, this function changes
9186  * wc->stage back to DROP_REFERENCE if we changed wc->stage
9187  * to UPDATE_BACKREF previously while processing the block.
9188  *
9189  * NOTE: return value 1 means we should stop walking up.
9190  */
9191 static noinline int walk_up_proc(struct btrfs_trans_handle *trans,
9192                                  struct btrfs_root *root,
9193                                  struct btrfs_path *path,
9194                                  struct walk_control *wc)
9195 {
9196         struct btrfs_fs_info *fs_info = root->fs_info;
9197         int ret;
9198         int level = wc->level;
9199         struct extent_buffer *eb = path->nodes[level];
9200         u64 parent = 0;
9201
9202         if (wc->stage == UPDATE_BACKREF) {
9203                 BUG_ON(wc->shared_level < level);
9204                 if (level < wc->shared_level)
9205                         goto out;
9206
9207                 ret = find_next_key(path, level + 1, &wc->update_progress);
9208                 if (ret > 0)
9209                         wc->update_ref = 0;
9210
9211                 wc->stage = DROP_REFERENCE;
9212                 wc->shared_level = -1;
9213                 path->slots[level] = 0;
9214
9215                 /*
9216                  * check reference count again if the block isn't locked.
9217                  * we should start walking down the tree again if reference
9218                  * count is one.
9219                  */
9220                 if (!path->locks[level]) {
9221                         BUG_ON(level == 0);
9222                         btrfs_tree_lock(eb);
9223                         btrfs_set_lock_blocking_write(eb);
9224                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
9225
9226                         ret = btrfs_lookup_extent_info(trans, fs_info,
9227                                                        eb->start, level, 1,
9228                                                        &wc->refs[level],
9229                                                        &wc->flags[level]);
9230                         if (ret < 0) {
9231                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
9232                                 path->locks[level] = 0;
9233                                 return ret;
9234                         }
9235                         BUG_ON(wc->refs[level] == 0);
9236                         if (wc->refs[level] == 1) {
9237                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
9238                                 path->locks[level] = 0;
9239                                 return 1;
9240                         }
9241                 }
9242         }
9243
9244         /* wc->stage == DROP_REFERENCE */
9245         BUG_ON(wc->refs[level] > 1 && !path->locks[level]);
9246
9247         if (wc->refs[level] == 1) {
9248                 if (level == 0) {
9249                         if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
9250                                 ret = btrfs_dec_ref(trans, root, eb, 1);
9251                         else
9252                                 ret = btrfs_dec_ref(trans, root, eb, 0);
9253                         BUG_ON(ret); /* -ENOMEM */
9254                         ret = btrfs_qgroup_trace_leaf_items(trans, eb);
9255                         if (ret) {
9256                                 btrfs_err_rl(fs_info,
9257                                              "error %d accounting leaf items. Quota is out of sync, rescan required.",
9258                                              ret);
9259                         }
9260                 }
9261                 /* make block locked assertion in clean_tree_block happy */
9262                 if (!path->locks[level] &&
9263                     btrfs_header_generation(eb) == trans->transid) {
9264                         btrfs_tree_lock(eb);
9265                         btrfs_set_lock_blocking_write(eb);
9266                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
9267                 }
9268                 clean_tree_block(fs_info, eb);
9269         }
9270
9271         if (eb == root->node) {
9272                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
9273                         parent = eb->start;
9274                 else if (root->root_key.objectid != btrfs_header_owner(eb))
9275                         goto owner_mismatch;
9276         } else {
9277                 if (wc->flags[level + 1] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
9278                         parent = path->nodes[level + 1]->start;
9279                 else if (root->root_key.objectid !=
9280                          btrfs_header_owner(path->nodes[level + 1]))
9281                         goto owner_mismatch;
9282         }
9283
9284         btrfs_free_tree_block(trans, root, eb, parent, wc->refs[level] == 1);
9285 out:
9286         wc->refs[level] = 0;
9287         wc->flags[level] = 0;
9288         return 0;
9289
9290 owner_mismatch:
9291         btrfs_err_rl(fs_info, "unexpected tree owner, have %llu expect %llu",
9292                      btrfs_header_owner(eb), root->root_key.objectid);
9293         return -EUCLEAN;
9294 }
9295
9296 static noinline int walk_down_tree(struct btrfs_trans_handle *trans,
9297                                    struct btrfs_root *root,
9298                                    struct btrfs_path *path,
9299                                    struct walk_control *wc)
9300 {
9301         int level = wc->level;
9302         int lookup_info = 1;
9303         int ret;
9304
9305         while (level >= 0) {
9306                 ret = walk_down_proc(trans, root, path, wc, lookup_info);
9307                 if (ret > 0)
9308                         break;
9309
9310                 if (level == 0)
9311                         break;
9312
9313                 if (path->slots[level] >=
9314                     btrfs_header_nritems(path->nodes[level]))
9315                         break;
9316
9317                 ret = do_walk_down(trans, root, path, wc, &lookup_info);
9318                 if (ret > 0) {
9319                         path->slots[level]++;
9320                         continue;
9321                 } else if (ret < 0)
9322                         return ret;
9323                 level = wc->level;
9324         }
9325         return 0;
9326 }
9327
9328 static noinline int walk_up_tree(struct btrfs_trans_handle *trans,
9329                                  struct btrfs_root *root,
9330                                  struct btrfs_path *path,
9331                                  struct walk_control *wc, int max_level)
9332 {
9333         int level = wc->level;
9334         int ret;
9335
9336         path->slots[level] = btrfs_header_nritems(path->nodes[level]);
9337         while (level < max_level && path->nodes[level]) {
9338                 wc->level = level;
9339                 if (path->slots[level] + 1 <
9340                     btrfs_header_nritems(path->nodes[level])) {
9341                         path->slots[level]++;
9342                         return 0;
9343                 } else {
9344                         ret = walk_up_proc(trans, root, path, wc);
9345                         if (ret > 0)
9346                                 return 0;
9347                         if (ret < 0)
9348                                 return ret;
9349
9350                         if (path->locks[level]) {
9351                                 btrfs_tree_unlock_rw(path->nodes[level],
9352                                                      path->locks[level]);
9353                                 path->locks[level] = 0;
9354                         }
9355                         free_extent_buffer(path->nodes[level]);
9356                         path->nodes[level] = NULL;
9357                         level++;
9358                 }
9359         }
9360         return 1;
9361 }
9362
9363 /*
9364  * drop a subvolume tree.
9365  *
9366  * this function traverses the tree freeing any blocks that only
9367  * referenced by the tree.
9368  *
9369  * when a shared tree block is found. this function decreases its
9370  * reference count by one. if update_ref is true, this function
9371  * also make sure backrefs for the shared block and all lower level
9372  * blocks are properly updated.
9373  *
9374  * If called with for_reloc == 0, may exit early with -EAGAIN
9375  */
9376 int btrfs_drop_snapshot(struct btrfs_root *root,
9377                          struct btrfs_block_rsv *block_rsv, int update_ref,
9378                          int for_reloc)
9379 {
9380         struct btrfs_fs_info *fs_info = root->fs_info;
9381         struct btrfs_path *path;
9382         struct btrfs_trans_handle *trans;
9383         struct btrfs_root *tree_root = fs_info->tree_root;
9384         struct btrfs_root_item *root_item = &root->root_item;
9385         struct walk_control *wc;
9386         struct btrfs_key key;
9387         int err = 0;
9388         int ret;
9389         int level;
9390         bool root_dropped = false;
9391
9392         btrfs_debug(fs_info, "Drop subvolume %llu", root->root_key.objectid);
9393
9394         path = btrfs_alloc_path();
9395         if (!path) {
9396                 err = -ENOMEM;
9397                 goto out;
9398         }
9399
9400         wc = kzalloc(sizeof(*wc), GFP_NOFS);
9401         if (!wc) {
9402                 btrfs_free_path(path);
9403                 err = -ENOMEM;
9404                 goto out;
9405         }
9406
9407         trans = btrfs_start_transaction(tree_root, 0);
9408         if (IS_ERR(trans)) {
9409                 err = PTR_ERR(trans);
9410                 goto out_free;
9411         }
9412
9413         err = btrfs_run_delayed_items(trans);
9414         if (err)
9415                 goto out_end_trans;
9416
9417         if (block_rsv)
9418                 trans->block_rsv = block_rsv;
9419
9420         /*
9421          * This will help us catch people modifying the fs tree while we're
9422          * dropping it.  It is unsafe to mess with the fs tree while it's being
9423          * dropped as we unlock the root node and parent nodes as we walk down
9424          * the tree, assuming nothing will change.  If something does change
9425          * then we'll have stale information and drop references to blocks we've
9426          * already dropped.
9427          */
9428         set_bit(BTRFS_ROOT_DELETING, &root->state);
9429         if (btrfs_disk_key_objectid(&root_item->drop_progress) == 0) {
9430                 level = btrfs_header_level(root->node);
9431                 path->nodes[level] = btrfs_lock_root_node(root);
9432                 btrfs_set_lock_blocking_write(path->nodes[level]);
9433                 path->slots[level] = 0;
9434                 path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
9435                 memset(&wc->update_progress, 0,
9436                        sizeof(wc->update_progress));
9437         } else {
9438                 btrfs_disk_key_to_cpu(&key, &root_item->drop_progress);
9439                 memcpy(&wc->update_progress, &key,
9440                        sizeof(wc->update_progress));
9441
9442                 level = root_item->drop_level;
9443                 BUG_ON(level == 0);
9444                 path->lowest_level = level;
9445                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
9446                 path->lowest_level = 0;
9447                 if (ret < 0) {
9448                         err = ret;
9449                         goto out_end_trans;
9450                 }
9451                 WARN_ON(ret > 0);
9452
9453                 /*
9454                  * unlock our path, this is safe because only this
9455                  * function is allowed to delete this snapshot
9456                  */
9457                 btrfs_unlock_up_safe(path, 0);
9458
9459                 level = btrfs_header_level(root->node);
9460                 while (1) {
9461                         btrfs_tree_lock(path->nodes[level]);
9462                         btrfs_set_lock_blocking_write(path->nodes[level]);
9463                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
9464
9465                         ret = btrfs_lookup_extent_info(trans, fs_info,
9466                                                 path->nodes[level]->start,
9467                                                 level, 1, &wc->refs[level],
9468                                                 &wc->flags[level]);
9469                         if (ret < 0) {
9470                                 err = ret;
9471                                 goto out_end_trans;
9472                         }
9473                         BUG_ON(wc->refs[level] == 0);
9474
9475                         if (level == root_item->drop_level)
9476                                 break;
9477
9478                         btrfs_tree_unlock(path->nodes[level]);
9479                         path->locks[level] = 0;
9480                         WARN_ON(wc->refs[level] != 1);
9481                         level--;
9482                 }
9483         }
9484
9485         wc->restarted = test_bit(BTRFS_ROOT_DEAD_TREE, &root->state);
9486         wc->level = level;
9487         wc->shared_level = -1;
9488         wc->stage = DROP_REFERENCE;
9489         wc->update_ref = update_ref;
9490         wc->keep_locks = 0;
9491         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(fs_info);
9492
9493         while (1) {
9494
9495                 ret = walk_down_tree(trans, root, path, wc);
9496                 if (ret < 0) {
9497                         err = ret;
9498                         break;
9499                 }
9500
9501                 ret = walk_up_tree(trans, root, path, wc, BTRFS_MAX_LEVEL);
9502                 if (ret < 0) {
9503                         err = ret;
9504                         break;
9505                 }
9506
9507                 if (ret > 0) {
9508                         BUG_ON(wc->stage != DROP_REFERENCE);
9509                         break;
9510                 }
9511
9512                 if (wc->stage == DROP_REFERENCE) {
9513                         wc->drop_level = wc->level;
9514                         btrfs_node_key_to_cpu(path->nodes[wc->drop_level],
9515                                               &wc->drop_progress,
9516                                               path->slots[wc->drop_level]);
9517                 }
9518                 btrfs_cpu_key_to_disk(&root_item->drop_progress,
9519                                       &wc->drop_progress);
9520                 root_item->drop_level = wc->drop_level;
9521
9522                 BUG_ON(wc->level == 0);
9523                 if (btrfs_should_end_transaction(trans) ||
9524                     (!for_reloc && btrfs_need_cleaner_sleep(fs_info))) {
9525                         ret = btrfs_update_root(trans, tree_root,
9526                                                 &root->root_key,
9527                                                 root_item);
9528                         if (ret) {
9529                                 btrfs_abort_transaction(trans, ret);
9530                                 err = ret;
9531                                 goto out_end_trans;
9532                         }
9533
9534                         btrfs_end_transaction_throttle(trans);
9535                         if (!for_reloc && btrfs_need_cleaner_sleep(fs_info)) {
9536                                 btrfs_debug(fs_info,
9537                                             "drop snapshot early exit");
9538                                 err = -EAGAIN;
9539                                 goto out_free;
9540                         }
9541
9542                         trans = btrfs_start_transaction(tree_root, 0);
9543                         if (IS_ERR(trans)) {
9544                                 err = PTR_ERR(trans);
9545                                 goto out_free;
9546                         }
9547                         if (block_rsv)
9548                                 trans->block_rsv = block_rsv;
9549                 }
9550         }
9551         btrfs_release_path(path);
9552         if (err)
9553                 goto out_end_trans;
9554
9555         ret = btrfs_del_root(trans, &root->root_key);
9556         if (ret) {
9557                 btrfs_abort_transaction(trans, ret);
9558                 err = ret;
9559                 goto out_end_trans;
9560         }
9561
9562         if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) {
9563                 ret = btrfs_find_root(tree_root, &root->root_key, path,
9564                                       NULL, NULL);
9565                 if (ret < 0) {
9566                         btrfs_abort_transaction(trans, ret);
9567                         err = ret;
9568                         goto out_end_trans;
9569                 } else if (ret > 0) {
9570                         /* if we fail to delete the orphan item this time
9571                          * around, it'll get picked up the next time.
