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_fs_info *fs_info,         \
  59                                  struct btrfs_space_info *sinfo,        \
  60                                  s64 bytes)                             \
  61 {                                                                       \
  62         lockdep_assert_held(&sinfo->lock);                              \
  63         trace_update_##name(fs_info, sinfo, sinfo->name, bytes);        \
  64         if (bytes < 0 && sinfo->name < -bytes) {                        \
  65                 WARN_ON(1);                                             \
  66                 sinfo->name = 0;                                        \
  67                 return;                                                 \
  68         }                                                               \
  69         sinfo->name += bytes;                                           \
  70 }
  71
  72 DECLARE_SPACE_INFO_UPDATE(bytes_may_use);
  73 DECLARE_SPACE_INFO_UPDATE(bytes_pinned);
  74
  75 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
  76                                struct btrfs_delayed_ref_node *node, u64 parent,
  77                                u64 root_objectid, u64 owner_objectid,
  78                                u64 owner_offset, int refs_to_drop,
  79                                struct btrfs_delayed_extent_op *extra_op);
  80 static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op,
  81                                     struct extent_buffer *leaf,
  82                                     struct btrfs_extent_item *ei);
  83 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
  84                                       u64 parent, u64 root_objectid,
  85                                       u64 flags, u64 owner, u64 offset,
  86                                       struct btrfs_key *ins, int ref_mod);
  87 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
  88                                      struct btrfs_delayed_ref_node *node,
  89                                      struct btrfs_delayed_extent_op *extent_op);
  90 static int do_chunk_alloc(struct btrfs_trans_handle *trans, u64 flags,
  91                           int force);
  92 static int find_next_key(struct btrfs_path *path, int level,
  93                          struct btrfs_key *key);
  94 static void dump_space_info(struct btrfs_fs_info *fs_info,
  95                             struct btrfs_space_info *info, u64 bytes,
  96                             int dump_block_groups);
  97 static int block_rsv_use_bytes(struct btrfs_block_rsv *block_rsv,
  98                                u64 num_bytes);
  99 static void space_info_add_new_bytes(struct btrfs_fs_info *fs_info,
 100                                      struct btrfs_space_info *space_info,
 101                                      u64 num_bytes);
 102 static void space_info_add_old_bytes(struct btrfs_fs_info *fs_info,
 103                                      struct btrfs_space_info *space_info,
 104                                      u64 num_bytes);
 105
 106 static noinline int
 107 block_group_cache_done(struct btrfs_block_group_cache *cache)
 108 {
 109         smp_mb();
 110         return cache->cached == BTRFS_CACHE_FINISHED ||
 111                 cache->cached == BTRFS_CACHE_ERROR;
 112 }
 113
 114 static int block_group_bits(struct btrfs_block_group_cache *cache, u64 bits)
 115 {
 116         return (cache->flags & bits) == bits;
 117 }
 118
 119 void btrfs_get_block_group(struct btrfs_block_group_cache *cache)
 120 {
 121         atomic_inc(&cache->count);
 122 }
 123
 124 void btrfs_put_block_group(struct btrfs_block_group_cache *cache)
 125 {
 126         if (atomic_dec_and_test(&cache->count)) {
 127                 WARN_ON(cache->pinned > 0);
 128                 WARN_ON(cache->reserved > 0);
 129
 130                 /*
 131                  * If not empty, someone is still holding mutex of
 132                  * full_stripe_lock, which can only be released by caller.
 133                  * And it will definitely cause use-after-free when caller
 134                  * tries to release full stripe lock.
 135                  *
 136                  * No better way to resolve, but only to warn.
 137                  */
 138                 WARN_ON(!RB_EMPTY_ROOT(&cache->full_stripe_locks_root.root));
 139                 kfree(cache->free_space_ctl);
 140                 kfree(cache);
 141         }
 142 }
 143
 144 /*
 145  * this adds the block group to the fs_info rb tree for the block group
 146  * cache
 147  */
 148 static int btrfs_add_block_group_cache(struct btrfs_fs_info *info,
 149                                 struct btrfs_block_group_cache *block_group)
 150 {
 151         struct rb_node **p;
 152         struct rb_node *parent = NULL;
 153         struct btrfs_block_group_cache *cache;
 154
 155         spin_lock(&info->block_group_cache_lock);
 156         p = &info->block_group_cache_tree.rb_node;
 157
 158         while (*p) {
 159                 parent = *p;
 160                 cache = rb_entry(parent, struct btrfs_block_group_cache,
 161                                  cache_node);
 162                 if (block_group->key.objectid < cache->key.objectid) {
 163                         p = &(*p)->rb_left;
 164                 } else if (block_group->key.objectid > cache->key.objectid) {
 165                         p = &(*p)->rb_right;
 166                 } else {
 167                         spin_unlock(&info->block_group_cache_lock);
 168                         return -EEXIST;
 169                 }
 170         }
 171
 172         rb_link_node(&block_group->cache_node, parent, p);
 173         rb_insert_color(&block_group->cache_node,
 174                         &info->block_group_cache_tree);
 175
 176         if (info->first_logical_byte > block_group->key.objectid)
 177                 info->first_logical_byte = block_group->key.objectid;
 178
 179         spin_unlock(&info->block_group_cache_lock);
 180
 181         return 0;
 182 }
 183
 184 /*
 185  * This will return the block group at or after bytenr if contains is 0, else
 186  * it will return the block group that contains the bytenr
 187  */
 188 static struct btrfs_block_group_cache *
 189 block_group_cache_tree_search(struct btrfs_fs_info *info, u64 bytenr,
 190                               int contains)
 191 {
 192         struct btrfs_block_group_cache *cache, *ret = NULL;
 193         struct rb_node *n;
 194         u64 end, start;
 195
 196         spin_lock(&info->block_group_cache_lock);
 197         n = info->block_group_cache_tree.rb_node;
 198
 199         while (n) {
 200                 cache = rb_entry(n, struct btrfs_block_group_cache,
 201                                  cache_node);
 202                 end = cache->key.objectid + cache->key.offset - 1;
 203                 start = cache->key.objectid;
 204
 205                 if (bytenr < start) {
 206                         if (!contains && (!ret || start < ret->key.objectid))
 207                                 ret = cache;
 208                         n = n->rb_left;
 209                 } else if (bytenr > start) {
 210                         if (contains && bytenr <= end) {
 211                                 ret = cache;
 212                                 break;
 213                         }
 214                         n = n->rb_right;
 215                 } else {
 216                         ret = cache;
 217                         break;
 218                 }
 219         }
 220         if (ret) {
 221                 btrfs_get_block_group(ret);
 222                 if (bytenr == 0 && info->first_logical_byte > ret->key.objectid)
 223                         info->first_logical_byte = ret->key.objectid;
 224         }
 225         spin_unlock(&info->block_group_cache_lock);
 226
 227         return ret;
 228 }
 229
 230 static int add_excluded_extent(struct btrfs_fs_info *fs_info,
 231                                u64 start, u64 num_bytes)
 232 {
 233         u64 end = start + num_bytes - 1;
 234         set_extent_bits(&fs_info->freed_extents[0],
 235                         start, end, EXTENT_UPTODATE);
 236         set_extent_bits(&fs_info->freed_extents[1],
 237                         start, end, EXTENT_UPTODATE);
 238         return 0;
 239 }
 240
 241 static void free_excluded_extents(struct btrfs_block_group_cache *cache)
 242 {
 243         struct btrfs_fs_info *fs_info = cache->fs_info;
 244         u64 start, end;
 245
 246         start = cache->key.objectid;
 247         end = start + cache->key.offset - 1;
 248
 249         clear_extent_bits(&fs_info->freed_extents[0],
 250                           start, end, EXTENT_UPTODATE);
 251         clear_extent_bits(&fs_info->freed_extents[1],
 252                           start, end, EXTENT_UPTODATE);
 253 }
 254
 255 static int exclude_super_stripes(struct btrfs_block_group_cache *cache)
 256 {
 257         struct btrfs_fs_info *fs_info = cache->fs_info;
 258         u64 bytenr;
 259         u64 *logical;
 260         int stripe_len;
 261         int i, nr, ret;
 262
 263         if (cache->key.objectid < BTRFS_SUPER_INFO_OFFSET) {
 264                 stripe_len = BTRFS_SUPER_INFO_OFFSET - cache->key.objectid;
 265                 cache->bytes_super += stripe_len;
 266                 ret = add_excluded_extent(fs_info, cache->key.objectid,
 267                                           stripe_len);
 268                 if (ret)
 269                         return ret;
 270         }
 271
 272         for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
 273                 bytenr = btrfs_sb_offset(i);
 274                 ret = btrfs_rmap_block(fs_info, cache->key.objectid,
 275                                        bytenr, &logical, &nr, &stripe_len);
 276                 if (ret)
 277                         return ret;
 278
 279                 while (nr--) {
 280                         u64 start, len;
 281
 282                         if (logical[nr] > cache->key.objectid +
 283                             cache->key.offset)
 284                                 continue;
 285
 286                         if (logical[nr] + stripe_len <= cache->key.objectid)
 287                                 continue;
 288
 289                         start = logical[nr];
 290                         if (start < cache->key.objectid) {
 291                                 start = cache->key.objectid;
 292                                 len = (logical[nr] + stripe_len) - start;
 293                         } else {
 294                                 len = min_t(u64, stripe_len,
 295                                             cache->key.objectid +
 296                                             cache->key.offset - start);
 297                         }
 298
 299                         cache->bytes_super += len;
 300                         ret = add_excluded_extent(fs_info, start, len);
 301                         if (ret) {
 302                                 kfree(logical);
 303                                 return ret;
 304                         }
 305                 }
 306
 307                 kfree(logical);
 308         }
 309         return 0;
 310 }
 311
 312 static struct btrfs_caching_control *
 313 get_caching_control(struct btrfs_block_group_cache *cache)
 314 {
 315         struct btrfs_caching_control *ctl;
 316
 317         spin_lock(&cache->lock);
 318         if (!cache->caching_ctl) {
 319                 spin_unlock(&cache->lock);
 320                 return NULL;
 321         }
 322
 323         ctl = cache->caching_ctl;
 324         refcount_inc(&ctl->count);
 325         spin_unlock(&cache->lock);
 326         return ctl;
 327 }
 328
 329 static void put_caching_control(struct btrfs_caching_control *ctl)
 330 {
 331         if (refcount_dec_and_test(&ctl->count))
 332                 kfree(ctl);
 333 }
 334
 335 #ifdef CONFIG_BTRFS_DEBUG
 336 static void fragment_free_space(struct btrfs_block_group_cache *block_group)
 337 {
 338         struct btrfs_fs_info *fs_info = block_group->fs_info;
 339         u64 start = block_group->key.objectid;
 340         u64 len = block_group->key.offset;
 341         u64 chunk = block_group->flags & BTRFS_BLOCK_GROUP_METADATA ?
 342                 fs_info->nodesize : fs_info->sectorsize;
 343         u64 step = chunk << 1;
 344
 345         while (len > chunk) {
 346                 btrfs_remove_free_space(block_group, start, chunk);
 347                 start += step;
 348                 if (len < step)
 349                         len = 0;
 350                 else
 351                         len -= step;
 352         }
 353 }
 354 #endif
 355
 356 /*
 357  * this is only called by cache_block_group, since we could have freed extents
 358  * we need to check the pinned_extents for any extents that can't be used yet
 359  * since their free space will be released as soon as the transaction commits.
 360  */
 361 u64 add_new_free_space(struct btrfs_block_group_cache *block_group,
 362                        u64 start, u64 end)
 363 {
 364         struct btrfs_fs_info *info = block_group->fs_info;
 365         u64 extent_start, extent_end, size, total_added = 0;
 366         int ret;
 367
 368         while (start < end) {
 369                 ret = find_first_extent_bit(info->pinned_extents, start,
 370                                             &extent_start, &extent_end,
 371                                             EXTENT_DIRTY | EXTENT_UPTODATE,
 372                                             NULL);
 373                 if (ret)
 374                         break;
 375
 376                 if (extent_start <= start) {
 377                         start = extent_end + 1;
 378                 } else if (extent_start > start && extent_start < end) {
 379                         size = extent_start - start;
 380                         total_added += size;
 381                         ret = btrfs_add_free_space(block_group, start,
 382                                                    size);
 383                         BUG_ON(ret); /* -ENOMEM or logic error */
 384                         start = extent_end + 1;
 385                 } else {
 386                         break;
 387                 }
 388         }
 389
 390         if (start < end) {
 391                 size = end - start;
 392                 total_added += size;
 393                 ret = btrfs_add_free_space(block_group, start, size);
 394                 BUG_ON(ret); /* -ENOMEM or logic error */
 395         }
 396
 397         return total_added;
 398 }
 399
 400 static int load_extent_tree_free(struct btrfs_caching_control *caching_ctl)
 401 {
 402         struct btrfs_block_group_cache *block_group = caching_ctl->block_group;
 403         struct btrfs_fs_info *fs_info = block_group->fs_info;
 404         struct btrfs_root *extent_root = fs_info->extent_root;
 405         struct btrfs_path *path;
 406         struct extent_buffer *leaf;
 407         struct btrfs_key key;
 408         u64 total_found = 0;
 409         u64 last = 0;
 410         u32 nritems;
 411         int ret;
 412         bool wakeup = true;
 413
 414         path = btrfs_alloc_path();
 415         if (!path)
 416                 return -ENOMEM;
 417
 418         last = max_t(u64, block_group->key.objectid, BTRFS_SUPER_INFO_OFFSET);
 419
 420 #ifdef CONFIG_BTRFS_DEBUG
 421         /*
 422          * If we're fragmenting we don't want to make anybody think we can
 423          * allocate from this block group until we've had a chance to fragment
 424          * the free space.
 425          */
 426         if (btrfs_should_fragment_free_space(block_group))
 427                 wakeup = false;
 428 #endif
 429         /*
 430          * We don't want to deadlock with somebody trying to allocate a new
 431          * extent for the extent root while also trying to search the extent
 432          * root to add free space.  So we skip locking and search the commit
 433          * root, since its read-only
 434          */
 435         path->skip_locking = 1;
 436         path->search_commit_root = 1;
 437         path->reada = READA_FORWARD;
 438
 439         key.objectid = last;
 440         key.offset = 0;
 441         key.type = BTRFS_EXTENT_ITEM_KEY;
 442
 443 next:
 444         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
 445         if (ret < 0)
 446                 goto out;
 447
 448         leaf = path->nodes[0];
 449         nritems = btrfs_header_nritems(leaf);
 450
 451         while (1) {
 452                 if (btrfs_fs_closing(fs_info) > 1) {
 453                         last = (u64)-1;
 454                         break;
 455                 }
 456
 457                 if (path->slots[0] < nritems) {
 458                         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
 459                 } else {
 460                         ret = find_next_key(path, 0, &key);
 461                         if (ret)
 462                                 break;
 463
 464                         if (need_resched() ||
 465                             rwsem_is_contended(&fs_info->commit_root_sem)) {
 466                                 if (wakeup)
 467                                         caching_ctl->progress = last;
 468                                 btrfs_release_path(path);
 469                                 up_read(&fs_info->commit_root_sem);
 470                                 mutex_unlock(&caching_ctl->mutex);
 471                                 cond_resched();
 472                                 mutex_lock(&caching_ctl->mutex);
 473                                 down_read(&fs_info->commit_root_sem);
 474                                 goto next;
 475                         }
 476
 477                         ret = btrfs_next_leaf(extent_root, path);
 478                         if (ret < 0)
 479                                 goto out;
 480                         if (ret)
 481                                 break;
 482                         leaf = path->nodes[0];
 483                         nritems = btrfs_header_nritems(leaf);
 484                         continue;
 485                 }
 486
 487                 if (key.objectid < last) {
 488                         key.objectid = last;
 489                         key.offset = 0;
 490                         key.type = BTRFS_EXTENT_ITEM_KEY;
 491
 492                         if (wakeup)
 493                                 caching_ctl->progress = last;
 494                         btrfs_release_path(path);
 495                         goto next;
 496                 }
 497
 498                 if (key.objectid < block_group->key.objectid) {
 499                         path->slots[0]++;
 500                         continue;
 501                 }
 502
 503                 if (key.objectid >= block_group->key.objectid +
 504                     block_group->key.offset)
 505                         break;
 506
 507                 if (key.type == BTRFS_EXTENT_ITEM_KEY ||
 508                     key.type == BTRFS_METADATA_ITEM_KEY) {
 509                         total_found += add_new_free_space(block_group, last,
 510                                                           key.objectid);
 511                         if (key.type == BTRFS_METADATA_ITEM_KEY)
 512                                 last = key.objectid +
 513                                         fs_info->nodesize;
 514                         else
 515                                 last = key.objectid + key.offset;
 516
 517                         if (total_found > CACHING_CTL_WAKE_UP) {
 518                                 total_found = 0;
 519                                 if (wakeup)
 520                                         wake_up(&caching_ctl->wait);
 521                         }
 522                 }
 523                 path->slots[0]++;
 524         }
 525         ret = 0;
 526
 527         total_found += add_new_free_space(block_group, last,
 528                                           block_group->key.objectid +
 529                                           block_group->key.offset);
 530         caching_ctl->progress = (u64)-1;
 531
 532 out:
 533         btrfs_free_path(path);
 534         return ret;
 535 }
 536
 537 static noinline void caching_thread(struct btrfs_work *work)
 538 {
 539         struct btrfs_block_group_cache *block_group;
 540         struct btrfs_fs_info *fs_info;
 541         struct btrfs_caching_control *caching_ctl;
 542         int ret;
 543
 544         caching_ctl = container_of(work, struct btrfs_caching_control, work);
 545         block_group = caching_ctl->block_group;
 546         fs_info = block_group->fs_info;
 547
 548         mutex_lock(&caching_ctl->mutex);
 549         down_read(&fs_info->commit_root_sem);
 550
 551         if (btrfs_fs_compat_ro(fs_info, FREE_SPACE_TREE))
 552                 ret = load_free_space_tree(caching_ctl);
 553         else
 554                 ret = load_extent_tree_free(caching_ctl);
 555
 556         spin_lock(&block_group->lock);
 557         block_group->caching_ctl = NULL;
 558         block_group->cached = ret ? BTRFS_CACHE_ERROR : BTRFS_CACHE_FINISHED;
 559         spin_unlock(&block_group->lock);
 560
 561 #ifdef CONFIG_BTRFS_DEBUG
 562         if (btrfs_should_fragment_free_space(block_group)) {
 563                 u64 bytes_used;
 564
 565                 spin_lock(&block_group->space_info->lock);
 566                 spin_lock(&block_group->lock);
 567                 bytes_used = block_group->key.offset -
 568                         btrfs_block_group_used(&block_group->item);
 569                 block_group->space_info->bytes_used += bytes_used >> 1;
 570                 spin_unlock(&block_group->lock);
 571                 spin_unlock(&block_group->space_info->lock);
 572                 fragment_free_space(block_group);
 573         }
 574 #endif
 575
 576         caching_ctl->progress = (u64)-1;
 577
 578         up_read(&fs_info->commit_root_sem);
 579         free_excluded_extents(block_group);
 580         mutex_unlock(&caching_ctl->mutex);
 581
 582         wake_up(&caching_ctl->wait);
 583
 584         put_caching_control(caching_ctl);
 585         btrfs_put_block_group(block_group);
 586 }
 587
 588 static int cache_block_group(struct btrfs_block_group_cache *cache,
 589                              int load_cache_only)
 590 {
 591         DEFINE_WAIT(wait);
 592         struct btrfs_fs_info *fs_info = cache->fs_info;
 593         struct btrfs_caching_control *caching_ctl;
 594         int ret = 0;
 595
 596         caching_ctl = kzalloc(sizeof(*caching_ctl), GFP_NOFS);
 597         if (!caching_ctl)
 598                 return -ENOMEM;
 599
 600         INIT_LIST_HEAD(&caching_ctl->list);
 601         mutex_init(&caching_ctl->mutex);
 602         init_waitqueue_head(&caching_ctl->wait);
 603         caching_ctl->block_group = cache;
 604         caching_ctl->progress = cache->key.objectid;
 605         refcount_set(&caching_ctl->count, 1);
 606         btrfs_init_work(&caching_ctl->work, btrfs_cache_helper,
 607                         caching_thread, NULL, NULL);
 608
 609         spin_lock(&cache->lock);
 610         /*
 611          * This should be a rare occasion, but this could happen I think in the
 612          * case where one thread starts to load the space cache info, and then
 613          * some other thread starts a transaction commit which tries to do an
 614          * allocation while the other thread is still loading the space cache
 615          * info.  The previous loop should have kept us from choosing this block
 616          * group, but if we've moved to the state where we will wait on caching
 617          * block groups we need to first check if we're doing a fast load here,
 618          * so we can wait for it to finish, otherwise we could end up allocating
 619          * from a block group who's cache gets evicted for one reason or
 620          * another.
 621          */
 622         while (cache->cached == BTRFS_CACHE_FAST) {
 623                 struct btrfs_caching_control *ctl;
 624
 625                 ctl = cache->caching_ctl;
 626                 refcount_inc(&ctl->count);
 627                 prepare_to_wait(&ctl->wait, &wait, TASK_UNINTERRUPTIBLE);
 628                 spin_unlock(&cache->lock);
 629
 630                 schedule();
 631
 632                 finish_wait(&ctl->wait, &wait);
 633                 put_caching_control(ctl);
 634                 spin_lock(&cache->lock);
 635         }
 636
 637         if (cache->cached != BTRFS_CACHE_NO) {
 638                 spin_unlock(&cache->lock);
 639                 kfree(caching_ctl);
 640                 return 0;
 641         }
 642         WARN_ON(cache->caching_ctl);
 643         cache->caching_ctl = caching_ctl;
 644         cache->cached = BTRFS_CACHE_FAST;
 645         spin_unlock(&cache->lock);
 646
 647         if (btrfs_test_opt(fs_info, SPACE_CACHE)) {
 648                 mutex_lock(&caching_ctl->mutex);
 649                 ret = load_free_space_cache(cache);
 650
 651                 spin_lock(&cache->lock);
 652                 if (ret == 1) {
 653                         cache->caching_ctl = NULL;
 654                         cache->cached = BTRFS_CACHE_FINISHED;
 655                         cache->last_byte_to_unpin = (u64)-1;
 656                         caching_ctl->progress = (u64)-1;
 657                 } else {
 658                         if (load_cache_only) {
 659                                 cache->caching_ctl = NULL;
 660                                 cache->cached = BTRFS_CACHE_NO;
 661                         } else {
 662                                 cache->cached = BTRFS_CACHE_STARTED;
 663                                 cache->has_caching_ctl = 1;
 664                         }
 665                 }
 666                 spin_unlock(&cache->lock);
 667 #ifdef CONFIG_BTRFS_DEBUG
 668                 if (ret == 1 &&
 669                     btrfs_should_fragment_free_space(cache)) {
 670                         u64 bytes_used;
 671
 672                         spin_lock(&cache->space_info->lock);
 673                         spin_lock(&cache->lock);
 674                         bytes_used = cache->key.offset -
 675                                 btrfs_block_group_used(&cache->item);
 676                         cache->space_info->bytes_used += bytes_used >> 1;
 677                         spin_unlock(&cache->lock);
 678                         spin_unlock(&cache->space_info->lock);
 679                         fragment_free_space(cache);
 680                 }
 681 #endif
 682                 mutex_unlock(&caching_ctl->mutex);
 683
 684                 wake_up(&caching_ctl->wait);
 685                 if (ret == 1) {
 686                         put_caching_control(caching_ctl);
 687                         free_excluded_extents(cache);
 688                         return 0;
 689                 }
 690         } else {
 691                 /*
 692                  * We're either using the free space tree or no caching at all.
 693                  * Set cached to the appropriate value and wakeup any waiters.
 694                  */
 695                 spin_lock(&cache->lock);
 696                 if (load_cache_only) {
 697                         cache->caching_ctl = NULL;
 698                         cache->cached = BTRFS_CACHE_NO;
 699                 } else {
 700                         cache->cached = BTRFS_CACHE_STARTED;
 701                         cache->has_caching_ctl = 1;
 702                 }
 703                 spin_unlock(&cache->lock);
 704                 wake_up(&caching_ctl->wait);
 705         }
 706
 707         if (load_cache_only) {
 708                 put_caching_control(caching_ctl);
 709                 return 0;
 710         }
 711
 712         down_write(&fs_info->commit_root_sem);
 713         refcount_inc(&caching_ctl->count);
 714         list_add_tail(&caching_ctl->list, &fs_info->caching_block_groups);
 715         up_write(&fs_info->commit_root_sem);
 716
 717         btrfs_get_block_group(cache);
 718
 719         btrfs_queue_work(fs_info->caching_workers, &caching_ctl->work);
 720
 721         return ret;
 722 }
 723
 724 /*
 725  * return the block group that starts at or after bytenr
 726  */
 727 static struct btrfs_block_group_cache *
 728 btrfs_lookup_first_block_group(struct btrfs_fs_info *info, u64 bytenr)
 729 {
 730         return block_group_cache_tree_search(info, bytenr, 0);
 731 }
 732
 733 /*
 734  * return the block group that contains the given bytenr
 735  */
 736 struct btrfs_block_group_cache *btrfs_lookup_block_group(
 737                                                  struct btrfs_fs_info *info,
 738                                                  u64 bytenr)
 739 {
 740         return block_group_cache_tree_search(info, bytenr, 1);
 741 }
 742
 743 static struct btrfs_space_info *__find_space_info(struct btrfs_fs_info *info,
 744                                                   u64 flags)
 745 {
 746         struct list_head *head = &info->space_info;
 747         struct btrfs_space_info *found;
 748
 749         flags &= BTRFS_BLOCK_GROUP_TYPE_MASK;
 750
 751         rcu_read_lock();
 752         list_for_each_entry_rcu(found, head, list) {
 753                 if (found->flags & flags) {
 754                         rcu_read_unlock();
 755                         return found;
 756                 }
 757         }
 758         rcu_read_unlock();
 759         return NULL;
 760 }
 761
 762 static u64 generic_ref_to_space_flags(struct btrfs_ref *ref)
 763 {
 764         if (ref->type == BTRFS_REF_METADATA) {
 765                 if (ref->tree_ref.root == BTRFS_CHUNK_TREE_OBJECTID)
 766                         return BTRFS_BLOCK_GROUP_SYSTEM;
 767                 else
 768                         return BTRFS_BLOCK_GROUP_METADATA;
 769         }
 770         return BTRFS_BLOCK_GROUP_DATA;
 771 }
 772
 773 static void add_pinned_bytes(struct btrfs_fs_info *fs_info,
 774                              struct btrfs_ref *ref)
 775 {
 776         struct btrfs_space_info *space_info;
 777         u64 flags = generic_ref_to_space_flags(ref);
 778
 779         space_info = __find_space_info(fs_info, flags);
 780         ASSERT(space_info);
 781         percpu_counter_add_batch(&space_info->total_bytes_pinned, ref->len,
 782                     BTRFS_TOTAL_BYTES_PINNED_BATCH);
 783 }
 784
 785 static void sub_pinned_bytes(struct btrfs_fs_info *fs_info,
 786                              struct btrfs_ref *ref)
 787 {
 788         struct btrfs_space_info *space_info;
 789         u64 flags = generic_ref_to_space_flags(ref);
 790
 791         space_info = __find_space_info(fs_info, flags);
 792         ASSERT(space_info);
 793         percpu_counter_add_batch(&space_info->total_bytes_pinned, -ref->len,
 794                     BTRFS_TOTAL_BYTES_PINNED_BATCH);
 795 }
 796
 797 /*
 798  * after adding space to the filesystem, we need to clear the full flags
 799  * on all the space infos.
 800  */
 801 void btrfs_clear_space_info_full(struct btrfs_fs_info *info)
 802 {
 803         struct list_head *head = &info->space_info;
 804         struct btrfs_space_info *found;
 805
 806         rcu_read_lock();
 807         list_for_each_entry_rcu(found, head, list)
 808                 found->full = 0;
 809         rcu_read_unlock();
 810 }
 811
 812 /* simple helper to search for an existing data extent at a given offset */
 813 int btrfs_lookup_data_extent(struct btrfs_fs_info *fs_info, u64 start, u64 len)
 814 {
 815         int ret;
 816         struct btrfs_key key;
 817         struct btrfs_path *path;
 818
 819         path = btrfs_alloc_path();
 820         if (!path)
 821                 return -ENOMEM;
 822
 823         key.objectid = start;
 824         key.offset = len;
 825         key.type = BTRFS_EXTENT_ITEM_KEY;
 826         ret = btrfs_search_slot(NULL, fs_info->extent_root, &key, path, 0, 0);
 827         btrfs_free_path(path);
 828         return ret;
 829 }
 830
 831 /*
 832  * helper function to lookup reference count and flags of a tree block.
 833  *
 834  * the head node for delayed ref is used to store the sum of all the
 835  * reference count modifications queued up in the rbtree. the head
 836  * node may also store the extent flags to set. This way you can check
 837  * to see what the reference count and extent flags would be if all of
 838  * the delayed refs are not processed.
 839  */
 840 int btrfs_lookup_extent_info(struct btrfs_trans_handle *trans,
 841                              struct btrfs_fs_info *fs_info, u64 bytenr,
 842                              u64 offset, int metadata, u64 *refs, u64 *flags)
 843 {
 844         struct btrfs_delayed_ref_head *head;
 845         struct btrfs_delayed_ref_root *delayed_refs;
 846         struct btrfs_path *path;
 847         struct btrfs_extent_item *ei;
 848         struct extent_buffer *leaf;
 849         struct btrfs_key key;
 850         u32 item_size;
 851         u64 num_refs;
 852         u64 extent_flags;
 853         int ret;
 854
 855         /*
 856          * If we don't have skinny metadata, don't bother doing anything
 857          * different
 858          */
 859         if (metadata && !btrfs_fs_incompat(fs_info, SKINNY_METADATA)) {
 860                 offset = fs_info->nodesize;
 861                 metadata = 0;
 862         }
 863
 864         path = btrfs_alloc_path();
 865         if (!path)
 866                 return -ENOMEM;
 867
 868         if (!trans) {
 869                 path->skip_locking = 1;
 870                 path->search_commit_root = 1;
 871         }
 872
 873 search_again:
 874         key.objectid = bytenr;
 875         key.offset = offset;
 876         if (metadata)
 877                 key.type = BTRFS_METADATA_ITEM_KEY;
 878         else
 879                 key.type = BTRFS_EXTENT_ITEM_KEY;
 880
 881         ret = btrfs_search_slot(trans, fs_info->extent_root, &key, path, 0, 0);
 882         if (ret < 0)
 883                 goto out_free;
 884
 885         if (ret > 0 && metadata && key.type == BTRFS_METADATA_ITEM_KEY) {
 886                 if (path->slots[0]) {
 887                         path->slots[0]--;
 888                         btrfs_item_key_to_cpu(path->nodes[0], &key,
 889                                               path->slots[0]);
 890                         if (key.objectid == bytenr &&
 891                             key.type == BTRFS_EXTENT_ITEM_KEY &&
 892                             key.offset == fs_info->nodesize)
 893                                 ret = 0;
 894                 }
 895         }
 896
 897         if (ret == 0) {
 898                 leaf = path->nodes[0];
 899                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
 900                 if (item_size >= sizeof(*ei)) {
 901                         ei = btrfs_item_ptr(leaf, path->slots[0],
 902                                             struct btrfs_extent_item);
 903                         num_refs = btrfs_extent_refs(leaf, ei);
 904                         extent_flags = btrfs_extent_flags(leaf, ei);
 905                 } else {
 906                         ret = -EINVAL;
 907                         btrfs_print_v0_err(fs_info);
 908                         if (trans)
 909                                 btrfs_abort_transaction(trans, ret);
 910                         else
 911                                 btrfs_handle_fs_error(fs_info, ret, NULL);
 912
 913                         goto out_free;
 914                 }
 915
 916                 BUG_ON(num_refs == 0);
 917         } else {
 918                 num_refs = 0;
 919                 extent_flags = 0;
 920                 ret = 0;
 921         }
 922
 923         if (!trans)
 924                 goto out;
 925
 926         delayed_refs = &trans->transaction->delayed_refs;
 927         spin_lock(&delayed_refs->lock);
 928         head = btrfs_find_delayed_ref_head(delayed_refs, bytenr);
 929         if (head) {
 930                 if (!mutex_trylock(&head->mutex)) {
 931                         refcount_inc(&head->refs);
 932                         spin_unlock(&delayed_refs->lock);
 933
 934                         btrfs_release_path(path);
 935
 936                         /*
 937                          * Mutex was contended, block until it's released and try
 938                          * again
 939                          */
 940                         mutex_lock(&head->mutex);
 941                         mutex_unlock(&head->mutex);
 942                         btrfs_put_delayed_ref_head(head);
 943                         goto search_again;
 944                 }
 945                 spin_lock(&head->lock);
 946                 if (head->extent_op && head->extent_op->update_flags)
 947                         extent_flags |= head->extent_op->flags_to_set;
 948                 else
 949                         BUG_ON(num_refs == 0);
 950
 951                 num_refs += head->ref_mod;
 952                 spin_unlock(&head->lock);
 953                 mutex_unlock(&head->mutex);
 954         }
 955         spin_unlock(&delayed_refs->lock);
 956 out:
 957         WARN_ON(num_refs == 0);
 958         if (refs)
 959                 *refs = num_refs;
 960         if (flags)
 961                 *flags = extent_flags;
 962 out_free:
 963         btrfs_free_path(path);
 964         return ret;
 965 }
 966
 967 /*
 968  * Back reference rules.  Back refs have three main goals:
 969  *
 970  * 1) differentiate between all holders of references to an extent so that
 971  *    when a reference is dropped we can make sure it was a valid reference
 972  *    before freeing the extent.
 973  *
 974  * 2) Provide enough information to quickly find the holders of an extent
 975  *    if we notice a given block is corrupted or bad.
 976  *
 977  * 3) Make it easy to migrate blocks for FS shrinking or storage pool
 978  *    maintenance.  This is actually the same as #2, but with a slightly
 979  *    different use case.
 980  *
 981  * There are two kinds of back refs. The implicit back refs is optimized
 982  * for pointers in non-shared tree blocks. For a given pointer in a block,
 983  * back refs of this kind provide information about the block's owner tree
 984  * and the pointer's key. These information allow us to find the block by
 985  * b-tree searching. The full back refs is for pointers in tree blocks not
 986  * referenced by their owner trees. The location of tree block is recorded
 987  * in the back refs. Actually the full back refs is generic, and can be
 988  * used in all cases the implicit back refs is used. The major shortcoming
 989  * of the full back refs is its overhead. Every time a tree block gets
 990  * COWed, we have to update back refs entry for all pointers in it.
 991  *
 992  * For a newly allocated tree block, we use implicit back refs for
 993  * pointers in it. This means most tree related operations only involve
 994  * implicit back refs. For a tree block created in old transaction, the
 995  * only way to drop a reference to it is COW it. So we can detect the
 996  * event that tree block loses its owner tree's reference and do the
 997  * back refs conversion.
 998  *
 999  * When a tree block is COWed through a tree, there are four cases:
1000  *
1001  * The reference count of the block is one and the tree is the block's
1002  * owner tree. Nothing to do in this case.
1003  *
1004  * The reference count of the block is one and the tree is not the
1005  * block's owner tree. In this case, full back refs is used for pointers
1006  * in the block. Remove these full back refs, add implicit back refs for
1007  * every pointers in the new block.
1008  *
1009  * The reference count of the block is greater than one and the tree is
1010  * the block's owner tree. In this case, implicit back refs is used for
1011  * pointers in the block. Add full back refs for every pointers in the
1012  * block, increase lower level extents' reference counts. The original
1013  * implicit back refs are entailed to the new block.
1014  *
1015  * The reference count of the block is greater than one and the tree is
1016  * not the block's owner tree. Add implicit back refs for every pointer in
1017  * the new block, increase lower level extents' reference count.
1018  *
1019  * Back Reference Key composing:
1020  *
1021  * The key objectid corresponds to the first byte in the extent,
1022  * The key type is used to differentiate between types of back refs.
1023  * There are different meanings of the key offset for different types
1024  * of back refs.
1025  *
1026  * File extents can be referenced by:
1027  *
1028  * - multiple snapshots, subvolumes, or different generations in one subvol
1029  * - different files inside a single subvolume
1030  * - different offsets inside a file (bookend extents in file.c)
1031  *
1032  * The extent ref structure for the implicit back refs has fields for:
1033  *
1034  * - Objectid of the subvolume root
1035  * - objectid of the file holding the reference
1036  * - original offset in the file
1037  * - how many bookend extents
1038  *
1039  * The key offset for the implicit back refs is hash of the first
1040  * three fields.
1041  *
1042  * The extent ref structure for the full back refs has field for:
1043  *
1044  * - number of pointers in the tree leaf
1045  *
1046  * The key offset for the implicit back refs is the first byte of
1047  * the tree leaf
1048  *
1049  * When a file extent is allocated, The implicit back refs is used.
1050  * the fields are filled in:
1051  *
1052  *     (root_key.objectid, inode objectid, offset in file, 1)
1053  *
1054  * When a file extent is removed file truncation, we find the
1055  * corresponding implicit back refs and check the following fields:
1056  *
1057  *     (btrfs_header_owner(leaf), inode objectid, offset in file)
1058  *
1059  * Btree extents can be referenced by:
1060  *
1061  * - Different subvolumes
1062  *
1063  * Both the implicit back refs and the full back refs for tree blocks
1064  * only consist of key. The key offset for the implicit back refs is
1065  * objectid of block's owner tree. The key offset for the full back refs
1066  * is the first byte of parent block.
1067  *
1068  * When implicit back refs is used, information about the lowest key and
1069  * level of the tree block are required. These information are stored in
1070  * tree block info structure.
1071  */
1072
1073 /*
1074  * is_data == BTRFS_REF_TYPE_BLOCK, tree block type is required,
1075  * is_data == BTRFS_REF_TYPE_DATA, data type is requiried,
1076  * is_data == BTRFS_REF_TYPE_ANY, either type is OK.
1077  */
1078 int btrfs_get_extent_inline_ref_type(const struct extent_buffer *eb,
1079                                      struct btrfs_extent_inline_ref *iref,
1080                                      enum btrfs_inline_ref_type is_data)
1081 {
1082         int type = btrfs_extent_inline_ref_type(eb, iref);
1083         u64 offset = btrfs_extent_inline_ref_offset(eb, iref);
1084
1085         if (type == BTRFS_TREE_BLOCK_REF_KEY ||
1086             type == BTRFS_SHARED_BLOCK_REF_KEY ||
1087             type == BTRFS_SHARED_DATA_REF_KEY ||
1088             type == BTRFS_EXTENT_DATA_REF_KEY) {
1089                 if (is_data == BTRFS_REF_TYPE_BLOCK) {
1090                         if (type == BTRFS_TREE_BLOCK_REF_KEY)
1091                                 return type;
1092                         if (type == BTRFS_SHARED_BLOCK_REF_KEY) {
1093                                 ASSERT(eb->fs_info);
1094                                 /*
1095                                  * Every shared one has parent tree
1096                                  * block, which must be aligned to
1097                                  * nodesize.
1098                                  */
1099                                 if (offset &&
1100                                     IS_ALIGNED(offset, eb->fs_info->nodesize))
1101                                         return type;
1102                         }
1103                 } else if (is_data == BTRFS_REF_TYPE_DATA) {
1104                         if (type == BTRFS_EXTENT_DATA_REF_KEY)
1105                                 return type;
1106                         if (type == BTRFS_SHARED_DATA_REF_KEY) {
1107                                 ASSERT(eb->fs_info);
1108                                 /*
1109                                  * Every shared one has parent tree
1110                                  * block, which must be aligned to
1111                                  * nodesize.
1112                                  */
1113                                 if (offset &&
1114                                     IS_ALIGNED(offset, eb->fs_info->nodesize))
1115                                         return type;
1116                         }
1117                 } else {
1118                         ASSERT(is_data == BTRFS_REF_TYPE_ANY);
1119                         return type;
1120                 }
1121         }
1122
1123         btrfs_print_leaf((struct extent_buffer *)eb);
1124         btrfs_err(eb->fs_info, "eb %llu invalid extent inline ref type %d",
1125                   eb->start, type);
1126         WARN_ON(1);
1127
1128         return BTRFS_REF_TYPE_INVALID;
1129 }
1130
1131 static u64 hash_extent_data_ref(u64 root_objectid, u64 owner, u64 offset)
1132 {
1133         u32 high_crc = ~(u32)0;
1134         u32 low_crc = ~(u32)0;
1135         __le64 lenum;
1136
1137         lenum = cpu_to_le64(root_objectid);
1138         high_crc = btrfs_crc32c(high_crc, &lenum, sizeof(lenum));
1139         lenum = cpu_to_le64(owner);
1140         low_crc = btrfs_crc32c(low_crc, &lenum, sizeof(lenum));
1141         lenum = cpu_to_le64(offset);
1142         low_crc = btrfs_crc32c(low_crc, &lenum, sizeof(lenum));
1143
1144         return ((u64)high_crc << 31) ^ (u64)low_crc;
1145 }
1146
1147 static u64 hash_extent_data_ref_item(struct extent_buffer *leaf,
1148                                      struct btrfs_extent_data_ref *ref)
1149 {
1150         return hash_extent_data_ref(btrfs_extent_data_ref_root(leaf, ref),
1151                                     btrfs_extent_data_ref_objectid(leaf, ref),
1152                                     btrfs_extent_data_ref_offset(leaf, ref));
1153 }
1154
1155 static int match_extent_data_ref(struct extent_buffer *leaf,
1156                                  struct btrfs_extent_data_ref *ref,
1157                                  u64 root_objectid, u64 owner, u64 offset)
1158 {
1159         if (btrfs_extent_data_ref_root(leaf, ref) != root_objectid ||
1160             btrfs_extent_data_ref_objectid(leaf, ref) != owner ||
1161             btrfs_extent_data_ref_offset(leaf, ref) != offset)
1162                 return 0;
1163         return 1;
1164 }
1165
1166 static noinline int lookup_extent_data_ref(struct btrfs_trans_handle *trans,
1167                                            struct btrfs_path *path,
1168                                            u64 bytenr, u64 parent,
1169                                            u64 root_objectid,
1170                                            u64 owner, u64 offset)
1171 {
1172         struct btrfs_root *root = trans->fs_info->extent_root;
1173         struct btrfs_key key;
1174         struct btrfs_extent_data_ref *ref;
1175         struct extent_buffer *leaf;
1176         u32 nritems;
1177         int ret;
1178         int recow;
1179         int err = -ENOENT;
1180
1181         key.objectid = bytenr;
1182         if (parent) {
1183                 key.type = BTRFS_SHARED_DATA_REF_KEY;
1184                 key.offset = parent;
1185         } else {
1186                 key.type = BTRFS_EXTENT_DATA_REF_KEY;
1187                 key.offset = hash_extent_data_ref(root_objectid,
1188                                                   owner, offset);
1189         }
1190 again:
1191         recow = 0;
1192         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1193         if (ret < 0) {
1194                 err = ret;
1195                 goto fail;
1196         }
1197
1198         if (parent) {
1199                 if (!ret)
1200                         return 0;
1201                 goto fail;
1202         }
1203
1204         leaf = path->nodes[0];
1205         nritems = btrfs_header_nritems(leaf);
1206         while (1) {
1207                 if (path->slots[0] >= nritems) {
1208                         ret = btrfs_next_leaf(root, path);
1209                         if (ret < 0)
1210                                 err = ret;
1211                         if (ret)
1212                                 goto fail;
1213
1214                         leaf = path->nodes[0];
1215                         nritems = btrfs_header_nritems(leaf);
1216                         recow = 1;
1217                 }
1218
1219                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1220                 if (key.objectid != bytenr ||
1221                     key.type != BTRFS_EXTENT_DATA_REF_KEY)
1222                         goto fail;
1223
1224                 ref = btrfs_item_ptr(leaf, path->slots[0],
1225                                      struct btrfs_extent_data_ref);
1226
1227                 if (match_extent_data_ref(leaf, ref, root_objectid,
1228                                           owner, offset)) {
1229                         if (recow) {
1230                                 btrfs_release_path(path);
1231                                 goto again;
1232                         }
1233                         err = 0;
1234                         break;
1235                 }
1236                 path->slots[0]++;
1237         }
1238 fail:
1239         return err;
1240 }
1241
1242 static noinline int insert_extent_data_ref(struct btrfs_trans_handle *trans,
1243                                            struct btrfs_path *path,
1244                                            u64 bytenr, u64 parent,
1245                                            u64 root_objectid, u64 owner,
1246                                            u64 offset, int refs_to_add)
1247 {
1248         struct btrfs_root *root = trans->fs_info->extent_root;
1249         struct btrfs_key key;
1250         struct extent_buffer *leaf;
1251         u32 size;
1252         u32 num_refs;
1253         int ret;
1254
1255         key.objectid = bytenr;
1256         if (parent) {
1257                 key.type = BTRFS_SHARED_DATA_REF_KEY;
1258                 key.offset = parent;
1259                 size = sizeof(struct btrfs_shared_data_ref);
1260         } else {
1261                 key.type = BTRFS_EXTENT_DATA_REF_KEY;
1262                 key.offset = hash_extent_data_ref(root_objectid,
1263                                                   owner, offset);
1264                 size = sizeof(struct btrfs_extent_data_ref);
1265         }
1266
1267         ret = btrfs_insert_empty_item(trans, root, path, &key, size);
1268         if (ret && ret != -EEXIST)
1269                 goto fail;
1270
1271         leaf = path->nodes[0];
1272         if (parent) {
1273                 struct btrfs_shared_data_ref *ref;
1274                 ref = btrfs_item_ptr(leaf, path->slots[0],
1275                                      struct btrfs_shared_data_ref);
1276                 if (ret == 0) {
1277                         btrfs_set_shared_data_ref_count(leaf, ref, refs_to_add);
1278                 } else {
1279                         num_refs = btrfs_shared_data_ref_count(leaf, ref);
1280                         num_refs += refs_to_add;
1281                         btrfs_set_shared_data_ref_count(leaf, ref, num_refs);
1282                 }
1283         } else {
1284                 struct btrfs_extent_data_ref *ref;
1285                 while (ret == -EEXIST) {
1286                         ref = btrfs_item_ptr(leaf, path->slots[0],
1287                                              struct btrfs_extent_data_ref);
1288                         if (match_extent_data_ref(leaf, ref, root_objectid,
1289                                                   owner, offset))
1290                                 break;
1291                         btrfs_release_path(path);
1292                         key.offset++;
1293                         ret = btrfs_insert_empty_item(trans, root, path, &key,
1294                                                       size);
1295                         if (ret && ret != -EEXIST)
1296                                 goto fail;
1297
1298                         leaf = path->nodes[0];
1299                 }
1300                 ref = btrfs_item_ptr(leaf, path->slots[0],
1301                                      struct btrfs_extent_data_ref);
1302                 if (ret == 0) {
1303                         btrfs_set_extent_data_ref_root(leaf, ref,
1304                                                        root_objectid);
1305                         btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
1306                         btrfs_set_extent_data_ref_offset(leaf, ref, offset);
1307                         btrfs_set_extent_data_ref_count(leaf, ref, refs_to_add);
1308                 } else {
1309                         num_refs = btrfs_extent_data_ref_count(leaf, ref);
1310                         num_refs += refs_to_add;
1311                         btrfs_set_extent_data_ref_count(leaf, ref, num_refs);
1312                 }
1313         }
1314         btrfs_mark_buffer_dirty(leaf);
1315         ret = 0;
1316 fail:
1317         btrfs_release_path(path);
1318         return ret;
1319 }
1320
1321 static noinline int remove_extent_data_ref(struct btrfs_trans_handle *trans,
1322                                            struct btrfs_path *path,
1323                                            int refs_to_drop, int *last_ref)
1324 {
1325         struct btrfs_key key;
1326         struct btrfs_extent_data_ref *ref1 = NULL;
1327         struct btrfs_shared_data_ref *ref2 = NULL;
1328         struct extent_buffer *leaf;
1329         u32 num_refs = 0;
1330         int ret = 0;
1331
1332         leaf = path->nodes[0];
1333         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1334
1335         if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
1336                 ref1 = btrfs_item_ptr(leaf, path->slots[0],
1337                                       struct btrfs_extent_data_ref);
1338                 num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1339         } else if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
1340                 ref2 = btrfs_item_ptr(leaf, path->slots[0],
1341                                       struct btrfs_shared_data_ref);
1342                 num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1343         } else if (unlikely(key.type == BTRFS_EXTENT_REF_V0_KEY)) {
1344                 btrfs_print_v0_err(trans->fs_info);
1345                 btrfs_abort_transaction(trans, -EINVAL);
1346                 return -EINVAL;
1347         } else {
1348                 BUG();
1349         }
1350
1351         BUG_ON(num_refs < refs_to_drop);
1352         num_refs -= refs_to_drop;
1353
1354         if (num_refs == 0) {
1355                 ret = btrfs_del_item(trans, trans->fs_info->extent_root, path);
1356                 *last_ref = 1;
1357         } else {
1358                 if (key.type == BTRFS_EXTENT_DATA_REF_KEY)
1359                         btrfs_set_extent_data_ref_count(leaf, ref1, num_refs);
1360                 else if (key.type == BTRFS_SHARED_DATA_REF_KEY)
1361                         btrfs_set_shared_data_ref_count(leaf, ref2, num_refs);
1362                 btrfs_mark_buffer_dirty(leaf);
1363         }
1364         return ret;
1365 }
1366
1367 static noinline u32 extent_data_ref_count(struct btrfs_path *path,
1368                                           struct btrfs_extent_inline_ref *iref)
1369 {
1370         struct btrfs_key key;
1371         struct extent_buffer *leaf;
1372         struct btrfs_extent_data_ref *ref1;
1373         struct btrfs_shared_data_ref *ref2;
1374         u32 num_refs = 0;
1375         int type;
1376
1377         leaf = path->nodes[0];
1378         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1379
1380         BUG_ON(key.type == BTRFS_EXTENT_REF_V0_KEY);
1381         if (iref) {
1382                 /*
1383                  * If type is invalid, we should have bailed out earlier than
1384                  * this call.
1385                  */
1386                 type = btrfs_get_extent_inline_ref_type(leaf, iref, BTRFS_REF_TYPE_DATA);
1387                 ASSERT(type != BTRFS_REF_TYPE_INVALID);
1388                 if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1389                         ref1 = (struct btrfs_extent_data_ref *)(&iref->offset);
1390                         num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1391                 } else {
1392                         ref2 = (struct btrfs_shared_data_ref *)(iref + 1);
1393                         num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1394                 }
1395         } else if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
1396                 ref1 = btrfs_item_ptr(leaf, path->slots[0],
1397                                       struct btrfs_extent_data_ref);
1398                 num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1399         } else if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
1400                 ref2 = btrfs_item_ptr(leaf, path->slots[0],
1401                                       struct btrfs_shared_data_ref);
1402                 num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1403         } else {
1404                 WARN_ON(1);
1405         }
1406         return num_refs;
1407 }
1408
1409 static noinline int lookup_tree_block_ref(struct btrfs_trans_handle *trans,
1410                                           struct btrfs_path *path,
1411                                           u64 bytenr, u64 parent,
1412                                           u64 root_objectid)
1413 {
1414         struct btrfs_root *root = trans->fs_info->extent_root;
1415         struct btrfs_key key;
1416         int ret;
1417
1418         key.objectid = bytenr;
1419         if (parent) {
1420                 key.type = BTRFS_SHARED_BLOCK_REF_KEY;
1421                 key.offset = parent;
1422         } else {
1423                 key.type = BTRFS_TREE_BLOCK_REF_KEY;
1424                 key.offset = root_objectid;
1425         }
1426
1427         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1428         if (ret > 0)
1429                 ret = -ENOENT;
1430         return ret;
1431 }
1432
1433 static noinline int insert_tree_block_ref(struct btrfs_trans_handle *trans,
1434                                           struct btrfs_path *path,
1435                                           u64 bytenr, u64 parent,
1436                                           u64 root_objectid)
1437 {
1438         struct btrfs_key key;
1439         int ret;
1440
1441         key.objectid = bytenr;
1442         if (parent) {
1443                 key.type = BTRFS_SHARED_BLOCK_REF_KEY;
1444                 key.offset = parent;
1445         } else {
1446                 key.type = BTRFS_TREE_BLOCK_REF_KEY;
1447                 key.offset = root_objectid;
1448         }
1449
1450         ret = btrfs_insert_empty_item(trans, trans->fs_info->extent_root,
1451                                       path, &key, 0);
1452         btrfs_release_path(path);
1453         return ret;
1454 }
1455
1456 static inline int extent_ref_type(u64 parent, u64 owner)
1457 {
1458         int type;
1459         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1460                 if (parent > 0)
1461                         type = BTRFS_SHARED_BLOCK_REF_KEY;
1462                 else
1463                         type = BTRFS_TREE_BLOCK_REF_KEY;
1464         } else {
1465                 if (parent > 0)
1466                         type = BTRFS_SHARED_DATA_REF_KEY;
1467                 else
1468                         type = BTRFS_EXTENT_DATA_REF_KEY;
1469         }
1470         return type;
1471 }
1472
1473 static int find_next_key(struct btrfs_path *path, int level,
1474                          struct btrfs_key *key)
1475
1476 {
1477         for (; level < BTRFS_MAX_LEVEL; level++) {
1478                 if (!path->nodes[level])
1479                         break;
1480                 if (path->slots[level] + 1 >=
1481                     btrfs_header_nritems(path->nodes[level]))
1482                         continue;
1483                 if (level == 0)
1484                         btrfs_item_key_to_cpu(path->nodes[level], key,
1485                                               path->slots[level] + 1);
1486                 else
1487                         btrfs_node_key_to_cpu(path->nodes[level], key,
1488                                               path->slots[level] + 1);
1489                 return 0;
1490         }
1491         return 1;
1492 }
1493
1494 /*
1495  * look for inline back ref. if back ref is found, *ref_ret is set
1496  * to the address of inline back ref, and 0 is returned.
1497  *
1498  * if back ref isn't found, *ref_ret is set to the address where it
1499  * should be inserted, and -ENOENT is returned.
1500  *
1501  * if insert is true and there are too many inline back refs, the path
1502  * points to the extent item, and -EAGAIN is returned.
1503  *
1504  * NOTE: inline back refs are ordered in the same way that back ref
1505  *       items in the tree are ordered.
1506  */
1507 static noinline_for_stack
1508 int lookup_inline_extent_backref(struct btrfs_trans_handle *trans,
1509                                  struct btrfs_path *path,
1510                                  struct btrfs_extent_inline_ref **ref_ret,
1511                                  u64 bytenr, u64 num_bytes,
1512                                  u64 parent, u64 root_objectid,
1513                                  u64 owner, u64 offset, int insert)
1514 {
1515         struct btrfs_fs_info *fs_info = trans->fs_info;
1516         struct btrfs_root *root = fs_info->extent_root;
1517         struct btrfs_key key;
1518         struct extent_buffer *leaf;
1519         struct btrfs_extent_item *ei;
1520         struct btrfs_extent_inline_ref *iref;
1521         u64 flags;
1522         u64 item_size;
1523         unsigned long ptr;
1524         unsigned long end;
1525         int extra_size;
1526         int type;
1527         int want;
1528         int ret;
1529         int err = 0;
1530         bool skinny_metadata = btrfs_fs_incompat(fs_info, SKINNY_METADATA);
1531         int needed;
1532
1533         key.objectid = bytenr;
1534         key.type = BTRFS_EXTENT_ITEM_KEY;
1535         key.offset = num_bytes;
1536
1537         want = extent_ref_type(parent, owner);
1538         if (insert) {
1539                 extra_size = btrfs_extent_inline_ref_size(want);
1540                 path->keep_locks = 1;
1541         } else
1542                 extra_size = -1;
1543
1544         /*
1545          * Owner is our level, so we can just add one to get the level for the
1546          * block we are interested in.
