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 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
  55                                struct btrfs_fs_info *fs_info,
  56                                 struct btrfs_delayed_ref_node *node, u64 parent,
  57                                 u64 root_objectid, u64 owner_objectid,
  58                                 u64 owner_offset, int refs_to_drop,
  59                                 struct btrfs_delayed_extent_op *extra_op);
  60 static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op,
  61                                     struct extent_buffer *leaf,
  62                                     struct btrfs_extent_item *ei);
  63 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
  64                                       struct btrfs_fs_info *fs_info,
  65                                       u64 parent, u64 root_objectid,
  66                                       u64 flags, u64 owner, u64 offset,
  67                                       struct btrfs_key *ins, int ref_mod);
  68 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
  69                                      struct btrfs_fs_info *fs_info,
  70                                      u64 parent, u64 root_objectid,
  71                                      u64 flags, struct btrfs_disk_key *key,
  72                                      int level, struct btrfs_key *ins);
  73 static int do_chunk_alloc(struct btrfs_trans_handle *trans,
  74                           struct btrfs_fs_info *fs_info, u64 flags,
  75                           int force);
  76 static int find_next_key(struct btrfs_path *path, int level,
  77                          struct btrfs_key *key);
  78 static void dump_space_info(struct btrfs_fs_info *fs_info,
  79                             struct btrfs_space_info *info, u64 bytes,
  80                             int dump_block_groups);
  81 static int block_rsv_use_bytes(struct btrfs_block_rsv *block_rsv,
  82                                u64 num_bytes);
  83 static void space_info_add_new_bytes(struct btrfs_fs_info *fs_info,
  84                                      struct btrfs_space_info *space_info,
  85                                      u64 num_bytes);
  86 static void space_info_add_old_bytes(struct btrfs_fs_info *fs_info,
  87                                      struct btrfs_space_info *space_info,
  88                                      u64 num_bytes);
  89
  90 static noinline int
  91 block_group_cache_done(struct btrfs_block_group_cache *cache)
  92 {
  93         smp_mb();
  94         return cache->cached == BTRFS_CACHE_FINISHED ||
  95                 cache->cached == BTRFS_CACHE_ERROR;
  96 }
  97
  98 static int block_group_bits(struct btrfs_block_group_cache *cache, u64 bits)
  99 {
 100         return (cache->flags & bits) == bits;
 101 }
 102
 103 void btrfs_get_block_group(struct btrfs_block_group_cache *cache)
 104 {
 105         atomic_inc(&cache->count);
 106 }
 107
 108 void btrfs_put_block_group(struct btrfs_block_group_cache *cache)
 109 {
 110         if (atomic_dec_and_test(&cache->count)) {
 111                 WARN_ON(cache->pinned > 0);
 112                 WARN_ON(cache->reserved > 0);
 113
 114                 /*
 115                  * If not empty, someone is still holding mutex of
 116                  * full_stripe_lock, which can only be released by caller.
 117                  * And it will definitely cause use-after-free when caller
 118                  * tries to release full stripe lock.
 119                  *
 120                  * No better way to resolve, but only to warn.
 121                  */
 122                 WARN_ON(!RB_EMPTY_ROOT(&cache->full_stripe_locks_root.root));
 123                 kfree(cache->free_space_ctl);
 124                 kfree(cache);
 125         }
 126 }
 127
 128 /*
 129  * this adds the block group to the fs_info rb tree for the block group
 130  * cache
 131  */
 132 static int btrfs_add_block_group_cache(struct btrfs_fs_info *info,
 133                                 struct btrfs_block_group_cache *block_group)
 134 {
 135         struct rb_node **p;
 136         struct rb_node *parent = NULL;
 137         struct btrfs_block_group_cache *cache;
 138
 139         spin_lock(&info->block_group_cache_lock);
 140         p = &info->block_group_cache_tree.rb_node;
 141
 142         while (*p) {
 143                 parent = *p;
 144                 cache = rb_entry(parent, struct btrfs_block_group_cache,
 145                                  cache_node);
 146                 if (block_group->key.objectid < cache->key.objectid) {
 147                         p = &(*p)->rb_left;
 148                 } else if (block_group->key.objectid > cache->key.objectid) {
 149                         p = &(*p)->rb_right;
 150                 } else {
 151                         spin_unlock(&info->block_group_cache_lock);
 152                         return -EEXIST;
 153                 }
 154         }
 155
 156         rb_link_node(&block_group->cache_node, parent, p);
 157         rb_insert_color(&block_group->cache_node,
 158                         &info->block_group_cache_tree);
 159
 160         if (info->first_logical_byte > block_group->key.objectid)
 161                 info->first_logical_byte = block_group->key.objectid;
 162
 163         spin_unlock(&info->block_group_cache_lock);
 164
 165         return 0;
 166 }
 167
 168 /*
 169  * This will return the block group at or after bytenr if contains is 0, else
 170  * it will return the block group that contains the bytenr
 171  */
 172 static struct btrfs_block_group_cache *
 173 block_group_cache_tree_search(struct btrfs_fs_info *info, u64 bytenr,
 174                               int contains)
 175 {
 176         struct btrfs_block_group_cache *cache, *ret = NULL;
 177         struct rb_node *n;
 178         u64 end, start;
 179
 180         spin_lock(&info->block_group_cache_lock);
 181         n = info->block_group_cache_tree.rb_node;
 182
 183         while (n) {
 184                 cache = rb_entry(n, struct btrfs_block_group_cache,
 185                                  cache_node);
 186                 end = cache->key.objectid + cache->key.offset - 1;
 187                 start = cache->key.objectid;
 188
 189                 if (bytenr < start) {
 190                         if (!contains && (!ret || start < ret->key.objectid))
 191                                 ret = cache;
 192                         n = n->rb_left;
 193                 } else if (bytenr > start) {
 194                         if (contains && bytenr <= end) {
 195                                 ret = cache;
 196                                 break;
 197                         }
 198                         n = n->rb_right;
 199                 } else {
 200                         ret = cache;
 201                         break;
 202                 }
 203         }
 204         if (ret) {
 205                 btrfs_get_block_group(ret);
 206                 if (bytenr == 0 && info->first_logical_byte > ret->key.objectid)
 207                         info->first_logical_byte = ret->key.objectid;
 208         }
 209         spin_unlock(&info->block_group_cache_lock);
 210
 211         return ret;
 212 }
 213
 214 static int add_excluded_extent(struct btrfs_fs_info *fs_info,
 215                                u64 start, u64 num_bytes)
 216 {
 217         u64 end = start + num_bytes - 1;
 218         set_extent_bits(&fs_info->freed_extents[0],
 219                         start, end, EXTENT_UPTODATE);
 220         set_extent_bits(&fs_info->freed_extents[1],
 221                         start, end, EXTENT_UPTODATE);
 222         return 0;
 223 }
 224
 225 static void free_excluded_extents(struct btrfs_fs_info *fs_info,
 226                                   struct btrfs_block_group_cache *cache)
 227 {
 228         u64 start, end;
 229
 230         start = cache->key.objectid;
 231         end = start + cache->key.offset - 1;
 232
 233         clear_extent_bits(&fs_info->freed_extents[0],
 234                           start, end, EXTENT_UPTODATE);
 235         clear_extent_bits(&fs_info->freed_extents[1],
 236                           start, end, EXTENT_UPTODATE);
 237 }
 238
 239 static int exclude_super_stripes(struct btrfs_fs_info *fs_info,
 240                                  struct btrfs_block_group_cache *cache)
 241 {
 242         u64 bytenr;
 243         u64 *logical;
 244         int stripe_len;
 245         int i, nr, ret;
 246
 247         if (cache->key.objectid < BTRFS_SUPER_INFO_OFFSET) {
 248                 stripe_len = BTRFS_SUPER_INFO_OFFSET - cache->key.objectid;
 249                 cache->bytes_super += stripe_len;
 250                 ret = add_excluded_extent(fs_info, cache->key.objectid,
 251                                           stripe_len);
 252                 if (ret)
 253                         return ret;
 254         }
 255
 256         for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
 257                 bytenr = btrfs_sb_offset(i);
 258                 ret = btrfs_rmap_block(fs_info, cache->key.objectid,
 259                                        bytenr, 0, &logical, &nr, &stripe_len);
 260                 if (ret)
 261                         return ret;
 262
 263                 while (nr--) {
 264                         u64 start, len;
 265
 266                         if (logical[nr] > cache->key.objectid +
 267                             cache->key.offset)
 268                                 continue;
 269
 270                         if (logical[nr] + stripe_len <= cache->key.objectid)
 271                                 continue;
 272
 273                         start = logical[nr];
 274                         if (start < cache->key.objectid) {
 275                                 start = cache->key.objectid;
 276                                 len = (logical[nr] + stripe_len) - start;
 277                         } else {
 278                                 len = min_t(u64, stripe_len,
 279                                             cache->key.objectid +
 280                                             cache->key.offset - start);
 281                         }
 282
 283                         cache->bytes_super += len;
 284                         ret = add_excluded_extent(fs_info, start, len);
 285                         if (ret) {
 286                                 kfree(logical);
 287                                 return ret;
 288                         }
 289                 }
 290
 291                 kfree(logical);
 292         }
 293         return 0;
 294 }
 295
 296 static struct btrfs_caching_control *
 297 get_caching_control(struct btrfs_block_group_cache *cache)
 298 {
 299         struct btrfs_caching_control *ctl;
 300
 301         spin_lock(&cache->lock);
 302         if (!cache->caching_ctl) {
 303                 spin_unlock(&cache->lock);
 304                 return NULL;
 305         }
 306
 307         ctl = cache->caching_ctl;
 308         refcount_inc(&ctl->count);
 309         spin_unlock(&cache->lock);
 310         return ctl;
 311 }
 312
 313 static void put_caching_control(struct btrfs_caching_control *ctl)
 314 {
 315         if (refcount_dec_and_test(&ctl->count))
 316                 kfree(ctl);
 317 }
 318
 319 #ifdef CONFIG_BTRFS_DEBUG
 320 static void fragment_free_space(struct btrfs_block_group_cache *block_group)
 321 {
 322         struct btrfs_fs_info *fs_info = block_group->fs_info;
 323         u64 start = block_group->key.objectid;
 324         u64 len = block_group->key.offset;
 325         u64 chunk = block_group->flags & BTRFS_BLOCK_GROUP_METADATA ?
 326                 fs_info->nodesize : fs_info->sectorsize;
 327         u64 step = chunk << 1;
 328
 329         while (len > chunk) {
 330                 btrfs_remove_free_space(block_group, start, chunk);
 331                 start += step;
 332                 if (len < step)
 333                         len = 0;
 334                 else
 335                         len -= step;
 336         }
 337 }
 338 #endif
 339
 340 /*
 341  * this is only called by cache_block_group, since we could have freed extents
 342  * we need to check the pinned_extents for any extents that can't be used yet
 343  * since their free space will be released as soon as the transaction commits.
 344  */
 345 u64 add_new_free_space(struct btrfs_block_group_cache *block_group,
 346                        struct btrfs_fs_info *info, u64 start, u64 end)
 347 {
 348         u64 extent_start, extent_end, size, total_added = 0;
 349         int ret;
 350
 351         while (start < end) {
 352                 ret = find_first_extent_bit(info->pinned_extents, start,
 353                                             &extent_start, &extent_end,
 354                                             EXTENT_DIRTY | EXTENT_UPTODATE,
 355                                             NULL);
 356                 if (ret)
 357                         break;
 358
 359                 if (extent_start <= start) {
 360                         start = extent_end + 1;
 361                 } else if (extent_start > start && extent_start < end) {
 362                         size = extent_start - start;
 363                         total_added += size;
 364                         ret = btrfs_add_free_space(block_group, start,
 365                                                    size);
 366                         BUG_ON(ret); /* -ENOMEM or logic error */
 367                         start = extent_end + 1;
 368                 } else {
 369                         break;
 370                 }
 371         }
 372
 373         if (start < end) {
 374                 size = end - start;
 375                 total_added += size;
 376                 ret = btrfs_add_free_space(block_group, start, size);
 377                 BUG_ON(ret); /* -ENOMEM or logic error */
 378         }
 379
 380         return total_added;
 381 }
 382
 383 static int load_extent_tree_free(struct btrfs_caching_control *caching_ctl)
 384 {
 385         struct btrfs_block_group_cache *block_group = caching_ctl->block_group;
 386         struct btrfs_fs_info *fs_info = block_group->fs_info;
 387         struct btrfs_root *extent_root = fs_info->extent_root;
 388         struct btrfs_path *path;
 389         struct extent_buffer *leaf;
 390         struct btrfs_key key;
 391         u64 total_found = 0;
 392         u64 last = 0;
 393         u32 nritems;
 394         int ret;
 395         bool wakeup = true;
 396
 397         path = btrfs_alloc_path();
 398         if (!path)
 399                 return -ENOMEM;
 400
 401         last = max_t(u64, block_group->key.objectid, BTRFS_SUPER_INFO_OFFSET);
 402
 403 #ifdef CONFIG_BTRFS_DEBUG
 404         /*
 405          * If we're fragmenting we don't want to make anybody think we can
 406          * allocate from this block group until we've had a chance to fragment
 407          * the free space.
 408          */
 409         if (btrfs_should_fragment_free_space(block_group))
 410                 wakeup = false;
 411 #endif
 412         /*
 413          * We don't want to deadlock with somebody trying to allocate a new
 414          * extent for the extent root while also trying to search the extent
 415          * root to add free space.  So we skip locking and search the commit
 416          * root, since its read-only
 417          */
 418         path->skip_locking = 1;
 419         path->search_commit_root = 1;
 420         path->reada = READA_FORWARD;
 421
 422         key.objectid = last;
 423         key.offset = 0;
 424         key.type = BTRFS_EXTENT_ITEM_KEY;
 425
 426 next:
 427         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
 428         if (ret < 0)
 429                 goto out;
 430
 431         leaf = path->nodes[0];
 432         nritems = btrfs_header_nritems(leaf);
 433
 434         while (1) {
 435                 if (btrfs_fs_closing(fs_info) > 1) {
 436                         last = (u64)-1;
 437                         break;
 438                 }
 439
 440                 if (path->slots[0] < nritems) {
 441                         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
 442                 } else {
 443                         ret = find_next_key(path, 0, &key);
 444                         if (ret)
 445                                 break;
 446
 447                         if (need_resched() ||
 448                             rwsem_is_contended(&fs_info->commit_root_sem)) {
 449                                 if (wakeup)
 450                                         caching_ctl->progress = last;
 451                                 btrfs_release_path(path);
 452                                 up_read(&fs_info->commit_root_sem);
 453                                 mutex_unlock(&caching_ctl->mutex);
 454                                 cond_resched();
 455                                 mutex_lock(&caching_ctl->mutex);
 456                                 down_read(&fs_info->commit_root_sem);
 457                                 goto next;
 458                         }
 459
 460                         ret = btrfs_next_leaf(extent_root, path);
 461                         if (ret < 0)
 462                                 goto out;
 463                         if (ret)
 464                                 break;
 465                         leaf = path->nodes[0];
 466                         nritems = btrfs_header_nritems(leaf);
 467                         continue;
 468                 }
 469
 470                 if (key.objectid < last) {
 471                         key.objectid = last;
 472                         key.offset = 0;
 473                         key.type = BTRFS_EXTENT_ITEM_KEY;
 474
 475                         if (wakeup)
 476                                 caching_ctl->progress = last;
 477                         btrfs_release_path(path);
 478                         goto next;
 479                 }
 480
 481                 if (key.objectid < block_group->key.objectid) {
 482                         path->slots[0]++;
 483                         continue;
 484                 }
 485
 486                 if (key.objectid >= block_group->key.objectid +
 487                     block_group->key.offset)
 488                         break;
 489
 490                 if (key.type == BTRFS_EXTENT_ITEM_KEY ||
 491                     key.type == BTRFS_METADATA_ITEM_KEY) {
 492                         total_found += add_new_free_space(block_group,
 493                                                           fs_info, last,
 494                                                           key.objectid);
 495                         if (key.type == BTRFS_METADATA_ITEM_KEY)
 496                                 last = key.objectid +
 497                                         fs_info->nodesize;
 498                         else
 499                                 last = key.objectid + key.offset;
 500
 501                         if (total_found > CACHING_CTL_WAKE_UP) {
 502                                 total_found = 0;
 503                                 if (wakeup)
 504                                         wake_up(&caching_ctl->wait);
 505                         }
 506                 }
 507                 path->slots[0]++;
 508         }
 509         ret = 0;
 510
 511         total_found += add_new_free_space(block_group, fs_info, last,
 512                                           block_group->key.objectid +
 513                                           block_group->key.offset);
 514         caching_ctl->progress = (u64)-1;
 515
 516 out:
 517         btrfs_free_path(path);
 518         return ret;
 519 }
 520
 521 static noinline void caching_thread(struct btrfs_work *work)
 522 {
 523         struct btrfs_block_group_cache *block_group;
 524         struct btrfs_fs_info *fs_info;
 525         struct btrfs_caching_control *caching_ctl;
 526         int ret;
 527
 528         caching_ctl = container_of(work, struct btrfs_caching_control, work);
 529         block_group = caching_ctl->block_group;
 530         fs_info = block_group->fs_info;
 531
 532         mutex_lock(&caching_ctl->mutex);
 533         down_read(&fs_info->commit_root_sem);
 534
 535         if (btrfs_fs_compat_ro(fs_info, FREE_SPACE_TREE))
 536                 ret = load_free_space_tree(caching_ctl);
 537         else
 538                 ret = load_extent_tree_free(caching_ctl);
 539
 540         spin_lock(&block_group->lock);
 541         block_group->caching_ctl = NULL;
 542         block_group->cached = ret ? BTRFS_CACHE_ERROR : BTRFS_CACHE_FINISHED;
 543         spin_unlock(&block_group->lock);
 544
 545 #ifdef CONFIG_BTRFS_DEBUG
 546         if (btrfs_should_fragment_free_space(block_group)) {
 547                 u64 bytes_used;
 548
 549                 spin_lock(&block_group->space_info->lock);
 550                 spin_lock(&block_group->lock);
 551                 bytes_used = block_group->key.offset -
 552                         btrfs_block_group_used(&block_group->item);
 553                 block_group->space_info->bytes_used += bytes_used >> 1;
 554                 spin_unlock(&block_group->lock);
 555                 spin_unlock(&block_group->space_info->lock);
 556                 fragment_free_space(block_group);
 557         }
 558 #endif
 559
 560         caching_ctl->progress = (u64)-1;
 561
 562         up_read(&fs_info->commit_root_sem);
 563         free_excluded_extents(fs_info, block_group);
 564         mutex_unlock(&caching_ctl->mutex);
 565
 566         wake_up(&caching_ctl->wait);
 567
 568         put_caching_control(caching_ctl);
 569         btrfs_put_block_group(block_group);
 570 }
 571
 572 static int cache_block_group(struct btrfs_block_group_cache *cache,
 573                              int load_cache_only)
 574 {
 575         DEFINE_WAIT(wait);
 576         struct btrfs_fs_info *fs_info = cache->fs_info;
 577         struct btrfs_caching_control *caching_ctl;
 578         int ret = 0;
 579
 580         caching_ctl = kzalloc(sizeof(*caching_ctl), GFP_NOFS);
 581         if (!caching_ctl)
 582                 return -ENOMEM;
 583
 584         INIT_LIST_HEAD(&caching_ctl->list);
 585         mutex_init(&caching_ctl->mutex);
 586         init_waitqueue_head(&caching_ctl->wait);
 587         caching_ctl->block_group = cache;
 588         caching_ctl->progress = cache->key.objectid;
 589         refcount_set(&caching_ctl->count, 1);
 590         btrfs_init_work(&caching_ctl->work, btrfs_cache_helper,
 591                         caching_thread, NULL, NULL);
 592
 593         spin_lock(&cache->lock);
 594         /*
 595          * This should be a rare occasion, but this could happen I think in the
 596          * case where one thread starts to load the space cache info, and then
 597          * some other thread starts a transaction commit which tries to do an
 598          * allocation while the other thread is still loading the space cache
 599          * info.  The previous loop should have kept us from choosing this block
 600          * group, but if we've moved to the state where we will wait on caching
 601          * block groups we need to first check if we're doing a fast load here,
 602          * so we can wait for it to finish, otherwise we could end up allocating
 603          * from a block group who's cache gets evicted for one reason or
 604          * another.
 605          */
 606         while (cache->cached == BTRFS_CACHE_FAST) {
 607                 struct btrfs_caching_control *ctl;
 608
 609                 ctl = cache->caching_ctl;
 610                 refcount_inc(&ctl->count);
 611                 prepare_to_wait(&ctl->wait, &wait, TASK_UNINTERRUPTIBLE);
 612                 spin_unlock(&cache->lock);
 613
 614                 schedule();
 615
 616                 finish_wait(&ctl->wait, &wait);
 617                 put_caching_control(ctl);
 618                 spin_lock(&cache->lock);
 619         }
 620
 621         if (cache->cached != BTRFS_CACHE_NO) {
 622                 spin_unlock(&cache->lock);
 623                 kfree(caching_ctl);
 624                 return 0;
 625         }
 626         WARN_ON(cache->caching_ctl);
 627         cache->caching_ctl = caching_ctl;
 628         cache->cached = BTRFS_CACHE_FAST;
 629         spin_unlock(&cache->lock);
 630
 631         if (btrfs_test_opt(fs_info, SPACE_CACHE)) {
 632                 mutex_lock(&caching_ctl->mutex);
 633                 ret = load_free_space_cache(fs_info, cache);
 634
 635                 spin_lock(&cache->lock);
 636                 if (ret == 1) {
 637                         cache->caching_ctl = NULL;
 638                         cache->cached = BTRFS_CACHE_FINISHED;
 639                         cache->last_byte_to_unpin = (u64)-1;
 640                         caching_ctl->progress = (u64)-1;
 641                 } else {
 642                         if (load_cache_only) {
 643                                 cache->caching_ctl = NULL;
 644                                 cache->cached = BTRFS_CACHE_NO;
 645                         } else {
 646                                 cache->cached = BTRFS_CACHE_STARTED;
 647                                 cache->has_caching_ctl = 1;
 648                         }
 649                 }
 650                 spin_unlock(&cache->lock);
 651 #ifdef CONFIG_BTRFS_DEBUG
 652                 if (ret == 1 &&
 653                     btrfs_should_fragment_free_space(cache)) {
 654                         u64 bytes_used;
 655
 656                         spin_lock(&cache->space_info->lock);
 657                         spin_lock(&cache->lock);
 658                         bytes_used = cache->key.offset -
 659                                 btrfs_block_group_used(&cache->item);
 660                         cache->space_info->bytes_used += bytes_used >> 1;
 661                         spin_unlock(&cache->lock);
 662                         spin_unlock(&cache->space_info->lock);
 663                         fragment_free_space(cache);
 664                 }
 665 #endif
 666                 mutex_unlock(&caching_ctl->mutex);
 667
 668                 wake_up(&caching_ctl->wait);
 669                 if (ret == 1) {
 670                         put_caching_control(caching_ctl);
 671                         free_excluded_extents(fs_info, cache);
 672                         return 0;
 673                 }
 674         } else {
 675                 /*
 676                  * We're either using the free space tree or no caching at all.
 677                  * Set cached to the appropriate value and wakeup any waiters.
 678                  */
 679                 spin_lock(&cache->lock);
 680                 if (load_cache_only) {
 681                         cache->caching_ctl = NULL;
 682                         cache->cached = BTRFS_CACHE_NO;
 683                 } else {
 684                         cache->cached = BTRFS_CACHE_STARTED;
 685                         cache->has_caching_ctl = 1;
 686                 }
 687                 spin_unlock(&cache->lock);
 688                 wake_up(&caching_ctl->wait);
 689         }
 690
 691         if (load_cache_only) {
 692                 put_caching_control(caching_ctl);
 693                 return 0;
 694         }
 695
 696         down_write(&fs_info->commit_root_sem);
 697         refcount_inc(&caching_ctl->count);
 698         list_add_tail(&caching_ctl->list, &fs_info->caching_block_groups);
 699         up_write(&fs_info->commit_root_sem);
 700
 701         btrfs_get_block_group(cache);
 702
 703         btrfs_queue_work(fs_info->caching_workers, &caching_ctl->work);
 704
 705         return ret;
 706 }
 707
 708 /*
 709  * return the block group that starts at or after bytenr
 710  */
 711 static struct btrfs_block_group_cache *
 712 btrfs_lookup_first_block_group(struct btrfs_fs_info *info, u64 bytenr)
 713 {
 714         return block_group_cache_tree_search(info, bytenr, 0);
 715 }
 716
 717 /*
 718  * return the block group that contains the given bytenr
 719  */
 720 struct btrfs_block_group_cache *btrfs_lookup_block_group(
 721                                                  struct btrfs_fs_info *info,
 722                                                  u64 bytenr)
 723 {
 724         return block_group_cache_tree_search(info, bytenr, 1);
 725 }
 726
 727 static struct btrfs_space_info *__find_space_info(struct btrfs_fs_info *info,
 728                                                   u64 flags)
 729 {
 730         struct list_head *head = &info->space_info;
 731         struct btrfs_space_info *found;
 732
 733         flags &= BTRFS_BLOCK_GROUP_TYPE_MASK;
 734
 735         rcu_read_lock();
 736         list_for_each_entry_rcu(found, head, list) {
 737                 if (found->flags & flags) {
 738                         rcu_read_unlock();
 739                         return found;
 740                 }
 741         }
 742         rcu_read_unlock();
 743         return NULL;
 744 }
 745
 746 static void add_pinned_bytes(struct btrfs_fs_info *fs_info, s64 num_bytes,
 747                              u64 owner, u64 root_objectid)
 748 {
 749         struct btrfs_space_info *space_info;
 750         u64 flags;
 751
 752         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
 753                 if (root_objectid == BTRFS_CHUNK_TREE_OBJECTID)
 754                         flags = BTRFS_BLOCK_GROUP_SYSTEM;
 755                 else
 756                         flags = BTRFS_BLOCK_GROUP_METADATA;
 757         } else {
 758                 flags = BTRFS_BLOCK_GROUP_DATA;
 759         }
 760
 761         space_info = __find_space_info(fs_info, flags);
 762         ASSERT(space_info);
 763         percpu_counter_add(&space_info->total_bytes_pinned, num_bytes);
 764 }
 765
 766 /*
 767  * after adding space to the filesystem, we need to clear the full flags
 768  * on all the space infos.
 769  */
 770 void btrfs_clear_space_info_full(struct btrfs_fs_info *info)
 771 {
 772         struct list_head *head = &info->space_info;
 773         struct btrfs_space_info *found;
 774
 775         rcu_read_lock();
 776         list_for_each_entry_rcu(found, head, list)
 777                 found->full = 0;
 778         rcu_read_unlock();
 779 }
 780
 781 /* simple helper to search for an existing data extent at a given offset */
 782 int btrfs_lookup_data_extent(struct btrfs_fs_info *fs_info, u64 start, u64 len)
 783 {
 784         int ret;
 785         struct btrfs_key key;
 786         struct btrfs_path *path;
 787
 788         path = btrfs_alloc_path();
 789         if (!path)
 790                 return -ENOMEM;
 791
 792         key.objectid = start;
 793         key.offset = len;
 794         key.type = BTRFS_EXTENT_ITEM_KEY;
 795         ret = btrfs_search_slot(NULL, fs_info->extent_root, &key, path, 0, 0);
 796         btrfs_free_path(path);
 797         return ret;
 798 }
 799
 800 /*
 801  * helper function to lookup reference count and flags of a tree block.
 802  *
 803  * the head node for delayed ref is used to store the sum of all the
 804  * reference count modifications queued up in the rbtree. the head
 805  * node may also store the extent flags to set. This way you can check
 806  * to see what the reference count and extent flags would be if all of
 807  * the delayed refs are not processed.