9572                          *
9573                          * The most common failure here is just -ENOENT.
9574                          */
9575                         btrfs_del_orphan_item(trans, tree_root,
9576                                               root->root_key.objectid);
9577                 }
9578         }
9579
9580         if (test_bit(BTRFS_ROOT_IN_RADIX, &root->state)) {
9581                 btrfs_add_dropped_root(trans, root);
9582         } else {
9583                 free_extent_buffer(root->node);
9584                 free_extent_buffer(root->commit_root);
9585                 btrfs_put_fs_root(root);
9586         }
9587         root_dropped = true;
9588 out_end_trans:
9589         btrfs_end_transaction_throttle(trans);
9590 out_free:
9591         kfree(wc);
9592         btrfs_free_path(path);
9593 out:
9594         /*
9595          * So if we need to stop dropping the snapshot for whatever reason we
9596          * need to make sure to add it back to the dead root list so that we
9597          * keep trying to do the work later.  This also cleans up roots if we
9598          * don't have it in the radix (like when we recover after a power fail
9599          * or unmount) so we don't leak memory.
9600          */
9601         if (!for_reloc && !root_dropped)
9602                 btrfs_add_dead_root(root);
9603         if (err && err != -EAGAIN)
9604                 btrfs_handle_fs_error(fs_info, err, NULL);
9605         return err;
9606 }
9607
9608 /*
9609  * drop subtree rooted at tree block 'node'.
9610  *
9611  * NOTE: this function will unlock and release tree block 'node'
9612  * only used by relocation code
9613  */
9614 int btrfs_drop_subtree(struct btrfs_trans_handle *trans,
9615                         struct btrfs_root *root,
9616                         struct extent_buffer *node,
9617                         struct extent_buffer *parent)
9618 {
9619         struct btrfs_fs_info *fs_info = root->fs_info;
9620         struct btrfs_path *path;
9621         struct walk_control *wc;
9622         int level;
9623         int parent_level;
9624         int ret = 0;
9625         int wret;
9626
9627         BUG_ON(root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID);
9628
9629         path = btrfs_alloc_path();
9630         if (!path)
9631                 return -ENOMEM;
9632
9633         wc = kzalloc(sizeof(*wc), GFP_NOFS);
9634         if (!wc) {
9635                 btrfs_free_path(path);
9636                 return -ENOMEM;
9637         }
9638
9639         btrfs_assert_tree_locked(parent);
9640         parent_level = btrfs_header_level(parent);
9641         extent_buffer_get(parent);
9642         path->nodes[parent_level] = parent;
9643         path->slots[parent_level] = btrfs_header_nritems(parent);
9644
9645         btrfs_assert_tree_locked(node);
9646         level = btrfs_header_level(node);
9647         path->nodes[level] = node;
9648         path->slots[level] = 0;
9649         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
9650
9651         wc->refs[parent_level] = 1;
9652         wc->flags[parent_level] = BTRFS_BLOCK_FLAG_FULL_BACKREF;
9653         wc->level = level;
9654         wc->shared_level = -1;
9655         wc->stage = DROP_REFERENCE;
9656         wc->update_ref = 0;
9657         wc->keep_locks = 1;
9658         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(fs_info);
9659
9660         while (1) {
9661                 wret = walk_down_tree(trans, root, path, wc);
9662                 if (wret < 0) {
9663                         ret = wret;
9664                         break;
9665                 }
9666
9667                 wret = walk_up_tree(trans, root, path, wc, parent_level);
9668                 if (wret < 0)
9669                         ret = wret;
9670                 if (wret != 0)
9671                         break;
9672         }
9673
9674         kfree(wc);
9675         btrfs_free_path(path);
9676         return ret;
9677 }
9678
9679 static u64 update_block_group_flags(struct btrfs_fs_info *fs_info, u64 flags)
9680 {
9681         u64 num_devices;
9682         u64 stripped;
9683
9684         /*
9685          * if restripe for this chunk_type is on pick target profile and
9686          * return, otherwise do the usual balance
9687          */
9688         stripped = get_restripe_target(fs_info, flags);
9689         if (stripped)
9690                 return extended_to_chunk(stripped);
9691
9692         num_devices = fs_info->fs_devices->rw_devices;
9693
9694         stripped = BTRFS_BLOCK_GROUP_RAID0 |
9695                 BTRFS_BLOCK_GROUP_RAID5 | BTRFS_BLOCK_GROUP_RAID6 |
9696                 BTRFS_BLOCK_GROUP_RAID1 | BTRFS_BLOCK_GROUP_RAID10;
9697
9698         if (num_devices == 1) {
9699                 stripped |= BTRFS_BLOCK_GROUP_DUP;
9700                 stripped = flags & ~stripped;
9701
9702                 /* turn raid0 into single device chunks */
9703                 if (flags & BTRFS_BLOCK_GROUP_RAID0)
9704                         return stripped;
9705
9706                 /* turn mirroring into duplication */
9707                 if (flags & (BTRFS_BLOCK_GROUP_RAID1 |
9708                              BTRFS_BLOCK_GROUP_RAID10))
9709                         return stripped | BTRFS_BLOCK_GROUP_DUP;
9710         } else {
9711                 /* they already had raid on here, just return */
9712                 if (flags & stripped)
9713                         return flags;
9714
9715                 stripped |= BTRFS_BLOCK_GROUP_DUP;
9716                 stripped = flags & ~stripped;
9717
9718                 /* switch duplicated blocks with raid1 */
9719                 if (flags & BTRFS_BLOCK_GROUP_DUP)
9720                         return stripped | BTRFS_BLOCK_GROUP_RAID1;
9721
9722                 /* this is drive concat, leave it alone */
9723         }
9724
9725         return flags;
9726 }
9727
9728 static int inc_block_group_ro(struct btrfs_block_group_cache *cache, int force)
9729 {
9730         struct btrfs_space_info *sinfo = cache->space_info;
9731         u64 num_bytes;
9732         u64 sinfo_used;
9733         u64 min_allocable_bytes;
9734         int ret = -ENOSPC;
9735
9736         /*
9737          * We need some metadata space and system metadata space for
9738          * allocating chunks in some corner cases until we force to set
9739          * it to be readonly.
9740          */
9741         if ((sinfo->flags &
9742              (BTRFS_BLOCK_GROUP_SYSTEM | BTRFS_BLOCK_GROUP_METADATA)) &&
9743             !force)
9744                 min_allocable_bytes = SZ_1M;
9745         else
9746                 min_allocable_bytes = 0;
9747
9748         spin_lock(&sinfo->lock);
9749         spin_lock(&cache->lock);
9750
9751         if (cache->ro) {
9752                 cache->ro++;
9753                 ret = 0;
9754                 goto out;
9755         }
9756
9757         num_bytes = cache->key.offset - cache->reserved - cache->pinned -
9758                     cache->bytes_super - btrfs_block_group_used(&cache->item);
9759         sinfo_used = btrfs_space_info_used(sinfo, true);
9760
9761         if (sinfo_used + num_bytes + min_allocable_bytes <=
9762             sinfo->total_bytes) {
9763                 sinfo->bytes_readonly += num_bytes;
9764                 cache->ro++;
9765                 list_add_tail(&cache->ro_list, &sinfo->ro_bgs);
9766                 ret = 0;
9767         }
9768 out:
9769         spin_unlock(&cache->lock);
9770         spin_unlock(&sinfo->lock);
9771         if (ret == -ENOSPC && btrfs_test_opt(cache->fs_info, ENOSPC_DEBUG)) {
9772                 btrfs_info(cache->fs_info,
9773                         "unable to make block group %llu ro",
9774                         cache->key.objectid);
9775                 btrfs_info(cache->fs_info,
9776                         "sinfo_used=%llu bg_num_bytes=%llu min_allocable=%llu",
9777                         sinfo_used, num_bytes, min_allocable_bytes);
9778                 dump_space_info(cache->fs_info, cache->space_info, 0, 0);
9779         }
9780         return ret;
9781 }
9782
9783 int btrfs_inc_block_group_ro(struct btrfs_block_group_cache *cache)
9784
9785 {
9786         struct btrfs_fs_info *fs_info = cache->fs_info;
9787         struct btrfs_trans_handle *trans;
9788         u64 alloc_flags;
9789         int ret;
9790
9791 again:
9792         trans = btrfs_join_transaction(fs_info->extent_root);
9793         if (IS_ERR(trans))
9794                 return PTR_ERR(trans);
9795
9796         /*
9797          * we're not allowed to set block groups readonly after the dirty
9798          * block groups cache has started writing.  If it already started,
9799          * back off and let this transaction commit
9800          */
9801         mutex_lock(&fs_info->ro_block_group_mutex);
9802         if (test_bit(BTRFS_TRANS_DIRTY_BG_RUN, &trans->transaction->flags)) {
9803                 u64 transid = trans->transid;
9804
9805                 mutex_unlock(&fs_info->ro_block_group_mutex);
9806                 btrfs_end_transaction(trans);
9807
9808                 ret = btrfs_wait_for_commit(fs_info, transid);
9809                 if (ret)
9810                         return ret;
9811                 goto again;
9812         }
9813
9814         /*
9815          * if we are changing raid levels, try to allocate a corresponding
9816          * block group with the new raid level.