1547          */
1548         if (skinny_metadata && owner < BTRFS_FIRST_FREE_OBJECTID) {
1549                 key.type = BTRFS_METADATA_ITEM_KEY;
1550                 key.offset = owner;
1551         }
1552
1553 again:
1554         ret = btrfs_search_slot(trans, root, &key, path, extra_size, 1);
1555         if (ret < 0) {
1556                 err = ret;
1557                 goto out;
1558         }
1559
1560         /*
1561          * We may be a newly converted file system which still has the old fat
1562          * extent entries for metadata, so try and see if we have one of those.
1563          */
1564         if (ret > 0 && skinny_metadata) {
1565                 skinny_metadata = false;
1566                 if (path->slots[0]) {
1567                         path->slots[0]--;
1568                         btrfs_item_key_to_cpu(path->nodes[0], &key,
1569                                               path->slots[0]);
1570                         if (key.objectid == bytenr &&
1571                             key.type == BTRFS_EXTENT_ITEM_KEY &&
1572                             key.offset == num_bytes)
1573                                 ret = 0;
1574                 }
1575                 if (ret) {
1576                         key.objectid = bytenr;
1577                         key.type = BTRFS_EXTENT_ITEM_KEY;
1578                         key.offset = num_bytes;
1579                         btrfs_release_path(path);
1580                         goto again;
1581                 }
1582         }
1583
1584         if (ret && !insert) {
1585                 err = -ENOENT;
1586                 goto out;
1587         } else if (WARN_ON(ret)) {
1588                 err = -EIO;
1589                 goto out;
1590         }
1591
1592         leaf = path->nodes[0];
1593         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1594         if (unlikely(item_size < sizeof(*ei))) {
1595                 err = -EINVAL;
1596                 btrfs_print_v0_err(fs_info);
1597                 btrfs_abort_transaction(trans, err);
1598                 goto out;
1599         }
1600
1601         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1602         flags = btrfs_extent_flags(leaf, ei);
1603
1604         ptr = (unsigned long)(ei + 1);
1605         end = (unsigned long)ei + item_size;
1606
1607         if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK && !skinny_metadata) {
1608                 ptr += sizeof(struct btrfs_tree_block_info);
1609                 BUG_ON(ptr > end);
1610         }
1611
1612         if (owner >= BTRFS_FIRST_FREE_OBJECTID)
1613                 needed = BTRFS_REF_TYPE_DATA;
1614         else
1615                 needed = BTRFS_REF_TYPE_BLOCK;
1616
1617         err = -ENOENT;
1618         while (1) {
1619                 if (ptr >= end) {
1620                         WARN_ON(ptr > end);
1621                         break;
1622                 }
1623                 iref = (struct btrfs_extent_inline_ref *)ptr;
1624                 type = btrfs_get_extent_inline_ref_type(leaf, iref, needed);
1625                 if (type == BTRFS_REF_TYPE_INVALID) {
1626                         err = -EUCLEAN;
1627                         goto out;
1628                 }
1629
1630                 if (want < type)
1631                         break;
1632                 if (want > type) {
1633                         ptr += btrfs_extent_inline_ref_size(type);
1634                         continue;
1635                 }
1636
1637                 if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1638                         struct btrfs_extent_data_ref *dref;
1639                         dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1640                         if (match_extent_data_ref(leaf, dref, root_objectid,
1641                                                   owner, offset)) {
1642                                 err = 0;
1643                                 break;
1644                         }
1645                         if (hash_extent_data_ref_item(leaf, dref) <
1646                             hash_extent_data_ref(root_objectid, owner, offset))
1647                                 break;
1648                 } else {
1649                         u64 ref_offset;
1650                         ref_offset = btrfs_extent_inline_ref_offset(leaf, iref);
1651                         if (parent > 0) {
1652                                 if (parent == ref_offset) {
1653                                         err = 0;
1654                                         break;
1655                                 }
1656                                 if (ref_offset < parent)
1657                                         break;
1658                         } else {
1659                                 if (root_objectid == ref_offset) {
1660                                         err = 0;
1661                                         break;
1662                                 }
1663                                 if (ref_offset < root_objectid)
1664                                         break;
1665                         }
1666                 }
1667                 ptr += btrfs_extent_inline_ref_size(type);
1668         }
1669         if (err == -ENOENT && insert) {
1670                 if (item_size + extra_size >=
1671                     BTRFS_MAX_EXTENT_ITEM_SIZE(root)) {
1672                         err = -EAGAIN;
1673                         goto out;
1674                 }
1675                 /*
1676                  * To add new inline back ref, we have to make sure
1677                  * there is no corresponding back ref item.
1678                  * For simplicity, we just do not add new inline back
1679                  * ref if there is any kind of item for this block
1680                  */
1681                 if (find_next_key(path, 0, &key) == 0 &&
1682                     key.objectid == bytenr &&
1683                     key.type < BTRFS_BLOCK_GROUP_ITEM_KEY) {
1684                         err = -EAGAIN;
1685                         goto out;
1686                 }
1687         }
1688         *ref_ret = (struct btrfs_extent_inline_ref *)ptr;
1689 out:
1690         if (insert) {
1691                 path->keep_locks = 0;
1692                 btrfs_unlock_up_safe(path, 1);
1693         }
1694         return err;
1695 }
1696
1697 /*
1698  * helper to add new inline back ref
1699  */
1700 static noinline_for_stack
1701 void setup_inline_extent_backref(struct btrfs_fs_info *fs_info,
1702                                  struct btrfs_path *path,
1703                                  struct btrfs_extent_inline_ref *iref,
1704                                  u64 parent, u64 root_objectid,
1705                                  u64 owner, u64 offset, int refs_to_add,
1706                                  struct btrfs_delayed_extent_op *extent_op)
1707 {
1708         struct extent_buffer *leaf;
1709         struct btrfs_extent_item *ei;
1710         unsigned long ptr;
1711         unsigned long end;
1712         unsigned long item_offset;
1713         u64 refs;
1714         int size;
1715         int type;
1716
1717         leaf = path->nodes[0];
1718         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1719         item_offset = (unsigned long)iref - (unsigned long)ei;
1720
1721         type = extent_ref_type(parent, owner);
1722         size = btrfs_extent_inline_ref_size(type);
1723
1724         btrfs_extend_item(path, size);
1725
1726         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1727         refs = btrfs_extent_refs(leaf, ei);
1728         refs += refs_to_add;
1729         btrfs_set_extent_refs(leaf, ei, refs);
1730         if (extent_op)
1731                 __run_delayed_extent_op(extent_op, leaf, ei);
1732
1733         ptr = (unsigned long)ei + item_offset;
1734         end = (unsigned long)ei + btrfs_item_size_nr(leaf, path->slots[0]);
1735         if (ptr < end - size)
1736                 memmove_extent_buffer(leaf, ptr + size, ptr,
1737                                       end - size - ptr);
1738
1739         iref = (struct btrfs_extent_inline_ref *)ptr;
1740         btrfs_set_extent_inline_ref_type(leaf, iref, type);
1741         if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1742                 struct btrfs_extent_data_ref *dref;
1743                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1744                 btrfs_set_extent_data_ref_root(leaf, dref, root_objectid);
1745                 btrfs_set_extent_data_ref_objectid(leaf, dref, owner);
1746                 btrfs_set_extent_data_ref_offset(leaf, dref, offset);
1747                 btrfs_set_extent_data_ref_count(leaf, dref, refs_to_add);
1748         } else if (type == BTRFS_SHARED_DATA_REF_KEY) {
1749                 struct btrfs_shared_data_ref *sref;
1750                 sref = (struct btrfs_shared_data_ref *)(iref + 1);
1751                 btrfs_set_shared_data_ref_count(leaf, sref, refs_to_add);
1752                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
1753         } else if (type == BTRFS_SHARED_BLOCK_REF_KEY) {
1754                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
1755         } else {
1756                 btrfs_set_extent_inline_ref_offset(leaf, iref, root_objectid);
1757         }
1758         btrfs_mark_buffer_dirty(leaf);
1759 }
1760
1761 static int lookup_extent_backref(struct btrfs_trans_handle *trans,
1762                                  struct btrfs_path *path,
1763                                  struct btrfs_extent_inline_ref **ref_ret,
1764                                  u64 bytenr, u64 num_bytes, u64 parent,
1765                                  u64 root_objectid, u64 owner, u64 offset)
1766 {
1767         int ret;
1768
1769         ret = lookup_inline_extent_backref(trans, path, ref_ret, bytenr,
1770                                            num_bytes, parent, root_objectid,
1771                                            owner, offset, 0);
1772         if (ret != -ENOENT)
1773                 return ret;
1774
1775         btrfs_release_path(path);
1776         *ref_ret = NULL;
1777
1778         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1779                 ret = lookup_tree_block_ref(trans, path, bytenr, parent,
1780                                             root_objectid);
1781         } else {
1782                 ret = lookup_extent_data_ref(trans, path, bytenr, parent,
1783                                              root_objectid, owner, offset);
1784         }
1785         return ret;
1786 }
1787
1788 /*
1789  * helper to update/remove inline back ref
1790  */
1791 static noinline_for_stack
1792 void update_inline_extent_backref(struct btrfs_path *path,
1793                                   struct btrfs_extent_inline_ref *iref,
1794                                   int refs_to_mod,
1795                                   struct btrfs_delayed_extent_op *extent_op,
1796                                   int *last_ref)
1797 {
1798         struct extent_buffer *leaf = path->nodes[0];
1799         struct btrfs_extent_item *ei;
1800         struct btrfs_extent_data_ref *dref = NULL;
1801         struct btrfs_shared_data_ref *sref = NULL;
1802         unsigned long ptr;
1803         unsigned long end;
1804         u32 item_size;
1805         int size;
1806         int type;
1807         u64 refs;
1808
1809         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1810         refs = btrfs_extent_refs(leaf, ei);
1811         WARN_ON(refs_to_mod < 0 && refs + refs_to_mod <= 0);
1812         refs += refs_to_mod;
1813         btrfs_set_extent_refs(leaf, ei, refs);
1814         if (extent_op)
1815                 __run_delayed_extent_op(extent_op, leaf, ei);
1816
1817         /*
1818          * If type is invalid, we should have bailed out after
1819          * lookup_inline_extent_backref().
1820          */
1821         type = btrfs_get_extent_inline_ref_type(leaf, iref, BTRFS_REF_TYPE_ANY);
1822         ASSERT(type != BTRFS_REF_TYPE_INVALID);
1823
1824         if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1825                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1826                 refs = btrfs_extent_data_ref_count(leaf, dref);
1827         } else if (type == BTRFS_SHARED_DATA_REF_KEY) {
1828                 sref = (struct btrfs_shared_data_ref *)(iref + 1);
1829                 refs = btrfs_shared_data_ref_count(leaf, sref);
1830         } else {
1831                 refs = 1;
1832                 BUG_ON(refs_to_mod != -1);
1833         }
1834
1835         BUG_ON(refs_to_mod < 0 && refs < -refs_to_mod);
1836         refs += refs_to_mod;
1837
1838         if (refs > 0) {
1839                 if (type == BTRFS_EXTENT_DATA_REF_KEY)
1840                         btrfs_set_extent_data_ref_count(leaf, dref, refs);
1841                 else
1842                         btrfs_set_shared_data_ref_count(leaf, sref, refs);
1843         } else {
1844                 *last_ref = 1;
1845                 size =  btrfs_extent_inline_ref_size(type);
1846                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1847                 ptr = (unsigned long)iref;
1848                 end = (unsigned long)ei + item_size;
1849                 if (ptr + size < end)
1850                         memmove_extent_buffer(leaf, ptr, ptr + size,
1851                                               end - ptr - size);
1852                 item_size -= size;
1853                 btrfs_truncate_item(path, item_size, 1);
1854         }
1855         btrfs_mark_buffer_dirty(leaf);
1856 }
1857
1858 static noinline_for_stack
1859 int insert_inline_extent_backref(struct btrfs_trans_handle *trans,
1860                                  struct btrfs_path *path,
1861                                  u64 bytenr, u64 num_bytes, u64 parent,
1862                                  u64 root_objectid, u64 owner,
1863                                  u64 offset, int refs_to_add,
1864                                  struct btrfs_delayed_extent_op *extent_op)
1865 {
1866         struct btrfs_extent_inline_ref *iref;
1867         int ret;
1868
1869         ret = lookup_inline_extent_backref(trans, path, &iref, bytenr,
1870                                            num_bytes, parent, root_objectid,
1871                                            owner, offset, 1);
1872         if (ret == 0) {
1873                 BUG_ON(owner < BTRFS_FIRST_FREE_OBJECTID);
1874                 update_inline_extent_backref(path, iref, refs_to_add,
1875                                              extent_op, NULL);
1876         } else if (ret == -ENOENT) {
1877                 setup_inline_extent_backref(trans->fs_info, path, iref, parent,
1878                                             root_objectid, owner, offset,
1879                                             refs_to_add, extent_op);
1880                 ret = 0;
1881         }
1882         return ret;
1883 }
1884
1885 static int insert_extent_backref(struct btrfs_trans_handle *trans,
1886                                  struct btrfs_path *path,
1887                                  u64 bytenr, u64 parent, u64 root_objectid,
1888                                  u64 owner, u64 offset, int refs_to_add)
1889 {
1890         int ret;
1891         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1892                 BUG_ON(refs_to_add != 1);
1893                 ret = insert_tree_block_ref(trans, path, bytenr, parent,
1894                                             root_objectid);
1895         } else {
1896                 ret = insert_extent_data_ref(trans, path, bytenr, parent,
1897                                              root_objectid, owner, offset,
1898                                              refs_to_add);
1899         }
1900         return ret;
1901 }
1902
1903 static int remove_extent_backref(struct btrfs_trans_handle *trans,
1904                                  struct btrfs_path *path,
1905                                  struct btrfs_extent_inline_ref *iref,
1906                                  int refs_to_drop, int is_data, int *last_ref)
1907 {
1908         int ret = 0;
1909
1910         BUG_ON(!is_data && refs_to_drop != 1);
1911         if (iref) {
1912                 update_inline_extent_backref(path, iref, -refs_to_drop, NULL,
1913                                              last_ref);
1914         } else if (is_data) {
1915                 ret = remove_extent_data_ref(trans, path, refs_to_drop,
1916                                              last_ref);
1917         } else {
1918                 *last_ref = 1;
1919                 ret = btrfs_del_item(trans, trans->fs_info->extent_root, path);
1920         }
1921         return ret;
1922 }
1923
1924 static int btrfs_issue_discard(struct block_device *bdev, u64 start, u64 len,
1925                                u64 *discarded_bytes)
1926 {
1927         int j, ret = 0;
1928         u64 bytes_left, end;
1929         u64 aligned_start = ALIGN(start, 1 << 9);
1930
1931         if (WARN_ON(start != aligned_start)) {
1932                 len -= aligned_start - start;
1933                 len = round_down(len, 1 << 9);
1934                 start = aligned_start;
1935         }
1936
1937         *discarded_bytes = 0;
1938
1939         if (!len)
1940                 return 0;
1941
1942         end = start + len;
1943         bytes_left = len;
1944
1945         /* Skip any superblocks on this device. */
1946         for (j = 0; j < BTRFS_SUPER_MIRROR_MAX; j++) {
1947                 u64 sb_start = btrfs_sb_offset(j);
1948                 u64 sb_end = sb_start + BTRFS_SUPER_INFO_SIZE;
1949                 u64 size = sb_start - start;
1950
1951                 if (!in_range(sb_start, start, bytes_left) &&
1952                     !in_range(sb_end, start, bytes_left) &&
1953                     !in_range(start, sb_start, BTRFS_SUPER_INFO_SIZE))
1954                         continue;
1955
1956                 /*
1957                  * Superblock spans beginning of range.  Adjust start and
1958                  * try again.
1959                  */
1960                 if (sb_start <= start) {
1961                         start += sb_end - start;
1962                         if (start > end) {
1963                                 bytes_left = 0;
1964                                 break;
1965                         }
1966                         bytes_left = end - start;
1967                         continue;
1968                 }
1969
1970                 if (size) {
1971                         ret = blkdev_issue_discard(bdev, start >> 9, size >> 9,
1972                                                    GFP_NOFS, 0);
1973                         if (!ret)
1974                                 *discarded_bytes += size;
1975                         else if (ret != -EOPNOTSUPP)
1976                                 return ret;
1977                 }
1978
1979                 start = sb_end;
1980                 if (start > end) {
1981                         bytes_left = 0;
1982                         break;
1983                 }
1984                 bytes_left = end - start;
1985         }
1986
1987         if (bytes_left) {
1988                 ret = blkdev_issue_discard(bdev, start >> 9, bytes_left >> 9,
1989                                            GFP_NOFS, 0);
1990                 if (!ret)
1991                         *discarded_bytes += bytes_left;
1992         }
1993         return ret;
1994 }
1995
1996 int btrfs_discard_extent(struct btrfs_fs_info *fs_info, u64 bytenr,
1997                          u64 num_bytes, u64 *actual_bytes)
1998 {
1999         int ret;
2000         u64 discarded_bytes = 0;
2001         struct btrfs_bio *bbio = NULL;
2002
2003
2004         /*
2005          * Avoid races with device replace and make sure our bbio has devices
2006          * associated to its stripes that don't go away while we are discarding.
2007          */
2008         btrfs_bio_counter_inc_blocked(fs_info);
2009         /* Tell the block device(s) that the sectors can be discarded */
2010         ret = btrfs_map_block(fs_info, BTRFS_MAP_DISCARD, bytenr, &num_bytes,
2011                               &bbio, 0);
2012         /* Error condition is -ENOMEM */
2013         if (!ret) {
2014                 struct btrfs_bio_stripe *stripe = bbio->stripes;
2015                 int i;
2016
2017
2018                 for (i = 0; i < bbio->num_stripes; i++, stripe++) {
2019                         u64 bytes;
2020                         struct request_queue *req_q;
2021
2022                         if (!stripe->dev->bdev) {
2023                                 ASSERT(btrfs_test_opt(fs_info, DEGRADED));
2024                                 continue;
2025                         }
2026                         req_q = bdev_get_queue(stripe->dev->bdev);
2027                         if (!blk_queue_discard(req_q))
2028                                 continue;
2029
2030                         ret = btrfs_issue_discard(stripe->dev->bdev,
2031                                                   stripe->physical,
2032                                                   stripe->length,
2033                                                   &bytes);
2034                         if (!ret)
2035                                 discarded_bytes += bytes;
2036                         else if (ret != -EOPNOTSUPP)
2037                                 break; /* Logic errors or -ENOMEM, or -EIO but I don't know how that could happen JDM */
2038
2039                         /*
2040                          * Just in case we get back EOPNOTSUPP for some reason,
2041                          * just ignore the return value so we don't screw up
2042                          * people calling discard_extent.
2043                          */
2044                         ret = 0;
2045                 }
2046                 btrfs_put_bbio(bbio);
2047         }
2048         btrfs_bio_counter_dec(fs_info);
2049
2050         if (actual_bytes)
2051                 *actual_bytes = discarded_bytes;
2052
2053
2054         if (ret == -EOPNOTSUPP)
2055                 ret = 0;
2056         return ret;
2057 }
2058
2059 /* Can return -ENOMEM */
2060 int btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
2061                          struct btrfs_ref *generic_ref)
2062 {
2063         struct btrfs_fs_info *fs_info = trans->fs_info;
2064         int old_ref_mod, new_ref_mod;
2065         int ret;
2066
2067         ASSERT(generic_ref->type != BTRFS_REF_NOT_SET &&
2068                generic_ref->action);
2069         BUG_ON(generic_ref->type == BTRFS_REF_METADATA &&
2070                generic_ref->tree_ref.root == BTRFS_TREE_LOG_OBJECTID);
2071
2072         if (generic_ref->type == BTRFS_REF_METADATA)
2073                 ret = btrfs_add_delayed_tree_ref(trans, generic_ref,
2074                                 NULL, &old_ref_mod, &new_ref_mod);
2075         else
2076                 ret = btrfs_add_delayed_data_ref(trans, generic_ref, 0,
2077                                                  &old_ref_mod, &new_ref_mod);
2078
2079         btrfs_ref_tree_mod(fs_info, generic_ref);
2080
2081         if (ret == 0 && old_ref_mod < 0 && new_ref_mod >= 0)
2082                 sub_pinned_bytes(fs_info, generic_ref);
2083
2084         return ret;
2085 }
2086
2087 /*
2088  * __btrfs_inc_extent_ref - insert backreference for a given extent
2089  *
2090  * @trans:          Handle of transaction
2091  *
2092  * @node:           The delayed ref node used to get the bytenr/length for
2093  *                  extent whose references are incremented.
2094  *
2095  * @parent:         If this is a shared extent (BTRFS_SHARED_DATA_REF_KEY/
2096  *                  BTRFS_SHARED_BLOCK_REF_KEY) then it holds the logical
2097  *                  bytenr of the parent block. Since new extents are always
2098  *                  created with indirect references, this will only be the case
2099  *                  when relocating a shared extent. In that case, root_objectid
2100  *                  will be BTRFS_TREE_RELOC_OBJECTID. Otheriwse, parent must
2101  *                  be 0
2102  *
2103  * @root_objectid:  The id of the root where this modification has originated,
2104  *                  this can be either one of the well-known metadata trees or
2105  *                  the subvolume id which references this extent.
2106  *
2107  * @owner:          For data extents it is the inode number of the owning file.
2108  *                  For metadata extents this parameter holds the level in the
2109  *                  tree of the extent.
2110  *
2111  * @offset:         For metadata extents the offset is ignored and is currently
2112  *                  always passed as 0. For data extents it is the fileoffset
2113  *                  this extent belongs to.
2114  *
2115  * @refs_to_add     Number of references to add
2116  *
2117  * @extent_op       Pointer to a structure, holding information necessary when
2118  *                  updating a tree block's flags
2119  *
2120  */
2121 static int __btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
2122                                   struct btrfs_delayed_ref_node *node,
2123                                   u64 parent, u64 root_objectid,
2124                                   u64 owner, u64 offset, int refs_to_add,
2125                                   struct btrfs_delayed_extent_op *extent_op)
2126 {
2127         struct btrfs_path *path;
2128         struct extent_buffer *leaf;
2129         struct btrfs_extent_item *item;
2130         struct btrfs_key key;
2131         u64 bytenr = node->bytenr;
2132         u64 num_bytes = node->num_bytes;
2133         u64 refs;
2134         int ret;
2135
2136         path = btrfs_alloc_path();
2137         if (!path)
2138                 return -ENOMEM;
2139
2140         path->reada = READA_FORWARD;
2141         path->leave_spinning = 1;
2142         /* this will setup the path even if it fails to insert the back ref */
2143         ret = insert_inline_extent_backref(trans, path, bytenr, num_bytes,
2144                                            parent, root_objectid, owner,
2145                                            offset, refs_to_add, extent_op);
2146         if ((ret < 0 && ret != -EAGAIN) || !ret)
2147                 goto out;
2148
2149         /*
2150          * Ok we had -EAGAIN which means we didn't have space to insert and
2151          * inline extent ref, so just update the reference count and add a
2152          * normal backref.
2153          */
2154         leaf = path->nodes[0];
2155         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
2156         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2157         refs = btrfs_extent_refs(leaf, item);
2158         btrfs_set_extent_refs(leaf, item, refs + refs_to_add);
2159         if (extent_op)
2160                 __run_delayed_extent_op(extent_op, leaf, item);
2161
2162         btrfs_mark_buffer_dirty(leaf);
2163         btrfs_release_path(path);
2164
2165         path->reada = READA_FORWARD;
2166         path->leave_spinning = 1;
2167         /* now insert the actual backref */
2168         ret = insert_extent_backref(trans, path, bytenr, parent, root_objectid,
2169                                     owner, offset, refs_to_add);
2170         if (ret)
2171                 btrfs_abort_transaction(trans, ret);
2172 out:
2173         btrfs_free_path(path);
2174         return ret;
2175 }
2176
2177 static int run_delayed_data_ref(struct btrfs_trans_handle *trans,
2178                                 struct btrfs_delayed_ref_node *node,
2179                                 struct btrfs_delayed_extent_op *extent_op,
2180                                 int insert_reserved)
2181 {
2182         int ret = 0;
2183         struct btrfs_delayed_data_ref *ref;
2184         struct btrfs_key ins;
2185         u64 parent = 0;
2186         u64 ref_root = 0;
2187         u64 flags = 0;
2188
2189         ins.objectid = node->bytenr;
2190         ins.offset = node->num_bytes;
2191         ins.type = BTRFS_EXTENT_ITEM_KEY;
2192
2193         ref = btrfs_delayed_node_to_data_ref(node);
2194         trace_run_delayed_data_ref(trans->fs_info, node, ref, node->action);
2195
2196         if (node->type == BTRFS_SHARED_DATA_REF_KEY)
2197                 parent = ref->parent;
2198         ref_root = ref->root;
2199
2200         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
2201                 if (extent_op)
2202                         flags |= extent_op->flags_to_set;
2203                 ret = alloc_reserved_file_extent(trans, parent, ref_root,
2204                                                  flags, ref->objectid,
2205                                                  ref->offset, &ins,
2206                                                  node->ref_mod);
2207         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
2208                 ret = __btrfs_inc_extent_ref(trans, node, parent, ref_root,
2209                                              ref->objectid, ref->offset,
2210                                              node->ref_mod, extent_op);
2211         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
2212                 ret = __btrfs_free_extent(trans, node, parent,
2213                                           ref_root, ref->objectid,
2214                                           ref->offset, node->ref_mod,
2215                                           extent_op);
2216         } else {
2217                 BUG();
2218         }
2219         return ret;
2220 }
2221
2222 static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op,
2223                                     struct extent_buffer *leaf,
2224                                     struct btrfs_extent_item *ei)
2225 {
2226         u64 flags = btrfs_extent_flags(leaf, ei);
2227         if (extent_op->update_flags) {
2228                 flags |= extent_op->flags_to_set;
2229                 btrfs_set_extent_flags(leaf, ei, flags);
2230         }
2231
2232         if (extent_op->update_key) {
2233                 struct btrfs_tree_block_info *bi;
2234                 BUG_ON(!(flags & BTRFS_EXTENT_FLAG_TREE_BLOCK));
2235                 bi = (struct btrfs_tree_block_info *)(ei + 1);
2236                 btrfs_set_tree_block_key(leaf, bi, &extent_op->key);
2237         }
2238 }
2239
2240 static int run_delayed_extent_op(struct btrfs_trans_handle *trans,
2241                                  struct btrfs_delayed_ref_head *head,
2242                                  struct btrfs_delayed_extent_op *extent_op)
2243 {
2244         struct btrfs_fs_info *fs_info = trans->fs_info;
2245         struct btrfs_key key;
2246         struct btrfs_path *path;
2247         struct btrfs_extent_item *ei;
2248         struct extent_buffer *leaf;
2249         u32 item_size;
2250         int ret;
2251         int err = 0;
2252         int metadata = !extent_op->is_data;
2253
2254         if (trans->aborted)
2255                 return 0;
2256
2257         if (metadata && !btrfs_fs_incompat(fs_info, SKINNY_METADATA))
2258                 metadata = 0;
2259
2260         path = btrfs_alloc_path();
2261         if (!path)
2262                 return -ENOMEM;
2263
2264         key.objectid = head->bytenr;
2265
2266         if (metadata) {
2267                 key.type = BTRFS_METADATA_ITEM_KEY;
2268                 key.offset = extent_op->level;
2269         } else {
2270                 key.type = BTRFS_EXTENT_ITEM_KEY;
2271                 key.offset = head->num_bytes;
2272         }
2273
2274 again:
2275         path->reada = READA_FORWARD;
2276         path->leave_spinning = 1;
2277         ret = btrfs_search_slot(trans, fs_info->extent_root, &key, path, 0, 1);
2278         if (ret < 0) {
2279                 err = ret;
2280                 goto out;
2281         }
2282         if (ret > 0) {
2283                 if (metadata) {
2284                         if (path->slots[0] > 0) {
2285                                 path->slots[0]--;
2286                                 btrfs_item_key_to_cpu(path->nodes[0], &key,
2287                                                       path->slots[0]);
2288                                 if (key.objectid == head->bytenr &&
2289                                     key.type == BTRFS_EXTENT_ITEM_KEY &&
2290                                     key.offset == head->num_bytes)
2291                                         ret = 0;
2292                         }
2293                         if (ret > 0) {
2294                                 btrfs_release_path(path);
2295                                 metadata = 0;
2296
2297                                 key.objectid = head->bytenr;
2298                                 key.offset = head->num_bytes;
2299                                 key.type = BTRFS_EXTENT_ITEM_KEY;
2300                                 goto again;
2301                         }
2302                 } else {
2303                         err = -EIO;
2304                         goto out;
2305                 }
2306         }
2307
2308         leaf = path->nodes[0];
2309         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2310
2311         if (unlikely(item_size < sizeof(*ei))) {
2312                 err = -EINVAL;
2313                 btrfs_print_v0_err(fs_info);
2314                 btrfs_abort_transaction(trans, err);
2315                 goto out;
2316         }
2317
2318         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2319         __run_delayed_extent_op(extent_op, leaf, ei);
2320
2321         btrfs_mark_buffer_dirty(leaf);
2322 out:
2323         btrfs_free_path(path);
2324         return err;
2325 }
2326
2327 static int run_delayed_tree_ref(struct btrfs_trans_handle *trans,
2328                                 struct btrfs_delayed_ref_node *node,
2329                                 struct btrfs_delayed_extent_op *extent_op,
2330                                 int insert_reserved)
2331 {
2332         int ret = 0;
2333         struct btrfs_delayed_tree_ref *ref;
2334         u64 parent = 0;
2335         u64 ref_root = 0;
2336
2337         ref = btrfs_delayed_node_to_tree_ref(node);
2338         trace_run_delayed_tree_ref(trans->fs_info, node, ref, node->action);
2339
2340         if (node->type == BTRFS_SHARED_BLOCK_REF_KEY)
2341                 parent = ref->parent;
2342         ref_root = ref->root;
2343
2344         if (node->ref_mod != 1) {
2345                 btrfs_err(trans->fs_info,
2346         "btree block(%llu) has %d references rather than 1: action %d ref_root %llu parent %llu",
2347                           node->bytenr, node->ref_mod, node->action, ref_root,
2348                           parent);
2349                 return -EIO;
2350         }
2351         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
2352                 BUG_ON(!extent_op || !extent_op->update_flags);
2353                 ret = alloc_reserved_tree_block(trans, node, extent_op);
2354         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
2355                 ret = __btrfs_inc_extent_ref(trans, node, parent, ref_root,
2356                                              ref->level, 0, 1, extent_op);
2357         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
2358                 ret = __btrfs_free_extent(trans, node, parent, ref_root,
2359                                           ref->level, 0, 1, extent_op);
2360         } else {
2361                 BUG();
2362         }
2363         return ret;
2364 }
2365
2366 /* helper function to actually process a single delayed ref entry */
2367 static int run_one_delayed_ref(struct btrfs_trans_handle *trans,
2368                                struct btrfs_delayed_ref_node *node,
2369                                struct btrfs_delayed_extent_op *extent_op,
2370                                int insert_reserved)
2371 {
2372         int ret = 0;
2373
2374         if (trans->aborted) {
2375                 if (insert_reserved)
2376                         btrfs_pin_extent(trans->fs_info, node->bytenr,
2377                                          node->num_bytes, 1);
2378                 return 0;
2379         }
2380
2381         if (node->type == BTRFS_TREE_BLOCK_REF_KEY ||
2382             node->type == BTRFS_SHARED_BLOCK_REF_KEY)
2383                 ret = run_delayed_tree_ref(trans, node, extent_op,
2384                                            insert_reserved);
2385         else if (node->type == BTRFS_EXTENT_DATA_REF_KEY ||
2386                  node->type == BTRFS_SHARED_DATA_REF_KEY)
2387                 ret = run_delayed_data_ref(trans, node, extent_op,
2388                                            insert_reserved);
2389         else
2390                 BUG();
2391         if (ret && insert_reserved)
2392                 btrfs_pin_extent(trans->fs_info, node->bytenr,
2393                                  node->num_bytes, 1);
2394         return ret;
2395 }
2396
2397 static inline struct btrfs_delayed_ref_node *
2398 select_delayed_ref(struct btrfs_delayed_ref_head *head)
2399 {
2400         struct btrfs_delayed_ref_node *ref;
2401
2402         if (RB_EMPTY_ROOT(&head->ref_tree.rb_root))
2403                 return NULL;
2404
2405         /*
2406          * Select a delayed ref of type BTRFS_ADD_DELAYED_REF first.
2407          * This is to prevent a ref count from going down to zero, which deletes
2408          * the extent item from the extent tree, when there still are references
2409          * to add, which would fail because they would not find the extent item.
2410          */
2411         if (!list_empty(&head->ref_add_list))
2412                 return list_first_entry(&head->ref_add_list,
2413                                 struct btrfs_delayed_ref_node, add_list);
2414
2415         ref = rb_entry(rb_first_cached(&head->ref_tree),
2416                        struct btrfs_delayed_ref_node, ref_node);
2417         ASSERT(list_empty(&ref->add_list));
2418         return ref;
2419 }
2420
2421 static void unselect_delayed_ref_head(struct btrfs_delayed_ref_root *delayed_refs,
2422                                       struct btrfs_delayed_ref_head *head)
2423 {
2424         spin_lock(&delayed_refs->lock);
2425         head->processing = 0;
2426         delayed_refs->num_heads_ready++;
2427         spin_unlock(&delayed_refs->lock);
2428         btrfs_delayed_ref_unlock(head);
2429 }
2430
2431 static struct btrfs_delayed_extent_op *cleanup_extent_op(
2432                                 struct btrfs_delayed_ref_head *head)
2433 {
2434         struct btrfs_delayed_extent_op *extent_op = head->extent_op;
2435
2436         if (!extent_op)
2437                 return NULL;
2438
2439         if (head->must_insert_reserved) {
2440                 head->extent_op = NULL;
2441                 btrfs_free_delayed_extent_op(extent_op);
2442                 return NULL;
2443         }
2444         return extent_op;
2445 }
2446
2447 static int run_and_cleanup_extent_op(struct btrfs_trans_handle *trans,
2448                                      struct btrfs_delayed_ref_head *head)
2449 {
2450         struct btrfs_delayed_extent_op *extent_op;
2451         int ret;
2452
2453         extent_op = cleanup_extent_op(head);
2454         if (!extent_op)
2455                 return 0;
2456         head->extent_op = NULL;
2457         spin_unlock(&head->lock);
2458         ret = run_delayed_extent_op(trans, head, extent_op);
2459         btrfs_free_delayed_extent_op(extent_op);
2460         return ret ? ret : 1;
2461 }
2462
2463 void btrfs_cleanup_ref_head_accounting(struct btrfs_fs_info *fs_info,
2464                                   struct btrfs_delayed_ref_root *delayed_refs,
2465                                   struct btrfs_delayed_ref_head *head)
2466 {
2467         int nr_items = 1;       /* Dropping this ref head update. */
2468
2469         if (head->total_ref_mod < 0) {
2470                 struct btrfs_space_info *space_info;
2471                 u64 flags;
2472
2473                 if (head->is_data)
2474                         flags = BTRFS_BLOCK_GROUP_DATA;
2475                 else if (head->is_system)
2476                         flags = BTRFS_BLOCK_GROUP_SYSTEM;
2477                 else
2478                         flags = BTRFS_BLOCK_GROUP_METADATA;
2479                 space_info = __find_space_info(fs_info, flags);
2480                 ASSERT(space_info);
2481                 percpu_counter_add_batch(&space_info->total_bytes_pinned,
2482                                    -head->num_bytes,
2483                                    BTRFS_TOTAL_BYTES_PINNED_BATCH);
2484
2485                 /*
2486                  * We had csum deletions accounted for in our delayed refs rsv,
2487                  * we need to drop the csum leaves for this update from our
2488                  * delayed_refs_rsv.
2489                  */
2490                 if (head->is_data) {
2491                         spin_lock(&delayed_refs->lock);
2492                         delayed_refs->pending_csums -= head->num_bytes;
2493                         spin_unlock(&delayed_refs->lock);
2494                         nr_items += btrfs_csum_bytes_to_leaves(fs_info,
2495                                 head->num_bytes);
2496                 }
2497         }
2498
2499         btrfs_delayed_refs_rsv_release(fs_info, nr_items);
2500 }
2501
2502 static int cleanup_ref_head(struct btrfs_trans_handle *trans,
2503                             struct btrfs_delayed_ref_head *head)
2504 {
2505
2506         struct btrfs_fs_info *fs_info = trans->fs_info;
2507         struct btrfs_delayed_ref_root *delayed_refs;
2508         int ret;
2509
2510         delayed_refs = &trans->transaction->delayed_refs;
2511
2512         ret = run_and_cleanup_extent_op(trans, head);
2513         if (ret < 0) {
2514                 unselect_delayed_ref_head(delayed_refs, head);
2515                 btrfs_debug(fs_info, "run_delayed_extent_op returned %d", ret);
2516                 return ret;
2517         } else if (ret) {
2518                 return ret;
2519         }
2520
2521         /*
2522          * Need to drop our head ref lock and re-acquire the delayed ref lock
2523          * and then re-check to make sure nobody got added.
2524          */
2525         spin_unlock(&head->lock);
2526         spin_lock(&delayed_refs->lock);
2527         spin_lock(&head->lock);
2528         if (!RB_EMPTY_ROOT(&head->ref_tree.rb_root) || head->extent_op) {
2529                 spin_unlock(&head->lock);
2530                 spin_unlock(&delayed_refs->lock);
2531                 return 1;
2532         }
2533         btrfs_delete_ref_head(delayed_refs, head);
2534         spin_unlock(&head->lock);
2535         spin_unlock(&delayed_refs->lock);
2536
2537         if (head->must_insert_reserved) {
2538                 btrfs_pin_extent(fs_info, head->bytenr,
2539                                  head->num_bytes, 1);
2540                 if (head->is_data) {
2541                         ret = btrfs_del_csums(trans, fs_info, head->bytenr,
2542                                               head->num_bytes);
2543                 }
2544         }
2545
2546         btrfs_cleanup_ref_head_accounting(fs_info, delayed_refs, head);
2547
2548         trace_run_delayed_ref_head(fs_info, head, 0);
2549         btrfs_delayed_ref_unlock(head);
2550         btrfs_put_delayed_ref_head(head);
2551         return 0;
2552 }
2553
2554 static struct btrfs_delayed_ref_head *btrfs_obtain_ref_head(
2555                                         struct btrfs_trans_handle *trans)
2556 {
2557         struct btrfs_delayed_ref_root *delayed_refs =
2558                 &trans->transaction->delayed_refs;
2559         struct btrfs_delayed_ref_head *head = NULL;
2560         int ret;
2561
2562         spin_lock(&delayed_refs->lock);
2563         head = btrfs_select_ref_head(delayed_refs);
2564         if (!head) {
2565                 spin_unlock(&delayed_refs->lock);
2566                 return head;
2567         }
2568
2569         /*
2570          * Grab the lock that says we are going to process all the refs for
2571          * this head
2572          */
2573         ret = btrfs_delayed_ref_lock(delayed_refs, head);
2574         spin_unlock(&delayed_refs->lock);
2575
2576         /*
2577          * We may have dropped the spin lock to get the head mutex lock, and
2578          * that might have given someone else time to free the head.  If that's
2579          * true, it has been removed from our list and we can move on.
2580          */
2581         if (ret == -EAGAIN)
2582                 head = ERR_PTR(-EAGAIN);
2583
2584         return head;
2585 }
2586
2587 static int btrfs_run_delayed_refs_for_head(struct btrfs_trans_handle *trans,
2588                                     struct btrfs_delayed_ref_head *locked_ref,
2589                                     unsigned long *run_refs)
2590 {
2591         struct btrfs_fs_info *fs_info = trans->fs_info;
2592         struct btrfs_delayed_ref_root *delayed_refs;
2593         struct btrfs_delayed_extent_op *extent_op;
2594         struct btrfs_delayed_ref_node *ref;
2595         int must_insert_reserved = 0;
2596         int ret;
2597
2598         delayed_refs = &trans->transaction->delayed_refs;
2599
2600         lockdep_assert_held(&locked_ref->mutex);
2601         lockdep_assert_held(&locked_ref->lock);
2602
2603         while ((ref = select_delayed_ref(locked_ref))) {
2604                 if (ref->seq &&
2605                     btrfs_check_delayed_seq(fs_info, ref->seq)) {
2606                         spin_unlock(&locked_ref->lock);
2607                         unselect_delayed_ref_head(delayed_refs, locked_ref);
2608                         return -EAGAIN;
2609                 }
2610
2611                 (*run_refs)++;
2612                 ref->in_tree = 0;
2613                 rb_erase_cached(&ref->ref_node, &locked_ref->ref_tree);
2614                 RB_CLEAR_NODE(&ref->ref_node);
2615                 if (!list_empty(&ref->add_list))
2616                         list_del(&ref->add_list);
2617                 /*
2618                  * When we play the delayed ref, also correct the ref_mod on
2619                  * head
2620                  */
2621                 switch (ref->action) {
2622                 case BTRFS_ADD_DELAYED_REF:
2623                 case BTRFS_ADD_DELAYED_EXTENT:
2624                         locked_ref->ref_mod -= ref->ref_mod;
2625                         break;
2626                 case BTRFS_DROP_DELAYED_REF:
2627                         locked_ref->ref_mod += ref->ref_mod;
2628                         break;
2629                 default:
2630                         WARN_ON(1);
2631                 }
2632                 atomic_dec(&delayed_refs->num_entries);
2633
2634                 /*
2635                  * Record the must_insert_reserved flag before we drop the
2636                  * spin lock.
2637                  */
2638                 must_insert_reserved = locked_ref->must_insert_reserved;
2639                 locked_ref->must_insert_reserved = 0;
2640
2641                 extent_op = locked_ref->extent_op;
2642                 locked_ref->extent_op = NULL;
2643                 spin_unlock(&locked_ref->lock);
2644
2645                 ret = run_one_delayed_ref(trans, ref, extent_op,
2646                                           must_insert_reserved);
2647
2648                 btrfs_free_delayed_extent_op(extent_op);
2649                 if (ret) {
2650                         unselect_delayed_ref_head(delayed_refs, locked_ref);
2651                         btrfs_put_delayed_ref(ref);
2652                         btrfs_debug(fs_info, "run_one_delayed_ref returned %d",
2653                                     ret);
2654                         return ret;
2655                 }
2656
2657                 btrfs_put_delayed_ref(ref);
2658                 cond_resched();
2659
2660                 spin_lock(&locked_ref->lock);
2661                 btrfs_merge_delayed_refs(trans, delayed_refs, locked_ref);
2662         }
2663
2664         return 0;
2665 }
2666
2667 /*
2668  * Returns 0 on success or if called with an already aborted transaction.
2669  * Returns -ENOMEM or -EIO on failure and will abort the transaction.
2670  */
2671 static noinline int __btrfs_run_delayed_refs(struct btrfs_trans_handle *trans,
2672                                              unsigned long nr)
2673 {
2674         struct btrfs_fs_info *fs_info = trans->fs_info;
2675         struct btrfs_delayed_ref_root *delayed_refs;
2676         struct btrfs_delayed_ref_head *locked_ref = NULL;
2677         ktime_t start = ktime_get();
2678         int ret;
2679         unsigned long count = 0;
2680         unsigned long actual_count = 0;
2681
2682         delayed_refs = &trans->transaction->delayed_refs;
2683         do {
2684                 if (!locked_ref) {
2685                         locked_ref = btrfs_obtain_ref_head(trans);
2686                         if (IS_ERR_OR_NULL(locked_ref)) {
2687                                 if (PTR_ERR(locked_ref) == -EAGAIN) {
2688                                         continue;
2689                                 } else {
2690                                         break;
2691                                 }
2692                         }
2693                         count++;
2694                 }
2695                 /*
2696                  * We need to try and merge add/drops of the same ref since we
2697                  * can run into issues with relocate dropping the implicit ref
2698                  * and then it being added back again before the drop can
2699                  * finish.  If we merged anything we need to re-loop so we can
2700                  * get a good ref.
2701                  * Or we can get node references of the same type that weren't
2702                  * merged when created due to bumps in the tree mod seq, and
2703                  * we need to merge them to prevent adding an inline extent
2704                  * backref before dropping it (triggering a BUG_ON at
2705                  * insert_inline_extent_backref()).
2706                  */
2707                 spin_lock(&locked_ref->lock);
2708                 btrfs_merge_delayed_refs(trans, delayed_refs, locked_ref);
2709
2710                 ret = btrfs_run_delayed_refs_for_head(trans, locked_ref,
2711                                                       &actual_count);
2712                 if (ret < 0 && ret != -EAGAIN) {
2713                         /*
2714                          * Error, btrfs_run_delayed_refs_for_head already
2715                          * unlocked everything so just bail out
2716                          */
2717                         return ret;
2718                 } else if (!ret) {
2719                         /*
2720                          * Success, perform the usual cleanup of a processed
2721                          * head
2722                          */
2723                         ret = cleanup_ref_head(trans, locked_ref);
2724                         if (ret > 0 ) {
2725                                 /* We dropped our lock, we need to loop. */
2726                                 ret = 0;
2727                                 continue;
2728                         } else if (ret) {
2729                                 return ret;
2730                         }
2731                 }
2732
2733                 /*
2734                  * Either success case or btrfs_run_delayed_refs_for_head
2735                  * returned -EAGAIN, meaning we need to select another head
2736                  */
2737
2738                 locked_ref = NULL;
2739                 cond_resched();
2740         } while ((nr != -1 && count < nr) || locked_ref);
2741
2742         /*
2743          * We don't want to include ref heads since we can have empty ref heads
2744          * and those will drastically skew our runtime down since we just do
2745          * accounting, no actual extent tree updates.
2746          */
2747         if (actual_count > 0) {
2748                 u64 runtime = ktime_to_ns(ktime_sub(ktime_get(), start));
2749                 u64 avg;
2750
2751                 /*
2752                  * We weigh the current average higher than our current runtime
2753                  * to avoid large swings in the average.
2754                  */
2755                 spin_lock(&delayed_refs->lock);
2756                 avg = fs_info->avg_delayed_ref_runtime * 3 + runtime;
2757                 fs_info->avg_delayed_ref_runtime = avg >> 2;    /* div by 4 */
2758                 spin_unlock(&delayed_refs->lock);
2759         }
2760         return 0;
2761 }
2762
2763 #ifdef SCRAMBLE_DELAYED_REFS
2764 /*
2765  * Normally delayed refs get processed in ascending bytenr order. This
2766  * correlates in most cases to the order added. To expose dependencies on this
2767  * order, we start to process the tree in the middle instead of the beginning
2768  */
2769 static u64 find_middle(struct rb_root *root)
2770 {
2771         struct rb_node *n = root->rb_node;
2772         struct btrfs_delayed_ref_node *entry;
2773         int alt = 1;
2774         u64 middle;
2775         u64 first = 0, last = 0;
2776
2777         n = rb_first(root);
2778         if (n) {
2779                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2780                 first = entry->bytenr;
2781         }
2782         n = rb_last(root);
2783         if (n) {
2784                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2785                 last = entry->bytenr;
2786         }
2787         n = root->rb_node;
2788
2789         while (n) {
2790                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2791                 WARN_ON(!entry->in_tree);
2792
2793                 middle = entry->bytenr;
2794
2795                 if (alt)
2796                         n = n->rb_left;
2797                 else
2798                         n = n->rb_right;
2799
2800                 alt = 1 - alt;
2801         }
2802         return middle;
2803 }
2804 #endif
2805
2806 static inline u64 heads_to_leaves(struct btrfs_fs_info *fs_info, u64 heads)
2807 {
2808         u64 num_bytes;
2809
2810         num_bytes = heads * (sizeof(struct btrfs_extent_item) +
2811                              sizeof(struct btrfs_extent_inline_ref));
2812         if (!btrfs_fs_incompat(fs_info, SKINNY_METADATA))
2813                 num_bytes += heads * sizeof(struct btrfs_tree_block_info);
2814
2815         /*
2816          * We don't ever fill up leaves all the way so multiply by 2 just to be
2817          * closer to what we're really going to want to use.
2818          */
2819         return div_u64(num_bytes, BTRFS_LEAF_DATA_SIZE(fs_info));
2820 }
2821
2822 /*
2823  * Takes the number of bytes to be csumm'ed and figures out how many leaves it
2824  * would require to store the csums for that many bytes.
2825  */
2826 u64 btrfs_csum_bytes_to_leaves(struct btrfs_fs_info *fs_info, u64 csum_bytes)
2827 {
2828         u64 csum_size;
2829         u64 num_csums_per_leaf;
2830         u64 num_csums;
2831
2832         csum_size = BTRFS_MAX_ITEM_SIZE(fs_info);
2833         num_csums_per_leaf = div64_u64(csum_size,
2834                         (u64)btrfs_super_csum_size(fs_info->super_copy));
2835         num_csums = div64_u64(csum_bytes, fs_info->sectorsize);
2836         num_csums += num_csums_per_leaf - 1;
2837         num_csums = div64_u64(num_csums, num_csums_per_leaf);
2838         return num_csums;
2839 }
2840
2841 bool btrfs_check_space_for_delayed_refs(struct btrfs_fs_info *fs_info)
2842 {
2843         struct btrfs_block_rsv *delayed_refs_rsv = &fs_info->delayed_refs_rsv;
2844         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
2845         bool ret = false;
2846         u64 reserved;
2847
2848         spin_lock(&global_rsv->lock);
2849         reserved = global_rsv->reserved;
2850         spin_unlock(&global_rsv->lock);
2851
2852         /*
2853          * Since the global reserve is just kind of magic we don't really want
2854          * to rely on it to save our bacon, so if our size is more than the
2855          * delayed_refs_rsv and the global rsv then it's time to think about
2856          * bailing.
2857          */
2858         spin_lock(&delayed_refs_rsv->lock);
2859         reserved += delayed_refs_rsv->reserved;
2860         if (delayed_refs_rsv->size >= reserved)
2861                 ret = true;
2862         spin_unlock(&delayed_refs_rsv->lock);
2863         return ret;
2864 }
2865
2866 int btrfs_should_throttle_delayed_refs(struct btrfs_trans_handle *trans)
2867 {
2868         u64 num_entries =
2869                 atomic_read(&trans->transaction->delayed_refs.num_entries);
2870         u64 avg_runtime;
2871         u64 val;
2872
2873         smp_mb();
2874         avg_runtime = trans->fs_info->avg_delayed_ref_runtime;
2875         val = num_entries * avg_runtime;
2876         if (val >= NSEC_PER_SEC)
2877                 return 1;
2878         if (val >= NSEC_PER_SEC / 2)
2879                 return 2;
2880
2881         return btrfs_check_space_for_delayed_refs(trans->fs_info);
2882 }
2883
2884 /*
2885  * this starts processing the delayed reference count updates and
2886  * extent insertions we have queued up so far.  count can be
2887  * 0, which means to process everything in the tree at the start
2888  * of the run (but not newly added entries), or it can be some target
2889  * number you'd like to process.
2890  *
2891  * Returns 0 on success or if called with an aborted transaction
2892  * Returns <0 on error and aborts the transaction
2893  */
2894 int btrfs_run_delayed_refs(struct btrfs_trans_handle *trans,
2895                            unsigned long count)
2896 {
2897         struct btrfs_fs_info *fs_info = trans->fs_info;
2898         struct rb_node *node;
2899         struct btrfs_delayed_ref_root *delayed_refs;
2900         struct btrfs_delayed_ref_head *head;
2901         int ret;
2902         int run_all = count == (unsigned long)-1;
2903
2904         /* We'll clean this up in btrfs_cleanup_transaction */
2905         if (trans->aborted)
2906                 return 0;
2907
2908         if (test_bit(BTRFS_FS_CREATING_FREE_SPACE_TREE, &fs_info->flags))
2909                 return 0;
2910
2911         delayed_refs = &trans->transaction->delayed_refs;
2912         if (count == 0)
2913                 count = atomic_read(&delayed_refs->num_entries) * 2;
2914
2915 again:
2916 #ifdef SCRAMBLE_DELAYED_REFS
2917         delayed_refs->run_delayed_start = find_middle(&delayed_refs->root);
2918 #endif
2919         ret = __btrfs_run_delayed_refs(trans, count);
2920         if (ret < 0) {
2921                 btrfs_abort_transaction(trans, ret);
2922                 return ret;
2923         }
2924
2925         if (run_all) {
2926                 btrfs_create_pending_block_groups(trans);
2927
2928                 spin_lock(&delayed_refs->lock);
2929                 node = rb_first_cached(&delayed_refs->href_root);
2930                 if (!node) {
2931                         spin_unlock(&delayed_refs->lock);
2932                         goto out;
2933                 }
2934                 head = rb_entry(node, struct btrfs_delayed_ref_head,
2935                                 href_node);
2936                 refcount_inc(&head->refs);
2937                 spin_unlock(&delayed_refs->lock);
2938
2939                 /* Mutex was contended, block until it's released and retry. */
2940                 mutex_lock(&head->mutex);
2941                 mutex_unlock(&head->mutex);
2942
2943                 btrfs_put_delayed_ref_head(head);
2944                 cond_resched();
2945                 goto again;
2946         }
2947 out:
2948         return 0;
2949 }
2950
2951 int btrfs_set_disk_extent_flags(struct btrfs_trans_handle *trans,
2952                                 u64 bytenr, u64 num_bytes, u64 flags,
2953                                 int level, int is_data)
2954 {
2955         struct btrfs_delayed_extent_op *extent_op;
2956         int ret;
2957
2958         extent_op = btrfs_alloc_delayed_extent_op();
2959         if (!extent_op)
2960                 return -ENOMEM;
2961
2962         extent_op->flags_to_set = flags;
2963         extent_op->update_flags = true;
2964         extent_op->update_key = false;
2965         extent_op->is_data = is_data ? true : false;
2966         extent_op->level = level;
2967
2968         ret = btrfs_add_delayed_extent_op(trans, bytenr, num_bytes, extent_op);
2969         if (ret)
2970                 btrfs_free_delayed_extent_op(extent_op);
2971         return ret;
2972 }
2973
2974 static noinline int check_delayed_ref(struct btrfs_root *root,
2975                                       struct btrfs_path *path,
2976                                       u64 objectid, u64 offset, u64 bytenr)
2977 {
2978         struct btrfs_delayed_ref_head *head;
2979         struct btrfs_delayed_ref_node *ref;
2980         struct btrfs_delayed_data_ref *data_ref;
2981         struct btrfs_delayed_ref_root *delayed_refs;
2982         struct btrfs_transaction *cur_trans;
2983         struct rb_node *node;
2984         int ret = 0;
2985
2986         spin_lock(&root->fs_info->trans_lock);
2987         cur_trans = root->fs_info->running_transaction;
2988         if (cur_trans)
2989                 refcount_inc(&cur_trans->use_count);
2990         spin_unlock(&root->fs_info->trans_lock);
2991         if (!cur_trans)
2992                 return 0;
2993
2994         delayed_refs = &cur_trans->delayed_refs;
2995         spin_lock(&delayed_refs->lock);
2996         head = btrfs_find_delayed_ref_head(delayed_refs, bytenr);
2997         if (!head) {
2998                 spin_unlock(&delayed_refs->lock);
2999                 btrfs_put_transaction(cur_trans);
3000                 return 0;
3001         }
3002
3003         if (!mutex_trylock(&head->mutex)) {
3004                 refcount_inc(&head->refs);
3005                 spin_unlock(&delayed_refs->lock);
3006
3007                 btrfs_release_path(path);
3008
3009                 /*
3010                  * Mutex was contended, block until it's released and let
3011                  * caller try again
3012                  */
3013                 mutex_lock(&head->mutex);
3014                 mutex_unlock(&head->mutex);
3015                 btrfs_put_delayed_ref_head(head);
3016                 btrfs_put_transaction(cur_trans);
3017                 return -EAGAIN;
3018         }
3019         spin_unlock(&delayed_refs->lock);
3020
3021         spin_lock(&head->lock);
3022         /*
3023          * XXX: We should replace this with a proper search function in the
3024          * future.