 808  */
 809 int btrfs_lookup_extent_info(struct btrfs_trans_handle *trans,
 810                              struct btrfs_fs_info *fs_info, u64 bytenr,
 811                              u64 offset, int metadata, u64 *refs, u64 *flags)
 812 {
 813         struct btrfs_delayed_ref_head *head;
 814         struct btrfs_delayed_ref_root *delayed_refs;
 815         struct btrfs_path *path;
 816         struct btrfs_extent_item *ei;
 817         struct extent_buffer *leaf;
 818         struct btrfs_key key;
 819         u32 item_size;
 820         u64 num_refs;
 821         u64 extent_flags;
 822         int ret;
 823
 824         /*
 825          * If we don't have skinny metadata, don't bother doing anything
 826          * different
 827          */
 828         if (metadata && !btrfs_fs_incompat(fs_info, SKINNY_METADATA)) {
 829                 offset = fs_info->nodesize;
 830                 metadata = 0;
 831         }
 832
 833         path = btrfs_alloc_path();
 834         if (!path)
 835                 return -ENOMEM;
 836
 837         if (!trans) {
 838                 path->skip_locking = 1;
 839                 path->search_commit_root = 1;
 840         }
 841
 842 search_again:
 843         key.objectid = bytenr;
 844         key.offset = offset;
 845         if (metadata)
 846                 key.type = BTRFS_METADATA_ITEM_KEY;
 847         else
 848                 key.type = BTRFS_EXTENT_ITEM_KEY;
 849
 850         ret = btrfs_search_slot(trans, fs_info->extent_root, &key, path, 0, 0);
 851         if (ret < 0)
 852                 goto out_free;
 853
 854         if (ret > 0 && metadata && key.type == BTRFS_METADATA_ITEM_KEY) {
 855                 if (path->slots[0]) {
 856                         path->slots[0]--;
 857                         btrfs_item_key_to_cpu(path->nodes[0], &key,
 858                                               path->slots[0]);
 859                         if (key.objectid == bytenr &&
 860                             key.type == BTRFS_EXTENT_ITEM_KEY &&
 861                             key.offset == fs_info->nodesize)
 862                                 ret = 0;
 863                 }
 864         }
 865
 866         if (ret == 0) {
 867                 leaf = path->nodes[0];
 868                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
 869                 if (item_size >= sizeof(*ei)) {
 870                         ei = btrfs_item_ptr(leaf, path->slots[0],
 871                                             struct btrfs_extent_item);
 872                         num_refs = btrfs_extent_refs(leaf, ei);
 873                         extent_flags = btrfs_extent_flags(leaf, ei);
 874                 } else {
 875 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
 876                         struct btrfs_extent_item_v0 *ei0;
 877                         BUG_ON(item_size != sizeof(*ei0));
 878                         ei0 = btrfs_item_ptr(leaf, path->slots[0],
 879                                              struct btrfs_extent_item_v0);
 880                         num_refs = btrfs_extent_refs_v0(leaf, ei0);
 881                         /* FIXME: this isn't correct for data */
 882                         extent_flags = BTRFS_BLOCK_FLAG_FULL_BACKREF;
 883 #else
 884                         BUG();
 885 #endif
 886                 }
 887                 BUG_ON(num_refs == 0);
 888         } else {
 889                 num_refs = 0;
 890                 extent_flags = 0;
 891                 ret = 0;
 892         }
 893
 894         if (!trans)
 895                 goto out;
 896
 897         delayed_refs = &trans->transaction->delayed_refs;
 898         spin_lock(&delayed_refs->lock);
 899         head = btrfs_find_delayed_ref_head(delayed_refs, bytenr);
 900         if (head) {
 901                 if (!mutex_trylock(&head->mutex)) {
 902                         refcount_inc(&head->refs);
 903                         spin_unlock(&delayed_refs->lock);
 904
 905                         btrfs_release_path(path);
 906
 907                         /*
 908                          * Mutex was contended, block until it's released and try
 909                          * again
 910                          */
 911                         mutex_lock(&head->mutex);
 912                         mutex_unlock(&head->mutex);
 913                         btrfs_put_delayed_ref_head(head);
 914                         goto search_again;
 915                 }
 916                 spin_lock(&head->lock);
 917                 if (head->extent_op && head->extent_op->update_flags)
 918                         extent_flags |= head->extent_op->flags_to_set;
 919                 else
 920                         BUG_ON(num_refs == 0);
 921
 922                 num_refs += head->ref_mod;
 923                 spin_unlock(&head->lock);
 924                 mutex_unlock(&head->mutex);
 925         }
 926         spin_unlock(&delayed_refs->lock);
 927 out:
 928         WARN_ON(num_refs == 0);
 929         if (refs)
 930                 *refs = num_refs;
 931         if (flags)
 932                 *flags = extent_flags;
 933 out_free:
 934         btrfs_free_path(path);
 935         return ret;
 936 }
 937
 938 /*
 939  * Back reference rules.  Back refs have three main goals:
 940  *
 941  * 1) differentiate between all holders of references to an extent so that
 942  *    when a reference is dropped we can make sure it was a valid reference
 943  *    before freeing the extent.
 944  *
 945  * 2) Provide enough information to quickly find the holders of an extent
 946  *    if we notice a given block is corrupted or bad.
 947  *
 948  * 3) Make it easy to migrate blocks for FS shrinking or storage pool
 949  *    maintenance.  This is actually the same as #2, but with a slightly
 950  *    different use case.
 951  *
 952  * There are two kinds of back refs. The implicit back refs is optimized
 953  * for pointers in non-shared tree blocks. For a given pointer in a block,
 954  * back refs of this kind provide information about the block's owner tree
 955  * and the pointer's key. These information allow us to find the block by
 956  * b-tree searching. The full back refs is for pointers in tree blocks not
 957  * referenced by their owner trees. The location of tree block is recorded
 958  * in the back refs. Actually the full back refs is generic, and can be
 959  * used in all cases the implicit back refs is used. The major shortcoming
 960  * of the full back refs is its overhead. Every time a tree block gets
 961  * COWed, we have to update back refs entry for all pointers in it.
 962  *
 963  * For a newly allocated tree block, we use implicit back refs for
 964  * pointers in it. This means most tree related operations only involve
 965  * implicit back refs. For a tree block created in old transaction, the
 966  * only way to drop a reference to it is COW it. So we can detect the
 967  * event that tree block loses its owner tree's reference and do the
 968  * back refs conversion.
 969  *
 970  * When a tree block is COWed through a tree, there are four cases:
 971  *
 972  * The reference count of the block is one and the tree is the block's
 973  * owner tree. Nothing to do in this case.
 974  *
 975  * The reference count of the block is one and the tree is not the
 976  * block's owner tree. In this case, full back refs is used for pointers
 977  * in the block. Remove these full back refs, add implicit back refs for
 978  * every pointers in the new block.
 979  *
 980  * The reference count of the block is greater than one and the tree is
 981  * the block's owner tree. In this case, implicit back refs is used for
 982  * pointers in the block. Add full back refs for every pointers in the
 983  * block, increase lower level extents' reference counts. The original
 984  * implicit back refs are entailed to the new block.
 985  *
 986  * The reference count of the block is greater than one and the tree is
 987  * not the block's owner tree. Add implicit back refs for every pointer in
 988  * the new block, increase lower level extents' reference count.
 989  *
 990  * Back Reference Key composing:
 991  *
 992  * The key objectid corresponds to the first byte in the extent,
 993  * The key type is used to differentiate between types of back refs.
 994  * There are different meanings of the key offset for different types
 995  * of back refs.
 996  *
 997  * File extents can be referenced by:
 998  *
 999  * - multiple snapshots, subvolumes, or different generations in one subvol
1000  * - different files inside a single subvolume
1001  * - different offsets inside a file (bookend extents in file.c)
1002  *
1003  * The extent ref structure for the implicit back refs has fields for:
1004  *
1005  * - Objectid of the subvolume root
1006  * - objectid of the file holding the reference
1007  * - original offset in the file
1008  * - how many bookend extents
1009  *
1010  * The key offset for the implicit back refs is hash of the first
1011  * three fields.
1012  *
1013  * The extent ref structure for the full back refs has field for:
1014  *
1015  * - number of pointers in the tree leaf
1016  *
1017  * The key offset for the implicit back refs is the first byte of
1018  * the tree leaf
1019  *
1020  * When a file extent is allocated, The implicit back refs is used.
1021  * the fields are filled in:
1022  *
1023  *     (root_key.objectid, inode objectid, offset in file, 1)
1024  *
1025  * When a file extent is removed file truncation, we find the
1026  * corresponding implicit back refs and check the following fields:
1027  *
1028  *     (btrfs_header_owner(leaf), inode objectid, offset in file)
1029  *
1030  * Btree extents can be referenced by:
1031  *
1032  * - Different subvolumes
1033  *
1034  * Both the implicit back refs and the full back refs for tree blocks
1035  * only consist of key. The key offset for the implicit back refs is
1036  * objectid of block's owner tree. The key offset for the full back refs
1037  * is the first byte of parent block.
1038  *
1039  * When implicit back refs is used, information about the lowest key and
1040  * level of the tree block are required. These information are stored in
1041  * tree block info structure.
1042  */
1043
1044 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1045 static int convert_extent_item_v0(struct btrfs_trans_handle *trans,
1046                                   struct btrfs_fs_info *fs_info,
1047                                   struct btrfs_path *path,
1048                                   u64 owner, u32 extra_size)
1049 {
1050         struct btrfs_root *root = fs_info->extent_root;
1051         struct btrfs_extent_item *item;
1052         struct btrfs_extent_item_v0 *ei0;
1053         struct btrfs_extent_ref_v0 *ref0;
1054         struct btrfs_tree_block_info *bi;
1055         struct extent_buffer *leaf;
1056         struct btrfs_key key;
1057         struct btrfs_key found_key;
1058         u32 new_size = sizeof(*item);
1059         u64 refs;
1060         int ret;
1061
1062         leaf = path->nodes[0];
1063         BUG_ON(btrfs_item_size_nr(leaf, path->slots[0]) != sizeof(*ei0));
1064
1065         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1066         ei0 = btrfs_item_ptr(leaf, path->slots[0],
1067                              struct btrfs_extent_item_v0);
1068         refs = btrfs_extent_refs_v0(leaf, ei0);
1069
1070         if (owner == (u64)-1) {
1071                 while (1) {
1072                         if (path->slots[0] >= btrfs_header_nritems(leaf)) {
1073                                 ret = btrfs_next_leaf(root, path);
1074                                 if (ret < 0)
1075                                         return ret;
1076                                 BUG_ON(ret > 0); /* Corruption */
1077                                 leaf = path->nodes[0];
1078                         }
1079                         btrfs_item_key_to_cpu(leaf, &found_key,
1080                                               path->slots[0]);
1081                         BUG_ON(key.objectid != found_key.objectid);
1082                         if (found_key.type != BTRFS_EXTENT_REF_V0_KEY) {
1083                                 path->slots[0]++;
1084                                 continue;
1085                         }
1086                         ref0 = btrfs_item_ptr(leaf, path->slots[0],
1087                                               struct btrfs_extent_ref_v0);
1088                         owner = btrfs_ref_objectid_v0(leaf, ref0);
1089                         break;
1090                 }
1091         }
1092         btrfs_release_path(path);
1093
1094         if (owner < BTRFS_FIRST_FREE_OBJECTID)
1095                 new_size += sizeof(*bi);
1096
1097         new_size -= sizeof(*ei0);
1098         ret = btrfs_search_slot(trans, root, &key, path,
1099                                 new_size + extra_size, 1);
1100         if (ret < 0)
1101                 return ret;
1102         BUG_ON(ret); /* Corruption */
1103
1104         btrfs_extend_item(fs_info, path, new_size);
1105
1106         leaf = path->nodes[0];
1107         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1108         btrfs_set_extent_refs(leaf, item, refs);
1109         /* FIXME: get real generation */
1110         btrfs_set_extent_generation(leaf, item, 0);
1111         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1112                 btrfs_set_extent_flags(leaf, item,
1113                                        BTRFS_EXTENT_FLAG_TREE_BLOCK |
1114                                        BTRFS_BLOCK_FLAG_FULL_BACKREF);
1115                 bi = (struct btrfs_tree_block_info *)(item + 1);
1116                 /* FIXME: get first key of the block */
1117                 memzero_extent_buffer(leaf, (unsigned long)bi, sizeof(*bi));
1118                 btrfs_set_tree_block_level(leaf, bi, (int)owner);
1119         } else {
1120                 btrfs_set_extent_flags(leaf, item, BTRFS_EXTENT_FLAG_DATA);
1121         }
1122         btrfs_mark_buffer_dirty(leaf);
1123         return 0;
1124 }
1125 #endif
1126
1127 /*
1128  * is_data == BTRFS_REF_TYPE_BLOCK, tree block type is required,
1129  * is_data == BTRFS_REF_TYPE_DATA, data type is requried,
1130  * is_data == BTRFS_REF_TYPE_ANY, either type is OK.
1131  */
1132 int btrfs_get_extent_inline_ref_type(const struct extent_buffer *eb,
1133                                      struct btrfs_extent_inline_ref *iref,
1134                                      enum btrfs_inline_ref_type is_data)
1135 {
1136         int type = btrfs_extent_inline_ref_type(eb, iref);
1137         u64 offset = btrfs_extent_inline_ref_offset(eb, iref);
1138
1139         if (type == BTRFS_TREE_BLOCK_REF_KEY ||
1140             type == BTRFS_SHARED_BLOCK_REF_KEY ||
1141             type == BTRFS_SHARED_DATA_REF_KEY ||
1142             type == BTRFS_EXTENT_DATA_REF_KEY) {
1143                 if (is_data == BTRFS_REF_TYPE_BLOCK) {
1144                         if (type == BTRFS_TREE_BLOCK_REF_KEY)
1145                                 return type;
1146                         if (type == BTRFS_SHARED_BLOCK_REF_KEY) {
1147                                 ASSERT(eb->fs_info);
1148                                 /*
1149                                  * Every shared one has parent tree
1150                                  * block, which must be aligned to
1151                                  * nodesize.
1152                                  */
1153                                 if (offset &&
1154                                     IS_ALIGNED(offset, eb->fs_info->nodesize))
1155                                         return type;
1156                         }
1157                 } else if (is_data == BTRFS_REF_TYPE_DATA) {
1158                         if (type == BTRFS_EXTENT_DATA_REF_KEY)
1159                                 return type;
1160                         if (type == BTRFS_SHARED_DATA_REF_KEY) {
1161                                 ASSERT(eb->fs_info);
1162                                 /*
1163                                  * Every shared one has parent tree
1164                                  * block, which must be aligned to
1165                                  * nodesize.
1166                                  */
1167                                 if (offset &&
1168                                     IS_ALIGNED(offset, eb->fs_info->nodesize))
1169                                         return type;
1170                         }
1171                 } else {
1172                         ASSERT(is_data == BTRFS_REF_TYPE_ANY);
1173                         return type;
1174                 }
1175         }
1176
1177         btrfs_print_leaf((struct extent_buffer *)eb);
1178         btrfs_err(eb->fs_info, "eb %llu invalid extent inline ref type %d",
1179                   eb->start, type);
1180         WARN_ON(1);
1181
1182         return BTRFS_REF_TYPE_INVALID;
1183 }
1184
1185 static u64 hash_extent_data_ref(u64 root_objectid, u64 owner, u64 offset)
1186 {
1187         u32 high_crc = ~(u32)0;
1188         u32 low_crc = ~(u32)0;
1189         __le64 lenum;
1190
1191         lenum = cpu_to_le64(root_objectid);
1192         high_crc = crc32c(high_crc, &lenum, sizeof(lenum));
1193         lenum = cpu_to_le64(owner);
1194         low_crc = crc32c(low_crc, &lenum, sizeof(lenum));
1195         lenum = cpu_to_le64(offset);
1196         low_crc = crc32c(low_crc, &lenum, sizeof(lenum));
1197
1198         return ((u64)high_crc << 31) ^ (u64)low_crc;
1199 }
1200
1201 static u64 hash_extent_data_ref_item(struct extent_buffer *leaf,
1202                                      struct btrfs_extent_data_ref *ref)
1203 {
1204         return hash_extent_data_ref(btrfs_extent_data_ref_root(leaf, ref),
1205                                     btrfs_extent_data_ref_objectid(leaf, ref),
1206                                     btrfs_extent_data_ref_offset(leaf, ref));
1207 }
1208
1209 static int match_extent_data_ref(struct extent_buffer *leaf,
1210                                  struct btrfs_extent_data_ref *ref,
1211                                  u64 root_objectid, u64 owner, u64 offset)
1212 {
1213         if (btrfs_extent_data_ref_root(leaf, ref) != root_objectid ||
1214             btrfs_extent_data_ref_objectid(leaf, ref) != owner ||
1215             btrfs_extent_data_ref_offset(leaf, ref) != offset)
1216                 return 0;
1217         return 1;
1218 }
1219
1220 static noinline int lookup_extent_data_ref(struct btrfs_trans_handle *trans,
1221                                            struct btrfs_fs_info *fs_info,
1222                                            struct btrfs_path *path,
1223                                            u64 bytenr, u64 parent,
1224                                            u64 root_objectid,
1225                                            u64 owner, u64 offset)
1226 {
1227         struct btrfs_root *root = fs_info->extent_root;
1228         struct btrfs_key key;
1229         struct btrfs_extent_data_ref *ref;
1230         struct extent_buffer *leaf;
1231         u32 nritems;
1232         int ret;
1233         int recow;
1234         int err = -ENOENT;
1235
1236         key.objectid = bytenr;
1237         if (parent) {
1238                 key.type = BTRFS_SHARED_DATA_REF_KEY;
1239                 key.offset = parent;
1240         } else {
1241                 key.type = BTRFS_EXTENT_DATA_REF_KEY;
1242                 key.offset = hash_extent_data_ref(root_objectid,
1243                                                   owner, offset);
1244         }
1245 again:
1246         recow = 0;
1247         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1248         if (ret < 0) {
1249                 err = ret;
1250                 goto fail;
1251         }
1252
1253         if (parent) {
1254                 if (!ret)
1255                         return 0;
1256 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1257                 key.type = BTRFS_EXTENT_REF_V0_KEY;
1258                 btrfs_release_path(path);
1259                 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1260                 if (ret < 0) {
1261                         err = ret;
1262                         goto fail;
1263                 }
1264                 if (!ret)
1265                         return 0;
1266 #endif
1267                 goto fail;
1268         }
1269
1270         leaf = path->nodes[0];
1271         nritems = btrfs_header_nritems(leaf);
1272         while (1) {
1273                 if (path->slots[0] >= nritems) {
1274                         ret = btrfs_next_leaf(root, path);
1275                         if (ret < 0)
1276                                 err = ret;
1277                         if (ret)
1278                                 goto fail;
1279
1280                         leaf = path->nodes[0];
1281                         nritems = btrfs_header_nritems(leaf);
1282                         recow = 1;
1283                 }
1284
1285                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1286                 if (key.objectid != bytenr ||
1287                     key.type != BTRFS_EXTENT_DATA_REF_KEY)
1288                         goto fail;
1289
1290                 ref = btrfs_item_ptr(leaf, path->slots[0],
1291                                      struct btrfs_extent_data_ref);
1292
1293                 if (match_extent_data_ref(leaf, ref, root_objectid,
1294                                           owner, offset)) {
1295                         if (recow) {
1296                                 btrfs_release_path(path);
1297                                 goto again;
1298                         }
1299                         err = 0;
1300                         break;
1301                 }
1302                 path->slots[0]++;
1303         }
1304 fail:
1305         return err;
1306 }
1307
1308 static noinline int insert_extent_data_ref(struct btrfs_trans_handle *trans,
1309                                            struct btrfs_fs_info *fs_info,
1310                                            struct btrfs_path *path,
1311                                            u64 bytenr, u64 parent,
1312                                            u64 root_objectid, u64 owner,
1313                                            u64 offset, int refs_to_add)
1314 {
1315         struct btrfs_root *root = fs_info->extent_root;
1316         struct btrfs_key key;
1317         struct extent_buffer *leaf;
1318         u32 size;
1319         u32 num_refs;
1320         int ret;
1321
1322         key.objectid = bytenr;
1323         if (parent) {
1324                 key.type = BTRFS_SHARED_DATA_REF_KEY;
1325                 key.offset = parent;
1326                 size = sizeof(struct btrfs_shared_data_ref);
1327         } else {
1328                 key.type = BTRFS_EXTENT_DATA_REF_KEY;
1329                 key.offset = hash_extent_data_ref(root_objectid,
1330                                                   owner, offset);
1331                 size = sizeof(struct btrfs_extent_data_ref);
1332         }
1333
1334         ret = btrfs_insert_empty_item(trans, root, path, &key, size);
1335         if (ret && ret != -EEXIST)
1336                 goto fail;
1337
1338         leaf = path->nodes[0];
1339         if (parent) {
1340                 struct btrfs_shared_data_ref *ref;
1341                 ref = btrfs_item_ptr(leaf, path->slots[0],
1342                                      struct btrfs_shared_data_ref);
1343                 if (ret == 0) {
1344                         btrfs_set_shared_data_ref_count(leaf, ref, refs_to_add);
1345                 } else {
1346                         num_refs = btrfs_shared_data_ref_count(leaf, ref);
1347                         num_refs += refs_to_add;
1348                         btrfs_set_shared_data_ref_count(leaf, ref, num_refs);
1349                 }
1350         } else {
1351                 struct btrfs_extent_data_ref *ref;
1352                 while (ret == -EEXIST) {
1353                         ref = btrfs_item_ptr(leaf, path->slots[0],
1354                                              struct btrfs_extent_data_ref);
1355                         if (match_extent_data_ref(leaf, ref, root_objectid,
1356                                                   owner, offset))
1357                                 break;
1358                         btrfs_release_path(path);
1359                         key.offset++;
1360                         ret = btrfs_insert_empty_item(trans, root, path, &key,
1361                                                       size);
1362                         if (ret && ret != -EEXIST)
1363                                 goto fail;
1364
1365                         leaf = path->nodes[0];
1366                 }
1367                 ref = btrfs_item_ptr(leaf, path->slots[0],
1368                                      struct btrfs_extent_data_ref);
1369                 if (ret == 0) {
1370                         btrfs_set_extent_data_ref_root(leaf, ref,
1371                                                        root_objectid);
1372                         btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
1373                         btrfs_set_extent_data_ref_offset(leaf, ref, offset);
1374                         btrfs_set_extent_data_ref_count(leaf, ref, refs_to_add);
1375                 } else {
1376                         num_refs = btrfs_extent_data_ref_count(leaf, ref);
1377                         num_refs += refs_to_add;
1378                         btrfs_set_extent_data_ref_count(leaf, ref, num_refs);
1379                 }
1380         }
1381         btrfs_mark_buffer_dirty(leaf);
1382         ret = 0;
1383 fail:
1384         btrfs_release_path(path);
1385         return ret;
1386 }
1387
1388 static noinline int remove_extent_data_ref(struct btrfs_trans_handle *trans,
1389                                            struct btrfs_fs_info *fs_info,
1390                                            struct btrfs_path *path,
1391                                            int refs_to_drop, int *last_ref)
1392 {
1393         struct btrfs_key key;
1394         struct btrfs_extent_data_ref *ref1 = NULL;
1395         struct btrfs_shared_data_ref *ref2 = NULL;
1396         struct extent_buffer *leaf;
1397         u32 num_refs = 0;
1398         int ret = 0;
1399
1400         leaf = path->nodes[0];
1401         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1402
1403         if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
1404                 ref1 = btrfs_item_ptr(leaf, path->slots[0],
1405                                       struct btrfs_extent_data_ref);
1406                 num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1407         } else if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
1408                 ref2 = btrfs_item_ptr(leaf, path->slots[0],
1409                                       struct btrfs_shared_data_ref);
1410                 num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1411 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1412         } else if (key.type == BTRFS_EXTENT_REF_V0_KEY) {
1413                 struct btrfs_extent_ref_v0 *ref0;
1414                 ref0 = btrfs_item_ptr(leaf, path->slots[0],
1415                                       struct btrfs_extent_ref_v0);
1416                 num_refs = btrfs_ref_count_v0(leaf, ref0);
1417 #endif
1418         } else {
1419                 BUG();
1420         }
1421
1422         BUG_ON(num_refs < refs_to_drop);
1423         num_refs -= refs_to_drop;
1424
1425         if (num_refs == 0) {
1426                 ret = btrfs_del_item(trans, fs_info->extent_root, path);
1427                 *last_ref = 1;
1428         } else {
1429                 if (key.type == BTRFS_EXTENT_DATA_REF_KEY)
1430                         btrfs_set_extent_data_ref_count(leaf, ref1, num_refs);
1431                 else if (key.type == BTRFS_SHARED_DATA_REF_KEY)
1432                         btrfs_set_shared_data_ref_count(leaf, ref2, num_refs);
1433 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1434                 else {
1435                         struct btrfs_extent_ref_v0 *ref0;
1436                         ref0 = btrfs_item_ptr(leaf, path->slots[0],
1437                                         struct btrfs_extent_ref_v0);
1438                         btrfs_set_ref_count_v0(leaf, ref0, num_refs);
1439                 }
1440 #endif
1441                 btrfs_mark_buffer_dirty(leaf);
1442         }
1443         return ret;
1444 }
1445
1446 static noinline u32 extent_data_ref_count(struct btrfs_path *path,
1447                                           struct btrfs_extent_inline_ref *iref)
1448 {
1449         struct btrfs_key key;
1450         struct extent_buffer *leaf;
1451         struct btrfs_extent_data_ref *ref1;
1452         struct btrfs_shared_data_ref *ref2;
1453         u32 num_refs = 0;
1454         int type;
1455
1456         leaf = path->nodes[0];
1457         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1458         if (iref) {
1459                 /*
1460                  * If type is invalid, we should have bailed out earlier than
1461                  * this call.
1462                  */
1463                 type = btrfs_get_extent_inline_ref_type(leaf, iref, BTRFS_REF_TYPE_DATA);
1464                 ASSERT(type != BTRFS_REF_TYPE_INVALID);
1465                 if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1466                         ref1 = (struct btrfs_extent_data_ref *)(&iref->offset);
1467                         num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1468                 } else {
1469                         ref2 = (struct btrfs_shared_data_ref *)(iref + 1);
1470                         num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1471                 }
1472         } else if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
1473                 ref1 = btrfs_item_ptr(leaf, path->slots[0],
1474                                       struct btrfs_extent_data_ref);
1475                 num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1476         } else if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
1477                 ref2 = btrfs_item_ptr(leaf, path->slots[0],
1478                                       struct btrfs_shared_data_ref);
1479                 num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1480 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1481         } else if (key.type == BTRFS_EXTENT_REF_V0_KEY) {
1482                 struct btrfs_extent_ref_v0 *ref0;
1483                 ref0 = btrfs_item_ptr(leaf, path->slots[0],
1484                                       struct btrfs_extent_ref_v0);
1485                 num_refs = btrfs_ref_count_v0(leaf, ref0);
1486 #endif
1487         } else {
1488                 WARN_ON(1);
1489         }
1490         return num_refs;
1491 }
1492
1493 static noinline int lookup_tree_block_ref(struct btrfs_trans_handle *trans,
1494                                           struct btrfs_fs_info *fs_info,
1495                                           struct btrfs_path *path,
1496                                           u64 bytenr, u64 parent,
1497                                           u64 root_objectid)
1498 {
1499         struct btrfs_root *root = fs_info->extent_root;
1500         struct btrfs_key key;
1501         int ret;
1502
1503         key.objectid = bytenr;
1504         if (parent) {
1505                 key.type = BTRFS_SHARED_BLOCK_REF_KEY;
1506                 key.offset = parent;
1507         } else {
1508                 key.type = BTRFS_TREE_BLOCK_REF_KEY;
1509                 key.offset = root_objectid;
1510         }
1511
1512         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1513         if (ret > 0)
1514                 ret = -ENOENT;
1515 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1516         if (ret == -ENOENT && parent) {
1517                 btrfs_release_path(path);
1518                 key.type = BTRFS_EXTENT_REF_V0_KEY;
1519                 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1520                 if (ret > 0)
1521                         ret = -ENOENT;
1522         }
1523 #endif
1524         return ret;
1525 }
1526
1527 static noinline int insert_tree_block_ref(struct btrfs_trans_handle *trans,
1528                                           struct btrfs_fs_info *fs_info,
1529                                           struct btrfs_path *path,
1530                                           u64 bytenr, u64 parent,
1531                                           u64 root_objectid)
1532 {
1533         struct btrfs_key key;
1534         int ret;
1535
1536         key.objectid = bytenr;
1537         if (parent) {
1538                 key.type = BTRFS_SHARED_BLOCK_REF_KEY;
1539                 key.offset = parent;
1540         } else {
1541                 key.type = BTRFS_TREE_BLOCK_REF_KEY;
1542                 key.offset = root_objectid;
1543         }
1544
1545         ret = btrfs_insert_empty_item(trans, fs_info->extent_root,
1546                                       path, &key, 0);
1547         btrfs_release_path(path);
1548         return ret;
1549 }
1550
1551 static inline int extent_ref_type(u64 parent, u64 owner)
1552 {
1553         int type;
1554         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1555                 if (parent > 0)
1556                         type = BTRFS_SHARED_BLOCK_REF_KEY;
1557                 else
1558                         type = BTRFS_TREE_BLOCK_REF_KEY;
1559         } else {
1560                 if (parent > 0)
1561                         type = BTRFS_SHARED_DATA_REF_KEY;
1562                 else
1563                         type = BTRFS_EXTENT_DATA_REF_KEY;
1564         }
1565         return type;
1566 }
1567
1568 static int find_next_key(struct btrfs_path *path, int level,
1569                          struct btrfs_key *key)
1570
1571 {
1572         for (; level < BTRFS_MAX_LEVEL; level++) {
1573                 if (!path->nodes[level])
1574                         break;
1575                 if (path->slots[level] + 1 >=
1576                     btrfs_header_nritems(path->nodes[level]))
1577                         continue;
1578                 if (level == 0)
1579                         btrfs_item_key_to_cpu(path->nodes[level], key,
1580                                               path->slots[level] + 1);
1581                 else
1582                         btrfs_node_key_to_cpu(path->nodes[level], key,
1583                                               path->slots[level] + 1);
1584                 return 0;
1585         }
1586         return 1;
1587 }
1588
1589 /*
1590  * look for inline back ref. if back ref is found, *ref_ret is set
1591  * to the address of inline back ref, and 0 is returned.
1592  *
1593  * if back ref isn't found, *ref_ret is set to the address where it
1594  * should be inserted, and -ENOENT is returned.
1595  *
1596  * if insert is true and there are too many inline back refs, the path
1597  * points to the extent item, and -EAGAIN is returned.
1598  *
1599  * NOTE: inline back refs are ordered in the same way that back ref
1600  *       items in the tree are ordered.
1601  */
1602 static noinline_for_stack
1603 int lookup_inline_extent_backref(struct btrfs_trans_handle *trans,
1604                                  struct btrfs_fs_info *fs_info,
1605                                  struct btrfs_path *path,
1606                                  struct btrfs_extent_inline_ref **ref_ret,
1607                                  u64 bytenr, u64 num_bytes,
1608                                  u64 parent, u64 root_objectid,
1609                                  u64 owner, u64 offset, int insert)
1610 {
1611         struct btrfs_root *root = fs_info->extent_root;
1612         struct btrfs_key key;
1613         struct extent_buffer *leaf;
1614         struct btrfs_extent_item *ei;
1615         struct btrfs_extent_inline_ref *iref;
1616         u64 flags;
1617         u64 item_size;
1618         unsigned long ptr;
1619         unsigned long end;
1620         int extra_size;
1621         int type;
1622         int want;
1623         int ret;
1624         int err = 0;
1625         bool skinny_metadata = btrfs_fs_incompat(fs_info, SKINNY_METADATA);
1626         int needed;
1627
1628         key.objectid = bytenr;
1629         key.type = BTRFS_EXTENT_ITEM_KEY;
1630         key.offset = num_bytes;
1631
1632         want = extent_ref_type(parent, owner);
1633         if (insert) {
1634                 extra_size = btrfs_extent_inline_ref_size(want);
1635                 path->keep_locks = 1;
1636         } else
1637                 extra_size = -1;
1638
1639         /*
1640          * Owner is our parent level, so we can just add one to get the level
1641          * for the block we are interested in.
