fs/btrfs/extent-tree.c

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