3025          */
3026         for (node = rb_first_cached(&head->ref_tree); node;
3027              node = rb_next(node)) {
3028                 ref = rb_entry(node, struct btrfs_delayed_ref_node, ref_node);
3029                 /* If it's a shared ref we know a cross reference exists */
3030                 if (ref->type != BTRFS_EXTENT_DATA_REF_KEY) {
3031                         ret = 1;
3032                         break;
3033                 }
3034
3035                 data_ref = btrfs_delayed_node_to_data_ref(ref);
3036
3037                 /*
3038                  * If our ref doesn't match the one we're currently looking at
3039                  * then we have a cross reference.
3040                  */
3041                 if (data_ref->root != root->root_key.objectid ||
3042                     data_ref->objectid != objectid ||
3043                     data_ref->offset != offset) {
3044                         ret = 1;
3045                         break;
3046                 }
3047         }
3048         spin_unlock(&head->lock);
3049         mutex_unlock(&head->mutex);
3050         btrfs_put_transaction(cur_trans);
3051         return ret;
3052 }
3053
3054 static noinline int check_committed_ref(struct btrfs_root *root,
3055                                         struct btrfs_path *path,
3056                                         u64 objectid, u64 offset, u64 bytenr)
3057 {
3058         struct btrfs_fs_info *fs_info = root->fs_info;
3059         struct btrfs_root *extent_root = fs_info->extent_root;
3060         struct extent_buffer *leaf;
3061         struct btrfs_extent_data_ref *ref;
3062         struct btrfs_extent_inline_ref *iref;
3063         struct btrfs_extent_item *ei;
3064         struct btrfs_key key;
3065         u32 item_size;
3066         int type;
3067         int ret;
3068
3069         key.objectid = bytenr;
3070         key.offset = (u64)-1;
3071         key.type = BTRFS_EXTENT_ITEM_KEY;
3072
3073         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
3074         if (ret < 0)
3075                 goto out;
3076         BUG_ON(ret == 0); /* Corruption */
3077
3078         ret = -ENOENT;
3079         if (path->slots[0] == 0)
3080                 goto out;
3081
3082         path->slots[0]--;
3083         leaf = path->nodes[0];
3084         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
3085
3086         if (key.objectid != bytenr || key.type != BTRFS_EXTENT_ITEM_KEY)
3087                 goto out;
3088
3089         ret = 1;
3090         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
3091         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
3092
3093         if (item_size != sizeof(*ei) +
3094             btrfs_extent_inline_ref_size(BTRFS_EXTENT_DATA_REF_KEY))
3095                 goto out;
3096
3097         if (btrfs_extent_generation(leaf, ei) <=
3098             btrfs_root_last_snapshot(&root->root_item))
3099                 goto out;
3100
3101         iref = (struct btrfs_extent_inline_ref *)(ei + 1);
3102
3103         type = btrfs_get_extent_inline_ref_type(leaf, iref, BTRFS_REF_TYPE_DATA);
3104         if (type != BTRFS_EXTENT_DATA_REF_KEY)
3105                 goto out;
3106
3107         ref = (struct btrfs_extent_data_ref *)(&iref->offset);
3108         if (btrfs_extent_refs(leaf, ei) !=
3109             btrfs_extent_data_ref_count(leaf, ref) ||
3110             btrfs_extent_data_ref_root(leaf, ref) !=
3111             root->root_key.objectid ||
3112             btrfs_extent_data_ref_objectid(leaf, ref) != objectid ||
3113             btrfs_extent_data_ref_offset(leaf, ref) != offset)
3114                 goto out;
3115
3116         ret = 0;
3117 out:
3118         return ret;
3119 }
3120
3121 int btrfs_cross_ref_exist(struct btrfs_root *root, u64 objectid, u64 offset,
3122                           u64 bytenr)
3123 {
3124         struct btrfs_path *path;
3125         int ret;
3126
3127         path = btrfs_alloc_path();
3128         if (!path)
3129                 return -ENOMEM;
3130
3131         do {
3132                 ret = check_committed_ref(root, path, objectid,
3133                                           offset, bytenr);
3134                 if (ret && ret != -ENOENT)
3135                         goto out;
3136
3137                 ret = check_delayed_ref(root, path, objectid, offset, bytenr);
3138         } while (ret == -EAGAIN);
3139
3140 out:
3141         btrfs_free_path(path);
3142         if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID)
3143                 WARN_ON(ret > 0);
3144         return ret;
3145 }
3146
3147 static int __btrfs_mod_ref(struct btrfs_trans_handle *trans,
3148                            struct btrfs_root *root,
3149                            struct extent_buffer *buf,
3150                            int full_backref, int inc)
3151 {
3152         struct btrfs_fs_info *fs_info = root->fs_info;
3153         u64 bytenr;
3154         u64 num_bytes;
3155         u64 parent;
3156         u64 ref_root;
3157         u32 nritems;
3158         struct btrfs_key key;
3159         struct btrfs_file_extent_item *fi;
3160         struct btrfs_ref generic_ref = { 0 };
3161         bool for_reloc = btrfs_header_flag(buf, BTRFS_HEADER_FLAG_RELOC);
3162         int i;
3163         int action;
3164         int level;
3165         int ret = 0;
3166
3167         if (btrfs_is_testing(fs_info))
3168                 return 0;
3169
3170         ref_root = btrfs_header_owner(buf);
3171         nritems = btrfs_header_nritems(buf);
3172         level = btrfs_header_level(buf);
3173
3174         if (!test_bit(BTRFS_ROOT_REF_COWS, &root->state) && level == 0)
3175                 return 0;
3176
3177         if (full_backref)
3178                 parent = buf->start;
3179         else
3180                 parent = 0;
3181         if (inc)
3182                 action = BTRFS_ADD_DELAYED_REF;
3183         else
3184                 action = BTRFS_DROP_DELAYED_REF;
3185
3186         for (i = 0; i < nritems; i++) {
3187                 if (level == 0) {
3188                         btrfs_item_key_to_cpu(buf, &key, i);
3189                         if (key.type != BTRFS_EXTENT_DATA_KEY)
3190                                 continue;
3191                         fi = btrfs_item_ptr(buf, i,
3192                                             struct btrfs_file_extent_item);
3193                         if (btrfs_file_extent_type(buf, fi) ==
3194                             BTRFS_FILE_EXTENT_INLINE)
3195                                 continue;
3196                         bytenr = btrfs_file_extent_disk_bytenr(buf, fi);
3197                         if (bytenr == 0)
3198                                 continue;
3199
3200                         num_bytes = btrfs_file_extent_disk_num_bytes(buf, fi);
3201                         key.offset -= btrfs_file_extent_offset(buf, fi);
3202                         btrfs_init_generic_ref(&generic_ref, action, bytenr,
3203                                                num_bytes, parent);
3204                         generic_ref.real_root = root->root_key.objectid;
3205                         btrfs_init_data_ref(&generic_ref, ref_root, key.objectid,
3206                                             key.offset);
3207                         generic_ref.skip_qgroup = for_reloc;
3208                         if (inc)
3209                                 ret = btrfs_inc_extent_ref(trans, &generic_ref);
3210                         else
3211                                 ret = btrfs_free_extent(trans, &generic_ref);
3212                         if (ret)
3213                                 goto fail;
3214                 } else {
3215                         bytenr = btrfs_node_blockptr(buf, i);
3216                         num_bytes = fs_info->nodesize;
3217                         btrfs_init_generic_ref(&generic_ref, action, bytenr,
3218                                                num_bytes, parent);
3219                         generic_ref.real_root = root->root_key.objectid;
3220                         btrfs_init_tree_ref(&generic_ref, level - 1, ref_root);
3221                         generic_ref.skip_qgroup = for_reloc;
3222                         if (inc)
3223                                 ret = btrfs_inc_extent_ref(trans, &generic_ref);
3224                         else
3225                                 ret = btrfs_free_extent(trans, &generic_ref);
3226                         if (ret)
3227                                 goto fail;
3228                 }
3229         }
3230         return 0;
3231 fail:
3232         return ret;
3233 }
3234
3235 int btrfs_inc_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
3236                   struct extent_buffer *buf, int full_backref)
3237 {
3238         return __btrfs_mod_ref(trans, root, buf, full_backref, 1);
3239 }
3240
3241 int btrfs_dec_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
3242                   struct extent_buffer *buf, int full_backref)
3243 {
3244         return __btrfs_mod_ref(trans, root, buf, full_backref, 0);
3245 }
3246
3247 static int write_one_cache_group(struct btrfs_trans_handle *trans,
3248                                  struct btrfs_path *path,
3249                                  struct btrfs_block_group_cache *cache)
3250 {
3251         struct btrfs_fs_info *fs_info = trans->fs_info;
3252         int ret;
3253         struct btrfs_root *extent_root = fs_info->extent_root;
3254         unsigned long bi;
3255         struct extent_buffer *leaf;
3256
3257         ret = btrfs_search_slot(trans, extent_root, &cache->key, path, 0, 1);
3258         if (ret) {
3259                 if (ret > 0)
3260                         ret = -ENOENT;
3261                 goto fail;
3262         }
3263
3264         leaf = path->nodes[0];
3265         bi = btrfs_item_ptr_offset(leaf, path->slots[0]);
3266         write_extent_buffer(leaf, &cache->item, bi, sizeof(cache->item));
3267         btrfs_mark_buffer_dirty(leaf);
3268 fail:
3269         btrfs_release_path(path);
3270         return ret;
3271
3272 }
3273
3274 static struct btrfs_block_group_cache *next_block_group(
3275                 struct btrfs_block_group_cache *cache)
3276 {
3277         struct btrfs_fs_info *fs_info = cache->fs_info;
3278         struct rb_node *node;
3279
3280         spin_lock(&fs_info->block_group_cache_lock);
3281
3282         /* If our block group was removed, we need a full search. */
3283         if (RB_EMPTY_NODE(&cache->cache_node)) {
3284                 const u64 next_bytenr = cache->key.objectid + cache->key.offset;
3285
3286                 spin_unlock(&fs_info->block_group_cache_lock);
3287                 btrfs_put_block_group(cache);
3288                 cache = btrfs_lookup_first_block_group(fs_info, next_bytenr); return cache;
3289         }
3290         node = rb_next(&cache->cache_node);
3291         btrfs_put_block_group(cache);
3292         if (node) {
3293                 cache = rb_entry(node, struct btrfs_block_group_cache,
3294                                  cache_node);
3295                 btrfs_get_block_group(cache);
3296         } else
3297                 cache = NULL;
3298         spin_unlock(&fs_info->block_group_cache_lock);
3299         return cache;
3300 }
3301
3302 static int cache_save_setup(struct btrfs_block_group_cache *block_group,
3303                             struct btrfs_trans_handle *trans,
3304                             struct btrfs_path *path)
3305 {
3306         struct btrfs_fs_info *fs_info = block_group->fs_info;
3307         struct btrfs_root *root = fs_info->tree_root;
3308         struct inode *inode = NULL;
3309         struct extent_changeset *data_reserved = NULL;
3310         u64 alloc_hint = 0;
3311         int dcs = BTRFS_DC_ERROR;
3312         u64 num_pages = 0;
3313         int retries = 0;
3314         int ret = 0;
3315
3316         /*
3317          * If this block group is smaller than 100 megs don't bother caching the
3318          * block group.
3319          */
3320         if (block_group->key.offset < (100 * SZ_1M)) {
3321                 spin_lock(&block_group->lock);
3322                 block_group->disk_cache_state = BTRFS_DC_WRITTEN;
3323                 spin_unlock(&block_group->lock);
3324                 return 0;
3325         }
3326
3327         if (trans->aborted)
3328                 return 0;
3329 again:
3330         inode = lookup_free_space_inode(block_group, path);
3331         if (IS_ERR(inode) && PTR_ERR(inode) != -ENOENT) {
3332                 ret = PTR_ERR(inode);
3333                 btrfs_release_path(path);
3334                 goto out;
3335         }
3336
3337         if (IS_ERR(inode)) {
3338                 BUG_ON(retries);
3339                 retries++;
3340
3341                 if (block_group->ro)
3342                         goto out_free;
3343
3344                 ret = create_free_space_inode(trans, block_group, path);
3345                 if (ret)
3346                         goto out_free;
3347                 goto again;
3348         }
3349
3350         /*
3351          * We want to set the generation to 0, that way if anything goes wrong
3352          * from here on out we know not to trust this cache when we load up next
3353          * time.
3354          */
3355         BTRFS_I(inode)->generation = 0;
3356         ret = btrfs_update_inode(trans, root, inode);
3357         if (ret) {
3358                 /*
3359                  * So theoretically we could recover from this, simply set the
3360                  * super cache generation to 0 so we know to invalidate the
3361                  * cache, but then we'd have to keep track of the block groups
3362                  * that fail this way so we know we _have_ to reset this cache
3363                  * before the next commit or risk reading stale cache.  So to
3364                  * limit our exposure to horrible edge cases lets just abort the
3365                  * transaction, this only happens in really bad situations
3366                  * anyway.
3367                  */
3368                 btrfs_abort_transaction(trans, ret);
3369                 goto out_put;
3370         }
3371         WARN_ON(ret);
3372
3373         /* We've already setup this transaction, go ahead and exit */
3374         if (block_group->cache_generation == trans->transid &&
3375             i_size_read(inode)) {
3376                 dcs = BTRFS_DC_SETUP;
3377                 goto out_put;
3378         }
3379
3380         if (i_size_read(inode) > 0) {
3381                 ret = btrfs_check_trunc_cache_free_space(fs_info,
3382                                         &fs_info->global_block_rsv);
3383                 if (ret)
3384                         goto out_put;
3385
3386                 ret = btrfs_truncate_free_space_cache(trans, NULL, inode);
3387                 if (ret)
3388                         goto out_put;
3389         }
3390
3391         spin_lock(&block_group->lock);
3392         if (block_group->cached != BTRFS_CACHE_FINISHED ||
3393             !btrfs_test_opt(fs_info, SPACE_CACHE)) {
3394                 /*
3395                  * don't bother trying to write stuff out _if_
3396                  * a) we're not cached,
3397                  * b) we're with nospace_cache mount option,
3398                  * c) we're with v2 space_cache (FREE_SPACE_TREE).
3399                  */
3400                 dcs = BTRFS_DC_WRITTEN;
3401                 spin_unlock(&block_group->lock);
3402                 goto out_put;
3403         }
3404         spin_unlock(&block_group->lock);
3405
3406         /*
3407          * We hit an ENOSPC when setting up the cache in this transaction, just
3408          * skip doing the setup, we've already cleared the cache so we're safe.
3409          */
3410         if (test_bit(BTRFS_TRANS_CACHE_ENOSPC, &trans->transaction->flags)) {
3411                 ret = -ENOSPC;
3412                 goto out_put;
3413         }
3414
3415         /*
3416          * Try to preallocate enough space based on how big the block group is.
3417          * Keep in mind this has to include any pinned space which could end up
3418          * taking up quite a bit since it's not folded into the other space
3419          * cache.
3420          */
3421         num_pages = div_u64(block_group->key.offset, SZ_256M);
3422         if (!num_pages)
3423                 num_pages = 1;
3424
3425         num_pages *= 16;
3426         num_pages *= PAGE_SIZE;
3427
3428         ret = btrfs_check_data_free_space(inode, &data_reserved, 0, num_pages);
3429         if (ret)
3430                 goto out_put;
3431
3432         ret = btrfs_prealloc_file_range_trans(inode, trans, 0, 0, num_pages,
3433                                               num_pages, num_pages,
3434                                               &alloc_hint);
3435         /*
3436          * Our cache requires contiguous chunks so that we don't modify a bunch
3437          * of metadata or split extents when writing the cache out, which means
3438          * we can enospc if we are heavily fragmented in addition to just normal
3439          * out of space conditions.  So if we hit this just skip setting up any
3440          * other block groups for this transaction, maybe we'll unpin enough
3441          * space the next time around.
3442          */
3443         if (!ret)
3444                 dcs = BTRFS_DC_SETUP;
3445         else if (ret == -ENOSPC)
3446                 set_bit(BTRFS_TRANS_CACHE_ENOSPC, &trans->transaction->flags);
3447
3448 out_put:
3449         iput(inode);
3450 out_free:
3451         btrfs_release_path(path);
3452 out:
3453         spin_lock(&block_group->lock);
3454         if (!ret && dcs == BTRFS_DC_SETUP)
3455                 block_group->cache_generation = trans->transid;
3456         block_group->disk_cache_state = dcs;
3457         spin_unlock(&block_group->lock);
3458
3459         extent_changeset_free(data_reserved);
3460         return ret;
3461 }
3462
3463 int btrfs_setup_space_cache(struct btrfs_trans_handle *trans)
3464 {
3465         struct btrfs_fs_info *fs_info = trans->fs_info;
3466         struct btrfs_block_group_cache *cache, *tmp;
3467         struct btrfs_transaction *cur_trans = trans->transaction;
3468         struct btrfs_path *path;
3469
3470         if (list_empty(&cur_trans->dirty_bgs) ||
3471             !btrfs_test_opt(fs_info, SPACE_CACHE))
3472                 return 0;
3473
3474         path = btrfs_alloc_path();
3475         if (!path)
3476                 return -ENOMEM;
3477
3478         /* Could add new block groups, use _safe just in case */
3479         list_for_each_entry_safe(cache, tmp, &cur_trans->dirty_bgs,
3480                                  dirty_list) {
3481                 if (cache->disk_cache_state == BTRFS_DC_CLEAR)
3482                         cache_save_setup(cache, trans, path);
3483         }
3484
3485         btrfs_free_path(path);
3486         return 0;
3487 }
3488
3489 /*
3490  * transaction commit does final block group cache writeback during a
3491  * critical section where nothing is allowed to change the FS.  This is
3492  * required in order for the cache to actually match the block group,
3493  * but can introduce a lot of latency into the commit.
3494  *
3495  * So, btrfs_start_dirty_block_groups is here to kick off block group
3496  * cache IO.  There's a chance we'll have to redo some of it if the
3497  * block group changes again during the commit, but it greatly reduces
3498  * the commit latency by getting rid of the easy block groups while
3499  * we're still allowing others to join the commit.
3500  */
3501 int btrfs_start_dirty_block_groups(struct btrfs_trans_handle *trans)
3502 {
3503         struct btrfs_fs_info *fs_info = trans->fs_info;
3504         struct btrfs_block_group_cache *cache;
3505         struct btrfs_transaction *cur_trans = trans->transaction;
3506         int ret = 0;
3507         int should_put;
3508         struct btrfs_path *path = NULL;
3509         LIST_HEAD(dirty);
3510         struct list_head *io = &cur_trans->io_bgs;
3511         int num_started = 0;
3512         int loops = 0;
3513
3514         spin_lock(&cur_trans->dirty_bgs_lock);
3515         if (list_empty(&cur_trans->dirty_bgs)) {
3516                 spin_unlock(&cur_trans->dirty_bgs_lock);
3517                 return 0;
3518         }
3519         list_splice_init(&cur_trans->dirty_bgs, &dirty);
3520         spin_unlock(&cur_trans->dirty_bgs_lock);
3521
3522 again:
3523         /*
3524          * make sure all the block groups on our dirty list actually
3525          * exist
3526          */
3527         btrfs_create_pending_block_groups(trans);
3528
3529         if (!path) {
3530                 path = btrfs_alloc_path();
3531                 if (!path)
3532                         return -ENOMEM;
3533         }
3534
3535         /*
3536          * cache_write_mutex is here only to save us from balance or automatic
3537          * removal of empty block groups deleting this block group while we are
3538          * writing out the cache
3539          */
3540         mutex_lock(&trans->transaction->cache_write_mutex);
3541         while (!list_empty(&dirty)) {
3542                 bool drop_reserve = true;
3543
3544                 cache = list_first_entry(&dirty,
3545                                          struct btrfs_block_group_cache,
3546                                          dirty_list);
3547                 /*
3548                  * this can happen if something re-dirties a block
3549                  * group that is already under IO.  Just wait for it to
3550                  * finish and then do it all again
3551                  */
3552                 if (!list_empty(&cache->io_list)) {
3553                         list_del_init(&cache->io_list);
3554                         btrfs_wait_cache_io(trans, cache, path);
3555                         btrfs_put_block_group(cache);
3556                 }
3557
3558
3559                 /*
3560                  * btrfs_wait_cache_io uses the cache->dirty_list to decide
3561                  * if it should update the cache_state.  Don't delete
3562                  * until after we wait.
3563                  *
3564                  * Since we're not running in the commit critical section
3565                  * we need the dirty_bgs_lock to protect from update_block_group
3566                  */
3567                 spin_lock(&cur_trans->dirty_bgs_lock);
3568                 list_del_init(&cache->dirty_list);
3569                 spin_unlock(&cur_trans->dirty_bgs_lock);
3570
3571                 should_put = 1;
3572
3573                 cache_save_setup(cache, trans, path);
3574
3575                 if (cache->disk_cache_state == BTRFS_DC_SETUP) {
3576                         cache->io_ctl.inode = NULL;
3577                         ret = btrfs_write_out_cache(trans, cache, path);
3578                         if (ret == 0 && cache->io_ctl.inode) {
3579                                 num_started++;
3580                                 should_put = 0;
3581
3582                                 /*
3583                                  * The cache_write_mutex is protecting the
3584                                  * io_list, also refer to the definition of
3585                                  * btrfs_transaction::io_bgs for more details
3586                                  */
3587                                 list_add_tail(&cache->io_list, io);
3588                         } else {
3589                                 /*
3590                                  * if we failed to write the cache, the
3591                                  * generation will be bad and life goes on
3592                                  */
3593                                 ret = 0;
3594                         }
3595                 }
3596                 if (!ret) {
3597                         ret = write_one_cache_group(trans, path, cache);
3598                         /*
3599                          * Our block group might still be attached to the list
3600                          * of new block groups in the transaction handle of some
3601                          * other task (struct btrfs_trans_handle->new_bgs). This
3602                          * means its block group item isn't yet in the extent
3603                          * tree. If this happens ignore the error, as we will
3604                          * try again later in the critical section of the
3605                          * transaction commit.
3606                          */
3607                         if (ret == -ENOENT) {
3608                                 ret = 0;
3609                                 spin_lock(&cur_trans->dirty_bgs_lock);
3610                                 if (list_empty(&cache->dirty_list)) {
3611                                         list_add_tail(&cache->dirty_list,
3612                                                       &cur_trans->dirty_bgs);
3613                                         btrfs_get_block_group(cache);
3614                                         drop_reserve = false;
3615                                 }
3616                                 spin_unlock(&cur_trans->dirty_bgs_lock);
3617                         } else if (ret) {
3618                                 btrfs_abort_transaction(trans, ret);
3619                         }
3620                 }
3621
3622                 /* if it's not on the io list, we need to put the block group */
3623                 if (should_put)
3624                         btrfs_put_block_group(cache);
3625                 if (drop_reserve)
3626                         btrfs_delayed_refs_rsv_release(fs_info, 1);
3627
3628                 if (ret)
3629                         break;
3630
3631                 /*
3632                  * Avoid blocking other tasks for too long. It might even save
3633                  * us from writing caches for block groups that are going to be
3634                  * removed.
3635                  */
3636                 mutex_unlock(&trans->transaction->cache_write_mutex);
3637                 mutex_lock(&trans->transaction->cache_write_mutex);
3638         }
3639         mutex_unlock(&trans->transaction->cache_write_mutex);
3640
3641         /*
3642          * go through delayed refs for all the stuff we've just kicked off
3643          * and then loop back (just once)
3644          */
3645         ret = btrfs_run_delayed_refs(trans, 0);
3646         if (!ret && loops == 0) {
3647                 loops++;
3648                 spin_lock(&cur_trans->dirty_bgs_lock);
3649                 list_splice_init(&cur_trans->dirty_bgs, &dirty);
3650                 /*
3651                  * dirty_bgs_lock protects us from concurrent block group
3652                  * deletes too (not just cache_write_mutex).
3653                  */
3654                 if (!list_empty(&dirty)) {
3655                         spin_unlock(&cur_trans->dirty_bgs_lock);
3656                         goto again;
3657                 }
3658                 spin_unlock(&cur_trans->dirty_bgs_lock);
3659         } else if (ret < 0) {
3660                 btrfs_cleanup_dirty_bgs(cur_trans, fs_info);
3661         }
3662
3663         btrfs_free_path(path);
3664         return ret;
3665 }
3666
3667 int btrfs_write_dirty_block_groups(struct btrfs_trans_handle *trans)
3668 {
3669         struct btrfs_fs_info *fs_info = trans->fs_info;
3670         struct btrfs_block_group_cache *cache;
3671         struct btrfs_transaction *cur_trans = trans->transaction;
3672         int ret = 0;
3673         int should_put;
3674         struct btrfs_path *path;
3675         struct list_head *io = &cur_trans->io_bgs;
3676         int num_started = 0;
3677
3678         path = btrfs_alloc_path();
3679         if (!path)
3680                 return -ENOMEM;
3681
3682         /*
3683          * Even though we are in the critical section of the transaction commit,
3684          * we can still have concurrent tasks adding elements to this
3685          * transaction's list of dirty block groups. These tasks correspond to
3686          * endio free space workers started when writeback finishes for a
3687          * space cache, which run inode.c:btrfs_finish_ordered_io(), and can
3688          * allocate new block groups as a result of COWing nodes of the root
3689          * tree when updating the free space inode. The writeback for the space
3690          * caches is triggered by an earlier call to
3691          * btrfs_start_dirty_block_groups() and iterations of the following
3692          * loop.
3693          * Also we want to do the cache_save_setup first and then run the
3694          * delayed refs to make sure we have the best chance at doing this all
3695          * in one shot.
3696          */
3697         spin_lock(&cur_trans->dirty_bgs_lock);
3698         while (!list_empty(&cur_trans->dirty_bgs)) {
3699                 cache = list_first_entry(&cur_trans->dirty_bgs,
3700                                          struct btrfs_block_group_cache,
3701                                          dirty_list);
3702
3703                 /*
3704                  * this can happen if cache_save_setup re-dirties a block
3705                  * group that is already under IO.  Just wait for it to
3706                  * finish and then do it all again
3707                  */
3708                 if (!list_empty(&cache->io_list)) {
3709                         spin_unlock(&cur_trans->dirty_bgs_lock);
3710                         list_del_init(&cache->io_list);
3711                         btrfs_wait_cache_io(trans, cache, path);
3712                         btrfs_put_block_group(cache);
3713                         spin_lock(&cur_trans->dirty_bgs_lock);
3714                 }
3715
3716                 /*
3717                  * don't remove from the dirty list until after we've waited
3718                  * on any pending IO
3719                  */
3720                 list_del_init(&cache->dirty_list);
3721                 spin_unlock(&cur_trans->dirty_bgs_lock);
3722                 should_put = 1;
3723
3724                 cache_save_setup(cache, trans, path);
3725
3726                 if (!ret)
3727                         ret = btrfs_run_delayed_refs(trans,
3728                                                      (unsigned long) -1);
3729
3730                 if (!ret && cache->disk_cache_state == BTRFS_DC_SETUP) {
3731                         cache->io_ctl.inode = NULL;
3732                         ret = btrfs_write_out_cache(trans, cache, path);
3733                         if (ret == 0 && cache->io_ctl.inode) {
3734                                 num_started++;
3735                                 should_put = 0;
3736                                 list_add_tail(&cache->io_list, io);
3737                         } else {
3738                                 /*
3739                                  * if we failed to write the cache, the
3740                                  * generation will be bad and life goes on
3741                                  */
3742                                 ret = 0;
3743                         }
3744                 }
3745                 if (!ret) {
3746                         ret = write_one_cache_group(trans, path, cache);
3747                         /*
3748                          * One of the free space endio workers might have
3749                          * created a new block group while updating a free space
3750                          * cache's inode (at inode.c:btrfs_finish_ordered_io())
3751                          * and hasn't released its transaction handle yet, in
3752                          * which case the new block group is still attached to
3753                          * its transaction handle and its creation has not
3754                          * finished yet (no block group item in the extent tree
3755                          * yet, etc). If this is the case, wait for all free
3756                          * space endio workers to finish and retry. This is a
3757                          * a very rare case so no need for a more efficient and
3758                          * complex approach.
3759                          */
3760                         if (ret == -ENOENT) {
3761                                 wait_event(cur_trans->writer_wait,
3762                                    atomic_read(&cur_trans->num_writers) == 1);
3763                                 ret = write_one_cache_group(trans, path, cache);
3764                         }
3765                         if (ret)
3766                                 btrfs_abort_transaction(trans, ret);
3767                 }
3768
3769                 /* if its not on the io list, we need to put the block group */
3770                 if (should_put)
3771                         btrfs_put_block_group(cache);
3772                 btrfs_delayed_refs_rsv_release(fs_info, 1);
3773                 spin_lock(&cur_trans->dirty_bgs_lock);
3774         }
3775         spin_unlock(&cur_trans->dirty_bgs_lock);
3776
3777         /*
3778          * Refer to the definition of io_bgs member for details why it's safe
3779          * to use it without any locking
3780          */
3781         while (!list_empty(io)) {
3782                 cache = list_first_entry(io, struct btrfs_block_group_cache,
3783                                          io_list);
3784                 list_del_init(&cache->io_list);
3785                 btrfs_wait_cache_io(trans, cache, path);
3786                 btrfs_put_block_group(cache);
3787         }
3788
3789         btrfs_free_path(path);
3790         return ret;
3791 }
3792
3793 int btrfs_extent_readonly(struct btrfs_fs_info *fs_info, u64 bytenr)
3794 {
3795         struct btrfs_block_group_cache *block_group;
3796         int readonly = 0;
3797
3798         block_group = btrfs_lookup_block_group(fs_info, bytenr);
3799         if (!block_group || block_group->ro)
3800                 readonly = 1;
3801         if (block_group)
3802                 btrfs_put_block_group(block_group);
3803         return readonly;
3804 }
3805
3806 bool btrfs_inc_nocow_writers(struct btrfs_fs_info *fs_info, u64 bytenr)
3807 {
3808         struct btrfs_block_group_cache *bg;
3809         bool ret = true;
3810
3811         bg = btrfs_lookup_block_group(fs_info, bytenr);
3812         if (!bg)
3813                 return false;
3814
3815         spin_lock(&bg->lock);
3816         if (bg->ro)
3817                 ret = false;
3818         else
3819                 atomic_inc(&bg->nocow_writers);
3820         spin_unlock(&bg->lock);
3821
3822         /* no put on block group, done by btrfs_dec_nocow_writers */
3823         if (!ret)
3824                 btrfs_put_block_group(bg);
3825
3826         return ret;
3827
3828 }
3829
3830 void btrfs_dec_nocow_writers(struct btrfs_fs_info *fs_info, u64 bytenr)
3831 {
3832         struct btrfs_block_group_cache *bg;
3833
3834         bg = btrfs_lookup_block_group(fs_info, bytenr);
3835         ASSERT(bg);
3836         if (atomic_dec_and_test(&bg->nocow_writers))
3837                 wake_up_var(&bg->nocow_writers);
3838         /*
3839          * Once for our lookup and once for the lookup done by a previous call
3840          * to btrfs_inc_nocow_writers()
3841          */
3842         btrfs_put_block_group(bg);
3843         btrfs_put_block_group(bg);
3844 }
3845
3846 void btrfs_wait_nocow_writers(struct btrfs_block_group_cache *bg)
3847 {
3848         wait_var_event(&bg->nocow_writers, !atomic_read(&bg->nocow_writers));
3849 }
3850
3851 static const char *alloc_name(u64 flags)
3852 {
3853         switch (flags) {
3854         case BTRFS_BLOCK_GROUP_METADATA|BTRFS_BLOCK_GROUP_DATA:
3855                 return "mixed";
3856         case BTRFS_BLOCK_GROUP_METADATA:
3857                 return "metadata";
3858         case BTRFS_BLOCK_GROUP_DATA:
3859                 return "data";
3860         case BTRFS_BLOCK_GROUP_SYSTEM:
3861                 return "system";
3862         default:
3863                 WARN_ON(1);
3864                 return "invalid-combination";
3865         };
3866 }
3867
3868 static int create_space_info(struct btrfs_fs_info *info, u64 flags)
3869 {
3870
3871         struct btrfs_space_info *space_info;
3872         int i;
3873         int ret;
3874
3875         space_info = kzalloc(sizeof(*space_info), GFP_NOFS);
3876         if (!space_info)
3877                 return -ENOMEM;
3878
3879         ret = percpu_counter_init(&space_info->total_bytes_pinned, 0,
3880                                  GFP_KERNEL);
3881         if (ret) {
3882                 kfree(space_info);
3883                 return ret;
3884         }
3885
3886         for (i = 0; i < BTRFS_NR_RAID_TYPES; i++)
3887                 INIT_LIST_HEAD(&space_info->block_groups[i]);
3888         init_rwsem(&space_info->groups_sem);
3889         spin_lock_init(&space_info->lock);
3890         space_info->flags = flags & BTRFS_BLOCK_GROUP_TYPE_MASK;
3891         space_info->force_alloc = CHUNK_ALLOC_NO_FORCE;
3892         init_waitqueue_head(&space_info->wait);
3893         INIT_LIST_HEAD(&space_info->ro_bgs);
3894         INIT_LIST_HEAD(&space_info->tickets);
3895         INIT_LIST_HEAD(&space_info->priority_tickets);
3896
3897         ret = kobject_init_and_add(&space_info->kobj, &space_info_ktype,
3898                                     info->space_info_kobj, "%s",
3899                                     alloc_name(space_info->flags));
3900         if (ret) {
3901                 kobject_put(&space_info->kobj);
3902                 return ret;
3903         }
3904
3905         list_add_rcu(&space_info->list, &info->space_info);
3906         if (flags & BTRFS_BLOCK_GROUP_DATA)
3907                 info->data_sinfo = space_info;
3908
3909         return ret;
3910 }
3911
3912 static void update_space_info(struct btrfs_fs_info *info, u64 flags,
3913                              u64 total_bytes, u64 bytes_used,
3914                              u64 bytes_readonly,
3915                              struct btrfs_space_info **space_info)
3916 {
3917         struct btrfs_space_info *found;
3918         int factor;
3919
3920         factor = btrfs_bg_type_to_factor(flags);
3921
3922         found = __find_space_info(info, flags);
3923         ASSERT(found);
3924         spin_lock(&found->lock);
3925         found->total_bytes += total_bytes;
3926         found->disk_total += total_bytes * factor;
3927         found->bytes_used += bytes_used;
3928         found->disk_used += bytes_used * factor;
3929         found->bytes_readonly += bytes_readonly;
3930         if (total_bytes > 0)
3931                 found->full = 0;
3932         space_info_add_new_bytes(info, found, total_bytes -
3933                                  bytes_used - bytes_readonly);
3934         spin_unlock(&found->lock);
3935         *space_info = found;
3936 }
3937
3938 static void set_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
3939 {
3940         u64 extra_flags = chunk_to_extended(flags) &
3941                                 BTRFS_EXTENDED_PROFILE_MASK;
3942
3943         write_seqlock(&fs_info->profiles_lock);
3944         if (flags & BTRFS_BLOCK_GROUP_DATA)
3945                 fs_info->avail_data_alloc_bits |= extra_flags;
3946         if (flags & BTRFS_BLOCK_GROUP_METADATA)
3947                 fs_info->avail_metadata_alloc_bits |= extra_flags;
3948         if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
3949                 fs_info->avail_system_alloc_bits |= extra_flags;
3950         write_sequnlock(&fs_info->profiles_lock);
3951 }
3952
3953 /*
3954  * returns target flags in extended format or 0 if restripe for this
3955  * chunk_type is not in progress
3956  *
3957  * should be called with balance_lock held
3958  */
3959 static u64 get_restripe_target(struct btrfs_fs_info *fs_info, u64 flags)
3960 {
3961         struct btrfs_balance_control *bctl = fs_info->balance_ctl;
3962         u64 target = 0;
3963
3964         if (!bctl)
3965                 return 0;
3966
3967         if (flags & BTRFS_BLOCK_GROUP_DATA &&
3968             bctl->data.flags & BTRFS_BALANCE_ARGS_CONVERT) {
3969                 target = BTRFS_BLOCK_GROUP_DATA | bctl->data.target;
3970         } else if (flags & BTRFS_BLOCK_GROUP_SYSTEM &&
3971                    bctl->sys.flags & BTRFS_BALANCE_ARGS_CONVERT) {
3972                 target = BTRFS_BLOCK_GROUP_SYSTEM | bctl->sys.target;
3973         } else if (flags & BTRFS_BLOCK_GROUP_METADATA &&
3974                    bctl->meta.flags & BTRFS_BALANCE_ARGS_CONVERT) {
3975                 target = BTRFS_BLOCK_GROUP_METADATA | bctl->meta.target;
3976         }
3977
3978         return target;
3979 }
3980
3981 /*
3982  * @flags: available profiles in extended format (see ctree.h)
3983  *
3984  * Returns reduced profile in chunk format.  If profile changing is in
3985  * progress (either running or paused) picks the target profile (if it's
3986  * already available), otherwise falls back to plain reducing.
3987  */
3988 static u64 btrfs_reduce_alloc_profile(struct btrfs_fs_info *fs_info, u64 flags)
3989 {
3990         u64 num_devices = fs_info->fs_devices->rw_devices;
3991         u64 target;
3992         u64 raid_type;
3993         u64 allowed = 0;
3994
3995         /*
3996          * see if restripe for this chunk_type is in progress, if so
3997          * try to reduce to the target profile
3998          */
3999         spin_lock(&fs_info->balance_lock);
4000         target = get_restripe_target(fs_info, flags);
4001         if (target) {
4002                 /* pick target profile only if it's already available */
4003                 if ((flags & target) & BTRFS_EXTENDED_PROFILE_MASK) {
4004                         spin_unlock(&fs_info->balance_lock);
4005                         return extended_to_chunk(target);
4006                 }
4007         }
4008         spin_unlock(&fs_info->balance_lock);
4009
4010         /* First, mask out the RAID levels which aren't possible */
4011         for (raid_type = 0; raid_type < BTRFS_NR_RAID_TYPES; raid_type++) {
4012                 if (num_devices >= btrfs_raid_array[raid_type].devs_min)
4013                         allowed |= btrfs_raid_array[raid_type].bg_flag;
4014         }
4015         allowed &= flags;
4016
4017         if (allowed & BTRFS_BLOCK_GROUP_RAID6)
4018                 allowed = BTRFS_BLOCK_GROUP_RAID6;
4019         else if (allowed & BTRFS_BLOCK_GROUP_RAID5)
4020                 allowed = BTRFS_BLOCK_GROUP_RAID5;
4021         else if (allowed & BTRFS_BLOCK_GROUP_RAID10)
4022                 allowed = BTRFS_BLOCK_GROUP_RAID10;
4023         else if (allowed & BTRFS_BLOCK_GROUP_RAID1)
4024                 allowed = BTRFS_BLOCK_GROUP_RAID1;
4025         else if (allowed & BTRFS_BLOCK_GROUP_RAID0)
4026                 allowed = BTRFS_BLOCK_GROUP_RAID0;
4027
4028         flags &= ~BTRFS_BLOCK_GROUP_PROFILE_MASK;
4029
4030         return extended_to_chunk(flags | allowed);
4031 }
4032
4033 static u64 get_alloc_profile(struct btrfs_fs_info *fs_info, u64 orig_flags)
4034 {
4035         unsigned seq;
4036         u64 flags;
4037
4038         do {
4039                 flags = orig_flags;
4040                 seq = read_seqbegin(&fs_info->profiles_lock);
4041
4042                 if (flags & BTRFS_BLOCK_GROUP_DATA)
4043                         flags |= fs_info->avail_data_alloc_bits;
4044                 else if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
4045                         flags |= fs_info->avail_system_alloc_bits;
4046                 else if (flags & BTRFS_BLOCK_GROUP_METADATA)
4047                         flags |= fs_info->avail_metadata_alloc_bits;
4048         } while (read_seqretry(&fs_info->profiles_lock, seq));
4049
4050         return btrfs_reduce_alloc_profile(fs_info, flags);
4051 }
4052
4053 static u64 get_alloc_profile_by_root(struct btrfs_root *root, int data)
4054 {
4055         struct btrfs_fs_info *fs_info = root->fs_info;
4056         u64 flags;
4057         u64 ret;
4058
4059         if (data)
4060                 flags = BTRFS_BLOCK_GROUP_DATA;
4061         else if (root == fs_info->chunk_root)
4062                 flags = BTRFS_BLOCK_GROUP_SYSTEM;
4063         else
4064                 flags = BTRFS_BLOCK_GROUP_METADATA;
4065
4066         ret = get_alloc_profile(fs_info, flags);
4067         return ret;
4068 }
4069
4070 u64 btrfs_data_alloc_profile(struct btrfs_fs_info *fs_info)
4071 {
4072         return get_alloc_profile(fs_info, BTRFS_BLOCK_GROUP_DATA);
4073 }
4074
4075 u64 btrfs_metadata_alloc_profile(struct btrfs_fs_info *fs_info)
4076 {
4077         return get_alloc_profile(fs_info, BTRFS_BLOCK_GROUP_METADATA);
4078 }
4079
4080 u64 btrfs_system_alloc_profile(struct btrfs_fs_info *fs_info)
4081 {
4082         return get_alloc_profile(fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
4083 }
4084
4085 static u64 btrfs_space_info_used(struct btrfs_space_info *s_info,
4086                                  bool may_use_included)
4087 {
4088         ASSERT(s_info);
4089         return s_info->bytes_used + s_info->bytes_reserved +
4090                 s_info->bytes_pinned + s_info->bytes_readonly +
4091                 (may_use_included ? s_info->bytes_may_use : 0);
4092 }
4093
4094 int btrfs_alloc_data_chunk_ondemand(struct btrfs_inode *inode, u64 bytes)
4095 {
4096         struct btrfs_root *root = inode->root;
4097         struct btrfs_fs_info *fs_info = root->fs_info;
4098         struct btrfs_space_info *data_sinfo = fs_info->data_sinfo;
4099         u64 used;
4100         int ret = 0;
4101         int need_commit = 2;
4102         int have_pinned_space;
4103
4104         /* make sure bytes are sectorsize aligned */
4105         bytes = ALIGN(bytes, fs_info->sectorsize);
4106
4107         if (btrfs_is_free_space_inode(inode)) {
4108                 need_commit = 0;
4109                 ASSERT(current->journal_info);
4110         }
4111
4112 again:
4113         /* make sure we have enough space to handle the data first */
4114         spin_lock(&data_sinfo->lock);
4115         used = btrfs_space_info_used(data_sinfo, true);
4116
4117         if (used + bytes > data_sinfo->total_bytes) {
4118                 struct btrfs_trans_handle *trans;
4119
4120                 /*
4121                  * if we don't have enough free bytes in this space then we need
4122                  * to alloc a new chunk.
4123                  */
4124                 if (!data_sinfo->full) {
4125                         u64 alloc_target;
4126
4127                         data_sinfo->force_alloc = CHUNK_ALLOC_FORCE;
4128                         spin_unlock(&data_sinfo->lock);
4129
4130                         alloc_target = btrfs_data_alloc_profile(fs_info);
4131                         /*
4132                          * It is ugly that we don't call nolock join
4133                          * transaction for the free space inode case here.
4134                          * But it is safe because we only do the data space
4135                          * reservation for the free space cache in the
4136                          * transaction context, the common join transaction
4137                          * just increase the counter of the current transaction
4138                          * handler, doesn't try to acquire the trans_lock of
4139                          * the fs.
4140                          */
4141                         trans = btrfs_join_transaction(root);
4142                         if (IS_ERR(trans))
4143                                 return PTR_ERR(trans);
4144
4145                         ret = do_chunk_alloc(trans, alloc_target,
4146                                              CHUNK_ALLOC_NO_FORCE);
4147                         btrfs_end_transaction(trans);
4148                         if (ret < 0) {
4149                                 if (ret != -ENOSPC)
4150                                         return ret;
4151                                 else {
4152                                         have_pinned_space = 1;
4153                                         goto commit_trans;
4154                                 }
4155                         }
4156
4157                         goto again;
4158                 }
4159
4160                 /*
4161                  * If we don't have enough pinned space to deal with this
4162                  * allocation, and no removed chunk in current transaction,
4163                  * don't bother committing the transaction.
4164                  */
4165                 have_pinned_space = __percpu_counter_compare(
4166                         &data_sinfo->total_bytes_pinned,
4167                         used + bytes - data_sinfo->total_bytes,
4168                         BTRFS_TOTAL_BYTES_PINNED_BATCH);
4169                 spin_unlock(&data_sinfo->lock);
4170
4171                 /* commit the current transaction and try again */
4172 commit_trans:
4173                 if (need_commit) {
4174                         need_commit--;
4175
4176                         if (need_commit > 0) {
4177                                 btrfs_start_delalloc_roots(fs_info, -1);
4178                                 btrfs_wait_ordered_roots(fs_info, U64_MAX, 0,
4179                                                          (u64)-1);
4180                         }
4181
4182                         trans = btrfs_join_transaction(root);
4183                         if (IS_ERR(trans))
4184                                 return PTR_ERR(trans);
4185                         if (have_pinned_space >= 0 ||
4186                             test_bit(BTRFS_TRANS_HAVE_FREE_BGS,
4187                                      &trans->transaction->flags) ||
4188                             need_commit > 0) {
4189                                 ret = btrfs_commit_transaction(trans);
4190                                 if (ret)
4191                                         return ret;
4192                                 /*
4193                                  * The cleaner kthread might still be doing iput
4194                                  * operations. Wait for it to finish so that
4195                                  * more space is released.  We don't need to
4196                                  * explicitly run the delayed iputs here because
4197                                  * the commit_transaction would have woken up
4198                                  * the cleaner.
4199                                  */
4200                                 ret = btrfs_wait_on_delayed_iputs(fs_info);
4201                                 if (ret)
4202                                         return ret;
4203                                 goto again;
4204                         } else {
4205                                 btrfs_end_transaction(trans);
4206                         }
4207                 }
4208
4209                 trace_btrfs_space_reservation(fs_info,
4210                                               "space_info:enospc",
4211                                               data_sinfo->flags, bytes, 1);
4212                 return -ENOSPC;
4213         }
4214         update_bytes_may_use(fs_info, data_sinfo, bytes);
4215         trace_btrfs_space_reservation(fs_info, "space_info",
4216                                       data_sinfo->flags, bytes, 1);
4217         spin_unlock(&data_sinfo->lock);
4218
4219         return 0;
4220 }
4221
4222 int btrfs_check_data_free_space(struct inode *inode,
4223                         struct extent_changeset **reserved, u64 start, u64 len)
4224 {
4225         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
4226         int ret;
4227
4228         /* align the range */
4229         len = round_up(start + len, fs_info->sectorsize) -
4230               round_down(start, fs_info->sectorsize);
4231         start = round_down(start, fs_info->sectorsize);
4232
4233         ret = btrfs_alloc_data_chunk_ondemand(BTRFS_I(inode), len);
4234         if (ret < 0)
4235                 return ret;
4236
4237         /* Use new btrfs_qgroup_reserve_data to reserve precious data space. */
4238         ret = btrfs_qgroup_reserve_data(inode, reserved, start, len);
4239         if (ret < 0)
4240                 btrfs_free_reserved_data_space_noquota(inode, start, len);
4241         else
4242                 ret = 0;
4243         return ret;
4244 }
4245
4246 /*
4247  * Called if we need to clear a data reservation for this inode
4248  * Normally in a error case.
4249  *
4250  * This one will *NOT* use accurate qgroup reserved space API, just for case
4251  * which we can't sleep and is sure it won't affect qgroup reserved space.
4252  * Like clear_bit_hook().
4253  */
4254 void btrfs_free_reserved_data_space_noquota(struct inode *inode, u64 start,
4255                                             u64 len)
4256 {
4257         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
4258         struct btrfs_space_info *data_sinfo;
4259
4260         /* Make sure the range is aligned to sectorsize */
4261         len = round_up(start + len, fs_info->sectorsize) -
4262               round_down(start, fs_info->sectorsize);
4263         start = round_down(start, fs_info->sectorsize);
4264
4265         data_sinfo = fs_info->data_sinfo;
4266         spin_lock(&data_sinfo->lock);
4267         update_bytes_may_use(fs_info, data_sinfo, -len);
4268         trace_btrfs_space_reservation(fs_info, "space_info",
4269                                       data_sinfo->flags, len, 0);
4270         spin_unlock(&data_sinfo->lock);
4271 }
4272
4273 /*
4274  * Called if we need to clear a data reservation for this inode
4275  * Normally in a error case.
4276  *
4277  * This one will handle the per-inode data rsv map for accurate reserved
4278  * space framework.
4279  */
4280 void btrfs_free_reserved_data_space(struct inode *inode,
4281                         struct extent_changeset *reserved, u64 start, u64 len)
4282 {
4283         struct btrfs_root *root = BTRFS_I(inode)->root;
4284
4285         /* Make sure the range is aligned to sectorsize */
4286         len = round_up(start + len, root->fs_info->sectorsize) -
4287               round_down(start, root->fs_info->sectorsize);
4288         start = round_down(start, root->fs_info->sectorsize);
4289
4290         btrfs_free_reserved_data_space_noquota(inode, start, len);
4291         btrfs_qgroup_free_data(inode, reserved, start, len);
4292 }
4293
4294 static void force_metadata_allocation(struct btrfs_fs_info *info)
4295 {
4296         struct list_head *head = &info->space_info;
4297         struct btrfs_space_info *found;
4298
4299         rcu_read_lock();
4300         list_for_each_entry_rcu(found, head, list) {
4301                 if (found->flags & BTRFS_BLOCK_GROUP_METADATA)
4302                         found->force_alloc = CHUNK_ALLOC_FORCE;
4303         }
4304         rcu_read_unlock();
4305 }
4306
4307 static inline u64 calc_global_rsv_need_space(struct btrfs_block_rsv *global)
4308 {
4309         return (global->size << 1);
4310 }
4311
4312 static int should_alloc_chunk(struct btrfs_fs_info *fs_info,
4313                               struct btrfs_space_info *sinfo, int force)
4314 {
4315         u64 bytes_used = btrfs_space_info_used(sinfo, false);
4316         u64 thresh;
4317
4318         if (force == CHUNK_ALLOC_FORCE)
4319                 return 1;
4320
4321         /*
4322          * in limited mode, we want to have some free space up to
4323          * about 1% of the FS size.
4324          */
4325         if (force == CHUNK_ALLOC_LIMITED) {
4326                 thresh = btrfs_super_total_bytes(fs_info->super_copy);
4327                 thresh = max_t(u64, SZ_64M, div_factor_fine(thresh, 1));
4328
4329                 if (sinfo->total_bytes - bytes_used < thresh)
4330                         return 1;
4331         }
4332
4333         if (bytes_used + SZ_2M < div_factor(sinfo->total_bytes, 8))
4334                 return 0;
4335         return 1;
4336 }
4337
4338 static u64 get_profile_num_devs(struct btrfs_fs_info *fs_info, u64 type)
4339 {
4340         u64 num_dev;
4341
4342         num_dev = btrfs_raid_array[btrfs_bg_flags_to_raid_index(type)].devs_max;
4343         if (!num_dev)
4344                 num_dev = fs_info->fs_devices->rw_devices;
4345
4346         return num_dev;
4347 }
4348
4349 /*
4350  * If @is_allocation is true, reserve space in the system space info necessary
4351  * for allocating a chunk, otherwise if it's false, reserve space necessary for
4352  * removing a chunk.
4353  */
4354 void check_system_chunk(struct btrfs_trans_handle *trans, u64 type)
4355 {
4356         struct btrfs_fs_info *fs_info = trans->fs_info;
4357         struct btrfs_space_info *info;
4358         u64 left;
4359         u64 thresh;
4360         int ret = 0;
4361         u64 num_devs;
4362
4363         /*
4364          * Needed because we can end up allocating a system chunk and for an
4365          * atomic and race free space reservation in the chunk block reserve.
4366          */
4367         lockdep_assert_held(&fs_info->chunk_mutex);
4368
4369         info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
4370         spin_lock(&info->lock);
4371         left = info->total_bytes - btrfs_space_info_used(info, true);
4372         spin_unlock(&info->lock);
4373
4374         num_devs = get_profile_num_devs(fs_info, type);
4375
4376         /* num_devs device items to update and 1 chunk item to add or remove */
4377         thresh = btrfs_calc_trunc_metadata_size(fs_info, num_devs) +
4378                 btrfs_calc_trans_metadata_size(fs_info, 1);
4379
4380         if (left < thresh && btrfs_test_opt(fs_info, ENOSPC_DEBUG)) {
4381                 btrfs_info(fs_info, "left=%llu, need=%llu, flags=%llu",
4382                            left, thresh, type);
4383                 dump_space_info(fs_info, info, 0, 0);
4384         }
4385
4386         if (left < thresh) {
4387                 u64 flags = btrfs_system_alloc_profile(fs_info);
4388
4389                 /*
4390                  * Ignore failure to create system chunk. We might end up not
4391                  * needing it, as we might not need to COW all nodes/leafs from
4392                  * the paths we visit in the chunk tree (they were already COWed
4393                  * or created in the current transaction for example).
4394                  */
4395                 ret = btrfs_alloc_chunk(trans, flags);
4396         }
4397
4398         if (!ret) {
4399                 ret = btrfs_block_rsv_add(fs_info->chunk_root,
4400                                           &fs_info->chunk_block_rsv,
4401                                           thresh, BTRFS_RESERVE_NO_FLUSH);
4402                 if (!ret)
4403                         trans->chunk_bytes_reserved += thresh;
4404         }
4405 }
4406
4407 /*
4408  * If force is CHUNK_ALLOC_FORCE:
4409  *    - return 1 if it successfully allocates a chunk,
4410  *    - return errors including -ENOSPC otherwise.