1642          */
1643         if (skinny_metadata && owner < BTRFS_FIRST_FREE_OBJECTID) {
1644                 key.type = BTRFS_METADATA_ITEM_KEY;
1645                 key.offset = owner;
1646         }
1647
1648 again:
1649         ret = btrfs_search_slot(trans, root, &key, path, extra_size, 1);
1650         if (ret < 0) {
1651                 err = ret;
1652                 goto out;
1653         }
1654
1655         /*
1656          * We may be a newly converted file system which still has the old fat
1657          * extent entries for metadata, so try and see if we have one of those.
1658          */
1659         if (ret > 0 && skinny_metadata) {
1660                 skinny_metadata = false;
1661                 if (path->slots[0]) {
1662                         path->slots[0]--;
1663                         btrfs_item_key_to_cpu(path->nodes[0], &key,
1664                                               path->slots[0]);
1665                         if (key.objectid == bytenr &&
1666                             key.type == BTRFS_EXTENT_ITEM_KEY &&
1667                             key.offset == num_bytes)
1668                                 ret = 0;
1669                 }
1670                 if (ret) {
1671                         key.objectid = bytenr;
1672                         key.type = BTRFS_EXTENT_ITEM_KEY;
1673                         key.offset = num_bytes;
1674                         btrfs_release_path(path);
1675                         goto again;
1676                 }
1677         }
1678
1679         if (ret && !insert) {
1680                 err = -ENOENT;
1681                 goto out;
1682         } else if (WARN_ON(ret)) {
1683                 err = -EIO;
1684                 goto out;
1685         }
1686
1687         leaf = path->nodes[0];
1688         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1689 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1690         if (item_size < sizeof(*ei)) {
1691                 if (!insert) {
1692                         err = -ENOENT;
1693                         goto out;
1694                 }
1695                 ret = convert_extent_item_v0(trans, fs_info, path, owner,
1696                                              extra_size);
1697                 if (ret < 0) {
1698                         err = ret;
1699                         goto out;
1700                 }
1701                 leaf = path->nodes[0];
1702                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1703         }
1704 #endif
1705         BUG_ON(item_size < sizeof(*ei));
1706
1707         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1708         flags = btrfs_extent_flags(leaf, ei);
1709
1710         ptr = (unsigned long)(ei + 1);
1711         end = (unsigned long)ei + item_size;
1712
1713         if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK && !skinny_metadata) {
1714                 ptr += sizeof(struct btrfs_tree_block_info);
1715                 BUG_ON(ptr > end);
1716         }
1717
1718         if (owner >= BTRFS_FIRST_FREE_OBJECTID)
1719                 needed = BTRFS_REF_TYPE_DATA;
1720         else
1721                 needed = BTRFS_REF_TYPE_BLOCK;
1722
1723         err = -ENOENT;
1724         while (1) {
1725                 if (ptr >= end) {
1726                         WARN_ON(ptr > end);
1727                         break;
1728                 }
1729                 iref = (struct btrfs_extent_inline_ref *)ptr;
1730                 type = btrfs_get_extent_inline_ref_type(leaf, iref, needed);
1731                 if (type == BTRFS_REF_TYPE_INVALID) {
1732                         err = -EINVAL;
1733                         goto out;
1734                 }
1735
1736                 if (want < type)
1737                         break;
1738                 if (want > type) {
1739                         ptr += btrfs_extent_inline_ref_size(type);
1740                         continue;
1741                 }
1742
1743                 if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1744                         struct btrfs_extent_data_ref *dref;
1745                         dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1746                         if (match_extent_data_ref(leaf, dref, root_objectid,
1747                                                   owner, offset)) {
1748                                 err = 0;
1749                                 break;
1750                         }
1751                         if (hash_extent_data_ref_item(leaf, dref) <
1752                             hash_extent_data_ref(root_objectid, owner, offset))
1753                                 break;
1754                 } else {
1755                         u64 ref_offset;
1756                         ref_offset = btrfs_extent_inline_ref_offset(leaf, iref);
1757                         if (parent > 0) {
1758                                 if (parent == ref_offset) {
1759                                         err = 0;
1760                                         break;
1761                                 }
1762                                 if (ref_offset < parent)
1763                                         break;
1764                         } else {
1765                                 if (root_objectid == ref_offset) {
1766                                         err = 0;
1767                                         break;
1768                                 }
1769                                 if (ref_offset < root_objectid)
1770                                         break;
1771                         }
1772                 }
1773                 ptr += btrfs_extent_inline_ref_size(type);
1774         }
1775         if (err == -ENOENT && insert) {
1776                 if (item_size + extra_size >=
1777                     BTRFS_MAX_EXTENT_ITEM_SIZE(root)) {
1778                         err = -EAGAIN;
1779                         goto out;
1780                 }
1781                 /*
1782                  * To add new inline back ref, we have to make sure
1783                  * there is no corresponding back ref item.
1784                  * For simplicity, we just do not add new inline back
1785                  * ref if there is any kind of item for this block
1786                  */
1787                 if (find_next_key(path, 0, &key) == 0 &&
1788                     key.objectid == bytenr &&
1789                     key.type < BTRFS_BLOCK_GROUP_ITEM_KEY) {
1790                         err = -EAGAIN;
1791                         goto out;
1792                 }
1793         }
1794         *ref_ret = (struct btrfs_extent_inline_ref *)ptr;
1795 out:
1796         if (insert) {
1797                 path->keep_locks = 0;
1798                 btrfs_unlock_up_safe(path, 1);
1799         }
1800         return err;
1801 }
1802
1803 /*
1804  * helper to add new inline back ref
1805  */
1806 static noinline_for_stack
1807 void setup_inline_extent_backref(struct btrfs_fs_info *fs_info,
1808                                  struct btrfs_path *path,
1809                                  struct btrfs_extent_inline_ref *iref,
1810                                  u64 parent, u64 root_objectid,
1811                                  u64 owner, u64 offset, int refs_to_add,
1812                                  struct btrfs_delayed_extent_op *extent_op)
1813 {
1814         struct extent_buffer *leaf;
1815         struct btrfs_extent_item *ei;
1816         unsigned long ptr;
1817         unsigned long end;
1818         unsigned long item_offset;
1819         u64 refs;
1820         int size;
1821         int type;
1822
1823         leaf = path->nodes[0];
1824         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1825         item_offset = (unsigned long)iref - (unsigned long)ei;
1826
1827         type = extent_ref_type(parent, owner);
1828         size = btrfs_extent_inline_ref_size(type);
1829
1830         btrfs_extend_item(fs_info, path, size);
1831
1832         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1833         refs = btrfs_extent_refs(leaf, ei);
1834         refs += refs_to_add;
1835         btrfs_set_extent_refs(leaf, ei, refs);
1836         if (extent_op)
1837                 __run_delayed_extent_op(extent_op, leaf, ei);
1838
1839         ptr = (unsigned long)ei + item_offset;
1840         end = (unsigned long)ei + btrfs_item_size_nr(leaf, path->slots[0]);
1841         if (ptr < end - size)
1842                 memmove_extent_buffer(leaf, ptr + size, ptr,
1843                                       end - size - ptr);
1844
1845         iref = (struct btrfs_extent_inline_ref *)ptr;
1846         btrfs_set_extent_inline_ref_type(leaf, iref, type);
1847         if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1848                 struct btrfs_extent_data_ref *dref;
1849                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1850                 btrfs_set_extent_data_ref_root(leaf, dref, root_objectid);
1851                 btrfs_set_extent_data_ref_objectid(leaf, dref, owner);
1852                 btrfs_set_extent_data_ref_offset(leaf, dref, offset);
1853                 btrfs_set_extent_data_ref_count(leaf, dref, refs_to_add);
1854         } else if (type == BTRFS_SHARED_DATA_REF_KEY) {
1855                 struct btrfs_shared_data_ref *sref;
1856                 sref = (struct btrfs_shared_data_ref *)(iref + 1);
1857                 btrfs_set_shared_data_ref_count(leaf, sref, refs_to_add);
1858                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
1859         } else if (type == BTRFS_SHARED_BLOCK_REF_KEY) {
1860                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
1861         } else {
1862                 btrfs_set_extent_inline_ref_offset(leaf, iref, root_objectid);
1863         }
1864         btrfs_mark_buffer_dirty(leaf);
1865 }
1866
1867 static int lookup_extent_backref(struct btrfs_trans_handle *trans,
1868                                  struct btrfs_fs_info *fs_info,
1869                                  struct btrfs_path *path,
1870                                  struct btrfs_extent_inline_ref **ref_ret,
1871                                  u64 bytenr, u64 num_bytes, u64 parent,
1872                                  u64 root_objectid, u64 owner, u64 offset)
1873 {
1874         int ret;
1875
1876         ret = lookup_inline_extent_backref(trans, fs_info, path, ref_ret,
1877                                            bytenr, num_bytes, parent,
1878                                            root_objectid, owner, offset, 0);
1879         if (ret != -ENOENT)
1880                 return ret;
1881
1882         btrfs_release_path(path);
1883         *ref_ret = NULL;
1884
1885         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1886                 ret = lookup_tree_block_ref(trans, fs_info, path, bytenr,
1887                                             parent, root_objectid);
1888         } else {
1889                 ret = lookup_extent_data_ref(trans, fs_info, path, bytenr,
1890                                              parent, root_objectid, owner,
1891                                              offset);
1892         }
1893         return ret;
1894 }
1895
1896 /*
1897  * helper to update/remove inline back ref
1898  */
1899 static noinline_for_stack
1900 void update_inline_extent_backref(struct btrfs_fs_info *fs_info,
1901                                   struct btrfs_path *path,
1902                                   struct btrfs_extent_inline_ref *iref,
1903                                   int refs_to_mod,
1904                                   struct btrfs_delayed_extent_op *extent_op,
1905                                   int *last_ref)
1906 {
1907         struct extent_buffer *leaf;
1908         struct btrfs_extent_item *ei;
1909         struct btrfs_extent_data_ref *dref = NULL;
1910         struct btrfs_shared_data_ref *sref = NULL;
1911         unsigned long ptr;
1912         unsigned long end;
1913         u32 item_size;
1914         int size;
1915         int type;
1916         u64 refs;
1917
1918         leaf = path->nodes[0];
1919         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1920         refs = btrfs_extent_refs(leaf, ei);
1921         WARN_ON(refs_to_mod < 0 && refs + refs_to_mod <= 0);
1922         refs += refs_to_mod;
1923         btrfs_set_extent_refs(leaf, ei, refs);
1924         if (extent_op)
1925                 __run_delayed_extent_op(extent_op, leaf, ei);
1926
1927         /*
1928          * If type is invalid, we should have bailed out after
1929          * lookup_inline_extent_backref().
1930          */
1931         type = btrfs_get_extent_inline_ref_type(leaf, iref, BTRFS_REF_TYPE_ANY);
1932         ASSERT(type != BTRFS_REF_TYPE_INVALID);
1933
1934         if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1935                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1936                 refs = btrfs_extent_data_ref_count(leaf, dref);
1937         } else if (type == BTRFS_SHARED_DATA_REF_KEY) {
1938                 sref = (struct btrfs_shared_data_ref *)(iref + 1);
1939                 refs = btrfs_shared_data_ref_count(leaf, sref);
1940         } else {
1941                 refs = 1;
1942                 BUG_ON(refs_to_mod != -1);
1943         }
1944
1945         BUG_ON(refs_to_mod < 0 && refs < -refs_to_mod);
1946         refs += refs_to_mod;
1947
1948         if (refs > 0) {
1949                 if (type == BTRFS_EXTENT_DATA_REF_KEY)
1950                         btrfs_set_extent_data_ref_count(leaf, dref, refs);
1951                 else
1952                         btrfs_set_shared_data_ref_count(leaf, sref, refs);
1953         } else {
1954                 *last_ref = 1;
1955                 size =  btrfs_extent_inline_ref_size(type);
1956                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1957                 ptr = (unsigned long)iref;
1958                 end = (unsigned long)ei + item_size;
1959                 if (ptr + size < end)
1960                         memmove_extent_buffer(leaf, ptr, ptr + size,
1961                                               end - ptr - size);
1962                 item_size -= size;
1963                 btrfs_truncate_item(fs_info, path, item_size, 1);
1964         }
1965         btrfs_mark_buffer_dirty(leaf);
1966 }
1967
1968 static noinline_for_stack
1969 int insert_inline_extent_backref(struct btrfs_trans_handle *trans,
1970                                  struct btrfs_fs_info *fs_info,
1971                                  struct btrfs_path *path,
1972                                  u64 bytenr, u64 num_bytes, u64 parent,
1973                                  u64 root_objectid, u64 owner,
1974                                  u64 offset, int refs_to_add,
1975                                  struct btrfs_delayed_extent_op *extent_op)
1976 {
1977         struct btrfs_extent_inline_ref *iref;
1978         int ret;
1979
1980         ret = lookup_inline_extent_backref(trans, fs_info, path, &iref,
1981                                            bytenr, num_bytes, parent,
1982                                            root_objectid, owner, offset, 1);
1983         if (ret == 0) {
1984                 BUG_ON(owner < BTRFS_FIRST_FREE_OBJECTID);
1985                 update_inline_extent_backref(fs_info, path, iref,
1986                                              refs_to_add, extent_op, NULL);
1987         } else if (ret == -ENOENT) {
1988                 setup_inline_extent_backref(fs_info, path, iref, parent,
1989                                             root_objectid, owner, offset,
1990                                             refs_to_add, extent_op);
1991                 ret = 0;
1992         }
1993         return ret;
1994 }
1995
1996 static int insert_extent_backref(struct btrfs_trans_handle *trans,
1997                                  struct btrfs_fs_info *fs_info,
1998                                  struct btrfs_path *path,
1999                                  u64 bytenr, u64 parent, u64 root_objectid,
2000                                  u64 owner, u64 offset, int refs_to_add)
2001 {
2002         int ret;
2003         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
2004                 BUG_ON(refs_to_add != 1);
2005                 ret = insert_tree_block_ref(trans, fs_info, path, bytenr,
2006                                             parent, root_objectid);
2007         } else {
2008                 ret = insert_extent_data_ref(trans, fs_info, path, bytenr,
2009                                              parent, root_objectid,
2010                                              owner, offset, refs_to_add);
2011         }
2012         return ret;
2013 }
2014
2015 static int remove_extent_backref(struct btrfs_trans_handle *trans,
2016                                  struct btrfs_fs_info *fs_info,
2017                                  struct btrfs_path *path,
2018                                  struct btrfs_extent_inline_ref *iref,
2019                                  int refs_to_drop, int is_data, int *last_ref)
2020 {
2021         int ret = 0;
2022
2023         BUG_ON(!is_data && refs_to_drop != 1);
2024         if (iref) {
2025                 update_inline_extent_backref(fs_info, path, iref,
2026                                              -refs_to_drop, NULL, last_ref);
2027         } else if (is_data) {
2028                 ret = remove_extent_data_ref(trans, fs_info, path, refs_to_drop,
2029                                              last_ref);
2030         } else {
2031                 *last_ref = 1;
2032                 ret = btrfs_del_item(trans, fs_info->extent_root, path);
2033         }
2034         return ret;
2035 }
2036
2037 #define in_range(b, first, len)        ((b) >= (first) && (b) < (first) + (len))
2038 static int btrfs_issue_discard(struct block_device *bdev, u64 start, u64 len,
2039                                u64 *discarded_bytes)
2040 {
2041         int j, ret = 0;
2042         u64 bytes_left, end;
2043         u64 aligned_start = ALIGN(start, 1 << 9);
2044
2045         if (WARN_ON(start != aligned_start)) {
2046                 len -= aligned_start - start;
2047                 len = round_down(len, 1 << 9);
2048                 start = aligned_start;
2049         }
2050
2051         *discarded_bytes = 0;
2052
2053         if (!len)
2054                 return 0;
2055
2056         end = start + len;
2057         bytes_left = len;
2058
2059         /* Skip any superblocks on this device. */
2060         for (j = 0; j < BTRFS_SUPER_MIRROR_MAX; j++) {
2061                 u64 sb_start = btrfs_sb_offset(j);
2062                 u64 sb_end = sb_start + BTRFS_SUPER_INFO_SIZE;
2063                 u64 size = sb_start - start;
2064
2065                 if (!in_range(sb_start, start, bytes_left) &&
2066                     !in_range(sb_end, start, bytes_left) &&
2067                     !in_range(start, sb_start, BTRFS_SUPER_INFO_SIZE))
2068                         continue;
2069
2070                 /*
2071                  * Superblock spans beginning of range.  Adjust start and
2072                  * try again.
2073                  */
2074                 if (sb_start <= start) {
2075                         start += sb_end - start;
2076                         if (start > end) {
2077                                 bytes_left = 0;
2078                                 break;
2079                         }
2080                         bytes_left = end - start;
2081                         continue;
2082                 }
2083
2084                 if (size) {
2085                         ret = blkdev_issue_discard(bdev, start >> 9, size >> 9,
2086                                                    GFP_NOFS, 0);
2087                         if (!ret)
2088                                 *discarded_bytes += size;
2089                         else if (ret != -EOPNOTSUPP)
2090                                 return ret;
2091                 }
2092
2093                 start = sb_end;
2094                 if (start > end) {
2095                         bytes_left = 0;
2096                         break;
2097                 }
2098                 bytes_left = end - start;
2099         }
2100
2101         if (bytes_left) {
2102                 ret = blkdev_issue_discard(bdev, start >> 9, bytes_left >> 9,
2103                                            GFP_NOFS, 0);
2104                 if (!ret)
2105                         *discarded_bytes += bytes_left;
2106         }
2107         return ret;
2108 }
2109
2110 int btrfs_discard_extent(struct btrfs_fs_info *fs_info, u64 bytenr,
2111                          u64 num_bytes, u64 *actual_bytes)
2112 {
2113         int ret;
2114         u64 discarded_bytes = 0;
2115         struct btrfs_bio *bbio = NULL;
2116
2117
2118         /*
2119          * Avoid races with device replace and make sure our bbio has devices
2120          * associated to its stripes that don't go away while we are discarding.
2121          */
2122         btrfs_bio_counter_inc_blocked(fs_info);
2123         /* Tell the block device(s) that the sectors can be discarded */
2124         ret = btrfs_map_block(fs_info, BTRFS_MAP_DISCARD, bytenr, &num_bytes,
2125                               &bbio, 0);
2126         /* Error condition is -ENOMEM */
2127         if (!ret) {
2128                 struct btrfs_bio_stripe *stripe = bbio->stripes;
2129                 int i;
2130
2131
2132                 for (i = 0; i < bbio->num_stripes; i++, stripe++) {
2133                         u64 bytes;
2134                         struct request_queue *req_q;
2135
2136                         if (!stripe->dev->bdev) {
2137                                 ASSERT(btrfs_test_opt(fs_info, DEGRADED));
2138                                 continue;
2139                         }
2140                         req_q = bdev_get_queue(stripe->dev->bdev);
2141                         if (!blk_queue_discard(req_q))
2142                                 continue;
2143
2144                         ret = btrfs_issue_discard(stripe->dev->bdev,
2145                                                   stripe->physical,
2146                                                   stripe->length,
2147                                                   &bytes);
2148                         if (!ret)
2149                                 discarded_bytes += bytes;
2150                         else if (ret != -EOPNOTSUPP)
2151                                 break; /* Logic errors or -ENOMEM, or -EIO but I don't know how that could happen JDM */
2152
2153                         /*
2154                          * Just in case we get back EOPNOTSUPP for some reason,
2155                          * just ignore the return value so we don't screw up
2156                          * people calling discard_extent.
2157                          */
2158                         ret = 0;
2159                 }
2160                 btrfs_put_bbio(bbio);
2161         }
2162         btrfs_bio_counter_dec(fs_info);
2163
2164         if (actual_bytes)
2165                 *actual_bytes = discarded_bytes;
2166
2167
2168         if (ret == -EOPNOTSUPP)
2169                 ret = 0;
2170         return ret;
2171 }
2172
2173 /* Can return -ENOMEM */
2174 int btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
2175                          struct btrfs_root *root,
2176                          u64 bytenr, u64 num_bytes, u64 parent,
2177                          u64 root_objectid, u64 owner, u64 offset)
2178 {
2179         struct btrfs_fs_info *fs_info = root->fs_info;
2180         int old_ref_mod, new_ref_mod;
2181         int ret;
2182
2183         BUG_ON(owner < BTRFS_FIRST_FREE_OBJECTID &&
2184                root_objectid == BTRFS_TREE_LOG_OBJECTID);
2185
2186         btrfs_ref_tree_mod(root, bytenr, num_bytes, parent, root_objectid,
2187                            owner, offset, BTRFS_ADD_DELAYED_REF);
2188
2189         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
2190                 ret = btrfs_add_delayed_tree_ref(fs_info, trans, bytenr,
2191                                                  num_bytes, parent,
2192                                                  root_objectid, (int)owner,
2193                                                  BTRFS_ADD_DELAYED_REF, NULL,
2194                                                  &old_ref_mod, &new_ref_mod);
2195         } else {
2196                 ret = btrfs_add_delayed_data_ref(fs_info, trans, bytenr,
2197                                                  num_bytes, parent,
2198                                                  root_objectid, owner, offset,
2199                                                  0, BTRFS_ADD_DELAYED_REF,
2200                                                  &old_ref_mod, &new_ref_mod);
2201         }
2202
2203         if (ret == 0 && old_ref_mod < 0 && new_ref_mod >= 0)
2204                 add_pinned_bytes(fs_info, -num_bytes, owner, root_objectid);
2205
2206         return ret;
2207 }
2208
2209 static int __btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
2210                                   struct btrfs_fs_info *fs_info,
2211                                   struct btrfs_delayed_ref_node *node,
2212                                   u64 parent, u64 root_objectid,
2213                                   u64 owner, u64 offset, int refs_to_add,
2214                                   struct btrfs_delayed_extent_op *extent_op)
2215 {
2216         struct btrfs_path *path;
2217         struct extent_buffer *leaf;
2218         struct btrfs_extent_item *item;
2219         struct btrfs_key key;
2220         u64 bytenr = node->bytenr;
2221         u64 num_bytes = node->num_bytes;
2222         u64 refs;
2223         int ret;
2224
2225         path = btrfs_alloc_path();
2226         if (!path)
2227                 return -ENOMEM;
2228
2229         path->reada = READA_FORWARD;
2230         path->leave_spinning = 1;
2231         /* this will setup the path even if it fails to insert the back ref */
2232         ret = insert_inline_extent_backref(trans, fs_info, path, bytenr,
2233                                            num_bytes, parent, root_objectid,
2234                                            owner, offset,
2235                                            refs_to_add, extent_op);
2236         if ((ret < 0 && ret != -EAGAIN) || !ret)
2237                 goto out;
2238
2239         /*
2240          * Ok we had -EAGAIN which means we didn't have space to insert and
2241          * inline extent ref, so just update the reference count and add a
2242          * normal backref.
2243          */
2244         leaf = path->nodes[0];
2245         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
2246         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2247         refs = btrfs_extent_refs(leaf, item);
2248         btrfs_set_extent_refs(leaf, item, refs + refs_to_add);
2249         if (extent_op)
2250                 __run_delayed_extent_op(extent_op, leaf, item);
2251
2252         btrfs_mark_buffer_dirty(leaf);
2253         btrfs_release_path(path);
2254
2255         path->reada = READA_FORWARD;
2256         path->leave_spinning = 1;
2257         /* now insert the actual backref */
2258         ret = insert_extent_backref(trans, fs_info, path, bytenr, parent,
2259                                     root_objectid, owner, offset, refs_to_add);
2260         if (ret)
2261                 btrfs_abort_transaction(trans, ret);
2262 out:
2263         btrfs_free_path(path);
2264         return ret;
2265 }
2266
2267 static int run_delayed_data_ref(struct btrfs_trans_handle *trans,
2268                                 struct btrfs_fs_info *fs_info,
2269                                 struct btrfs_delayed_ref_node *node,
2270                                 struct btrfs_delayed_extent_op *extent_op,
2271                                 int insert_reserved)
2272 {
2273         int ret = 0;
2274         struct btrfs_delayed_data_ref *ref;
2275         struct btrfs_key ins;
2276         u64 parent = 0;
2277         u64 ref_root = 0;
2278         u64 flags = 0;
2279
2280         ins.objectid = node->bytenr;
2281         ins.offset = node->num_bytes;
2282         ins.type = BTRFS_EXTENT_ITEM_KEY;
2283
2284         ref = btrfs_delayed_node_to_data_ref(node);
2285         trace_run_delayed_data_ref(fs_info, node, ref, node->action);
2286
2287         if (node->type == BTRFS_SHARED_DATA_REF_KEY)
2288                 parent = ref->parent;
2289         ref_root = ref->root;
2290
2291         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
2292                 if (extent_op)
2293                         flags |= extent_op->flags_to_set;
2294                 ret = alloc_reserved_file_extent(trans, fs_info,
2295                                                  parent, ref_root, flags,
2296                                                  ref->objectid, ref->offset,
2297                                                  &ins, node->ref_mod);
2298         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
2299                 ret = __btrfs_inc_extent_ref(trans, fs_info, node, parent,
2300                                              ref_root, ref->objectid,
2301                                              ref->offset, node->ref_mod,
2302                                              extent_op);
2303         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
2304                 ret = __btrfs_free_extent(trans, fs_info, node, parent,
2305                                           ref_root, ref->objectid,
2306                                           ref->offset, node->ref_mod,
2307                                           extent_op);
2308         } else {
2309                 BUG();
2310         }
2311         return ret;
2312 }
2313
2314 static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op,
2315                                     struct extent_buffer *leaf,
2316                                     struct btrfs_extent_item *ei)
2317 {
2318         u64 flags = btrfs_extent_flags(leaf, ei);
2319         if (extent_op->update_flags) {
2320                 flags |= extent_op->flags_to_set;
2321                 btrfs_set_extent_flags(leaf, ei, flags);
2322         }
2323
2324         if (extent_op->update_key) {
2325                 struct btrfs_tree_block_info *bi;
2326                 BUG_ON(!(flags & BTRFS_EXTENT_FLAG_TREE_BLOCK));
2327                 bi = (struct btrfs_tree_block_info *)(ei + 1);
2328                 btrfs_set_tree_block_key(leaf, bi, &extent_op->key);
2329         }
2330 }
2331
2332 static int run_delayed_extent_op(struct btrfs_trans_handle *trans,
2333                                  struct btrfs_fs_info *fs_info,
2334                                  struct btrfs_delayed_ref_head *head,
2335                                  struct btrfs_delayed_extent_op *extent_op)
2336 {
2337         struct btrfs_key key;
2338         struct btrfs_path *path;
2339         struct btrfs_extent_item *ei;
2340         struct extent_buffer *leaf;
2341         u32 item_size;
2342         int ret;
2343         int err = 0;
2344         int metadata = !extent_op->is_data;
2345
2346         if (trans->aborted)
2347                 return 0;
2348
2349         if (metadata && !btrfs_fs_incompat(fs_info, SKINNY_METADATA))
2350                 metadata = 0;
2351
2352         path = btrfs_alloc_path();
2353         if (!path)
2354                 return -ENOMEM;
2355
2356         key.objectid = head->bytenr;
2357
2358         if (metadata) {
2359                 key.type = BTRFS_METADATA_ITEM_KEY;
2360                 key.offset = extent_op->level;
2361         } else {
2362                 key.type = BTRFS_EXTENT_ITEM_KEY;
2363                 key.offset = head->num_bytes;
2364         }
2365
2366 again:
2367         path->reada = READA_FORWARD;
2368         path->leave_spinning = 1;
2369         ret = btrfs_search_slot(trans, fs_info->extent_root, &key, path, 0, 1);
2370         if (ret < 0) {
2371                 err = ret;
2372                 goto out;
2373         }
2374         if (ret > 0) {
2375                 if (metadata) {
2376                         if (path->slots[0] > 0) {
2377                                 path->slots[0]--;
2378                                 btrfs_item_key_to_cpu(path->nodes[0], &key,
2379                                                       path->slots[0]);
2380                                 if (key.objectid == head->bytenr &&
2381                                     key.type == BTRFS_EXTENT_ITEM_KEY &&
2382                                     key.offset == head->num_bytes)
2383                                         ret = 0;
2384                         }
2385                         if (ret > 0) {
2386                                 btrfs_release_path(path);
2387                                 metadata = 0;
2388
2389                                 key.objectid = head->bytenr;
2390                                 key.offset = head->num_bytes;
2391                                 key.type = BTRFS_EXTENT_ITEM_KEY;
2392                                 goto again;
2393                         }
2394                 } else {
2395                         err = -EIO;
2396                         goto out;
2397                 }
2398         }
2399
2400         leaf = path->nodes[0];
2401         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2402 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
2403         if (item_size < sizeof(*ei)) {
2404                 ret = convert_extent_item_v0(trans, fs_info, path, (u64)-1, 0);
2405                 if (ret < 0) {
2406                         err = ret;
2407                         goto out;
2408                 }
2409                 leaf = path->nodes[0];
2410                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2411         }
2412 #endif
2413         BUG_ON(item_size < sizeof(*ei));
2414         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2415         __run_delayed_extent_op(extent_op, leaf, ei);
2416
2417         btrfs_mark_buffer_dirty(leaf);
2418 out:
2419         btrfs_free_path(path);
2420         return err;
2421 }
2422
2423 static int run_delayed_tree_ref(struct btrfs_trans_handle *trans,
2424                                 struct btrfs_fs_info *fs_info,
2425                                 struct btrfs_delayed_ref_node *node,
2426                                 struct btrfs_delayed_extent_op *extent_op,
2427                                 int insert_reserved)
2428 {
2429         int ret = 0;
2430         struct btrfs_delayed_tree_ref *ref;
2431         struct btrfs_key ins;
2432         u64 parent = 0;
2433         u64 ref_root = 0;
2434         bool skinny_metadata = btrfs_fs_incompat(fs_info, SKINNY_METADATA);
2435
2436         ref = btrfs_delayed_node_to_tree_ref(node);
2437         trace_run_delayed_tree_ref(fs_info, node, ref, node->action);
2438
2439         if (node->type == BTRFS_SHARED_BLOCK_REF_KEY)
2440                 parent = ref->parent;
2441         ref_root = ref->root;
2442
2443         ins.objectid = node->bytenr;
2444         if (skinny_metadata) {
2445                 ins.offset = ref->level;
2446                 ins.type = BTRFS_METADATA_ITEM_KEY;
2447         } else {
2448                 ins.offset = node->num_bytes;
2449                 ins.type = BTRFS_EXTENT_ITEM_KEY;
2450         }
2451
2452         if (node->ref_mod != 1) {
2453                 btrfs_err(fs_info,
2454         "btree block(%llu) has %d references rather than 1: action %d ref_root %llu parent %llu",
2455                           node->bytenr, node->ref_mod, node->action, ref_root,
2456                           parent);
2457                 return -EIO;
2458         }
2459         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
2460                 BUG_ON(!extent_op || !extent_op->update_flags);
2461                 ret = alloc_reserved_tree_block(trans, fs_info,
2462                                                 parent, ref_root,
2463                                                 extent_op->flags_to_set,
2464                                                 &extent_op->key,
2465                                                 ref->level, &ins);
2466         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
2467                 ret = __btrfs_inc_extent_ref(trans, fs_info, node,
2468                                              parent, ref_root,
2469                                              ref->level, 0, 1,
2470                                              extent_op);
2471         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
2472                 ret = __btrfs_free_extent(trans, fs_info, node,
2473                                           parent, ref_root,
2474                                           ref->level, 0, 1, extent_op);
2475         } else {
2476                 BUG();
2477         }
2478         return ret;
2479 }
2480
2481 /* helper function to actually process a single delayed ref entry */
2482 static int run_one_delayed_ref(struct btrfs_trans_handle *trans,
2483                                struct btrfs_fs_info *fs_info,
2484                                struct btrfs_delayed_ref_node *node,
2485                                struct btrfs_delayed_extent_op *extent_op,
2486                                int insert_reserved)
2487 {
2488         int ret = 0;
2489
2490         if (trans->aborted) {
2491                 if (insert_reserved)
2492                         btrfs_pin_extent(fs_info, node->bytenr,
2493                                          node->num_bytes, 1);
2494                 return 0;
2495         }
2496
2497         if (node->type == BTRFS_TREE_BLOCK_REF_KEY ||
2498             node->type == BTRFS_SHARED_BLOCK_REF_KEY)
2499                 ret = run_delayed_tree_ref(trans, fs_info, node, extent_op,
2500                                            insert_reserved);
2501         else if (node->type == BTRFS_EXTENT_DATA_REF_KEY ||
2502                  node->type == BTRFS_SHARED_DATA_REF_KEY)
2503                 ret = run_delayed_data_ref(trans, fs_info, node, extent_op,
2504                                            insert_reserved);
2505         else
2506                 BUG();
2507         return ret;
2508 }
2509
2510 static inline struct btrfs_delayed_ref_node *
2511 select_delayed_ref(struct btrfs_delayed_ref_head *head)
2512 {
2513         struct btrfs_delayed_ref_node *ref;
2514
2515         if (RB_EMPTY_ROOT(&head->ref_tree))
2516                 return NULL;
2517
2518         /*
2519          * Select a delayed ref of type BTRFS_ADD_DELAYED_REF first.