9817          */
9818         alloc_flags = update_block_group_flags(fs_info, cache->flags);
9819         if (alloc_flags != cache->flags) {
9820                 ret = do_chunk_alloc(trans, alloc_flags,
9821                                      CHUNK_ALLOC_FORCE);
9822                 /*
9823                  * ENOSPC is allowed here, we may have enough space
9824                  * already allocated at the new raid level to
9825                  * carry on
9826                  */
9827                 if (ret == -ENOSPC)
9828                         ret = 0;
9829                 if (ret < 0)
9830                         goto out;
9831         }
9832
9833         ret = inc_block_group_ro(cache, 0);
9834         if (!ret)
9835                 goto out;
9836         alloc_flags = get_alloc_profile(fs_info, cache->space_info->flags);
9837         ret = do_chunk_alloc(trans, alloc_flags, CHUNK_ALLOC_FORCE);
9838         if (ret < 0)
9839                 goto out;
9840         ret = inc_block_group_ro(cache, 0);
9841 out:
9842         if (cache->flags & BTRFS_BLOCK_GROUP_SYSTEM) {
9843                 alloc_flags = update_block_group_flags(fs_info, cache->flags);
9844                 mutex_lock(&fs_info->chunk_mutex);
9845                 check_system_chunk(trans, alloc_flags);
9846                 mutex_unlock(&fs_info->chunk_mutex);
9847         }
9848         mutex_unlock(&fs_info->ro_block_group_mutex);
9849
9850         btrfs_end_transaction(trans);
9851         return ret;
9852 }
9853
9854 int btrfs_force_chunk_alloc(struct btrfs_trans_handle *trans, u64 type)
9855 {
9856         u64 alloc_flags = get_alloc_profile(trans->fs_info, type);
9857
9858         return do_chunk_alloc(trans, alloc_flags, CHUNK_ALLOC_FORCE);
9859 }
9860
9861 /*
9862  * helper to account the unused space of all the readonly block group in the
9863  * space_info. takes mirrors into account.
9864  */
9865 u64 btrfs_account_ro_block_groups_free_space(struct btrfs_space_info *sinfo)
9866 {
9867         struct btrfs_block_group_cache *block_group;
9868         u64 free_bytes = 0;
9869         int factor;
9870
9871         /* It's df, we don't care if it's racy */
9872         if (list_empty(&sinfo->ro_bgs))
9873                 return 0;
9874
9875         spin_lock(&sinfo->lock);
9876         list_for_each_entry(block_group, &sinfo->ro_bgs, ro_list) {
9877                 spin_lock(&block_group->lock);
9878
9879                 if (!block_group->ro) {
9880                         spin_unlock(&block_group->lock);
9881                         continue;
9882                 }
9883
9884                 factor = btrfs_bg_type_to_factor(block_group->flags);
9885                 free_bytes += (block_group->key.offset -
9886                                btrfs_block_group_used(&block_group->item)) *
9887                                factor;
9888
9889                 spin_unlock(&block_group->lock);
9890         }
9891         spin_unlock(&sinfo->lock);
9892
9893         return free_bytes;
9894 }
9895
9896 void btrfs_dec_block_group_ro(struct btrfs_block_group_cache *cache)
9897 {
9898         struct btrfs_space_info *sinfo = cache->space_info;
9899         u64 num_bytes;
9900
9901         BUG_ON(!cache->ro);
9902
9903         spin_lock(&sinfo->lock);
9904         spin_lock(&cache->lock);
9905         if (!--cache->ro) {
9906                 num_bytes = cache->key.offset - cache->reserved -
9907                             cache->pinned - cache->bytes_super -
9908                             btrfs_block_group_used(&cache->item);
9909                 sinfo->bytes_readonly -= num_bytes;
9910                 list_del_init(&cache->ro_list);
9911         }
9912         spin_unlock(&cache->lock);
9913         spin_unlock(&sinfo->lock);
9914 }
9915
9916 /*
9917  * Checks to see if it's even possible to relocate this block group.
9918  *
9919  * @return - -1 if it's not a good idea to relocate this block group, 0 if its
9920  * ok to go ahead and try.
9921  */
9922 int btrfs_can_relocate(struct btrfs_fs_info *fs_info, u64 bytenr)
9923 {
9924         struct btrfs_root *root = fs_info->extent_root;
9925         struct btrfs_block_group_cache *block_group;
9926         struct btrfs_space_info *space_info;
9927         struct btrfs_fs_devices *fs_devices = fs_info->fs_devices;
9928         struct btrfs_device *device;
9929         struct btrfs_trans_handle *trans;
9930         u64 min_free;
9931         u64 dev_min = 1;
9932         u64 dev_nr = 0;
9933         u64 target;
9934         int debug;
9935         int index;
9936         int full = 0;
9937         int ret = 0;
9938
9939         debug = btrfs_test_opt(fs_info, ENOSPC_DEBUG);
9940
9941         block_group = btrfs_lookup_block_group(fs_info, bytenr);
9942
9943         /* odd, couldn't find the block group, leave it alone */
9944         if (!block_group) {
9945                 if (debug)
9946                         btrfs_warn(fs_info,
9947                                    "can't find block group for bytenr %llu",
9948                                    bytenr);
9949                 return -1;
9950         }
9951
9952         min_free = btrfs_block_group_used(&block_group->item);
9953
9954         /* no bytes used, we're good */
9955         if (!min_free)
9956                 goto out;
9957
9958         space_info = block_group->space_info;
9959         spin_lock(&space_info->lock);
9960
9961         full = space_info->full;
9962
9963         /*
9964          * if this is the last block group we have in this space, we can't
9965          * relocate it unless we're able to allocate a new chunk below.
9966          *
9967          * Otherwise, we need to make sure we have room in the space to handle
9968          * all of the extents from this block group.  If we can, we're good
9969          */
9970         if ((space_info->total_bytes != block_group->key.offset) &&
9971             (btrfs_space_info_used(space_info, false) + min_free <
9972              space_info->total_bytes)) {
9973                 spin_unlock(&space_info->lock);
9974                 goto out;
9975         }
9976         spin_unlock(&space_info->lock);
9977
9978         /*
9979          * ok we don't have enough space, but maybe we have free space on our
9980          * devices to allocate new chunks for relocation, so loop through our
9981          * alloc devices and guess if we have enough space.  if this block
9982          * group is going to be restriped, run checks against the target
9983          * profile instead of the current one.
9984          */
9985         ret = -1;
9986
9987         /*
9988          * index:
9989          *      0: raid10
9990          *      1: raid1
9991          *      2: dup
9992          *      3: raid0
9993          *      4: single
9994          */
9995         target = get_restripe_target(fs_info, block_group->flags);
9996         if (target) {
9997                 index = btrfs_bg_flags_to_raid_index(extended_to_chunk(target));
9998         } else {
9999                 /*
10000                  * this is just a balance, so if we were marked as full
10001                  * we know there is no space for a new chunk
10002                  */
10003                 if (full) {
10004                         if (debug)
10005                                 btrfs_warn(fs_info,
10006                                            "no space to alloc new chunk for block group %llu",
10007                                            block_group->key.objectid);
10008                         goto out;
10009                 }
10010
10011                 index = btrfs_bg_flags_to_raid_index(block_group->flags);
10012         }
10013
10014         if (index == BTRFS_RAID_RAID10) {
10015                 dev_min = 4;
10016                 /* Divide by 2 */
10017                 min_free >>= 1;
10018         } else if (index == BTRFS_RAID_RAID1) {
10019                 dev_min = 2;
10020         } else if (index == BTRFS_RAID_DUP) {
10021                 /* Multiply by 2 */
10022                 min_free <<= 1;
10023         } else if (index == BTRFS_RAID_RAID0) {
10024                 dev_min = fs_devices->rw_devices;
10025                 min_free = div64_u64(min_free, dev_min);
10026         }
10027
10028         /* We need to do this so that we can look at pending chunks */
10029         trans = btrfs_join_transaction(root);
10030         if (IS_ERR(trans)) {
10031                 ret = PTR_ERR(trans);
10032                 goto out;
10033         }
10034
10035         mutex_lock(&fs_info->chunk_mutex);
10036         list_for_each_entry(device, &fs_devices->alloc_list, dev_alloc_list) {
10037                 u64 dev_offset;
10038
10039                 /*
10040                  * check to make sure we can actually find a chunk with enough
10041                  * space to fit our block group in.
10042                  */
10043                 if (device->total_bytes > device->bytes_used + min_free &&
10044                     !test_bit(BTRFS_DEV_STATE_REPLACE_TGT, &device->dev_state)) {
10045                         ret = find_free_dev_extent(trans, device, min_free,
10046                                                    &dev_offset, NULL);
10047                         if (!ret)
10048                                 dev_nr++;
10049
10050                         if (dev_nr >= dev_min)
10051                                 break;
10052
10053                         ret = -1;
10054                 }
10055         }
10056         if (debug && ret == -1)
10057                 btrfs_warn(fs_info,
10058                            "no space to allocate a new chunk for block group %llu",
10059                            block_group->key.objectid);
10060         mutex_unlock(&fs_info->chunk_mutex);
10061         btrfs_end_transaction(trans);
10062 out:
10063         btrfs_put_block_group(block_group);
10064         return ret;
10065 }
10066
10067 static int find_first_block_group(struct btrfs_fs_info *fs_info,
10068                                   struct btrfs_path *path,
10069                                   struct btrfs_key *key)
10070 {
10071         struct btrfs_root *root = fs_info->extent_root;
10072         int ret = 0;
10073         struct btrfs_key found_key;
10074         struct extent_buffer *leaf;
10075         struct btrfs_block_group_item bg;
10076         u64 flags;
10077         int slot;
10078
10079         ret = btrfs_search_slot(NULL, root, key, path, 0, 0);
10080         if (ret < 0)
10081                 goto out;
10082
10083         while (1) {
10084                 slot = path->slots[0];
10085                 leaf = path->nodes[0];
10086                 if (slot >= btrfs_header_nritems(leaf)) {
10087                         ret = btrfs_next_leaf(root, path);
10088                         if (ret == 0)
10089                                 continue;
10090                         if (ret < 0)
10091                                 goto out;
10092                         break;
10093                 }
10094                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
10095
10096                 if (found_key.objectid >= key->objectid &&
10097                     found_key.type == BTRFS_BLOCK_GROUP_ITEM_KEY) {
10098                         struct extent_map_tree *em_tree;
10099                         struct extent_map *em;
10100
10101                         em_tree = &root->fs_info->mapping_tree.map_tree;
10102                         read_lock(&em_tree->lock);
10103                         em = lookup_extent_mapping(em_tree, found_key.objectid,
10104                                                    found_key.offset);
10105                         read_unlock(&em_tree->lock);
10106                         if (!em) {
10107                                 btrfs_err(fs_info,
10108                         "logical %llu len %llu found bg but no related chunk",
10109                                           found_key.objectid, found_key.offset);
10110                                 ret = -ENOENT;
10111                         } else if (em->start != found_key.objectid ||
10112                                    em->len != found_key.offset) {
10113                                 btrfs_err(fs_info,
10114                 "block group %llu len %llu mismatch with chunk %llu len %llu",
10115                                           found_key.objectid, found_key.offset,
10116                                           em->start, em->len);
10117                                 ret = -EUCLEAN;
10118                         } else {
10119                                 read_extent_buffer(leaf, &bg,
10120                                         btrfs_item_ptr_offset(leaf, slot),
10121                                         sizeof(bg));
10122                                 flags = btrfs_block_group_flags(&bg) &
10123                                         BTRFS_BLOCK_GROUP_TYPE_MASK;
10124
10125                                 if (flags != (em->map_lookup->type &
10126                                               BTRFS_BLOCK_GROUP_TYPE_MASK)) {
10127                                         btrfs_err(fs_info,
10128 "block group %llu len %llu type flags 0x%llx mismatch with chunk type flags 0x%llx",
10129                                                 found_key.objectid,
10130                                                 found_key.offset, flags,
10131                                                 (BTRFS_BLOCK_GROUP_TYPE_MASK &
10132                                                  em->map_lookup->type));
10133                                         ret = -EUCLEAN;
10134                                 } else {
10135                                         ret = 0;
10136                                 }
10137                         }
10138                         free_extent_map(em);
10139                         goto out;
10140                 }
10141                 path->slots[0]++;
10142         }
10143 out:
10144         return ret;
10145 }
10146
10147 void btrfs_put_block_group_cache(struct btrfs_fs_info *info)
10148 {
10149         struct btrfs_block_group_cache *block_group;
10150         u64 last = 0;
10151
10152         while (1) {
10153                 struct inode *inode;
10154
10155                 block_group = btrfs_lookup_first_block_group(info, last);
10156                 while (block_group) {
10157                         wait_block_group_cache_done(block_group);
10158                         spin_lock(&block_group->lock);
10159                         if (block_group->iref)
10160                                 break;
10161                         spin_unlock(&block_group->lock);
10162                         block_group = next_block_group(info, block_group);
10163                 }
10164                 if (!block_group) {
10165                         if (last == 0)
10166                                 break;
10167                         last = 0;
10168                         continue;
10169                 }
10170
10171                 inode = block_group->inode;
10172                 block_group->iref = 0;
10173                 block_group->inode = NULL;
10174                 spin_unlock(&block_group->lock);
10175                 ASSERT(block_group->io_ctl.inode == NULL);
10176                 iput(inode);
10177                 last = block_group->key.objectid + block_group->key.offset;
10178                 btrfs_put_block_group(block_group);
10179         }
10180 }
10181
10182 /*
10183  * Must be called only after stopping all workers, since we could have block
10184  * group caching kthreads running, and therefore they could race with us if we
10185  * freed the block groups before stopping them.