4411  * If force is NOT CHUNK_ALLOC_FORCE:
4412  *    - return 0 if it doesn't need to allocate a new chunk,
4413  *    - return 1 if it successfully allocates a chunk,
4414  *    - return errors including -ENOSPC otherwise.
4415  */
4416 static int do_chunk_alloc(struct btrfs_trans_handle *trans, u64 flags,
4417                           int force)
4418 {
4419         struct btrfs_fs_info *fs_info = trans->fs_info;
4420         struct btrfs_space_info *space_info;
4421         bool wait_for_alloc = false;
4422         bool should_alloc = false;
4423         int ret = 0;
4424
4425         /* Don't re-enter if we're already allocating a chunk */
4426         if (trans->allocating_chunk)
4427                 return -ENOSPC;
4428
4429         space_info = __find_space_info(fs_info, flags);
4430         ASSERT(space_info);
4431
4432         do {
4433                 spin_lock(&space_info->lock);
4434                 if (force < space_info->force_alloc)
4435                         force = space_info->force_alloc;
4436                 should_alloc = should_alloc_chunk(fs_info, space_info, force);
4437                 if (space_info->full) {
4438                         /* No more free physical space */
4439                         if (should_alloc)
4440                                 ret = -ENOSPC;
4441                         else
4442                                 ret = 0;
4443                         spin_unlock(&space_info->lock);
4444                         return ret;
4445                 } else if (!should_alloc) {
4446                         spin_unlock(&space_info->lock);
4447                         return 0;
4448                 } else if (space_info->chunk_alloc) {
4449                         /*
4450                          * Someone is already allocating, so we need to block
4451                          * until this someone is finished and then loop to
4452                          * recheck if we should continue with our allocation
4453                          * attempt.
4454                          */
4455                         wait_for_alloc = true;
4456                         spin_unlock(&space_info->lock);
4457                         mutex_lock(&fs_info->chunk_mutex);
4458                         mutex_unlock(&fs_info->chunk_mutex);
4459                 } else {
4460                         /* Proceed with allocation */
4461                         space_info->chunk_alloc = 1;
4462                         wait_for_alloc = false;
4463                         spin_unlock(&space_info->lock);
4464                 }
4465
4466                 cond_resched();
4467         } while (wait_for_alloc);
4468
4469         mutex_lock(&fs_info->chunk_mutex);
4470         trans->allocating_chunk = true;
4471
4472         /*
4473          * If we have mixed data/metadata chunks we want to make sure we keep
4474          * allocating mixed chunks instead of individual chunks.
4475          */
4476         if (btrfs_mixed_space_info(space_info))
4477                 flags |= (BTRFS_BLOCK_GROUP_DATA | BTRFS_BLOCK_GROUP_METADATA);
4478
4479         /*
4480          * if we're doing a data chunk, go ahead and make sure that
4481          * we keep a reasonable number of metadata chunks allocated in the
4482          * FS as well.
4483          */
4484         if (flags & BTRFS_BLOCK_GROUP_DATA && fs_info->metadata_ratio) {
4485                 fs_info->data_chunk_allocations++;
4486                 if (!(fs_info->data_chunk_allocations %
4487                       fs_info->metadata_ratio))
4488                         force_metadata_allocation(fs_info);
4489         }
4490
4491         /*
4492          * Check if we have enough space in SYSTEM chunk because we may need
4493          * to update devices.
4494          */
4495         check_system_chunk(trans, flags);
4496
4497         ret = btrfs_alloc_chunk(trans, flags);
4498         trans->allocating_chunk = false;
4499
4500         spin_lock(&space_info->lock);
4501         if (ret < 0) {
4502                 if (ret == -ENOSPC)
4503                         space_info->full = 1;
4504                 else
4505                         goto out;
4506         } else {
4507                 ret = 1;
4508                 space_info->max_extent_size = 0;
4509         }
4510
4511         space_info->force_alloc = CHUNK_ALLOC_NO_FORCE;
4512 out:
4513         space_info->chunk_alloc = 0;
4514         spin_unlock(&space_info->lock);
4515         mutex_unlock(&fs_info->chunk_mutex);
4516         /*
4517          * When we allocate a new chunk we reserve space in the chunk block
4518          * reserve to make sure we can COW nodes/leafs in the chunk tree or
4519          * add new nodes/leafs to it if we end up needing to do it when
4520          * inserting the chunk item and updating device items as part of the
4521          * second phase of chunk allocation, performed by
4522          * btrfs_finish_chunk_alloc(). So make sure we don't accumulate a
4523          * large number of new block groups to create in our transaction
4524          * handle's new_bgs list to avoid exhausting the chunk block reserve
4525          * in extreme cases - like having a single transaction create many new
4526          * block groups when starting to write out the free space caches of all
4527          * the block groups that were made dirty during the lifetime of the
4528          * transaction.
4529          */
4530         if (trans->chunk_bytes_reserved >= (u64)SZ_2M)
4531                 btrfs_create_pending_block_groups(trans);
4532
4533         return ret;
4534 }
4535
4536 static int can_overcommit(struct btrfs_fs_info *fs_info,
4537                           struct btrfs_space_info *space_info, u64 bytes,
4538                           enum btrfs_reserve_flush_enum flush,
4539                           bool system_chunk)
4540 {
4541         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
4542         u64 profile;
4543         u64 space_size;
4544         u64 avail;
4545         u64 used;
4546         int factor;
4547
4548         /* Don't overcommit when in mixed mode. */
4549         if (space_info->flags & BTRFS_BLOCK_GROUP_DATA)
4550                 return 0;
4551
4552         if (system_chunk)
4553                 profile = btrfs_system_alloc_profile(fs_info);
4554         else
4555                 profile = btrfs_metadata_alloc_profile(fs_info);
4556
4557         used = btrfs_space_info_used(space_info, false);
4558
4559         /*
4560          * We only want to allow over committing if we have lots of actual space
4561          * free, but if we don't have enough space to handle the global reserve
4562          * space then we could end up having a real enospc problem when trying
4563          * to allocate a chunk or some other such important allocation.
4564          */
4565         spin_lock(&global_rsv->lock);
4566         space_size = calc_global_rsv_need_space(global_rsv);
4567         spin_unlock(&global_rsv->lock);
4568         if (used + space_size >= space_info->total_bytes)
4569                 return 0;
4570
4571         used += space_info->bytes_may_use;
4572
4573         avail = atomic64_read(&fs_info->free_chunk_space);
4574
4575         /*
4576          * If we have dup, raid1 or raid10 then only half of the free
4577          * space is actually usable.  For raid56, the space info used
4578          * doesn't include the parity drive, so we don't have to
4579          * change the math
4580          */
4581         factor = btrfs_bg_type_to_factor(profile);
4582         avail = div_u64(avail, factor);
4583
4584         /*
4585          * If we aren't flushing all things, let us overcommit up to
4586          * 1/2th of the space. If we can flush, don't let us overcommit
4587          * too much, let it overcommit up to 1/8 of the space.
4588          */
4589         if (flush == BTRFS_RESERVE_FLUSH_ALL)
4590                 avail >>= 3;
4591         else
4592                 avail >>= 1;
4593
4594         if (used + bytes < space_info->total_bytes + avail)
4595                 return 1;
4596         return 0;
4597 }
4598
4599 static void btrfs_writeback_inodes_sb_nr(struct btrfs_fs_info *fs_info,
4600                                          unsigned long nr_pages, int nr_items)
4601 {
4602         struct super_block *sb = fs_info->sb;
4603
4604         if (down_read_trylock(&sb->s_umount)) {
4605                 writeback_inodes_sb_nr(sb, nr_pages, WB_REASON_FS_FREE_SPACE);
4606                 up_read(&sb->s_umount);
4607         } else {
4608                 /*
4609                  * We needn't worry the filesystem going from r/w to r/o though
4610                  * we don't acquire ->s_umount mutex, because the filesystem
4611                  * should guarantee the delalloc inodes list be empty after
4612                  * the filesystem is readonly(all dirty pages are written to
4613                  * the disk).
4614                  */
4615                 btrfs_start_delalloc_roots(fs_info, nr_items);
4616                 if (!current->journal_info)
4617                         btrfs_wait_ordered_roots(fs_info, nr_items, 0, (u64)-1);
4618         }
4619 }
4620
4621 static inline u64 calc_reclaim_items_nr(struct btrfs_fs_info *fs_info,
4622                                         u64 to_reclaim)
4623 {
4624         u64 bytes;
4625         u64 nr;
4626
4627         bytes = btrfs_calc_trans_metadata_size(fs_info, 1);
4628         nr = div64_u64(to_reclaim, bytes);
4629         if (!nr)
4630                 nr = 1;
4631         return nr;
4632 }
4633
4634 #define EXTENT_SIZE_PER_ITEM    SZ_256K
4635
4636 /*
4637  * shrink metadata reservation for delalloc
4638  */
4639 static void shrink_delalloc(struct btrfs_fs_info *fs_info, u64 to_reclaim,
4640                             u64 orig, bool wait_ordered)
4641 {
4642         struct btrfs_space_info *space_info;
4643         struct btrfs_trans_handle *trans;
4644         u64 delalloc_bytes;
4645         u64 dio_bytes;
4646         u64 async_pages;
4647         u64 items;
4648         long time_left;
4649         unsigned long nr_pages;
4650         int loops;
4651
4652         /* Calc the number of the pages we need flush for space reservation */
4653         items = calc_reclaim_items_nr(fs_info, to_reclaim);
4654         to_reclaim = items * EXTENT_SIZE_PER_ITEM;
4655
4656         trans = (struct btrfs_trans_handle *)current->journal_info;
4657         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
4658
4659         delalloc_bytes = percpu_counter_sum_positive(
4660                                                 &fs_info->delalloc_bytes);
4661         dio_bytes = percpu_counter_sum_positive(&fs_info->dio_bytes);
4662         if (delalloc_bytes == 0 && dio_bytes == 0) {
4663                 if (trans)
4664                         return;
4665                 if (wait_ordered)
4666                         btrfs_wait_ordered_roots(fs_info, items, 0, (u64)-1);
4667                 return;
4668         }
4669
4670         /*
4671          * If we are doing more ordered than delalloc we need to just wait on
4672          * ordered extents, otherwise we'll waste time trying to flush delalloc
4673          * that likely won't give us the space back we need.
4674          */
4675         if (dio_bytes > delalloc_bytes)
4676                 wait_ordered = true;
4677
4678         loops = 0;
4679         while ((delalloc_bytes || dio_bytes) && loops < 3) {
4680                 nr_pages = min(delalloc_bytes, to_reclaim) >> PAGE_SHIFT;
4681
4682                 /*
4683                  * Triggers inode writeback for up to nr_pages. This will invoke
4684                  * ->writepages callback and trigger delalloc filling
4685                  *  (btrfs_run_delalloc_range()).
4686                  */
4687                 btrfs_writeback_inodes_sb_nr(fs_info, nr_pages, items);
4688
4689                 /*
4690                  * We need to wait for the compressed pages to start before
4691                  * we continue.
4692                  */
4693                 async_pages = atomic_read(&fs_info->async_delalloc_pages);
4694                 if (!async_pages)
4695                         goto skip_async;
4696
4697                 /*
4698                  * Calculate how many compressed pages we want to be written
4699                  * before we continue. I.e if there are more async pages than we
4700                  * require wait_event will wait until nr_pages are written.
4701                  */
4702                 if (async_pages <= nr_pages)
4703                         async_pages = 0;
4704                 else
4705                         async_pages -= nr_pages;
4706
4707                 wait_event(fs_info->async_submit_wait,
4708                            atomic_read(&fs_info->async_delalloc_pages) <=
4709                            (int)async_pages);
4710 skip_async:
4711                 spin_lock(&space_info->lock);
4712                 if (list_empty(&space_info->tickets) &&
4713                     list_empty(&space_info->priority_tickets)) {
4714                         spin_unlock(&space_info->lock);
4715                         break;
4716                 }
4717                 spin_unlock(&space_info->lock);
4718
4719                 loops++;
4720                 if (wait_ordered && !trans) {
4721                         btrfs_wait_ordered_roots(fs_info, items, 0, (u64)-1);
4722                 } else {
4723                         time_left = schedule_timeout_killable(1);
4724                         if (time_left)
4725                                 break;
4726                 }
4727                 delalloc_bytes = percpu_counter_sum_positive(
4728                                                 &fs_info->delalloc_bytes);
4729                 dio_bytes = percpu_counter_sum_positive(&fs_info->dio_bytes);
4730         }
4731 }
4732
4733 struct reserve_ticket {
4734         u64 orig_bytes;
4735         u64 bytes;
4736         int error;
4737         struct list_head list;
4738         wait_queue_head_t wait;
4739 };
4740
4741 /**
4742  * maybe_commit_transaction - possibly commit the transaction if its ok to
4743  * @root - the root we're allocating for
4744  * @bytes - the number of bytes we want to reserve
4745  * @force - force the commit
4746  *
4747  * This will check to make sure that committing the transaction will actually
4748  * get us somewhere and then commit the transaction if it does.  Otherwise it
4749  * will return -ENOSPC.
4750  */
4751 static int may_commit_transaction(struct btrfs_fs_info *fs_info,
4752                                   struct btrfs_space_info *space_info)
4753 {
4754         struct reserve_ticket *ticket = NULL;
4755         struct btrfs_block_rsv *delayed_rsv = &fs_info->delayed_block_rsv;
4756         struct btrfs_block_rsv *delayed_refs_rsv = &fs_info->delayed_refs_rsv;
4757         struct btrfs_trans_handle *trans;
4758         u64 bytes_needed;
4759         u64 reclaim_bytes = 0;
4760
4761         trans = (struct btrfs_trans_handle *)current->journal_info;
4762         if (trans)
4763                 return -EAGAIN;
4764
4765         spin_lock(&space_info->lock);
4766         if (!list_empty(&space_info->priority_tickets))
4767                 ticket = list_first_entry(&space_info->priority_tickets,
4768                                           struct reserve_ticket, list);
4769         else if (!list_empty(&space_info->tickets))
4770                 ticket = list_first_entry(&space_info->tickets,
4771                                           struct reserve_ticket, list);
4772         bytes_needed = (ticket) ? ticket->bytes : 0;
4773         spin_unlock(&space_info->lock);
4774
4775         if (!bytes_needed)
4776                 return 0;
4777
4778         trans = btrfs_join_transaction(fs_info->extent_root);
4779         if (IS_ERR(trans))
4780                 return PTR_ERR(trans);
4781
4782         /*
4783          * See if there is enough pinned space to make this reservation, or if
4784          * we have block groups that are going to be freed, allowing us to
4785          * possibly do a chunk allocation the next loop through.
4786          */
4787         if (test_bit(BTRFS_TRANS_HAVE_FREE_BGS, &trans->transaction->flags) ||
4788             __percpu_counter_compare(&space_info->total_bytes_pinned,
4789                                      bytes_needed,
4790                                      BTRFS_TOTAL_BYTES_PINNED_BATCH) >= 0)
4791                 goto commit;
4792
4793         /*
4794          * See if there is some space in the delayed insertion reservation for
4795          * this reservation.
4796          */
4797         if (space_info != delayed_rsv->space_info)
4798                 goto enospc;
4799
4800         spin_lock(&delayed_rsv->lock);
4801         reclaim_bytes += delayed_rsv->reserved;
4802         spin_unlock(&delayed_rsv->lock);
4803
4804         spin_lock(&delayed_refs_rsv->lock);
4805         reclaim_bytes += delayed_refs_rsv->reserved;
4806         spin_unlock(&delayed_refs_rsv->lock);
4807         if (reclaim_bytes >= bytes_needed)
4808                 goto commit;
4809         bytes_needed -= reclaim_bytes;
4810
4811         if (__percpu_counter_compare(&space_info->total_bytes_pinned,
4812                                    bytes_needed,
4813                                    BTRFS_TOTAL_BYTES_PINNED_BATCH) < 0)
4814                 goto enospc;
4815
4816 commit:
4817         return btrfs_commit_transaction(trans);
4818 enospc:
4819         btrfs_end_transaction(trans);
4820         return -ENOSPC;
4821 }
4822
4823 /*
4824  * Try to flush some data based on policy set by @state. This is only advisory
4825  * and may fail for various reasons. The caller is supposed to examine the
4826  * state of @space_info to detect the outcome.
4827  */
4828 static void flush_space(struct btrfs_fs_info *fs_info,
4829                        struct btrfs_space_info *space_info, u64 num_bytes,
4830                        int state)
4831 {
4832         struct btrfs_root *root = fs_info->extent_root;
4833         struct btrfs_trans_handle *trans;
4834         int nr;
4835         int ret = 0;
4836
4837         switch (state) {
4838         case FLUSH_DELAYED_ITEMS_NR:
4839         case FLUSH_DELAYED_ITEMS:
4840                 if (state == FLUSH_DELAYED_ITEMS_NR)
4841                         nr = calc_reclaim_items_nr(fs_info, num_bytes) * 2;
4842                 else
4843                         nr = -1;
4844
4845                 trans = btrfs_join_transaction(root);
4846                 if (IS_ERR(trans)) {
4847                         ret = PTR_ERR(trans);
4848                         break;
4849                 }
4850                 ret = btrfs_run_delayed_items_nr(trans, nr);
4851                 btrfs_end_transaction(trans);
4852                 break;
4853         case FLUSH_DELALLOC:
4854         case FLUSH_DELALLOC_WAIT:
4855                 shrink_delalloc(fs_info, num_bytes * 2, num_bytes,
4856                                 state == FLUSH_DELALLOC_WAIT);
4857                 break;
4858         case FLUSH_DELAYED_REFS_NR:
4859         case FLUSH_DELAYED_REFS:
4860                 trans = btrfs_join_transaction(root);
4861                 if (IS_ERR(trans)) {
4862                         ret = PTR_ERR(trans);
4863                         break;
4864                 }
4865                 if (state == FLUSH_DELAYED_REFS_NR)
4866                         nr = calc_reclaim_items_nr(fs_info, num_bytes);
4867                 else
4868                         nr = 0;
4869                 btrfs_run_delayed_refs(trans, nr);
4870                 btrfs_end_transaction(trans);
4871                 break;
4872         case ALLOC_CHUNK:
4873         case ALLOC_CHUNK_FORCE:
4874                 trans = btrfs_join_transaction(root);
4875                 if (IS_ERR(trans)) {
4876                         ret = PTR_ERR(trans);
4877                         break;
4878                 }
4879                 ret = do_chunk_alloc(trans,
4880                                      btrfs_metadata_alloc_profile(fs_info),
4881                                      (state == ALLOC_CHUNK) ?
4882                                       CHUNK_ALLOC_NO_FORCE : CHUNK_ALLOC_FORCE);
4883                 btrfs_end_transaction(trans);
4884                 if (ret > 0 || ret == -ENOSPC)
4885                         ret = 0;
4886                 break;
4887         case COMMIT_TRANS:
4888                 /*
4889                  * If we have pending delayed iputs then we could free up a
4890                  * bunch of pinned space, so make sure we run the iputs before
4891                  * we do our pinned bytes check below.
4892                  */
4893                 btrfs_run_delayed_iputs(fs_info);
4894                 btrfs_wait_on_delayed_iputs(fs_info);
4895
4896                 ret = may_commit_transaction(fs_info, space_info);
4897                 break;
4898         default:
4899                 ret = -ENOSPC;
4900                 break;
4901         }
4902
4903         trace_btrfs_flush_space(fs_info, space_info->flags, num_bytes, state,
4904                                 ret);
4905         return;
4906 }
4907
4908 static inline u64
4909 btrfs_calc_reclaim_metadata_size(struct btrfs_fs_info *fs_info,
4910                                  struct btrfs_space_info *space_info,
4911                                  bool system_chunk)
4912 {
4913         struct reserve_ticket *ticket;
4914         u64 used;
4915         u64 expected;
4916         u64 to_reclaim = 0;
4917
4918         list_for_each_entry(ticket, &space_info->tickets, list)
4919                 to_reclaim += ticket->bytes;
4920         list_for_each_entry(ticket, &space_info->priority_tickets, list)
4921                 to_reclaim += ticket->bytes;
4922         if (to_reclaim)
4923                 return to_reclaim;
4924
4925         to_reclaim = min_t(u64, num_online_cpus() * SZ_1M, SZ_16M);
4926         if (can_overcommit(fs_info, space_info, to_reclaim,
4927                            BTRFS_RESERVE_FLUSH_ALL, system_chunk))
4928                 return 0;
4929
4930         used = btrfs_space_info_used(space_info, true);
4931
4932         if (can_overcommit(fs_info, space_info, SZ_1M,
4933                            BTRFS_RESERVE_FLUSH_ALL, system_chunk))
4934                 expected = div_factor_fine(space_info->total_bytes, 95);
4935         else
4936                 expected = div_factor_fine(space_info->total_bytes, 90);
4937
4938         if (used > expected)
4939                 to_reclaim = used - expected;
4940         else
4941                 to_reclaim = 0;
4942         to_reclaim = min(to_reclaim, space_info->bytes_may_use +
4943                                      space_info->bytes_reserved);
4944         return to_reclaim;
4945 }
4946
4947 static inline int need_do_async_reclaim(struct btrfs_fs_info *fs_info,
4948                                         struct btrfs_space_info *space_info,
4949                                         u64 used, bool system_chunk)
4950 {
4951         u64 thresh = div_factor_fine(space_info->total_bytes, 98);
4952
4953         /* If we're just plain full then async reclaim just slows us down. */
4954         if ((space_info->bytes_used + space_info->bytes_reserved) >= thresh)
4955                 return 0;
4956
4957         if (!btrfs_calc_reclaim_metadata_size(fs_info, space_info,
4958                                               system_chunk))
4959                 return 0;
4960
4961         return (used >= thresh && !btrfs_fs_closing(fs_info) &&
4962                 !test_bit(BTRFS_FS_STATE_REMOUNTING, &fs_info->fs_state));
4963 }
4964
4965 static bool wake_all_tickets(struct list_head *head)
4966 {
4967         struct reserve_ticket *ticket;
4968
4969         while (!list_empty(head)) {
4970                 ticket = list_first_entry(head, struct reserve_ticket, list);
4971                 list_del_init(&ticket->list);
4972                 ticket->error = -ENOSPC;
4973                 wake_up(&ticket->wait);
4974                 if (ticket->bytes != ticket->orig_bytes)
4975                         return true;
4976         }
4977         return false;
4978 }
4979
4980 /*
4981  * This is for normal flushers, we can wait all goddamned day if we want to.  We
4982  * will loop and continuously try to flush as long as we are making progress.
4983  * We count progress as clearing off tickets each time we have to loop.
4984  */
4985 static void btrfs_async_reclaim_metadata_space(struct work_struct *work)
4986 {
4987         struct btrfs_fs_info *fs_info;
4988         struct btrfs_space_info *space_info;
4989         u64 to_reclaim;
4990         int flush_state;
4991         int commit_cycles = 0;
4992         u64 last_tickets_id;
4993
4994         fs_info = container_of(work, struct btrfs_fs_info, async_reclaim_work);
4995         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
4996
4997         spin_lock(&space_info->lock);
4998         to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info, space_info,
4999                                                       false);
5000         if (!to_reclaim) {
5001                 space_info->flush = 0;
5002                 spin_unlock(&space_info->lock);
5003                 return;
5004         }
5005         last_tickets_id = space_info->tickets_id;
5006         spin_unlock(&space_info->lock);
5007
5008         flush_state = FLUSH_DELAYED_ITEMS_NR;
5009         do {
5010                 flush_space(fs_info, space_info, to_reclaim, flush_state);
5011                 spin_lock(&space_info->lock);
5012                 if (list_empty(&space_info->tickets)) {
5013                         space_info->flush = 0;
5014                         spin_unlock(&space_info->lock);
5015                         return;
5016                 }
5017                 to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info,
5018                                                               space_info,
5019                                                               false);
5020                 if (last_tickets_id == space_info->tickets_id) {
5021                         flush_state++;
5022                 } else {
5023                         last_tickets_id = space_info->tickets_id;
5024                         flush_state = FLUSH_DELAYED_ITEMS_NR;
5025                         if (commit_cycles)
5026                                 commit_cycles--;
5027                 }
5028
5029                 /*
5030                  * We don't want to force a chunk allocation until we've tried
5031                  * pretty hard to reclaim space.  Think of the case where we
5032                  * freed up a bunch of space and so have a lot of pinned space
5033                  * to reclaim.  We would rather use that than possibly create a
5034                  * underutilized metadata chunk.  So if this is our first run
5035                  * through the flushing state machine skip ALLOC_CHUNK_FORCE and
5036                  * commit the transaction.  If nothing has changed the next go
5037                  * around then we can force a chunk allocation.
5038                  */
5039                 if (flush_state == ALLOC_CHUNK_FORCE && !commit_cycles)
5040                         flush_state++;
5041
5042                 if (flush_state > COMMIT_TRANS) {
5043                         commit_cycles++;
5044                         if (commit_cycles > 2) {
5045                                 if (wake_all_tickets(&space_info->tickets)) {
5046                                         flush_state = FLUSH_DELAYED_ITEMS_NR;
5047                                         commit_cycles--;
5048                                 } else {
5049                                         space_info->flush = 0;
5050                                 }
5051                         } else {
5052                                 flush_state = FLUSH_DELAYED_ITEMS_NR;
5053                         }
5054                 }
5055                 spin_unlock(&space_info->lock);
5056         } while (flush_state <= COMMIT_TRANS);
5057 }
5058
5059 void btrfs_init_async_reclaim_work(struct work_struct *work)
5060 {
5061         INIT_WORK(work, btrfs_async_reclaim_metadata_space);
5062 }
5063
5064 static const enum btrfs_flush_state priority_flush_states[] = {
5065         FLUSH_DELAYED_ITEMS_NR,
5066         FLUSH_DELAYED_ITEMS,
5067         ALLOC_CHUNK,
5068 };
5069
5070 static void priority_reclaim_metadata_space(struct btrfs_fs_info *fs_info,
5071                                             struct btrfs_space_info *space_info,
5072                                             struct reserve_ticket *ticket)
5073 {
5074         u64 to_reclaim;
5075         int flush_state;
5076
5077         spin_lock(&space_info->lock);
5078         to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info, space_info,
5079                                                       false);
5080         if (!to_reclaim) {
5081                 spin_unlock(&space_info->lock);
5082                 return;
5083         }
5084         spin_unlock(&space_info->lock);
5085
5086         flush_state = 0;
5087         do {
5088                 flush_space(fs_info, space_info, to_reclaim,
5089                             priority_flush_states[flush_state]);
5090                 flush_state++;
5091                 spin_lock(&space_info->lock);
5092                 if (ticket->bytes == 0) {
5093                         spin_unlock(&space_info->lock);
5094                         return;
5095                 }
5096                 spin_unlock(&space_info->lock);
5097         } while (flush_state < ARRAY_SIZE(priority_flush_states));
5098 }
5099
5100 static int wait_reserve_ticket(struct btrfs_fs_info *fs_info,
5101                                struct btrfs_space_info *space_info,
5102                                struct reserve_ticket *ticket)
5103
5104 {
5105         DEFINE_WAIT(wait);
5106         u64 reclaim_bytes = 0;
5107         int ret = 0;
5108
5109         spin_lock(&space_info->lock);
5110         while (ticket->bytes > 0 && ticket->error == 0) {
5111                 ret = prepare_to_wait_event(&ticket->wait, &wait, TASK_KILLABLE);
5112                 if (ret) {
5113                         ret = -EINTR;
5114                         break;
5115                 }
5116                 spin_unlock(&space_info->lock);
5117
5118                 schedule();
5119
5120                 finish_wait(&ticket->wait, &wait);
5121                 spin_lock(&space_info->lock);
5122         }
5123         if (!ret)
5124                 ret = ticket->error;
5125         if (!list_empty(&ticket->list))
5126                 list_del_init(&ticket->list);
5127         if (ticket->bytes && ticket->bytes < ticket->orig_bytes)
5128                 reclaim_bytes = ticket->orig_bytes - ticket->bytes;
5129         spin_unlock(&space_info->lock);
5130
5131         if (reclaim_bytes)
5132                 space_info_add_old_bytes(fs_info, space_info, reclaim_bytes);
5133         return ret;
5134 }
5135
5136 /**
5137  * reserve_metadata_bytes - try to reserve bytes from the block_rsv's space
5138  * @root - the root we're allocating for
5139  * @space_info - the space info we want to allocate from
5140  * @orig_bytes - the number of bytes we want
5141  * @flush - whether or not we can flush to make our reservation
5142  *
5143  * This will reserve orig_bytes number of bytes from the space info associated
5144  * with the block_rsv.  If there is not enough space it will make an attempt to
5145  * flush out space to make room.  It will do this by flushing delalloc if
5146  * possible or committing the transaction.  If flush is 0 then no attempts to
5147  * regain reservations will be made and this will fail if there is not enough
5148  * space already.
5149  */
5150 static int __reserve_metadata_bytes(struct btrfs_fs_info *fs_info,
5151                                     struct btrfs_space_info *space_info,
5152                                     u64 orig_bytes,
5153                                     enum btrfs_reserve_flush_enum flush,
5154                                     bool system_chunk)
5155 {
5156         struct reserve_ticket ticket;
5157         u64 used;
5158         u64 reclaim_bytes = 0;
5159         int ret = 0;
5160
5161         ASSERT(orig_bytes);
5162         ASSERT(!current->journal_info || flush != BTRFS_RESERVE_FLUSH_ALL);
5163
5164         spin_lock(&space_info->lock);
5165         ret = -ENOSPC;
5166         used = btrfs_space_info_used(space_info, true);
5167
5168         /*
5169          * If we have enough space then hooray, make our reservation and carry
5170          * on.  If not see if we can overcommit, and if we can, hooray carry on.
5171          * If not things get more complicated.
5172          */
5173         if (used + orig_bytes <= space_info->total_bytes) {
5174                 update_bytes_may_use(fs_info, space_info, orig_bytes);
5175                 trace_btrfs_space_reservation(fs_info, "space_info",
5176                                               space_info->flags, orig_bytes, 1);
5177                 ret = 0;
5178         } else if (can_overcommit(fs_info, space_info, orig_bytes, flush,
5179                                   system_chunk)) {
5180                 update_bytes_may_use(fs_info, space_info, orig_bytes);
5181                 trace_btrfs_space_reservation(fs_info, "space_info",
5182                                               space_info->flags, orig_bytes, 1);
5183                 ret = 0;
5184         }
5185
5186         /*
5187          * If we couldn't make a reservation then setup our reservation ticket
5188          * and kick the async worker if it's not already running.
5189          *
5190          * If we are a priority flusher then we just need to add our ticket to
5191          * the list and we will do our own flushing further down.
5192          */
5193         if (ret && flush != BTRFS_RESERVE_NO_FLUSH) {
5194                 ticket.orig_bytes = orig_bytes;
5195                 ticket.bytes = orig_bytes;
5196                 ticket.error = 0;
5197                 init_waitqueue_head(&ticket.wait);
5198                 if (flush == BTRFS_RESERVE_FLUSH_ALL) {
5199                         list_add_tail(&ticket.list, &space_info->tickets);
5200                         if (!space_info->flush) {
5201                                 space_info->flush = 1;
5202                                 trace_btrfs_trigger_flush(fs_info,
5203                                                           space_info->flags,
5204                                                           orig_bytes, flush,
5205                                                           "enospc");
5206                                 queue_work(system_unbound_wq,
5207                                            &fs_info->async_reclaim_work);
5208                         }
5209                 } else {
5210                         list_add_tail(&ticket.list,
5211                                       &space_info->priority_tickets);
5212                 }
5213         } else if (!ret && space_info->flags & BTRFS_BLOCK_GROUP_METADATA) {
5214                 used += orig_bytes;
5215                 /*
5216                  * We will do the space reservation dance during log replay,
5217                  * which means we won't have fs_info->fs_root set, so don't do
5218                  * the async reclaim as we will panic.
5219                  */
5220                 if (!test_bit(BTRFS_FS_LOG_RECOVERING, &fs_info->flags) &&
5221                     need_do_async_reclaim(fs_info, space_info,
5222                                           used, system_chunk) &&
5223                     !work_busy(&fs_info->async_reclaim_work)) {
5224                         trace_btrfs_trigger_flush(fs_info, space_info->flags,
5225                                                   orig_bytes, flush, "preempt");
5226                         queue_work(system_unbound_wq,
5227                                    &fs_info->async_reclaim_work);
5228                 }
5229         }
5230         spin_unlock(&space_info->lock);
5231         if (!ret || flush == BTRFS_RESERVE_NO_FLUSH)
5232                 return ret;
5233
5234         if (flush == BTRFS_RESERVE_FLUSH_ALL)
5235                 return wait_reserve_ticket(fs_info, space_info, &ticket);
5236
5237         ret = 0;
5238         priority_reclaim_metadata_space(fs_info, space_info, &ticket);
5239         spin_lock(&space_info->lock);
5240         if (ticket.bytes) {
5241                 if (ticket.bytes < orig_bytes)
5242                         reclaim_bytes = orig_bytes - ticket.bytes;
5243                 list_del_init(&ticket.list);
5244                 ret = -ENOSPC;
5245         }
5246         spin_unlock(&space_info->lock);
5247
5248         if (reclaim_bytes)
5249                 space_info_add_old_bytes(fs_info, space_info, reclaim_bytes);
5250         ASSERT(list_empty(&ticket.list));
5251         return ret;
5252 }
5253
5254 /**
5255  * reserve_metadata_bytes - try to reserve bytes from the block_rsv's space
5256  * @root - the root we're allocating for
5257  * @block_rsv - the block_rsv we're allocating for
5258  * @orig_bytes - the number of bytes we want
5259  * @flush - whether or not we can flush to make our reservation
5260  *
5261  * This will reserve orig_bytes number of bytes from the space info associated
5262  * with the block_rsv.  If there is not enough space it will make an attempt to
5263  * flush out space to make room.  It will do this by flushing delalloc if
5264  * possible or committing the transaction.  If flush is 0 then no attempts to
5265  * regain reservations will be made and this will fail if there is not enough
5266  * space already.
5267  */
5268 static int reserve_metadata_bytes(struct btrfs_root *root,
5269                                   struct btrfs_block_rsv *block_rsv,
5270                                   u64 orig_bytes,
5271                                   enum btrfs_reserve_flush_enum flush)
5272 {
5273         struct btrfs_fs_info *fs_info = root->fs_info;
5274         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
5275         int ret;
5276         bool system_chunk = (root == fs_info->chunk_root);
5277
5278         ret = __reserve_metadata_bytes(fs_info, block_rsv->space_info,
5279                                        orig_bytes, flush, system_chunk);
5280         if (ret == -ENOSPC &&
5281             unlikely(root->orphan_cleanup_state == ORPHAN_CLEANUP_STARTED)) {
5282                 if (block_rsv != global_rsv &&
5283                     !block_rsv_use_bytes(global_rsv, orig_bytes))
5284                         ret = 0;
5285         }
5286         if (ret == -ENOSPC) {
5287                 trace_btrfs_space_reservation(fs_info, "space_info:enospc",
5288                                               block_rsv->space_info->flags,
5289                                               orig_bytes, 1);
5290
5291                 if (btrfs_test_opt(fs_info, ENOSPC_DEBUG))
5292                         dump_space_info(fs_info, block_rsv->space_info,
5293                                         orig_bytes, 0);
5294         }
5295         return ret;
5296 }
5297
5298 static struct btrfs_block_rsv *get_block_rsv(
5299                                         const struct btrfs_trans_handle *trans,
5300                                         const struct btrfs_root *root)
5301 {
5302         struct btrfs_fs_info *fs_info = root->fs_info;
5303         struct btrfs_block_rsv *block_rsv = NULL;
5304
5305         if (test_bit(BTRFS_ROOT_REF_COWS, &root->state) ||
5306             (root == fs_info->csum_root && trans->adding_csums) ||
5307             (root == fs_info->uuid_root))
5308                 block_rsv = trans->block_rsv;
5309
5310         if (!block_rsv)
5311                 block_rsv = root->block_rsv;
5312
5313         if (!block_rsv)
5314                 block_rsv = &fs_info->empty_block_rsv;
5315
5316         return block_rsv;
5317 }
5318
5319 static int block_rsv_use_bytes(struct btrfs_block_rsv *block_rsv,
5320                                u64 num_bytes)
5321 {
5322         int ret = -ENOSPC;
5323         spin_lock(&block_rsv->lock);
5324         if (block_rsv->reserved >= num_bytes) {
5325                 block_rsv->reserved -= num_bytes;
5326                 if (block_rsv->reserved < block_rsv->size)
5327                         block_rsv->full = 0;
5328                 ret = 0;
5329         }
5330         spin_unlock(&block_rsv->lock);
5331         return ret;
5332 }
5333
5334 static void block_rsv_add_bytes(struct btrfs_block_rsv *block_rsv,
5335                                 u64 num_bytes, bool update_size)
5336 {
5337         spin_lock(&block_rsv->lock);
5338         block_rsv->reserved += num_bytes;
5339         if (update_size)
5340                 block_rsv->size += num_bytes;
5341         else if (block_rsv->reserved >= block_rsv->size)
5342                 block_rsv->full = 1;
5343         spin_unlock(&block_rsv->lock);
5344 }
5345
5346 int btrfs_cond_migrate_bytes(struct btrfs_fs_info *fs_info,
5347                              struct btrfs_block_rsv *dest, u64 num_bytes,
5348                              int min_factor)
5349 {
5350         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
5351         u64 min_bytes;
5352
5353         if (global_rsv->space_info != dest->space_info)
5354                 return -ENOSPC;
5355
5356         spin_lock(&global_rsv->lock);
5357         min_bytes = div_factor(global_rsv->size, min_factor);
5358         if (global_rsv->reserved < min_bytes + num_bytes) {
5359                 spin_unlock(&global_rsv->lock);
5360                 return -ENOSPC;
5361         }
5362         global_rsv->reserved -= num_bytes;
5363         if (global_rsv->reserved < global_rsv->size)
5364                 global_rsv->full = 0;
5365         spin_unlock(&global_rsv->lock);
5366
5367         block_rsv_add_bytes(dest, num_bytes, true);
5368         return 0;
5369 }
5370
5371 /**
5372  * btrfs_migrate_to_delayed_refs_rsv - transfer bytes to our delayed refs rsv.
5373  * @fs_info - the fs info for our fs.
5374  * @src - the source block rsv to transfer from.
5375  * @num_bytes - the number of bytes to transfer.
5376  *
5377  * This transfers up to the num_bytes amount from the src rsv to the
5378  * delayed_refs_rsv.  Any extra bytes are returned to the space info.
5379  */
5380 void btrfs_migrate_to_delayed_refs_rsv(struct btrfs_fs_info *fs_info,
5381                                        struct btrfs_block_rsv *src,
5382                                        u64 num_bytes)
5383 {
5384         struct btrfs_block_rsv *delayed_refs_rsv = &fs_info->delayed_refs_rsv;
5385         u64 to_free = 0;
5386
5387         spin_lock(&src->lock);
5388         src->reserved -= num_bytes;
5389         src->size -= num_bytes;
5390         spin_unlock(&src->lock);
5391
5392         spin_lock(&delayed_refs_rsv->lock);
5393         if (delayed_refs_rsv->size > delayed_refs_rsv->reserved) {
5394                 u64 delta = delayed_refs_rsv->size -
5395                         delayed_refs_rsv->reserved;
5396                 if (num_bytes > delta) {
5397                         to_free = num_bytes - delta;
5398                         num_bytes = delta;
5399                 }
5400         } else {
5401                 to_free = num_bytes;
5402                 num_bytes = 0;
5403         }
5404
5405         if (num_bytes)
5406                 delayed_refs_rsv->reserved += num_bytes;
5407         if (delayed_refs_rsv->reserved >= delayed_refs_rsv->size)
5408                 delayed_refs_rsv->full = 1;
5409         spin_unlock(&delayed_refs_rsv->lock);
5410
5411         if (num_bytes)
5412                 trace_btrfs_space_reservation(fs_info, "delayed_refs_rsv",
5413                                               0, num_bytes, 1);
5414         if (to_free)
5415                 space_info_add_old_bytes(fs_info, delayed_refs_rsv->space_info,
5416                                          to_free);
5417 }
5418
5419 /**
5420  * btrfs_delayed_refs_rsv_refill - refill based on our delayed refs usage.
5421  * @fs_info - the fs_info for our fs.
5422  * @flush - control how we can flush for this reservation.
5423  *
5424  * This will refill the delayed block_rsv up to 1 items size worth of space and
5425  * will return -ENOSPC if we can't make the reservation.
5426  */
5427 int btrfs_delayed_refs_rsv_refill(struct btrfs_fs_info *fs_info,
5428                                   enum btrfs_reserve_flush_enum flush)
5429 {
5430         struct btrfs_block_rsv *block_rsv = &fs_info->delayed_refs_rsv;
5431         u64 limit = btrfs_calc_trans_metadata_size(fs_info, 1);
5432         u64 num_bytes = 0;
5433         int ret = -ENOSPC;
5434
5435         spin_lock(&block_rsv->lock);
5436         if (block_rsv->reserved < block_rsv->size) {
5437                 num_bytes = block_rsv->size - block_rsv->reserved;
5438                 num_bytes = min(num_bytes, limit);
5439         }
5440         spin_unlock(&block_rsv->lock);
5441
5442         if (!num_bytes)
5443                 return 0;
5444
5445         ret = reserve_metadata_bytes(fs_info->extent_root, block_rsv,
5446                                      num_bytes, flush);
5447         if (ret)
5448                 return ret;
5449         block_rsv_add_bytes(block_rsv, num_bytes, 0);
5450         trace_btrfs_space_reservation(fs_info, "delayed_refs_rsv",
5451                                       0, num_bytes, 1);
5452         return 0;
5453 }
5454
5455 /*
5456  * This is for space we already have accounted in space_info->bytes_may_use, so
5457  * basically when we're returning space from block_rsv's.
5458  */
5459 static void space_info_add_old_bytes(struct btrfs_fs_info *fs_info,
5460                                      struct btrfs_space_info *space_info,
5461                                      u64 num_bytes)
5462 {
5463         struct reserve_ticket *ticket;
5464         struct list_head *head;
5465         u64 used;
5466         enum btrfs_reserve_flush_enum flush = BTRFS_RESERVE_NO_FLUSH;
5467         bool check_overcommit = false;
5468
5469         spin_lock(&space_info->lock);
5470         head = &space_info->priority_tickets;
5471
5472         /*
5473          * If we are over our limit then we need to check and see if we can
5474          * overcommit, and if we can't then we just need to free up our space
5475          * and not satisfy any requests.
5476          */
5477         used = btrfs_space_info_used(space_info, true);
5478         if (used - num_bytes >= space_info->total_bytes)
5479                 check_overcommit = true;
5480 again:
5481         while (!list_empty(head) && num_bytes) {
5482                 ticket = list_first_entry(head, struct reserve_ticket,
5483                                           list);
5484                 /*
5485                  * We use 0 bytes because this space is already reserved, so
5486                  * adding the ticket space would be a double count.
5487                  */
5488                 if (check_overcommit &&
5489                     !can_overcommit(fs_info, space_info, 0, flush, false))
5490                         break;
5491                 if (num_bytes >= ticket->bytes) {
5492                         list_del_init(&ticket->list);
5493                         num_bytes -= ticket->bytes;
5494                         ticket->bytes = 0;
5495                         space_info->tickets_id++;
5496                         wake_up(&ticket->wait);
5497                 } else {
5498                         ticket->bytes -= num_bytes;
5499                         num_bytes = 0;
5500                 }
5501         }
5502
5503         if (num_bytes && head == &space_info->priority_tickets) {
5504                 head = &space_info->tickets;
5505                 flush = BTRFS_RESERVE_FLUSH_ALL;
5506                 goto again;
5507         }
5508         update_bytes_may_use(fs_info, space_info, -num_bytes);
5509         trace_btrfs_space_reservation(fs_info, "space_info",
5510                                       space_info->flags, num_bytes, 0);
5511         spin_unlock(&space_info->lock);
5512 }
5513
5514 /*
5515  * This is for newly allocated space that isn't accounted in
5516  * space_info->bytes_may_use yet.  So if we allocate a chunk or unpin an extent
5517  * we use this helper.
5518  */
5519 static void space_info_add_new_bytes(struct btrfs_fs_info *fs_info,
5520                                      struct btrfs_space_info *space_info,
5521                                      u64 num_bytes)
5522 {
5523         struct reserve_ticket *ticket;
5524         struct list_head *head = &space_info->priority_tickets;
5525
5526 again:
5527         while (!list_empty(head) && num_bytes) {
5528                 ticket = list_first_entry(head, struct reserve_ticket,
5529                                           list);
5530                 if (num_bytes >= ticket->bytes) {
5531                         trace_btrfs_space_reservation(fs_info, "space_info",
5532                                                       space_info->flags,
5533                                                       ticket->bytes, 1);
5534                         list_del_init(&ticket->list);
5535                         num_bytes -= ticket->bytes;
5536                         update_bytes_may_use(fs_info, space_info,
5537                                              ticket->bytes);
5538                         ticket->bytes = 0;
5539                         space_info->tickets_id++;
5540                         wake_up(&ticket->wait);
5541                 } else {
5542                         trace_btrfs_space_reservation(fs_info, "space_info",
5543                                                       space_info->flags,
5544                                                       num_bytes, 1);
5545                         update_bytes_may_use(fs_info, space_info, num_bytes);
5546                         ticket->bytes -= num_bytes;
5547                         num_bytes = 0;
5548                 }
5549         }
5550
5551         if (num_bytes && head == &space_info->priority_tickets) {
5552                 head = &space_info->tickets;
5553                 goto again;
5554         }
5555 }
5556
5557 static u64 block_rsv_release_bytes(struct btrfs_fs_info *fs_info,
5558                                     struct btrfs_block_rsv *block_rsv,
5559                                     struct btrfs_block_rsv *dest, u64 num_bytes,
5560                                     u64 *qgroup_to_release_ret)
5561 {
5562         struct btrfs_space_info *space_info = block_rsv->space_info;
5563         u64 qgroup_to_release = 0;
5564         u64 ret;
5565
5566         spin_lock(&block_rsv->lock);
5567         if (num_bytes == (u64)-1) {
5568                 num_bytes = block_rsv->size;
5569                 qgroup_to_release = block_rsv->qgroup_rsv_size;
5570         }
5571         block_rsv->size -= num_bytes;
5572         if (block_rsv->reserved >= block_rsv->size) {
5573                 num_bytes = block_rsv->reserved - block_rsv->size;
5574                 block_rsv->reserved = block_rsv->size;
5575                 block_rsv->full = 1;
5576         } else {
5577                 num_bytes = 0;
5578         }
5579         if (block_rsv->qgroup_rsv_reserved >= block_rsv->qgroup_rsv_size) {
5580                 qgroup_to_release = block_rsv->qgroup_rsv_reserved -
5581                                     block_rsv->qgroup_rsv_size;
5582                 block_rsv->qgroup_rsv_reserved = block_rsv->qgroup_rsv_size;
5583         } else {
5584                 qgroup_to_release = 0;
5585         }
5586         spin_unlock(&block_rsv->lock);
5587
5588         ret = num_bytes;
5589         if (num_bytes > 0) {
5590                 if (dest) {
5591                         spin_lock(&dest->lock);
5592                         if (!dest->full) {
5593                                 u64 bytes_to_add;
5594
5595                                 bytes_to_add = dest->size - dest->reserved;
5596                                 bytes_to_add = min(num_bytes, bytes_to_add);
5597                                 dest->reserved += bytes_to_add;
5598                                 if (dest->reserved >= dest->size)
5599                                         dest->full = 1;
5600                                 num_bytes -= bytes_to_add;
5601                         }
5602                         spin_unlock(&dest->lock);
5603                 }
5604                 if (num_bytes)
5605                         space_info_add_old_bytes(fs_info, space_info,
5606                                                  num_bytes);
5607         }
5608         if (qgroup_to_release_ret)
5609                 *qgroup_to_release_ret = qgroup_to_release;
5610         return ret;
5611 }
5612
5613 int btrfs_block_rsv_migrate(struct btrfs_block_rsv *src,
5614                             struct btrfs_block_rsv *dst, u64 num_bytes,
5615                             bool update_size)
5616 {
5617         int ret;
5618
5619         ret = block_rsv_use_bytes(src, num_bytes);
5620         if (ret)
5621                 return ret;
5622
5623         block_rsv_add_bytes(dst, num_bytes, update_size);
5624         return 0;
5625 }
5626
5627 void btrfs_init_block_rsv(struct btrfs_block_rsv *rsv, unsigned short type)
5628 {
5629         memset(rsv, 0, sizeof(*rsv));
5630         spin_lock_init(&rsv->lock);
5631         rsv->type = type;
5632 }
5633
5634 void btrfs_init_metadata_block_rsv(struct btrfs_fs_info *fs_info,
5635                                    struct btrfs_block_rsv *rsv,
5636                                    unsigned short type)
5637 {
5638         btrfs_init_block_rsv(rsv, type);
5639         rsv->space_info = __find_space_info(fs_info,
5640                                             BTRFS_BLOCK_GROUP_METADATA);
5641 }
5642
5643 struct btrfs_block_rsv *btrfs_alloc_block_rsv(struct btrfs_fs_info *fs_info,
5644                                               unsigned short type)
5645 {
5646         struct btrfs_block_rsv *block_rsv;
5647
5648         block_rsv = kmalloc(sizeof(*block_rsv), GFP_NOFS);
5649         if (!block_rsv)
5650                 return NULL;
5651
5652         btrfs_init_metadata_block_rsv(fs_info, block_rsv, type);
5653         return block_rsv;
5654 }
5655
5656 void btrfs_free_block_rsv(struct btrfs_fs_info *fs_info,
5657                           struct btrfs_block_rsv *rsv)
5658 {
5659         if (!rsv)
5660                 return;
5661         btrfs_block_rsv_release(fs_info, rsv, (u64)-1);
5662         kfree(rsv);
5663 }
5664
5665 int btrfs_block_rsv_add(struct btrfs_root *root,
5666                         struct btrfs_block_rsv *block_rsv, u64 num_bytes,
5667                         enum btrfs_reserve_flush_enum flush)
5668 {
5669         int ret;
5670
5671         if (num_bytes == 0)
5672                 return 0;
5673
5674         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
5675         if (!ret)
5676                 block_rsv_add_bytes(block_rsv, num_bytes, true);
5677
5678         return ret;
5679 }
5680
5681 int btrfs_block_rsv_check(struct btrfs_block_rsv *block_rsv, int min_factor)
5682 {
5683         u64 num_bytes = 0;
5684         int ret = -ENOSPC;
5685
5686         if (!block_rsv)
5687                 return 0;
5688
5689         spin_lock(&block_rsv->lock);
5690         num_bytes = div_factor(block_rsv->size, min_factor);
5691         if (block_rsv->reserved >= num_bytes)
5692                 ret = 0;
5693         spin_unlock(&block_rsv->lock);
5694
5695         return ret;
5696 }
5697
5698 int btrfs_block_rsv_refill(struct btrfs_root *root,
5699                            struct btrfs_block_rsv *block_rsv, u64 min_reserved,
5700                            enum btrfs_reserve_flush_enum flush)
5701 {
5702         u64 num_bytes = 0;
5703         int ret = -ENOSPC;
5704
5705         if (!block_rsv)
5706                 return 0;
5707
5708         spin_lock(&block_rsv->lock);
5709         num_bytes = min_reserved;
5710         if (block_rsv->reserved >= num_bytes)
5711                 ret = 0;
5712         else
5713                 num_bytes -= block_rsv->reserved;
5714         spin_unlock(&block_rsv->lock);
5715
5716         if (!ret)
5717                 return 0;
5718
5719         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
5720         if (!ret) {
5721                 block_rsv_add_bytes(block_rsv, num_bytes, false);
5722                 return 0;
5723         }
5724
5725         return ret;
5726 }
5727
5728 static u64 __btrfs_block_rsv_release(struct btrfs_fs_info *fs_info,
5729                                      struct btrfs_block_rsv *block_rsv,
5730                                      u64 num_bytes, u64 *qgroup_to_release)
5731 {
5732         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
5733         struct btrfs_block_rsv *delayed_rsv = &fs_info->delayed_refs_rsv;
5734         struct btrfs_block_rsv *target = delayed_rsv;
5735
5736         if (target->full || target == block_rsv)
5737                 target = global_rsv;
5738
5739         if (block_rsv->space_info != target->space_info)
5740                 target = NULL;
5741
5742         return block_rsv_release_bytes(fs_info, block_rsv, target, num_bytes,
5743                                        qgroup_to_release);
5744 }
5745
5746 void btrfs_block_rsv_release(struct btrfs_fs_info *fs_info,
5747                              struct btrfs_block_rsv *block_rsv,
5748                              u64 num_bytes)
5749 {
5750         __btrfs_block_rsv_release(fs_info, block_rsv, num_bytes, NULL);
5751 }
5752
5753 /**
5754  * btrfs_inode_rsv_release - release any excessive reservation.
5755  * @inode - the inode we need to release from.
5756  * @qgroup_free - free or convert qgroup meta.
5757  *   Unlike normal operation, qgroup meta reservation needs to know if we are
5758  *   freeing qgroup reservation or just converting it into per-trans.  Normally
5759  *   @qgroup_free is true for error handling, and false for normal release.
5760  *
5761  * This is the same as btrfs_block_rsv_release, except that it handles the
5762  * tracepoint for the reservation.
5763  */
5764 static void btrfs_inode_rsv_release(struct btrfs_inode *inode, bool qgroup_free)
5765 {
5766         struct btrfs_fs_info *fs_info = inode->root->fs_info;
5767         struct btrfs_block_rsv *block_rsv = &inode->block_rsv;
5768         u64 released = 0;
5769         u64 qgroup_to_release = 0;
5770
5771         /*
5772          * Since we statically set the block_rsv->size we just want to say we
5773          * are releasing 0 bytes, and then we'll just get the reservation over
5774          * the size free'd.
5775          */
5776         released = __btrfs_block_rsv_release(fs_info, block_rsv, 0,
5777                                              &qgroup_to_release);
5778         if (released > 0)
5779                 trace_btrfs_space_reservation(fs_info, "delalloc",
5780                                               btrfs_ino(inode), released, 0);
5781         if (qgroup_free)
5782                 btrfs_qgroup_free_meta_prealloc(inode->root, qgroup_to_release);
5783         else
5784                 btrfs_qgroup_convert_reserved_meta(inode->root,
5785                                                    qgroup_to_release);
5786 }
5787
5788 /**
5789  * btrfs_delayed_refs_rsv_release - release a ref head's reservation.
5790  * @fs_info - the fs_info for our fs.
5791  * @nr - the number of items to drop.
5792  *
5793  * This drops the delayed ref head's count from the delayed refs rsv and frees
5794  * any excess reservation we had.