2520          * This is to prevent a ref count from going down to zero, which deletes
2521          * the extent item from the extent tree, when there still are references
2522          * to add, which would fail because they would not find the extent item.
2523          */
2524         if (!list_empty(&head->ref_add_list))
2525                 return list_first_entry(&head->ref_add_list,
2526                                 struct btrfs_delayed_ref_node, add_list);
2527
2528         ref = rb_entry(rb_first(&head->ref_tree),
2529                        struct btrfs_delayed_ref_node, ref_node);
2530         ASSERT(list_empty(&ref->add_list));
2531         return ref;
2532 }
2533
2534 static void unselect_delayed_ref_head(struct btrfs_delayed_ref_root *delayed_refs,
2535                                       struct btrfs_delayed_ref_head *head)
2536 {
2537         spin_lock(&delayed_refs->lock);
2538         head->processing = 0;
2539         delayed_refs->num_heads_ready++;
2540         spin_unlock(&delayed_refs->lock);
2541         btrfs_delayed_ref_unlock(head);
2542 }
2543
2544 static int cleanup_extent_op(struct btrfs_trans_handle *trans,
2545                              struct btrfs_fs_info *fs_info,
2546                              struct btrfs_delayed_ref_head *head)
2547 {
2548         struct btrfs_delayed_extent_op *extent_op = head->extent_op;
2549         int ret;
2550
2551         if (!extent_op)
2552                 return 0;
2553         head->extent_op = NULL;
2554         if (head->must_insert_reserved) {
2555                 btrfs_free_delayed_extent_op(extent_op);
2556                 return 0;
2557         }
2558         spin_unlock(&head->lock);
2559         ret = run_delayed_extent_op(trans, fs_info, head, extent_op);
2560         btrfs_free_delayed_extent_op(extent_op);
2561         return ret ? ret : 1;
2562 }
2563
2564 static int cleanup_ref_head(struct btrfs_trans_handle *trans,
2565                             struct btrfs_fs_info *fs_info,
2566                             struct btrfs_delayed_ref_head *head)
2567 {
2568         struct btrfs_delayed_ref_root *delayed_refs;
2569         int ret;
2570
2571         delayed_refs = &trans->transaction->delayed_refs;
2572
2573         ret = cleanup_extent_op(trans, fs_info, head);
2574         if (ret < 0) {
2575                 unselect_delayed_ref_head(delayed_refs, head);
2576                 btrfs_debug(fs_info, "run_delayed_extent_op returned %d", ret);
2577                 return ret;
2578         } else if (ret) {
2579                 return ret;
2580         }
2581
2582         /*
2583          * Need to drop our head ref lock and re-acquire the delayed ref lock
2584          * and then re-check to make sure nobody got added.
2585          */
2586         spin_unlock(&head->lock);
2587         spin_lock(&delayed_refs->lock);
2588         spin_lock(&head->lock);
2589         if (!RB_EMPTY_ROOT(&head->ref_tree) || head->extent_op) {
2590                 spin_unlock(&head->lock);
2591                 spin_unlock(&delayed_refs->lock);
2592                 return 1;
2593         }
2594         delayed_refs->num_heads--;
2595         rb_erase(&head->href_node, &delayed_refs->href_root);
2596         RB_CLEAR_NODE(&head->href_node);
2597         spin_unlock(&delayed_refs->lock);
2598         spin_unlock(&head->lock);
2599         atomic_dec(&delayed_refs->num_entries);
2600
2601         trace_run_delayed_ref_head(fs_info, head, 0);
2602
2603         if (head->total_ref_mod < 0) {
2604                 struct btrfs_space_info *space_info;
2605                 u64 flags;
2606
2607                 if (head->is_data)
2608                         flags = BTRFS_BLOCK_GROUP_DATA;
2609                 else if (head->is_system)
2610                         flags = BTRFS_BLOCK_GROUP_SYSTEM;
2611                 else
2612                         flags = BTRFS_BLOCK_GROUP_METADATA;
2613                 space_info = __find_space_info(fs_info, flags);
2614                 ASSERT(space_info);
2615                 percpu_counter_add(&space_info->total_bytes_pinned,
2616                                    -head->num_bytes);
2617
2618                 if (head->is_data) {
2619                         spin_lock(&delayed_refs->lock);
2620                         delayed_refs->pending_csums -= head->num_bytes;
2621                         spin_unlock(&delayed_refs->lock);
2622                 }
2623         }
2624
2625         if (head->must_insert_reserved) {
2626                 btrfs_pin_extent(fs_info, head->bytenr,
2627                                  head->num_bytes, 1);
2628                 if (head->is_data) {
2629                         ret = btrfs_del_csums(trans, fs_info, head->bytenr,
2630                                               head->num_bytes);
2631                 }
2632         }
2633
2634         /* Also free its reserved qgroup space */
2635         btrfs_qgroup_free_delayed_ref(fs_info, head->qgroup_ref_root,
2636                                       head->qgroup_reserved);
2637         btrfs_delayed_ref_unlock(head);
2638         btrfs_put_delayed_ref_head(head);
2639         return 0;
2640 }
2641
2642 /*
2643  * Returns 0 on success or if called with an already aborted transaction.
2644  * Returns -ENOMEM or -EIO on failure and will abort the transaction.
2645  */
2646 static noinline int __btrfs_run_delayed_refs(struct btrfs_trans_handle *trans,
2647                                              unsigned long nr)
2648 {
2649         struct btrfs_fs_info *fs_info = trans->fs_info;
2650         struct btrfs_delayed_ref_root *delayed_refs;
2651         struct btrfs_delayed_ref_node *ref;
2652         struct btrfs_delayed_ref_head *locked_ref = NULL;
2653         struct btrfs_delayed_extent_op *extent_op;
2654         ktime_t start = ktime_get();
2655         int ret;
2656         unsigned long count = 0;
2657         unsigned long actual_count = 0;
2658         int must_insert_reserved = 0;
2659
2660         delayed_refs = &trans->transaction->delayed_refs;
2661         while (1) {
2662                 if (!locked_ref) {
2663                         if (count >= nr)
2664                                 break;
2665
2666                         spin_lock(&delayed_refs->lock);
2667                         locked_ref = btrfs_select_ref_head(trans);
2668                         if (!locked_ref) {
2669                                 spin_unlock(&delayed_refs->lock);
2670                                 break;
2671                         }
2672
2673                         /* grab the lock that says we are going to process
2674                          * all the refs for this head */
2675                         ret = btrfs_delayed_ref_lock(trans, locked_ref);
2676                         spin_unlock(&delayed_refs->lock);
2677                         /*
2678                          * we may have dropped the spin lock to get the head
2679                          * mutex lock, and that might have given someone else
2680                          * time to free the head.  If that's true, it has been
2681                          * removed from our list and we can move on.
2682                          */
2683                         if (ret == -EAGAIN) {
2684                                 locked_ref = NULL;
2685                                 count++;
2686                                 continue;
2687                         }
2688                 }
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, fs_info, delayed_refs,
2704                                          locked_ref);
2705
2706                 /*
2707                  * locked_ref is the head node, so we have to go one
2708                  * node back for any delayed ref updates
2709                  */
2710                 ref = select_delayed_ref(locked_ref);
2711
2712                 if (ref && ref->seq &&
2713                     btrfs_check_delayed_seq(fs_info, delayed_refs, ref->seq)) {
2714                         spin_unlock(&locked_ref->lock);
2715                         unselect_delayed_ref_head(delayed_refs, locked_ref);
2716                         locked_ref = NULL;
2717                         cond_resched();
2718                         count++;
2719                         continue;
2720                 }
2721
2722                 /*
2723                  * We're done processing refs in this ref_head, clean everything
2724                  * up and move on to the next ref_head.
2725                  */
2726                 if (!ref) {
2727                         ret = cleanup_ref_head(trans, fs_info, locked_ref);
2728                         if (ret > 0 ) {
2729                                 /* We dropped our lock, we need to loop. */
2730                                 ret = 0;
2731                                 continue;
2732                         } else if (ret) {
2733                                 return ret;
2734                         }
2735                         locked_ref = NULL;
2736                         count++;
2737                         continue;
2738                 }
2739
2740                 actual_count++;
2741                 ref->in_tree = 0;
2742                 rb_erase(&ref->ref_node, &locked_ref->ref_tree);
2743                 RB_CLEAR_NODE(&ref->ref_node);
2744                 if (!list_empty(&ref->add_list))
2745                         list_del(&ref->add_list);
2746                 /*
2747                  * When we play the delayed ref, also correct the ref_mod on
2748                  * head
2749                  */
2750                 switch (ref->action) {
2751                 case BTRFS_ADD_DELAYED_REF:
2752                 case BTRFS_ADD_DELAYED_EXTENT:
2753                         locked_ref->ref_mod -= ref->ref_mod;
2754                         break;
2755                 case BTRFS_DROP_DELAYED_REF:
2756                         locked_ref->ref_mod += ref->ref_mod;
2757                         break;
2758                 default:
2759                         WARN_ON(1);
2760                 }
2761                 atomic_dec(&delayed_refs->num_entries);
2762
2763                 /*
2764                  * Record the must-insert_reserved flag before we drop the spin
2765                  * lock.
2766                  */
2767                 must_insert_reserved = locked_ref->must_insert_reserved;
2768                 locked_ref->must_insert_reserved = 0;
2769
2770                 extent_op = locked_ref->extent_op;
2771                 locked_ref->extent_op = NULL;
2772                 spin_unlock(&locked_ref->lock);
2773
2774                 ret = run_one_delayed_ref(trans, fs_info, ref, extent_op,
2775                                           must_insert_reserved);
2776
2777                 btrfs_free_delayed_extent_op(extent_op);
2778                 if (ret) {
2779                         unselect_delayed_ref_head(delayed_refs, locked_ref);
2780                         btrfs_put_delayed_ref(ref);
2781                         btrfs_debug(fs_info, "run_one_delayed_ref returned %d",
2782                                     ret);
2783                         return ret;
2784                 }
2785
2786                 btrfs_put_delayed_ref(ref);
2787                 count++;
2788                 cond_resched();
2789         }
2790
2791         /*
2792          * We don't want to include ref heads since we can have empty ref heads
2793          * and those will drastically skew our runtime down since we just do
2794          * accounting, no actual extent tree updates.
2795          */
2796         if (actual_count > 0) {
2797                 u64 runtime = ktime_to_ns(ktime_sub(ktime_get(), start));
2798                 u64 avg;
2799
2800                 /*
2801                  * We weigh the current average higher than our current runtime
2802                  * to avoid large swings in the average.
2803                  */
2804                 spin_lock(&delayed_refs->lock);
2805                 avg = fs_info->avg_delayed_ref_runtime * 3 + runtime;
2806                 fs_info->avg_delayed_ref_runtime = avg >> 2;    /* div by 4 */
2807                 spin_unlock(&delayed_refs->lock);
2808         }
2809         return 0;
2810 }
2811
2812 #ifdef SCRAMBLE_DELAYED_REFS
2813 /*
2814  * Normally delayed refs get processed in ascending bytenr order. This
2815  * correlates in most cases to the order added. To expose dependencies on this
2816  * order, we start to process the tree in the middle instead of the beginning
2817  */
2818 static u64 find_middle(struct rb_root *root)
2819 {
2820         struct rb_node *n = root->rb_node;
2821         struct btrfs_delayed_ref_node *entry;
2822         int alt = 1;
2823         u64 middle;
2824         u64 first = 0, last = 0;
2825
2826         n = rb_first(root);
2827         if (n) {
2828                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2829                 first = entry->bytenr;
2830         }
2831         n = rb_last(root);
2832         if (n) {
2833                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2834                 last = entry->bytenr;
2835         }
2836         n = root->rb_node;
2837
2838         while (n) {
2839                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2840                 WARN_ON(!entry->in_tree);
2841
2842                 middle = entry->bytenr;
2843
2844                 if (alt)
2845                         n = n->rb_left;
2846                 else
2847                         n = n->rb_right;
2848
2849                 alt = 1 - alt;
2850         }
2851         return middle;
2852 }
2853 #endif
2854
2855 static inline u64 heads_to_leaves(struct btrfs_fs_info *fs_info, u64 heads)
2856 {
2857         u64 num_bytes;
2858
2859         num_bytes = heads * (sizeof(struct btrfs_extent_item) +
2860                              sizeof(struct btrfs_extent_inline_ref));
2861         if (!btrfs_fs_incompat(fs_info, SKINNY_METADATA))
2862                 num_bytes += heads * sizeof(struct btrfs_tree_block_info);
2863
2864         /*
2865          * We don't ever fill up leaves all the way so multiply by 2 just to be
2866          * closer to what we're really going to want to use.
2867          */
2868         return div_u64(num_bytes, BTRFS_LEAF_DATA_SIZE(fs_info));
2869 }
2870
2871 /*
2872  * Takes the number of bytes to be csumm'ed and figures out how many leaves it
2873  * would require to store the csums for that many bytes.
2874  */
2875 u64 btrfs_csum_bytes_to_leaves(struct btrfs_fs_info *fs_info, u64 csum_bytes)
2876 {
2877         u64 csum_size;
2878         u64 num_csums_per_leaf;
2879         u64 num_csums;
2880
2881         csum_size = BTRFS_MAX_ITEM_SIZE(fs_info);
2882         num_csums_per_leaf = div64_u64(csum_size,
2883                         (u64)btrfs_super_csum_size(fs_info->super_copy));
2884         num_csums = div64_u64(csum_bytes, fs_info->sectorsize);
2885         num_csums += num_csums_per_leaf - 1;
2886         num_csums = div64_u64(num_csums, num_csums_per_leaf);
2887         return num_csums;
2888 }
2889
2890 int btrfs_check_space_for_delayed_refs(struct btrfs_trans_handle *trans,
2891                                        struct btrfs_fs_info *fs_info)
2892 {
2893         struct btrfs_block_rsv *global_rsv;
2894         u64 num_heads = trans->transaction->delayed_refs.num_heads_ready;
2895         u64 csum_bytes = trans->transaction->delayed_refs.pending_csums;
2896         unsigned int num_dirty_bgs = trans->transaction->num_dirty_bgs;
2897         u64 num_bytes, num_dirty_bgs_bytes;
2898         int ret = 0;
2899
2900         num_bytes = btrfs_calc_trans_metadata_size(fs_info, 1);
2901         num_heads = heads_to_leaves(fs_info, num_heads);
2902         if (num_heads > 1)
2903                 num_bytes += (num_heads - 1) * fs_info->nodesize;
2904         num_bytes <<= 1;
2905         num_bytes += btrfs_csum_bytes_to_leaves(fs_info, csum_bytes) *
2906                                                         fs_info->nodesize;
2907         num_dirty_bgs_bytes = btrfs_calc_trans_metadata_size(fs_info,
2908                                                              num_dirty_bgs);
2909         global_rsv = &fs_info->global_block_rsv;
2910
2911         /*
2912          * If we can't allocate any more chunks lets make sure we have _lots_ of
2913          * wiggle room since running delayed refs can create more delayed refs.
2914          */
2915         if (global_rsv->space_info->full) {
2916                 num_dirty_bgs_bytes <<= 1;
2917                 num_bytes <<= 1;
2918         }
2919
2920         spin_lock(&global_rsv->lock);
2921         if (global_rsv->reserved <= num_bytes + num_dirty_bgs_bytes)
2922                 ret = 1;
2923         spin_unlock(&global_rsv->lock);
2924         return ret;
2925 }
2926
2927 int btrfs_should_throttle_delayed_refs(struct btrfs_trans_handle *trans,
2928                                        struct btrfs_fs_info *fs_info)
2929 {
2930         u64 num_entries =
2931                 atomic_read(&trans->transaction->delayed_refs.num_entries);
2932         u64 avg_runtime;
2933         u64 val;
2934
2935         smp_mb();
2936         avg_runtime = fs_info->avg_delayed_ref_runtime;
2937         val = num_entries * avg_runtime;
2938         if (val >= NSEC_PER_SEC)
2939                 return 1;
2940         if (val >= NSEC_PER_SEC / 2)
2941                 return 2;
2942
2943         return btrfs_check_space_for_delayed_refs(trans, fs_info);
2944 }
2945
2946 struct async_delayed_refs {
2947         struct btrfs_root *root;
2948         u64 transid;
2949         int count;
2950         int error;
2951         int sync;
2952         struct completion wait;
2953         struct btrfs_work work;
2954 };
2955
2956 static inline struct async_delayed_refs *
2957 to_async_delayed_refs(struct btrfs_work *work)
2958 {
2959         return container_of(work, struct async_delayed_refs, work);
2960 }
2961
2962 static void delayed_ref_async_start(struct btrfs_work *work)
2963 {
2964         struct async_delayed_refs *async = to_async_delayed_refs(work);
2965         struct btrfs_trans_handle *trans;
2966         struct btrfs_fs_info *fs_info = async->root->fs_info;
2967         int ret;
2968
2969         /* if the commit is already started, we don't need to wait here */
2970         if (btrfs_transaction_blocked(fs_info))
2971                 goto done;
2972
2973         trans = btrfs_join_transaction(async->root);
2974         if (IS_ERR(trans)) {
2975                 async->error = PTR_ERR(trans);
2976                 goto done;
2977         }
2978
2979         /*
2980          * trans->sync means that when we call end_transaction, we won't
2981          * wait on delayed refs
2982          */
2983         trans->sync = true;
2984
2985         /* Don't bother flushing if we got into a different transaction */
2986         if (trans->transid > async->transid)
2987                 goto end;
2988
2989         ret = btrfs_run_delayed_refs(trans, async->count);
2990         if (ret)
2991                 async->error = ret;
2992 end:
2993         ret = btrfs_end_transaction(trans);
2994         if (ret && !async->error)
2995                 async->error = ret;
2996 done:
2997         if (async->sync)
2998                 complete(&async->wait);
2999         else
3000                 kfree(async);
3001 }
3002
3003 int btrfs_async_run_delayed_refs(struct btrfs_fs_info *fs_info,
3004                                  unsigned long count, u64 transid, int wait)
3005 {
3006         struct async_delayed_refs *async;
3007         int ret;
3008
3009         async = kmalloc(sizeof(*async), GFP_NOFS);
3010         if (!async)
3011                 return -ENOMEM;
3012
3013         async->root = fs_info->tree_root;
3014         async->count = count;
3015         async->error = 0;
3016         async->transid = transid;
3017         if (wait)
3018                 async->sync = 1;
3019         else
3020                 async->sync = 0;
3021         init_completion(&async->wait);
3022
3023         btrfs_init_work(&async->work, btrfs_extent_refs_helper,
3024                         delayed_ref_async_start, NULL, NULL);
3025
3026         btrfs_queue_work(fs_info->extent_workers, &async->work);
3027
3028         if (wait) {
3029                 wait_for_completion(&async->wait);
3030                 ret = async->error;
3031                 kfree(async);
3032                 return ret;
3033         }
3034         return 0;
3035 }
3036
3037 /*
3038  * this starts processing the delayed reference count updates and
3039  * extent insertions we have queued up so far.  count can be
3040  * 0, which means to process everything in the tree at the start
3041  * of the run (but not newly added entries), or it can be some target
3042  * number you'd like to process.
3043  *
3044  * Returns 0 on success or if called with an aborted transaction
3045  * Returns <0 on error and aborts the transaction
3046  */
3047 int btrfs_run_delayed_refs(struct btrfs_trans_handle *trans,
3048                            unsigned long count)
3049 {
3050         struct btrfs_fs_info *fs_info = trans->fs_info;
3051         struct rb_node *node;
3052         struct btrfs_delayed_ref_root *delayed_refs;
3053         struct btrfs_delayed_ref_head *head;
3054         int ret;
3055         int run_all = count == (unsigned long)-1;
3056         bool can_flush_pending_bgs = trans->can_flush_pending_bgs;
3057
3058         /* We'll clean this up in btrfs_cleanup_transaction */
3059         if (trans->aborted)
3060                 return 0;
3061
3062         if (test_bit(BTRFS_FS_CREATING_FREE_SPACE_TREE, &fs_info->flags))
3063                 return 0;
3064
3065         delayed_refs = &trans->transaction->delayed_refs;
3066         if (count == 0)
3067                 count = atomic_read(&delayed_refs->num_entries) * 2;
3068
3069 again:
3070 #ifdef SCRAMBLE_DELAYED_REFS
3071         delayed_refs->run_delayed_start = find_middle(&delayed_refs->root);
3072 #endif
3073         trans->can_flush_pending_bgs = false;
3074         ret = __btrfs_run_delayed_refs(trans, count);
3075         if (ret < 0) {
3076                 btrfs_abort_transaction(trans, ret);
3077                 return ret;
3078         }
3079
3080         if (run_all) {
3081                 if (!list_empty(&trans->new_bgs))
3082                         btrfs_create_pending_block_groups(trans);
3083
3084                 spin_lock(&delayed_refs->lock);
3085                 node = rb_first(&delayed_refs->href_root);
3086                 if (!node) {
3087                         spin_unlock(&delayed_refs->lock);
3088                         goto out;
3089                 }
3090                 head = rb_entry(node, struct btrfs_delayed_ref_head,
3091                                 href_node);
3092                 refcount_inc(&head->refs);
3093                 spin_unlock(&delayed_refs->lock);
3094
3095                 /* Mutex was contended, block until it's released and retry. */
3096                 mutex_lock(&head->mutex);
3097                 mutex_unlock(&head->mutex);
3098
3099                 btrfs_put_delayed_ref_head(head);
3100                 cond_resched();
3101                 goto again;
3102         }
3103 out:
3104         trans->can_flush_pending_bgs = can_flush_pending_bgs;
3105         return 0;
3106 }
3107
3108 int btrfs_set_disk_extent_flags(struct btrfs_trans_handle *trans,
3109                                 struct btrfs_fs_info *fs_info,
3110                                 u64 bytenr, u64 num_bytes, u64 flags,
3111                                 int level, int is_data)
3112 {
3113         struct btrfs_delayed_extent_op *extent_op;
3114         int ret;
3115
3116         extent_op = btrfs_alloc_delayed_extent_op();
3117         if (!extent_op)
3118                 return -ENOMEM;
3119
3120         extent_op->flags_to_set = flags;
3121         extent_op->update_flags = true;
3122         extent_op->update_key = false;
3123         extent_op->is_data = is_data ? true : false;
3124         extent_op->level = level;
3125
3126         ret = btrfs_add_delayed_extent_op(fs_info, trans, bytenr,
3127                                           num_bytes, extent_op);
3128         if (ret)
3129                 btrfs_free_delayed_extent_op(extent_op);
3130         return ret;
3131 }
3132
3133 static noinline int check_delayed_ref(struct btrfs_root *root,
3134                                       struct btrfs_path *path,
3135                                       u64 objectid, u64 offset, u64 bytenr)
3136 {
3137         struct btrfs_delayed_ref_head *head;
3138         struct btrfs_delayed_ref_node *ref;
3139         struct btrfs_delayed_data_ref *data_ref;
3140         struct btrfs_delayed_ref_root *delayed_refs;
3141         struct btrfs_transaction *cur_trans;
3142         struct rb_node *node;
3143         int ret = 0;
3144
3145         cur_trans = root->fs_info->running_transaction;
3146         if (!cur_trans)
3147                 return 0;
3148
3149         delayed_refs = &cur_trans->delayed_refs;
3150         spin_lock(&delayed_refs->lock);
3151         head = btrfs_find_delayed_ref_head(delayed_refs, bytenr);
3152         if (!head) {
3153                 spin_unlock(&delayed_refs->lock);
3154                 return 0;
3155         }
3156
3157         if (!mutex_trylock(&head->mutex)) {
3158                 refcount_inc(&head->refs);
3159                 spin_unlock(&delayed_refs->lock);
3160
3161                 btrfs_release_path(path);
3162
3163                 /*
3164                  * Mutex was contended, block until it's released and let
3165                  * caller try again
3166                  */
3167                 mutex_lock(&head->mutex);
3168                 mutex_unlock(&head->mutex);
3169                 btrfs_put_delayed_ref_head(head);
3170                 return -EAGAIN;
3171         }
3172         spin_unlock(&delayed_refs->lock);
3173
3174         spin_lock(&head->lock);
3175         /*
3176          * XXX: We should replace this with a proper search function in the
3177          * future.
3178          */
3179         for (node = rb_first(&head->ref_tree); node; node = rb_next(node)) {
3180                 ref = rb_entry(node, struct btrfs_delayed_ref_node, ref_node);
3181                 /* If it's a shared ref we know a cross reference exists */
3182                 if (ref->type != BTRFS_EXTENT_DATA_REF_KEY) {
3183                         ret = 1;
3184                         break;
3185                 }
3186
3187                 data_ref = btrfs_delayed_node_to_data_ref(ref);
3188
3189                 /*
3190                  * If our ref doesn't match the one we're currently looking at
3191                  * then we have a cross reference.