10186  */
10187 int btrfs_free_block_groups(struct btrfs_fs_info *info)
10188 {
10189         struct btrfs_block_group_cache *block_group;
10190         struct btrfs_space_info *space_info;
10191         struct btrfs_caching_control *caching_ctl;
10192         struct rb_node *n;
10193
10194         down_write(&info->commit_root_sem);
10195         while (!list_empty(&info->caching_block_groups)) {
10196                 caching_ctl = list_entry(info->caching_block_groups.next,
10197                                          struct btrfs_caching_control, list);
10198                 list_del(&caching_ctl->list);
10199                 put_caching_control(caching_ctl);
10200         }
10201         up_write(&info->commit_root_sem);
10202
10203         spin_lock(&info->unused_bgs_lock);
10204         while (!list_empty(&info->unused_bgs)) {
10205                 block_group = list_first_entry(&info->unused_bgs,
10206                                                struct btrfs_block_group_cache,
10207                                                bg_list);
10208                 list_del_init(&block_group->bg_list);
10209                 btrfs_put_block_group(block_group);
10210         }
10211         spin_unlock(&info->unused_bgs_lock);
10212
10213         spin_lock(&info->block_group_cache_lock);
10214         while ((n = rb_last(&info->block_group_cache_tree)) != NULL) {
10215                 block_group = rb_entry(n, struct btrfs_block_group_cache,
10216                                        cache_node);
10217                 rb_erase(&block_group->cache_node,
10218                          &info->block_group_cache_tree);
10219                 RB_CLEAR_NODE(&block_group->cache_node);
10220                 spin_unlock(&info->block_group_cache_lock);
10221
10222                 down_write(&block_group->space_info->groups_sem);
10223                 list_del(&block_group->list);
10224                 up_write(&block_group->space_info->groups_sem);
10225
10226                 /*
10227                  * We haven't cached this block group, which means we could
10228                  * possibly have excluded extents on this block group.
10229                  */
10230                 if (block_group->cached == BTRFS_CACHE_NO ||
10231                     block_group->cached == BTRFS_CACHE_ERROR)
10232                         free_excluded_extents(block_group);
10233
10234                 btrfs_remove_free_space_cache(block_group);
10235                 ASSERT(block_group->cached != BTRFS_CACHE_STARTED);
10236                 ASSERT(list_empty(&block_group->dirty_list));
10237                 ASSERT(list_empty(&block_group->io_list));
10238                 ASSERT(list_empty(&block_group->bg_list));
10239                 ASSERT(atomic_read(&block_group->count) == 1);
10240                 btrfs_put_block_group(block_group);
10241
10242                 spin_lock(&info->block_group_cache_lock);
10243         }
10244         spin_unlock(&info->block_group_cache_lock);
10245
10246         /* now that all the block groups are freed, go through and
10247          * free all the space_info structs.  This is only called during
10248          * the final stages of unmount, and so we know nobody is
10249          * using them.  We call synchronize_rcu() once before we start,
10250          * just to be on the safe side.
10251          */
10252         synchronize_rcu();
10253
10254         release_global_block_rsv(info);
10255
10256         while (!list_empty(&info->space_info)) {
10257                 int i;
10258
10259                 space_info = list_entry(info->space_info.next,
10260                                         struct btrfs_space_info,
10261                                         list);
10262
10263                 /*
10264                  * Do not hide this behind enospc_debug, this is actually
10265                  * important and indicates a real bug if this happens.
10266                  */
10267                 if (WARN_ON(space_info->bytes_pinned > 0 ||
10268                             space_info->bytes_reserved > 0 ||
10269                             space_info->bytes_may_use > 0))
10270                         dump_space_info(info, space_info, 0, 0);
10271                 list_del(&space_info->list);
10272                 for (i = 0; i < BTRFS_NR_RAID_TYPES; i++) {
10273                         struct kobject *kobj;
10274                         kobj = space_info->block_group_kobjs[i];
10275                         space_info->block_group_kobjs[i] = NULL;
10276                         if (kobj) {
10277                                 kobject_del(kobj);
10278                                 kobject_put(kobj);
10279                         }
10280                 }
10281                 kobject_del(&space_info->kobj);
10282                 kobject_put(&space_info->kobj);
10283         }
10284         return 0;
10285 }
10286
10287 /* link_block_group will queue up kobjects to add when we're reclaim-safe */
10288 void btrfs_add_raid_kobjects(struct btrfs_fs_info *fs_info)
10289 {
10290         struct btrfs_space_info *space_info;
10291         struct raid_kobject *rkobj;
10292         LIST_HEAD(list);
10293         int index;
10294         int ret = 0;
10295
10296         spin_lock(&fs_info->pending_raid_kobjs_lock);
10297         list_splice_init(&fs_info->pending_raid_kobjs, &list);
10298         spin_unlock(&fs_info->pending_raid_kobjs_lock);
10299
10300         list_for_each_entry(rkobj, &list, list) {
10301                 space_info = __find_space_info(fs_info, rkobj->flags);
10302                 index = btrfs_bg_flags_to_raid_index(rkobj->flags);
10303
10304                 ret = kobject_add(&rkobj->kobj, &space_info->kobj,
10305                                   "%s", get_raid_name(index));
10306                 if (ret) {
10307                         kobject_put(&rkobj->kobj);
10308                         break;
10309                 }
10310         }
10311         if (ret)
10312                 btrfs_warn(fs_info,
10313                            "failed to add kobject for block cache, ignoring");
10314 }
10315
10316 static void link_block_group(struct btrfs_block_group_cache *cache)
10317 {
10318         struct btrfs_space_info *space_info = cache->space_info;
10319         struct btrfs_fs_info *fs_info = cache->fs_info;
10320         int index = btrfs_bg_flags_to_raid_index(cache->flags);
10321         bool first = false;
10322
10323         down_write(&space_info->groups_sem);
10324         if (list_empty(&space_info->block_groups[index]))
10325                 first = true;
10326         list_add_tail(&cache->list, &space_info->block_groups[index]);
10327         up_write(&space_info->groups_sem);
10328
10329         if (first) {
10330                 struct raid_kobject *rkobj = kzalloc(sizeof(*rkobj), GFP_NOFS);
10331                 if (!rkobj) {
10332                         btrfs_warn(cache->fs_info,
10333                                 "couldn't alloc memory for raid level kobject");
10334                         return;
10335                 }
10336                 rkobj->flags = cache->flags;
10337                 kobject_init(&rkobj->kobj, &btrfs_raid_ktype);
10338
10339                 spin_lock(&fs_info->pending_raid_kobjs_lock);
10340                 list_add_tail(&rkobj->list, &fs_info->pending_raid_kobjs);
10341                 spin_unlock(&fs_info->pending_raid_kobjs_lock);
10342                 space_info->block_group_kobjs[index] = &rkobj->kobj;
10343         }
10344 }
10345
10346 static struct btrfs_block_group_cache *
10347 btrfs_create_block_group_cache(struct btrfs_fs_info *fs_info,
10348                                u64 start, u64 size)
10349 {
10350         struct btrfs_block_group_cache *cache;
10351
10352         cache = kzalloc(sizeof(*cache), GFP_NOFS);
10353         if (!cache)
10354                 return NULL;
10355
10356         cache->free_space_ctl = kzalloc(sizeof(*cache->free_space_ctl),
10357                                         GFP_NOFS);
10358         if (!cache->free_space_ctl) {
10359                 kfree(cache);
10360                 return NULL;
10361         }
10362
10363         cache->key.objectid = start;
10364         cache->key.offset = size;
10365         cache->key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
10366
10367         cache->fs_info = fs_info;
10368         cache->full_stripe_len = btrfs_full_stripe_len(fs_info, start);
10369         set_free_space_tree_thresholds(cache);
10370
10371         atomic_set(&cache->count, 1);
10372         spin_lock_init(&cache->lock);
10373         init_rwsem(&cache->data_rwsem);
10374         INIT_LIST_HEAD(&cache->list);
10375         INIT_LIST_HEAD(&cache->cluster_list);
10376         INIT_LIST_HEAD(&cache->bg_list);
10377         INIT_LIST_HEAD(&cache->ro_list);
10378         INIT_LIST_HEAD(&cache->dirty_list);
10379         INIT_LIST_HEAD(&cache->io_list);
10380         btrfs_init_free_space_ctl(cache);
10381         atomic_set(&cache->trimming, 0);
10382         mutex_init(&cache->free_space_lock);
10383         btrfs_init_full_stripe_locks_tree(&cache->full_stripe_locks_root);
10384
10385         return cache;
10386 }
10387
10388
10389 /*
10390  * Iterate all chunks and verify that each of them has the corresponding block
10391  * group
10392  */
10393 static int check_chunk_block_group_mappings(struct btrfs_fs_info *fs_info)
10394 {
10395         struct btrfs_mapping_tree *map_tree = &fs_info->mapping_tree;
10396         struct extent_map *em;
10397         struct btrfs_block_group_cache *bg;
10398         u64 start = 0;
10399         int ret = 0;
10400
10401         while (1) {
10402                 read_lock(&map_tree->map_tree.lock);
10403                 /*
10404                  * lookup_extent_mapping will return the first extent map
10405                  * intersecting the range, so setting @len to 1 is enough to
10406                  * get the first chunk.