5795  */
5796 void btrfs_delayed_refs_rsv_release(struct btrfs_fs_info *fs_info, int nr)
5797 {
5798         struct btrfs_block_rsv *block_rsv = &fs_info->delayed_refs_rsv;
5799         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
5800         u64 num_bytes = btrfs_calc_trans_metadata_size(fs_info, nr);
5801         u64 released = 0;
5802
5803         released = block_rsv_release_bytes(fs_info, block_rsv, global_rsv,
5804                                            num_bytes, NULL);
5805         if (released)
5806                 trace_btrfs_space_reservation(fs_info, "delayed_refs_rsv",
5807                                               0, released, 0);
5808 }
5809
5810 static void update_global_block_rsv(struct btrfs_fs_info *fs_info)
5811 {
5812         struct btrfs_block_rsv *block_rsv = &fs_info->global_block_rsv;
5813         struct btrfs_space_info *sinfo = block_rsv->space_info;
5814         u64 num_bytes;
5815
5816         /*
5817          * The global block rsv is based on the size of the extent tree, the
5818          * checksum tree and the root tree.  If the fs is empty we want to set
5819          * it to a minimal amount for safety.
5820          */
5821         num_bytes = btrfs_root_used(&fs_info->extent_root->root_item) +
5822                 btrfs_root_used(&fs_info->csum_root->root_item) +
5823                 btrfs_root_used(&fs_info->tree_root->root_item);
5824         num_bytes = max_t(u64, num_bytes, SZ_16M);
5825
5826         spin_lock(&sinfo->lock);
5827         spin_lock(&block_rsv->lock);
5828
5829         block_rsv->size = min_t(u64, num_bytes, SZ_512M);
5830
5831         if (block_rsv->reserved < block_rsv->size) {
5832                 num_bytes = btrfs_space_info_used(sinfo, true);
5833                 if (sinfo->total_bytes > num_bytes) {
5834                         num_bytes = sinfo->total_bytes - num_bytes;
5835                         num_bytes = min(num_bytes,
5836                                         block_rsv->size - block_rsv->reserved);
5837                         block_rsv->reserved += num_bytes;
5838                         update_bytes_may_use(fs_info, sinfo, num_bytes);
5839                         trace_btrfs_space_reservation(fs_info, "space_info",
5840                                                       sinfo->flags, num_bytes,
5841                                                       1);
5842                 }
5843         } else if (block_rsv->reserved > block_rsv->size) {
5844                 num_bytes = block_rsv->reserved - block_rsv->size;
5845                 update_bytes_may_use(fs_info, sinfo, -num_bytes);
5846                 trace_btrfs_space_reservation(fs_info, "space_info",
5847                                       sinfo->flags, num_bytes, 0);
5848                 block_rsv->reserved = block_rsv->size;
5849         }
5850
5851         if (block_rsv->reserved == block_rsv->size)
5852                 block_rsv->full = 1;
5853         else
5854                 block_rsv->full = 0;
5855
5856         spin_unlock(&block_rsv->lock);
5857         spin_unlock(&sinfo->lock);
5858 }
5859
5860 static void init_global_block_rsv(struct btrfs_fs_info *fs_info)
5861 {
5862         struct btrfs_space_info *space_info;
5863
5864         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
5865         fs_info->chunk_block_rsv.space_info = space_info;
5866
5867         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
5868         fs_info->global_block_rsv.space_info = space_info;
5869         fs_info->trans_block_rsv.space_info = space_info;
5870         fs_info->empty_block_rsv.space_info = space_info;
5871         fs_info->delayed_block_rsv.space_info = space_info;
5872         fs_info->delayed_refs_rsv.space_info = space_info;
5873
5874         fs_info->extent_root->block_rsv = &fs_info->delayed_refs_rsv;
5875         fs_info->csum_root->block_rsv = &fs_info->delayed_refs_rsv;
5876         fs_info->dev_root->block_rsv = &fs_info->global_block_rsv;
5877         fs_info->tree_root->block_rsv = &fs_info->global_block_rsv;
5878         if (fs_info->quota_root)
5879                 fs_info->quota_root->block_rsv = &fs_info->global_block_rsv;
5880         fs_info->chunk_root->block_rsv = &fs_info->chunk_block_rsv;
5881
5882         update_global_block_rsv(fs_info);
5883 }
5884
5885 static void release_global_block_rsv(struct btrfs_fs_info *fs_info)
5886 {
5887         block_rsv_release_bytes(fs_info, &fs_info->global_block_rsv, NULL,
5888                                 (u64)-1, NULL);
5889         WARN_ON(fs_info->trans_block_rsv.size > 0);
5890         WARN_ON(fs_info->trans_block_rsv.reserved > 0);
5891         WARN_ON(fs_info->chunk_block_rsv.size > 0);
5892         WARN_ON(fs_info->chunk_block_rsv.reserved > 0);
5893         WARN_ON(fs_info->delayed_block_rsv.size > 0);
5894         WARN_ON(fs_info->delayed_block_rsv.reserved > 0);
5895         WARN_ON(fs_info->delayed_refs_rsv.reserved > 0);
5896         WARN_ON(fs_info->delayed_refs_rsv.size > 0);
5897 }
5898
5899 /*
5900  * btrfs_update_delayed_refs_rsv - adjust the size of the delayed refs rsv
5901  * @trans - the trans that may have generated delayed refs
5902  *
5903  * This is to be called anytime we may have adjusted trans->delayed_ref_updates,
5904  * it'll calculate the additional size and add it to the delayed_refs_rsv.
5905  */
5906 void btrfs_update_delayed_refs_rsv(struct btrfs_trans_handle *trans)
5907 {
5908         struct btrfs_fs_info *fs_info = trans->fs_info;
5909         struct btrfs_block_rsv *delayed_rsv = &fs_info->delayed_refs_rsv;
5910         u64 num_bytes;
5911
5912         if (!trans->delayed_ref_updates)
5913                 return;
5914
5915         num_bytes = btrfs_calc_trans_metadata_size(fs_info,
5916                                                    trans->delayed_ref_updates);
5917         spin_lock(&delayed_rsv->lock);
5918         delayed_rsv->size += num_bytes;
5919         delayed_rsv->full = 0;
5920         spin_unlock(&delayed_rsv->lock);
5921         trans->delayed_ref_updates = 0;
5922 }
5923
5924 /*
5925  * To be called after all the new block groups attached to the transaction
5926  * handle have been created (btrfs_create_pending_block_groups()).
5927  */
5928 void btrfs_trans_release_chunk_metadata(struct btrfs_trans_handle *trans)
5929 {
5930         struct btrfs_fs_info *fs_info = trans->fs_info;
5931
5932         if (!trans->chunk_bytes_reserved)
5933                 return;
5934
5935         WARN_ON_ONCE(!list_empty(&trans->new_bgs));
5936
5937         block_rsv_release_bytes(fs_info, &fs_info->chunk_block_rsv, NULL,
5938                                 trans->chunk_bytes_reserved, NULL);
5939         trans->chunk_bytes_reserved = 0;
5940 }
5941
5942 /*
5943  * btrfs_subvolume_reserve_metadata() - reserve space for subvolume operation
5944  * root: the root of the parent directory
5945  * rsv: block reservation
5946  * items: the number of items that we need do reservation
5947  * use_global_rsv: allow fallback to the global block reservation
5948  *
5949  * This function is used to reserve the space for snapshot/subvolume
5950  * creation and deletion. Those operations are different with the
5951  * common file/directory operations, they change two fs/file trees
5952  * and root tree, the number of items that the qgroup reserves is
5953  * different with the free space reservation. So we can not use
5954  * the space reservation mechanism in start_transaction().
5955  */
5956 int btrfs_subvolume_reserve_metadata(struct btrfs_root *root,
5957                                      struct btrfs_block_rsv *rsv, int items,
5958                                      bool use_global_rsv)
5959 {
5960         u64 qgroup_num_bytes = 0;
5961         u64 num_bytes;
5962         int ret;
5963         struct btrfs_fs_info *fs_info = root->fs_info;
5964         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
5965
5966         if (test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags)) {
5967                 /* One for parent inode, two for dir entries */
5968                 qgroup_num_bytes = 3 * fs_info->nodesize;
5969                 ret = btrfs_qgroup_reserve_meta_prealloc(root,
5970                                 qgroup_num_bytes, true);
5971                 if (ret)
5972                         return ret;
5973         }
5974
5975         num_bytes = btrfs_calc_trans_metadata_size(fs_info, items);
5976         rsv->space_info = __find_space_info(fs_info,
5977                                             BTRFS_BLOCK_GROUP_METADATA);
5978         ret = btrfs_block_rsv_add(root, rsv, num_bytes,
5979                                   BTRFS_RESERVE_FLUSH_ALL);
5980
5981         if (ret == -ENOSPC && use_global_rsv)
5982                 ret = btrfs_block_rsv_migrate(global_rsv, rsv, num_bytes, true);
5983
5984         if (ret && qgroup_num_bytes)
5985                 btrfs_qgroup_free_meta_prealloc(root, qgroup_num_bytes);
5986
5987         return ret;
5988 }
5989
5990 void btrfs_subvolume_release_metadata(struct btrfs_fs_info *fs_info,
5991                                       struct btrfs_block_rsv *rsv)
5992 {
5993         btrfs_block_rsv_release(fs_info, rsv, (u64)-1);
5994 }
5995
5996 static void btrfs_calculate_inode_block_rsv_size(struct btrfs_fs_info *fs_info,
5997                                                  struct btrfs_inode *inode)
5998 {
5999         struct btrfs_block_rsv *block_rsv = &inode->block_rsv;
6000         u64 reserve_size = 0;
6001         u64 qgroup_rsv_size = 0;
6002         u64 csum_leaves;
6003         unsigned outstanding_extents;
6004
6005         lockdep_assert_held(&inode->lock);
6006         outstanding_extents = inode->outstanding_extents;
6007         if (outstanding_extents)
6008                 reserve_size = btrfs_calc_trans_metadata_size(fs_info,
6009                                                 outstanding_extents + 1);
6010         csum_leaves = btrfs_csum_bytes_to_leaves(fs_info,
6011                                                  inode->csum_bytes);
6012         reserve_size += btrfs_calc_trans_metadata_size(fs_info,
6013                                                        csum_leaves);
6014         /*
6015          * For qgroup rsv, the calculation is very simple:
6016          * account one nodesize for each outstanding extent
6017          *
6018          * This is overestimating in most cases.
6019          */
6020         qgroup_rsv_size = (u64)outstanding_extents * fs_info->nodesize;
6021
6022         spin_lock(&block_rsv->lock);
6023         block_rsv->size = reserve_size;
6024         block_rsv->qgroup_rsv_size = qgroup_rsv_size;
6025         spin_unlock(&block_rsv->lock);
6026 }
6027
6028 static void calc_inode_reservations(struct btrfs_fs_info *fs_info,
6029                                     u64 num_bytes, u64 *meta_reserve,
6030                                     u64 *qgroup_reserve)
6031 {
6032         u64 nr_extents = count_max_extents(num_bytes);
6033         u64 csum_leaves = btrfs_csum_bytes_to_leaves(fs_info, num_bytes);
6034
6035         /* We add one for the inode update at finish ordered time */
6036         *meta_reserve = btrfs_calc_trans_metadata_size(fs_info,
6037                                                 nr_extents + csum_leaves + 1);
6038         *qgroup_reserve = nr_extents * fs_info->nodesize;
6039 }
6040
6041 int btrfs_delalloc_reserve_metadata(struct btrfs_inode *inode, u64 num_bytes)
6042 {
6043         struct btrfs_root *root = inode->root;
6044         struct btrfs_fs_info *fs_info = root->fs_info;
6045         struct btrfs_block_rsv *block_rsv = &inode->block_rsv;
6046         u64 meta_reserve, qgroup_reserve;
6047         unsigned nr_extents;
6048         enum btrfs_reserve_flush_enum flush = BTRFS_RESERVE_FLUSH_ALL;
6049         int ret = 0;
6050         bool delalloc_lock = true;
6051
6052         /* If we are a free space inode we need to not flush since we will be in
6053          * the middle of a transaction commit.  We also don't need the delalloc
6054          * mutex since we won't race with anybody.  We need this mostly to make
6055          * lockdep shut its filthy mouth.
6056          *
6057          * If we have a transaction open (can happen if we call truncate_block
6058          * from truncate), then we need FLUSH_LIMIT so we don't deadlock.
6059          */
6060         if (btrfs_is_free_space_inode(inode)) {
6061                 flush = BTRFS_RESERVE_NO_FLUSH;
6062                 delalloc_lock = false;
6063         } else {
6064                 if (current->journal_info)
6065                         flush = BTRFS_RESERVE_FLUSH_LIMIT;
6066
6067                 if (btrfs_transaction_in_commit(fs_info))
6068                         schedule_timeout(1);
6069         }
6070
6071         if (delalloc_lock)
6072                 mutex_lock(&inode->delalloc_mutex);
6073
6074         num_bytes = ALIGN(num_bytes, fs_info->sectorsize);
6075
6076         /*
6077          * We always want to do it this way, every other way is wrong and ends
6078          * in tears.  Pre-reserving the amount we are going to add will always
6079          * be the right way, because otherwise if we have enough parallelism we
6080          * could end up with thousands of inodes all holding little bits of
6081          * reservations they were able to make previously and the only way to
6082          * reclaim that space is to ENOSPC out the operations and clear
6083          * everything out and try again, which is bad.  This way we just
6084          * over-reserve slightly, and clean up the mess when we are done.
6085          */
6086         calc_inode_reservations(fs_info, num_bytes, &meta_reserve,
6087                                 &qgroup_reserve);
6088         ret = btrfs_qgroup_reserve_meta_prealloc(root, qgroup_reserve, true);
6089         if (ret)
6090                 goto out_fail;
6091         ret = reserve_metadata_bytes(root, block_rsv, meta_reserve, flush);
6092         if (ret)
6093                 goto out_qgroup;
6094
6095         /*
6096          * Now we need to update our outstanding extents and csum bytes _first_
6097          * and then add the reservation to the block_rsv.  This keeps us from
6098          * racing with an ordered completion or some such that would think it
6099          * needs to free the reservation we just made.
6100          */
6101         spin_lock(&inode->lock);
6102         nr_extents = count_max_extents(num_bytes);
6103         btrfs_mod_outstanding_extents(inode, nr_extents);
6104         inode->csum_bytes += num_bytes;
6105         btrfs_calculate_inode_block_rsv_size(fs_info, inode);
6106         spin_unlock(&inode->lock);
6107
6108         /* Now we can safely add our space to our block rsv */
6109         block_rsv_add_bytes(block_rsv, meta_reserve, false);
6110         trace_btrfs_space_reservation(root->fs_info, "delalloc",
6111                                       btrfs_ino(inode), meta_reserve, 1);
6112
6113         spin_lock(&block_rsv->lock);
6114         block_rsv->qgroup_rsv_reserved += qgroup_reserve;
6115         spin_unlock(&block_rsv->lock);
6116
6117         if (delalloc_lock)
6118                 mutex_unlock(&inode->delalloc_mutex);
6119         return 0;
6120 out_qgroup:
6121         btrfs_qgroup_free_meta_prealloc(root, qgroup_reserve);
6122 out_fail:
6123         btrfs_inode_rsv_release(inode, true);
6124         if (delalloc_lock)
6125                 mutex_unlock(&inode->delalloc_mutex);
6126         return ret;
6127 }
6128
6129 /**
6130  * btrfs_delalloc_release_metadata - release a metadata reservation for an inode
6131  * @inode: the inode to release the reservation for.
6132  * @num_bytes: the number of bytes we are releasing.
6133  * @qgroup_free: free qgroup reservation or convert it to per-trans reservation
6134  *
6135  * This will release the metadata reservation for an inode.  This can be called
6136  * once we complete IO for a given set of bytes to release their metadata
6137  * reservations, or on error for the same reason.
6138  */
6139 void btrfs_delalloc_release_metadata(struct btrfs_inode *inode, u64 num_bytes,
6140                                      bool qgroup_free)
6141 {
6142         struct btrfs_fs_info *fs_info = inode->root->fs_info;
6143
6144         num_bytes = ALIGN(num_bytes, fs_info->sectorsize);
6145         spin_lock(&inode->lock);
6146         inode->csum_bytes -= num_bytes;
6147         btrfs_calculate_inode_block_rsv_size(fs_info, inode);
6148         spin_unlock(&inode->lock);
6149
6150         if (btrfs_is_testing(fs_info))
6151                 return;
6152
6153         btrfs_inode_rsv_release(inode, qgroup_free);
6154 }
6155
6156 /**
6157  * btrfs_delalloc_release_extents - release our outstanding_extents
6158  * @inode: the inode to balance the reservation for.
6159  * @num_bytes: the number of bytes we originally reserved with
6160  * @qgroup_free: do we need to free qgroup meta reservation or convert them.
6161  *
6162  * When we reserve space we increase outstanding_extents for the extents we may
6163  * add.  Once we've set the range as delalloc or created our ordered extents we
6164  * have outstanding_extents to track the real usage, so we use this to free our
6165  * temporarily tracked outstanding_extents.  This _must_ be used in conjunction
6166  * with btrfs_delalloc_reserve_metadata.
6167  */
6168 void btrfs_delalloc_release_extents(struct btrfs_inode *inode, u64 num_bytes,
6169                                     bool qgroup_free)
6170 {
6171         struct btrfs_fs_info *fs_info = inode->root->fs_info;
6172         unsigned num_extents;
6173
6174         spin_lock(&inode->lock);
6175         num_extents = count_max_extents(num_bytes);
6176         btrfs_mod_outstanding_extents(inode, -num_extents);
6177         btrfs_calculate_inode_block_rsv_size(fs_info, inode);
6178         spin_unlock(&inode->lock);
6179
6180         if (btrfs_is_testing(fs_info))
6181                 return;
6182
6183         btrfs_inode_rsv_release(inode, qgroup_free);
6184 }
6185
6186 /**
6187  * btrfs_delalloc_reserve_space - reserve data and metadata space for
6188  * delalloc
6189  * @inode: inode we're writing to
6190  * @start: start range we are writing to
6191  * @len: how long the range we are writing to
6192  * @reserved: mandatory parameter, record actually reserved qgroup ranges of
6193  *            current reservation.
6194  *
6195  * This will do the following things
6196  *
6197  * o reserve space in data space info for num bytes
6198  *   and reserve precious corresponding qgroup space
6199  *   (Done in check_data_free_space)
6200  *
6201  * o reserve space for metadata space, based on the number of outstanding
6202  *   extents and how much csums will be needed
6203  *   also reserve metadata space in a per root over-reserve method.
6204  * o add to the inodes->delalloc_bytes
6205  * o add it to the fs_info's delalloc inodes list.
6206  *   (Above 3 all done in delalloc_reserve_metadata)
6207  *
6208  * Return 0 for success
6209  * Return <0 for error(-ENOSPC or -EQUOT)
6210  */
6211 int btrfs_delalloc_reserve_space(struct inode *inode,
6212                         struct extent_changeset **reserved, u64 start, u64 len)
6213 {
6214         int ret;
6215
6216         ret = btrfs_check_data_free_space(inode, reserved, start, len);
6217         if (ret < 0)
6218                 return ret;
6219         ret = btrfs_delalloc_reserve_metadata(BTRFS_I(inode), len);
6220         if (ret < 0)
6221                 btrfs_free_reserved_data_space(inode, *reserved, start, len);
6222         return ret;
6223 }
6224
6225 /**
6226  * btrfs_delalloc_release_space - release data and metadata space for delalloc
6227  * @inode: inode we're releasing space for
6228  * @start: start position of the space already reserved
6229  * @len: the len of the space already reserved
6230  * @release_bytes: the len of the space we consumed or didn't use
6231  *
6232  * This function will release the metadata space that was not used and will
6233  * decrement ->delalloc_bytes and remove it from the fs_info delalloc_inodes
6234  * list if there are no delalloc bytes left.
6235  * Also it will handle the qgroup reserved space.
6236  */
6237 void btrfs_delalloc_release_space(struct inode *inode,
6238                                   struct extent_changeset *reserved,
6239                                   u64 start, u64 len, bool qgroup_free)
6240 {
6241         btrfs_delalloc_release_metadata(BTRFS_I(inode), len, qgroup_free);
6242         btrfs_free_reserved_data_space(inode, reserved, start, len);
6243 }
6244
6245 static int update_block_group(struct btrfs_trans_handle *trans,
6246                               u64 bytenr, u64 num_bytes, int alloc)
6247 {
6248         struct btrfs_fs_info *info = trans->fs_info;
6249         struct btrfs_block_group_cache *cache = NULL;
6250         u64 total = num_bytes;
6251         u64 old_val;
6252         u64 byte_in_group;
6253         int factor;
6254         int ret = 0;
6255
6256         /* block accounting for super block */
6257         spin_lock(&info->delalloc_root_lock);
6258         old_val = btrfs_super_bytes_used(info->super_copy);
6259         if (alloc)
6260                 old_val += num_bytes;
6261         else
6262                 old_val -= num_bytes;
6263         btrfs_set_super_bytes_used(info->super_copy, old_val);
6264         spin_unlock(&info->delalloc_root_lock);
6265
6266         while (total) {
6267                 cache = btrfs_lookup_block_group(info, bytenr);
6268                 if (!cache) {
6269                         ret = -ENOENT;
6270                         break;
6271                 }
6272                 factor = btrfs_bg_type_to_factor(cache->flags);
6273
6274                 /*
6275                  * If this block group has free space cache written out, we
6276                  * need to make sure to load it if we are removing space.  This
6277                  * is because we need the unpinning stage to actually add the
6278                  * space back to the block group, otherwise we will leak space.
6279                  */
6280                 if (!alloc && cache->cached == BTRFS_CACHE_NO)
6281                         cache_block_group(cache, 1);
6282
6283                 byte_in_group = bytenr - cache->key.objectid;
6284                 WARN_ON(byte_in_group > cache->key.offset);
6285
6286                 spin_lock(&cache->space_info->lock);
6287                 spin_lock(&cache->lock);
6288
6289                 if (btrfs_test_opt(info, SPACE_CACHE) &&
6290                     cache->disk_cache_state < BTRFS_DC_CLEAR)
6291                         cache->disk_cache_state = BTRFS_DC_CLEAR;
6292
6293                 old_val = btrfs_block_group_used(&cache->item);
6294                 num_bytes = min(total, cache->key.offset - byte_in_group);
6295                 if (alloc) {
6296                         old_val += num_bytes;
6297                         btrfs_set_block_group_used(&cache->item, old_val);
6298                         cache->reserved -= num_bytes;
6299                         cache->space_info->bytes_reserved -= num_bytes;
6300                         cache->space_info->bytes_used += num_bytes;
6301                         cache->space_info->disk_used += num_bytes * factor;
6302                         spin_unlock(&cache->lock);
6303                         spin_unlock(&cache->space_info->lock);
6304                 } else {
6305                         old_val -= num_bytes;
6306                         btrfs_set_block_group_used(&cache->item, old_val);
6307                         cache->pinned += num_bytes;
6308                         update_bytes_pinned(info, cache->space_info, num_bytes);
6309                         cache->space_info->bytes_used -= num_bytes;
6310                         cache->space_info->disk_used -= num_bytes * factor;
6311                         spin_unlock(&cache->lock);
6312                         spin_unlock(&cache->space_info->lock);
6313
6314                         trace_btrfs_space_reservation(info, "pinned",
6315                                                       cache->space_info->flags,
6316                                                       num_bytes, 1);
6317                         percpu_counter_add_batch(&cache->space_info->total_bytes_pinned,
6318                                            num_bytes,
6319                                            BTRFS_TOTAL_BYTES_PINNED_BATCH);
6320                         set_extent_dirty(info->pinned_extents,
6321                                          bytenr, bytenr + num_bytes - 1,
6322                                          GFP_NOFS | __GFP_NOFAIL);
6323                 }
6324
6325                 spin_lock(&trans->transaction->dirty_bgs_lock);
6326                 if (list_empty(&cache->dirty_list)) {
6327                         list_add_tail(&cache->dirty_list,
6328                                       &trans->transaction->dirty_bgs);
6329                         trans->delayed_ref_updates++;
6330                         btrfs_get_block_group(cache);
6331                 }
6332                 spin_unlock(&trans->transaction->dirty_bgs_lock);
6333
6334                 /*
6335                  * No longer have used bytes in this block group, queue it for
6336                  * deletion. We do this after adding the block group to the
6337                  * dirty list to avoid races between cleaner kthread and space
6338                  * cache writeout.
6339                  */
6340                 if (!alloc && old_val == 0)
6341                         btrfs_mark_bg_unused(cache);
6342
6343                 btrfs_put_block_group(cache);
6344                 total -= num_bytes;
6345                 bytenr += num_bytes;
6346         }
6347
6348         /* Modified block groups are accounted for in the delayed_refs_rsv. */
6349         btrfs_update_delayed_refs_rsv(trans);
6350         return ret;
6351 }
6352
6353 static u64 first_logical_byte(struct btrfs_fs_info *fs_info, u64 search_start)
6354 {
6355         struct btrfs_block_group_cache *cache;
6356         u64 bytenr;
6357
6358         spin_lock(&fs_info->block_group_cache_lock);
6359         bytenr = fs_info->first_logical_byte;
6360         spin_unlock(&fs_info->block_group_cache_lock);
6361
6362         if (bytenr < (u64)-1)
6363                 return bytenr;
6364
6365         cache = btrfs_lookup_first_block_group(fs_info, search_start);
6366         if (!cache)
6367                 return 0;
6368
6369         bytenr = cache->key.objectid;
6370         btrfs_put_block_group(cache);
6371
6372         return bytenr;
6373 }
6374
6375 static int pin_down_extent(struct btrfs_block_group_cache *cache,
6376                            u64 bytenr, u64 num_bytes, int reserved)
6377 {
6378         struct btrfs_fs_info *fs_info = cache->fs_info;
6379
6380         spin_lock(&cache->space_info->lock);
6381         spin_lock(&cache->lock);
6382         cache->pinned += num_bytes;
6383         update_bytes_pinned(fs_info, cache->space_info, num_bytes);
6384         if (reserved) {
6385                 cache->reserved -= num_bytes;
6386                 cache->space_info->bytes_reserved -= num_bytes;
6387         }
6388         spin_unlock(&cache->lock);
6389         spin_unlock(&cache->space_info->lock);
6390
6391         trace_btrfs_space_reservation(fs_info, "pinned",
6392                                       cache->space_info->flags, num_bytes, 1);
6393         percpu_counter_add_batch(&cache->space_info->total_bytes_pinned,
6394                     num_bytes, BTRFS_TOTAL_BYTES_PINNED_BATCH);
6395         set_extent_dirty(fs_info->pinned_extents, bytenr,
6396                          bytenr + num_bytes - 1, GFP_NOFS | __GFP_NOFAIL);
6397         return 0;
6398 }
6399
6400 /*
6401  * this function must be called within transaction
6402  */
6403 int btrfs_pin_extent(struct btrfs_fs_info *fs_info,
6404                      u64 bytenr, u64 num_bytes, int reserved)
6405 {
6406         struct btrfs_block_group_cache *cache;
6407
6408         cache = btrfs_lookup_block_group(fs_info, bytenr);
6409         BUG_ON(!cache); /* Logic error */
6410
6411         pin_down_extent(cache, bytenr, num_bytes, reserved);
6412
6413         btrfs_put_block_group(cache);
6414         return 0;
6415 }
6416
6417 /*
6418  * this function must be called within transaction
6419  */
6420 int btrfs_pin_extent_for_log_replay(struct btrfs_fs_info *fs_info,
6421                                     u64 bytenr, u64 num_bytes)
6422 {
6423         struct btrfs_block_group_cache *cache;
6424         int ret;
6425
6426         cache = btrfs_lookup_block_group(fs_info, bytenr);
6427         if (!cache)
6428                 return -EINVAL;
6429
6430         /*
6431          * pull in the free space cache (if any) so that our pin
6432          * removes the free space from the cache.  We have load_only set
6433          * to one because the slow code to read in the free extents does check
6434          * the pinned extents.
6435          */
6436         cache_block_group(cache, 1);
6437
6438         pin_down_extent(cache, bytenr, num_bytes, 0);
6439
6440         /* remove us from the free space cache (if we're there at all) */
6441         ret = btrfs_remove_free_space(cache, bytenr, num_bytes);
6442         btrfs_put_block_group(cache);
6443         return ret;
6444 }
6445
6446 static int __exclude_logged_extent(struct btrfs_fs_info *fs_info,
6447                                    u64 start, u64 num_bytes)
6448 {
6449         int ret;
6450         struct btrfs_block_group_cache *block_group;
6451         struct btrfs_caching_control *caching_ctl;
6452
6453         block_group = btrfs_lookup_block_group(fs_info, start);
6454         if (!block_group)
6455                 return -EINVAL;
6456
6457         cache_block_group(block_group, 0);
6458         caching_ctl = get_caching_control(block_group);
6459
6460         if (!caching_ctl) {
6461                 /* Logic error */
6462                 BUG_ON(!block_group_cache_done(block_group));
6463                 ret = btrfs_remove_free_space(block_group, start, num_bytes);
6464         } else {
6465                 mutex_lock(&caching_ctl->mutex);
6466
6467                 if (start >= caching_ctl->progress) {
6468                         ret = add_excluded_extent(fs_info, start, num_bytes);
6469                 } else if (start + num_bytes <= caching_ctl->progress) {
6470                         ret = btrfs_remove_free_space(block_group,
6471                                                       start, num_bytes);
6472                 } else {
6473                         num_bytes = caching_ctl->progress - start;
6474                         ret = btrfs_remove_free_space(block_group,
6475                                                       start, num_bytes);
6476                         if (ret)
6477                                 goto out_lock;
6478
6479                         num_bytes = (start + num_bytes) -
6480                                 caching_ctl->progress;
6481                         start = caching_ctl->progress;
6482                         ret = add_excluded_extent(fs_info, start, num_bytes);
6483                 }
6484 out_lock:
6485                 mutex_unlock(&caching_ctl->mutex);
6486                 put_caching_control(caching_ctl);
6487         }
6488         btrfs_put_block_group(block_group);
6489         return ret;
6490 }
6491
6492 int btrfs_exclude_logged_extents(struct extent_buffer *eb)
6493 {
6494         struct btrfs_fs_info *fs_info = eb->fs_info;
6495         struct btrfs_file_extent_item *item;
6496         struct btrfs_key key;
6497         int found_type;
6498         int i;
6499         int ret = 0;
6500
6501         if (!btrfs_fs_incompat(fs_info, MIXED_GROUPS))
6502                 return 0;
6503
6504         for (i = 0; i < btrfs_header_nritems(eb); i++) {
6505                 btrfs_item_key_to_cpu(eb, &key, i);
6506                 if (key.type != BTRFS_EXTENT_DATA_KEY)
6507                         continue;
6508                 item = btrfs_item_ptr(eb, i, struct btrfs_file_extent_item);
6509                 found_type = btrfs_file_extent_type(eb, item);
6510                 if (found_type == BTRFS_FILE_EXTENT_INLINE)
6511                         continue;
6512                 if (btrfs_file_extent_disk_bytenr(eb, item) == 0)
6513                         continue;
6514                 key.objectid = btrfs_file_extent_disk_bytenr(eb, item);
6515                 key.offset = btrfs_file_extent_disk_num_bytes(eb, item);
6516                 ret = __exclude_logged_extent(fs_info, key.objectid, key.offset);
6517                 if (ret)
6518                         break;
6519         }
6520
6521         return ret;
6522 }
6523
6524 static void
6525 btrfs_inc_block_group_reservations(struct btrfs_block_group_cache *bg)
6526 {
6527         atomic_inc(&bg->reservations);
6528 }
6529
6530 void btrfs_dec_block_group_reservations(struct btrfs_fs_info *fs_info,
6531                                         const u64 start)
6532 {
6533         struct btrfs_block_group_cache *bg;
6534
6535         bg = btrfs_lookup_block_group(fs_info, start);
6536         ASSERT(bg);
6537         if (atomic_dec_and_test(&bg->reservations))
6538                 wake_up_var(&bg->reservations);
6539         btrfs_put_block_group(bg);
6540 }
6541
6542 void btrfs_wait_block_group_reservations(struct btrfs_block_group_cache *bg)
6543 {
6544         struct btrfs_space_info *space_info = bg->space_info;
6545
6546         ASSERT(bg->ro);
6547
6548         if (!(bg->flags & BTRFS_BLOCK_GROUP_DATA))
6549                 return;
6550
6551         /*
6552          * Our block group is read only but before we set it to read only,
6553          * some task might have had allocated an extent from it already, but it
6554          * has not yet created a respective ordered extent (and added it to a
6555          * root's list of ordered extents).
6556          * Therefore wait for any task currently allocating extents, since the
6557          * block group's reservations counter is incremented while a read lock
6558          * on the groups' semaphore is held and decremented after releasing
6559          * the read access on that semaphore and creating the ordered extent.
6560          */
6561         down_write(&space_info->groups_sem);
6562         up_write(&space_info->groups_sem);
6563
6564         wait_var_event(&bg->reservations, !atomic_read(&bg->reservations));
6565 }
6566
6567 /**
6568  * btrfs_add_reserved_bytes - update the block_group and space info counters
6569  * @cache:      The cache we are manipulating
6570  * @ram_bytes:  The number of bytes of file content, and will be same to
6571  *              @num_bytes except for the compress path.
6572  * @num_bytes:  The number of bytes in question
6573  * @delalloc:   The blocks are allocated for the delalloc write
6574  *
6575  * This is called by the allocator when it reserves space. If this is a
6576  * reservation and the block group has become read only we cannot make the
6577  * reservation and return -EAGAIN, otherwise this function always succeeds.
6578  */
6579 static int btrfs_add_reserved_bytes(struct btrfs_block_group_cache *cache,
6580                                     u64 ram_bytes, u64 num_bytes, int delalloc)
6581 {
6582         struct btrfs_space_info *space_info = cache->space_info;
6583         int ret = 0;
6584
6585         spin_lock(&space_info->lock);
6586         spin_lock(&cache->lock);
6587         if (cache->ro) {
6588                 ret = -EAGAIN;
6589         } else {
6590                 cache->reserved += num_bytes;
6591                 space_info->bytes_reserved += num_bytes;
6592                 update_bytes_may_use(cache->fs_info, space_info, -ram_bytes);
6593                 if (delalloc)
6594                         cache->delalloc_bytes += num_bytes;
6595         }
6596         spin_unlock(&cache->lock);
6597         spin_unlock(&space_info->lock);
6598         return ret;
6599 }
6600
6601 /**
6602  * btrfs_free_reserved_bytes - update the block_group and space info counters
6603  * @cache:      The cache we are manipulating
6604  * @num_bytes:  The number of bytes in question
6605  * @delalloc:   The blocks are allocated for the delalloc write
6606  *
6607  * This is called by somebody who is freeing space that was never actually used
6608  * on disk.  For example if you reserve some space for a new leaf in transaction
6609  * A and before transaction A commits you free that leaf, you call this with
6610  * reserve set to 0 in order to clear the reservation.
6611  */
6612
6613 static void btrfs_free_reserved_bytes(struct btrfs_block_group_cache *cache,
6614                                       u64 num_bytes, int delalloc)
6615 {
6616         struct btrfs_space_info *space_info = cache->space_info;
6617
6618         spin_lock(&space_info->lock);
6619         spin_lock(&cache->lock);
6620         if (cache->ro)
6621                 space_info->bytes_readonly += num_bytes;
6622         cache->reserved -= num_bytes;
6623         space_info->bytes_reserved -= num_bytes;
6624         space_info->max_extent_size = 0;
6625
6626         if (delalloc)
6627                 cache->delalloc_bytes -= num_bytes;
6628         spin_unlock(&cache->lock);
6629         spin_unlock(&space_info->lock);
6630 }
6631 void btrfs_prepare_extent_commit(struct btrfs_fs_info *fs_info)
6632 {
6633         struct btrfs_caching_control *next;
6634         struct btrfs_caching_control *caching_ctl;
6635         struct btrfs_block_group_cache *cache;
6636
6637         down_write(&fs_info->commit_root_sem);
6638
6639         list_for_each_entry_safe(caching_ctl, next,
6640                                  &fs_info->caching_block_groups, list) {
6641                 cache = caching_ctl->block_group;
6642                 if (block_group_cache_done(cache)) {
6643                         cache->last_byte_to_unpin = (u64)-1;
6644                         list_del_init(&caching_ctl->list);
6645                         put_caching_control(caching_ctl);
6646                 } else {
6647                         cache->last_byte_to_unpin = caching_ctl->progress;
6648                 }
6649         }
6650
6651         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
6652                 fs_info->pinned_extents = &fs_info->freed_extents[1];
6653         else
6654                 fs_info->pinned_extents = &fs_info->freed_extents[0];
6655
6656         up_write(&fs_info->commit_root_sem);
6657
6658         update_global_block_rsv(fs_info);
6659 }
6660
6661 /*
6662  * Returns the free cluster for the given space info and sets empty_cluster to
6663  * what it should be based on the mount options.
6664  */
6665 static struct btrfs_free_cluster *
6666 fetch_cluster_info(struct btrfs_fs_info *fs_info,
6667                    struct btrfs_space_info *space_info, u64 *empty_cluster)
6668 {
6669         struct btrfs_free_cluster *ret = NULL;
6670
6671         *empty_cluster = 0;
6672         if (btrfs_mixed_space_info(space_info))
6673                 return ret;
6674
6675         if (space_info->flags & BTRFS_BLOCK_GROUP_METADATA) {
6676                 ret = &fs_info->meta_alloc_cluster;
6677                 if (btrfs_test_opt(fs_info, SSD))
6678                         *empty_cluster = SZ_2M;
6679                 else
6680                         *empty_cluster = SZ_64K;
6681         } else if ((space_info->flags & BTRFS_BLOCK_GROUP_DATA) &&
6682                    btrfs_test_opt(fs_info, SSD_SPREAD)) {
6683                 *empty_cluster = SZ_2M;
6684                 ret = &fs_info->data_alloc_cluster;
6685         }
6686
6687         return ret;
6688 }
6689
6690 static int unpin_extent_range(struct btrfs_fs_info *fs_info,
6691                               u64 start, u64 end,
6692                               const bool return_free_space)
6693 {
6694         struct btrfs_block_group_cache *cache = NULL;
6695         struct btrfs_space_info *space_info;
6696         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
6697         struct btrfs_free_cluster *cluster = NULL;
6698         u64 len;
6699         u64 total_unpinned = 0;
6700         u64 empty_cluster = 0;
6701         bool readonly;
6702
6703         while (start <= end) {
6704                 readonly = false;
6705                 if (!cache ||
6706                     start >= cache->key.objectid + cache->key.offset) {
6707                         if (cache)
6708                                 btrfs_put_block_group(cache);
6709                         total_unpinned = 0;
6710                         cache = btrfs_lookup_block_group(fs_info, start);
6711                         BUG_ON(!cache); /* Logic error */
6712
6713                         cluster = fetch_cluster_info(fs_info,
6714                                                      cache->space_info,
6715                                                      &empty_cluster);
6716                         empty_cluster <<= 1;
6717                 }
6718
6719                 len = cache->key.objectid + cache->key.offset - start;
6720                 len = min(len, end + 1 - start);
6721
6722                 if (start < cache->last_byte_to_unpin) {
6723                         len = min(len, cache->last_byte_to_unpin - start);
6724                         if (return_free_space)
6725                                 btrfs_add_free_space(cache, start, len);
6726                 }
6727
6728                 start += len;
6729                 total_unpinned += len;
6730                 space_info = cache->space_info;
6731
6732                 /*
6733                  * If this space cluster has been marked as fragmented and we've
6734                  * unpinned enough in this block group to potentially allow a
6735                  * cluster to be created inside of it go ahead and clear the
6736                  * fragmented check.
6737                  */
6738                 if (cluster && cluster->fragmented &&
6739                     total_unpinned > empty_cluster) {
6740                         spin_lock(&cluster->lock);
6741                         cluster->fragmented = 0;
6742                         spin_unlock(&cluster->lock);
6743                 }
6744
6745                 spin_lock(&space_info->lock);
6746                 spin_lock(&cache->lock);
6747                 cache->pinned -= len;
6748                 update_bytes_pinned(fs_info, space_info, -len);
6749
6750                 trace_btrfs_space_reservation(fs_info, "pinned",
6751                                               space_info->flags, len, 0);
6752                 space_info->max_extent_size = 0;
6753                 percpu_counter_add_batch(&space_info->total_bytes_pinned,
6754                             -len, BTRFS_TOTAL_BYTES_PINNED_BATCH);
6755                 if (cache->ro) {
6756                         space_info->bytes_readonly += len;
6757                         readonly = true;
6758                 }
6759                 spin_unlock(&cache->lock);
6760                 if (!readonly && return_free_space &&
6761                     global_rsv->space_info == space_info) {
6762                         u64 to_add = len;
6763
6764                         spin_lock(&global_rsv->lock);
6765                         if (!global_rsv->full) {
6766                                 to_add = min(len, global_rsv->size -
6767                                              global_rsv->reserved);
6768                                 global_rsv->reserved += to_add;
6769                                 update_bytes_may_use(fs_info, space_info,
6770                                                      to_add);
6771                                 if (global_rsv->reserved >= global_rsv->size)
6772                                         global_rsv->full = 1;
6773                                 trace_btrfs_space_reservation(fs_info,
6774                                                               "space_info",
6775                                                               space_info->flags,
6776                                                               to_add, 1);
6777                                 len -= to_add;
6778                         }
6779                         spin_unlock(&global_rsv->lock);
6780                         /* Add to any tickets we may have */
6781                         if (len)
6782                                 space_info_add_new_bytes(fs_info, space_info,
6783                                                          len);
6784                 }
6785                 spin_unlock(&space_info->lock);
6786         }
6787
6788         if (cache)
6789                 btrfs_put_block_group(cache);
6790         return 0;
6791 }
6792
6793 int btrfs_finish_extent_commit(struct btrfs_trans_handle *trans)
6794 {
6795         struct btrfs_fs_info *fs_info = trans->fs_info;
6796         struct btrfs_block_group_cache *block_group, *tmp;
6797         struct list_head *deleted_bgs;
6798         struct extent_io_tree *unpin;
6799         u64 start;
6800         u64 end;
6801         int ret;
6802
6803         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
6804                 unpin = &fs_info->freed_extents[1];
6805         else
6806                 unpin = &fs_info->freed_extents[0];
6807
6808         while (!trans->aborted) {
6809                 struct extent_state *cached_state = NULL;
6810
6811                 mutex_lock(&fs_info->unused_bg_unpin_mutex);
6812                 ret = find_first_extent_bit(unpin, 0, &start, &end,
6813                                             EXTENT_DIRTY, &cached_state);
6814                 if (ret) {
6815                         mutex_unlock(&fs_info->unused_bg_unpin_mutex);
6816                         break;
6817                 }
6818
6819                 if (btrfs_test_opt(fs_info, DISCARD))
6820                         ret = btrfs_discard_extent(fs_info, start,
6821                                                    end + 1 - start, NULL);
6822
6823                 clear_extent_dirty(unpin, start, end, &cached_state);
6824                 unpin_extent_range(fs_info, start, end, true);
6825                 mutex_unlock(&fs_info->unused_bg_unpin_mutex);
6826                 free_extent_state(cached_state);
6827                 cond_resched();
6828         }
6829
6830         /*
6831          * Transaction is finished.  We don't need the lock anymore.  We
6832          * do need to clean up the block groups in case of a transaction
6833          * abort.
6834          */
6835         deleted_bgs = &trans->transaction->deleted_bgs;
6836         list_for_each_entry_safe(block_group, tmp, deleted_bgs, bg_list) {
6837                 u64 trimmed = 0;
6838
6839                 ret = -EROFS;
6840                 if (!trans->aborted)
6841                         ret = btrfs_discard_extent(fs_info,
6842                                                    block_group->key.objectid,
6843                                                    block_group->key.offset,
6844                                                    &trimmed);
6845
6846                 list_del_init(&block_group->bg_list);
6847                 btrfs_put_block_group_trimming(block_group);
6848                 btrfs_put_block_group(block_group);
6849
6850                 if (ret) {
6851                         const char *errstr = btrfs_decode_error(ret);
6852                         btrfs_warn(fs_info,
6853                            "discard failed while removing blockgroup: errno=%d %s",
6854                                    ret, errstr);
6855                 }
6856         }
6857
6858         return 0;
6859 }
6860
6861 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
6862                                struct btrfs_delayed_ref_node *node, u64 parent,
6863                                u64 root_objectid, u64 owner_objectid,
6864                                u64 owner_offset, int refs_to_drop,
6865                                struct btrfs_delayed_extent_op *extent_op)
6866 {
6867         struct btrfs_fs_info *info = trans->fs_info;
6868         struct btrfs_key key;
6869         struct btrfs_path *path;
6870         struct btrfs_root *extent_root = info->extent_root;
6871         struct extent_buffer *leaf;
6872         struct btrfs_extent_item *ei;
6873         struct btrfs_extent_inline_ref *iref;
6874         int ret;
6875         int is_data;
6876         int extent_slot = 0;
6877         int found_extent = 0;
6878         int num_to_del = 1;
6879         u32 item_size;
6880         u64 refs;
6881         u64 bytenr = node->bytenr;
6882         u64 num_bytes = node->num_bytes;
6883         int last_ref = 0;
6884         bool skinny_metadata = btrfs_fs_incompat(info, SKINNY_METADATA);
6885
6886         path = btrfs_alloc_path();
6887         if (!path)
6888                 return -ENOMEM;
6889
6890         path->reada = READA_FORWARD;
6891         path->leave_spinning = 1;
6892
6893         is_data = owner_objectid >= BTRFS_FIRST_FREE_OBJECTID;
6894         BUG_ON(!is_data && refs_to_drop != 1);
6895
6896         if (is_data)
6897                 skinny_metadata = false;
6898
6899         ret = lookup_extent_backref(trans, path, &iref, bytenr, num_bytes,
6900                                     parent, root_objectid, owner_objectid,
6901                                     owner_offset);
6902         if (ret == 0) {
6903                 extent_slot = path->slots[0];
6904                 while (extent_slot >= 0) {
6905                         btrfs_item_key_to_cpu(path->nodes[0], &key,
6906                                               extent_slot);
6907                         if (key.objectid != bytenr)
6908                                 break;
6909                         if (key.type == BTRFS_EXTENT_ITEM_KEY &&
6910                             key.offset == num_bytes) {
6911                                 found_extent = 1;
6912                                 break;
6913                         }
6914                         if (key.type == BTRFS_METADATA_ITEM_KEY &&
6915                             key.offset == owner_objectid) {
6916                                 found_extent = 1;
6917                                 break;
6918                         }
6919                         if (path->slots[0] - extent_slot > 5)
6920                                 break;
6921                         extent_slot--;
6922                 }
6923
6924                 if (!found_extent) {
6925                         BUG_ON(iref);
6926                         ret = remove_extent_backref(trans, path, NULL,
6927                                                     refs_to_drop,
6928                                                     is_data, &last_ref);
6929                         if (ret) {
6930                                 btrfs_abort_transaction(trans, ret);
6931                                 goto out;
6932                         }
6933                         btrfs_release_path(path);
6934                         path->leave_spinning = 1;
6935
6936                         key.objectid = bytenr;
6937                         key.type = BTRFS_EXTENT_ITEM_KEY;
6938                         key.offset = num_bytes;
6939
6940                         if (!is_data && skinny_metadata) {
6941                                 key.type = BTRFS_METADATA_ITEM_KEY;
6942                                 key.offset = owner_objectid;
6943                         }
6944
6945                         ret = btrfs_search_slot(trans, extent_root,
6946                                                 &key, path, -1, 1);
6947                         if (ret > 0 && skinny_metadata && path->slots[0]) {
6948                                 /*
6949                                  * Couldn't find our skinny metadata item,
6950                                  * see if we have ye olde extent item.
6951                                  */
6952                                 path->slots[0]--;
6953                                 btrfs_item_key_to_cpu(path->nodes[0], &key,
6954                                                       path->slots[0]);
6955                                 if (key.objectid == bytenr &&
6956                                     key.type == BTRFS_EXTENT_ITEM_KEY &&
6957                                     key.offset == num_bytes)
6958                                         ret = 0;
6959                         }
6960
6961                         if (ret > 0 && skinny_metadata) {
6962                                 skinny_metadata = false;
6963                                 key.objectid = bytenr;
6964                                 key.type = BTRFS_EXTENT_ITEM_KEY;
6965                                 key.offset = num_bytes;
6966                                 btrfs_release_path(path);
6967                                 ret = btrfs_search_slot(trans, extent_root,
6968                                                         &key, path, -1, 1);
6969                         }
6970
6971                         if (ret) {
6972                                 btrfs_err(info,
6973                                           "umm, got %d back from search, was looking for %llu",
6974                                           ret, bytenr);
6975                                 if (ret > 0)
6976                                         btrfs_print_leaf(path->nodes[0]);
6977                         }
6978                         if (ret < 0) {
6979                                 btrfs_abort_transaction(trans, ret);
6980                                 goto out;
6981                         }
6982                         extent_slot = path->slots[0];
6983                 }
6984         } else if (WARN_ON(ret == -ENOENT)) {
6985                 btrfs_print_leaf(path->nodes[0]);
6986                 btrfs_err(info,
6987                         "unable to find ref byte nr %llu parent %llu root %llu  owner %llu offset %llu",
6988                         bytenr, parent, root_objectid, owner_objectid,
6989                         owner_offset);
6990                 btrfs_abort_transaction(trans, ret);
6991                 goto out;
6992         } else {
6993                 btrfs_abort_transaction(trans, ret);
6994                 goto out;
6995         }
6996
6997         leaf = path->nodes[0];
6998         item_size = btrfs_item_size_nr(leaf, extent_slot);
6999         if (unlikely(item_size < sizeof(*ei))) {
7000                 ret = -EINVAL;
7001                 btrfs_print_v0_err(info);
7002                 btrfs_abort_transaction(trans, ret);
7003                 goto out;
7004         }
7005         ei = btrfs_item_ptr(leaf, extent_slot,
7006                             struct btrfs_extent_item);
7007         if (owner_objectid < BTRFS_FIRST_FREE_OBJECTID &&
7008             key.type == BTRFS_EXTENT_ITEM_KEY) {
7009                 struct btrfs_tree_block_info *bi;
7010                 BUG_ON(item_size < sizeof(*ei) + sizeof(*bi));
7011                 bi = (struct btrfs_tree_block_info *)(ei + 1);
7012                 WARN_ON(owner_objectid != btrfs_tree_block_level(leaf, bi));
7013         }
7014
7015         refs = btrfs_extent_refs(leaf, ei);
7016         if (refs < refs_to_drop) {
7017                 btrfs_err(info,
7018                           "trying to drop %d refs but we only have %Lu for bytenr %Lu",
7019                           refs_to_drop, refs, bytenr);
7020                 ret = -EINVAL;
7021                 btrfs_abort_transaction(trans, ret);
7022                 goto out;
7023         }
7024         refs -= refs_to_drop;
7025
7026         if (refs > 0) {
7027                 if (extent_op)
7028                         __run_delayed_extent_op(extent_op, leaf, ei);
7029                 /*
7030                  * In the case of inline back ref, reference count will
7031                  * be updated by remove_extent_backref
7032                  */
7033                 if (iref) {
7034                         BUG_ON(!found_extent);
7035                 } else {
7036                         btrfs_set_extent_refs(leaf, ei, refs);
7037                         btrfs_mark_buffer_dirty(leaf);
7038                 }
7039                 if (found_extent) {
7040                         ret = remove_extent_backref(trans, path, iref,
7041                                                     refs_to_drop, is_data,
7042                                                     &last_ref);
7043                         if (ret) {
7044                                 btrfs_abort_transaction(trans, ret);
7045                                 goto out;
7046                         }
7047                 }
7048         } else {
7049                 if (found_extent) {
7050                         BUG_ON(is_data && refs_to_drop !=
7051                                extent_data_ref_count(path, iref));
7052                         if (iref) {
7053                                 BUG_ON(path->slots[0] != extent_slot);
7054                         } else {
7055                                 BUG_ON(path->slots[0] != extent_slot + 1);
7056                                 path->slots[0] = extent_slot;
7057                                 num_to_del = 2;
7058                         }
7059                 }
7060
7061                 last_ref = 1;
7062                 ret = btrfs_del_items(trans, extent_root, path, path->slots[0],
7063                                       num_to_del);
7064                 if (ret) {
7065                         btrfs_abort_transaction(trans, ret);
7066                         goto out;
7067                 }
7068                 btrfs_release_path(path);
7069
7070                 if (is_data) {
7071                         ret = btrfs_del_csums(trans, info, bytenr, num_bytes);
7072                         if (ret) {
7073                                 btrfs_abort_transaction(trans, ret);
7074                                 goto out;
7075                         }
7076                 }
7077
7078                 ret = add_to_free_space_tree(trans, bytenr, num_bytes);
7079                 if (ret) {
7080                         btrfs_abort_transaction(trans, ret);
7081                         goto out;
7082                 }
7083
7084                 ret = update_block_group(trans, bytenr, num_bytes, 0);
7085                 if (ret) {
7086                         btrfs_abort_transaction(trans, ret);
7087                         goto out;
7088                 }
7089         }
7090         btrfs_release_path(path);
7091
7092 out:
7093         btrfs_free_path(path);
7094         return ret;
7095 }
7096
7097 /*
7098  * when we free an block, it is possible (and likely) that we free the last
7099  * delayed ref for that extent as well.  This searches the delayed ref tree for
7100  * a given extent, and if there are no other delayed refs to be processed, it
7101  * removes it from the tree.