3192                  */
3193                 if (data_ref->root != root->root_key.objectid ||
3194                     data_ref->objectid != objectid ||
3195                     data_ref->offset != offset) {
3196                         ret = 1;
3197                         break;
3198                 }
3199         }
3200         spin_unlock(&head->lock);
3201         mutex_unlock(&head->mutex);
3202         return ret;
3203 }
3204
3205 static noinline int check_committed_ref(struct btrfs_root *root,
3206                                         struct btrfs_path *path,
3207                                         u64 objectid, u64 offset, u64 bytenr)
3208 {
3209         struct btrfs_fs_info *fs_info = root->fs_info;
3210         struct btrfs_root *extent_root = fs_info->extent_root;
3211         struct extent_buffer *leaf;
3212         struct btrfs_extent_data_ref *ref;
3213         struct btrfs_extent_inline_ref *iref;
3214         struct btrfs_extent_item *ei;
3215         struct btrfs_key key;
3216         u32 item_size;
3217         int type;
3218         int ret;
3219
3220         key.objectid = bytenr;
3221         key.offset = (u64)-1;
3222         key.type = BTRFS_EXTENT_ITEM_KEY;
3223
3224         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
3225         if (ret < 0)
3226                 goto out;
3227         BUG_ON(ret == 0); /* Corruption */
3228
3229         ret = -ENOENT;
3230         if (path->slots[0] == 0)
3231                 goto out;
3232
3233         path->slots[0]--;
3234         leaf = path->nodes[0];
3235         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
3236
3237         if (key.objectid != bytenr || key.type != BTRFS_EXTENT_ITEM_KEY)
3238                 goto out;
3239
3240         ret = 1;
3241         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
3242 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
3243         if (item_size < sizeof(*ei)) {
3244                 WARN_ON(item_size != sizeof(struct btrfs_extent_item_v0));
3245                 goto out;
3246         }
3247 #endif
3248         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
3249
3250         if (item_size != sizeof(*ei) +
3251             btrfs_extent_inline_ref_size(BTRFS_EXTENT_DATA_REF_KEY))
3252                 goto out;
3253
3254         if (btrfs_extent_generation(leaf, ei) <=
3255             btrfs_root_last_snapshot(&root->root_item))
3256                 goto out;
3257
3258         iref = (struct btrfs_extent_inline_ref *)(ei + 1);
3259
3260         type = btrfs_get_extent_inline_ref_type(leaf, iref, BTRFS_REF_TYPE_DATA);
3261         if (type != BTRFS_EXTENT_DATA_REF_KEY)
3262                 goto out;
3263
3264         ref = (struct btrfs_extent_data_ref *)(&iref->offset);
3265         if (btrfs_extent_refs(leaf, ei) !=
3266             btrfs_extent_data_ref_count(leaf, ref) ||
3267             btrfs_extent_data_ref_root(leaf, ref) !=
3268             root->root_key.objectid ||
3269             btrfs_extent_data_ref_objectid(leaf, ref) != objectid ||
3270             btrfs_extent_data_ref_offset(leaf, ref) != offset)
3271                 goto out;
3272
3273         ret = 0;
3274 out:
3275         return ret;
3276 }
3277
3278 int btrfs_cross_ref_exist(struct btrfs_root *root, u64 objectid, u64 offset,
3279                           u64 bytenr)
3280 {
3281         struct btrfs_path *path;
3282         int ret;
3283         int ret2;
3284
3285         path = btrfs_alloc_path();
3286         if (!path)
3287                 return -ENOENT;
3288
3289         do {
3290                 ret = check_committed_ref(root, path, objectid,
3291                                           offset, bytenr);
3292                 if (ret && ret != -ENOENT)
3293                         goto out;
3294
3295                 ret2 = check_delayed_ref(root, path, objectid,
3296                                          offset, bytenr);
3297         } while (ret2 == -EAGAIN);
3298
3299         if (ret2 && ret2 != -ENOENT) {
3300                 ret = ret2;
3301                 goto out;
3302         }
3303
3304         if (ret != -ENOENT || ret2 != -ENOENT)
3305                 ret = 0;
3306 out:
3307         btrfs_free_path(path);
3308         if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID)
3309                 WARN_ON(ret > 0);
3310         return ret;
3311 }
3312
3313 static int __btrfs_mod_ref(struct btrfs_trans_handle *trans,
3314                            struct btrfs_root *root,
3315                            struct extent_buffer *buf,
3316                            int full_backref, int inc)
3317 {
3318         struct btrfs_fs_info *fs_info = root->fs_info;
3319         u64 bytenr;
3320         u64 num_bytes;
3321         u64 parent;
3322         u64 ref_root;
3323         u32 nritems;
3324         struct btrfs_key key;
3325         struct btrfs_file_extent_item *fi;
3326         int i;
3327         int level;
3328         int ret = 0;
3329         int (*process_func)(struct btrfs_trans_handle *,
3330                             struct btrfs_root *,
3331                             u64, u64, u64, u64, u64, u64);
3332
3333
3334         if (btrfs_is_testing(fs_info))
3335                 return 0;
3336
3337         ref_root = btrfs_header_owner(buf);
3338         nritems = btrfs_header_nritems(buf);
3339         level = btrfs_header_level(buf);
3340
3341         if (!test_bit(BTRFS_ROOT_REF_COWS, &root->state) && level == 0)
3342                 return 0;
3343
3344         if (inc)
3345                 process_func = btrfs_inc_extent_ref;
3346         else
3347                 process_func = btrfs_free_extent;
3348
3349         if (full_backref)
3350                 parent = buf->start;
3351         else
3352                 parent = 0;
3353
3354         for (i = 0; i < nritems; i++) {
3355                 if (level == 0) {
3356                         btrfs_item_key_to_cpu(buf, &key, i);
3357                         if (key.type != BTRFS_EXTENT_DATA_KEY)
3358                                 continue;
3359                         fi = btrfs_item_ptr(buf, i,
3360                                             struct btrfs_file_extent_item);
3361                         if (btrfs_file_extent_type(buf, fi) ==
3362                             BTRFS_FILE_EXTENT_INLINE)
3363                                 continue;
3364                         bytenr = btrfs_file_extent_disk_bytenr(buf, fi);
3365                         if (bytenr == 0)
3366                                 continue;
3367
3368                         num_bytes = btrfs_file_extent_disk_num_bytes(buf, fi);
3369                         key.offset -= btrfs_file_extent_offset(buf, fi);
3370                         ret = process_func(trans, root, bytenr, num_bytes,
3371                                            parent, ref_root, key.objectid,
3372                                            key.offset);
3373                         if (ret)
3374                                 goto fail;
3375                 } else {
3376                         bytenr = btrfs_node_blockptr(buf, i);
3377                         num_bytes = fs_info->nodesize;
3378                         ret = process_func(trans, root, bytenr, num_bytes,
3379                                            parent, ref_root, level - 1, 0);
3380                         if (ret)
3381                                 goto fail;
3382                 }
3383         }
3384         return 0;
3385 fail:
3386         return ret;
3387 }
3388
3389 int btrfs_inc_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
3390                   struct extent_buffer *buf, int full_backref)
3391 {
3392         return __btrfs_mod_ref(trans, root, buf, full_backref, 1);
3393 }
3394
3395 int btrfs_dec_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
3396                   struct extent_buffer *buf, int full_backref)
3397 {
3398         return __btrfs_mod_ref(trans, root, buf, full_backref, 0);
3399 }
3400
3401 static int write_one_cache_group(struct btrfs_trans_handle *trans,
3402                                  struct btrfs_fs_info *fs_info,
3403                                  struct btrfs_path *path,
3404                                  struct btrfs_block_group_cache *cache)
3405 {
3406         int ret;
3407         struct btrfs_root *extent_root = fs_info->extent_root;
3408         unsigned long bi;
3409         struct extent_buffer *leaf;
3410
3411         ret = btrfs_search_slot(trans, extent_root, &cache->key, path, 0, 1);
3412         if (ret) {
3413                 if (ret > 0)
3414                         ret = -ENOENT;
3415                 goto fail;
3416         }
3417
3418         leaf = path->nodes[0];
3419         bi = btrfs_item_ptr_offset(leaf, path->slots[0]);
3420         write_extent_buffer(leaf, &cache->item, bi, sizeof(cache->item));
3421         btrfs_mark_buffer_dirty(leaf);
3422 fail:
3423         btrfs_release_path(path);
3424         return ret;
3425
3426 }
3427
3428 static struct btrfs_block_group_cache *
3429 next_block_group(struct btrfs_fs_info *fs_info,
3430                  struct btrfs_block_group_cache *cache)
3431 {
3432         struct rb_node *node;
3433
3434         spin_lock(&fs_info->block_group_cache_lock);
3435
3436         /* If our block group was removed, we need a full search. */
3437         if (RB_EMPTY_NODE(&cache->cache_node)) {
3438                 const u64 next_bytenr = cache->key.objectid + cache->key.offset;
3439
3440                 spin_unlock(&fs_info->block_group_cache_lock);
3441                 btrfs_put_block_group(cache);
3442                 cache = btrfs_lookup_first_block_group(fs_info, next_bytenr); return cache;
3443         }
3444         node = rb_next(&cache->cache_node);
3445         btrfs_put_block_group(cache);
3446         if (node) {
3447                 cache = rb_entry(node, struct btrfs_block_group_cache,
3448                                  cache_node);
3449                 btrfs_get_block_group(cache);
3450         } else
3451                 cache = NULL;
3452         spin_unlock(&fs_info->block_group_cache_lock);
3453         return cache;
3454 }
3455
3456 static int cache_save_setup(struct btrfs_block_group_cache *block_group,
3457                             struct btrfs_trans_handle *trans,
3458                             struct btrfs_path *path)
3459 {
3460         struct btrfs_fs_info *fs_info = block_group->fs_info;
3461         struct btrfs_root *root = fs_info->tree_root;
3462         struct inode *inode = NULL;
3463         struct extent_changeset *data_reserved = NULL;
3464         u64 alloc_hint = 0;
3465         int dcs = BTRFS_DC_ERROR;
3466         u64 num_pages = 0;
3467         int retries = 0;
3468         int ret = 0;
3469
3470         /*
3471          * If this block group is smaller than 100 megs don't bother caching the
3472          * block group.
3473          */
3474         if (block_group->key.offset < (100 * SZ_1M)) {
3475                 spin_lock(&block_group->lock);
3476                 block_group->disk_cache_state = BTRFS_DC_WRITTEN;
3477                 spin_unlock(&block_group->lock);
3478                 return 0;
3479         }
3480
3481         if (trans->aborted)
3482                 return 0;
3483 again:
3484         inode = lookup_free_space_inode(fs_info, block_group, path);
3485         if (IS_ERR(inode) && PTR_ERR(inode) != -ENOENT) {
3486                 ret = PTR_ERR(inode);
3487                 btrfs_release_path(path);
3488                 goto out;
3489         }
3490
3491         if (IS_ERR(inode)) {
3492                 BUG_ON(retries);
3493                 retries++;
3494
3495                 if (block_group->ro)
3496                         goto out_free;
3497
3498                 ret = create_free_space_inode(fs_info, trans, block_group,
3499                                               path);
3500                 if (ret)
3501                         goto out_free;
3502                 goto again;
3503         }
3504
3505         /*
3506          * We want to set the generation to 0, that way if anything goes wrong
3507          * from here on out we know not to trust this cache when we load up next
3508          * time.
3509          */
3510         BTRFS_I(inode)->generation = 0;
3511         ret = btrfs_update_inode(trans, root, inode);
3512         if (ret) {
3513                 /*
3514                  * So theoretically we could recover from this, simply set the
3515                  * super cache generation to 0 so we know to invalidate the
3516                  * cache, but then we'd have to keep track of the block groups
3517                  * that fail this way so we know we _have_ to reset this cache
3518                  * before the next commit or risk reading stale cache.  So to
3519                  * limit our exposure to horrible edge cases lets just abort the
3520                  * transaction, this only happens in really bad situations
3521                  * anyway.
3522                  */
3523                 btrfs_abort_transaction(trans, ret);
3524                 goto out_put;
3525         }
3526         WARN_ON(ret);
3527
3528         /* We've already setup this transaction, go ahead and exit */
3529         if (block_group->cache_generation == trans->transid &&
3530             i_size_read(inode)) {
3531                 dcs = BTRFS_DC_SETUP;
3532                 goto out_put;
3533         }
3534
3535         if (i_size_read(inode) > 0) {
3536                 ret = btrfs_check_trunc_cache_free_space(fs_info,
3537                                         &fs_info->global_block_rsv);
3538                 if (ret)
3539                         goto out_put;
3540
3541                 ret = btrfs_truncate_free_space_cache(trans, NULL, inode);
3542                 if (ret)
3543                         goto out_put;
3544         }
3545
3546         spin_lock(&block_group->lock);
3547         if (block_group->cached != BTRFS_CACHE_FINISHED ||
3548             !btrfs_test_opt(fs_info, SPACE_CACHE)) {
3549                 /*
3550                  * don't bother trying to write stuff out _if_
3551                  * a) we're not cached,
3552                  * b) we're with nospace_cache mount option,
3553                  * c) we're with v2 space_cache (FREE_SPACE_TREE).
3554                  */
3555                 dcs = BTRFS_DC_WRITTEN;
3556                 spin_unlock(&block_group->lock);
3557                 goto out_put;
3558         }
3559         spin_unlock(&block_group->lock);
3560
3561         /*
3562          * We hit an ENOSPC when setting up the cache in this transaction, just
3563          * skip doing the setup, we've already cleared the cache so we're safe.
3564          */
3565         if (test_bit(BTRFS_TRANS_CACHE_ENOSPC, &trans->transaction->flags)) {
3566                 ret = -ENOSPC;
3567                 goto out_put;
3568         }
3569
3570         /*
3571          * Try to preallocate enough space based on how big the block group is.
3572          * Keep in mind this has to include any pinned space which could end up
3573          * taking up quite a bit since it's not folded into the other space
3574          * cache.
3575          */
3576         num_pages = div_u64(block_group->key.offset, SZ_256M);
3577         if (!num_pages)
3578                 num_pages = 1;
3579
3580         num_pages *= 16;
3581         num_pages *= PAGE_SIZE;
3582
3583         ret = btrfs_check_data_free_space(inode, &data_reserved, 0, num_pages);
3584         if (ret)
3585                 goto out_put;
3586
3587         ret = btrfs_prealloc_file_range_trans(inode, trans, 0, 0, num_pages,
3588                                               num_pages, num_pages,
3589                                               &alloc_hint);
3590         /*
3591          * Our cache requires contiguous chunks so that we don't modify a bunch
3592          * of metadata or split extents when writing the cache out, which means
3593          * we can enospc if we are heavily fragmented in addition to just normal
3594          * out of space conditions.  So if we hit this just skip setting up any
3595          * other block groups for this transaction, maybe we'll unpin enough
3596          * space the next time around.
3597          */
3598         if (!ret)
3599                 dcs = BTRFS_DC_SETUP;
3600         else if (ret == -ENOSPC)
3601                 set_bit(BTRFS_TRANS_CACHE_ENOSPC, &trans->transaction->flags);
3602
3603 out_put:
3604         iput(inode);
3605 out_free:
3606         btrfs_release_path(path);
3607 out:
3608         spin_lock(&block_group->lock);
3609         if (!ret && dcs == BTRFS_DC_SETUP)
3610                 block_group->cache_generation = trans->transid;
3611         block_group->disk_cache_state = dcs;
3612         spin_unlock(&block_group->lock);
3613
3614         extent_changeset_free(data_reserved);
3615         return ret;
3616 }
3617
3618 int btrfs_setup_space_cache(struct btrfs_trans_handle *trans,
3619                             struct btrfs_fs_info *fs_info)
3620 {
3621         struct btrfs_block_group_cache *cache, *tmp;
3622         struct btrfs_transaction *cur_trans = trans->transaction;
3623         struct btrfs_path *path;
3624
3625         if (list_empty(&cur_trans->dirty_bgs) ||
3626             !btrfs_test_opt(fs_info, SPACE_CACHE))
3627                 return 0;
3628
3629         path = btrfs_alloc_path();
3630         if (!path)
3631                 return -ENOMEM;
3632
3633         /* Could add new block groups, use _safe just in case */
3634         list_for_each_entry_safe(cache, tmp, &cur_trans->dirty_bgs,
3635                                  dirty_list) {
3636                 if (cache->disk_cache_state == BTRFS_DC_CLEAR)
3637                         cache_save_setup(cache, trans, path);
3638         }
3639
3640         btrfs_free_path(path);
3641         return 0;
3642 }
3643
3644 /*
3645  * transaction commit does final block group cache writeback during a
3646  * critical section where nothing is allowed to change the FS.  This is
3647  * required in order for the cache to actually match the block group,
3648  * but can introduce a lot of latency into the commit.
3649  *
3650  * So, btrfs_start_dirty_block_groups is here to kick off block group
3651  * cache IO.  There's a chance we'll have to redo some of it if the
3652  * block group changes again during the commit, but it greatly reduces
3653  * the commit latency by getting rid of the easy block groups while
3654  * we're still allowing others to join the commit.
3655  */
3656 int btrfs_start_dirty_block_groups(struct btrfs_trans_handle *trans)
3657 {
3658         struct btrfs_fs_info *fs_info = trans->fs_info;
3659         struct btrfs_block_group_cache *cache;
3660         struct btrfs_transaction *cur_trans = trans->transaction;
3661         int ret = 0;
3662         int should_put;
3663         struct btrfs_path *path = NULL;
3664         LIST_HEAD(dirty);
3665         struct list_head *io = &cur_trans->io_bgs;
3666         int num_started = 0;
3667         int loops = 0;
3668
3669         spin_lock(&cur_trans->dirty_bgs_lock);
3670         if (list_empty(&cur_trans->dirty_bgs)) {
3671                 spin_unlock(&cur_trans->dirty_bgs_lock);
3672                 return 0;
3673         }
3674         list_splice_init(&cur_trans->dirty_bgs, &dirty);
3675         spin_unlock(&cur_trans->dirty_bgs_lock);
3676
3677 again:
3678         /*
3679          * make sure all the block groups on our dirty list actually
3680          * exist
3681          */
3682         btrfs_create_pending_block_groups(trans);
3683
3684         if (!path) {
3685                 path = btrfs_alloc_path();
3686                 if (!path)
3687                         return -ENOMEM;
3688         }
3689
3690         /*
3691          * cache_write_mutex is here only to save us from balance or automatic
3692          * removal of empty block groups deleting this block group while we are
3693          * writing out the cache
3694          */
3695         mutex_lock(&trans->transaction->cache_write_mutex);
3696         while (!list_empty(&dirty)) {
3697                 cache = list_first_entry(&dirty,
3698                                          struct btrfs_block_group_cache,
3699                                          dirty_list);
3700                 /*
3701                  * this can happen if something re-dirties a block
3702                  * group that is already under IO.  Just wait for it to
3703                  * finish and then do it all again
3704                  */
3705                 if (!list_empty(&cache->io_list)) {
3706                         list_del_init(&cache->io_list);
3707                         btrfs_wait_cache_io(trans, cache, path);
3708                         btrfs_put_block_group(cache);
3709                 }
3710
3711
3712                 /*
3713                  * btrfs_wait_cache_io uses the cache->dirty_list to decide
3714                  * if it should update the cache_state.  Don't delete
3715                  * until after we wait.
3716                  *
3717                  * Since we're not running in the commit critical section
3718                  * we need the dirty_bgs_lock to protect from update_block_group
3719                  */
3720                 spin_lock(&cur_trans->dirty_bgs_lock);
3721                 list_del_init(&cache->dirty_list);
3722                 spin_unlock(&cur_trans->dirty_bgs_lock);
3723
3724                 should_put = 1;
3725
3726                 cache_save_setup(cache, trans, path);
3727
3728                 if (cache->disk_cache_state == BTRFS_DC_SETUP) {
3729                         cache->io_ctl.inode = NULL;
3730                         ret = btrfs_write_out_cache(fs_info, trans,
3731                                                     cache, path);
3732                         if (ret == 0 && cache->io_ctl.inode) {
3733                                 num_started++;
3734                                 should_put = 0;
3735
3736                                 /*
3737                                  * The cache_write_mutex is protecting the
3738                                  * io_list, also refer to the definition of
3739                                  * btrfs_transaction::io_bgs for more details
3740                                  */
3741                                 list_add_tail(&cache->io_list, io);
3742                         } else {
3743                                 /*
3744                                  * if we failed to write the cache, the
3745                                  * generation will be bad and life goes on
3746                                  */
3747                                 ret = 0;
3748                         }
3749                 }
3750                 if (!ret) {
3751                         ret = write_one_cache_group(trans, fs_info,
3752                                                     path, cache);
3753                         /*
3754                          * Our block group might still be attached to the list
3755                          * of new block groups in the transaction handle of some
3756                          * other task (struct btrfs_trans_handle->new_bgs). This
3757                          * means its block group item isn't yet in the extent
3758                          * tree. If this happens ignore the error, as we will
3759                          * try again later in the critical section of the
3760                          * transaction commit.
3761                          */
3762                         if (ret == -ENOENT) {
3763                                 ret = 0;
3764                                 spin_lock(&cur_trans->dirty_bgs_lock);
3765                                 if (list_empty(&cache->dirty_list)) {
3766                                         list_add_tail(&cache->dirty_list,
3767                                                       &cur_trans->dirty_bgs);
3768                                         btrfs_get_block_group(cache);
3769                                 }
3770                                 spin_unlock(&cur_trans->dirty_bgs_lock);
3771                         } else if (ret) {
3772                                 btrfs_abort_transaction(trans, ret);
3773                         }
3774                 }
3775
3776                 /* if its not on the io list, we need to put the block group */
3777                 if (should_put)
3778                         btrfs_put_block_group(cache);
3779
3780                 if (ret)
3781                         break;
3782
3783                 /*
3784                  * Avoid blocking other tasks for too long. It might even save
3785                  * us from writing caches for block groups that are going to be
3786                  * removed.
3787                  */
3788                 mutex_unlock(&trans->transaction->cache_write_mutex);
3789                 mutex_lock(&trans->transaction->cache_write_mutex);
3790         }
3791         mutex_unlock(&trans->transaction->cache_write_mutex);
3792
3793         /*
3794          * go through delayed refs for all the stuff we've just kicked off
3795          * and then loop back (just once)
3796          */
3797         ret = btrfs_run_delayed_refs(trans, 0);
3798         if (!ret && loops == 0) {
3799                 loops++;
3800                 spin_lock(&cur_trans->dirty_bgs_lock);
3801                 list_splice_init(&cur_trans->dirty_bgs, &dirty);
3802                 /*
3803                  * dirty_bgs_lock protects us from concurrent block group
3804                  * deletes too (not just cache_write_mutex).
3805                  */
3806                 if (!list_empty(&dirty)) {
3807                         spin_unlock(&cur_trans->dirty_bgs_lock);
3808                         goto again;
3809                 }
3810                 spin_unlock(&cur_trans->dirty_bgs_lock);
3811         } else if (ret < 0) {
3812                 btrfs_cleanup_dirty_bgs(cur_trans, fs_info);
3813         }
3814
3815         btrfs_free_path(path);
3816         return ret;
3817 }
3818
3819 int btrfs_write_dirty_block_groups(struct btrfs_trans_handle *trans,
3820                                    struct btrfs_fs_info *fs_info)
3821 {
3822         struct btrfs_block_group_cache *cache;
3823         struct btrfs_transaction *cur_trans = trans->transaction;
3824         int ret = 0;
3825         int should_put;
3826         struct btrfs_path *path;
3827         struct list_head *io = &cur_trans->io_bgs;
3828         int num_started = 0;
3829
3830         path = btrfs_alloc_path();
3831         if (!path)
3832                 return -ENOMEM;
3833
3834         /*
3835          * Even though we are in the critical section of the transaction commit,
3836          * we can still have concurrent tasks adding elements to this
3837          * transaction's list of dirty block groups. These tasks correspond to
3838          * endio free space workers started when writeback finishes for a
3839          * space cache, which run inode.c:btrfs_finish_ordered_io(), and can
3840          * allocate new block groups as a result of COWing nodes of the root
3841          * tree when updating the free space inode. The writeback for the space
3842          * caches is triggered by an earlier call to
3843          * btrfs_start_dirty_block_groups() and iterations of the following
3844          * loop.
3845          * Also we want to do the cache_save_setup first and then run the
3846          * delayed refs to make sure we have the best chance at doing this all
3847          * in one shot.
3848          */
3849         spin_lock(&cur_trans->dirty_bgs_lock);
3850         while (!list_empty(&cur_trans->dirty_bgs)) {
3851                 cache = list_first_entry(&cur_trans->dirty_bgs,
3852                                          struct btrfs_block_group_cache,
3853                                          dirty_list);
3854
3855                 /*
3856                  * this can happen if cache_save_setup re-dirties a block
3857                  * group that is already under IO.  Just wait for it to
3858                  * finish and then do it all again
3859                  */
3860                 if (!list_empty(&cache->io_list)) {
3861                         spin_unlock(&cur_trans->dirty_bgs_lock);
3862                         list_del_init(&cache->io_list);
3863                         btrfs_wait_cache_io(trans, cache, path);
3864                         btrfs_put_block_group(cache);
3865                         spin_lock(&cur_trans->dirty_bgs_lock);
3866                 }
3867
3868                 /*
3869                  * don't remove from the dirty list until after we've waited
3870                  * on any pending IO
3871                  */
3872                 list_del_init(&cache->dirty_list);
3873                 spin_unlock(&cur_trans->dirty_bgs_lock);
3874                 should_put = 1;
3875
3876                 cache_save_setup(cache, trans, path);
3877
3878                 if (!ret)
3879                         ret = btrfs_run_delayed_refs(trans,
3880                                                      (unsigned long) -1);
3881
3882                 if (!ret && cache->disk_cache_state == BTRFS_DC_SETUP) {
3883                         cache->io_ctl.inode = NULL;
3884                         ret = btrfs_write_out_cache(fs_info, trans,
3885                                                     cache, path);
3886                         if (ret == 0 && cache->io_ctl.inode) {
3887                                 num_started++;
3888                                 should_put = 0;
3889                                 list_add_tail(&cache->io_list, io);
3890                         } else {
3891                                 /*
3892                                  * if we failed to write the cache, the
3893                                  * generation will be bad and life goes on
3894                                  */
3895                                 ret = 0;
3896                         }
3897                 }
3898                 if (!ret) {
3899                         ret = write_one_cache_group(trans, fs_info,
3900                                                     path, cache);
3901                         /*
3902                          * One of the free space endio workers might have
3903                          * created a new block group while updating a free space
3904                          * cache's inode (at inode.c:btrfs_finish_ordered_io())
3905                          * and hasn't released its transaction handle yet, in
3906                          * which case the new block group is still attached to
3907                          * its transaction handle and its creation has not
3908                          * finished yet (no block group item in the extent tree
3909                          * yet, etc). If this is the case, wait for all free
3910                          * space endio workers to finish and retry. This is a
3911                          * a very rare case so no need for a more efficient and
3912                          * complex approach.