10407                  */
10408                 em = lookup_extent_mapping(&map_tree->map_tree, start, 1);
10409                 read_unlock(&map_tree->map_tree.lock);
10410                 if (!em)
10411                         break;
10412
10413                 bg = btrfs_lookup_block_group(fs_info, em->start);
10414                 if (!bg) {
10415                         btrfs_err(fs_info,
10416         "chunk start=%llu len=%llu doesn't have corresponding block group",
10417                                      em->start, em->len);
10418                         ret = -EUCLEAN;
10419                         free_extent_map(em);
10420                         break;
10421                 }
10422                 if (bg->key.objectid != em->start ||
10423                     bg->key.offset != em->len ||
10424                     (bg->flags & BTRFS_BLOCK_GROUP_TYPE_MASK) !=
10425                     (em->map_lookup->type & BTRFS_BLOCK_GROUP_TYPE_MASK)) {
10426                         btrfs_err(fs_info,
10427 "chunk start=%llu len=%llu flags=0x%llx doesn't match block group start=%llu len=%llu flags=0x%llx",
10428                                 em->start, em->len,
10429                                 em->map_lookup->type & BTRFS_BLOCK_GROUP_TYPE_MASK,
10430                                 bg->key.objectid, bg->key.offset,
10431                                 bg->flags & BTRFS_BLOCK_GROUP_TYPE_MASK);
10432                         ret = -EUCLEAN;
10433                         free_extent_map(em);
10434                         btrfs_put_block_group(bg);
10435                         break;
10436                 }
10437                 start = em->start + em->len;
10438                 free_extent_map(em);
10439                 btrfs_put_block_group(bg);
10440         }
10441         return ret;
10442 }
10443
10444 int btrfs_read_block_groups(struct btrfs_fs_info *info)
10445 {
10446         struct btrfs_path *path;
10447         int ret;
10448         struct btrfs_block_group_cache *cache;
10449         struct btrfs_space_info *space_info;
10450         struct btrfs_key key;
10451         struct btrfs_key found_key;
10452         struct extent_buffer *leaf;
10453         int need_clear = 0;
10454         u64 cache_gen;
10455         u64 feature;
10456         int mixed;
10457
10458         feature = btrfs_super_incompat_flags(info->super_copy);
10459         mixed = !!(feature & BTRFS_FEATURE_INCOMPAT_MIXED_GROUPS);
10460
10461         key.objectid = 0;
10462         key.offset = 0;
10463         key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
10464         path = btrfs_alloc_path();
10465         if (!path)
10466                 return -ENOMEM;
10467         path->reada = READA_FORWARD;
10468
10469         cache_gen = btrfs_super_cache_generation(info->super_copy);
10470         if (btrfs_test_opt(info, SPACE_CACHE) &&
10471             btrfs_super_generation(info->super_copy) != cache_gen)
10472                 need_clear = 1;
10473         if (btrfs_test_opt(info, CLEAR_CACHE))
10474                 need_clear = 1;
10475
10476         while (1) {
10477                 ret = find_first_block_group(info, path, &key);
10478                 if (ret > 0)
10479                         break;
10480                 if (ret != 0)
10481                         goto error;
10482
10483                 leaf = path->nodes[0];
10484                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
10485
10486                 cache = btrfs_create_block_group_cache(info, found_key.objectid,
10487                                                        found_key.offset);
10488                 if (!cache) {
10489                         ret = -ENOMEM;
10490                         goto error;
10491                 }
10492
10493                 if (need_clear) {
10494                         /*
10495                          * When we mount with old space cache, we need to
10496                          * set BTRFS_DC_CLEAR and set dirty flag.
10497                          *
10498                          * a) Setting 'BTRFS_DC_CLEAR' makes sure that we
10499                          *    truncate the old free space cache inode and
10500                          *    setup a new one.
10501                          * b) Setting 'dirty flag' makes sure that we flush
10502                          *    the new space cache info onto disk.
10503                          */
10504                         if (btrfs_test_opt(info, SPACE_CACHE))
10505                                 cache->disk_cache_state = BTRFS_DC_CLEAR;
10506                 }
10507
10508                 read_extent_buffer(leaf, &cache->item,
10509                                    btrfs_item_ptr_offset(leaf, path->slots[0]),
10510                                    sizeof(cache->item));
10511                 cache->flags = btrfs_block_group_flags(&cache->item);
10512                 if (!mixed &&
10513                     ((cache->flags & BTRFS_BLOCK_GROUP_METADATA) &&
10514                     (cache->flags & BTRFS_BLOCK_GROUP_DATA))) {
10515                         btrfs_err(info,
10516 "bg %llu is a mixed block group but filesystem hasn't enabled mixed block groups",
10517                                   cache->key.objectid);
10518                         ret = -EINVAL;
10519                         goto error;
10520                 }
10521
10522                 key.objectid = found_key.objectid + found_key.offset;
10523                 btrfs_release_path(path);
10524
10525                 /*
10526                  * We need to exclude the super stripes now so that the space
10527                  * info has super bytes accounted for, otherwise we'll think
10528                  * we have more space than we actually do.
10529                  */
10530                 ret = exclude_super_stripes(cache);
10531                 if (ret) {
10532                         /*
10533                          * We may have excluded something, so call this just in
10534                          * case.
10535                          */
10536                         free_excluded_extents(cache);
10537                         btrfs_put_block_group(cache);
10538                         goto error;
10539                 }
10540
10541                 /*
10542                  * check for two cases, either we are full, and therefore
10543                  * don't need to bother with the caching work since we won't
10544                  * find any space, or we are empty, and we can just add all
10545                  * the space in and be done with it.  This saves us _a_lot_ of
10546                  * time, particularly in the full case.
10547                  */
10548                 if (found_key.offset == btrfs_block_group_used(&cache->item)) {
10549                         cache->last_byte_to_unpin = (u64)-1;
10550                         cache->cached = BTRFS_CACHE_FINISHED;
10551                         free_excluded_extents(cache);
10552                 } else if (btrfs_block_group_used(&cache->item) == 0) {
10553                         cache->last_byte_to_unpin = (u64)-1;
10554                         cache->cached = BTRFS_CACHE_FINISHED;
10555                         add_new_free_space(cache, found_key.objectid,
10556                                            found_key.objectid +
10557                                            found_key.offset);
10558                         free_excluded_extents(cache);
10559                 }
10560
10561                 ret = btrfs_add_block_group_cache(info, cache);
10562                 if (ret) {
10563                         btrfs_remove_free_space_cache(cache);
10564                         btrfs_put_block_group(cache);
10565                         goto error;
10566                 }
10567
10568                 trace_btrfs_add_block_group(info, cache, 0);
10569                 update_space_info(info, cache->flags, found_key.offset,
10570                                   btrfs_block_group_used(&cache->item),
10571                                   cache->bytes_super, &space_info);
10572
10573                 cache->space_info = space_info;
10574
10575                 link_block_group(cache);
10576
10577                 set_avail_alloc_bits(info, cache->flags);
10578                 if (btrfs_chunk_readonly(info, cache->key.objectid)) {
10579                         inc_block_group_ro(cache, 1);
10580                 } else if (btrfs_block_group_used(&cache->item) == 0) {
10581                         ASSERT(list_empty(&cache->bg_list));
10582                         btrfs_mark_bg_unused(cache);
10583                 }
10584         }
10585
10586         list_for_each_entry_rcu(space_info, &info->space_info, list) {
10587                 if (!(get_alloc_profile(info, space_info->flags) &
10588                       (BTRFS_BLOCK_GROUP_RAID10 |
10589                        BTRFS_BLOCK_GROUP_RAID1 |
10590                        BTRFS_BLOCK_GROUP_RAID5 |
10591                        BTRFS_BLOCK_GROUP_RAID6 |
10592                        BTRFS_BLOCK_GROUP_DUP)))
10593                         continue;
10594                 /*
10595                  * avoid allocating from un-mirrored block group if there are
10596                  * mirrored block groups.
10597                  */
10598                 list_for_each_entry(cache,
10599                                 &space_info->block_groups[BTRFS_RAID_RAID0],
10600                                 list)
10601                         inc_block_group_ro(cache, 1);
10602                 list_for_each_entry(cache,
10603                                 &space_info->block_groups[BTRFS_RAID_SINGLE],
10604                                 list)
10605                         inc_block_group_ro(cache, 1);
10606         }
10607
10608         btrfs_add_raid_kobjects(info);
10609         init_global_block_rsv(info);
10610         ret = check_chunk_block_group_mappings(info);
10611 error:
10612         btrfs_free_path(path);
10613         return ret;
10614 }
10615
10616 void btrfs_create_pending_block_groups(struct btrfs_trans_handle *trans)
10617 {
10618         struct btrfs_fs_info *fs_info = trans->fs_info;
10619         struct btrfs_block_group_cache *block_group;
10620         struct btrfs_root *extent_root = fs_info->extent_root;
10621         struct btrfs_block_group_item item;
10622         struct btrfs_key key;
10623         int ret = 0;
10624
10625         if (!trans->can_flush_pending_bgs)
10626                 return;
10627
10628         while (!list_empty(&trans->new_bgs)) {
10629                 block_group = list_first_entry(&trans->new_bgs,
10630                                                struct btrfs_block_group_cache,
10631                                                bg_list);
10632                 if (ret)
10633                         goto next;
10634
10635                 spin_lock(&block_group->lock);
10636                 memcpy(&item, &block_group->item, sizeof(item));
10637                 memcpy(&key, &block_group->key, sizeof(key));
10638                 spin_unlock(&block_group->lock);
10639
10640                 ret = btrfs_insert_item(trans, extent_root, &key, &item,
10641                                         sizeof(item));
10642                 if (ret)
10643                         btrfs_abort_transaction(trans, ret);
10644                 ret = btrfs_finish_chunk_alloc(trans, key.objectid, key.offset);
10645                 if (ret)
10646                         btrfs_abort_transaction(trans, ret);
10647                 add_block_group_free_space(trans, block_group);
10648                 /* already aborted the transaction if it failed. */
10649 next:
10650                 btrfs_delayed_refs_rsv_release(fs_info, 1);
10651                 list_del_init(&block_group->bg_list);
10652         }
10653         btrfs_trans_release_chunk_metadata(trans);
10654 }
10655
10656 int btrfs_make_block_group(struct btrfs_trans_handle *trans, u64 bytes_used,
10657                            u64 type, u64 chunk_offset, u64 size)
10658 {
10659         struct btrfs_fs_info *fs_info = trans->fs_info;
10660         struct btrfs_block_group_cache *cache;
10661         int ret;
10662
10663         btrfs_set_log_full_commit(fs_info, trans);
10664
10665         cache = btrfs_create_block_group_cache(fs_info, chunk_offset, size);
10666         if (!cache)
10667                 return -ENOMEM;
10668
10669         btrfs_set_block_group_used(&cache->item, bytes_used);
10670         btrfs_set_block_group_chunk_objectid(&cache->item,
10671                                              BTRFS_FIRST_CHUNK_TREE_OBJECTID);
10672         btrfs_set_block_group_flags(&cache->item, type);
10673
10674         cache->flags = type;
10675         cache->last_byte_to_unpin = (u64)-1;
10676         cache->cached = BTRFS_CACHE_FINISHED;
10677         cache->needs_free_space = 1;
10678         ret = exclude_super_stripes(cache);
10679         if (ret) {
10680                 /*
10681                  * We may have excluded something, so call this just in
10682                  * case.