7102  */
7103 static noinline int check_ref_cleanup(struct btrfs_trans_handle *trans,
7104                                       u64 bytenr)
7105 {
7106         struct btrfs_delayed_ref_head *head;
7107         struct btrfs_delayed_ref_root *delayed_refs;
7108         int ret = 0;
7109
7110         delayed_refs = &trans->transaction->delayed_refs;
7111         spin_lock(&delayed_refs->lock);
7112         head = btrfs_find_delayed_ref_head(delayed_refs, bytenr);
7113         if (!head)
7114                 goto out_delayed_unlock;
7115
7116         spin_lock(&head->lock);
7117         if (!RB_EMPTY_ROOT(&head->ref_tree.rb_root))
7118                 goto out;
7119
7120         if (cleanup_extent_op(head) != NULL)
7121                 goto out;
7122
7123         /*
7124          * waiting for the lock here would deadlock.  If someone else has it
7125          * locked they are already in the process of dropping it anyway
7126          */
7127         if (!mutex_trylock(&head->mutex))
7128                 goto out;
7129
7130         btrfs_delete_ref_head(delayed_refs, head);
7131         head->processing = 0;
7132
7133         spin_unlock(&head->lock);
7134         spin_unlock(&delayed_refs->lock);
7135
7136         BUG_ON(head->extent_op);
7137         if (head->must_insert_reserved)
7138                 ret = 1;
7139
7140         btrfs_cleanup_ref_head_accounting(trans->fs_info, delayed_refs, head);
7141         mutex_unlock(&head->mutex);
7142         btrfs_put_delayed_ref_head(head);
7143         return ret;
7144 out:
7145         spin_unlock(&head->lock);
7146
7147 out_delayed_unlock:
7148         spin_unlock(&delayed_refs->lock);
7149         return 0;
7150 }
7151
7152 void btrfs_free_tree_block(struct btrfs_trans_handle *trans,
7153                            struct btrfs_root *root,
7154                            struct extent_buffer *buf,
7155                            u64 parent, int last_ref)
7156 {
7157         struct btrfs_fs_info *fs_info = root->fs_info;
7158         struct btrfs_ref generic_ref = { 0 };
7159         int pin = 1;
7160         int ret;
7161
7162         btrfs_init_generic_ref(&generic_ref, BTRFS_DROP_DELAYED_REF,
7163                                buf->start, buf->len, parent);
7164         btrfs_init_tree_ref(&generic_ref, btrfs_header_level(buf),
7165                             root->root_key.objectid);
7166
7167         if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
7168                 int old_ref_mod, new_ref_mod;
7169
7170                 btrfs_ref_tree_mod(fs_info, &generic_ref);
7171                 ret = btrfs_add_delayed_tree_ref(trans, &generic_ref, NULL,
7172                                                  &old_ref_mod, &new_ref_mod);
7173                 BUG_ON(ret); /* -ENOMEM */
7174                 pin = old_ref_mod >= 0 && new_ref_mod < 0;
7175         }
7176
7177         if (last_ref && btrfs_header_generation(buf) == trans->transid) {
7178                 struct btrfs_block_group_cache *cache;
7179
7180                 if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
7181                         ret = check_ref_cleanup(trans, buf->start);
7182                         if (!ret)
7183                                 goto out;
7184                 }
7185
7186                 pin = 0;
7187                 cache = btrfs_lookup_block_group(fs_info, buf->start);
7188
7189                 if (btrfs_header_flag(buf, BTRFS_HEADER_FLAG_WRITTEN)) {
7190                         pin_down_extent(cache, buf->start, buf->len, 1);
7191                         btrfs_put_block_group(cache);
7192                         goto out;
7193                 }
7194
7195                 WARN_ON(test_bit(EXTENT_BUFFER_DIRTY, &buf->bflags));
7196
7197                 btrfs_add_free_space(cache, buf->start, buf->len);
7198                 btrfs_free_reserved_bytes(cache, buf->len, 0);
7199                 btrfs_put_block_group(cache);
7200                 trace_btrfs_reserved_extent_free(fs_info, buf->start, buf->len);
7201         }
7202 out:
7203         if (pin)
7204                 add_pinned_bytes(fs_info, &generic_ref);
7205
7206         if (last_ref) {
7207                 /*
7208                  * Deleting the buffer, clear the corrupt flag since it doesn't
7209                  * matter anymore.
7210                  */
7211                 clear_bit(EXTENT_BUFFER_CORRUPT, &buf->bflags);
7212         }
7213 }
7214
7215 /* Can return -ENOMEM */
7216 int btrfs_free_extent(struct btrfs_trans_handle *trans, struct btrfs_ref *ref)
7217 {
7218         struct btrfs_fs_info *fs_info = trans->fs_info;
7219         int old_ref_mod, new_ref_mod;
7220         int ret;
7221
7222         if (btrfs_is_testing(fs_info))
7223                 return 0;
7224
7225         /*
7226          * tree log blocks never actually go into the extent allocation
7227          * tree, just update pinning info and exit early.
7228          */
7229         if ((ref->type == BTRFS_REF_METADATA &&
7230              ref->tree_ref.root == BTRFS_TREE_LOG_OBJECTID) ||
7231             (ref->type == BTRFS_REF_DATA &&
7232              ref->data_ref.ref_root == BTRFS_TREE_LOG_OBJECTID)) {
7233                 /* unlocks the pinned mutex */
7234                 btrfs_pin_extent(fs_info, ref->bytenr, ref->len, 1);
7235                 old_ref_mod = new_ref_mod = 0;
7236                 ret = 0;
7237         } else if (ref->type == BTRFS_REF_METADATA) {
7238                 ret = btrfs_add_delayed_tree_ref(trans, ref, NULL,
7239                                                  &old_ref_mod, &new_ref_mod);
7240         } else {
7241                 ret = btrfs_add_delayed_data_ref(trans, ref, 0,
7242                                                  &old_ref_mod, &new_ref_mod);
7243         }
7244
7245         if (!((ref->type == BTRFS_REF_METADATA &&
7246                ref->tree_ref.root == BTRFS_TREE_LOG_OBJECTID) ||
7247               (ref->type == BTRFS_REF_DATA &&
7248                ref->data_ref.ref_root == BTRFS_TREE_LOG_OBJECTID)))
7249                 btrfs_ref_tree_mod(fs_info, ref);
7250
7251         if (ret == 0 && old_ref_mod >= 0 && new_ref_mod < 0)
7252                 add_pinned_bytes(fs_info, ref);
7253
7254         return ret;
7255 }
7256
7257 /*
7258  * when we wait for progress in the block group caching, its because
7259  * our allocation attempt failed at least once.  So, we must sleep
7260  * and let some progress happen before we try again.
7261  *
7262  * This function will sleep at least once waiting for new free space to
7263  * show up, and then it will check the block group free space numbers
7264  * for our min num_bytes.  Another option is to have it go ahead
7265  * and look in the rbtree for a free extent of a given size, but this
7266  * is a good start.
7267  *
7268  * Callers of this must check if cache->cached == BTRFS_CACHE_ERROR before using
7269  * any of the information in this block group.
7270  */
7271 static noinline void
7272 wait_block_group_cache_progress(struct btrfs_block_group_cache *cache,
7273                                 u64 num_bytes)
7274 {
7275         struct btrfs_caching_control *caching_ctl;
7276
7277         caching_ctl = get_caching_control(cache);
7278         if (!caching_ctl)
7279                 return;
7280
7281         wait_event(caching_ctl->wait, block_group_cache_done(cache) ||
7282                    (cache->free_space_ctl->free_space >= num_bytes));
7283
7284         put_caching_control(caching_ctl);
7285 }
7286
7287 static noinline int
7288 wait_block_group_cache_done(struct btrfs_block_group_cache *cache)
7289 {
7290         struct btrfs_caching_control *caching_ctl;
7291         int ret = 0;
7292
7293         caching_ctl = get_caching_control(cache);
7294         if (!caching_ctl)
7295                 return (cache->cached == BTRFS_CACHE_ERROR) ? -EIO : 0;
7296
7297         wait_event(caching_ctl->wait, block_group_cache_done(cache));
7298         if (cache->cached == BTRFS_CACHE_ERROR)
7299                 ret = -EIO;
7300         put_caching_control(caching_ctl);
7301         return ret;
7302 }
7303
7304 enum btrfs_loop_type {
7305         LOOP_CACHING_NOWAIT = 0,
7306         LOOP_CACHING_WAIT = 1,
7307         LOOP_ALLOC_CHUNK = 2,
7308         LOOP_NO_EMPTY_SIZE = 3,
7309 };
7310
7311 static inline void
7312 btrfs_lock_block_group(struct btrfs_block_group_cache *cache,
7313                        int delalloc)
7314 {
7315         if (delalloc)
7316                 down_read(&cache->data_rwsem);
7317 }
7318
7319 static inline void
7320 btrfs_grab_block_group(struct btrfs_block_group_cache *cache,
7321                        int delalloc)
7322 {
7323         btrfs_get_block_group(cache);
7324         if (delalloc)
7325                 down_read(&cache->data_rwsem);
7326 }
7327
7328 static struct btrfs_block_group_cache *
7329 btrfs_lock_cluster(struct btrfs_block_group_cache *block_group,
7330                    struct btrfs_free_cluster *cluster,
7331                    int delalloc)
7332 {
7333         struct btrfs_block_group_cache *used_bg = NULL;
7334
7335         spin_lock(&cluster->refill_lock);
7336         while (1) {
7337                 used_bg = cluster->block_group;
7338                 if (!used_bg)
7339                         return NULL;
7340
7341                 if (used_bg == block_group)
7342                         return used_bg;
7343
7344                 btrfs_get_block_group(used_bg);
7345
7346                 if (!delalloc)
7347                         return used_bg;
7348
7349                 if (down_read_trylock(&used_bg->data_rwsem))
7350                         return used_bg;
7351
7352                 spin_unlock(&cluster->refill_lock);
7353
7354                 /* We should only have one-level nested. */
7355                 down_read_nested(&used_bg->data_rwsem, SINGLE_DEPTH_NESTING);
7356
7357                 spin_lock(&cluster->refill_lock);
7358                 if (used_bg == cluster->block_group)
7359                         return used_bg;
7360
7361                 up_read(&used_bg->data_rwsem);
7362                 btrfs_put_block_group(used_bg);
7363         }
7364 }
7365
7366 static inline void
7367 btrfs_release_block_group(struct btrfs_block_group_cache *cache,
7368                          int delalloc)
7369 {
7370         if (delalloc)
7371                 up_read(&cache->data_rwsem);
7372         btrfs_put_block_group(cache);
7373 }
7374
7375 /*
7376  * Structure used internally for find_free_extent() function.  Wraps needed
7377  * parameters.
7378  */
7379 struct find_free_extent_ctl {
7380         /* Basic allocation info */
7381         u64 ram_bytes;
7382         u64 num_bytes;
7383         u64 empty_size;
7384         u64 flags;
7385         int delalloc;
7386
7387         /* Where to start the search inside the bg */
7388         u64 search_start;
7389
7390         /* For clustered allocation */
7391         u64 empty_cluster;
7392
7393         bool have_caching_bg;
7394         bool orig_have_caching_bg;
7395
7396         /* RAID index, converted from flags */
7397         int index;
7398
7399         /*
7400          * Current loop number, check find_free_extent_update_loop() for details
7401          */
7402         int loop;
7403
7404         /*
7405          * Whether we're refilling a cluster, if true we need to re-search
7406          * current block group but don't try to refill the cluster again.
7407          */
7408         bool retry_clustered;
7409
7410         /*
7411          * Whether we're updating free space cache, if true we need to re-search
7412          * current block group but don't try updating free space cache again.
7413          */
7414         bool retry_unclustered;
7415
7416         /* If current block group is cached */
7417         int cached;
7418
7419         /* Max contiguous hole found */
7420         u64 max_extent_size;
7421
7422         /* Total free space from free space cache, not always contiguous */
7423         u64 total_free_space;
7424
7425         /* Found result */
7426         u64 found_offset;
7427 };
7428
7429
7430 /*
7431  * Helper function for find_free_extent().
7432  *
7433  * Return -ENOENT to inform caller that we need fallback to unclustered mode.
7434  * Return -EAGAIN to inform caller that we need to re-search this block group
7435  * Return >0 to inform caller that we find nothing
7436  * Return 0 means we have found a location and set ffe_ctl->found_offset.
7437  */
7438 static int find_free_extent_clustered(struct btrfs_block_group_cache *bg,
7439                 struct btrfs_free_cluster *last_ptr,
7440                 struct find_free_extent_ctl *ffe_ctl,
7441                 struct btrfs_block_group_cache **cluster_bg_ret)
7442 {
7443         struct btrfs_block_group_cache *cluster_bg;
7444         u64 aligned_cluster;
7445         u64 offset;
7446         int ret;
7447
7448         cluster_bg = btrfs_lock_cluster(bg, last_ptr, ffe_ctl->delalloc);
7449         if (!cluster_bg)
7450                 goto refill_cluster;
7451         if (cluster_bg != bg && (cluster_bg->ro ||
7452             !block_group_bits(cluster_bg, ffe_ctl->flags)))
7453                 goto release_cluster;
7454
7455         offset = btrfs_alloc_from_cluster(cluster_bg, last_ptr,
7456                         ffe_ctl->num_bytes, cluster_bg->key.objectid,
7457                         &ffe_ctl->max_extent_size);
7458         if (offset) {
7459                 /* We have a block, we're done */
7460                 spin_unlock(&last_ptr->refill_lock);
7461                 trace_btrfs_reserve_extent_cluster(cluster_bg,
7462                                 ffe_ctl->search_start, ffe_ctl->num_bytes);
7463                 *cluster_bg_ret = cluster_bg;
7464                 ffe_ctl->found_offset = offset;
7465                 return 0;
7466         }
7467         WARN_ON(last_ptr->block_group != cluster_bg);
7468
7469 release_cluster:
7470         /*
7471          * If we are on LOOP_NO_EMPTY_SIZE, we can't set up a new clusters, so
7472          * lets just skip it and let the allocator find whatever block it can
7473          * find. If we reach this point, we will have tried the cluster
7474          * allocator plenty of times and not have found anything, so we are
7475          * likely way too fragmented for the clustering stuff to find anything.
7476          *
7477          * However, if the cluster is taken from the current block group,
7478          * release the cluster first, so that we stand a better chance of
7479          * succeeding in the unclustered allocation.
7480          */
7481         if (ffe_ctl->loop >= LOOP_NO_EMPTY_SIZE && cluster_bg != bg) {
7482                 spin_unlock(&last_ptr->refill_lock);
7483                 btrfs_release_block_group(cluster_bg, ffe_ctl->delalloc);
7484                 return -ENOENT;
7485         }
7486
7487         /* This cluster didn't work out, free it and start over */
7488         btrfs_return_cluster_to_free_space(NULL, last_ptr);
7489
7490         if (cluster_bg != bg)
7491                 btrfs_release_block_group(cluster_bg, ffe_ctl->delalloc);
7492
7493 refill_cluster:
7494         if (ffe_ctl->loop >= LOOP_NO_EMPTY_SIZE) {
7495                 spin_unlock(&last_ptr->refill_lock);
7496                 return -ENOENT;
7497         }
7498
7499         aligned_cluster = max_t(u64,
7500                         ffe_ctl->empty_cluster + ffe_ctl->empty_size,
7501                         bg->full_stripe_len);
7502         ret = btrfs_find_space_cluster(bg, last_ptr, ffe_ctl->search_start,
7503                         ffe_ctl->num_bytes, aligned_cluster);
7504         if (ret == 0) {
7505                 /* Now pull our allocation out of this cluster */
7506                 offset = btrfs_alloc_from_cluster(bg, last_ptr,
7507                                 ffe_ctl->num_bytes, ffe_ctl->search_start,
7508                                 &ffe_ctl->max_extent_size);
7509                 if (offset) {
7510                         /* We found one, proceed */
7511                         spin_unlock(&last_ptr->refill_lock);
7512                         trace_btrfs_reserve_extent_cluster(bg,
7513                                         ffe_ctl->search_start,
7514                                         ffe_ctl->num_bytes);
7515                         ffe_ctl->found_offset = offset;
7516                         return 0;
7517                 }
7518         } else if (!ffe_ctl->cached && ffe_ctl->loop > LOOP_CACHING_NOWAIT &&
7519                    !ffe_ctl->retry_clustered) {
7520                 spin_unlock(&last_ptr->refill_lock);
7521
7522                 ffe_ctl->retry_clustered = true;
7523                 wait_block_group_cache_progress(bg, ffe_ctl->num_bytes +
7524                                 ffe_ctl->empty_cluster + ffe_ctl->empty_size);
7525                 return -EAGAIN;
7526         }
7527         /*
7528          * At this point we either didn't find a cluster or we weren't able to
7529          * allocate a block from our cluster.  Free the cluster we've been
7530          * trying to use, and go to the next block group.
7531          */
7532         btrfs_return_cluster_to_free_space(NULL, last_ptr);
7533         spin_unlock(&last_ptr->refill_lock);
7534         return 1;
7535 }
7536
7537 /*
7538  * Return >0 to inform caller that we find nothing
7539  * Return 0 when we found an free extent and set ffe_ctrl->found_offset
7540  * Return -EAGAIN to inform caller that we need to re-search this block group
7541  */
7542 static int find_free_extent_unclustered(struct btrfs_block_group_cache *bg,
7543                 struct btrfs_free_cluster *last_ptr,
7544                 struct find_free_extent_ctl *ffe_ctl)
7545 {
7546         u64 offset;
7547
7548         /*
7549          * We are doing an unclustered allocation, set the fragmented flag so
7550          * we don't bother trying to setup a cluster again until we get more
7551          * space.
7552          */
7553         if (unlikely(last_ptr)) {
7554                 spin_lock(&last_ptr->lock);
7555                 last_ptr->fragmented = 1;
7556                 spin_unlock(&last_ptr->lock);
7557         }
7558         if (ffe_ctl->cached) {
7559                 struct btrfs_free_space_ctl *free_space_ctl;
7560
7561                 free_space_ctl = bg->free_space_ctl;
7562                 spin_lock(&free_space_ctl->tree_lock);
7563                 if (free_space_ctl->free_space <
7564                     ffe_ctl->num_bytes + ffe_ctl->empty_cluster +
7565                     ffe_ctl->empty_size) {
7566                         ffe_ctl->total_free_space = max_t(u64,
7567                                         ffe_ctl->total_free_space,
7568                                         free_space_ctl->free_space);
7569                         spin_unlock(&free_space_ctl->tree_lock);
7570                         return 1;
7571                 }
7572                 spin_unlock(&free_space_ctl->tree_lock);
7573         }
7574
7575         offset = btrfs_find_space_for_alloc(bg, ffe_ctl->search_start,
7576                         ffe_ctl->num_bytes, ffe_ctl->empty_size,
7577                         &ffe_ctl->max_extent_size);
7578
7579         /*
7580          * If we didn't find a chunk, and we haven't failed on this block group
7581          * before, and this block group is in the middle of caching and we are
7582          * ok with waiting, then go ahead and wait for progress to be made, and
7583          * set @retry_unclustered to true.
7584          *
7585          * If @retry_unclustered is true then we've already waited on this
7586          * block group once and should move on to the next block group.
7587          */
7588         if (!offset && !ffe_ctl->retry_unclustered && !ffe_ctl->cached &&
7589             ffe_ctl->loop > LOOP_CACHING_NOWAIT) {
7590                 wait_block_group_cache_progress(bg, ffe_ctl->num_bytes +
7591                                                 ffe_ctl->empty_size);
7592                 ffe_ctl->retry_unclustered = true;
7593                 return -EAGAIN;
7594         } else if (!offset) {
7595                 return 1;
7596         }
7597         ffe_ctl->found_offset = offset;
7598         return 0;
7599 }
7600
7601 /*
7602  * Return >0 means caller needs to re-search for free extent
7603  * Return 0 means we have the needed free extent.
7604  * Return <0 means we failed to locate any free extent.
7605  */
7606 static int find_free_extent_update_loop(struct btrfs_fs_info *fs_info,
7607                                         struct btrfs_free_cluster *last_ptr,
7608                                         struct btrfs_key *ins,
7609                                         struct find_free_extent_ctl *ffe_ctl,
7610                                         int full_search, bool use_cluster)
7611 {
7612         struct btrfs_root *root = fs_info->extent_root;
7613         int ret;
7614
7615         if ((ffe_ctl->loop == LOOP_CACHING_NOWAIT) &&
7616             ffe_ctl->have_caching_bg && !ffe_ctl->orig_have_caching_bg)
7617                 ffe_ctl->orig_have_caching_bg = true;
7618
7619         if (!ins->objectid && ffe_ctl->loop >= LOOP_CACHING_WAIT &&
7620             ffe_ctl->have_caching_bg)
7621                 return 1;
7622
7623         if (!ins->objectid && ++(ffe_ctl->index) < BTRFS_NR_RAID_TYPES)
7624                 return 1;
7625
7626         if (ins->objectid) {
7627                 if (!use_cluster && last_ptr) {
7628                         spin_lock(&last_ptr->lock);
7629                         last_ptr->window_start = ins->objectid;
7630                         spin_unlock(&last_ptr->lock);
7631                 }
7632                 return 0;
7633         }
7634
7635         /*
7636          * LOOP_CACHING_NOWAIT, search partially cached block groups, kicking
7637          *                      caching kthreads as we move along
7638          * LOOP_CACHING_WAIT, search everything, and wait if our bg is caching
7639          * LOOP_ALLOC_CHUNK, force a chunk allocation and try again
7640          * LOOP_NO_EMPTY_SIZE, set empty_size and empty_cluster to 0 and try
7641          *                     again
7642          */
7643         if (ffe_ctl->loop < LOOP_NO_EMPTY_SIZE) {
7644                 ffe_ctl->index = 0;
7645                 if (ffe_ctl->loop == LOOP_CACHING_NOWAIT) {
7646                         /*
7647                          * We want to skip the LOOP_CACHING_WAIT step if we
7648                          * don't have any uncached bgs and we've already done a
7649                          * full search through.
7650                          */
7651                         if (ffe_ctl->orig_have_caching_bg || !full_search)
7652                                 ffe_ctl->loop = LOOP_CACHING_WAIT;
7653                         else
7654                                 ffe_ctl->loop = LOOP_ALLOC_CHUNK;
7655                 } else {
7656                         ffe_ctl->loop++;
7657                 }
7658
7659                 if (ffe_ctl->loop == LOOP_ALLOC_CHUNK) {
7660                         struct btrfs_trans_handle *trans;
7661                         int exist = 0;
7662
7663                         trans = current->journal_info;
7664                         if (trans)
7665                                 exist = 1;
7666                         else
7667                                 trans = btrfs_join_transaction(root);
7668
7669                         if (IS_ERR(trans)) {
7670                                 ret = PTR_ERR(trans);
7671                                 return ret;
7672                         }
7673
7674                         ret = do_chunk_alloc(trans, ffe_ctl->flags,
7675                                              CHUNK_ALLOC_FORCE);
7676
7677                         /*
7678                          * If we can't allocate a new chunk we've already looped
7679                          * through at least once, move on to the NO_EMPTY_SIZE
7680                          * case.
7681                          */
7682                         if (ret == -ENOSPC)
7683                                 ffe_ctl->loop = LOOP_NO_EMPTY_SIZE;
7684
7685                         /* Do not bail out on ENOSPC since we can do more. */
7686                         if (ret < 0 && ret != -ENOSPC)
7687                                 btrfs_abort_transaction(trans, ret);
7688                         else
7689                                 ret = 0;
7690                         if (!exist)
7691                                 btrfs_end_transaction(trans);
7692                         if (ret)
7693                                 return ret;
7694                 }
7695
7696                 if (ffe_ctl->loop == LOOP_NO_EMPTY_SIZE) {
7697                         /*
7698                          * Don't loop again if we already have no empty_size and
7699                          * no empty_cluster.
7700                          */
7701                         if (ffe_ctl->empty_size == 0 &&
7702                             ffe_ctl->empty_cluster == 0)
7703                                 return -ENOSPC;
7704                         ffe_ctl->empty_size = 0;
7705                         ffe_ctl->empty_cluster = 0;
7706                 }
7707                 return 1;
7708         }
7709         return -ENOSPC;
7710 }
7711
7712 /*
7713  * walks the btree of allocated extents and find a hole of a given size.
7714  * The key ins is changed to record the hole:
7715  * ins->objectid == start position
7716  * ins->flags = BTRFS_EXTENT_ITEM_KEY
7717  * ins->offset == the size of the hole.
7718  * Any available blocks before search_start are skipped.
7719  *
7720  * If there is no suitable free space, we will record the max size of
7721  * the free space extent currently.
7722  *
7723  * The overall logic and call chain:
7724  *
7725  * find_free_extent()
7726  * |- Iterate through all block groups
7727  * |  |- Get a valid block group
7728  * |  |- Try to do clustered allocation in that block group
7729  * |  |- Try to do unclustered allocation in that block group
7730  * |  |- Check if the result is valid
7731  * |  |  |- If valid, then exit
7732  * |  |- Jump to next block group
7733  * |
7734  * |- Push harder to find free extents
7735  *    |- If not found, re-iterate all block groups
7736  */
7737 static noinline int find_free_extent(struct btrfs_fs_info *fs_info,
7738                                 u64 ram_bytes, u64 num_bytes, u64 empty_size,
7739                                 u64 hint_byte, struct btrfs_key *ins,
7740                                 u64 flags, int delalloc)
7741 {
7742         int ret = 0;
7743         struct btrfs_free_cluster *last_ptr = NULL;
7744         struct btrfs_block_group_cache *block_group = NULL;
7745         struct find_free_extent_ctl ffe_ctl = {0};
7746         struct btrfs_space_info *space_info;
7747         bool use_cluster = true;
7748         bool full_search = false;
7749
7750         WARN_ON(num_bytes < fs_info->sectorsize);
7751
7752         ffe_ctl.ram_bytes = ram_bytes;
7753         ffe_ctl.num_bytes = num_bytes;
7754         ffe_ctl.empty_size = empty_size;
7755         ffe_ctl.flags = flags;
7756         ffe_ctl.search_start = 0;
7757         ffe_ctl.retry_clustered = false;
7758         ffe_ctl.retry_unclustered = false;
7759         ffe_ctl.delalloc = delalloc;
7760         ffe_ctl.index = btrfs_bg_flags_to_raid_index(flags);
7761         ffe_ctl.have_caching_bg = false;
7762         ffe_ctl.orig_have_caching_bg = false;
7763         ffe_ctl.found_offset = 0;
7764
7765         ins->type = BTRFS_EXTENT_ITEM_KEY;
7766         ins->objectid = 0;
7767         ins->offset = 0;
7768
7769         trace_find_free_extent(fs_info, num_bytes, empty_size, flags);
7770
7771         space_info = __find_space_info(fs_info, flags);
7772         if (!space_info) {
7773                 btrfs_err(fs_info, "No space info for %llu", flags);
7774                 return -ENOSPC;
7775         }
7776
7777         /*
7778          * If our free space is heavily fragmented we may not be able to make
7779          * big contiguous allocations, so instead of doing the expensive search
7780          * for free space, simply return ENOSPC with our max_extent_size so we
7781          * can go ahead and search for a more manageable chunk.
7782          *
7783          * If our max_extent_size is large enough for our allocation simply
7784          * disable clustering since we will likely not be able to find enough
7785          * space to create a cluster and induce latency trying.
7786          */
7787         if (unlikely(space_info->max_extent_size)) {
7788                 spin_lock(&space_info->lock);
7789                 if (space_info->max_extent_size &&
7790                     num_bytes > space_info->max_extent_size) {
7791                         ins->offset = space_info->max_extent_size;
7792                         spin_unlock(&space_info->lock);
7793                         return -ENOSPC;
7794                 } else if (space_info->max_extent_size) {
7795                         use_cluster = false;
7796                 }
7797                 spin_unlock(&space_info->lock);
7798         }
7799
7800         last_ptr = fetch_cluster_info(fs_info, space_info,
7801                                       &ffe_ctl.empty_cluster);
7802         if (last_ptr) {
7803                 spin_lock(&last_ptr->lock);
7804                 if (last_ptr->block_group)
7805                         hint_byte = last_ptr->window_start;
7806                 if (last_ptr->fragmented) {
7807                         /*
7808                          * We still set window_start so we can keep track of the
7809                          * last place we found an allocation to try and save
7810                          * some time.
7811                          */
7812                         hint_byte = last_ptr->window_start;
7813                         use_cluster = false;
7814                 }
7815                 spin_unlock(&last_ptr->lock);
7816         }
7817
7818         ffe_ctl.search_start = max(ffe_ctl.search_start,
7819                                    first_logical_byte(fs_info, 0));
7820         ffe_ctl.search_start = max(ffe_ctl.search_start, hint_byte);
7821         if (ffe_ctl.search_start == hint_byte) {
7822                 block_group = btrfs_lookup_block_group(fs_info,
7823                                                        ffe_ctl.search_start);
7824                 /*
7825                  * we don't want to use the block group if it doesn't match our
7826                  * allocation bits, or if its not cached.
7827                  *
7828                  * However if we are re-searching with an ideal block group
7829                  * picked out then we don't care that the block group is cached.
7830                  */
7831                 if (block_group && block_group_bits(block_group, flags) &&
7832                     block_group->cached != BTRFS_CACHE_NO) {
7833                         down_read(&space_info->groups_sem);
7834                         if (list_empty(&block_group->list) ||
7835                             block_group->ro) {
7836                                 /*
7837                                  * someone is removing this block group,
7838                                  * we can't jump into the have_block_group
7839                                  * target because our list pointers are not
7840                                  * valid
7841                                  */
7842                                 btrfs_put_block_group(block_group);
7843                                 up_read(&space_info->groups_sem);
7844                         } else {
7845                                 ffe_ctl.index = btrfs_bg_flags_to_raid_index(
7846                                                 block_group->flags);
7847                                 btrfs_lock_block_group(block_group, delalloc);
7848                                 goto have_block_group;
7849                         }
7850                 } else if (block_group) {
7851                         btrfs_put_block_group(block_group);
7852                 }
7853         }
7854 search:
7855         ffe_ctl.have_caching_bg = false;
7856         if (ffe_ctl.index == btrfs_bg_flags_to_raid_index(flags) ||
7857             ffe_ctl.index == 0)
7858                 full_search = true;
7859         down_read(&space_info->groups_sem);
7860         list_for_each_entry(block_group,
7861                             &space_info->block_groups[ffe_ctl.index], list) {
7862                 /* If the block group is read-only, we can skip it entirely. */
7863                 if (unlikely(block_group->ro))
7864                         continue;
7865
7866                 btrfs_grab_block_group(block_group, delalloc);
7867                 ffe_ctl.search_start = block_group->key.objectid;
7868
7869                 /*
7870                  * this can happen if we end up cycling through all the
7871                  * raid types, but we want to make sure we only allocate
7872                  * for the proper type.
7873                  */
7874                 if (!block_group_bits(block_group, flags)) {
7875                         u64 extra = BTRFS_BLOCK_GROUP_DUP |
7876                                 BTRFS_BLOCK_GROUP_RAID1_MASK |
7877                                 BTRFS_BLOCK_GROUP_RAID56_MASK |
7878                                 BTRFS_BLOCK_GROUP_RAID10;
7879
7880                         /*
7881                          * if they asked for extra copies and this block group
7882                          * doesn't provide them, bail.  This does allow us to
7883                          * fill raid0 from raid1.
7884                          */
7885                         if ((flags & extra) && !(block_group->flags & extra))
7886                                 goto loop;
7887                 }
7888
7889 have_block_group:
7890                 ffe_ctl.cached = block_group_cache_done(block_group);
7891                 if (unlikely(!ffe_ctl.cached)) {
7892                         ffe_ctl.have_caching_bg = true;
7893                         ret = cache_block_group(block_group, 0);
7894                         BUG_ON(ret < 0);
7895                         ret = 0;
7896                 }
7897
7898                 if (unlikely(block_group->cached == BTRFS_CACHE_ERROR))
7899                         goto loop;
7900
7901                 /*
7902                  * Ok we want to try and use the cluster allocator, so
7903                  * lets look there
7904                  */
7905                 if (last_ptr && use_cluster) {
7906                         struct btrfs_block_group_cache *cluster_bg = NULL;
7907
7908                         ret = find_free_extent_clustered(block_group, last_ptr,
7909                                                          &ffe_ctl, &cluster_bg);
7910
7911                         if (ret == 0) {
7912                                 if (cluster_bg && cluster_bg != block_group) {
7913                                         btrfs_release_block_group(block_group,
7914                                                                   delalloc);
7915                                         block_group = cluster_bg;
7916                                 }
7917                                 goto checks;
7918                         } else if (ret == -EAGAIN) {
7919                                 goto have_block_group;
7920                         } else if (ret > 0) {
7921                                 goto loop;
7922                         }
7923                         /* ret == -ENOENT case falls through */
7924                 }
7925
7926                 ret = find_free_extent_unclustered(block_group, last_ptr,
7927                                                    &ffe_ctl);
7928                 if (ret == -EAGAIN)
7929                         goto have_block_group;
7930                 else if (ret > 0)
7931                         goto loop;
7932                 /* ret == 0 case falls through */
7933 checks:
7934                 ffe_ctl.search_start = round_up(ffe_ctl.found_offset,
7935                                              fs_info->stripesize);
7936
7937                 /* move on to the next group */
7938                 if (ffe_ctl.search_start + num_bytes >
7939                     block_group->key.objectid + block_group->key.offset) {
7940                         btrfs_add_free_space(block_group, ffe_ctl.found_offset,
7941                                              num_bytes);
7942                         goto loop;
7943                 }
7944
7945                 if (ffe_ctl.found_offset < ffe_ctl.search_start)
7946                         btrfs_add_free_space(block_group, ffe_ctl.found_offset,
7947                                 ffe_ctl.search_start - ffe_ctl.found_offset);
7948
7949                 ret = btrfs_add_reserved_bytes(block_group, ram_bytes,
7950                                 num_bytes, delalloc);
7951                 if (ret == -EAGAIN) {
7952                         btrfs_add_free_space(block_group, ffe_ctl.found_offset,
7953                                              num_bytes);
7954                         goto loop;
7955                 }
7956                 btrfs_inc_block_group_reservations(block_group);
7957
7958                 /* we are all good, lets return */
7959                 ins->objectid = ffe_ctl.search_start;
7960                 ins->offset = num_bytes;
7961
7962                 trace_btrfs_reserve_extent(block_group, ffe_ctl.search_start,
7963                                            num_bytes);
7964                 btrfs_release_block_group(block_group, delalloc);
7965                 break;
7966 loop:
7967                 ffe_ctl.retry_clustered = false;
7968                 ffe_ctl.retry_unclustered = false;
7969                 BUG_ON(btrfs_bg_flags_to_raid_index(block_group->flags) !=
7970                        ffe_ctl.index);
7971                 btrfs_release_block_group(block_group, delalloc);
7972                 cond_resched();
7973         }
7974         up_read(&space_info->groups_sem);
7975
7976         ret = find_free_extent_update_loop(fs_info, last_ptr, ins, &ffe_ctl,
7977                                            full_search, use_cluster);
7978         if (ret > 0)
7979                 goto search;
7980
7981         if (ret == -ENOSPC) {
7982                 /*
7983                  * Use ffe_ctl->total_free_space as fallback if we can't find
7984                  * any contiguous hole.
7985                  */
7986                 if (!ffe_ctl.max_extent_size)
7987                         ffe_ctl.max_extent_size = ffe_ctl.total_free_space;
7988                 spin_lock(&space_info->lock);
7989                 space_info->max_extent_size = ffe_ctl.max_extent_size;
7990                 spin_unlock(&space_info->lock);
7991                 ins->offset = ffe_ctl.max_extent_size;
7992         }
7993         return ret;
7994 }
7995
7996 #define DUMP_BLOCK_RSV(fs_info, rsv_name)                               \
7997 do {                                                                    \
7998         struct btrfs_block_rsv *__rsv = &(fs_info)->rsv_name;           \
7999         spin_lock(&__rsv->lock);                                        \
8000         btrfs_info(fs_info, #rsv_name ": size %llu reserved %llu",      \
8001                    __rsv->size, __rsv->reserved);                       \
8002         spin_unlock(&__rsv->lock);                                      \
8003 } while (0)
8004
8005 static void dump_space_info(struct btrfs_fs_info *fs_info,
8006                             struct btrfs_space_info *info, u64 bytes,
8007                             int dump_block_groups)
8008 {
8009         struct btrfs_block_group_cache *cache;
8010         int index = 0;
8011
8012         spin_lock(&info->lock);
8013         btrfs_info(fs_info, "space_info %llu has %llu free, is %sfull",
8014                    info->flags,
8015                    info->total_bytes - btrfs_space_info_used(info, true),
8016                    info->full ? "" : "not ");
8017         btrfs_info(fs_info,
8018                 "space_info total=%llu, used=%llu, pinned=%llu, reserved=%llu, may_use=%llu, readonly=%llu",
8019                 info->total_bytes, info->bytes_used, info->bytes_pinned,
8020                 info->bytes_reserved, info->bytes_may_use,
8021                 info->bytes_readonly);
8022         spin_unlock(&info->lock);
8023
8024         DUMP_BLOCK_RSV(fs_info, global_block_rsv);
8025         DUMP_BLOCK_RSV(fs_info, trans_block_rsv);
8026         DUMP_BLOCK_RSV(fs_info, chunk_block_rsv);
8027         DUMP_BLOCK_RSV(fs_info, delayed_block_rsv);
8028         DUMP_BLOCK_RSV(fs_info, delayed_refs_rsv);
8029
8030         if (!dump_block_groups)
8031                 return;
8032
8033         down_read(&info->groups_sem);
8034 again:
8035         list_for_each_entry(cache, &info->block_groups[index], list) {
8036                 spin_lock(&cache->lock);
8037                 btrfs_info(fs_info,
8038                         "block group %llu has %llu bytes, %llu used %llu pinned %llu reserved %s",
8039                         cache->key.objectid, cache->key.offset,
8040                         btrfs_block_group_used(&cache->item), cache->pinned,
8041                         cache->reserved, cache->ro ? "[readonly]" : "");
8042                 btrfs_dump_free_space(cache, bytes);
8043                 spin_unlock(&cache->lock);
8044         }
8045         if (++index < BTRFS_NR_RAID_TYPES)
8046                 goto again;
8047         up_read(&info->groups_sem);
8048 }
8049
8050 /*
8051  * btrfs_reserve_extent - entry point to the extent allocator. Tries to find a
8052  *                        hole that is at least as big as @num_bytes.
8053  *
8054  * @root           -    The root that will contain this extent
8055  *
8056  * @ram_bytes      -    The amount of space in ram that @num_bytes take. This
8057  *                      is used for accounting purposes. This value differs
8058  *                      from @num_bytes only in the case of compressed extents.
8059  *
8060  * @num_bytes      -    Number of bytes to allocate on-disk.
8061  *
8062  * @min_alloc_size -    Indicates the minimum amount of space that the
8063  *                      allocator should try to satisfy. In some cases
8064  *                      @num_bytes may be larger than what is required and if
8065  *                      the filesystem is fragmented then allocation fails.
8066  *                      However, the presence of @min_alloc_size gives a
8067  *                      chance to try and satisfy the smaller allocation.
8068  *
8069  * @empty_size     -    A hint that you plan on doing more COW. This is the
8070  *                      size in bytes the allocator should try to find free
8071  *                      next to the block it returns.  This is just a hint and
8072  *                      may be ignored by the allocator.
8073  *
8074  * @hint_byte      -    Hint to the allocator to start searching above the byte
8075  *                      address passed. It might be ignored.
8076  *
8077  * @ins            -    This key is modified to record the found hole. It will
8078  *                      have the following values:
8079  *                      ins->objectid == start position
8080  *                      ins->flags = BTRFS_EXTENT_ITEM_KEY
8081  *                      ins->offset == the size of the hole.
8082  *
8083  * @is_data        -    Boolean flag indicating whether an extent is
8084  *                      allocated for data (true) or metadata (false)
8085  *
8086  * @delalloc       -    Boolean flag indicating whether this allocation is for
8087  *                      delalloc or not. If 'true' data_rwsem of block groups
8088  *                      is going to be acquired.
8089  *
8090  *
8091  * Returns 0 when an allocation succeeded or < 0 when an error occurred. In
8092  * case -ENOSPC is returned then @ins->offset will contain the size of the
8093  * largest available hole the allocator managed to find.
8094  */
8095 int btrfs_reserve_extent(struct btrfs_root *root, u64 ram_bytes,
8096                          u64 num_bytes, u64 min_alloc_size,
8097                          u64 empty_size, u64 hint_byte,
8098                          struct btrfs_key *ins, int is_data, int delalloc)
8099 {
8100         struct btrfs_fs_info *fs_info = root->fs_info;
8101         bool final_tried = num_bytes == min_alloc_size;
8102         u64 flags;
8103         int ret;
8104
8105         flags = get_alloc_profile_by_root(root, is_data);
8106 again:
8107         WARN_ON(num_bytes < fs_info->sectorsize);
8108         ret = find_free_extent(fs_info, ram_bytes, num_bytes, empty_size,
8109                                hint_byte, ins, flags, delalloc);
8110         if (!ret && !is_data) {
8111                 btrfs_dec_block_group_reservations(fs_info, ins->objectid);
8112         } else if (ret == -ENOSPC) {
8113                 if (!final_tried && ins->offset) {
8114                         num_bytes = min(num_bytes >> 1, ins->offset);
8115                         num_bytes = round_down(num_bytes,
8116                                                fs_info->sectorsize);
8117                         num_bytes = max(num_bytes, min_alloc_size);
8118                         ram_bytes = num_bytes;
8119                         if (num_bytes == min_alloc_size)
8120                                 final_tried = true;
8121                         goto again;
8122                 } else if (btrfs_test_opt(fs_info, ENOSPC_DEBUG)) {
8123                         struct btrfs_space_info *sinfo;
8124
8125                         sinfo = __find_space_info(fs_info, flags);
8126                         btrfs_err(fs_info,
8127                                   "allocation failed flags %llu, wanted %llu",
8128                                   flags, num_bytes);
8129                         if (sinfo)
8130                                 dump_space_info(fs_info, sinfo, num_bytes, 1);
8131                 }
8132         }
8133
8134         return ret;
8135 }
8136
8137 static int __btrfs_free_reserved_extent(struct btrfs_fs_info *fs_info,
8138                                         u64 start, u64 len,
8139                                         int pin, int delalloc)
8140 {
8141         struct btrfs_block_group_cache *cache;
8142         int ret = 0;
8143
8144         cache = btrfs_lookup_block_group(fs_info, start);
8145         if (!cache) {
8146                 btrfs_err(fs_info, "Unable to find block group for %llu",
8147                           start);
8148                 return -ENOSPC;
8149         }
8150
8151         if (pin)
8152                 pin_down_extent(cache, start, len, 1);
8153         else {
8154                 if (btrfs_test_opt(fs_info, DISCARD))
8155                         ret = btrfs_discard_extent(fs_info, start, len, NULL);
8156                 btrfs_add_free_space(cache, start, len);
8157                 btrfs_free_reserved_bytes(cache, len, delalloc);
8158                 trace_btrfs_reserved_extent_free(fs_info, start, len);
8159         }
8160
8161         btrfs_put_block_group(cache);
8162         return ret;
8163 }
8164
8165 int btrfs_free_reserved_extent(struct btrfs_fs_info *fs_info,
8166                                u64 start, u64 len, int delalloc)
8167 {
8168         return __btrfs_free_reserved_extent(fs_info, start, len, 0, delalloc);
8169 }
8170
8171 int btrfs_free_and_pin_reserved_extent(struct btrfs_fs_info *fs_info,
8172                                        u64 start, u64 len)
8173 {
8174         return __btrfs_free_reserved_extent(fs_info, start, len, 1, 0);
8175 }
8176
8177 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
8178                                       u64 parent, u64 root_objectid,
8179                                       u64 flags, u64 owner, u64 offset,
8180                                       struct btrfs_key *ins, int ref_mod)
8181 {
8182         struct btrfs_fs_info *fs_info = trans->fs_info;
8183         int ret;
8184         struct btrfs_extent_item *extent_item;
8185         struct btrfs_extent_inline_ref *iref;
8186         struct btrfs_path *path;
8187         struct extent_buffer *leaf;
8188         int type;
8189         u32 size;
8190
8191         if (parent > 0)
8192                 type = BTRFS_SHARED_DATA_REF_KEY;
8193         else
8194                 type = BTRFS_EXTENT_DATA_REF_KEY;
8195
8196         size = sizeof(*extent_item) + btrfs_extent_inline_ref_size(type);
8197
8198         path = btrfs_alloc_path();
8199         if (!path)
8200                 return -ENOMEM;
8201
8202         path->leave_spinning = 1;
8203         ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path,
8204                                       ins, size);
8205         if (ret) {
8206                 btrfs_free_path(path);
8207                 return ret;
8208         }
8209
8210         leaf = path->nodes[0];
8211         extent_item = btrfs_item_ptr(leaf, path->slots[0],
8212                                      struct btrfs_extent_item);
8213         btrfs_set_extent_refs(leaf, extent_item, ref_mod);
8214         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
8215         btrfs_set_extent_flags(leaf, extent_item,
8216                                flags | BTRFS_EXTENT_FLAG_DATA);
8217
8218         iref = (struct btrfs_extent_inline_ref *)(extent_item + 1);
8219         btrfs_set_extent_inline_ref_type(leaf, iref, type);
8220         if (parent > 0) {
8221                 struct btrfs_shared_data_ref *ref;
8222                 ref = (struct btrfs_shared_data_ref *)(iref + 1);
8223                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
8224                 btrfs_set_shared_data_ref_count(leaf, ref, ref_mod);
8225         } else {
8226                 struct btrfs_extent_data_ref *ref;
8227                 ref = (struct btrfs_extent_data_ref *)(&iref->offset);
8228                 btrfs_set_extent_data_ref_root(leaf, ref, root_objectid);
8229                 btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
8230                 btrfs_set_extent_data_ref_offset(leaf, ref, offset);
8231                 btrfs_set_extent_data_ref_count(leaf, ref, ref_mod);
8232         }
8233
8234         btrfs_mark_buffer_dirty(path->nodes[0]);
8235         btrfs_free_path(path);
8236
8237         ret = remove_from_free_space_tree(trans, ins->objectid, ins->offset);
8238         if (ret)
8239                 return ret;
8240
8241         ret = update_block_group(trans, ins->objectid, ins->offset, 1);
8242         if (ret) { /* -ENOENT, logic error */
8243                 btrfs_err(fs_info, "update block group failed for %llu %llu",
8244                         ins->objectid, ins->offset);
8245                 BUG();
8246         }
8247         trace_btrfs_reserved_extent_alloc(fs_info, ins->objectid, ins->offset);
8248         return ret;
8249 }
8250
8251 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
8252                                      struct btrfs_delayed_ref_node *node,
8253                                      struct btrfs_delayed_extent_op *extent_op)
8254 {
8255         struct btrfs_fs_info *fs_info = trans->fs_info;
8256         int ret;
8257         struct btrfs_extent_item *extent_item;
8258         struct btrfs_key extent_key;
8259         struct btrfs_tree_block_info *block_info;
8260         struct btrfs_extent_inline_ref *iref;
8261         struct btrfs_path *path;
8262         struct extent_buffer *leaf;
8263         struct btrfs_delayed_tree_ref *ref;
8264         u32 size = sizeof(*extent_item) + sizeof(*iref);
8265         u64 num_bytes;
8266         u64 flags = extent_op->flags_to_set;
8267         bool skinny_metadata = btrfs_fs_incompat(fs_info, SKINNY_METADATA);
8268
8269         ref = btrfs_delayed_node_to_tree_ref(node);
8270
8271         extent_key.objectid = node->bytenr;
8272         if (skinny_metadata) {
8273                 extent_key.offset = ref->level;
8274                 extent_key.type = BTRFS_METADATA_ITEM_KEY;
8275                 num_bytes = fs_info->nodesize;
8276         } else {
8277                 extent_key.offset = node->num_bytes;
8278                 extent_key.type = BTRFS_EXTENT_ITEM_KEY;
8279                 size += sizeof(*block_info);
8280                 num_bytes = node->num_bytes;
8281         }
8282
8283         path = btrfs_alloc_path();
8284         if (!path)
8285                 return -ENOMEM;
8286
8287         path->leave_spinning = 1;
8288         ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path,
8289                                       &extent_key, size);
8290         if (ret) {
8291                 btrfs_free_path(path);
8292                 return ret;
8293         }
8294
8295         leaf = path->nodes[0];
8296         extent_item = btrfs_item_ptr(leaf, path->slots[0],
8297                                      struct btrfs_extent_item);
8298         btrfs_set_extent_refs(leaf, extent_item, 1);
8299         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
8300         btrfs_set_extent_flags(leaf, extent_item,
8301                                flags | BTRFS_EXTENT_FLAG_TREE_BLOCK);
8302
8303         if (skinny_metadata) {
8304                 iref = (struct btrfs_extent_inline_ref *)(extent_item + 1);
8305         } else {
8306                 block_info = (struct btrfs_tree_block_info *)(extent_item + 1);
8307                 btrfs_set_tree_block_key(leaf, block_info, &extent_op->key);
8308                 btrfs_set_tree_block_level(leaf, block_info, ref->level);
8309                 iref = (struct btrfs_extent_inline_ref *)(block_info + 1);
8310         }
8311
8312         if (node->type == BTRFS_SHARED_BLOCK_REF_KEY) {
8313                 BUG_ON(!(flags & BTRFS_BLOCK_FLAG_FULL_BACKREF));
8314                 btrfs_set_extent_inline_ref_type(leaf, iref,
8315                                                  BTRFS_SHARED_BLOCK_REF_KEY);
8316                 btrfs_set_extent_inline_ref_offset(leaf, iref, ref->parent);
8317         } else {
8318                 btrfs_set_extent_inline_ref_type(leaf, iref,
8319                                                  BTRFS_TREE_BLOCK_REF_KEY);
8320                 btrfs_set_extent_inline_ref_offset(leaf, iref, ref->root);
8321         }
8322
8323         btrfs_mark_buffer_dirty(leaf);
8324         btrfs_free_path(path);
8325
8326         ret = remove_from_free_space_tree(trans, extent_key.objectid,
8327                                           num_bytes);
8328         if (ret)
8329                 return ret;
8330
8331         ret = update_block_group(trans, extent_key.objectid,
8332                                  fs_info->nodesize, 1);
8333         if (ret) { /* -ENOENT, logic error */
8334                 btrfs_err(fs_info, "update block group failed for %llu %llu",
8335                         extent_key.objectid, extent_key.offset);
8336                 BUG();
8337         }
8338
8339         trace_btrfs_reserved_extent_alloc(fs_info, extent_key.objectid,
8340                                           fs_info->nodesize);
8341         return ret;
8342 }
8343
8344 int btrfs_alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
8345                                      struct btrfs_root *root, u64 owner,
8346                                      u64 offset, u64 ram_bytes,
8347                                      struct btrfs_key *ins)
8348 {
8349         struct btrfs_ref generic_ref = { 0 };
8350         int ret;
8351
8352         BUG_ON(root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID);
8353
8354         btrfs_init_generic_ref(&generic_ref, BTRFS_ADD_DELAYED_EXTENT,
8355                                ins->objectid, ins->offset, 0);
8356         btrfs_init_data_ref(&generic_ref, root->root_key.objectid, owner, offset);
8357         btrfs_ref_tree_mod(root->fs_info, &generic_ref);
8358         ret = btrfs_add_delayed_data_ref(trans, &generic_ref,
8359                                          ram_bytes, NULL, NULL);
8360         return ret;
8361 }
8362
8363 /*
8364  * this is used by the tree logging recovery code.  It records that
8365  * an extent has been allocated and makes sure to clear the free
8366  * space cache bits as well
8367  */
8368 int btrfs_alloc_logged_file_extent(struct btrfs_trans_handle *trans,
8369                                    u64 root_objectid, u64 owner, u64 offset,
8370                                    struct btrfs_key *ins)
8371 {
8372         struct btrfs_fs_info *fs_info = trans->fs_info;
8373         int ret;
8374         struct btrfs_block_group_cache *block_group;
8375         struct btrfs_space_info *space_info;
8376
8377         /*
8378          * Mixed block groups will exclude before processing the log so we only
8379          * need to do the exclude dance if this fs isn't mixed.
8380          */
8381         if (!btrfs_fs_incompat(fs_info, MIXED_GROUPS)) {
8382                 ret = __exclude_logged_extent(fs_info, ins->objectid,
8383                                               ins->offset);
8384                 if (ret)
8385                         return ret;
8386         }
8387
8388         block_group = btrfs_lookup_block_group(fs_info, ins->objectid);
8389         if (!block_group)
8390                 return -EINVAL;
8391
8392         space_info = block_group->space_info;
8393         spin_lock(&space_info->lock);
8394         spin_lock(&block_group->lock);
8395         space_info->bytes_reserved += ins->offset;
8396         block_group->reserved += ins->offset;
8397         spin_unlock(&block_group->lock);
8398         spin_unlock(&space_info->lock);
8399
8400         ret = alloc_reserved_file_extent(trans, 0, root_objectid, 0, owner,
8401                                          offset, ins, 1);
8402         btrfs_put_block_group(block_group);
8403         return ret;
8404 }
8405
8406 static struct extent_buffer *
8407 btrfs_init_new_buffer(struct btrfs_trans_handle *trans, struct btrfs_root *root,
8408                       u64 bytenr, int level, u64 owner)
8409 {
8410         struct btrfs_fs_info *fs_info = root->fs_info;
8411         struct extent_buffer *buf;
8412
8413         buf = btrfs_find_create_tree_block(fs_info, bytenr);
8414         if (IS_ERR(buf))
8415                 return buf;
8416
8417         /*
8418          * Extra safety check in case the extent tree is corrupted and extent
8419          * allocator chooses to use a tree block which is already used and
8420          * locked.
8421          */
8422         if (buf->lock_owner == current->pid) {
8423                 btrfs_err_rl(fs_info,
8424 "tree block %llu owner %llu already locked by pid=%d, extent tree corruption detected",
8425                         buf->start, btrfs_header_owner(buf), current->pid);
8426                 free_extent_buffer(buf);
8427                 return ERR_PTR(-EUCLEAN);
8428         }
8429
8430         btrfs_set_buffer_lockdep_class(root->root_key.objectid, buf, level);
8431         btrfs_tree_lock(buf);
8432         btrfs_clean_tree_block(buf);
8433         clear_bit(EXTENT_BUFFER_STALE, &buf->bflags);
8434
8435         btrfs_set_lock_blocking_write(buf);
8436         set_extent_buffer_uptodate(buf);
8437
8438         memzero_extent_buffer(buf, 0, sizeof(struct btrfs_header));
8439         btrfs_set_header_level(buf, level);
8440         btrfs_set_header_bytenr(buf, buf->start);
8441         btrfs_set_header_generation(buf, trans->transid);
8442         btrfs_set_header_backref_rev(buf, BTRFS_MIXED_BACKREF_REV);
8443         btrfs_set_header_owner(buf, owner);
8444         write_extent_buffer_fsid(buf, fs_info->fs_devices->metadata_uuid);
8445         write_extent_buffer_chunk_tree_uuid(buf, fs_info->chunk_tree_uuid);
8446         if (root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID) {
8447                 buf->log_index = root->log_transid % 2;
8448                 /*
8449                  * we allow two log transactions at a time, use different
8450                  * EXTENT bit to differentiate dirty pages.
8451                  */
8452                 if (buf->log_index == 0)
8453                         set_extent_dirty(&root->dirty_log_pages, buf->start,
8454                                         buf->start + buf->len - 1, GFP_NOFS);
8455                 else
8456                         set_extent_new(&root->dirty_log_pages, buf->start,
8457                                         buf->start + buf->len - 1);
8458         } else {
8459                 buf->log_index = -1;
8460                 set_extent_dirty(&trans->transaction->dirty_pages, buf->start,
8461                          buf->start + buf->len - 1, GFP_NOFS);
8462         }
8463         trans->dirty = true;
8464         /* this returns a buffer locked for blocking */
8465         return buf;
8466 }
8467
8468 static struct btrfs_block_rsv *
8469 use_block_rsv(struct btrfs_trans_handle *trans,
8470               struct btrfs_root *root, u32 blocksize)
8471 {
8472         struct btrfs_fs_info *fs_info = root->fs_info;
8473         struct btrfs_block_rsv *block_rsv;
8474         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
8475         int ret;
8476         bool global_updated = false;
8477
8478         block_rsv = get_block_rsv(trans, root);
8479
8480         if (unlikely(block_rsv->size == 0))
8481                 goto try_reserve;
8482 again:
8483         ret = block_rsv_use_bytes(block_rsv, blocksize);
8484         if (!ret)
8485                 return block_rsv;
8486
8487         if (block_rsv->failfast)
8488                 return ERR_PTR(ret);
8489
8490         if (block_rsv->type == BTRFS_BLOCK_RSV_GLOBAL && !global_updated) {
8491                 global_updated = true;
8492                 update_global_block_rsv(fs_info);
8493                 goto again;
8494         }
8495
8496         /*
8497          * The global reserve still exists to save us from ourselves, so don't
8498          * warn_on if we are short on our delayed refs reserve.