3913                          */
3914                         if (ret == -ENOENT) {
3915                                 wait_event(cur_trans->writer_wait,
3916                                    atomic_read(&cur_trans->num_writers) == 1);
3917                                 ret = write_one_cache_group(trans, fs_info,
3918                                                             path, cache);
3919                         }
3920                         if (ret)
3921                                 btrfs_abort_transaction(trans, ret);
3922                 }
3923
3924                 /* if its not on the io list, we need to put the block group */
3925                 if (should_put)
3926                         btrfs_put_block_group(cache);
3927                 spin_lock(&cur_trans->dirty_bgs_lock);
3928         }
3929         spin_unlock(&cur_trans->dirty_bgs_lock);
3930
3931         /*
3932          * Refer to the definition of io_bgs member for details why it's safe
3933          * to use it without any locking
3934          */
3935         while (!list_empty(io)) {
3936                 cache = list_first_entry(io, struct btrfs_block_group_cache,
3937                                          io_list);
3938                 list_del_init(&cache->io_list);
3939                 btrfs_wait_cache_io(trans, cache, path);
3940                 btrfs_put_block_group(cache);
3941         }
3942
3943         btrfs_free_path(path);
3944         return ret;
3945 }
3946
3947 int btrfs_extent_readonly(struct btrfs_fs_info *fs_info, u64 bytenr)
3948 {
3949         struct btrfs_block_group_cache *block_group;
3950         int readonly = 0;
3951
3952         block_group = btrfs_lookup_block_group(fs_info, bytenr);
3953         if (!block_group || block_group->ro)
3954                 readonly = 1;
3955         if (block_group)
3956                 btrfs_put_block_group(block_group);
3957         return readonly;
3958 }
3959
3960 bool btrfs_inc_nocow_writers(struct btrfs_fs_info *fs_info, u64 bytenr)
3961 {
3962         struct btrfs_block_group_cache *bg;
3963         bool ret = true;
3964
3965         bg = btrfs_lookup_block_group(fs_info, bytenr);
3966         if (!bg)
3967                 return false;
3968
3969         spin_lock(&bg->lock);
3970         if (bg->ro)
3971                 ret = false;
3972         else
3973                 atomic_inc(&bg->nocow_writers);
3974         spin_unlock(&bg->lock);
3975
3976         /* no put on block group, done by btrfs_dec_nocow_writers */
3977         if (!ret)
3978                 btrfs_put_block_group(bg);
3979
3980         return ret;
3981
3982 }
3983
3984 void btrfs_dec_nocow_writers(struct btrfs_fs_info *fs_info, u64 bytenr)
3985 {
3986         struct btrfs_block_group_cache *bg;
3987
3988         bg = btrfs_lookup_block_group(fs_info, bytenr);
3989         ASSERT(bg);
3990         if (atomic_dec_and_test(&bg->nocow_writers))
3991                 wake_up_atomic_t(&bg->nocow_writers);
3992         /*
3993          * Once for our lookup and once for the lookup done by a previous call
3994          * to btrfs_inc_nocow_writers()
3995          */
3996         btrfs_put_block_group(bg);
3997         btrfs_put_block_group(bg);
3998 }
3999
4000 void btrfs_wait_nocow_writers(struct btrfs_block_group_cache *bg)
4001 {
4002         wait_on_atomic_t(&bg->nocow_writers, atomic_t_wait,
4003                          TASK_UNINTERRUPTIBLE);
4004 }
4005
4006 static const char *alloc_name(u64 flags)
4007 {
4008         switch (flags) {
4009         case BTRFS_BLOCK_GROUP_METADATA|BTRFS_BLOCK_GROUP_DATA:
4010                 return "mixed";
4011         case BTRFS_BLOCK_GROUP_METADATA:
4012                 return "metadata";
4013         case BTRFS_BLOCK_GROUP_DATA:
4014                 return "data";
4015         case BTRFS_BLOCK_GROUP_SYSTEM:
4016                 return "system";
4017         default:
4018                 WARN_ON(1);
4019                 return "invalid-combination";
4020         };
4021 }
4022
4023 static int create_space_info(struct btrfs_fs_info *info, u64 flags,
4024                              struct btrfs_space_info **new)
4025 {
4026
4027         struct btrfs_space_info *space_info;
4028         int i;
4029         int ret;
4030
4031         space_info = kzalloc(sizeof(*space_info), GFP_NOFS);
4032         if (!space_info)
4033                 return -ENOMEM;
4034
4035         ret = percpu_counter_init(&space_info->total_bytes_pinned, 0,
4036                                  GFP_KERNEL);
4037         if (ret) {
4038                 kfree(space_info);
4039                 return ret;
4040         }
4041
4042         for (i = 0; i < BTRFS_NR_RAID_TYPES; i++)
4043                 INIT_LIST_HEAD(&space_info->block_groups[i]);
4044         init_rwsem(&space_info->groups_sem);
4045         spin_lock_init(&space_info->lock);
4046         space_info->flags = flags & BTRFS_BLOCK_GROUP_TYPE_MASK;
4047         space_info->force_alloc = CHUNK_ALLOC_NO_FORCE;
4048         init_waitqueue_head(&space_info->wait);
4049         INIT_LIST_HEAD(&space_info->ro_bgs);
4050         INIT_LIST_HEAD(&space_info->tickets);
4051         INIT_LIST_HEAD(&space_info->priority_tickets);
4052
4053         ret = kobject_init_and_add(&space_info->kobj, &space_info_ktype,
4054                                     info->space_info_kobj, "%s",
4055                                     alloc_name(space_info->flags));
4056         if (ret) {
4057                 percpu_counter_destroy(&space_info->total_bytes_pinned);
4058                 kfree(space_info);
4059                 return ret;
4060         }
4061
4062         *new = space_info;
4063         list_add_rcu(&space_info->list, &info->space_info);
4064         if (flags & BTRFS_BLOCK_GROUP_DATA)
4065                 info->data_sinfo = space_info;
4066
4067         return ret;
4068 }
4069
4070 static void update_space_info(struct btrfs_fs_info *info, u64 flags,
4071                              u64 total_bytes, u64 bytes_used,
4072                              u64 bytes_readonly,
4073                              struct btrfs_space_info **space_info)
4074 {
4075         struct btrfs_space_info *found;
4076         int factor;
4077
4078         if (flags & (BTRFS_BLOCK_GROUP_DUP | BTRFS_BLOCK_GROUP_RAID1 |
4079                      BTRFS_BLOCK_GROUP_RAID10))
4080                 factor = 2;
4081         else
4082                 factor = 1;
4083
4084         found = __find_space_info(info, flags);
4085         ASSERT(found);
4086         spin_lock(&found->lock);
4087         found->total_bytes += total_bytes;
4088         found->disk_total += total_bytes * factor;
4089         found->bytes_used += bytes_used;
4090         found->disk_used += bytes_used * factor;
4091         found->bytes_readonly += bytes_readonly;
4092         if (total_bytes > 0)
4093                 found->full = 0;
4094         space_info_add_new_bytes(info, found, total_bytes -
4095                                  bytes_used - bytes_readonly);
4096         spin_unlock(&found->lock);
4097         *space_info = found;
4098 }
4099
4100 static void set_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
4101 {
4102         u64 extra_flags = chunk_to_extended(flags) &
4103                                 BTRFS_EXTENDED_PROFILE_MASK;
4104
4105         write_seqlock(&fs_info->profiles_lock);
4106         if (flags & BTRFS_BLOCK_GROUP_DATA)
4107                 fs_info->avail_data_alloc_bits |= extra_flags;
4108         if (flags & BTRFS_BLOCK_GROUP_METADATA)
4109                 fs_info->avail_metadata_alloc_bits |= extra_flags;
4110         if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
4111                 fs_info->avail_system_alloc_bits |= extra_flags;
4112         write_sequnlock(&fs_info->profiles_lock);
4113 }
4114
4115 /*
4116  * returns target flags in extended format or 0 if restripe for this
4117  * chunk_type is not in progress
4118  *
4119  * should be called with either volume_mutex or balance_lock held
4120  */
4121 static u64 get_restripe_target(struct btrfs_fs_info *fs_info, u64 flags)
4122 {
4123         struct btrfs_balance_control *bctl = fs_info->balance_ctl;
4124         u64 target = 0;
4125
4126         if (!bctl)
4127                 return 0;
4128
4129         if (flags & BTRFS_BLOCK_GROUP_DATA &&
4130             bctl->data.flags & BTRFS_BALANCE_ARGS_CONVERT) {
4131                 target = BTRFS_BLOCK_GROUP_DATA | bctl->data.target;
4132         } else if (flags & BTRFS_BLOCK_GROUP_SYSTEM &&
4133                    bctl->sys.flags & BTRFS_BALANCE_ARGS_CONVERT) {
4134                 target = BTRFS_BLOCK_GROUP_SYSTEM | bctl->sys.target;
4135         } else if (flags & BTRFS_BLOCK_GROUP_METADATA &&
4136                    bctl->meta.flags & BTRFS_BALANCE_ARGS_CONVERT) {
4137                 target = BTRFS_BLOCK_GROUP_METADATA | bctl->meta.target;
4138         }
4139
4140         return target;
4141 }
4142
4143 /*
4144  * @flags: available profiles in extended format (see ctree.h)
4145  *
4146  * Returns reduced profile in chunk format.  If profile changing is in
4147  * progress (either running or paused) picks the target profile (if it's
4148  * already available), otherwise falls back to plain reducing.
4149  */
4150 static u64 btrfs_reduce_alloc_profile(struct btrfs_fs_info *fs_info, u64 flags)
4151 {
4152         u64 num_devices = fs_info->fs_devices->rw_devices;
4153         u64 target;
4154         u64 raid_type;
4155         u64 allowed = 0;
4156
4157         /*
4158          * see if restripe for this chunk_type is in progress, if so
4159          * try to reduce to the target profile
4160          */
4161         spin_lock(&fs_info->balance_lock);
4162         target = get_restripe_target(fs_info, flags);
4163         if (target) {
4164                 /* pick target profile only if it's already available */
4165                 if ((flags & target) & BTRFS_EXTENDED_PROFILE_MASK) {
4166                         spin_unlock(&fs_info->balance_lock);
4167                         return extended_to_chunk(target);
4168                 }
4169         }
4170         spin_unlock(&fs_info->balance_lock);
4171
4172         /* First, mask out the RAID levels which aren't possible */
4173         for (raid_type = 0; raid_type < BTRFS_NR_RAID_TYPES; raid_type++) {
4174                 if (num_devices >= btrfs_raid_array[raid_type].devs_min)
4175                         allowed |= btrfs_raid_group[raid_type];
4176         }
4177         allowed &= flags;
4178
4179         if (allowed & BTRFS_BLOCK_GROUP_RAID6)
4180                 allowed = BTRFS_BLOCK_GROUP_RAID6;
4181         else if (allowed & BTRFS_BLOCK_GROUP_RAID5)
4182                 allowed = BTRFS_BLOCK_GROUP_RAID5;
4183         else if (allowed & BTRFS_BLOCK_GROUP_RAID10)
4184                 allowed = BTRFS_BLOCK_GROUP_RAID10;
4185         else if (allowed & BTRFS_BLOCK_GROUP_RAID1)
4186                 allowed = BTRFS_BLOCK_GROUP_RAID1;
4187         else if (allowed & BTRFS_BLOCK_GROUP_RAID0)
4188                 allowed = BTRFS_BLOCK_GROUP_RAID0;
4189
4190         flags &= ~BTRFS_BLOCK_GROUP_PROFILE_MASK;
4191
4192         return extended_to_chunk(flags | allowed);
4193 }
4194
4195 static u64 get_alloc_profile(struct btrfs_fs_info *fs_info, u64 orig_flags)
4196 {
4197         unsigned seq;
4198         u64 flags;
4199
4200         do {
4201                 flags = orig_flags;
4202                 seq = read_seqbegin(&fs_info->profiles_lock);
4203
4204                 if (flags & BTRFS_BLOCK_GROUP_DATA)
4205                         flags |= fs_info->avail_data_alloc_bits;
4206                 else if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
4207                         flags |= fs_info->avail_system_alloc_bits;
4208                 else if (flags & BTRFS_BLOCK_GROUP_METADATA)
4209                         flags |= fs_info->avail_metadata_alloc_bits;
4210         } while (read_seqretry(&fs_info->profiles_lock, seq));
4211
4212         return btrfs_reduce_alloc_profile(fs_info, flags);
4213 }
4214
4215 static u64 get_alloc_profile_by_root(struct btrfs_root *root, int data)
4216 {
4217         struct btrfs_fs_info *fs_info = root->fs_info;
4218         u64 flags;
4219         u64 ret;
4220
4221         if (data)
4222                 flags = BTRFS_BLOCK_GROUP_DATA;
4223         else if (root == fs_info->chunk_root)
4224                 flags = BTRFS_BLOCK_GROUP_SYSTEM;
4225         else
4226                 flags = BTRFS_BLOCK_GROUP_METADATA;
4227
4228         ret = get_alloc_profile(fs_info, flags);
4229         return ret;
4230 }
4231
4232 u64 btrfs_data_alloc_profile(struct btrfs_fs_info *fs_info)
4233 {
4234         return get_alloc_profile(fs_info, BTRFS_BLOCK_GROUP_DATA);
4235 }
4236
4237 u64 btrfs_metadata_alloc_profile(struct btrfs_fs_info *fs_info)
4238 {
4239         return get_alloc_profile(fs_info, BTRFS_BLOCK_GROUP_METADATA);
4240 }
4241
4242 u64 btrfs_system_alloc_profile(struct btrfs_fs_info *fs_info)
4243 {
4244         return get_alloc_profile(fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
4245 }
4246
4247 static u64 btrfs_space_info_used(struct btrfs_space_info *s_info,
4248                                  bool may_use_included)
4249 {
4250         ASSERT(s_info);
4251         return s_info->bytes_used + s_info->bytes_reserved +
4252                 s_info->bytes_pinned + s_info->bytes_readonly +
4253                 (may_use_included ? s_info->bytes_may_use : 0);
4254 }
4255
4256 int btrfs_alloc_data_chunk_ondemand(struct btrfs_inode *inode, u64 bytes)
4257 {
4258         struct btrfs_root *root = inode->root;
4259         struct btrfs_fs_info *fs_info = root->fs_info;
4260         struct btrfs_space_info *data_sinfo = fs_info->data_sinfo;
4261         u64 used;
4262         int ret = 0;
4263         int need_commit = 2;
4264         int have_pinned_space;
4265
4266         /* make sure bytes are sectorsize aligned */
4267         bytes = ALIGN(bytes, fs_info->sectorsize);
4268
4269         if (btrfs_is_free_space_inode(inode)) {
4270                 need_commit = 0;
4271                 ASSERT(current->journal_info);
4272         }
4273
4274 again:
4275         /* make sure we have enough space to handle the data first */
4276         spin_lock(&data_sinfo->lock);
4277         used = btrfs_space_info_used(data_sinfo, true);
4278
4279         if (used + bytes > data_sinfo->total_bytes) {
4280                 struct btrfs_trans_handle *trans;
4281
4282                 /*
4283                  * if we don't have enough free bytes in this space then we need
4284                  * to alloc a new chunk.
4285                  */
4286                 if (!data_sinfo->full) {
4287                         u64 alloc_target;
4288
4289                         data_sinfo->force_alloc = CHUNK_ALLOC_FORCE;
4290                         spin_unlock(&data_sinfo->lock);
4291
4292                         alloc_target = btrfs_data_alloc_profile(fs_info);
4293                         /*
4294                          * It is ugly that we don't call nolock join
4295                          * transaction for the free space inode case here.
4296                          * But it is safe because we only do the data space
4297                          * reservation for the free space cache in the
4298                          * transaction context, the common join transaction
4299                          * just increase the counter of the current transaction
4300                          * handler, doesn't try to acquire the trans_lock of
4301                          * the fs.
4302                          */
4303                         trans = btrfs_join_transaction(root);
4304                         if (IS_ERR(trans))
4305                                 return PTR_ERR(trans);
4306
4307                         ret = do_chunk_alloc(trans, fs_info, alloc_target,
4308                                              CHUNK_ALLOC_NO_FORCE);
4309                         btrfs_end_transaction(trans);
4310                         if (ret < 0) {
4311                                 if (ret != -ENOSPC)
4312                                         return ret;
4313                                 else {
4314                                         have_pinned_space = 1;
4315                                         goto commit_trans;
4316                                 }
4317                         }
4318
4319                         goto again;
4320                 }
4321
4322                 /*
4323                  * If we don't have enough pinned space to deal with this
4324                  * allocation, and no removed chunk in current transaction,
4325                  * don't bother committing the transaction.
4326                  */
4327                 have_pinned_space = percpu_counter_compare(
4328                         &data_sinfo->total_bytes_pinned,
4329                         used + bytes - data_sinfo->total_bytes);
4330                 spin_unlock(&data_sinfo->lock);
4331
4332                 /* commit the current transaction and try again */
4333 commit_trans:
4334                 if (need_commit) {
4335                         need_commit--;
4336
4337                         if (need_commit > 0) {
4338                                 btrfs_start_delalloc_roots(fs_info, 0, -1);
4339                                 btrfs_wait_ordered_roots(fs_info, U64_MAX, 0,
4340                                                          (u64)-1);
4341                         }
4342
4343                         trans = btrfs_join_transaction(root);
4344                         if (IS_ERR(trans))
4345                                 return PTR_ERR(trans);
4346                         if (have_pinned_space >= 0 ||
4347                             test_bit(BTRFS_TRANS_HAVE_FREE_BGS,
4348                                      &trans->transaction->flags) ||
4349                             need_commit > 0) {
4350                                 ret = btrfs_commit_transaction(trans);
4351                                 if (ret)
4352                                         return ret;
4353                                 /*
4354                                  * The cleaner kthread might still be doing iput
4355                                  * operations. Wait for it to finish so that
4356                                  * more space is released.
4357                                  */
4358                                 mutex_lock(&fs_info->cleaner_delayed_iput_mutex);
4359                                 mutex_unlock(&fs_info->cleaner_delayed_iput_mutex);
4360                                 goto again;
4361                         } else {
4362                                 btrfs_end_transaction(trans);
4363                         }
4364                 }
4365
4366                 trace_btrfs_space_reservation(fs_info,
4367                                               "space_info:enospc",
4368                                               data_sinfo->flags, bytes, 1);
4369                 return -ENOSPC;
4370         }
4371         data_sinfo->bytes_may_use += bytes;
4372         trace_btrfs_space_reservation(fs_info, "space_info",
4373                                       data_sinfo->flags, bytes, 1);
4374         spin_unlock(&data_sinfo->lock);
4375
4376         return ret;
4377 }
4378
4379 int btrfs_check_data_free_space(struct inode *inode,
4380                         struct extent_changeset **reserved, u64 start, u64 len)
4381 {
4382         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
4383         int ret;
4384
4385         /* align the range */
4386         len = round_up(start + len, fs_info->sectorsize) -
4387               round_down(start, fs_info->sectorsize);
4388         start = round_down(start, fs_info->sectorsize);
4389
4390         ret = btrfs_alloc_data_chunk_ondemand(BTRFS_I(inode), len);
4391         if (ret < 0)
4392                 return ret;
4393
4394         /* Use new btrfs_qgroup_reserve_data to reserve precious data space. */
4395         ret = btrfs_qgroup_reserve_data(inode, reserved, start, len);
4396         if (ret < 0)
4397                 btrfs_free_reserved_data_space_noquota(inode, start, len);
4398         else
4399                 ret = 0;
4400         return ret;
4401 }
4402
4403 /*
4404  * Called if we need to clear a data reservation for this inode
4405  * Normally in a error case.
4406  *
4407  * This one will *NOT* use accurate qgroup reserved space API, just for case
4408  * which we can't sleep and is sure it won't affect qgroup reserved space.
4409  * Like clear_bit_hook().
4410  */
4411 void btrfs_free_reserved_data_space_noquota(struct inode *inode, u64 start,
4412                                             u64 len)
4413 {
4414         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
4415         struct btrfs_space_info *data_sinfo;
4416
4417         /* Make sure the range is aligned to sectorsize */
4418         len = round_up(start + len, fs_info->sectorsize) -
4419               round_down(start, fs_info->sectorsize);
4420         start = round_down(start, fs_info->sectorsize);
4421
4422         data_sinfo = fs_info->data_sinfo;
4423         spin_lock(&data_sinfo->lock);
4424         if (WARN_ON(data_sinfo->bytes_may_use < len))
4425                 data_sinfo->bytes_may_use = 0;
4426         else
4427                 data_sinfo->bytes_may_use -= len;
4428         trace_btrfs_space_reservation(fs_info, "space_info",
4429                                       data_sinfo->flags, len, 0);
4430         spin_unlock(&data_sinfo->lock);
4431 }
4432
4433 /*
4434  * Called if we need to clear a data reservation for this inode
4435  * Normally in a error case.
4436  *
4437  * This one will handle the per-inode data rsv map for accurate reserved
4438  * space framework.
4439  */
4440 void btrfs_free_reserved_data_space(struct inode *inode,
4441                         struct extent_changeset *reserved, u64 start, u64 len)
4442 {
4443         struct btrfs_root *root = BTRFS_I(inode)->root;
4444
4445         /* Make sure the range is aligned to sectorsize */
4446         len = round_up(start + len, root->fs_info->sectorsize) -
4447               round_down(start, root->fs_info->sectorsize);
4448         start = round_down(start, root->fs_info->sectorsize);
4449
4450         btrfs_free_reserved_data_space_noquota(inode, start, len);
4451         btrfs_qgroup_free_data(inode, reserved, start, len);
4452 }
4453
4454 static void force_metadata_allocation(struct btrfs_fs_info *info)
4455 {
4456         struct list_head *head = &info->space_info;
4457         struct btrfs_space_info *found;
4458
4459         rcu_read_lock();
4460         list_for_each_entry_rcu(found, head, list) {
4461                 if (found->flags & BTRFS_BLOCK_GROUP_METADATA)
4462                         found->force_alloc = CHUNK_ALLOC_FORCE;
4463         }
4464         rcu_read_unlock();
4465 }
4466
4467 static inline u64 calc_global_rsv_need_space(struct btrfs_block_rsv *global)
4468 {
4469         return (global->size << 1);
4470 }
4471
4472 static int should_alloc_chunk(struct btrfs_fs_info *fs_info,
4473                               struct btrfs_space_info *sinfo, int force)
4474 {
4475         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
4476         u64 bytes_used = btrfs_space_info_used(sinfo, false);
4477         u64 thresh;
4478
4479         if (force == CHUNK_ALLOC_FORCE)
4480                 return 1;
4481
4482         /*
4483          * We need to take into account the global rsv because for all intents
4484          * and purposes it's used space.  Don't worry about locking the
4485          * global_rsv, it doesn't change except when the transaction commits.
4486          */
4487         if (sinfo->flags & BTRFS_BLOCK_GROUP_METADATA)
4488                 bytes_used += calc_global_rsv_need_space(global_rsv);
4489
4490         /*
4491          * in limited mode, we want to have some free space up to
4492          * about 1% of the FS size.
4493          */
4494         if (force == CHUNK_ALLOC_LIMITED) {
4495                 thresh = btrfs_super_total_bytes(fs_info->super_copy);
4496                 thresh = max_t(u64, SZ_64M, div_factor_fine(thresh, 1));
4497
4498                 if (sinfo->total_bytes - bytes_used < thresh)
4499                         return 1;
4500         }
4501
4502         if (bytes_used + SZ_2M < div_factor(sinfo->total_bytes, 8))
4503                 return 0;
4504         return 1;
4505 }
4506
4507 static u64 get_profile_num_devs(struct btrfs_fs_info *fs_info, u64 type)
4508 {
4509         u64 num_dev;
4510
4511         if (type & (BTRFS_BLOCK_GROUP_RAID10 |
4512                     BTRFS_BLOCK_GROUP_RAID0 |
4513                     BTRFS_BLOCK_GROUP_RAID5 |
4514                     BTRFS_BLOCK_GROUP_RAID6))
4515                 num_dev = fs_info->fs_devices->rw_devices;
4516         else if (type & BTRFS_BLOCK_GROUP_RAID1)
4517                 num_dev = 2;
4518         else
4519                 num_dev = 1;    /* DUP or single */
4520
4521         return num_dev;
4522 }
4523
4524 /*
4525  * If @is_allocation is true, reserve space in the system space info necessary
4526  * for allocating a chunk, otherwise if it's false, reserve space necessary for
4527  * removing a chunk.
4528  */
4529 void check_system_chunk(struct btrfs_trans_handle *trans,
4530                         struct btrfs_fs_info *fs_info, u64 type)
4531 {
4532         struct btrfs_space_info *info;
4533         u64 left;
4534         u64 thresh;
4535         int ret = 0;
4536         u64 num_devs;
4537
4538         /*
4539          * Needed because we can end up allocating a system chunk and for an
4540          * atomic and race free space reservation in the chunk block reserve.
4541          */
4542         lockdep_assert_held(&fs_info->chunk_mutex);
4543
4544         info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
4545         spin_lock(&info->lock);
4546         left = info->total_bytes - btrfs_space_info_used(info, true);
4547         spin_unlock(&info->lock);
4548
4549         num_devs = get_profile_num_devs(fs_info, type);
4550
4551         /* num_devs device items to update and 1 chunk item to add or remove */
4552         thresh = btrfs_calc_trunc_metadata_size(fs_info, num_devs) +
4553                 btrfs_calc_trans_metadata_size(fs_info, 1);
4554
4555         if (left < thresh && btrfs_test_opt(fs_info, ENOSPC_DEBUG)) {
4556                 btrfs_info(fs_info, "left=%llu, need=%llu, flags=%llu",
4557                            left, thresh, type);
4558                 dump_space_info(fs_info, info, 0, 0);
4559         }
4560
4561         if (left < thresh) {
4562                 u64 flags = btrfs_system_alloc_profile(fs_info);
4563
4564                 /*
4565                  * Ignore failure to create system chunk. We might end up not
4566                  * needing it, as we might not need to COW all nodes/leafs from
4567                  * the paths we visit in the chunk tree (they were already COWed
4568                  * or created in the current transaction for example).
4569                  */
4570                 ret = btrfs_alloc_chunk(trans, fs_info, flags);
4571         }
4572
4573         if (!ret) {
4574                 ret = btrfs_block_rsv_add(fs_info->chunk_root,
4575                                           &fs_info->chunk_block_rsv,
4576                                           thresh, BTRFS_RESERVE_NO_FLUSH);
4577                 if (!ret)
4578                         trans->chunk_bytes_reserved += thresh;
4579         }
4580 }
4581
4582 /*
4583  * If force is CHUNK_ALLOC_FORCE:
4584  *    - return 1 if it successfully allocates a chunk,
4585  *    - return errors including -ENOSPC otherwise.
4586  * If force is NOT CHUNK_ALLOC_FORCE:
4587  *    - return 0 if it doesn't need to allocate a new chunk,
4588  *    - return 1 if it successfully allocates a chunk,
4589  *    - return errors including -ENOSPC otherwise.
4590  */
4591 static int do_chunk_alloc(struct btrfs_trans_handle *trans,
4592                           struct btrfs_fs_info *fs_info, u64 flags, int force)
4593 {
4594         struct btrfs_space_info *space_info;
4595         int wait_for_alloc = 0;
4596         int ret = 0;
4597
4598         /* Don't re-enter if we're already allocating a chunk */
4599         if (trans->allocating_chunk)
4600                 return -ENOSPC;
4601
4602         space_info = __find_space_info(fs_info, flags);
4603         ASSERT(space_info);
4604
4605 again:
4606         spin_lock(&space_info->lock);
4607         if (force < space_info->force_alloc)
4608                 force = space_info->force_alloc;
4609         if (space_info->full) {
4610                 if (should_alloc_chunk(fs_info, space_info, force))
4611                         ret = -ENOSPC;
4612                 else
4613                         ret = 0;
4614                 spin_unlock(&space_info->lock);
4615                 return ret;
4616         }
4617
4618         if (!should_alloc_chunk(fs_info, space_info, force)) {
4619                 spin_unlock(&space_info->lock);
4620                 return 0;
4621         } else if (space_info->chunk_alloc) {
4622                 wait_for_alloc = 1;
4623         } else {
4624                 space_info->chunk_alloc = 1;
4625         }
4626
4627         spin_unlock(&space_info->lock);
4628
4629         mutex_lock(&fs_info->chunk_mutex);
4630
4631         /*
4632          * The chunk_mutex is held throughout the entirety of a chunk
4633          * allocation, so once we've acquired the chunk_mutex we know that the
4634          * other guy is done and we need to recheck and see if we should
4635          * allocate.
4636          */
4637         if (wait_for_alloc) {
4638                 mutex_unlock(&fs_info->chunk_mutex);
4639                 wait_for_alloc = 0;
4640                 cond_resched();
4641                 goto again;
4642         }
4643
4644         trans->allocating_chunk = true;
4645
4646         /*
4647          * If we have mixed data/metadata chunks we want to make sure we keep
4648          * allocating mixed chunks instead of individual chunks.
4649          */
4650         if (btrfs_mixed_space_info(space_info))
4651                 flags |= (BTRFS_BLOCK_GROUP_DATA | BTRFS_BLOCK_GROUP_METADATA);
4652
4653         /*
4654          * if we're doing a data chunk, go ahead and make sure that
4655          * we keep a reasonable number of metadata chunks allocated in the
4656          * FS as well.
4657          */
4658         if (flags & BTRFS_BLOCK_GROUP_DATA && fs_info->metadata_ratio) {
4659                 fs_info->data_chunk_allocations++;
4660                 if (!(fs_info->data_chunk_allocations %
4661                       fs_info->metadata_ratio))
4662                         force_metadata_allocation(fs_info);
4663         }
4664
4665         /*
4666          * Check if we have enough space in SYSTEM chunk because we may need
4667          * to update devices.
4668          */
4669         check_system_chunk(trans, fs_info, flags);
4670
4671         ret = btrfs_alloc_chunk(trans, fs_info, flags);
4672         trans->allocating_chunk = false;
4673
4674         spin_lock(&space_info->lock);
4675         if (ret < 0 && ret != -ENOSPC)
4676                 goto out;
4677         if (ret)
4678                 space_info->full = 1;
4679         else
4680                 ret = 1;
4681
4682         space_info->force_alloc = CHUNK_ALLOC_NO_FORCE;
4683 out:
4684         space_info->chunk_alloc = 0;
4685         spin_unlock(&space_info->lock);
4686         mutex_unlock(&fs_info->chunk_mutex);
4687         /*
4688          * When we allocate a new chunk we reserve space in the chunk block
4689          * reserve to make sure we can COW nodes/leafs in the chunk tree or
4690          * add new nodes/leafs to it if we end up needing to do it when
4691          * inserting the chunk item and updating device items as part of the
4692          * second phase of chunk allocation, performed by
4693          * btrfs_finish_chunk_alloc(). So make sure we don't accumulate a
4694          * large number of new block groups to create in our transaction
4695          * handle's new_bgs list to avoid exhausting the chunk block reserve
4696          * in extreme cases - like having a single transaction create many new
4697          * block groups when starting to write out the free space caches of all
4698          * the block groups that were made dirty during the lifetime of the
4699          * transaction.
4700          */
4701         if (trans->can_flush_pending_bgs &&
4702             trans->chunk_bytes_reserved >= (u64)SZ_2M) {
4703                 btrfs_create_pending_block_groups(trans);
4704                 btrfs_trans_release_chunk_metadata(trans);
4705         }
4706         return ret;
4707 }
4708
4709 static int can_overcommit(struct btrfs_fs_info *fs_info,
4710                           struct btrfs_space_info *space_info, u64 bytes,
4711                           enum btrfs_reserve_flush_enum flush,
4712                           bool system_chunk)
4713 {
4714         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
4715         u64 profile;
4716         u64 space_size;
4717         u64 avail;
4718         u64 used;
4719
4720         /* Don't overcommit when in mixed mode. */
4721         if (space_info->flags & BTRFS_BLOCK_GROUP_DATA)
4722                 return 0;
4723
4724         if (system_chunk)
4725                 profile = btrfs_system_alloc_profile(fs_info);
4726         else
4727                 profile = btrfs_metadata_alloc_profile(fs_info);
4728
4729         used = btrfs_space_info_used(space_info, false);
4730
4731         /*
4732          * We only want to allow over committing if we have lots of actual space
4733          * free, but if we don't have enough space to handle the global reserve
4734          * space then we could end up having a real enospc problem when trying
4735          * to allocate a chunk or some other such important allocation.
4736          */
4737         spin_lock(&global_rsv->lock);
4738         space_size = calc_global_rsv_need_space(global_rsv);
4739         spin_unlock(&global_rsv->lock);
4740         if (used + space_size >= space_info->total_bytes)
4741                 return 0;
4742
4743         used += space_info->bytes_may_use;
4744
4745         avail = atomic64_read(&fs_info->free_chunk_space);
4746
4747         /*
4748          * If we have dup, raid1 or raid10 then only half of the free
4749          * space is actually useable.  For raid56, the space info used
4750          * doesn't include the parity drive, so we don't have to
4751          * change the math
4752          */
4753         if (profile & (BTRFS_BLOCK_GROUP_DUP |
4754                        BTRFS_BLOCK_GROUP_RAID1 |
4755                        BTRFS_BLOCK_GROUP_RAID10))
4756                 avail >>= 1;
4757
4758         /*
4759          * If we aren't flushing all things, let us overcommit up to
4760          * 1/2th of the space. If we can flush, don't let us overcommit
4761          * too much, let it overcommit up to 1/8 of the space.