10683                  */
10684                 free_excluded_extents(cache);
10685                 btrfs_put_block_group(cache);
10686                 return ret;
10687         }
10688
10689         add_new_free_space(cache, chunk_offset, chunk_offset + size);
10690
10691         free_excluded_extents(cache);
10692
10693 #ifdef CONFIG_BTRFS_DEBUG
10694         if (btrfs_should_fragment_free_space(cache)) {
10695                 u64 new_bytes_used = size - bytes_used;
10696
10697                 bytes_used += new_bytes_used >> 1;
10698                 fragment_free_space(cache);
10699         }
10700 #endif
10701         /*
10702          * Ensure the corresponding space_info object is created and
10703          * assigned to our block group. We want our bg to be added to the rbtree
10704          * with its ->space_info set.
10705          */
10706         cache->space_info = __find_space_info(fs_info, cache->flags);
10707         ASSERT(cache->space_info);
10708
10709         ret = btrfs_add_block_group_cache(fs_info, cache);
10710         if (ret) {
10711                 btrfs_remove_free_space_cache(cache);
10712                 btrfs_put_block_group(cache);
10713                 return ret;
10714         }
10715
10716         /*
10717          * Now that our block group has its ->space_info set and is inserted in
10718          * the rbtree, update the space info's counters.
10719          */
10720         trace_btrfs_add_block_group(fs_info, cache, 1);
10721         update_space_info(fs_info, cache->flags, size, bytes_used,
10722                                 cache->bytes_super, &cache->space_info);
10723         update_global_block_rsv(fs_info);
10724
10725         link_block_group(cache);
10726
10727         list_add_tail(&cache->bg_list, &trans->new_bgs);
10728         trans->delayed_ref_updates++;
10729         btrfs_update_delayed_refs_rsv(trans);
10730
10731         set_avail_alloc_bits(fs_info, type);
10732         return 0;
10733 }
10734
10735 static void clear_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
10736 {
10737         u64 extra_flags = chunk_to_extended(flags) &
10738                                 BTRFS_EXTENDED_PROFILE_MASK;
10739
10740         write_seqlock(&fs_info->profiles_lock);
10741         if (flags & BTRFS_BLOCK_GROUP_DATA)
10742                 fs_info->avail_data_alloc_bits &= ~extra_flags;
10743         if (flags & BTRFS_BLOCK_GROUP_METADATA)
10744                 fs_info->avail_metadata_alloc_bits &= ~extra_flags;
10745         if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
10746                 fs_info->avail_system_alloc_bits &= ~extra_flags;
10747         write_sequnlock(&fs_info->profiles_lock);
10748 }
10749
10750 int btrfs_remove_block_group(struct btrfs_trans_handle *trans,
10751                              u64 group_start, struct extent_map *em)
10752 {
10753         struct btrfs_fs_info *fs_info = trans->fs_info;
10754         struct btrfs_root *root = fs_info->extent_root;
10755         struct btrfs_path *path;
10756         struct btrfs_block_group_cache *block_group;
10757         struct btrfs_free_cluster *cluster;
10758         struct btrfs_root *tree_root = fs_info->tree_root;
10759         struct btrfs_key key;
10760         struct inode *inode;
10761         struct kobject *kobj = NULL;
10762         int ret;
10763         int index;
10764         int factor;
10765         struct btrfs_caching_control *caching_ctl = NULL;
10766         bool remove_em;
10767         bool remove_rsv = false;
10768
10769         block_group = btrfs_lookup_block_group(fs_info, group_start);
10770         BUG_ON(!block_group);
10771         BUG_ON(!block_group->ro);
10772
10773         trace_btrfs_remove_block_group(block_group);
10774         /*
10775          * Free the reserved super bytes from this block group before
10776          * remove it.
10777          */
10778         free_excluded_extents(block_group);
10779         btrfs_free_ref_tree_range(fs_info, block_group->key.objectid,
10780                                   block_group->key.offset);
10781
10782         memcpy(&key, &block_group->key, sizeof(key));
10783         index = btrfs_bg_flags_to_raid_index(block_group->flags);
10784         factor = btrfs_bg_type_to_factor(block_group->flags);
10785
10786         /* make sure this block group isn't part of an allocation cluster */
10787         cluster = &fs_info->data_alloc_cluster;
10788         spin_lock(&cluster->refill_lock);
10789         btrfs_return_cluster_to_free_space(block_group, cluster);
10790         spin_unlock(&cluster->refill_lock);
10791
10792         /*
10793          * make sure this block group isn't part of a metadata
10794          * allocation cluster
10795          */
10796         cluster = &fs_info->meta_alloc_cluster;
10797         spin_lock(&cluster->refill_lock);
10798         btrfs_return_cluster_to_free_space(block_group, cluster);
10799         spin_unlock(&cluster->refill_lock);
10800
10801         path = btrfs_alloc_path();
10802         if (!path) {
10803                 ret = -ENOMEM;
10804                 goto out;
10805         }
10806
10807         /*
10808          * get the inode first so any iput calls done for the io_list
10809          * aren't the final iput (no unlinks allowed now)
10810          */
10811         inode = lookup_free_space_inode(fs_info, block_group, path);
10812
10813         mutex_lock(&trans->transaction->cache_write_mutex);
10814         /*
10815          * Make sure our free space cache IO is done before removing the
10816          * free space inode
10817          */
10818         spin_lock(&trans->transaction->dirty_bgs_lock);
10819         if (!list_empty(&block_group->io_list)) {
10820                 list_del_init(&block_group->io_list);
10821
10822                 WARN_ON(!IS_ERR(inode) && inode != block_group->io_ctl.inode);
10823
10824                 spin_unlock(&trans->transaction->dirty_bgs_lock);
10825                 btrfs_wait_cache_io(trans, block_group, path);
10826                 btrfs_put_block_group(block_group);
10827                 spin_lock(&trans->transaction->dirty_bgs_lock);
10828         }
10829
10830         if (!list_empty(&block_group->dirty_list)) {
10831                 list_del_init(&block_group->dirty_list);
10832                 remove_rsv = true;
10833                 btrfs_put_block_group(block_group);
10834         }
10835         spin_unlock(&trans->transaction->dirty_bgs_lock);
10836         mutex_unlock(&trans->transaction->cache_write_mutex);
10837
10838         if (!IS_ERR(inode)) {
10839                 ret = btrfs_orphan_add(trans, BTRFS_I(inode));
10840                 if (ret) {
10841                         btrfs_add_delayed_iput(inode);
10842                         goto out;
10843                 }
10844                 clear_nlink(inode);
10845                 /* One for the block groups ref */
10846                 spin_lock(&block_group->lock);
10847                 if (block_group->iref) {
10848                         block_group->iref = 0;
10849                         block_group->inode = NULL;
10850                         spin_unlock(&block_group->lock);
10851                         iput(inode);
10852                 } else {
10853                         spin_unlock(&block_group->lock);
10854                 }
10855                 /* One for our lookup ref */
10856                 btrfs_add_delayed_iput(inode);
10857         }
10858
10859         key.objectid = BTRFS_FREE_SPACE_OBJECTID;
10860         key.offset = block_group->key.objectid;
10861         key.type = 0;
10862
10863         ret = btrfs_search_slot(trans, tree_root, &key, path, -1, 1);
10864         if (ret < 0)
10865                 goto out;
10866         if (ret > 0)
10867                 btrfs_release_path(path);
10868         if (ret == 0) {
10869                 ret = btrfs_del_item(trans, tree_root, path);
10870                 if (ret)
10871                         goto out;
10872                 btrfs_release_path(path);
10873         }
10874
10875         spin_lock(&fs_info->block_group_cache_lock);
10876         rb_erase(&block_group->cache_node,
10877                  &fs_info->block_group_cache_tree);
10878         RB_CLEAR_NODE(&block_group->cache_node);
10879
10880         if (fs_info->first_logical_byte == block_group->key.objectid)
10881                 fs_info->first_logical_byte = (u64)-1;
10882         spin_unlock(&fs_info->block_group_cache_lock);
10883
10884         down_write(&block_group->space_info->groups_sem);
10885         /*
10886          * we must use list_del_init so people can check to see if they
10887          * are still on the list after taking the semaphore
10888          */
10889         list_del_init(&block_group->list);
10890         if (list_empty(&block_group->space_info->block_groups[index])) {
10891                 kobj = block_group->space_info->block_group_kobjs[index];
10892                 block_group->space_info->block_group_kobjs[index] = NULL;
10893                 clear_avail_alloc_bits(fs_info, block_group->flags);
10894         }
10895         up_write(&block_group->space_info->groups_sem);
10896         if (kobj) {
10897                 kobject_del(kobj);
10898                 kobject_put(kobj);
10899         }
10900
10901         if (block_group->has_caching_ctl)
10902                 caching_ctl = get_caching_control(block_group);
10903         if (block_group->cached == BTRFS_CACHE_STARTED)
10904                 wait_block_group_cache_done(block_group);
10905         if (block_group->has_caching_ctl) {
10906                 down_write(&fs_info->commit_root_sem);
10907                 if (!caching_ctl) {
10908                         struct btrfs_caching_control *ctl;
10909
10910                         list_for_each_entry(ctl,
10911                                     &fs_info->caching_block_groups, list)
10912                                 if (ctl->block_group == block_group) {
10913                                         caching_ctl = ctl;
10914                                         refcount_inc(&caching_ctl->count);
10915                                         break;
10916                                 }
10917                 }
10918                 if (caching_ctl)
10919                         list_del_init(&caching_ctl->list);
10920                 up_write(&fs_info->commit_root_sem);
10921                 if (caching_ctl) {
10922                         /* Once for the caching bgs list and once for us. */
10923                         put_caching_control(caching_ctl);
10924                         put_caching_control(caching_ctl);
10925                 }
10926         }
10927
10928         spin_lock(&trans->transaction->dirty_bgs_lock);
10929         WARN_ON(!list_empty(&block_group->dirty_list));
10930         WARN_ON(!list_empty(&block_group->io_list));
10931         spin_unlock(&trans->transaction->dirty_bgs_lock);
10932
10933         btrfs_remove_free_space_cache(block_group);
10934
10935         spin_lock(&block_group->space_info->lock);
10936         list_del_init(&block_group->ro_list);
10937
10938         if (btrfs_test_opt(fs_info, ENOSPC_DEBUG)) {
10939                 WARN_ON(block_group->space_info->total_bytes
10940                         < block_group->key.offset);
10941                 WARN_ON(block_group->space_info->bytes_readonly
10942                         < block_group->key.offset);
10943                 WARN_ON(block_group->space_info->disk_total
10944                         < block_group->key.offset * factor);
10945         }
10946         block_group->space_info->total_bytes -= block_group->key.offset;
10947         block_group->space_info->bytes_readonly -= block_group->key.offset;
10948         block_group->space_info->disk_total -= block_group->key.offset * factor;
10949
10950         spin_unlock(&block_group->space_info->lock);
10951
10952         memcpy(&key, &block_group->key, sizeof(key));
10953
10954         mutex_lock(&fs_info->chunk_mutex);
10955         if (!list_empty(&em->list)) {
10956                 /* We're in the transaction->pending_chunks list. */
10957                 free_extent_map(em);
10958         }
10959         spin_lock(&block_group->lock);
10960         block_group->removed = 1;
10961         /*
10962          * At this point trimming can't start on this block group, because we
10963          * removed the block group from the tree fs_info->block_group_cache_tree
10964          * so no one can't find it anymore and even if someone already got this
10965          * block group before we removed it from the rbtree, they have already
10966          * incremented block_group->trimming - if they didn't, they won't find
10967          * any free space entries because we already removed them all when we
10968          * called btrfs_remove_free_space_cache().