8499          */
8500         if (block_rsv->type != BTRFS_BLOCK_RSV_DELREFS &&
8501             btrfs_test_opt(fs_info, ENOSPC_DEBUG)) {
8502                 static DEFINE_RATELIMIT_STATE(_rs,
8503                                 DEFAULT_RATELIMIT_INTERVAL * 10,
8504                                 /*DEFAULT_RATELIMIT_BURST*/ 1);
8505                 if (__ratelimit(&_rs))
8506                         WARN(1, KERN_DEBUG
8507                                 "BTRFS: block rsv returned %d\n", ret);
8508         }
8509 try_reserve:
8510         ret = reserve_metadata_bytes(root, block_rsv, blocksize,
8511                                      BTRFS_RESERVE_NO_FLUSH);
8512         if (!ret)
8513                 return block_rsv;
8514         /*
8515          * If we couldn't reserve metadata bytes try and use some from
8516          * the global reserve if its space type is the same as the global
8517          * reservation.
8518          */
8519         if (block_rsv->type != BTRFS_BLOCK_RSV_GLOBAL &&
8520             block_rsv->space_info == global_rsv->space_info) {
8521                 ret = block_rsv_use_bytes(global_rsv, blocksize);
8522                 if (!ret)
8523                         return global_rsv;
8524         }
8525         return ERR_PTR(ret);
8526 }
8527
8528 static void unuse_block_rsv(struct btrfs_fs_info *fs_info,
8529                             struct btrfs_block_rsv *block_rsv, u32 blocksize)
8530 {
8531         block_rsv_add_bytes(block_rsv, blocksize, false);
8532         block_rsv_release_bytes(fs_info, block_rsv, NULL, 0, NULL);
8533 }
8534
8535 /*
8536  * finds a free extent and does all the dirty work required for allocation
8537  * returns the tree buffer or an ERR_PTR on error.
8538  */
8539 struct extent_buffer *btrfs_alloc_tree_block(struct btrfs_trans_handle *trans,
8540                                              struct btrfs_root *root,
8541                                              u64 parent, u64 root_objectid,
8542                                              const struct btrfs_disk_key *key,
8543                                              int level, u64 hint,
8544                                              u64 empty_size)
8545 {
8546         struct btrfs_fs_info *fs_info = root->fs_info;
8547         struct btrfs_key ins;
8548         struct btrfs_block_rsv *block_rsv;
8549         struct extent_buffer *buf;
8550         struct btrfs_delayed_extent_op *extent_op;
8551         struct btrfs_ref generic_ref = { 0 };
8552         u64 flags = 0;
8553         int ret;
8554         u32 blocksize = fs_info->nodesize;
8555         bool skinny_metadata = btrfs_fs_incompat(fs_info, SKINNY_METADATA);
8556
8557 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
8558         if (btrfs_is_testing(fs_info)) {
8559                 buf = btrfs_init_new_buffer(trans, root, root->alloc_bytenr,
8560                                             level, root_objectid);
8561                 if (!IS_ERR(buf))
8562                         root->alloc_bytenr += blocksize;
8563                 return buf;
8564         }
8565 #endif
8566
8567         block_rsv = use_block_rsv(trans, root, blocksize);
8568         if (IS_ERR(block_rsv))
8569                 return ERR_CAST(block_rsv);
8570
8571         ret = btrfs_reserve_extent(root, blocksize, blocksize, blocksize,
8572                                    empty_size, hint, &ins, 0, 0);
8573         if (ret)
8574                 goto out_unuse;
8575
8576         buf = btrfs_init_new_buffer(trans, root, ins.objectid, level,
8577                                     root_objectid);
8578         if (IS_ERR(buf)) {
8579                 ret = PTR_ERR(buf);
8580                 goto out_free_reserved;
8581         }
8582
8583         if (root_objectid == BTRFS_TREE_RELOC_OBJECTID) {
8584                 if (parent == 0)
8585                         parent = ins.objectid;
8586                 flags |= BTRFS_BLOCK_FLAG_FULL_BACKREF;
8587         } else
8588                 BUG_ON(parent > 0);
8589
8590         if (root_objectid != BTRFS_TREE_LOG_OBJECTID) {
8591                 extent_op = btrfs_alloc_delayed_extent_op();
8592                 if (!extent_op) {
8593                         ret = -ENOMEM;
8594                         goto out_free_buf;
8595                 }
8596                 if (key)
8597                         memcpy(&extent_op->key, key, sizeof(extent_op->key));
8598                 else
8599                         memset(&extent_op->key, 0, sizeof(extent_op->key));
8600                 extent_op->flags_to_set = flags;
8601                 extent_op->update_key = skinny_metadata ? false : true;
8602                 extent_op->update_flags = true;
8603                 extent_op->is_data = false;
8604                 extent_op->level = level;
8605
8606                 btrfs_init_generic_ref(&generic_ref, BTRFS_ADD_DELAYED_EXTENT,
8607                                        ins.objectid, ins.offset, parent);
8608                 generic_ref.real_root = root->root_key.objectid;
8609                 btrfs_init_tree_ref(&generic_ref, level, root_objectid);
8610                 btrfs_ref_tree_mod(fs_info, &generic_ref);
8611                 ret = btrfs_add_delayed_tree_ref(trans, &generic_ref,
8612                                                  extent_op, NULL, NULL);
8613                 if (ret)
8614                         goto out_free_delayed;
8615         }
8616         return buf;
8617
8618 out_free_delayed:
8619         btrfs_free_delayed_extent_op(extent_op);
8620 out_free_buf:
8621         free_extent_buffer(buf);
8622 out_free_reserved:
8623         btrfs_free_reserved_extent(fs_info, ins.objectid, ins.offset, 0);
8624 out_unuse:
8625         unuse_block_rsv(fs_info, block_rsv, blocksize);
8626         return ERR_PTR(ret);
8627 }
8628
8629 struct walk_control {
8630         u64 refs[BTRFS_MAX_LEVEL];
8631         u64 flags[BTRFS_MAX_LEVEL];
8632         struct btrfs_key update_progress;
8633         struct btrfs_key drop_progress;
8634         int drop_level;
8635         int stage;
8636         int level;
8637         int shared_level;
8638         int update_ref;
8639         int keep_locks;
8640         int reada_slot;
8641         int reada_count;
8642         int restarted;
8643 };
8644
8645 #define DROP_REFERENCE  1
8646 #define UPDATE_BACKREF  2
8647
8648 static noinline void reada_walk_down(struct btrfs_trans_handle *trans,
8649                                      struct btrfs_root *root,
8650                                      struct walk_control *wc,
8651                                      struct btrfs_path *path)
8652 {
8653         struct btrfs_fs_info *fs_info = root->fs_info;
8654         u64 bytenr;
8655         u64 generation;
8656         u64 refs;
8657         u64 flags;
8658         u32 nritems;
8659         struct btrfs_key key;
8660         struct extent_buffer *eb;
8661         int ret;
8662         int slot;
8663         int nread = 0;
8664
8665         if (path->slots[wc->level] < wc->reada_slot) {
8666                 wc->reada_count = wc->reada_count * 2 / 3;
8667                 wc->reada_count = max(wc->reada_count, 2);
8668         } else {
8669                 wc->reada_count = wc->reada_count * 3 / 2;
8670                 wc->reada_count = min_t(int, wc->reada_count,
8671                                         BTRFS_NODEPTRS_PER_BLOCK(fs_info));
8672         }
8673
8674         eb = path->nodes[wc->level];
8675         nritems = btrfs_header_nritems(eb);
8676
8677         for (slot = path->slots[wc->level]; slot < nritems; slot++) {
8678                 if (nread >= wc->reada_count)
8679                         break;
8680
8681                 cond_resched();
8682                 bytenr = btrfs_node_blockptr(eb, slot);
8683                 generation = btrfs_node_ptr_generation(eb, slot);
8684
8685                 if (slot == path->slots[wc->level])
8686                         goto reada;
8687
8688                 if (wc->stage == UPDATE_BACKREF &&
8689                     generation <= root->root_key.offset)
8690                         continue;
8691
8692                 /* We don't lock the tree block, it's OK to be racy here */
8693                 ret = btrfs_lookup_extent_info(trans, fs_info, bytenr,
8694                                                wc->level - 1, 1, &refs,
8695                                                &flags);
8696                 /* We don't care about errors in readahead. */
8697                 if (ret < 0)
8698                         continue;
8699                 BUG_ON(refs == 0);
8700
8701                 if (wc->stage == DROP_REFERENCE) {
8702                         if (refs == 1)
8703                                 goto reada;
8704
8705                         if (wc->level == 1 &&
8706                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
8707                                 continue;
8708                         if (!wc->update_ref ||
8709                             generation <= root->root_key.offset)
8710                                 continue;
8711                         btrfs_node_key_to_cpu(eb, &key, slot);
8712                         ret = btrfs_comp_cpu_keys(&key,
8713                                                   &wc->update_progress);
8714                         if (ret < 0)
8715                                 continue;
8716                 } else {
8717                         if (wc->level == 1 &&
8718                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
8719                                 continue;
8720                 }
8721 reada:
8722                 readahead_tree_block(fs_info, bytenr);
8723                 nread++;
8724         }
8725         wc->reada_slot = slot;
8726 }
8727
8728 /*
8729  * helper to process tree block while walking down the tree.
8730  *
8731  * when wc->stage == UPDATE_BACKREF, this function updates
8732  * back refs for pointers in the block.
8733  *
8734  * NOTE: return value 1 means we should stop walking down.
8735  */
8736 static noinline int walk_down_proc(struct btrfs_trans_handle *trans,
8737                                    struct btrfs_root *root,
8738                                    struct btrfs_path *path,
8739                                    struct walk_control *wc, int lookup_info)
8740 {
8741         struct btrfs_fs_info *fs_info = root->fs_info;
8742         int level = wc->level;
8743         struct extent_buffer *eb = path->nodes[level];
8744         u64 flag = BTRFS_BLOCK_FLAG_FULL_BACKREF;
8745         int ret;
8746
8747         if (wc->stage == UPDATE_BACKREF &&
8748             btrfs_header_owner(eb) != root->root_key.objectid)
8749                 return 1;
8750
8751         /*
8752          * when reference count of tree block is 1, it won't increase
8753          * again. once full backref flag is set, we never clear it.
8754          */
8755         if (lookup_info &&
8756             ((wc->stage == DROP_REFERENCE && wc->refs[level] != 1) ||
8757              (wc->stage == UPDATE_BACKREF && !(wc->flags[level] & flag)))) {
8758                 BUG_ON(!path->locks[level]);
8759                 ret = btrfs_lookup_extent_info(trans, fs_info,
8760                                                eb->start, level, 1,
8761                                                &wc->refs[level],
8762                                                &wc->flags[level]);
8763                 BUG_ON(ret == -ENOMEM);
8764                 if (ret)
8765                         return ret;
8766                 BUG_ON(wc->refs[level] == 0);
8767         }
8768
8769         if (wc->stage == DROP_REFERENCE) {
8770                 if (wc->refs[level] > 1)
8771                         return 1;
8772
8773                 if (path->locks[level] && !wc->keep_locks) {
8774                         btrfs_tree_unlock_rw(eb, path->locks[level]);
8775                         path->locks[level] = 0;
8776                 }
8777                 return 0;
8778         }
8779
8780         /* wc->stage == UPDATE_BACKREF */
8781         if (!(wc->flags[level] & flag)) {
8782                 BUG_ON(!path->locks[level]);
8783                 ret = btrfs_inc_ref(trans, root, eb, 1);
8784                 BUG_ON(ret); /* -ENOMEM */
8785                 ret = btrfs_dec_ref(trans, root, eb, 0);
8786                 BUG_ON(ret); /* -ENOMEM */
8787                 ret = btrfs_set_disk_extent_flags(trans, eb->start,
8788                                                   eb->len, flag,
8789                                                   btrfs_header_level(eb), 0);
8790                 BUG_ON(ret); /* -ENOMEM */
8791                 wc->flags[level] |= flag;
8792         }
8793
8794         /*
8795          * the block is shared by multiple trees, so it's not good to
8796          * keep the tree lock
8797          */
8798         if (path->locks[level] && level > 0) {
8799                 btrfs_tree_unlock_rw(eb, path->locks[level]);
8800                 path->locks[level] = 0;
8801         }
8802         return 0;
8803 }
8804
8805 /*
8806  * This is used to verify a ref exists for this root to deal with a bug where we
8807  * would have a drop_progress key that hadn't been updated properly.
8808  */
8809 static int check_ref_exists(struct btrfs_trans_handle *trans,
8810                             struct btrfs_root *root, u64 bytenr, u64 parent,
8811                             int level)
8812 {
8813         struct btrfs_path *path;
8814         struct btrfs_extent_inline_ref *iref;
8815         int ret;
8816
8817         path = btrfs_alloc_path();
8818         if (!path)
8819                 return -ENOMEM;
8820
8821         ret = lookup_extent_backref(trans, path, &iref, bytenr,
8822                                     root->fs_info->nodesize, parent,
8823                                     root->root_key.objectid, level, 0);
8824         btrfs_free_path(path);
8825         if (ret == -ENOENT)
8826                 return 0;
8827         if (ret < 0)
8828                 return ret;
8829         return 1;
8830 }
8831
8832 /*
8833  * helper to process tree block pointer.
8834  *
8835  * when wc->stage == DROP_REFERENCE, this function checks
8836  * reference count of the block pointed to. if the block
8837  * is shared and we need update back refs for the subtree
8838  * rooted at the block, this function changes wc->stage to
8839  * UPDATE_BACKREF. if the block is shared and there is no
8840  * need to update back, this function drops the reference
8841  * to the block.
8842  *
8843  * NOTE: return value 1 means we should stop walking down.
8844  */
8845 static noinline int do_walk_down(struct btrfs_trans_handle *trans,
8846                                  struct btrfs_root *root,
8847                                  struct btrfs_path *path,
8848                                  struct walk_control *wc, int *lookup_info)
8849 {
8850         struct btrfs_fs_info *fs_info = root->fs_info;
8851         u64 bytenr;
8852         u64 generation;
8853         u64 parent;
8854         struct btrfs_key key;
8855         struct btrfs_key first_key;
8856         struct btrfs_ref ref = { 0 };
8857         struct extent_buffer *next;
8858         int level = wc->level;
8859         int reada = 0;
8860         int ret = 0;
8861         bool need_account = false;
8862
8863         generation = btrfs_node_ptr_generation(path->nodes[level],
8864                                                path->slots[level]);
8865         /*
8866          * if the lower level block was created before the snapshot
8867          * was created, we know there is no need to update back refs
8868          * for the subtree
8869          */
8870         if (wc->stage == UPDATE_BACKREF &&
8871             generation <= root->root_key.offset) {
8872                 *lookup_info = 1;
8873                 return 1;
8874         }
8875
8876         bytenr = btrfs_node_blockptr(path->nodes[level], path->slots[level]);
8877         btrfs_node_key_to_cpu(path->nodes[level], &first_key,
8878                               path->slots[level]);
8879
8880         next = find_extent_buffer(fs_info, bytenr);
8881         if (!next) {
8882                 next = btrfs_find_create_tree_block(fs_info, bytenr);
8883                 if (IS_ERR(next))
8884                         return PTR_ERR(next);
8885
8886                 btrfs_set_buffer_lockdep_class(root->root_key.objectid, next,
8887                                                level - 1);
8888                 reada = 1;
8889         }
8890         btrfs_tree_lock(next);
8891         btrfs_set_lock_blocking_write(next);
8892
8893         ret = btrfs_lookup_extent_info(trans, fs_info, bytenr, level - 1, 1,
8894                                        &wc->refs[level - 1],
8895                                        &wc->flags[level - 1]);
8896         if (ret < 0)
8897                 goto out_unlock;
8898
8899         if (unlikely(wc->refs[level - 1] == 0)) {
8900                 btrfs_err(fs_info, "Missing references.");
8901                 ret = -EIO;
8902                 goto out_unlock;
8903         }
8904         *lookup_info = 0;
8905
8906         if (wc->stage == DROP_REFERENCE) {
8907                 if (wc->refs[level - 1] > 1) {
8908                         need_account = true;
8909                         if (level == 1 &&
8910                             (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
8911                                 goto skip;
8912
8913                         if (!wc->update_ref ||
8914                             generation <= root->root_key.offset)
8915                                 goto skip;
8916
8917                         btrfs_node_key_to_cpu(path->nodes[level], &key,
8918                                               path->slots[level]);
8919                         ret = btrfs_comp_cpu_keys(&key, &wc->update_progress);
8920                         if (ret < 0)
8921                                 goto skip;
8922
8923                         wc->stage = UPDATE_BACKREF;
8924                         wc->shared_level = level - 1;
8925                 }
8926         } else {
8927                 if (level == 1 &&
8928                     (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
8929                         goto skip;
8930         }
8931
8932         if (!btrfs_buffer_uptodate(next, generation, 0)) {
8933                 btrfs_tree_unlock(next);
8934                 free_extent_buffer(next);
8935                 next = NULL;
8936                 *lookup_info = 1;
8937         }
8938
8939         if (!next) {
8940                 if (reada && level == 1)
8941                         reada_walk_down(trans, root, wc, path);
8942                 next = read_tree_block(fs_info, bytenr, generation, level - 1,
8943                                        &first_key);
8944                 if (IS_ERR(next)) {
8945                         return PTR_ERR(next);
8946                 } else if (!extent_buffer_uptodate(next)) {
8947                         free_extent_buffer(next);
8948                         return -EIO;
8949                 }
8950                 btrfs_tree_lock(next);
8951                 btrfs_set_lock_blocking_write(next);
8952         }
8953
8954         level--;
8955         ASSERT(level == btrfs_header_level(next));
8956         if (level != btrfs_header_level(next)) {
8957                 btrfs_err(root->fs_info, "mismatched level");
8958                 ret = -EIO;
8959                 goto out_unlock;
8960         }
8961         path->nodes[level] = next;
8962         path->slots[level] = 0;
8963         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
8964         wc->level = level;
8965         if (wc->level == 1)
8966                 wc->reada_slot = 0;
8967         return 0;
8968 skip:
8969         wc->refs[level - 1] = 0;
8970         wc->flags[level - 1] = 0;
8971         if (wc->stage == DROP_REFERENCE) {
8972                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF) {
8973                         parent = path->nodes[level]->start;
8974                 } else {
8975                         ASSERT(root->root_key.objectid ==
8976                                btrfs_header_owner(path->nodes[level]));
8977                         if (root->root_key.objectid !=
8978                             btrfs_header_owner(path->nodes[level])) {
8979                                 btrfs_err(root->fs_info,
8980                                                 "mismatched block owner");
8981                                 ret = -EIO;
8982                                 goto out_unlock;
8983                         }
8984                         parent = 0;
8985                 }
8986
8987                 /*
8988                  * If we had a drop_progress we need to verify the refs are set
8989                  * as expected.  If we find our ref then we know that from here
8990                  * on out everything should be correct, and we can clear the
8991                  * ->restarted flag.
8992                  */
8993                 if (wc->restarted) {
8994                         ret = check_ref_exists(trans, root, bytenr, parent,
8995                                                level - 1);
8996                         if (ret < 0)
8997                                 goto out_unlock;
8998                         if (ret == 0)
8999                                 goto no_delete;
9000                         ret = 0;
9001                         wc->restarted = 0;
9002                 }
9003
9004                 /*
9005                  * Reloc tree doesn't contribute to qgroup numbers, and we have
9006                  * already accounted them at merge time (replace_path),
9007                  * thus we could skip expensive subtree trace here.
9008                  */
9009                 if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID &&
9010                     need_account) {
9011                         ret = btrfs_qgroup_trace_subtree(trans, next,
9012                                                          generation, level - 1);
9013                         if (ret) {
9014                                 btrfs_err_rl(fs_info,
9015                                              "Error %d accounting shared subtree. Quota is out of sync, rescan required.",
9016                                              ret);
9017                         }
9018                 }
9019
9020                 /*
9021                  * We need to update the next key in our walk control so we can
9022                  * update the drop_progress key accordingly.  We don't care if
9023                  * find_next_key doesn't find a key because that means we're at
9024                  * the end and are going to clean up now.
9025                  */
9026                 wc->drop_level = level;
9027                 find_next_key(path, level, &wc->drop_progress);
9028
9029                 btrfs_init_generic_ref(&ref, BTRFS_DROP_DELAYED_REF, bytenr,
9030                                        fs_info->nodesize, parent);
9031                 btrfs_init_tree_ref(&ref, level - 1, root->root_key.objectid);
9032                 ret = btrfs_free_extent(trans, &ref);
9033                 if (ret)
9034                         goto out_unlock;
9035         }
9036 no_delete:
9037         *lookup_info = 1;
9038         ret = 1;
9039
9040 out_unlock:
9041         btrfs_tree_unlock(next);
9042         free_extent_buffer(next);
9043
9044         return ret;
9045 }
9046
9047 /*
9048  * helper to process tree block while walking up the tree.
9049  *
9050  * when wc->stage == DROP_REFERENCE, this function drops
9051  * reference count on the block.
9052  *
9053  * when wc->stage == UPDATE_BACKREF, this function changes
9054  * wc->stage back to DROP_REFERENCE if we changed wc->stage
9055  * to UPDATE_BACKREF previously while processing the block.
9056  *
9057  * NOTE: return value 1 means we should stop walking up.
9058  */
9059 static noinline int walk_up_proc(struct btrfs_trans_handle *trans,
9060                                  struct btrfs_root *root,
9061                                  struct btrfs_path *path,
9062                                  struct walk_control *wc)
9063 {
9064         struct btrfs_fs_info *fs_info = root->fs_info;
9065         int ret;
9066         int level = wc->level;
9067         struct extent_buffer *eb = path->nodes[level];
9068         u64 parent = 0;
9069
9070         if (wc->stage == UPDATE_BACKREF) {
9071                 BUG_ON(wc->shared_level < level);
9072                 if (level < wc->shared_level)
9073                         goto out;
9074
9075                 ret = find_next_key(path, level + 1, &wc->update_progress);
9076                 if (ret > 0)
9077                         wc->update_ref = 0;
9078
9079                 wc->stage = DROP_REFERENCE;
9080                 wc->shared_level = -1;
9081                 path->slots[level] = 0;
9082
9083                 /*
9084                  * check reference count again if the block isn't locked.
9085                  * we should start walking down the tree again if reference
9086                  * count is one.
9087                  */
9088                 if (!path->locks[level]) {
9089                         BUG_ON(level == 0);
9090                         btrfs_tree_lock(eb);
9091                         btrfs_set_lock_blocking_write(eb);
9092                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
9093
9094                         ret = btrfs_lookup_extent_info(trans, fs_info,
9095                                                        eb->start, level, 1,
9096                                                        &wc->refs[level],
9097                                                        &wc->flags[level]);
9098                         if (ret < 0) {
9099                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
9100                                 path->locks[level] = 0;
9101                                 return ret;
9102                         }
9103                         BUG_ON(wc->refs[level] == 0);
9104                         if (wc->refs[level] == 1) {
9105                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
9106                                 path->locks[level] = 0;
9107                                 return 1;
9108                         }
9109                 }
9110         }
9111
9112         /* wc->stage == DROP_REFERENCE */
9113         BUG_ON(wc->refs[level] > 1 && !path->locks[level]);
9114
9115         if (wc->refs[level] == 1) {
9116                 if (level == 0) {
9117                         if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
9118                                 ret = btrfs_dec_ref(trans, root, eb, 1);
9119                         else
9120                                 ret = btrfs_dec_ref(trans, root, eb, 0);
9121                         BUG_ON(ret); /* -ENOMEM */
9122                         if (is_fstree(root->root_key.objectid)) {
9123                                 ret = btrfs_qgroup_trace_leaf_items(trans, eb);
9124                                 if (ret) {
9125                                         btrfs_err_rl(fs_info,
9126         "error %d accounting leaf items, quota is out of sync, rescan required",
9127                                              ret);
9128                                 }
9129                         }
9130                 }
9131                 /* make block locked assertion in btrfs_clean_tree_block happy */
9132                 if (!path->locks[level] &&
9133                     btrfs_header_generation(eb) == trans->transid) {
9134                         btrfs_tree_lock(eb);
9135                         btrfs_set_lock_blocking_write(eb);
9136                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
9137                 }
9138                 btrfs_clean_tree_block(eb);
9139         }
9140
9141         if (eb == root->node) {
9142                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
9143                         parent = eb->start;
9144                 else if (root->root_key.objectid != btrfs_header_owner(eb))
9145                         goto owner_mismatch;
9146         } else {
9147                 if (wc->flags[level + 1] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
9148                         parent = path->nodes[level + 1]->start;
9149                 else if (root->root_key.objectid !=
9150                          btrfs_header_owner(path->nodes[level + 1]))
9151                         goto owner_mismatch;
9152         }
9153
9154         btrfs_free_tree_block(trans, root, eb, parent, wc->refs[level] == 1);
9155 out:
9156         wc->refs[level] = 0;
9157         wc->flags[level] = 0;
9158         return 0;
9159
9160 owner_mismatch:
9161         btrfs_err_rl(fs_info, "unexpected tree owner, have %llu expect %llu",
9162                      btrfs_header_owner(eb), root->root_key.objectid);
9163         return -EUCLEAN;
9164 }
9165
9166 static noinline int walk_down_tree(struct btrfs_trans_handle *trans,
9167                                    struct btrfs_root *root,
9168                                    struct btrfs_path *path,
9169                                    struct walk_control *wc)
9170 {
9171         int level = wc->level;
9172         int lookup_info = 1;
9173         int ret;
9174
9175         while (level >= 0) {
9176                 ret = walk_down_proc(trans, root, path, wc, lookup_info);
9177                 if (ret > 0)
9178                         break;
9179
9180                 if (level == 0)
9181                         break;
9182
9183                 if (path->slots[level] >=
9184                     btrfs_header_nritems(path->nodes[level]))
9185                         break;
9186
9187                 ret = do_walk_down(trans, root, path, wc, &lookup_info);
9188                 if (ret > 0) {
9189                         path->slots[level]++;
9190                         continue;
9191                 } else if (ret < 0)
9192                         return ret;
9193                 level = wc->level;
9194         }
9195         return 0;
9196 }
9197
9198 static noinline int walk_up_tree(struct btrfs_trans_handle *trans,
9199                                  struct btrfs_root *root,
9200                                  struct btrfs_path *path,
9201                                  struct walk_control *wc, int max_level)
9202 {
9203         int level = wc->level;
9204         int ret;
9205
9206         path->slots[level] = btrfs_header_nritems(path->nodes[level]);
9207         while (level < max_level && path->nodes[level]) {
9208                 wc->level = level;
9209                 if (path->slots[level] + 1 <
9210                     btrfs_header_nritems(path->nodes[level])) {
9211                         path->slots[level]++;
9212                         return 0;
9213                 } else {
9214                         ret = walk_up_proc(trans, root, path, wc);
9215                         if (ret > 0)
9216                                 return 0;
9217                         if (ret < 0)
9218                                 return ret;
9219
9220                         if (path->locks[level]) {
9221                                 btrfs_tree_unlock_rw(path->nodes[level],
9222                                                      path->locks[level]);
9223                                 path->locks[level] = 0;
9224                         }
9225                         free_extent_buffer(path->nodes[level]);
9226                         path->nodes[level] = NULL;
9227                         level++;
9228                 }
9229         }
9230         return 1;
9231 }
9232
9233 /*
9234  * drop a subvolume tree.
9235  *
9236  * this function traverses the tree freeing any blocks that only
9237  * referenced by the tree.
9238  *
9239  * when a shared tree block is found. this function decreases its
9240  * reference count by one. if update_ref is true, this function
9241  * also make sure backrefs for the shared block and all lower level
9242  * blocks are properly updated.
9243  *
9244  * If called with for_reloc == 0, may exit early with -EAGAIN
9245  */
9246 int btrfs_drop_snapshot(struct btrfs_root *root,
9247                          struct btrfs_block_rsv *block_rsv, int update_ref,
9248                          int for_reloc)
9249 {
9250         struct btrfs_fs_info *fs_info = root->fs_info;
9251         struct btrfs_path *path;
9252         struct btrfs_trans_handle *trans;
9253         struct btrfs_root *tree_root = fs_info->tree_root;
9254         struct btrfs_root_item *root_item = &root->root_item;
9255         struct walk_control *wc;
9256         struct btrfs_key key;
9257         int err = 0;
9258         int ret;
9259         int level;
9260         bool root_dropped = false;
9261
9262         btrfs_debug(fs_info, "Drop subvolume %llu", root->root_key.objectid);
9263
9264         path = btrfs_alloc_path();
9265         if (!path) {
9266                 err = -ENOMEM;
9267                 goto out;
9268         }
9269
9270         wc = kzalloc(sizeof(*wc), GFP_NOFS);
9271         if (!wc) {
9272                 btrfs_free_path(path);
9273                 err = -ENOMEM;
9274                 goto out;
9275         }
9276
9277         trans = btrfs_start_transaction(tree_root, 0);
9278         if (IS_ERR(trans)) {
9279                 err = PTR_ERR(trans);
9280                 goto out_free;
9281         }
9282
9283         err = btrfs_run_delayed_items(trans);
9284         if (err)
9285                 goto out_end_trans;
9286
9287         if (block_rsv)
9288                 trans->block_rsv = block_rsv;
9289
9290         /*
9291          * This will help us catch people modifying the fs tree while we're
9292          * dropping it.  It is unsafe to mess with the fs tree while it's being
9293          * dropped as we unlock the root node and parent nodes as we walk down
9294          * the tree, assuming nothing will change.  If something does change
9295          * then we'll have stale information and drop references to blocks we've
9296          * already dropped.
9297          */
9298         set_bit(BTRFS_ROOT_DELETING, &root->state);
9299         if (btrfs_disk_key_objectid(&root_item->drop_progress) == 0) {
9300                 level = btrfs_header_level(root->node);
9301                 path->nodes[level] = btrfs_lock_root_node(root);
9302                 btrfs_set_lock_blocking_write(path->nodes[level]);
9303                 path->slots[level] = 0;
9304                 path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
9305                 memset(&wc->update_progress, 0,
9306                        sizeof(wc->update_progress));
9307         } else {
9308                 btrfs_disk_key_to_cpu(&key, &root_item->drop_progress);
9309                 memcpy(&wc->update_progress, &key,
9310                        sizeof(wc->update_progress));
9311
9312                 level = root_item->drop_level;
9313                 BUG_ON(level == 0);
9314                 path->lowest_level = level;
9315                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
9316                 path->lowest_level = 0;
9317                 if (ret < 0) {
9318                         err = ret;
9319                         goto out_end_trans;
9320                 }
9321                 WARN_ON(ret > 0);
9322
9323                 /*
9324                  * unlock our path, this is safe because only this
9325                  * function is allowed to delete this snapshot
9326                  */
9327                 btrfs_unlock_up_safe(path, 0);
9328
9329                 level = btrfs_header_level(root->node);
9330                 while (1) {
9331                         btrfs_tree_lock(path->nodes[level]);
9332                         btrfs_set_lock_blocking_write(path->nodes[level]);
9333                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
9334
9335                         ret = btrfs_lookup_extent_info(trans, fs_info,
9336                                                 path->nodes[level]->start,
9337                                                 level, 1, &wc->refs[level],
9338                                                 &wc->flags[level]);
9339                         if (ret < 0) {
9340                                 err = ret;
9341                                 goto out_end_trans;
9342                         }
9343                         BUG_ON(wc->refs[level] == 0);
9344
9345                         if (level == root_item->drop_level)
9346                                 break;
9347
9348                         btrfs_tree_unlock(path->nodes[level]);
9349                         path->locks[level] = 0;
9350                         WARN_ON(wc->refs[level] != 1);
9351                         level--;
9352                 }
9353         }
9354
9355         wc->restarted = test_bit(BTRFS_ROOT_DEAD_TREE, &root->state);
9356         wc->level = level;
9357         wc->shared_level = -1;
9358         wc->stage = DROP_REFERENCE;
9359         wc->update_ref = update_ref;
9360         wc->keep_locks = 0;
9361         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(fs_info);
9362
9363         while (1) {
9364
9365                 ret = walk_down_tree(trans, root, path, wc);
9366                 if (ret < 0) {
9367                         err = ret;
9368                         break;
9369                 }
9370
9371                 ret = walk_up_tree(trans, root, path, wc, BTRFS_MAX_LEVEL);
9372                 if (ret < 0) {
9373                         err = ret;
9374                         break;
9375                 }
9376
9377                 if (ret > 0) {
9378                         BUG_ON(wc->stage != DROP_REFERENCE);
9379                         break;
9380                 }
9381
9382                 if (wc->stage == DROP_REFERENCE) {
9383                         wc->drop_level = wc->level;
9384                         btrfs_node_key_to_cpu(path->nodes[wc->drop_level],
9385                                               &wc->drop_progress,
9386                                               path->slots[wc->drop_level]);
9387                 }
9388                 btrfs_cpu_key_to_disk(&root_item->drop_progress,
9389                                       &wc->drop_progress);
9390                 root_item->drop_level = wc->drop_level;
9391
9392                 BUG_ON(wc->level == 0);
9393                 if (btrfs_should_end_transaction(trans) ||
9394                     (!for_reloc && btrfs_need_cleaner_sleep(fs_info))) {
9395                         ret = btrfs_update_root(trans, tree_root,
9396                                                 &root->root_key,
9397                                                 root_item);
9398                         if (ret) {
9399                                 btrfs_abort_transaction(trans, ret);
9400                                 err = ret;
9401                                 goto out_end_trans;
9402                         }
9403
9404                         btrfs_end_transaction_throttle(trans);
9405                         if (!for_reloc && btrfs_need_cleaner_sleep(fs_info)) {
9406                                 btrfs_debug(fs_info,
9407                                             "drop snapshot early exit");
9408                                 err = -EAGAIN;
9409                                 goto out_free;
9410                         }
9411
9412                         trans = btrfs_start_transaction(tree_root, 0);
9413                         if (IS_ERR(trans)) {
9414                                 err = PTR_ERR(trans);
9415                                 goto out_free;
9416                         }
9417                         if (block_rsv)
9418                                 trans->block_rsv = block_rsv;
9419                 }
9420         }
9421         btrfs_release_path(path);
9422         if (err)
9423                 goto out_end_trans;
9424
9425         ret = btrfs_del_root(trans, &root->root_key);
9426         if (ret) {
9427                 btrfs_abort_transaction(trans, ret);
9428                 err = ret;
9429                 goto out_end_trans;
9430         }
9431
9432         if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) {
9433                 ret = btrfs_find_root(tree_root, &root->root_key, path,
9434                                       NULL, NULL);
9435                 if (ret < 0) {
9436                         btrfs_abort_transaction(trans, ret);
9437                         err = ret;
9438                         goto out_end_trans;
9439                 } else if (ret > 0) {
9440                         /* if we fail to delete the orphan item this time
9441                          * around, it'll get picked up the next time.
9442                          *
9443                          * The most common failure here is just -ENOENT.
9444                          */
9445                         btrfs_del_orphan_item(trans, tree_root,
9446                                               root->root_key.objectid);
9447                 }
9448         }
9449
9450         if (test_bit(BTRFS_ROOT_IN_RADIX, &root->state)) {
9451                 btrfs_add_dropped_root(trans, root);
9452         } else {
9453                 free_extent_buffer(root->node);
9454                 free_extent_buffer(root->commit_root);
9455                 btrfs_put_fs_root(root);
9456         }
9457         root_dropped = true;
9458 out_end_trans:
9459         btrfs_end_transaction_throttle(trans);
9460 out_free:
9461         kfree(wc);
9462         btrfs_free_path(path);
9463 out:
9464         /*
9465          * So if we need to stop dropping the snapshot for whatever reason we
9466          * need to make sure to add it back to the dead root list so that we
9467          * keep trying to do the work later.  This also cleans up roots if we
9468          * don't have it in the radix (like when we recover after a power fail
9469          * or unmount) so we don't leak memory.
9470          */
9471         if (!for_reloc && !root_dropped)
9472                 btrfs_add_dead_root(root);
9473         if (err && err != -EAGAIN)
9474                 btrfs_handle_fs_error(fs_info, err, NULL);
9475         return err;
9476 }
9477
9478 /*
9479  * drop subtree rooted at tree block 'node'.
9480  *
9481  * NOTE: this function will unlock and release tree block 'node'
9482  * only used by relocation code
9483  */
9484 int btrfs_drop_subtree(struct btrfs_trans_handle *trans,
9485                         struct btrfs_root *root,
9486                         struct extent_buffer *node,
9487                         struct extent_buffer *parent)
9488 {
9489         struct btrfs_fs_info *fs_info = root->fs_info;
9490         struct btrfs_path *path;
9491         struct walk_control *wc;
9492         int level;
9493         int parent_level;
9494         int ret = 0;
9495         int wret;
9496
9497         BUG_ON(root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID);
9498
9499         path = btrfs_alloc_path();
9500         if (!path)
9501                 return -ENOMEM;
9502
9503         wc = kzalloc(sizeof(*wc), GFP_NOFS);
9504         if (!wc) {
9505                 btrfs_free_path(path);
9506                 return -ENOMEM;
9507         }
9508
9509         btrfs_assert_tree_locked(parent);
9510         parent_level = btrfs_header_level(parent);
9511         extent_buffer_get(parent);
9512         path->nodes[parent_level] = parent;
9513         path->slots[parent_level] = btrfs_header_nritems(parent);
9514
9515         btrfs_assert_tree_locked(node);
9516         level = btrfs_header_level(node);
9517         path->nodes[level] = node;
9518         path->slots[level] = 0;
9519         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
9520
9521         wc->refs[parent_level] = 1;
9522         wc->flags[parent_level] = BTRFS_BLOCK_FLAG_FULL_BACKREF;
9523         wc->level = level;
9524         wc->shared_level = -1;
9525         wc->stage = DROP_REFERENCE;
9526         wc->update_ref = 0;
9527         wc->keep_locks = 1;
9528         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(fs_info);
9529
9530         while (1) {
9531                 wret = walk_down_tree(trans, root, path, wc);
9532                 if (wret < 0) {
9533                         ret = wret;
9534                         break;
9535                 }
9536
9537                 wret = walk_up_tree(trans, root, path, wc, parent_level);
9538                 if (wret < 0)
9539                         ret = wret;
9540                 if (wret != 0)
9541                         break;
9542         }
9543
9544         kfree(wc);
9545         btrfs_free_path(path);
9546         return ret;
9547 }
9548
9549 static u64 update_block_group_flags(struct btrfs_fs_info *fs_info, u64 flags)
9550 {
9551         u64 num_devices;
9552         u64 stripped;
9553
9554         /*
9555          * if restripe for this chunk_type is on pick target profile and
9556          * return, otherwise do the usual balance
9557          */
9558         stripped = get_restripe_target(fs_info, flags);
9559         if (stripped)
9560                 return extended_to_chunk(stripped);
9561
9562         num_devices = fs_info->fs_devices->rw_devices;
9563
9564         stripped = BTRFS_BLOCK_GROUP_RAID0 | BTRFS_BLOCK_GROUP_RAID56_MASK |
9565                 BTRFS_BLOCK_GROUP_RAID1_MASK | BTRFS_BLOCK_GROUP_RAID10;
9566
9567         if (num_devices == 1) {
9568                 stripped |= BTRFS_BLOCK_GROUP_DUP;
9569                 stripped = flags & ~stripped;
9570
9571                 /* turn raid0 into single device chunks */
9572                 if (flags & BTRFS_BLOCK_GROUP_RAID0)
9573                         return stripped;
9574
9575                 /* turn mirroring into duplication */
9576                 if (flags & (BTRFS_BLOCK_GROUP_RAID1_MASK |
9577                              BTRFS_BLOCK_GROUP_RAID10))
9578                         return stripped | BTRFS_BLOCK_GROUP_DUP;
9579         } else {
9580                 /* they already had raid on here, just return */
9581                 if (flags & stripped)
9582                         return flags;
9583
9584                 stripped |= BTRFS_BLOCK_GROUP_DUP;
9585                 stripped = flags & ~stripped;
9586
9587                 /* switch duplicated blocks with raid1 */
9588                 if (flags & BTRFS_BLOCK_GROUP_DUP)
9589                         return stripped | BTRFS_BLOCK_GROUP_RAID1;
9590
9591                 /* this is drive concat, leave it alone */
9592         }
9593
9594         return flags;
9595 }
9596
9597 static int inc_block_group_ro(struct btrfs_block_group_cache *cache, int force)
9598 {
9599         struct btrfs_space_info *sinfo = cache->space_info;
9600         u64 num_bytes;
9601         u64 sinfo_used;
9602         u64 min_allocable_bytes;
9603         int ret = -ENOSPC;
9604
9605         /*
9606          * We need some metadata space and system metadata space for
9607          * allocating chunks in some corner cases until we force to set
9608          * it to be readonly.
9609          */
9610         if ((sinfo->flags &
9611              (BTRFS_BLOCK_GROUP_SYSTEM | BTRFS_BLOCK_GROUP_METADATA)) &&
9612             !force)
9613                 min_allocable_bytes = SZ_1M;
9614         else
9615                 min_allocable_bytes = 0;
9616
9617         spin_lock(&sinfo->lock);
9618         spin_lock(&cache->lock);
9619
9620         if (cache->ro) {
9621                 cache->ro++;
9622                 ret = 0;
9623                 goto out;
9624         }
9625
9626         num_bytes = cache->key.offset - cache->reserved - cache->pinned -
9627                     cache->bytes_super - btrfs_block_group_used(&cache->item);
9628         sinfo_used = btrfs_space_info_used(sinfo, true);
9629
9630         if (sinfo_used + num_bytes + min_allocable_bytes <=
9631             sinfo->total_bytes) {
9632                 sinfo->bytes_readonly += num_bytes;
9633                 cache->ro++;
9634                 list_add_tail(&cache->ro_list, &sinfo->ro_bgs);
9635                 ret = 0;
9636         }
9637 out:
9638         spin_unlock(&cache->lock);
9639         spin_unlock(&sinfo->lock);
9640         if (ret == -ENOSPC && btrfs_test_opt(cache->fs_info, ENOSPC_DEBUG)) {
9641                 btrfs_info(cache->fs_info,
9642                         "unable to make block group %llu ro",
9643                         cache->key.objectid);
9644                 btrfs_info(cache->fs_info,
9645                         "sinfo_used=%llu bg_num_bytes=%llu min_allocable=%llu",
9646                         sinfo_used, num_bytes, min_allocable_bytes);
9647                 dump_space_info(cache->fs_info, cache->space_info, 0, 0);
9648         }
9649         return ret;
9650 }
9651
9652 int btrfs_inc_block_group_ro(struct btrfs_block_group_cache *cache)
9653
9654 {
9655         struct btrfs_fs_info *fs_info = cache->fs_info;
9656         struct btrfs_trans_handle *trans;
9657         u64 alloc_flags;
9658         int ret;
9659
9660 again:
9661         trans = btrfs_join_transaction(fs_info->extent_root);
9662         if (IS_ERR(trans))
9663                 return PTR_ERR(trans);
9664
9665         /*
9666          * we're not allowed to set block groups readonly after the dirty
9667          * block groups cache has started writing.  If it already started,
9668          * back off and let this transaction commit
9669          */
9670         mutex_lock(&fs_info->ro_block_group_mutex);
9671         if (test_bit(BTRFS_TRANS_DIRTY_BG_RUN, &trans->transaction->flags)) {
9672                 u64 transid = trans->transid;
9673
9674                 mutex_unlock(&fs_info->ro_block_group_mutex);
9675                 btrfs_end_transaction(trans);
9676
9677                 ret = btrfs_wait_for_commit(fs_info, transid);
9678                 if (ret)
9679                         return ret;
9680                 goto again;
9681         }
9682
9683         /*
9684          * if we are changing raid levels, try to allocate a corresponding
9685          * block group with the new raid level.
9686          */
9687         alloc_flags = update_block_group_flags(fs_info, cache->flags);
9688         if (alloc_flags != cache->flags) {
9689                 ret = do_chunk_alloc(trans, alloc_flags,
9690                                      CHUNK_ALLOC_FORCE);
9691                 /*
9692                  * ENOSPC is allowed here, we may have enough space
9693                  * already allocated at the new raid level to
9694                  * carry on
9695                  */
9696                 if (ret == -ENOSPC)
9697                         ret = 0;
9698                 if (ret < 0)
9699                         goto out;
9700         }
9701
9702         ret = inc_block_group_ro(cache, 0);
9703         if (!ret)
9704                 goto out;
9705         alloc_flags = get_alloc_profile(fs_info, cache->space_info->flags);
9706         ret = do_chunk_alloc(trans, alloc_flags, CHUNK_ALLOC_FORCE);
9707         if (ret < 0)
9708                 goto out;
9709         ret = inc_block_group_ro(cache, 0);
9710 out:
9711         if (cache->flags & BTRFS_BLOCK_GROUP_SYSTEM) {
9712                 alloc_flags = update_block_group_flags(fs_info, cache->flags);
9713                 mutex_lock(&fs_info->chunk_mutex);
9714                 check_system_chunk(trans, alloc_flags);
9715                 mutex_unlock(&fs_info->chunk_mutex);
9716         }
9717         mutex_unlock(&fs_info->ro_block_group_mutex);
9718
9719         btrfs_end_transaction(trans);
9720         return ret;
9721 }
9722
9723 int btrfs_force_chunk_alloc(struct btrfs_trans_handle *trans, u64 type)
9724 {
9725         u64 alloc_flags = get_alloc_profile(trans->fs_info, type);
9726
9727         return do_chunk_alloc(trans, alloc_flags, CHUNK_ALLOC_FORCE);
9728 }
9729
9730 /*
9731  * helper to account the unused space of all the readonly block group in the
9732  * space_info. takes mirrors into account.
9733  */
9734 u64 btrfs_account_ro_block_groups_free_space(struct btrfs_space_info *sinfo)
9735 {
9736         struct btrfs_block_group_cache *block_group;
9737         u64 free_bytes = 0;
9738         int factor;
9739
9740         /* It's df, we don't care if it's racy */
9741         if (list_empty(&sinfo->ro_bgs))
9742                 return 0;
9743
9744         spin_lock(&sinfo->lock);
9745         list_for_each_entry(block_group, &sinfo->ro_bgs, ro_list) {
9746                 spin_lock(&block_group->lock);
9747
9748                 if (!block_group->ro) {
9749                         spin_unlock(&block_group->lock);
9750                         continue;
9751                 }
9752
9753                 factor = btrfs_bg_type_to_factor(block_group->flags);
9754                 free_bytes += (block_group->key.offset -
9755                                btrfs_block_group_used(&block_group->item)) *
9756                                factor;
9757
9758                 spin_unlock(&block_group->lock);
9759         }
9760         spin_unlock(&sinfo->lock);
9761
9762         return free_bytes;
9763 }
9764
9765 void btrfs_dec_block_group_ro(struct btrfs_block_group_cache *cache)
9766 {
9767         struct btrfs_space_info *sinfo = cache->space_info;
9768         u64 num_bytes;
9769
9770         BUG_ON(!cache->ro);
9771
9772         spin_lock(&sinfo->lock);
9773         spin_lock(&cache->lock);
9774         if (!--cache->ro) {
9775                 num_bytes = cache->key.offset - cache->reserved -
9776                             cache->pinned - cache->bytes_super -
9777                             btrfs_block_group_used(&cache->item);
9778                 sinfo->bytes_readonly -= num_bytes;
9779                 list_del_init(&cache->ro_list);
9780         }
9781         spin_unlock(&cache->lock);
9782         spin_unlock(&sinfo->lock);
9783 }
9784
9785 /*
9786  * Checks to see if it's even possible to relocate this block group.
9787  *
9788  * @return - -1 if it's not a good idea to relocate this block group, 0 if its
9789  * ok to go ahead and try.
9790  */
9791 int btrfs_can_relocate(struct btrfs_fs_info *fs_info, u64 bytenr)
9792 {
9793         struct btrfs_block_group_cache *block_group;
9794         struct btrfs_space_info *space_info;
9795         struct btrfs_fs_devices *fs_devices = fs_info->fs_devices;
9796         struct btrfs_device *device;
9797         u64 min_free;
9798         u64 dev_min = 1;
9799         u64 dev_nr = 0;
9800         u64 target;
9801         int debug;
9802         int index;
9803         int full = 0;
9804         int ret = 0;
9805
9806         debug = btrfs_test_opt(fs_info, ENOSPC_DEBUG);
9807
9808         block_group = btrfs_lookup_block_group(fs_info, bytenr);
9809
9810         /* odd, couldn't find the block group, leave it alone */
9811         if (!block_group) {
9812                 if (debug)
9813                         btrfs_warn(fs_info,
9814                                    "can't find block group for bytenr %llu",
9815                                    bytenr);
9816                 return -1;
9817         }
9818
9819         min_free = btrfs_block_group_used(&block_group->item);
9820
9821         /* no bytes used, we're good */
9822         if (!min_free)
9823                 goto out;
9824
9825         space_info = block_group->space_info;
9826         spin_lock(&space_info->lock);
9827
9828         full = space_info->full;
9829
9830         /*
9831          * if this is the last block group we have in this space, we can't
9832          * relocate it unless we're able to allocate a new chunk below.
9833          *
9834          * Otherwise, we need to make sure we have room in the space to handle
9835          * all of the extents from this block group.  If we can, we're good
9836          */
9837         if ((space_info->total_bytes != block_group->key.offset) &&
9838             (btrfs_space_info_used(space_info, false) + min_free <
9839              space_info->total_bytes)) {
9840                 spin_unlock(&space_info->lock);
9841                 goto out;
9842         }
9843         spin_unlock(&space_info->lock);
9844
9845         /*
9846          * ok we don't have enough space, but maybe we have free space on our
9847          * devices to allocate new chunks for relocation, so loop through our
9848          * alloc devices and guess if we have enough space.  if this block
9849          * group is going to be restriped, run checks against the target
9850          * profile instead of the current one.
9851          */
9852         ret = -1;
9853
9854         /*
9855          * index:
9856          *      0: raid10
9857          *      1: raid1
9858          *      2: dup
9859          *      3: raid0
9860          *      4: single
9861          */
9862         target = get_restripe_target(fs_info, block_group->flags);
9863         if (target) {
9864                 index = btrfs_bg_flags_to_raid_index(extended_to_chunk(target));
9865         } else {
9866                 /*
9867                  * this is just a balance, so if we were marked as full
9868                  * we know there is no space for a new chunk
9869                  */
9870                 if (full) {
9871                         if (debug)
9872                                 btrfs_warn(fs_info,
9873                                            "no space to alloc new chunk for block group %llu",
9874                                            block_group->key.objectid);
9875                         goto out;
9876                 }
9877
9878                 index = btrfs_bg_flags_to_raid_index(block_group->flags);
9879         }
9880
9881         if (index == BTRFS_RAID_RAID10) {
9882                 dev_min = 4;
9883                 /* Divide by 2 */
9884                 min_free >>= 1;
9885         } else if (index == BTRFS_RAID_RAID1) {
9886                 dev_min = 2;
9887         } else if (index == BTRFS_RAID_DUP) {
9888                 /* Multiply by 2 */
9889                 min_free <<= 1;
9890         } else if (index == BTRFS_RAID_RAID0) {
9891                 dev_min = fs_devices->rw_devices;
9892                 min_free = div64_u64(min_free, dev_min);
9893         }
9894
9895         mutex_lock(&fs_info->chunk_mutex);
9896         list_for_each_entry(device, &fs_devices->alloc_list, dev_alloc_list) {
9897                 u64 dev_offset;
9898
9899                 /*
9900                  * check to make sure we can actually find a chunk with enough
9901                  * space to fit our block group in.
9902                  */
9903                 if (device->total_bytes > device->bytes_used + min_free &&
9904                     !test_bit(BTRFS_DEV_STATE_REPLACE_TGT, &device->dev_state)) {
9905                         ret = find_free_dev_extent(device, min_free,
9906                                                    &dev_offset, NULL);
9907                         if (!ret)
9908                                 dev_nr++;
9909
9910                         if (dev_nr >= dev_min)
9911                                 break;
9912
9913                         ret = -1;
9914                 }
9915         }
9916         if (debug && ret == -1)
9917                 btrfs_warn(fs_info,
9918                            "no space to allocate a new chunk for block group %llu",
9919                            block_group->key.objectid);
9920         mutex_unlock(&fs_info->chunk_mutex);
9921 out:
9922         btrfs_put_block_group(block_group);
9923         return ret;
9924 }
9925
9926 static int find_first_block_group(struct btrfs_fs_info *fs_info,
9927                                   struct btrfs_path *path,
9928                                   struct btrfs_key *key)
9929 {
9930         struct btrfs_root *root = fs_info->extent_root;
9931         int ret = 0;
9932         struct btrfs_key found_key;
9933         struct extent_buffer *leaf;
9934         struct btrfs_block_group_item bg;
9935         u64 flags;
9936         int slot;
9937
9938         ret = btrfs_search_slot(NULL, root, key, path, 0, 0);
9939         if (ret < 0)
9940                 goto out;
9941
9942         while (1) {
9943                 slot = path->slots[0];
9944                 leaf = path->nodes[0];
9945                 if (slot >= btrfs_header_nritems(leaf)) {
9946                         ret = btrfs_next_leaf(root, path);
9947                         if (ret == 0)
9948                                 continue;
9949                         if (ret < 0)
9950                                 goto out;
9951                         break;
9952                 }
9953                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
9954
9955                 if (found_key.objectid >= key->objectid &&
9956                     found_key.type == BTRFS_BLOCK_GROUP_ITEM_KEY) {
9957                         struct extent_map_tree *em_tree;
9958                         struct extent_map *em;
9959
9960                         em_tree = &root->fs_info->mapping_tree;
9961                         read_lock(&em_tree->lock);
9962                         em = lookup_extent_mapping(em_tree, found_key.objectid,
9963                                                    found_key.offset);
9964                         read_unlock(&em_tree->lock);
9965                         if (!em) {
9966                                 btrfs_err(fs_info,
9967                         "logical %llu len %llu found bg but no related chunk",
9968                                           found_key.objectid, found_key.offset);
9969                                 ret = -ENOENT;
9970                         } else if (em->start != found_key.objectid ||
9971                                    em->len != found_key.offset) {
9972                                 btrfs_err(fs_info,
9973                 "block group %llu len %llu mismatch with chunk %llu len %llu",
9974                                           found_key.objectid, found_key.offset,
9975                                           em->start, em->len);
9976                                 ret = -EUCLEAN;
9977                         } else {
9978                                 read_extent_buffer(leaf, &bg,
9979                                         btrfs_item_ptr_offset(leaf, slot),
9980                                         sizeof(bg));
9981                                 flags = btrfs_block_group_flags(&bg) &
9982                                         BTRFS_BLOCK_GROUP_TYPE_MASK;
9983
9984                                 if (flags != (em->map_lookup->type &
9985                                               BTRFS_BLOCK_GROUP_TYPE_MASK)) {
9986                                         btrfs_err(fs_info,
9987 "block group %llu len %llu type flags 0x%llx mismatch with chunk type flags 0x%llx",
9988                                                 found_key.objectid,
9989                                                 found_key.offset, flags,
9990                                                 (BTRFS_BLOCK_GROUP_TYPE_MASK &
9991                                                  em->map_lookup->type));
9992                                         ret = -EUCLEAN;
9993                                 } else {
9994                                         ret = 0;
9995                                 }
9996                         }
9997                         free_extent_map(em);
9998                         goto out;
9999                 }
10000                 path->slots[0]++;
10001         }
10002 out:
10003         return ret;
10004 }
10005
10006 void btrfs_put_block_group_cache(struct btrfs_fs_info *info)
10007 {
10008         struct btrfs_block_group_cache *block_group;
10009         u64 last = 0;
10010
10011         while (1) {
10012                 struct inode *inode;
10013
10014                 block_group = btrfs_lookup_first_block_group(info, last);
10015                 while (block_group) {
10016                         wait_block_group_cache_done(block_group);
10017                         spin_lock(&block_group->lock);
10018                         if (block_group->iref)
10019                                 break;
10020                         spin_unlock(&block_group->lock);
10021                         block_group = next_block_group(block_group);
10022                 }
10023                 if (!block_group) {
10024                         if (last == 0)
10025                                 break;
10026                         last = 0;
10027                         continue;
10028                 }
10029
10030                 inode = block_group->inode;
10031                 block_group->iref = 0;
10032                 block_group->inode = NULL;
10033                 spin_unlock(&block_group->lock);
10034                 ASSERT(block_group->io_ctl.inode == NULL);
10035                 iput(inode);
10036                 last = block_group->key.objectid + block_group->key.offset;
10037                 btrfs_put_block_group(block_group);
10038         }
10039 }
10040
10041 /*
10042  * Must be called only after stopping all workers, since we could have block
10043  * group caching kthreads running, and therefore they could race with us if we
10044  * freed the block groups before stopping them.