4762          */
4763         if (flush == BTRFS_RESERVE_FLUSH_ALL)
4764                 avail >>= 3;
4765         else
4766                 avail >>= 1;
4767
4768         if (used + bytes < space_info->total_bytes + avail)
4769                 return 1;
4770         return 0;
4771 }
4772
4773 static void btrfs_writeback_inodes_sb_nr(struct btrfs_fs_info *fs_info,
4774                                          unsigned long nr_pages, int nr_items)
4775 {
4776         struct super_block *sb = fs_info->sb;
4777
4778         if (down_read_trylock(&sb->s_umount)) {
4779                 writeback_inodes_sb_nr(sb, nr_pages, WB_REASON_FS_FREE_SPACE);
4780                 up_read(&sb->s_umount);
4781         } else {
4782                 /*
4783                  * We needn't worry the filesystem going from r/w to r/o though
4784                  * we don't acquire ->s_umount mutex, because the filesystem
4785                  * should guarantee the delalloc inodes list be empty after
4786                  * the filesystem is readonly(all dirty pages are written to
4787                  * the disk).
4788                  */
4789                 btrfs_start_delalloc_roots(fs_info, 0, nr_items);
4790                 if (!current->journal_info)
4791                         btrfs_wait_ordered_roots(fs_info, nr_items, 0, (u64)-1);
4792         }
4793 }
4794
4795 static inline u64 calc_reclaim_items_nr(struct btrfs_fs_info *fs_info,
4796                                         u64 to_reclaim)
4797 {
4798         u64 bytes;
4799         u64 nr;
4800
4801         bytes = btrfs_calc_trans_metadata_size(fs_info, 1);
4802         nr = div64_u64(to_reclaim, bytes);
4803         if (!nr)
4804                 nr = 1;
4805         return nr;
4806 }
4807
4808 #define EXTENT_SIZE_PER_ITEM    SZ_256K
4809
4810 /*
4811  * shrink metadata reservation for delalloc
4812  */
4813 static void shrink_delalloc(struct btrfs_fs_info *fs_info, u64 to_reclaim,
4814                             u64 orig, bool wait_ordered)
4815 {
4816         struct btrfs_space_info *space_info;
4817         struct btrfs_trans_handle *trans;
4818         u64 delalloc_bytes;
4819         u64 max_reclaim;
4820         u64 items;
4821         long time_left;
4822         unsigned long nr_pages;
4823         int loops;
4824
4825         /* Calc the number of the pages we need flush for space reservation */
4826         items = calc_reclaim_items_nr(fs_info, to_reclaim);
4827         to_reclaim = items * EXTENT_SIZE_PER_ITEM;
4828
4829         trans = (struct btrfs_trans_handle *)current->journal_info;
4830         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
4831
4832         delalloc_bytes = percpu_counter_sum_positive(
4833                                                 &fs_info->delalloc_bytes);
4834         if (delalloc_bytes == 0) {
4835                 if (trans)
4836                         return;
4837                 if (wait_ordered)
4838                         btrfs_wait_ordered_roots(fs_info, items, 0, (u64)-1);
4839                 return;
4840         }
4841
4842         loops = 0;
4843         while (delalloc_bytes && loops < 3) {
4844                 max_reclaim = min(delalloc_bytes, to_reclaim);
4845                 nr_pages = max_reclaim >> PAGE_SHIFT;
4846                 btrfs_writeback_inodes_sb_nr(fs_info, nr_pages, items);
4847                 /*
4848                  * We need to wait for the async pages to actually start before
4849                  * we do anything.
4850                  */
4851                 max_reclaim = atomic_read(&fs_info->async_delalloc_pages);
4852                 if (!max_reclaim)
4853                         goto skip_async;
4854
4855                 if (max_reclaim <= nr_pages)
4856                         max_reclaim = 0;
4857                 else
4858                         max_reclaim -= nr_pages;
4859
4860                 wait_event(fs_info->async_submit_wait,
4861                            atomic_read(&fs_info->async_delalloc_pages) <=
4862                            (int)max_reclaim);
4863 skip_async:
4864                 spin_lock(&space_info->lock);
4865                 if (list_empty(&space_info->tickets) &&
4866                     list_empty(&space_info->priority_tickets)) {
4867                         spin_unlock(&space_info->lock);
4868                         break;
4869                 }
4870                 spin_unlock(&space_info->lock);
4871
4872                 loops++;
4873                 if (wait_ordered && !trans) {
4874                         btrfs_wait_ordered_roots(fs_info, items, 0, (u64)-1);
4875                 } else {
4876                         time_left = schedule_timeout_killable(1);
4877                         if (time_left)
4878                                 break;
4879                 }
4880                 delalloc_bytes = percpu_counter_sum_positive(
4881                                                 &fs_info->delalloc_bytes);
4882         }
4883 }
4884
4885 struct reserve_ticket {
4886         u64 bytes;
4887         int error;
4888         struct list_head list;
4889         wait_queue_head_t wait;
4890 };
4891
4892 /**
4893  * maybe_commit_transaction - possibly commit the transaction if its ok to
4894  * @root - the root we're allocating for
4895  * @bytes - the number of bytes we want to reserve
4896  * @force - force the commit
4897  *
4898  * This will check to make sure that committing the transaction will actually
4899  * get us somewhere and then commit the transaction if it does.  Otherwise it
4900  * will return -ENOSPC.
4901  */
4902 static int may_commit_transaction(struct btrfs_fs_info *fs_info,
4903                                   struct btrfs_space_info *space_info)
4904 {
4905         struct reserve_ticket *ticket = NULL;
4906         struct btrfs_block_rsv *delayed_rsv = &fs_info->delayed_block_rsv;
4907         struct btrfs_trans_handle *trans;
4908         u64 bytes;
4909
4910         trans = (struct btrfs_trans_handle *)current->journal_info;
4911         if (trans)
4912                 return -EAGAIN;
4913
4914         spin_lock(&space_info->lock);
4915         if (!list_empty(&space_info->priority_tickets))
4916                 ticket = list_first_entry(&space_info->priority_tickets,
4917                                           struct reserve_ticket, list);
4918         else if (!list_empty(&space_info->tickets))
4919                 ticket = list_first_entry(&space_info->tickets,
4920                                           struct reserve_ticket, list);
4921         bytes = (ticket) ? ticket->bytes : 0;
4922         spin_unlock(&space_info->lock);
4923
4924         if (!bytes)
4925                 return 0;
4926
4927         /* See if there is enough pinned space to make this reservation */
4928         if (percpu_counter_compare(&space_info->total_bytes_pinned,
4929                                    bytes) >= 0)
4930                 goto commit;
4931
4932         /*
4933          * See if there is some space in the delayed insertion reservation for
4934          * this reservation.
4935          */
4936         if (space_info != delayed_rsv->space_info)
4937                 return -ENOSPC;
4938
4939         spin_lock(&delayed_rsv->lock);
4940         if (delayed_rsv->size > bytes)
4941                 bytes = 0;
4942         else
4943                 bytes -= delayed_rsv->size;
4944         spin_unlock(&delayed_rsv->lock);
4945
4946         if (percpu_counter_compare(&space_info->total_bytes_pinned,
4947                                    bytes) < 0) {
4948                 return -ENOSPC;
4949         }
4950
4951 commit:
4952         trans = btrfs_join_transaction(fs_info->extent_root);
4953         if (IS_ERR(trans))
4954                 return -ENOSPC;
4955
4956         return btrfs_commit_transaction(trans);
4957 }
4958
4959 /*
4960  * Try to flush some data based on policy set by @state. This is only advisory
4961  * and may fail for various reasons. The caller is supposed to examine the
4962  * state of @space_info to detect the outcome.
4963  */
4964 static void flush_space(struct btrfs_fs_info *fs_info,
4965                        struct btrfs_space_info *space_info, u64 num_bytes,
4966                        int state)
4967 {
4968         struct btrfs_root *root = fs_info->extent_root;
4969         struct btrfs_trans_handle *trans;
4970         int nr;
4971         int ret = 0;
4972
4973         switch (state) {
4974         case FLUSH_DELAYED_ITEMS_NR:
4975         case FLUSH_DELAYED_ITEMS:
4976                 if (state == FLUSH_DELAYED_ITEMS_NR)
4977                         nr = calc_reclaim_items_nr(fs_info, num_bytes) * 2;
4978                 else
4979                         nr = -1;
4980
4981                 trans = btrfs_join_transaction(root);
4982                 if (IS_ERR(trans)) {
4983                         ret = PTR_ERR(trans);
4984                         break;
4985                 }
4986                 ret = btrfs_run_delayed_items_nr(trans, nr);
4987                 btrfs_end_transaction(trans);
4988                 break;
4989         case FLUSH_DELALLOC:
4990         case FLUSH_DELALLOC_WAIT:
4991                 shrink_delalloc(fs_info, num_bytes * 2, num_bytes,
4992                                 state == FLUSH_DELALLOC_WAIT);
4993                 break;
4994         case ALLOC_CHUNK:
4995                 trans = btrfs_join_transaction(root);
4996                 if (IS_ERR(trans)) {
4997                         ret = PTR_ERR(trans);
4998                         break;
4999                 }
5000                 ret = do_chunk_alloc(trans, fs_info,
5001                                      btrfs_metadata_alloc_profile(fs_info),
5002                                      CHUNK_ALLOC_NO_FORCE);
5003                 btrfs_end_transaction(trans);
5004                 if (ret > 0 || ret == -ENOSPC)
5005                         ret = 0;
5006                 break;
5007         case COMMIT_TRANS:
5008                 ret = may_commit_transaction(fs_info, space_info);
5009                 break;
5010         default:
5011                 ret = -ENOSPC;
5012                 break;
5013         }
5014
5015         trace_btrfs_flush_space(fs_info, space_info->flags, num_bytes, state,
5016                                 ret);
5017         return;
5018 }
5019
5020 static inline u64
5021 btrfs_calc_reclaim_metadata_size(struct btrfs_fs_info *fs_info,
5022                                  struct btrfs_space_info *space_info,
5023                                  bool system_chunk)
5024 {
5025         struct reserve_ticket *ticket;
5026         u64 used;
5027         u64 expected;
5028         u64 to_reclaim = 0;
5029
5030         list_for_each_entry(ticket, &space_info->tickets, list)
5031                 to_reclaim += ticket->bytes;
5032         list_for_each_entry(ticket, &space_info->priority_tickets, list)
5033                 to_reclaim += ticket->bytes;
5034         if (to_reclaim)
5035                 return to_reclaim;
5036
5037         to_reclaim = min_t(u64, num_online_cpus() * SZ_1M, SZ_16M);
5038         if (can_overcommit(fs_info, space_info, to_reclaim,
5039                            BTRFS_RESERVE_FLUSH_ALL, system_chunk))
5040                 return 0;
5041
5042         used = btrfs_space_info_used(space_info, true);
5043
5044         if (can_overcommit(fs_info, space_info, SZ_1M,
5045                            BTRFS_RESERVE_FLUSH_ALL, system_chunk))
5046                 expected = div_factor_fine(space_info->total_bytes, 95);
5047         else
5048                 expected = div_factor_fine(space_info->total_bytes, 90);
5049
5050         if (used > expected)
5051                 to_reclaim = used - expected;
5052         else
5053                 to_reclaim = 0;
5054         to_reclaim = min(to_reclaim, space_info->bytes_may_use +
5055                                      space_info->bytes_reserved);
5056         return to_reclaim;
5057 }
5058
5059 static inline int need_do_async_reclaim(struct btrfs_fs_info *fs_info,
5060                                         struct btrfs_space_info *space_info,
5061                                         u64 used, bool system_chunk)
5062 {
5063         u64 thresh = div_factor_fine(space_info->total_bytes, 98);
5064
5065         /* If we're just plain full then async reclaim just slows us down. */
5066         if ((space_info->bytes_used + space_info->bytes_reserved) >= thresh)
5067                 return 0;
5068
5069         if (!btrfs_calc_reclaim_metadata_size(fs_info, space_info,
5070                                               system_chunk))
5071                 return 0;
5072
5073         return (used >= thresh && !btrfs_fs_closing(fs_info) &&
5074                 !test_bit(BTRFS_FS_STATE_REMOUNTING, &fs_info->fs_state));
5075 }
5076
5077 static void wake_all_tickets(struct list_head *head)
5078 {
5079         struct reserve_ticket *ticket;
5080
5081         while (!list_empty(head)) {
5082                 ticket = list_first_entry(head, struct reserve_ticket, list);
5083                 list_del_init(&ticket->list);
5084                 ticket->error = -ENOSPC;
5085                 wake_up(&ticket->wait);
5086         }
5087 }
5088
5089 /*
5090  * This is for normal flushers, we can wait all goddamned day if we want to.  We
5091  * will loop and continuously try to flush as long as we are making progress.
5092  * We count progress as clearing off tickets each time we have to loop.
5093  */
5094 static void btrfs_async_reclaim_metadata_space(struct work_struct *work)
5095 {
5096         struct btrfs_fs_info *fs_info;
5097         struct btrfs_space_info *space_info;
5098         u64 to_reclaim;
5099         int flush_state;
5100         int commit_cycles = 0;
5101         u64 last_tickets_id;
5102
5103         fs_info = container_of(work, struct btrfs_fs_info, async_reclaim_work);
5104         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
5105
5106         spin_lock(&space_info->lock);
5107         to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info, space_info,
5108                                                       false);
5109         if (!to_reclaim) {
5110                 space_info->flush = 0;
5111                 spin_unlock(&space_info->lock);
5112                 return;
5113         }
5114         last_tickets_id = space_info->tickets_id;
5115         spin_unlock(&space_info->lock);
5116
5117         flush_state = FLUSH_DELAYED_ITEMS_NR;
5118         do {
5119                 flush_space(fs_info, space_info, to_reclaim, flush_state);
5120                 spin_lock(&space_info->lock);
5121                 if (list_empty(&space_info->tickets)) {
5122                         space_info->flush = 0;
5123                         spin_unlock(&space_info->lock);
5124                         return;
5125                 }
5126                 to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info,
5127                                                               space_info,
5128                                                               false);
5129                 if (last_tickets_id == space_info->tickets_id) {
5130                         flush_state++;
5131                 } else {
5132                         last_tickets_id = space_info->tickets_id;
5133                         flush_state = FLUSH_DELAYED_ITEMS_NR;
5134                         if (commit_cycles)
5135                                 commit_cycles--;
5136                 }
5137
5138                 if (flush_state > COMMIT_TRANS) {
5139                         commit_cycles++;
5140                         if (commit_cycles > 2) {
5141                                 wake_all_tickets(&space_info->tickets);
5142                                 space_info->flush = 0;
5143                         } else {
5144                                 flush_state = FLUSH_DELAYED_ITEMS_NR;
5145                         }
5146                 }
5147                 spin_unlock(&space_info->lock);
5148         } while (flush_state <= COMMIT_TRANS);
5149 }
5150
5151 void btrfs_init_async_reclaim_work(struct work_struct *work)
5152 {
5153         INIT_WORK(work, btrfs_async_reclaim_metadata_space);
5154 }
5155
5156 static void priority_reclaim_metadata_space(struct btrfs_fs_info *fs_info,
5157                                             struct btrfs_space_info *space_info,
5158                                             struct reserve_ticket *ticket)
5159 {
5160         u64 to_reclaim;
5161         int flush_state = FLUSH_DELAYED_ITEMS_NR;
5162
5163         spin_lock(&space_info->lock);
5164         to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info, space_info,
5165                                                       false);
5166         if (!to_reclaim) {
5167                 spin_unlock(&space_info->lock);
5168                 return;
5169         }
5170         spin_unlock(&space_info->lock);
5171
5172         do {
5173                 flush_space(fs_info, space_info, to_reclaim, flush_state);
5174                 flush_state++;
5175                 spin_lock(&space_info->lock);
5176                 if (ticket->bytes == 0) {
5177                         spin_unlock(&space_info->lock);
5178                         return;
5179                 }
5180                 spin_unlock(&space_info->lock);
5181
5182                 /*
5183                  * Priority flushers can't wait on delalloc without
5184                  * deadlocking.
5185                  */
5186                 if (flush_state == FLUSH_DELALLOC ||
5187                     flush_state == FLUSH_DELALLOC_WAIT)
5188                         flush_state = ALLOC_CHUNK;
5189         } while (flush_state < COMMIT_TRANS);
5190 }
5191
5192 static int wait_reserve_ticket(struct btrfs_fs_info *fs_info,
5193                                struct btrfs_space_info *space_info,
5194                                struct reserve_ticket *ticket, u64 orig_bytes)
5195
5196 {
5197         DEFINE_WAIT(wait);
5198         int ret = 0;
5199
5200         spin_lock(&space_info->lock);
5201         while (ticket->bytes > 0 && ticket->error == 0) {
5202                 ret = prepare_to_wait_event(&ticket->wait, &wait, TASK_KILLABLE);
5203                 if (ret) {
5204                         ret = -EINTR;
5205                         break;
5206                 }
5207                 spin_unlock(&space_info->lock);
5208
5209                 schedule();
5210
5211                 finish_wait(&ticket->wait, &wait);
5212                 spin_lock(&space_info->lock);
5213         }
5214         if (!ret)
5215                 ret = ticket->error;
5216         if (!list_empty(&ticket->list))
5217                 list_del_init(&ticket->list);
5218         if (ticket->bytes && ticket->bytes < orig_bytes) {
5219                 u64 num_bytes = orig_bytes - ticket->bytes;
5220                 space_info->bytes_may_use -= num_bytes;
5221                 trace_btrfs_space_reservation(fs_info, "space_info",
5222                                               space_info->flags, num_bytes, 0);
5223         }
5224         spin_unlock(&space_info->lock);
5225
5226         return ret;
5227 }
5228
5229 /**
5230  * reserve_metadata_bytes - try to reserve bytes from the block_rsv's space
5231  * @root - the root we're allocating for
5232  * @space_info - the space info we want to allocate from
5233  * @orig_bytes - the number of bytes we want
5234  * @flush - whether or not we can flush to make our reservation
5235  *
5236  * This will reserve orig_bytes number of bytes from the space info associated
5237  * with the block_rsv.  If there is not enough space it will make an attempt to
5238  * flush out space to make room.  It will do this by flushing delalloc if
5239  * possible or committing the transaction.  If flush is 0 then no attempts to
5240  * regain reservations will be made and this will fail if there is not enough
5241  * space already.
5242  */
5243 static int __reserve_metadata_bytes(struct btrfs_fs_info *fs_info,
5244                                     struct btrfs_space_info *space_info,
5245                                     u64 orig_bytes,
5246                                     enum btrfs_reserve_flush_enum flush,
5247                                     bool system_chunk)
5248 {
5249         struct reserve_ticket ticket;
5250         u64 used;
5251         int ret = 0;
5252
5253         ASSERT(orig_bytes);
5254         ASSERT(!current->journal_info || flush != BTRFS_RESERVE_FLUSH_ALL);
5255
5256         spin_lock(&space_info->lock);
5257         ret = -ENOSPC;
5258         used = btrfs_space_info_used(space_info, true);
5259
5260         /*
5261          * If we have enough space then hooray, make our reservation and carry
5262          * on.  If not see if we can overcommit, and if we can, hooray carry on.
5263          * If not things get more complicated.
5264          */
5265         if (used + orig_bytes <= space_info->total_bytes) {
5266                 space_info->bytes_may_use += orig_bytes;
5267                 trace_btrfs_space_reservation(fs_info, "space_info",
5268                                               space_info->flags, orig_bytes, 1);
5269                 ret = 0;
5270         } else if (can_overcommit(fs_info, space_info, orig_bytes, flush,
5271                                   system_chunk)) {
5272                 space_info->bytes_may_use += orig_bytes;
5273                 trace_btrfs_space_reservation(fs_info, "space_info",
5274                                               space_info->flags, orig_bytes, 1);
5275                 ret = 0;
5276         }
5277
5278         /*
5279          * If we couldn't make a reservation then setup our reservation ticket
5280          * and kick the async worker if it's not already running.
5281          *
5282          * If we are a priority flusher then we just need to add our ticket to
5283          * the list and we will do our own flushing further down.
5284          */
5285         if (ret && flush != BTRFS_RESERVE_NO_FLUSH) {
5286                 ticket.bytes = orig_bytes;
5287                 ticket.error = 0;
5288                 init_waitqueue_head(&ticket.wait);
5289                 if (flush == BTRFS_RESERVE_FLUSH_ALL) {
5290                         list_add_tail(&ticket.list, &space_info->tickets);
5291                         if (!space_info->flush) {
5292                                 space_info->flush = 1;
5293                                 trace_btrfs_trigger_flush(fs_info,
5294                                                           space_info->flags,
5295                                                           orig_bytes, flush,
5296                                                           "enospc");
5297                                 queue_work(system_unbound_wq,
5298                                            &fs_info->async_reclaim_work);
5299                         }
5300                 } else {
5301                         list_add_tail(&ticket.list,
5302                                       &space_info->priority_tickets);
5303                 }
5304         } else if (!ret && space_info->flags & BTRFS_BLOCK_GROUP_METADATA) {
5305                 used += orig_bytes;
5306                 /*
5307                  * We will do the space reservation dance during log replay,
5308                  * which means we won't have fs_info->fs_root set, so don't do
5309                  * the async reclaim as we will panic.
5310                  */
5311                 if (!test_bit(BTRFS_FS_LOG_RECOVERING, &fs_info->flags) &&
5312                     need_do_async_reclaim(fs_info, space_info,
5313                                           used, system_chunk) &&
5314                     !work_busy(&fs_info->async_reclaim_work)) {
5315                         trace_btrfs_trigger_flush(fs_info, space_info->flags,
5316                                                   orig_bytes, flush, "preempt");
5317                         queue_work(system_unbound_wq,
5318                                    &fs_info->async_reclaim_work);
5319                 }
5320         }
5321         spin_unlock(&space_info->lock);
5322         if (!ret || flush == BTRFS_RESERVE_NO_FLUSH)
5323                 return ret;
5324
5325         if (flush == BTRFS_RESERVE_FLUSH_ALL)
5326                 return wait_reserve_ticket(fs_info, space_info, &ticket,
5327                                            orig_bytes);
5328
5329         ret = 0;
5330         priority_reclaim_metadata_space(fs_info, space_info, &ticket);
5331         spin_lock(&space_info->lock);
5332         if (ticket.bytes) {
5333                 if (ticket.bytes < orig_bytes) {
5334                         u64 num_bytes = orig_bytes - ticket.bytes;
5335                         space_info->bytes_may_use -= num_bytes;
5336                         trace_btrfs_space_reservation(fs_info, "space_info",
5337                                                       space_info->flags,
5338                                                       num_bytes, 0);
5339
5340                 }
5341                 list_del_init(&ticket.list);
5342                 ret = -ENOSPC;
5343         }
5344         spin_unlock(&space_info->lock);
5345         ASSERT(list_empty(&ticket.list));
5346         return ret;
5347 }
5348
5349 /**
5350  * reserve_metadata_bytes - try to reserve bytes from the block_rsv's space
5351  * @root - the root we're allocating for
5352  * @block_rsv - the block_rsv we're allocating for
5353  * @orig_bytes - the number of bytes we want
5354  * @flush - whether or not we can flush to make our reservation
5355  *
5356  * This will reserve orgi_bytes number of bytes from the space info associated
5357  * with the block_rsv.  If there is not enough space it will make an attempt to
5358  * flush out space to make room.  It will do this by flushing delalloc if
5359  * possible or committing the transaction.  If flush is 0 then no attempts to
5360  * regain reservations will be made and this will fail if there is not enough
5361  * space already.
5362  */
5363 static int reserve_metadata_bytes(struct btrfs_root *root,
5364                                   struct btrfs_block_rsv *block_rsv,
5365                                   u64 orig_bytes,
5366                                   enum btrfs_reserve_flush_enum flush)
5367 {
5368         struct btrfs_fs_info *fs_info = root->fs_info;
5369         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
5370         int ret;
5371         bool system_chunk = (root == fs_info->chunk_root);
5372
5373         ret = __reserve_metadata_bytes(fs_info, block_rsv->space_info,
5374                                        orig_bytes, flush, system_chunk);
5375         if (ret == -ENOSPC &&
5376             unlikely(root->orphan_cleanup_state == ORPHAN_CLEANUP_STARTED)) {
5377                 if (block_rsv != global_rsv &&
5378                     !block_rsv_use_bytes(global_rsv, orig_bytes))
5379                         ret = 0;
5380         }
5381         if (ret == -ENOSPC) {
5382                 trace_btrfs_space_reservation(fs_info, "space_info:enospc",
5383                                               block_rsv->space_info->flags,
5384                                               orig_bytes, 1);
5385
5386                 if (btrfs_test_opt(fs_info, ENOSPC_DEBUG))
5387                         dump_space_info(fs_info, block_rsv->space_info,
5388                                         orig_bytes, 0);
5389         }
5390         return ret;
5391 }
5392
5393 static struct btrfs_block_rsv *get_block_rsv(
5394                                         const struct btrfs_trans_handle *trans,
5395                                         const struct btrfs_root *root)
5396 {
5397         struct btrfs_fs_info *fs_info = root->fs_info;
5398         struct btrfs_block_rsv *block_rsv = NULL;
5399
5400         if (test_bit(BTRFS_ROOT_REF_COWS, &root->state) ||
5401             (root == fs_info->csum_root && trans->adding_csums) ||
5402             (root == fs_info->uuid_root))
5403                 block_rsv = trans->block_rsv;
5404
5405         if (!block_rsv)
5406                 block_rsv = root->block_rsv;
5407
5408         if (!block_rsv)
5409                 block_rsv = &fs_info->empty_block_rsv;
5410
5411         return block_rsv;
5412 }
5413
5414 static int block_rsv_use_bytes(struct btrfs_block_rsv *block_rsv,
5415                                u64 num_bytes)
5416 {
5417         int ret = -ENOSPC;
5418         spin_lock(&block_rsv->lock);
5419         if (block_rsv->reserved >= num_bytes) {
5420                 block_rsv->reserved -= num_bytes;
5421                 if (block_rsv->reserved < block_rsv->size)
5422                         block_rsv->full = 0;
5423                 ret = 0;
5424         }
5425         spin_unlock(&block_rsv->lock);
5426         return ret;
5427 }
5428
5429 static void block_rsv_add_bytes(struct btrfs_block_rsv *block_rsv,
5430                                 u64 num_bytes, int update_size)
5431 {
5432         spin_lock(&block_rsv->lock);
5433         block_rsv->reserved += num_bytes;
5434         if (update_size)
5435                 block_rsv->size += num_bytes;
5436         else if (block_rsv->reserved >= block_rsv->size)
5437                 block_rsv->full = 1;
5438         spin_unlock(&block_rsv->lock);
5439 }
5440
5441 int btrfs_cond_migrate_bytes(struct btrfs_fs_info *fs_info,
5442                              struct btrfs_block_rsv *dest, u64 num_bytes,
5443                              int min_factor)
5444 {
5445         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
5446         u64 min_bytes;
5447
5448         if (global_rsv->space_info != dest->space_info)
5449                 return -ENOSPC;
5450
5451         spin_lock(&global_rsv->lock);
5452         min_bytes = div_factor(global_rsv->size, min_factor);
5453         if (global_rsv->reserved < min_bytes + num_bytes) {
5454                 spin_unlock(&global_rsv->lock);
5455                 return -ENOSPC;
5456         }
5457         global_rsv->reserved -= num_bytes;
5458         if (global_rsv->reserved < global_rsv->size)
5459                 global_rsv->full = 0;
5460         spin_unlock(&global_rsv->lock);
5461
5462         block_rsv_add_bytes(dest, num_bytes, 1);
5463         return 0;
5464 }
5465
5466 /*
5467  * This is for space we already have accounted in space_info->bytes_may_use, so
5468  * basically when we're returning space from block_rsv's.
5469  */
5470 static void space_info_add_old_bytes(struct btrfs_fs_info *fs_info,
5471                                      struct btrfs_space_info *space_info,
5472                                      u64 num_bytes)
5473 {
5474         struct reserve_ticket *ticket;
5475         struct list_head *head;
5476         u64 used;
5477         enum btrfs_reserve_flush_enum flush = BTRFS_RESERVE_NO_FLUSH;
5478         bool check_overcommit = false;
5479
5480         spin_lock(&space_info->lock);
5481         head = &space_info->priority_tickets;
5482
5483         /*
5484          * If we are over our limit then we need to check and see if we can
5485          * overcommit, and if we can't then we just need to free up our space
5486          * and not satisfy any requests.
5487          */
5488         used = btrfs_space_info_used(space_info, true);
5489         if (used - num_bytes >= space_info->total_bytes)
5490                 check_overcommit = true;
5491 again:
5492         while (!list_empty(head) && num_bytes) {
5493                 ticket = list_first_entry(head, struct reserve_ticket,
5494                                           list);
5495                 /*
5496                  * We use 0 bytes because this space is already reserved, so
5497                  * adding the ticket space would be a double count.
5498                  */
5499                 if (check_overcommit &&
5500                     !can_overcommit(fs_info, space_info, 0, flush, false))
5501                         break;
5502                 if (num_bytes >= ticket->bytes) {
5503                         list_del_init(&ticket->list);
5504                         num_bytes -= ticket->bytes;
5505                         ticket->bytes = 0;
5506                         space_info->tickets_id++;
5507                         wake_up(&ticket->wait);
5508                 } else {
5509                         ticket->bytes -= num_bytes;
5510                         num_bytes = 0;
5511                 }
5512         }
5513
5514         if (num_bytes && head == &space_info->priority_tickets) {
5515                 head = &space_info->tickets;
5516                 flush = BTRFS_RESERVE_FLUSH_ALL;
5517                 goto again;
5518         }
5519         space_info->bytes_may_use -= num_bytes;
5520         trace_btrfs_space_reservation(fs_info, "space_info",
5521                                       space_info->flags, num_bytes, 0);
5522         spin_unlock(&space_info->lock);
5523 }
5524
5525 /*
5526  * This is for newly allocated space that isn't accounted in
5527  * space_info->bytes_may_use yet.  So if we allocate a chunk or unpin an extent
5528  * we use this helper.