10969          *
10970          * And we must not remove the extent map from the fs_info->mapping_tree
10971          * to prevent the same logical address range and physical device space
10972          * ranges from being reused for a new block group. This is because our
10973          * fs trim operation (btrfs_trim_fs() / btrfs_ioctl_fitrim()) is
10974          * completely transactionless, so while it is trimming a range the
10975          * currently running transaction might finish and a new one start,
10976          * allowing for new block groups to be created that can reuse the same
10977          * physical device locations unless we take this special care.
10978          *
10979          * There may also be an implicit trim operation if the file system
10980          * is mounted with -odiscard. The same protections must remain
10981          * in place until the extents have been discarded completely when
10982          * the transaction commit has completed.
10983          */
10984         remove_em = (atomic_read(&block_group->trimming) == 0);
10985         /*
10986          * Make sure a trimmer task always sees the em in the pinned_chunks list
10987          * if it sees block_group->removed == 1 (needs to lock block_group->lock
10988          * before checking block_group->removed).
10989          */
10990         if (!remove_em) {
10991                 /*
10992                  * Our em might be in trans->transaction->pending_chunks which
10993                  * is protected by fs_info->chunk_mutex ([lock|unlock]_chunks),
10994                  * and so is the fs_info->pinned_chunks list.
10995                  *
10996                  * So at this point we must be holding the chunk_mutex to avoid
10997                  * any races with chunk allocation (more specifically at
10998                  * volumes.c:contains_pending_extent()), to ensure it always
10999                  * sees the em, either in the pending_chunks list or in the
11000                  * pinned_chunks list.
11001                  */
11002                 list_move_tail(&em->list, &fs_info->pinned_chunks);
11003         }
11004         spin_unlock(&block_group->lock);
11005
11006         if (remove_em) {
11007                 struct extent_map_tree *em_tree;
11008
11009                 em_tree = &fs_info->mapping_tree.map_tree;
11010                 write_lock(&em_tree->lock);
11011                 /*
11012                  * The em might be in the pending_chunks list, so make sure the
11013                  * chunk mutex is locked, since remove_extent_mapping() will
11014                  * delete us from that list.
11015                  */
11016                 remove_extent_mapping(em_tree, em);
11017                 write_unlock(&em_tree->lock);
11018                 /* once for the tree */
11019                 free_extent_map(em);
11020         }
11021
11022         mutex_unlock(&fs_info->chunk_mutex);
11023
11024         ret = remove_block_group_free_space(trans, block_group);
11025         if (ret)
11026                 goto out;
11027
11028         btrfs_put_block_group(block_group);
11029         btrfs_put_block_group(block_group);
11030
11031         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
11032         if (ret > 0)
11033                 ret = -EIO;
11034         if (ret < 0)
11035                 goto out;
11036
11037         ret = btrfs_del_item(trans, root, path);
11038 out:
11039         if (remove_rsv)
11040                 btrfs_delayed_refs_rsv_release(fs_info, 1);
11041         btrfs_free_path(path);
11042         return ret;
11043 }
11044
11045 struct btrfs_trans_handle *
11046 btrfs_start_trans_remove_block_group(struct btrfs_fs_info *fs_info,
11047                                      const u64 chunk_offset)
11048 {
11049         struct extent_map_tree *em_tree = &fs_info->mapping_tree.map_tree;
11050         struct extent_map *em;
11051         struct map_lookup *map;
11052         unsigned int num_items;
11053
11054         read_lock(&em_tree->lock);
11055         em = lookup_extent_mapping(em_tree, chunk_offset, 1);
11056         read_unlock(&em_tree->lock);
11057         ASSERT(em && em->start == chunk_offset);
11058
11059         /*
11060          * We need to reserve 3 + N units from the metadata space info in order
11061          * to remove a block group (done at btrfs_remove_chunk() and at
11062          * btrfs_remove_block_group()), which are used for:
11063          *
11064          * 1 unit for adding the free space inode's orphan (located in the tree
11065          * of tree roots).
11066          * 1 unit for deleting the block group item (located in the extent
11067          * tree).
11068          * 1 unit for deleting the free space item (located in tree of tree
11069          * roots).
11070          * N units for deleting N device extent items corresponding to each
11071          * stripe (located in the device tree).
11072          *
11073          * In order to remove a block group we also need to reserve units in the
11074          * system space info in order to update the chunk tree (update one or
11075          * more device items and remove one chunk item), but this is done at
11076          * btrfs_remove_chunk() through a call to check_system_chunk().
11077          */
11078         map = em->map_lookup;
11079         num_items = 3 + map->num_stripes;
11080         free_extent_map(em);
11081
11082         return btrfs_start_transaction_fallback_global_rsv(fs_info->extent_root,
11083                                                            num_items, 1);
11084 }
11085
11086 /*
11087  * Process the unused_bgs list and remove any that don't have any allocated
11088  * space inside of them.
11089  */
11090 void btrfs_delete_unused_bgs(struct btrfs_fs_info *fs_info)
11091 {
11092         struct btrfs_block_group_cache *block_group;
11093         struct btrfs_space_info *space_info;
11094         struct btrfs_trans_handle *trans;
11095         int ret = 0;
11096
11097         if (!test_bit(BTRFS_FS_OPEN, &fs_info->flags))
11098                 return;
11099
11100         spin_lock(&fs_info->unused_bgs_lock);
11101         while (!list_empty(&fs_info->unused_bgs)) {
11102                 u64 start, end;
11103                 int trimming;
11104
11105                 block_group = list_first_entry(&fs_info->unused_bgs,
11106                                                struct btrfs_block_group_cache,
11107                                                bg_list);
11108                 list_del_init(&block_group->bg_list);
11109
11110                 space_info = block_group->space_info;
11111
11112                 if (ret || btrfs_mixed_space_info(space_info)) {
11113                         btrfs_put_block_group(block_group);
11114                         continue;
11115                 }
11116                 spin_unlock(&fs_info->unused_bgs_lock);
11117
11118                 mutex_lock(&fs_info->delete_unused_bgs_mutex);
11119
11120                 /* Don't want to race with allocators so take the groups_sem */
11121                 down_write(&space_info->groups_sem);
11122                 spin_lock(&block_group->lock);
11123                 if (block_group->reserved || block_group->pinned ||
11124                     btrfs_block_group_used(&block_group->item) ||
11125                     block_group->ro ||
11126                     list_is_singular(&block_group->list)) {
11127                         /*
11128                          * We want to bail if we made new allocations or have
11129                          * outstanding allocations in this block group.  We do
11130                          * the ro check in case balance is currently acting on
11131                          * this block group.
11132                          */
11133                         trace_btrfs_skip_unused_block_group(block_group);
11134                         spin_unlock(&block_group->lock);
11135                         up_write(&space_info->groups_sem);
11136                         goto next;
11137                 }
11138                 spin_unlock(&block_group->lock);
11139
11140                 /* We don't want to force the issue, only flip if it's ok. */
11141                 ret = inc_block_group_ro(block_group, 0);
11142                 up_write(&space_info->groups_sem);
11143                 if (ret < 0) {
11144                         ret = 0;
11145                         goto next;
11146                 }
11147
11148                 /*
11149                  * Want to do this before we do anything else so we can recover
11150                  * properly if we fail to join the transaction.
11151                  */
11152                 trans = btrfs_start_trans_remove_block_group(fs_info,
11153                                                      block_group->key.objectid);
11154                 if (IS_ERR(trans)) {
11155                         btrfs_dec_block_group_ro(block_group);
11156                         ret = PTR_ERR(trans);
11157                         goto next;
11158                 }
11159
11160                 /*
11161                  * We could have pending pinned extents for this block group,
11162                  * just delete them, we don't care about them anymore.
11163                  */
11164                 start = block_group->key.objectid;
11165                 end = start + block_group->key.offset - 1;
11166                 /*
11167                  * Hold the unused_bg_unpin_mutex lock to avoid racing with
11168                  * btrfs_finish_extent_commit(). If we are at transaction N,
11169                  * another task might be running finish_extent_commit() for the
11170                  * previous transaction N - 1, and have seen a range belonging
11171                  * to the block group in freed_extents[] before we were able to
11172                  * clear the whole block group range from freed_extents[]. This
11173                  * means that task can lookup for the block group after we
11174                  * unpinned it from freed_extents[] and removed it, leading to
11175                  * a BUG_ON() at btrfs_unpin_extent_range().
11176                  */
11177                 mutex_lock(&fs_info->unused_bg_unpin_mutex);
11178                 ret = clear_extent_bits(&fs_info->freed_extents[0], start, end,
11179                                   EXTENT_DIRTY);
11180                 if (ret) {
11181                         mutex_unlock(&fs_info->unused_bg_unpin_mutex);
11182                         btrfs_dec_block_group_ro(block_group);
11183                         goto end_trans;
11184                 }
11185                 ret = clear_extent_bits(&fs_info->freed_extents[1], start, end,
11186                                   EXTENT_DIRTY);
11187                 if (ret) {
11188                         mutex_unlock(&fs_info->unused_bg_unpin_mutex);
11189                         btrfs_dec_block_group_ro(block_group);
11190                         goto end_trans;
11191                 }
11192                 mutex_unlock(&fs_info->unused_bg_unpin_mutex);
11193
11194                 /* Reset pinned so btrfs_put_block_group doesn't complain */
11195                 spin_lock(&space_info->lock);
11196                 spin_lock(&block_group->lock);
11197
11198                 update_bytes_pinned(space_info, -block_group->pinned);
11199                 space_info->bytes_readonly += block_group->pinned;
11200                 percpu_counter_add_batch(&space_info->total_bytes_pinned,
11201                                    -block_group->pinned,
11202                                    BTRFS_TOTAL_BYTES_PINNED_BATCH);
11203                 block_group->pinned = 0;
11204
11205                 spin_unlock(&block_group->lock);
11206                 spin_unlock(&space_info->lock);
11207
11208                 /* DISCARD can flip during remount */
11209                 trimming = btrfs_test_opt(fs_info, DISCARD);
11210
11211                 /* Implicit trim during transaction commit. */
11212                 if (trimming)
11213                         btrfs_get_block_group_trimming(block_group);
11214
11215                 /*
11216                  * Btrfs_remove_chunk will abort the transaction if things go
11217                  * horribly wrong.
11218                  */
11219                 ret = btrfs_remove_chunk(trans, block_group->key.objectid);
11220
11221                 if (ret) {
11222                         if (trimming)
11223                                 btrfs_put_block_group_trimming(block_group);
11224                         goto end_trans;
11225                 }
11226
11227                 /*
11228                  * If we're not mounted with -odiscard, we can just forget
11229                  * about this block group. Otherwise we'll need to wait
11230                  * until transaction commit to do the actual discard.