10045  */
10046 int btrfs_free_block_groups(struct btrfs_fs_info *info)
10047 {
10048         struct btrfs_block_group_cache *block_group;
10049         struct btrfs_space_info *space_info;
10050         struct btrfs_caching_control *caching_ctl;
10051         struct rb_node *n;
10052
10053         down_write(&info->commit_root_sem);
10054         while (!list_empty(&info->caching_block_groups)) {
10055                 caching_ctl = list_entry(info->caching_block_groups.next,
10056                                          struct btrfs_caching_control, list);
10057                 list_del(&caching_ctl->list);
10058                 put_caching_control(caching_ctl);
10059         }
10060         up_write(&info->commit_root_sem);
10061
10062         spin_lock(&info->unused_bgs_lock);
10063         while (!list_empty(&info->unused_bgs)) {
10064                 block_group = list_first_entry(&info->unused_bgs,
10065                                                struct btrfs_block_group_cache,
10066                                                bg_list);
10067                 list_del_init(&block_group->bg_list);
10068                 btrfs_put_block_group(block_group);
10069         }
10070         spin_unlock(&info->unused_bgs_lock);
10071
10072         spin_lock(&info->block_group_cache_lock);
10073         while ((n = rb_last(&info->block_group_cache_tree)) != NULL) {
10074                 block_group = rb_entry(n, struct btrfs_block_group_cache,
10075                                        cache_node);
10076                 rb_erase(&block_group->cache_node,
10077                          &info->block_group_cache_tree);
10078                 RB_CLEAR_NODE(&block_group->cache_node);
10079                 spin_unlock(&info->block_group_cache_lock);
10080
10081                 down_write(&block_group->space_info->groups_sem);
10082                 list_del(&block_group->list);
10083                 up_write(&block_group->space_info->groups_sem);
10084
10085                 /*
10086                  * We haven't cached this block group, which means we could
10087                  * possibly have excluded extents on this block group.
10088                  */
10089                 if (block_group->cached == BTRFS_CACHE_NO ||
10090                     block_group->cached == BTRFS_CACHE_ERROR)
10091                         free_excluded_extents(block_group);
10092
10093                 btrfs_remove_free_space_cache(block_group);
10094                 ASSERT(block_group->cached != BTRFS_CACHE_STARTED);
10095                 ASSERT(list_empty(&block_group->dirty_list));
10096                 ASSERT(list_empty(&block_group->io_list));
10097                 ASSERT(list_empty(&block_group->bg_list));
10098                 ASSERT(atomic_read(&block_group->count) == 1);
10099                 btrfs_put_block_group(block_group);
10100
10101                 spin_lock(&info->block_group_cache_lock);
10102         }
10103         spin_unlock(&info->block_group_cache_lock);
10104
10105         /* now that all the block groups are freed, go through and
10106          * free all the space_info structs.  This is only called during
10107          * the final stages of unmount, and so we know nobody is
10108          * using them.  We call synchronize_rcu() once before we start,
10109          * just to be on the safe side.
10110          */
10111         synchronize_rcu();
10112
10113         release_global_block_rsv(info);
10114
10115         while (!list_empty(&info->space_info)) {
10116                 int i;
10117
10118                 space_info = list_entry(info->space_info.next,
10119                                         struct btrfs_space_info,
10120                                         list);
10121
10122                 /*
10123                  * Do not hide this behind enospc_debug, this is actually
10124                  * important and indicates a real bug if this happens.
10125                  */
10126                 if (WARN_ON(space_info->bytes_pinned > 0 ||
10127                             space_info->bytes_reserved > 0 ||
10128                             space_info->bytes_may_use > 0))
10129                         dump_space_info(info, space_info, 0, 0);
10130                 list_del(&space_info->list);
10131                 for (i = 0; i < BTRFS_NR_RAID_TYPES; i++) {
10132                         struct kobject *kobj;
10133                         kobj = space_info->block_group_kobjs[i];
10134                         space_info->block_group_kobjs[i] = NULL;
10135                         if (kobj) {
10136                                 kobject_del(kobj);
10137                                 kobject_put(kobj);
10138                         }
10139                 }
10140                 kobject_del(&space_info->kobj);
10141                 kobject_put(&space_info->kobj);
10142         }
10143         return 0;
10144 }
10145
10146 /* link_block_group will queue up kobjects to add when we're reclaim-safe */
10147 void btrfs_add_raid_kobjects(struct btrfs_fs_info *fs_info)
10148 {
10149         struct btrfs_space_info *space_info;
10150         struct raid_kobject *rkobj;
10151         LIST_HEAD(list);
10152         int ret = 0;
10153
10154         spin_lock(&fs_info->pending_raid_kobjs_lock);
10155         list_splice_init(&fs_info->pending_raid_kobjs, &list);
10156         spin_unlock(&fs_info->pending_raid_kobjs_lock);
10157
10158         list_for_each_entry(rkobj, &list, list) {
10159                 space_info = __find_space_info(fs_info, rkobj->flags);
10160
10161                 ret = kobject_add(&rkobj->kobj, &space_info->kobj,
10162                                 "%s", btrfs_bg_type_to_raid_name(rkobj->flags));
10163                 if (ret) {
10164                         kobject_put(&rkobj->kobj);
10165                         break;
10166                 }
10167         }
10168         if (ret)
10169                 btrfs_warn(fs_info,
10170                            "failed to add kobject for block cache, ignoring");
10171 }
10172
10173 static void link_block_group(struct btrfs_block_group_cache *cache)
10174 {
10175         struct btrfs_space_info *space_info = cache->space_info;
10176         struct btrfs_fs_info *fs_info = cache->fs_info;
10177         int index = btrfs_bg_flags_to_raid_index(cache->flags);
10178         bool first = false;
10179
10180         down_write(&space_info->groups_sem);
10181         if (list_empty(&space_info->block_groups[index]))
10182                 first = true;
10183         list_add_tail(&cache->list, &space_info->block_groups[index]);
10184         up_write(&space_info->groups_sem);
10185
10186         if (first) {
10187                 struct raid_kobject *rkobj = kzalloc(sizeof(*rkobj), GFP_NOFS);
10188                 if (!rkobj) {
10189                         btrfs_warn(cache->fs_info,
10190                                 "couldn't alloc memory for raid level kobject");
10191                         return;
10192                 }
10193                 rkobj->flags = cache->flags;
10194                 kobject_init(&rkobj->kobj, &btrfs_raid_ktype);
10195
10196                 spin_lock(&fs_info->pending_raid_kobjs_lock);
10197                 list_add_tail(&rkobj->list, &fs_info->pending_raid_kobjs);
10198                 spin_unlock(&fs_info->pending_raid_kobjs_lock);
10199                 space_info->block_group_kobjs[index] = &rkobj->kobj;
10200         }
10201 }
10202
10203 static struct btrfs_block_group_cache *
10204 btrfs_create_block_group_cache(struct btrfs_fs_info *fs_info,
10205                                u64 start, u64 size)
10206 {
10207         struct btrfs_block_group_cache *cache;
10208
10209         cache = kzalloc(sizeof(*cache), GFP_NOFS);
10210         if (!cache)
10211                 return NULL;
10212
10213         cache->free_space_ctl = kzalloc(sizeof(*cache->free_space_ctl),
10214                                         GFP_NOFS);
10215         if (!cache->free_space_ctl) {
10216                 kfree(cache);
10217                 return NULL;
10218         }
10219
10220         cache->key.objectid = start;
10221         cache->key.offset = size;
10222         cache->key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
10223
10224         cache->fs_info = fs_info;
10225         cache->full_stripe_len = btrfs_full_stripe_len(fs_info, start);
10226         set_free_space_tree_thresholds(cache);
10227
10228         atomic_set(&cache->count, 1);
10229         spin_lock_init(&cache->lock);
10230         init_rwsem(&cache->data_rwsem);
10231         INIT_LIST_HEAD(&cache->list);
10232         INIT_LIST_HEAD(&cache->cluster_list);
10233         INIT_LIST_HEAD(&cache->bg_list);
10234         INIT_LIST_HEAD(&cache->ro_list);
10235         INIT_LIST_HEAD(&cache->dirty_list);
10236         INIT_LIST_HEAD(&cache->io_list);
10237         btrfs_init_free_space_ctl(cache);
10238         atomic_set(&cache->trimming, 0);
10239         mutex_init(&cache->free_space_lock);
10240         btrfs_init_full_stripe_locks_tree(&cache->full_stripe_locks_root);
10241
10242         return cache;
10243 }
10244
10245
10246 /*
10247  * Iterate all chunks and verify that each of them has the corresponding block
10248  * group
10249  */
10250 static int check_chunk_block_group_mappings(struct btrfs_fs_info *fs_info)
10251 {
10252         struct extent_map_tree *map_tree = &fs_info->mapping_tree;
10253         struct extent_map *em;
10254         struct btrfs_block_group_cache *bg;
10255         u64 start = 0;
10256         int ret = 0;
10257
10258         while (1) {
10259                 read_lock(&map_tree->lock);
10260                 /*
10261                  * lookup_extent_mapping will return the first extent map
10262                  * intersecting the range, so setting @len to 1 is enough to
10263                  * get the first chunk.
10264                  */
10265                 em = lookup_extent_mapping(map_tree, start, 1);
10266                 read_unlock(&map_tree->lock);
10267                 if (!em)
10268                         break;
10269
10270                 bg = btrfs_lookup_block_group(fs_info, em->start);
10271                 if (!bg) {
10272                         btrfs_err(fs_info,
10273         "chunk start=%llu len=%llu doesn't have corresponding block group",
10274                                      em->start, em->len);
10275                         ret = -EUCLEAN;
10276                         free_extent_map(em);
10277                         break;
10278                 }
10279                 if (bg->key.objectid != em->start ||
10280                     bg->key.offset != em->len ||
10281                     (bg->flags & BTRFS_BLOCK_GROUP_TYPE_MASK) !=
10282                     (em->map_lookup->type & BTRFS_BLOCK_GROUP_TYPE_MASK)) {
10283                         btrfs_err(fs_info,
10284 "chunk start=%llu len=%llu flags=0x%llx doesn't match block group start=%llu len=%llu flags=0x%llx",
10285                                 em->start, em->len,
10286                                 em->map_lookup->type & BTRFS_BLOCK_GROUP_TYPE_MASK,
10287                                 bg->key.objectid, bg->key.offset,
10288                                 bg->flags & BTRFS_BLOCK_GROUP_TYPE_MASK);
10289                         ret = -EUCLEAN;
10290                         free_extent_map(em);
10291                         btrfs_put_block_group(bg);
10292                         break;
10293                 }
10294                 start = em->start + em->len;
10295                 free_extent_map(em);
10296                 btrfs_put_block_group(bg);
10297         }
10298         return ret;
10299 }
10300
10301 int btrfs_read_block_groups(struct btrfs_fs_info *info)
10302 {
10303         struct btrfs_path *path;
10304         int ret;
10305         struct btrfs_block_group_cache *cache;
10306         struct btrfs_space_info *space_info;
10307         struct btrfs_key key;
10308         struct btrfs_key found_key;
10309         struct extent_buffer *leaf;
10310         int need_clear = 0;
10311         u64 cache_gen;
10312         u64 feature;
10313         int mixed;
10314
10315         feature = btrfs_super_incompat_flags(info->super_copy);
10316         mixed = !!(feature & BTRFS_FEATURE_INCOMPAT_MIXED_GROUPS);
10317
10318         key.objectid = 0;
10319         key.offset = 0;
10320         key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
10321         path = btrfs_alloc_path();
10322         if (!path)
10323                 return -ENOMEM;
10324         path->reada = READA_FORWARD;
10325
10326         cache_gen = btrfs_super_cache_generation(info->super_copy);
10327         if (btrfs_test_opt(info, SPACE_CACHE) &&
10328             btrfs_super_generation(info->super_copy) != cache_gen)
10329                 need_clear = 1;
10330         if (btrfs_test_opt(info, CLEAR_CACHE))
10331                 need_clear = 1;
10332
10333         while (1) {
10334                 ret = find_first_block_group(info, path, &key);
10335                 if (ret > 0)
10336                         break;
10337                 if (ret != 0)
10338                         goto error;
10339
10340                 leaf = path->nodes[0];
10341                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
10342
10343                 cache = btrfs_create_block_group_cache(info, found_key.objectid,
10344                                                        found_key.offset);
10345                 if (!cache) {
10346                         ret = -ENOMEM;
10347                         goto error;
10348                 }
10349
10350                 if (need_clear) {
10351                         /*
10352                          * When we mount with old space cache, we need to
10353                          * set BTRFS_DC_CLEAR and set dirty flag.
10354                          *
10355                          * a) Setting 'BTRFS_DC_CLEAR' makes sure that we
10356                          *    truncate the old free space cache inode and
10357                          *    setup a new one.
10358                          * b) Setting 'dirty flag' makes sure that we flush
10359                          *    the new space cache info onto disk.
10360                          */
10361                         if (btrfs_test_opt(info, SPACE_CACHE))
10362                                 cache->disk_cache_state = BTRFS_DC_CLEAR;
10363                 }
10364
10365                 read_extent_buffer(leaf, &cache->item,
10366                                    btrfs_item_ptr_offset(leaf, path->slots[0]),
10367                                    sizeof(cache->item));
10368                 cache->flags = btrfs_block_group_flags(&cache->item);
10369                 if (!mixed &&
10370                     ((cache->flags & BTRFS_BLOCK_GROUP_METADATA) &&
10371                     (cache->flags & BTRFS_BLOCK_GROUP_DATA))) {
10372                         btrfs_err(info,
10373 "bg %llu is a mixed block group but filesystem hasn't enabled mixed block groups",
10374                                   cache->key.objectid);
10375                         ret = -EINVAL;
10376                         goto error;
10377                 }
10378
10379                 key.objectid = found_key.objectid + found_key.offset;
10380                 btrfs_release_path(path);
10381
10382                 /*
10383                  * We need to exclude the super stripes now so that the space
10384                  * info has super bytes accounted for, otherwise we'll think
10385                  * we have more space than we actually do.
10386                  */
10387                 ret = exclude_super_stripes(cache);
10388                 if (ret) {
10389                         /*
10390                          * We may have excluded something, so call this just in
10391                          * case.
10392                          */
10393                         free_excluded_extents(cache);
10394                         btrfs_put_block_group(cache);
10395                         goto error;
10396                 }
10397
10398                 /*
10399                  * check for two cases, either we are full, and therefore
10400                  * don't need to bother with the caching work since we won't
10401                  * find any space, or we are empty, and we can just add all
10402                  * the space in and be done with it.  This saves us _a_lot_ of
10403                  * time, particularly in the full case.
10404                  */
10405                 if (found_key.offset == btrfs_block_group_used(&cache->item)) {
10406                         cache->last_byte_to_unpin = (u64)-1;
10407                         cache->cached = BTRFS_CACHE_FINISHED;
10408                         free_excluded_extents(cache);
10409                 } else if (btrfs_block_group_used(&cache->item) == 0) {
10410                         cache->last_byte_to_unpin = (u64)-1;
10411                         cache->cached = BTRFS_CACHE_FINISHED;
10412                         add_new_free_space(cache, found_key.objectid,
10413                                            found_key.objectid +
10414                                            found_key.offset);
10415                         free_excluded_extents(cache);
10416                 }
10417
10418                 ret = btrfs_add_block_group_cache(info, cache);
10419                 if (ret) {
10420                         btrfs_remove_free_space_cache(cache);
10421                         btrfs_put_block_group(cache);
10422                         goto error;
10423                 }
10424
10425                 trace_btrfs_add_block_group(info, cache, 0);
10426                 update_space_info(info, cache->flags, found_key.offset,
10427                                   btrfs_block_group_used(&cache->item),
10428                                   cache->bytes_super, &space_info);
10429
10430                 cache->space_info = space_info;
10431
10432                 link_block_group(cache);
10433
10434                 set_avail_alloc_bits(info, cache->flags);
10435                 if (btrfs_chunk_readonly(info, cache->key.objectid)) {
10436                         inc_block_group_ro(cache, 1);
10437                 } else if (btrfs_block_group_used(&cache->item) == 0) {
10438                         ASSERT(list_empty(&cache->bg_list));
10439                         btrfs_mark_bg_unused(cache);
10440                 }
10441         }
10442
10443         list_for_each_entry_rcu(space_info, &info->space_info, list) {
10444                 if (!(get_alloc_profile(info, space_info->flags) &
10445                       (BTRFS_BLOCK_GROUP_RAID10 |
10446                        BTRFS_BLOCK_GROUP_RAID1_MASK |
10447                        BTRFS_BLOCK_GROUP_RAID56_MASK |
10448                        BTRFS_BLOCK_GROUP_DUP)))
10449                         continue;
10450                 /*
10451                  * avoid allocating from un-mirrored block group if there are
10452                  * mirrored block groups.
10453                  */
10454                 list_for_each_entry(cache,
10455                                 &space_info->block_groups[BTRFS_RAID_RAID0],
10456                                 list)
10457                         inc_block_group_ro(cache, 1);
10458                 list_for_each_entry(cache,
10459                                 &space_info->block_groups[BTRFS_RAID_SINGLE],
10460                                 list)
10461                         inc_block_group_ro(cache, 1);
10462         }
10463
10464         btrfs_add_raid_kobjects(info);
10465         init_global_block_rsv(info);
10466         ret = check_chunk_block_group_mappings(info);
10467 error:
10468         btrfs_free_path(path);
10469         return ret;
10470 }
10471
10472 void btrfs_create_pending_block_groups(struct btrfs_trans_handle *trans)
10473 {
10474         struct btrfs_fs_info *fs_info = trans->fs_info;
10475         struct btrfs_block_group_cache *block_group;
10476         struct btrfs_root *extent_root = fs_info->extent_root;
10477         struct btrfs_block_group_item item;
10478         struct btrfs_key key;
10479         int ret = 0;
10480
10481         if (!trans->can_flush_pending_bgs)
10482                 return;
10483
10484         while (!list_empty(&trans->new_bgs)) {
10485                 block_group = list_first_entry(&trans->new_bgs,
10486                                                struct btrfs_block_group_cache,
10487                                                bg_list);
10488                 if (ret)
10489                         goto next;
10490
10491                 spin_lock(&block_group->lock);
10492                 memcpy(&item, &block_group->item, sizeof(item));
10493                 memcpy(&key, &block_group->key, sizeof(key));
10494                 spin_unlock(&block_group->lock);
10495
10496                 ret = btrfs_insert_item(trans, extent_root, &key, &item,
10497                                         sizeof(item));
10498                 if (ret)
10499                         btrfs_abort_transaction(trans, ret);
10500                 ret = btrfs_finish_chunk_alloc(trans, key.objectid, key.offset);
10501                 if (ret)
10502                         btrfs_abort_transaction(trans, ret);
10503                 add_block_group_free_space(trans, block_group);
10504                 /* already aborted the transaction if it failed. */
10505 next:
10506                 btrfs_delayed_refs_rsv_release(fs_info, 1);
10507                 list_del_init(&block_group->bg_list);
10508         }
10509         btrfs_trans_release_chunk_metadata(trans);
10510 }
10511
10512 int btrfs_make_block_group(struct btrfs_trans_handle *trans, u64 bytes_used,
10513                            u64 type, u64 chunk_offset, u64 size)
10514 {
10515         struct btrfs_fs_info *fs_info = trans->fs_info;
10516         struct btrfs_block_group_cache *cache;
10517         int ret;
10518
10519         btrfs_set_log_full_commit(trans);
10520
10521         cache = btrfs_create_block_group_cache(fs_info, chunk_offset, size);
10522         if (!cache)
10523                 return -ENOMEM;
10524
10525         btrfs_set_block_group_used(&cache->item, bytes_used);
10526         btrfs_set_block_group_chunk_objectid(&cache->item,
10527                                              BTRFS_FIRST_CHUNK_TREE_OBJECTID);
10528         btrfs_set_block_group_flags(&cache->item, type);
10529
10530         cache->flags = type;
10531         cache->last_byte_to_unpin = (u64)-1;
10532         cache->cached = BTRFS_CACHE_FINISHED;
10533         cache->needs_free_space = 1;
10534         ret = exclude_super_stripes(cache);
10535         if (ret) {
10536                 /*
10537                  * We may have excluded something, so call this just in
10538                  * case.
10539                  */
10540                 free_excluded_extents(cache);
10541                 btrfs_put_block_group(cache);
10542                 return ret;
10543         }
10544
10545         add_new_free_space(cache, chunk_offset, chunk_offset + size);
10546
10547         free_excluded_extents(cache);
10548
10549 #ifdef CONFIG_BTRFS_DEBUG
10550         if (btrfs_should_fragment_free_space(cache)) {
10551                 u64 new_bytes_used = size - bytes_used;
10552
10553                 bytes_used += new_bytes_used >> 1;
10554                 fragment_free_space(cache);
10555         }
10556 #endif
10557         /*
10558          * Ensure the corresponding space_info object is created and
10559          * assigned to our block group. We want our bg to be added to the rbtree
10560          * with its ->space_info set.
10561          */
10562         cache->space_info = __find_space_info(fs_info, cache->flags);
10563         ASSERT(cache->space_info);
10564
10565         ret = btrfs_add_block_group_cache(fs_info, cache);
10566         if (ret) {
10567                 btrfs_remove_free_space_cache(cache);
10568                 btrfs_put_block_group(cache);
10569                 return ret;
10570         }
10571
10572         /*
10573          * Now that our block group has its ->space_info set and is inserted in
10574          * the rbtree, update the space info's counters.
10575          */
10576         trace_btrfs_add_block_group(fs_info, cache, 1);
10577         update_space_info(fs_info, cache->flags, size, bytes_used,
10578                                 cache->bytes_super, &cache->space_info);
10579         update_global_block_rsv(fs_info);
10580
10581         link_block_group(cache);
10582
10583         list_add_tail(&cache->bg_list, &trans->new_bgs);
10584         trans->delayed_ref_updates++;
10585         btrfs_update_delayed_refs_rsv(trans);
10586
10587         set_avail_alloc_bits(fs_info, type);
10588         return 0;
10589 }
10590
10591 static void clear_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
10592 {
10593         u64 extra_flags = chunk_to_extended(flags) &
10594                                 BTRFS_EXTENDED_PROFILE_MASK;
10595
10596         write_seqlock(&fs_info->profiles_lock);
10597         if (flags & BTRFS_BLOCK_GROUP_DATA)
10598                 fs_info->avail_data_alloc_bits &= ~extra_flags;
10599         if (flags & BTRFS_BLOCK_GROUP_METADATA)
10600                 fs_info->avail_metadata_alloc_bits &= ~extra_flags;
10601         if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
10602                 fs_info->avail_system_alloc_bits &= ~extra_flags;
10603         write_sequnlock(&fs_info->profiles_lock);
10604 }
10605
10606 /*
10607  * Clear incompat bits for the following feature(s):
10608  *
10609  * - RAID56 - in case there's neither RAID5 nor RAID6 profile block group
10610  *            in the whole filesystem
10611  */
10612 static void clear_incompat_bg_bits(struct btrfs_fs_info *fs_info, u64 flags)
10613 {
10614         if (flags & BTRFS_BLOCK_GROUP_RAID56_MASK) {
10615                 struct list_head *head = &fs_info->space_info;
10616                 struct btrfs_space_info *sinfo;
10617
10618                 list_for_each_entry_rcu(sinfo, head, list) {
10619                         bool found = false;
10620
10621                         down_read(&sinfo->groups_sem);
10622                         if (!list_empty(&sinfo->block_groups[BTRFS_RAID_RAID5]))
10623                                 found = true;
10624                         if (!list_empty(&sinfo->block_groups[BTRFS_RAID_RAID6]))
10625                                 found = true;
10626                         up_read(&sinfo->groups_sem);
10627
10628                         if (found)
10629                                 return;
10630                 }
10631                 btrfs_clear_fs_incompat(fs_info, RAID56);
10632         }
10633 }
10634
10635 int btrfs_remove_block_group(struct btrfs_trans_handle *trans,
10636                              u64 group_start, struct extent_map *em)
10637 {
10638         struct btrfs_fs_info *fs_info = trans->fs_info;
10639         struct btrfs_root *root = fs_info->extent_root;
10640         struct btrfs_path *path;
10641         struct btrfs_block_group_cache *block_group;
10642         struct btrfs_free_cluster *cluster;
10643         struct btrfs_root *tree_root = fs_info->tree_root;
10644         struct btrfs_key key;
10645         struct inode *inode;
10646         struct kobject *kobj = NULL;
10647         int ret;
10648         int index;
10649         int factor;
10650         struct btrfs_caching_control *caching_ctl = NULL;
10651         bool remove_em;
10652         bool remove_rsv = false;
10653
10654         block_group = btrfs_lookup_block_group(fs_info, group_start);
10655         BUG_ON(!block_group);
10656         BUG_ON(!block_group->ro);
10657
10658         trace_btrfs_remove_block_group(block_group);
10659         /*
10660          * Free the reserved super bytes from this block group before
10661          * remove it.
10662          */
10663         free_excluded_extents(block_group);
10664         btrfs_free_ref_tree_range(fs_info, block_group->key.objectid,
10665                                   block_group->key.offset);
10666
10667         memcpy(&key, &block_group->key, sizeof(key));
10668         index = btrfs_bg_flags_to_raid_index(block_group->flags);
10669         factor = btrfs_bg_type_to_factor(block_group->flags);
10670
10671         /* make sure this block group isn't part of an allocation cluster */
10672         cluster = &fs_info->data_alloc_cluster;
10673         spin_lock(&cluster->refill_lock);
10674         btrfs_return_cluster_to_free_space(block_group, cluster);
10675         spin_unlock(&cluster->refill_lock);
10676
10677         /*
10678          * make sure this block group isn't part of a metadata
10679          * allocation cluster
10680          */
10681         cluster = &fs_info->meta_alloc_cluster;
10682         spin_lock(&cluster->refill_lock);
10683         btrfs_return_cluster_to_free_space(block_group, cluster);
10684         spin_unlock(&cluster->refill_lock);
10685
10686         path = btrfs_alloc_path();
10687         if (!path) {
10688                 ret = -ENOMEM;
10689                 goto out;
10690         }
10691
10692         /*
10693          * get the inode first so any iput calls done for the io_list
10694          * aren't the final iput (no unlinks allowed now)
10695          */
10696         inode = lookup_free_space_inode(block_group, path);
10697
10698         mutex_lock(&trans->transaction->cache_write_mutex);
10699         /*
10700          * Make sure our free space cache IO is done before removing the
10701          * free space inode
10702          */
10703         spin_lock(&trans->transaction->dirty_bgs_lock);
10704         if (!list_empty(&block_group->io_list)) {
10705                 list_del_init(&block_group->io_list);
10706
10707                 WARN_ON(!IS_ERR(inode) && inode != block_group->io_ctl.inode);
10708
10709                 spin_unlock(&trans->transaction->dirty_bgs_lock);
10710                 btrfs_wait_cache_io(trans, block_group, path);
10711                 btrfs_put_block_group(block_group);
10712                 spin_lock(&trans->transaction->dirty_bgs_lock);
10713         }
10714
10715         if (!list_empty(&block_group->dirty_list)) {
10716                 list_del_init(&block_group->dirty_list);
10717                 remove_rsv = true;
10718                 btrfs_put_block_group(block_group);
10719         }
10720         spin_unlock(&trans->transaction->dirty_bgs_lock);
10721         mutex_unlock(&trans->transaction->cache_write_mutex);
10722
10723         if (!IS_ERR(inode)) {
10724                 ret = btrfs_orphan_add(trans, BTRFS_I(inode));
10725                 if (ret) {
10726                         btrfs_add_delayed_iput(inode);
10727                         goto out;
10728                 }
10729                 clear_nlink(inode);
10730                 /* One for the block groups ref */
10731                 spin_lock(&block_group->lock);
10732                 if (block_group->iref) {
10733                         block_group->iref = 0;
10734                         block_group->inode = NULL;
10735                         spin_unlock(&block_group->lock);
10736                         iput(inode);
10737                 } else {
10738                         spin_unlock(&block_group->lock);
10739                 }
10740                 /* One for our lookup ref */
10741                 btrfs_add_delayed_iput(inode);
10742         }
10743
10744         key.objectid = BTRFS_FREE_SPACE_OBJECTID;
10745         key.offset = block_group->key.objectid;
10746         key.type = 0;
10747
10748         ret = btrfs_search_slot(trans, tree_root, &key, path, -1, 1);
10749         if (ret < 0)
10750                 goto out;
10751         if (ret > 0)
10752                 btrfs_release_path(path);
10753         if (ret == 0) {
10754                 ret = btrfs_del_item(trans, tree_root, path);
10755                 if (ret)
10756                         goto out;
10757                 btrfs_release_path(path);
10758         }
10759
10760         spin_lock(&fs_info->block_group_cache_lock);
10761         rb_erase(&block_group->cache_node,
10762                  &fs_info->block_group_cache_tree);
10763         RB_CLEAR_NODE(&block_group->cache_node);
10764
10765         if (fs_info->first_logical_byte == block_group->key.objectid)
10766                 fs_info->first_logical_byte = (u64)-1;
10767         spin_unlock(&fs_info->block_group_cache_lock);
10768
10769         down_write(&block_group->space_info->groups_sem);
10770         /*
10771          * we must use list_del_init so people can check to see if they
10772          * are still on the list after taking the semaphore
10773          */
10774         list_del_init(&block_group->list);
10775         if (list_empty(&block_group->space_info->block_groups[index])) {
10776                 kobj = block_group->space_info->block_group_kobjs[index];
10777                 block_group->space_info->block_group_kobjs[index] = NULL;
10778                 clear_avail_alloc_bits(fs_info, block_group->flags);
10779         }
10780         up_write(&block_group->space_info->groups_sem);
10781         clear_incompat_bg_bits(fs_info, block_group->flags);
10782         if (kobj) {
10783                 kobject_del(kobj);
10784                 kobject_put(kobj);
10785         }
10786
10787         if (block_group->has_caching_ctl)
10788                 caching_ctl = get_caching_control(block_group);
10789         if (block_group->cached == BTRFS_CACHE_STARTED)
10790                 wait_block_group_cache_done(block_group);
10791         if (block_group->has_caching_ctl) {
10792                 down_write(&fs_info->commit_root_sem);
10793                 if (!caching_ctl) {
10794                         struct btrfs_caching_control *ctl;
10795
10796                         list_for_each_entry(ctl,
10797                                     &fs_info->caching_block_groups, list)
10798                                 if (ctl->block_group == block_group) {
10799                                         caching_ctl = ctl;
10800                                         refcount_inc(&caching_ctl->count);
10801                                         break;
10802                                 }
10803                 }
10804                 if (caching_ctl)
10805                         list_del_init(&caching_ctl->list);
10806                 up_write(&fs_info->commit_root_sem);
10807                 if (caching_ctl) {
10808                         /* Once for the caching bgs list and once for us. */
10809                         put_caching_control(caching_ctl);
10810                         put_caching_control(caching_ctl);
10811                 }
10812         }
10813
10814         spin_lock(&trans->transaction->dirty_bgs_lock);
10815         WARN_ON(!list_empty(&block_group->dirty_list));
10816         WARN_ON(!list_empty(&block_group->io_list));
10817         spin_unlock(&trans->transaction->dirty_bgs_lock);
10818
10819         btrfs_remove_free_space_cache(block_group);
10820
10821         spin_lock(&block_group->space_info->lock);
10822         list_del_init(&block_group->ro_list);
10823
10824         if (btrfs_test_opt(fs_info, ENOSPC_DEBUG)) {
10825                 WARN_ON(block_group->space_info->total_bytes
10826                         < block_group->key.offset);
10827                 WARN_ON(block_group->space_info->bytes_readonly
10828                         < block_group->key.offset);
10829                 WARN_ON(block_group->space_info->disk_total
10830                         < block_group->key.offset * factor);
10831         }
10832         block_group->space_info->total_bytes -= block_group->key.offset;
10833         block_group->space_info->bytes_readonly -= block_group->key.offset;
10834         block_group->space_info->disk_total -= block_group->key.offset * factor;
10835
10836         spin_unlock(&block_group->space_info->lock);
10837
10838         memcpy(&key, &block_group->key, sizeof(key));
10839
10840         mutex_lock(&fs_info->chunk_mutex);
10841         spin_lock(&block_group->lock);
10842         block_group->removed = 1;
10843         /*
10844          * At this point trimming can't start on this block group, because we
10845          * removed the block group from the tree fs_info->block_group_cache_tree
10846          * so no one can't find it anymore and even if someone already got this
10847          * block group before we removed it from the rbtree, they have already
10848          * incremented block_group->trimming - if they didn't, they won't find
10849          * any free space entries because we already removed them all when we
10850          * called btrfs_remove_free_space_cache().
10851          *
10852          * And we must not remove the extent map from the fs_info->mapping_tree
10853          * to prevent the same logical address range and physical device space
10854          * ranges from being reused for a new block group. This is because our
10855          * fs trim operation (btrfs_trim_fs() / btrfs_ioctl_fitrim()) is
10856          * completely transactionless, so while it is trimming a range the
10857          * currently running transaction might finish and a new one start,
10858          * allowing for new block groups to be created that can reuse the same
10859          * physical device locations unless we take this special care.
10860          *
10861          * There may also be an implicit trim operation if the file system
10862          * is mounted with -odiscard. The same protections must remain
10863          * in place until the extents have been discarded completely when
10864          * the transaction commit has completed.
10865          */
10866         remove_em = (atomic_read(&block_group->trimming) == 0);
10867         spin_unlock(&block_group->lock);
10868
10869         mutex_unlock(&fs_info->chunk_mutex);
10870
10871         ret = remove_block_group_free_space(trans, block_group);
10872         if (ret)
10873                 goto out;
10874
10875         btrfs_put_block_group(block_group);
10876         btrfs_put_block_group(block_group);
10877
10878         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
10879         if (ret > 0)
10880                 ret = -EIO;
10881         if (ret < 0)
10882                 goto out;
10883
10884         ret = btrfs_del_item(trans, root, path);
10885         if (ret)
10886                 goto out;
10887
10888         if (remove_em) {
10889                 struct extent_map_tree *em_tree;
10890
10891                 em_tree = &fs_info->mapping_tree;
10892                 write_lock(&em_tree->lock);
10893                 remove_extent_mapping(em_tree, em);
10894                 write_unlock(&em_tree->lock);
10895                 /* once for the tree */
10896                 free_extent_map(em);
10897         }
10898 out:
10899         if (remove_rsv)
10900                 btrfs_delayed_refs_rsv_release(fs_info, 1);
10901         btrfs_free_path(path);
10902         return ret;
10903 }
10904
10905 struct btrfs_trans_handle *
10906 btrfs_start_trans_remove_block_group(struct btrfs_fs_info *fs_info,
10907                                      const u64 chunk_offset)
10908 {
10909         struct extent_map_tree *em_tree = &fs_info->mapping_tree;
10910         struct extent_map *em;
10911         struct map_lookup *map;
10912         unsigned int num_items;
10913
10914         read_lock(&em_tree->lock);
10915         em = lookup_extent_mapping(em_tree, chunk_offset, 1);
10916         read_unlock(&em_tree->lock);
10917         ASSERT(em && em->start == chunk_offset);
10918
10919         /*
10920          * We need to reserve 3 + N units from the metadata space info in order
10921          * to remove a block group (done at btrfs_remove_chunk() and at
10922          * btrfs_remove_block_group()), which are used for:
10923          *
10924          * 1 unit for adding the free space inode's orphan (located in the tree
10925          * of tree roots).
10926          * 1 unit for deleting the block group item (located in the extent
10927          * tree).
10928          * 1 unit for deleting the free space item (located in tree of tree
10929          * roots).
10930          * N units for deleting N device extent items corresponding to each
10931          * stripe (located in the device tree).
10932          *
10933          * In order to remove a block group we also need to reserve units in the
10934          * system space info in order to update the chunk tree (update one or
10935          * more device items and remove one chunk item), but this is done at
10936          * btrfs_remove_chunk() through a call to check_system_chunk().
10937          */
10938         map = em->map_lookup;
10939         num_items = 3 + map->num_stripes;
10940         free_extent_map(em);
10941
10942         return btrfs_start_transaction_fallback_global_rsv(fs_info->extent_root,
10943                                                            num_items, 1);
10944 }
10945
10946 /*
10947  * Process the unused_bgs list and remove any that don't have any allocated
10948  * space inside of them.
10949  */
10950 void btrfs_delete_unused_bgs(struct btrfs_fs_info *fs_info)
10951 {
10952         struct btrfs_block_group_cache *block_group;
10953         struct btrfs_space_info *space_info;
10954         struct btrfs_trans_handle *trans;
10955         int ret = 0;
10956
10957         if (!test_bit(BTRFS_FS_OPEN, &fs_info->flags))
10958                 return;
10959
10960         spin_lock(&fs_info->unused_bgs_lock);
10961         while (!list_empty(&fs_info->unused_bgs)) {
10962                 u64 start, end;
10963                 int trimming;
10964
10965                 block_group = list_first_entry(&fs_info->unused_bgs,
10966                                                struct btrfs_block_group_cache,
10967                                                bg_list);
10968                 list_del_init(&block_group->bg_list);
10969
10970                 space_info = block_group->space_info;
10971
10972                 if (ret || btrfs_mixed_space_info(space_info)) {
10973                         btrfs_put_block_group(block_group);
10974                         continue;
10975                 }
10976                 spin_unlock(&fs_info->unused_bgs_lock);
10977
10978                 mutex_lock(&fs_info->delete_unused_bgs_mutex);
10979
10980                 /* Don't want to race with allocators so take the groups_sem */
10981                 down_write(&space_info->groups_sem);
10982                 spin_lock(&block_group->lock);
10983                 if (block_group->reserved || block_group->pinned ||
10984                     btrfs_block_group_used(&block_group->item) ||
10985                     block_group->ro ||
10986                     list_is_singular(&block_group->list)) {
10987                         /*
10988                          * We want to bail if we made new allocations or have
10989                          * outstanding allocations in this block group.  We do
10990                          * the ro check in case balance is currently acting on
10991                          * this block group.
10992                          */
10993                         trace_btrfs_skip_unused_block_group(block_group);
10994                         spin_unlock(&block_group->lock);
10995                         up_write(&space_info->groups_sem);
10996                         goto next;
10997                 }
10998                 spin_unlock(&block_group->lock);
10999
11000                 /* We don't want to force the issue, only flip if it's ok. */
11001                 ret = inc_block_group_ro(block_group, 0);
11002                 up_write(&space_info->groups_sem);
11003                 if (ret < 0) {
11004                         ret = 0;
11005                         goto next;
11006                 }
11007
11008                 /*
11009                  * Want to do this before we do anything else so we can recover
11010                  * properly if we fail to join the transaction.
11011                  */
11012                 trans = btrfs_start_trans_remove_block_group(fs_info,
11013                                                      block_group->key.objectid);
11014                 if (IS_ERR(trans)) {
11015                         btrfs_dec_block_group_ro(block_group);
11016                         ret = PTR_ERR(trans);
11017                         goto next;
11018                 }
11019
11020                 /*
11021                  * We could have pending pinned extents for this block group,
11022                  * just delete them, we don't care about them anymore.
11023                  */
11024                 start = block_group->key.objectid;
11025                 end = start + block_group->key.offset - 1;
11026                 /*
11027                  * Hold the unused_bg_unpin_mutex lock to avoid racing with
11028                  * btrfs_finish_extent_commit(). If we are at transaction N,
11029                  * another task might be running finish_extent_commit() for the
11030                  * previous transaction N - 1, and have seen a range belonging
11031                  * to the block group in freed_extents[] before we were able to
11032                  * clear the whole block group range from freed_extents[]. This
11033                  * means that task can lookup for the block group after we
11034                  * unpinned it from freed_extents[] and removed it, leading to
11035                  * a BUG_ON() at btrfs_unpin_extent_range().
11036                  */
11037                 mutex_lock(&fs_info->unused_bg_unpin_mutex);
11038                 ret = clear_extent_bits(&fs_info->freed_extents[0], start, end,
11039                                   EXTENT_DIRTY);
11040                 if (ret) {
11041                         mutex_unlock(&fs_info->unused_bg_unpin_mutex);
11042                         btrfs_dec_block_group_ro(block_group);
11043                         goto end_trans;
11044                 }
11045                 ret = clear_extent_bits(&fs_info->freed_extents[1], start, end,
11046                                   EXTENT_DIRTY);
11047                 if (ret) {
11048                         mutex_unlock(&fs_info->unused_bg_unpin_mutex);
11049                         btrfs_dec_block_group_ro(block_group);
11050                         goto end_trans;
11051                 }
11052                 mutex_unlock(&fs_info->unused_bg_unpin_mutex);
11053
11054                 /* Reset pinned so btrfs_put_block_group doesn't complain */
11055                 spin_lock(&space_info->lock);
11056                 spin_lock(&block_group->lock);
11057
11058                 update_bytes_pinned(fs_info, space_info, -block_group->pinned);
11059                 space_info->bytes_readonly += block_group->pinned;
11060                 percpu_counter_add_batch(&space_info->total_bytes_pinned,
11061                                    -block_group->pinned,
11062                                    BTRFS_TOTAL_BYTES_PINNED_BATCH);
11063                 block_group->pinned = 0;
11064
11065                 spin_unlock(&block_group->lock);
11066                 spin_unlock(&space_info->lock);
11067
11068                 /* DISCARD can flip during remount */
11069                 trimming = btrfs_test_opt(fs_info, DISCARD);
11070
11071                 /* Implicit trim during transaction commit. */
11072                 if (trimming)
11073                         btrfs_get_block_group_trimming(block_group);
11074
11075                 /*
11076                  * Btrfs_remove_chunk will abort the transaction if things go
11077                  * horribly wrong.
11078                  */
11079                 ret = btrfs_remove_chunk(trans, block_group->key.objectid);
11080
11081                 if (ret) {
11082                         if (trimming)
11083                                 btrfs_put_block_group_trimming(block_group);
11084                         goto end_trans;
11085                 }
11086
11087                 /*
11088                  * If we're not mounted with -odiscard, we can just forget
11089                  * about this block group. Otherwise we'll need to wait
11090                  * until transaction commit to do the actual discard.
11091                  */
11092                 if (trimming) {
11093                         spin_lock(&fs_info->unused_bgs_lock);
11094                         /*
11095                          * A concurrent scrub might have added us to the list
11096                          * fs_info->unused_bgs, so use a list_move operation
11097                          * to add the block group to the deleted_bgs list.
11098                          */
11099                         list_move(&block_group->bg_list,
11100                                   &trans->transaction->deleted_bgs);
11101                         spin_unlock(&fs_info->unused_bgs_lock);
11102                         btrfs_get_block_group(block_group);
11103                 }
11104 end_trans:
11105                 btrfs_end_transaction(trans);
11106 next:
11107                 mutex_unlock(&fs_info->delete_unused_bgs_mutex);
11108                 btrfs_put_block_group(block_group);
11109                 spin_lock(&fs_info->unused_bgs_lock);
11110         }
11111         spin_unlock(&fs_info->unused_bgs_lock);
11112 }
11113
11114 int btrfs_init_space_info(struct btrfs_fs_info *fs_info)
11115 {
11116         struct btrfs_super_block *disk_super;
11117         u64 features;
11118         u64 flags;
11119         int mixed = 0;
11120         int ret;
11121
11122         disk_super = fs_info->super_copy;
11123         if (!btrfs_super_root(disk_super))
11124                 return -EINVAL;
11125
11126         features = btrfs_super_incompat_flags(disk_super);
11127         if (features & BTRFS_FEATURE_INCOMPAT_MIXED_GROUPS)
11128                 mixed = 1;
11129
11130         flags = BTRFS_BLOCK_GROUP_SYSTEM;
11131         ret = create_space_info(fs_info, flags);
11132         if (ret)
11133                 goto out;
11134
11135         if (mixed) {
11136                 flags = BTRFS_BLOCK_GROUP_METADATA | BTRFS_BLOCK_GROUP_DATA;
11137                 ret = create_space_info(fs_info, flags);
11138         } else {
11139                 flags = BTRFS_BLOCK_GROUP_METADATA;
11140                 ret = create_space_info(fs_info, flags);
11141                 if (ret)
11142                         goto out;
11143
11144                 flags = BTRFS_BLOCK_GROUP_DATA;
11145                 ret = create_space_info(fs_info, flags);
11146         }
11147 out:
11148         return ret;
11149 }
11150
11151 int btrfs_error_unpin_extent_range(struct btrfs_fs_info *fs_info,
11152                                    u64 start, u64 end)
11153 {
11154         return unpin_extent_range(fs_info, start, end, false);
11155 }
11156
11157 /*
11158  * It used to be that old block groups would be left around forever.
11159  * Iterating over them would be enough to trim unused space.  Since we
11160  * now automatically remove them, we also need to iterate over unallocated
11161  * space.
11162  *
11163  * We don't want a transaction for this since the discard may take a
11164  * substantial amount of time.  We don't require that a transaction be
11165  * running, but we do need to take a running transaction into account
11166  * to ensure that we're not discarding chunks that were released or
11167  * allocated in the current transaction.
11168  *
11169  * Holding the chunks lock will prevent other threads from allocating
11170  * or releasing chunks, but it won't prevent a running transaction
11171  * from committing and releasing the memory that the pending chunks
11172  * list head uses.  For that, we need to take a reference to the
11173  * transaction and hold the commit root sem.  We only need to hold
11174  * it while performing the free space search since we have already
11175  * held back allocations.
11176  */
11177 static int btrfs_trim_free_extents(struct btrfs_device *device, u64 *trimmed)
11178 {
11179         u64 start = SZ_1M, len = 0, end = 0;
11180         int ret;
11181
11182         *trimmed = 0;
11183
11184         /* Discard not supported = nothing to do. */
11185         if (!blk_queue_discard(bdev_get_queue(device->bdev)))
11186                 return 0;
11187
11188         /* Not writable = nothing to do. */
11189         if (!test_bit(BTRFS_DEV_STATE_WRITEABLE, &device->dev_state))
11190                 return 0;
11191
11192         /* No free space = nothing to do. */
11193         if (device->total_bytes <= device->bytes_used)
11194                 return 0;
11195
11196         ret = 0;
11197
11198         while (1) {
11199                 struct btrfs_fs_info *fs_info = device->fs_info;
11200                 u64 bytes;
11201
11202                 ret = mutex_lock_interruptible(&fs_info->chunk_mutex);
11203                 if (ret)
11204                         break;
11205
11206                 find_first_clear_extent_bit(&device->alloc_state, start,
11207                                             &start, &end,
11208                                             CHUNK_TRIMMED | CHUNK_ALLOCATED);
11209
11210                 /* Ensure we skip the reserved area in the first 1M */
11211                 start = max_t(u64, start, SZ_1M);
11212
11213                 /*
11214                  * If find_first_clear_extent_bit find a range that spans the
11215                  * end of the device it will set end to -1, in this case it's up
11216                  * to the caller to trim the value to the size of the device.
11217                  */
11218                 end = min(end, device->total_bytes - 1);
11219
11220                 len = end - start + 1;
11221
11222                 /* We didn't find any extents */
11223                 if (!len) {
11224                         mutex_unlock(&fs_info->chunk_mutex);
11225                         ret = 0;
11226                         break;
11227                 }
11228
11229                 ret = btrfs_issue_discard(device->bdev, start, len,
11230                                           &bytes);
11231                 if (!ret)
11232                         set_extent_bits(&device->alloc_state, start,
11233                                         start + bytes - 1,
11234                                         CHUNK_TRIMMED);
11235                 mutex_unlock(&fs_info->chunk_mutex);
11236
11237                 if (ret)
11238                         break;
11239
11240                 start += len;
11241                 *trimmed += bytes;
11242
11243                 if (fatal_signal_pending(current)) {
11244                         ret = -ERESTARTSYS;
11245                         break;
11246                 }
11247
11248                 cond_resched();
11249         }
11250
11251         return ret;
11252 }
11253
11254 /*
11255  * Trim the whole filesystem by:
11256  * 1) trimming the free space in each block group
11257  * 2) trimming the unallocated space on each device
11258  *
11259  * This will also continue trimming even if a block group or device encounters
11260  * an error.  The return value will be the last error, or 0 if nothing bad
11261  * happens.
11262  */
11263 int btrfs_trim_fs(struct btrfs_fs_info *fs_info, struct fstrim_range *range)
11264 {
11265         struct btrfs_block_group_cache *cache = NULL;
11266         struct btrfs_device *device;
11267         struct list_head *devices;
11268         u64 group_trimmed;
11269         u64 start;
11270         u64 end;
11271         u64 trimmed = 0;
11272         u64 bg_failed = 0;
11273         u64 dev_failed = 0;
11274         int bg_ret = 0;
11275         int dev_ret = 0;
11276         int ret = 0;
11277
11278         cache = btrfs_lookup_first_block_group(fs_info, range->start);
11279         for (; cache; cache = next_block_group(cache)) {
11280                 if (cache->key.objectid >= (range->start + range->len)) {
11281                         btrfs_put_block_group(cache);
11282                         break;
11283                 }
11284
11285                 start = max(range->start, cache->key.objectid);
11286                 end = min(range->start + range->len,
11287                                 cache->key.objectid + cache->key.offset);
11288
11289                 if (end - start >= range->minlen) {
11290                         if (!block_group_cache_done(cache)) {
11291                                 ret = cache_block_group(cache, 0);
11292                                 if (ret) {
11293                                         bg_failed++;
11294                                         bg_ret = ret;
11295                                         continue;
11296                                 }
11297                                 ret = wait_block_group_cache_done(cache);
11298                                 if (ret) {
11299                                         bg_failed++;
11300                                         bg_ret = ret;
11301                                         continue;
11302                                 }
11303                         }
11304                         ret = btrfs_trim_block_group(cache,
11305                                                      &group_trimmed,
11306                                                      start,
11307                                                      end,
11308                                                      range->minlen);
11309
11310                         trimmed += group_trimmed;
11311                         if (ret) {
11312                                 bg_failed++;
11313                                 bg_ret = ret;
11314                                 continue;
11315                         }
11316                 }
11317         }
11318
11319         if (bg_failed)
11320                 btrfs_warn(fs_info,
11321                         "failed to trim %llu block group(s), last error %d",
11322                         bg_failed, bg_ret);
11323         mutex_lock(&fs_info->fs_devices->device_list_mutex);
11324         devices = &fs_info->fs_devices->devices;
11325         list_for_each_entry(device, devices, dev_list) {
11326                 ret = btrfs_trim_free_extents(device, &group_trimmed);
11327                 if (ret) {
11328                         dev_failed++;
11329                         dev_ret = ret;
11330                         break;
11331                 }
11332
11333                 trimmed += group_trimmed;
11334         }
11335         mutex_unlock(&fs_info->fs_devices->device_list_mutex);
11336
11337         if (dev_failed)
11338                 btrfs_warn(fs_info,
11339                         "failed to trim %llu device(s), last error %d",
11340                         dev_failed, dev_ret);
11341         range->len = trimmed;
11342         if (bg_ret)
11343                 return bg_ret;
11344         return dev_ret;
11345 }
11346
11347 /*
11348  * btrfs_{start,end}_write_no_snapshotting() are similar to
11349  * mnt_{want,drop}_write(), they are used to prevent some tasks from writing
11350  * data into the page cache through nocow before the subvolume is snapshoted,
11351  * but flush the data into disk after the snapshot creation, or to prevent
11352  * operations while snapshotting is ongoing and that cause the snapshot to be
11353  * inconsistent (writes followed by expanding truncates for example).
11354  */
11355 void btrfs_end_write_no_snapshotting(struct btrfs_root *root)
11356 {
11357         percpu_counter_dec(&root->subv_writers->counter);
11358         cond_wake_up(&root->subv_writers->wait);
11359 }
11360
11361 int btrfs_start_write_no_snapshotting(struct btrfs_root *root)
11362 {
11363         if (atomic_read(&root->will_be_snapshotted))
11364                 return 0;
11365
11366         percpu_counter_inc(&root->subv_writers->counter);
11367         /*
11368          * Make sure counter is updated before we check for snapshot creation.
11369          */
11370         smp_mb();
11371         if (atomic_read(&root->will_be_snapshotted)) {
11372                 btrfs_end_write_no_snapshotting(root);
11373                 return 0;
11374         }
11375         return 1;
11376 }
11377
11378 void btrfs_wait_for_snapshot_creation(struct btrfs_root *root)
11379 {
11380         while (true) {
11381                 int ret;
11382
11383                 ret = btrfs_start_write_no_snapshotting(root);
11384                 if (ret)
11385                         break;
11386                 wait_var_event(&root->will_be_snapshotted,
11387                                !atomic_read(&root->will_be_snapshotted));
11388         }
11389 }
11390
11391 void btrfs_mark_bg_unused(struct btrfs_block_group_cache *bg)
11392 {
11393         struct btrfs_fs_info *fs_info = bg->fs_info;
11394
11395         spin_lock(&fs_info->unused_bgs_lock);
11396         if (list_empty(&bg->bg_list)) {
11397                 btrfs_get_block_group(bg);
11398                 trace_btrfs_add_unused_block_group(bg);
11399                 list_add_tail(&bg->bg_list, &fs_info->unused_bgs);
11400         }
11401         spin_unlock(&fs_info->unused_bgs_lock);
11402 }