11231                  */
11232                 if (trimming) {
11233                         spin_lock(&fs_info->unused_bgs_lock);
11234                         /*
11235                          * A concurrent scrub might have added us to the list
11236                          * fs_info->unused_bgs, so use a list_move operation
11237                          * to add the block group to the deleted_bgs list.
11238                          */
11239                         list_move(&block_group->bg_list,
11240                                   &trans->transaction->deleted_bgs);
11241                         spin_unlock(&fs_info->unused_bgs_lock);
11242                         btrfs_get_block_group(block_group);
11243                 }
11244 end_trans:
11245                 btrfs_end_transaction(trans);
11246 next:
11247                 mutex_unlock(&fs_info->delete_unused_bgs_mutex);
11248                 btrfs_put_block_group(block_group);
11249                 spin_lock(&fs_info->unused_bgs_lock);
11250         }
11251         spin_unlock(&fs_info->unused_bgs_lock);
11252 }
11253
11254 int btrfs_init_space_info(struct btrfs_fs_info *fs_info)
11255 {
11256         struct btrfs_super_block *disk_super;
11257         u64 features;
11258         u64 flags;
11259         int mixed = 0;
11260         int ret;
11261
11262         disk_super = fs_info->super_copy;
11263         if (!btrfs_super_root(disk_super))
11264                 return -EINVAL;
11265
11266         features = btrfs_super_incompat_flags(disk_super);
11267         if (features & BTRFS_FEATURE_INCOMPAT_MIXED_GROUPS)
11268                 mixed = 1;
11269
11270         flags = BTRFS_BLOCK_GROUP_SYSTEM;
11271         ret = create_space_info(fs_info, flags);
11272         if (ret)
11273                 goto out;
11274
11275         if (mixed) {
11276                 flags = BTRFS_BLOCK_GROUP_METADATA | BTRFS_BLOCK_GROUP_DATA;
11277                 ret = create_space_info(fs_info, flags);
11278         } else {
11279                 flags = BTRFS_BLOCK_GROUP_METADATA;
11280                 ret = create_space_info(fs_info, flags);
11281                 if (ret)
11282                         goto out;
11283
11284                 flags = BTRFS_BLOCK_GROUP_DATA;
11285                 ret = create_space_info(fs_info, flags);
11286         }
11287 out:
11288         return ret;
11289 }
11290
11291 int btrfs_error_unpin_extent_range(struct btrfs_fs_info *fs_info,
11292                                    u64 start, u64 end)
11293 {
11294         return unpin_extent_range(fs_info, start, end, false);
11295 }
11296
11297 /*
11298  * It used to be that old block groups would be left around forever.
11299  * Iterating over them would be enough to trim unused space.  Since we
11300  * now automatically remove them, we also need to iterate over unallocated
11301  * space.
11302  *
11303  * We don't want a transaction for this since the discard may take a
11304  * substantial amount of time.  We don't require that a transaction be
11305  * running, but we do need to take a running transaction into account
11306  * to ensure that we're not discarding chunks that were released or
11307  * allocated in the current transaction.
11308  *
11309  * Holding the chunks lock will prevent other threads from allocating
11310  * or releasing chunks, but it won't prevent a running transaction
11311  * from committing and releasing the memory that the pending chunks
11312  * list head uses.  For that, we need to take a reference to the
11313  * transaction and hold the commit root sem.  We only need to hold
11314  * it while performing the free space search since we have already
11315  * held back allocations.
11316  */
11317 static int btrfs_trim_free_extents(struct btrfs_device *device,
11318                                    u64 minlen, u64 *trimmed)
11319 {
11320         u64 start = 0, len = 0;
11321         int ret;
11322
11323         *trimmed = 0;
11324
11325         /* Discard not supported = nothing to do. */
11326         if (!blk_queue_discard(bdev_get_queue(device->bdev)))
11327                 return 0;
11328
11329         /* Not writable = nothing to do. */
11330         if (!test_bit(BTRFS_DEV_STATE_WRITEABLE, &device->dev_state))
11331                 return 0;
11332
11333         /* No free space = nothing to do. */
11334         if (device->total_bytes <= device->bytes_used)
11335                 return 0;
11336
11337         ret = 0;
11338
11339         while (1) {
11340                 struct btrfs_fs_info *fs_info = device->fs_info;
11341                 struct btrfs_transaction *trans;
11342                 u64 bytes;
11343
11344                 ret = mutex_lock_interruptible(&fs_info->chunk_mutex);
11345                 if (ret)
11346                         break;
11347
11348                 ret = down_read_killable(&fs_info->commit_root_sem);
11349                 if (ret) {
11350                         mutex_unlock(&fs_info->chunk_mutex);
11351                         break;
11352                 }
11353
11354                 spin_lock(&fs_info->trans_lock);
11355                 trans = fs_info->running_transaction;
11356                 if (trans)
11357                         refcount_inc(&trans->use_count);
11358                 spin_unlock(&fs_info->trans_lock);
11359
11360                 if (!trans)
11361                         up_read(&fs_info->commit_root_sem);
11362
11363                 ret = find_free_dev_extent_start(trans, device, minlen, start,
11364                                                  &start, &len);
11365                 if (trans) {
11366                         up_read(&fs_info->commit_root_sem);
11367                         btrfs_put_transaction(trans);
11368                 }
11369
11370                 if (ret) {
11371                         mutex_unlock(&fs_info->chunk_mutex);
11372                         if (ret == -ENOSPC)
11373                                 ret = 0;
11374                         break;
11375                 }
11376
11377                 ret = btrfs_issue_discard(device->bdev, start, len, &bytes);
11378                 mutex_unlock(&fs_info->chunk_mutex);
11379
11380                 if (ret)
11381                         break;
11382
11383                 start += len;
11384                 *trimmed += bytes;
11385
11386                 if (fatal_signal_pending(current)) {
11387                         ret = -ERESTARTSYS;
11388                         break;
11389                 }
11390
11391                 cond_resched();
11392         }
11393
11394         return ret;
11395 }
11396
11397 /*
11398  * Trim the whole filesystem by:
11399  * 1) trimming the free space in each block group
11400  * 2) trimming the unallocated space on each device
11401  *
11402  * This will also continue trimming even if a block group or device encounters
11403  * an error.  The return value will be the last error, or 0 if nothing bad
11404  * happens.
11405  */
11406 int btrfs_trim_fs(struct btrfs_fs_info *fs_info, struct fstrim_range *range)
11407 {
11408         struct btrfs_block_group_cache *cache = NULL;
11409         struct btrfs_device *device;
11410         struct list_head *devices;
11411         u64 group_trimmed;
11412         u64 start;
11413         u64 end;
11414         u64 trimmed = 0;
11415         u64 bg_failed = 0;
11416         u64 dev_failed = 0;
11417         int bg_ret = 0;
11418         int dev_ret = 0;
11419         int ret = 0;
11420
11421         cache = btrfs_lookup_first_block_group(fs_info, range->start);
11422         for (; cache; cache = next_block_group(fs_info, cache)) {
11423                 if (cache->key.objectid >= (range->start + range->len)) {
11424                         btrfs_put_block_group(cache);
11425                         break;
11426                 }
11427
11428                 start = max(range->start, cache->key.objectid);
11429                 end = min(range->start + range->len,
11430                                 cache->key.objectid + cache->key.offset);
11431
11432                 if (end - start >= range->minlen) {
11433                         if (!block_group_cache_done(cache)) {
11434                                 ret = cache_block_group(cache, 0);
11435                                 if (ret) {
11436                                         bg_failed++;
11437                                         bg_ret = ret;
11438                                         continue;
11439                                 }
11440                                 ret = wait_block_group_cache_done(cache);
11441                                 if (ret) {
11442                                         bg_failed++;
11443                                         bg_ret = ret;
11444                                         continue;
11445                                 }
11446                         }
11447                         ret = btrfs_trim_block_group(cache,
11448                                                      &group_trimmed,
11449                                                      start,
11450                                                      end,
11451                                                      range->minlen);
11452
11453                         trimmed += group_trimmed;
11454                         if (ret) {
11455                                 bg_failed++;
11456                                 bg_ret = ret;
11457                                 continue;
11458                         }
11459                 }
11460         }
11461
11462         if (bg_failed)
11463                 btrfs_warn(fs_info,
11464                         "failed to trim %llu block group(s), last error %d",
11465                         bg_failed, bg_ret);
11466         mutex_lock(&fs_info->fs_devices->device_list_mutex);
11467         devices = &fs_info->fs_devices->devices;
11468         list_for_each_entry(device, devices, dev_list) {
11469                 ret = btrfs_trim_free_extents(device, range->minlen,
11470                                               &group_trimmed);
11471                 if (ret) {
11472                         dev_failed++;
11473                         dev_ret = ret;
11474                         break;
11475                 }
11476
11477                 trimmed += group_trimmed;
11478         }
11479         mutex_unlock(&fs_info->fs_devices->device_list_mutex);
11480
11481         if (dev_failed)
11482                 btrfs_warn(fs_info,
11483                         "failed to trim %llu device(s), last error %d",
11484                         dev_failed, dev_ret);
11485         range->len = trimmed;
11486         if (bg_ret)
11487                 return bg_ret;
11488         return dev_ret;
11489 }
11490
11491 /*
11492  * btrfs_{start,end}_write_no_snapshotting() are similar to
11493  * mnt_{want,drop}_write(), they are used to prevent some tasks from writing
11494  * data into the page cache through nocow before the subvolume is snapshoted,
11495  * but flush the data into disk after the snapshot creation, or to prevent
11496  * operations while snapshotting is ongoing and that cause the snapshot to be
11497  * inconsistent (writes followed by expanding truncates for example).
11498  */
11499 void btrfs_end_write_no_snapshotting(struct btrfs_root *root)
11500 {
11501         percpu_counter_dec(&root->subv_writers->counter);
11502         cond_wake_up(&root->subv_writers->wait);
11503 }
11504
11505 int btrfs_start_write_no_snapshotting(struct btrfs_root *root)
11506 {
11507         if (atomic_read(&root->will_be_snapshotted))
11508                 return 0;
11509
11510         percpu_counter_inc(&root->subv_writers->counter);
11511         /*
11512          * Make sure counter is updated before we check for snapshot creation.
11513          */
11514         smp_mb();
11515         if (atomic_read(&root->will_be_snapshotted)) {
11516                 btrfs_end_write_no_snapshotting(root);
11517                 return 0;
11518         }
11519         return 1;
11520 }
11521
11522 void btrfs_wait_for_snapshot_creation(struct btrfs_root *root)
11523 {
11524         while (true) {
11525                 int ret;
11526
11527                 ret = btrfs_start_write_no_snapshotting(root);
11528                 if (ret)
11529                         break;
11530                 wait_var_event(&root->will_be_snapshotted,
11531                                !atomic_read(&root->will_be_snapshotted));
11532         }
11533 }
11534
11535 void btrfs_mark_bg_unused(struct btrfs_block_group_cache *bg)
11536 {
11537         struct btrfs_fs_info *fs_info = bg->fs_info;
11538
11539         spin_lock(&fs_info->unused_bgs_lock);
11540         if (list_empty(&bg->bg_list)) {
11541                 btrfs_get_block_group(bg);
11542                 trace_btrfs_add_unused_block_group(bg);
11543                 list_add_tail(&bg->bg_list, &fs_info->unused_bgs);
11544         }
11545         spin_unlock(&fs_info->unused_bgs_lock);
11546 }