5529  */
5530 static void space_info_add_new_bytes(struct btrfs_fs_info *fs_info,
5531                                      struct btrfs_space_info *space_info,
5532                                      u64 num_bytes)
5533 {
5534         struct reserve_ticket *ticket;
5535         struct list_head *head = &space_info->priority_tickets;
5536
5537 again:
5538         while (!list_empty(head) && num_bytes) {
5539                 ticket = list_first_entry(head, struct reserve_ticket,
5540                                           list);
5541                 if (num_bytes >= ticket->bytes) {
5542                         trace_btrfs_space_reservation(fs_info, "space_info",
5543                                                       space_info->flags,
5544                                                       ticket->bytes, 1);
5545                         list_del_init(&ticket->list);
5546                         num_bytes -= ticket->bytes;
5547                         space_info->bytes_may_use += ticket->bytes;
5548                         ticket->bytes = 0;
5549                         space_info->tickets_id++;
5550                         wake_up(&ticket->wait);
5551                 } else {
5552                         trace_btrfs_space_reservation(fs_info, "space_info",
5553                                                       space_info->flags,
5554                                                       num_bytes, 1);
5555                         space_info->bytes_may_use += num_bytes;
5556                         ticket->bytes -= num_bytes;
5557                         num_bytes = 0;
5558                 }
5559         }
5560
5561         if (num_bytes && head == &space_info->priority_tickets) {
5562                 head = &space_info->tickets;
5563                 goto again;
5564         }
5565 }
5566
5567 static u64 block_rsv_release_bytes(struct btrfs_fs_info *fs_info,
5568                                     struct btrfs_block_rsv *block_rsv,
5569                                     struct btrfs_block_rsv *dest, u64 num_bytes,
5570                                     u64 *qgroup_to_release_ret)
5571 {
5572         struct btrfs_space_info *space_info = block_rsv->space_info;
5573         u64 qgroup_to_release = 0;
5574         u64 ret;
5575
5576         spin_lock(&block_rsv->lock);
5577         if (num_bytes == (u64)-1) {
5578                 num_bytes = block_rsv->size;
5579                 qgroup_to_release = block_rsv->qgroup_rsv_size;
5580         }
5581         block_rsv->size -= num_bytes;
5582         if (block_rsv->reserved >= block_rsv->size) {
5583                 num_bytes = block_rsv->reserved - block_rsv->size;
5584                 block_rsv->reserved = block_rsv->size;
5585                 block_rsv->full = 1;
5586         } else {
5587                 num_bytes = 0;
5588         }
5589         if (block_rsv->qgroup_rsv_reserved >= block_rsv->qgroup_rsv_size) {
5590                 qgroup_to_release = block_rsv->qgroup_rsv_reserved -
5591                                     block_rsv->qgroup_rsv_size;
5592                 block_rsv->qgroup_rsv_reserved = block_rsv->qgroup_rsv_size;
5593         } else {
5594                 qgroup_to_release = 0;
5595         }
5596         spin_unlock(&block_rsv->lock);
5597
5598         ret = num_bytes;
5599         if (num_bytes > 0) {
5600                 if (dest) {
5601                         spin_lock(&dest->lock);
5602                         if (!dest->full) {
5603                                 u64 bytes_to_add;
5604
5605                                 bytes_to_add = dest->size - dest->reserved;
5606                                 bytes_to_add = min(num_bytes, bytes_to_add);
5607                                 dest->reserved += bytes_to_add;
5608                                 if (dest->reserved >= dest->size)
5609                                         dest->full = 1;
5610                                 num_bytes -= bytes_to_add;
5611                         }
5612                         spin_unlock(&dest->lock);
5613                 }
5614                 if (num_bytes)
5615                         space_info_add_old_bytes(fs_info, space_info,
5616                                                  num_bytes);
5617         }
5618         if (qgroup_to_release_ret)
5619                 *qgroup_to_release_ret = qgroup_to_release;
5620         return ret;
5621 }
5622
5623 int btrfs_block_rsv_migrate(struct btrfs_block_rsv *src,
5624                             struct btrfs_block_rsv *dst, u64 num_bytes,
5625                             int update_size)
5626 {
5627         int ret;
5628
5629         ret = block_rsv_use_bytes(src, num_bytes);
5630         if (ret)
5631                 return ret;
5632
5633         block_rsv_add_bytes(dst, num_bytes, update_size);
5634         return 0;
5635 }
5636
5637 void btrfs_init_block_rsv(struct btrfs_block_rsv *rsv, unsigned short type)
5638 {
5639         memset(rsv, 0, sizeof(*rsv));
5640         spin_lock_init(&rsv->lock);
5641         rsv->type = type;
5642 }
5643
5644 void btrfs_init_metadata_block_rsv(struct btrfs_fs_info *fs_info,
5645                                    struct btrfs_block_rsv *rsv,
5646                                    unsigned short type)
5647 {
5648         btrfs_init_block_rsv(rsv, type);
5649         rsv->space_info = __find_space_info(fs_info,
5650                                             BTRFS_BLOCK_GROUP_METADATA);
5651 }
5652
5653 struct btrfs_block_rsv *btrfs_alloc_block_rsv(struct btrfs_fs_info *fs_info,
5654                                               unsigned short type)
5655 {
5656         struct btrfs_block_rsv *block_rsv;
5657
5658         block_rsv = kmalloc(sizeof(*block_rsv), GFP_NOFS);
5659         if (!block_rsv)
5660                 return NULL;
5661
5662         btrfs_init_metadata_block_rsv(fs_info, block_rsv, type);
5663         return block_rsv;
5664 }
5665
5666 void btrfs_free_block_rsv(struct btrfs_fs_info *fs_info,
5667                           struct btrfs_block_rsv *rsv)
5668 {
5669         if (!rsv)
5670                 return;
5671         btrfs_block_rsv_release(fs_info, rsv, (u64)-1);
5672         kfree(rsv);
5673 }
5674
5675 void __btrfs_free_block_rsv(struct btrfs_block_rsv *rsv)
5676 {
5677         kfree(rsv);
5678 }
5679
5680 int btrfs_block_rsv_add(struct btrfs_root *root,
5681                         struct btrfs_block_rsv *block_rsv, u64 num_bytes,
5682                         enum btrfs_reserve_flush_enum flush)
5683 {
5684         int ret;
5685
5686         if (num_bytes == 0)
5687                 return 0;
5688
5689         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
5690         if (!ret) {
5691                 block_rsv_add_bytes(block_rsv, num_bytes, 1);
5692                 return 0;
5693         }
5694
5695         return ret;
5696 }
5697
5698 int btrfs_block_rsv_check(struct btrfs_block_rsv *block_rsv, int min_factor)
5699 {
5700         u64 num_bytes = 0;
5701         int ret = -ENOSPC;
5702
5703         if (!block_rsv)
5704                 return 0;
5705
5706         spin_lock(&block_rsv->lock);
5707         num_bytes = div_factor(block_rsv->size, min_factor);
5708         if (block_rsv->reserved >= num_bytes)
5709                 ret = 0;
5710         spin_unlock(&block_rsv->lock);
5711
5712         return ret;
5713 }
5714
5715 int btrfs_block_rsv_refill(struct btrfs_root *root,
5716                            struct btrfs_block_rsv *block_rsv, u64 min_reserved,
5717                            enum btrfs_reserve_flush_enum flush)
5718 {
5719         u64 num_bytes = 0;
5720         int ret = -ENOSPC;
5721
5722         if (!block_rsv)
5723                 return 0;
5724
5725         spin_lock(&block_rsv->lock);
5726         num_bytes = min_reserved;
5727         if (block_rsv->reserved >= num_bytes)
5728                 ret = 0;
5729         else
5730                 num_bytes -= block_rsv->reserved;
5731         spin_unlock(&block_rsv->lock);
5732
5733         if (!ret)
5734                 return 0;
5735
5736         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
5737         if (!ret) {
5738                 block_rsv_add_bytes(block_rsv, num_bytes, 0);
5739                 return 0;
5740         }
5741
5742         return ret;
5743 }
5744
5745 /**
5746  * btrfs_inode_rsv_refill - refill the inode block rsv.
5747  * @inode - the inode we are refilling.
5748  * @flush - the flusing restriction.
5749  *
5750  * Essentially the same as btrfs_block_rsv_refill, except it uses the
5751  * block_rsv->size as the minimum size.  We'll either refill the missing amount
5752  * or return if we already have enough space.  This will also handle the resreve
5753  * tracepoint for the reserved amount.
5754  */
5755 static int btrfs_inode_rsv_refill(struct btrfs_inode *inode,
5756                                   enum btrfs_reserve_flush_enum flush)
5757 {
5758         struct btrfs_root *root = inode->root;
5759         struct btrfs_block_rsv *block_rsv = &inode->block_rsv;
5760         u64 num_bytes = 0;
5761         u64 qgroup_num_bytes = 0;
5762         int ret = -ENOSPC;
5763
5764         spin_lock(&block_rsv->lock);
5765         if (block_rsv->reserved < block_rsv->size)
5766                 num_bytes = block_rsv->size - block_rsv->reserved;
5767         if (block_rsv->qgroup_rsv_reserved < block_rsv->qgroup_rsv_size)
5768                 qgroup_num_bytes = block_rsv->qgroup_rsv_size -
5769                                    block_rsv->qgroup_rsv_reserved;
5770         spin_unlock(&block_rsv->lock);
5771
5772         if (num_bytes == 0)
5773                 return 0;
5774
5775         ret = btrfs_qgroup_reserve_meta_prealloc(root, qgroup_num_bytes, true);
5776         if (ret)
5777                 return ret;
5778         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
5779         if (!ret) {
5780                 block_rsv_add_bytes(block_rsv, num_bytes, 0);
5781                 trace_btrfs_space_reservation(root->fs_info, "delalloc",
5782                                               btrfs_ino(inode), num_bytes, 1);
5783
5784                 /* Don't forget to increase qgroup_rsv_reserved */
5785                 spin_lock(&block_rsv->lock);
5786                 block_rsv->qgroup_rsv_reserved += qgroup_num_bytes;
5787                 spin_unlock(&block_rsv->lock);
5788         } else
5789                 btrfs_qgroup_free_meta_prealloc(root, qgroup_num_bytes);
5790         return ret;
5791 }
5792
5793 /**
5794  * btrfs_inode_rsv_release - release any excessive reservation.
5795  * @inode - the inode we need to release from.
5796  * @qgroup_free - free or convert qgroup meta.
5797  *   Unlike normal operation, qgroup meta reservation needs to know if we are
5798  *   freeing qgroup reservation or just converting it into per-trans.  Normally
5799  *   @qgroup_free is true for error handling, and false for normal release.
5800  *
5801  * This is the same as btrfs_block_rsv_release, except that it handles the
5802  * tracepoint for the reservation.
5803  */
5804 static void btrfs_inode_rsv_release(struct btrfs_inode *inode, bool qgroup_free)
5805 {
5806         struct btrfs_fs_info *fs_info = inode->root->fs_info;
5807         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
5808         struct btrfs_block_rsv *block_rsv = &inode->block_rsv;
5809         u64 released = 0;
5810         u64 qgroup_to_release = 0;
5811
5812         /*
5813          * Since we statically set the block_rsv->size we just want to say we
5814          * are releasing 0 bytes, and then we'll just get the reservation over
5815          * the size free'd.
5816          */
5817         released = block_rsv_release_bytes(fs_info, block_rsv, global_rsv, 0,
5818                                            &qgroup_to_release);
5819         if (released > 0)
5820                 trace_btrfs_space_reservation(fs_info, "delalloc",
5821                                               btrfs_ino(inode), released, 0);
5822         if (qgroup_free)
5823                 btrfs_qgroup_free_meta_prealloc(inode->root, qgroup_to_release);
5824         else
5825                 btrfs_qgroup_convert_reserved_meta(inode->root,
5826                                                    qgroup_to_release);
5827 }
5828
5829 void btrfs_block_rsv_release(struct btrfs_fs_info *fs_info,
5830                              struct btrfs_block_rsv *block_rsv,
5831                              u64 num_bytes)
5832 {
5833         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
5834
5835         if (global_rsv == block_rsv ||
5836             block_rsv->space_info != global_rsv->space_info)
5837                 global_rsv = NULL;
5838         block_rsv_release_bytes(fs_info, block_rsv, global_rsv, num_bytes, NULL);
5839 }
5840
5841 static void update_global_block_rsv(struct btrfs_fs_info *fs_info)
5842 {
5843         struct btrfs_block_rsv *block_rsv = &fs_info->global_block_rsv;
5844         struct btrfs_space_info *sinfo = block_rsv->space_info;
5845         u64 num_bytes;
5846
5847         /*
5848          * The global block rsv is based on the size of the extent tree, the
5849          * checksum tree and the root tree.  If the fs is empty we want to set
5850          * it to a minimal amount for safety.
5851          */
5852         num_bytes = btrfs_root_used(&fs_info->extent_root->root_item) +
5853                 btrfs_root_used(&fs_info->csum_root->root_item) +
5854                 btrfs_root_used(&fs_info->tree_root->root_item);
5855         num_bytes = max_t(u64, num_bytes, SZ_16M);
5856
5857         spin_lock(&sinfo->lock);
5858         spin_lock(&block_rsv->lock);
5859
5860         block_rsv->size = min_t(u64, num_bytes, SZ_512M);
5861
5862         if (block_rsv->reserved < block_rsv->size) {
5863                 num_bytes = btrfs_space_info_used(sinfo, true);
5864                 if (sinfo->total_bytes > num_bytes) {
5865                         num_bytes = sinfo->total_bytes - num_bytes;
5866                         num_bytes = min(num_bytes,
5867                                         block_rsv->size - block_rsv->reserved);
5868                         block_rsv->reserved += num_bytes;
5869                         sinfo->bytes_may_use += num_bytes;
5870                         trace_btrfs_space_reservation(fs_info, "space_info",
5871                                                       sinfo->flags, num_bytes,
5872                                                       1);
5873                 }
5874         } else if (block_rsv->reserved > block_rsv->size) {
5875                 num_bytes = block_rsv->reserved - block_rsv->size;
5876                 sinfo->bytes_may_use -= num_bytes;
5877                 trace_btrfs_space_reservation(fs_info, "space_info",
5878                                       sinfo->flags, num_bytes, 0);
5879                 block_rsv->reserved = block_rsv->size;
5880         }
5881
5882         if (block_rsv->reserved == block_rsv->size)
5883                 block_rsv->full = 1;
5884         else
5885                 block_rsv->full = 0;
5886
5887         spin_unlock(&block_rsv->lock);
5888         spin_unlock(&sinfo->lock);
5889 }
5890
5891 static void init_global_block_rsv(struct btrfs_fs_info *fs_info)
5892 {
5893         struct btrfs_space_info *space_info;
5894
5895         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
5896         fs_info->chunk_block_rsv.space_info = space_info;
5897
5898         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
5899         fs_info->global_block_rsv.space_info = space_info;
5900         fs_info->trans_block_rsv.space_info = space_info;
5901         fs_info->empty_block_rsv.space_info = space_info;
5902         fs_info->delayed_block_rsv.space_info = space_info;
5903
5904         fs_info->extent_root->block_rsv = &fs_info->global_block_rsv;
5905         fs_info->csum_root->block_rsv = &fs_info->global_block_rsv;
5906         fs_info->dev_root->block_rsv = &fs_info->global_block_rsv;
5907         fs_info->tree_root->block_rsv = &fs_info->global_block_rsv;
5908         if (fs_info->quota_root)
5909                 fs_info->quota_root->block_rsv = &fs_info->global_block_rsv;
5910         fs_info->chunk_root->block_rsv = &fs_info->chunk_block_rsv;
5911
5912         update_global_block_rsv(fs_info);
5913 }
5914
5915 static void release_global_block_rsv(struct btrfs_fs_info *fs_info)
5916 {
5917         block_rsv_release_bytes(fs_info, &fs_info->global_block_rsv, NULL,
5918                                 (u64)-1, NULL);
5919         WARN_ON(fs_info->trans_block_rsv.size > 0);
5920         WARN_ON(fs_info->trans_block_rsv.reserved > 0);
5921         WARN_ON(fs_info->chunk_block_rsv.size > 0);
5922         WARN_ON(fs_info->chunk_block_rsv.reserved > 0);
5923         WARN_ON(fs_info->delayed_block_rsv.size > 0);
5924         WARN_ON(fs_info->delayed_block_rsv.reserved > 0);
5925 }
5926
5927
5928 /*
5929  * To be called after all the new block groups attached to the transaction
5930  * handle have been created (btrfs_create_pending_block_groups()).
5931  */
5932 void btrfs_trans_release_chunk_metadata(struct btrfs_trans_handle *trans)
5933 {
5934         struct btrfs_fs_info *fs_info = trans->fs_info;
5935
5936         if (!trans->chunk_bytes_reserved)
5937                 return;
5938
5939         WARN_ON_ONCE(!list_empty(&trans->new_bgs));
5940
5941         block_rsv_release_bytes(fs_info, &fs_info->chunk_block_rsv, NULL,
5942                                 trans->chunk_bytes_reserved, NULL);
5943         trans->chunk_bytes_reserved = 0;
5944 }
5945
5946 /* Can only return 0 or -ENOSPC */
5947 int btrfs_orphan_reserve_metadata(struct btrfs_trans_handle *trans,
5948                                   struct btrfs_inode *inode)
5949 {
5950         struct btrfs_fs_info *fs_info = btrfs_sb(inode->vfs_inode.i_sb);
5951         struct btrfs_root *root = inode->root;
5952         /*
5953          * We always use trans->block_rsv here as we will have reserved space
5954          * for our orphan when starting the transaction, using get_block_rsv()
5955          * here will sometimes make us choose the wrong block rsv as we could be
5956          * doing a reloc inode for a non refcounted root.
5957          */
5958         struct btrfs_block_rsv *src_rsv = trans->block_rsv;
5959         struct btrfs_block_rsv *dst_rsv = root->orphan_block_rsv;
5960
5961         /*
5962          * We need to hold space in order to delete our orphan item once we've
5963          * added it, so this takes the reservation so we can release it later
5964          * when we are truly done with the orphan item.
5965          */
5966         u64 num_bytes = btrfs_calc_trans_metadata_size(fs_info, 1);
5967
5968         trace_btrfs_space_reservation(fs_info, "orphan", btrfs_ino(inode),
5969                         num_bytes, 1);
5970         return btrfs_block_rsv_migrate(src_rsv, dst_rsv, num_bytes, 1);
5971 }
5972
5973 void btrfs_orphan_release_metadata(struct btrfs_inode *inode)
5974 {
5975         struct btrfs_fs_info *fs_info = btrfs_sb(inode->vfs_inode.i_sb);
5976         struct btrfs_root *root = inode->root;
5977         u64 num_bytes = btrfs_calc_trans_metadata_size(fs_info, 1);
5978
5979         trace_btrfs_space_reservation(fs_info, "orphan", btrfs_ino(inode),
5980                         num_bytes, 0);
5981         btrfs_block_rsv_release(fs_info, root->orphan_block_rsv, num_bytes);
5982 }
5983
5984 /*
5985  * btrfs_subvolume_reserve_metadata() - reserve space for subvolume operation
5986  * root: the root of the parent directory
5987  * rsv: block reservation
5988  * items: the number of items that we need do reservation
5989  * qgroup_reserved: used to return the reserved size in qgroup
5990  *
5991  * This function is used to reserve the space for snapshot/subvolume
5992  * creation and deletion. Those operations are different with the
5993  * common file/directory operations, they change two fs/file trees
5994  * and root tree, the number of items that the qgroup reserves is
5995  * different with the free space reservation. So we can not use
5996  * the space reservation mechanism in start_transaction().
5997  */
5998 int btrfs_subvolume_reserve_metadata(struct btrfs_root *root,
5999                                      struct btrfs_block_rsv *rsv,
6000                                      int items,
6001                                      u64 *qgroup_reserved,
6002                                      bool use_global_rsv)
6003 {
6004         u64 num_bytes;
6005         int ret;
6006         struct btrfs_fs_info *fs_info = root->fs_info;
6007         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
6008
6009         if (test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags)) {
6010                 /* One for parent inode, two for dir entries */
6011                 num_bytes = 3 * fs_info->nodesize;
6012                 ret = btrfs_qgroup_reserve_meta_prealloc(root, num_bytes, true);
6013                 if (ret)
6014                         return ret;
6015         } else {
6016                 num_bytes = 0;
6017         }
6018
6019         *qgroup_reserved = num_bytes;
6020
6021         num_bytes = btrfs_calc_trans_metadata_size(fs_info, items);
6022         rsv->space_info = __find_space_info(fs_info,
6023                                             BTRFS_BLOCK_GROUP_METADATA);
6024         ret = btrfs_block_rsv_add(root, rsv, num_bytes,
6025                                   BTRFS_RESERVE_FLUSH_ALL);
6026
6027         if (ret == -ENOSPC && use_global_rsv)
6028                 ret = btrfs_block_rsv_migrate(global_rsv, rsv, num_bytes, 1);
6029
6030         if (ret && *qgroup_reserved)
6031                 btrfs_qgroup_free_meta_prealloc(root, *qgroup_reserved);
6032
6033         return ret;
6034 }
6035
6036 void btrfs_subvolume_release_metadata(struct btrfs_fs_info *fs_info,
6037                                       struct btrfs_block_rsv *rsv)
6038 {
6039         btrfs_block_rsv_release(fs_info, rsv, (u64)-1);
6040 }
6041
6042 static void btrfs_calculate_inode_block_rsv_size(struct btrfs_fs_info *fs_info,
6043                                                  struct btrfs_inode *inode)
6044 {
6045         struct btrfs_block_rsv *block_rsv = &inode->block_rsv;
6046         u64 reserve_size = 0;
6047         u64 qgroup_rsv_size = 0;
6048         u64 csum_leaves;
6049         unsigned outstanding_extents;
6050
6051         lockdep_assert_held(&inode->lock);
6052         outstanding_extents = inode->outstanding_extents;
6053         if (outstanding_extents)
6054                 reserve_size = btrfs_calc_trans_metadata_size(fs_info,
6055                                                 outstanding_extents + 1);
6056         csum_leaves = btrfs_csum_bytes_to_leaves(fs_info,
6057                                                  inode->csum_bytes);
6058         reserve_size += btrfs_calc_trans_metadata_size(fs_info,
6059                                                        csum_leaves);
6060         /*
6061          * For qgroup rsv, the calculation is very simple:
6062          * account one nodesize for each outstanding extent
6063          *
6064          * This is overestimating in most cases.
6065          */
6066         qgroup_rsv_size = outstanding_extents * fs_info->nodesize;
6067
6068         spin_lock(&block_rsv->lock);
6069         block_rsv->size = reserve_size;
6070         block_rsv->qgroup_rsv_size = qgroup_rsv_size;
6071         spin_unlock(&block_rsv->lock);
6072 }
6073
6074 int btrfs_delalloc_reserve_metadata(struct btrfs_inode *inode, u64 num_bytes)
6075 {
6076         struct btrfs_fs_info *fs_info = btrfs_sb(inode->vfs_inode.i_sb);
6077         unsigned nr_extents;
6078         enum btrfs_reserve_flush_enum flush = BTRFS_RESERVE_FLUSH_ALL;
6079         int ret = 0;
6080         bool delalloc_lock = true;
6081
6082         /* If we are a free space inode we need to not flush since we will be in
6083          * the middle of a transaction commit.  We also don't need the delalloc
6084          * mutex since we won't race with anybody.  We need this mostly to make
6085          * lockdep shut its filthy mouth.
6086          *
6087          * If we have a transaction open (can happen if we call truncate_block
6088          * from truncate), then we need FLUSH_LIMIT so we don't deadlock.
6089          */
6090         if (btrfs_is_free_space_inode(inode)) {
6091                 flush = BTRFS_RESERVE_NO_FLUSH;
6092                 delalloc_lock = false;
6093         } else {
6094                 if (current->journal_info)
6095                         flush = BTRFS_RESERVE_FLUSH_LIMIT;
6096
6097                 if (btrfs_transaction_in_commit(fs_info))
6098                         schedule_timeout(1);
6099         }
6100
6101         if (delalloc_lock)
6102                 mutex_lock(&inode->delalloc_mutex);
6103
6104         num_bytes = ALIGN(num_bytes, fs_info->sectorsize);
6105
6106         /* Add our new extents and calculate the new rsv size. */
6107         spin_lock(&inode->lock);
6108         nr_extents = count_max_extents(num_bytes);
6109         btrfs_mod_outstanding_extents(inode, nr_extents);
6110         inode->csum_bytes += num_bytes;
6111         btrfs_calculate_inode_block_rsv_size(fs_info, inode);
6112         spin_unlock(&inode->lock);
6113
6114         ret = btrfs_inode_rsv_refill(inode, flush);
6115         if (unlikely(ret))
6116                 goto out_fail;
6117
6118         if (delalloc_lock)
6119                 mutex_unlock(&inode->delalloc_mutex);
6120         return 0;
6121
6122 out_fail:
6123         spin_lock(&inode->lock);
6124         nr_extents = count_max_extents(num_bytes);
6125         btrfs_mod_outstanding_extents(inode, -nr_extents);
6126         inode->csum_bytes -= num_bytes;
6127         btrfs_calculate_inode_block_rsv_size(fs_info, inode);
6128         spin_unlock(&inode->lock);
6129
6130         btrfs_inode_rsv_release(inode, true);
6131         if (delalloc_lock)
6132                 mutex_unlock(&inode->delalloc_mutex);
6133         return ret;
6134 }
6135
6136 /**
6137  * btrfs_delalloc_release_metadata - release a metadata reservation for an inode
6138  * @inode: the inode to release the reservation for.
6139  * @num_bytes: the number of bytes we are releasing.
6140  * @qgroup_free: free qgroup reservation or convert it to per-trans reservation
6141  *
6142  * This will release the metadata reservation for an inode.  This can be called
6143  * once we complete IO for a given set of bytes to release their metadata
6144  * reservations, or on error for the same reason.
6145  */
6146 void btrfs_delalloc_release_metadata(struct btrfs_inode *inode, u64 num_bytes,
6147                                      bool qgroup_free)
6148 {
6149         struct btrfs_fs_info *fs_info = btrfs_sb(inode->vfs_inode.i_sb);
6150
6151         num_bytes = ALIGN(num_bytes, fs_info->sectorsize);
6152         spin_lock(&inode->lock);
6153         inode->csum_bytes -= num_bytes;
6154         btrfs_calculate_inode_block_rsv_size(fs_info, inode);
6155         spin_unlock(&inode->lock);
6156
6157         if (btrfs_is_testing(fs_info))
6158                 return;
6159
6160         btrfs_inode_rsv_release(inode, qgroup_free);
6161 }
6162
6163 /**
6164  * btrfs_delalloc_release_extents - release our outstanding_extents
6165  * @inode: the inode to balance the reservation for.
6166  * @num_bytes: the number of bytes we originally reserved with
6167  * @qgroup_free: do we need to free qgroup meta reservation or convert them.
6168  *
6169  * When we reserve space we increase outstanding_extents for the extents we may
6170  * add.  Once we've set the range as delalloc or created our ordered extents we
6171  * have outstanding_extents to track the real usage, so we use this to free our
6172  * temporarily tracked outstanding_extents.  This _must_ be used in conjunction
6173  * with btrfs_delalloc_reserve_metadata.
6174  */
6175 void btrfs_delalloc_release_extents(struct btrfs_inode *inode, u64 num_bytes,
6176                                     bool qgroup_free)
6177 {
6178         struct btrfs_fs_info *fs_info = btrfs_sb(inode->vfs_inode.i_sb);
6179         unsigned num_extents;
6180
6181         spin_lock(&inode->lock);
6182         num_extents = count_max_extents(num_bytes);
6183         btrfs_mod_outstanding_extents(inode, -num_extents);
6184         btrfs_calculate_inode_block_rsv_size(fs_info, inode);
6185         spin_unlock(&inode->lock);
6186
6187         if (btrfs_is_testing(fs_info))
6188                 return;
6189
6190         btrfs_inode_rsv_release(inode, qgroup_free);
6191 }
6192
6193 /**
6194  * btrfs_delalloc_reserve_space - reserve data and metadata space for
6195  * delalloc
6196  * @inode: inode we're writing to
6197  * @start: start range we are writing to
6198  * @len: how long the range we are writing to
6199  * @reserved: mandatory parameter, record actually reserved qgroup ranges of
6200  *            current reservation.
6201  *
6202  * This will do the following things
6203  *
6204  * o reserve space in data space info for num bytes
6205  *   and reserve precious corresponding qgroup space
6206  *   (Done in check_data_free_space)
6207  *
6208  * o reserve space for metadata space, based on the number of outstanding
6209  *   extents and how much csums will be needed
6210  *   also reserve metadata space in a per root over-reserve method.
6211  * o add to the inodes->delalloc_bytes
6212  * o add it to the fs_info's delalloc inodes list.
6213  *   (Above 3 all done in delalloc_reserve_metadata)
6214  *
6215  * Return 0 for success
6216  * Return <0 for error(-ENOSPC or -EQUOT)
6217  */
6218 int btrfs_delalloc_reserve_space(struct inode *inode,
6219                         struct extent_changeset **reserved, u64 start, u64 len)
6220 {
6221         int ret;
6222
6223         ret = btrfs_check_data_free_space(inode, reserved, start, len);
6224         if (ret < 0)
6225                 return ret;
6226         ret = btrfs_delalloc_reserve_metadata(BTRFS_I(inode), len);
6227         if (ret < 0)
6228                 btrfs_free_reserved_data_space(inode, *reserved, start, len);
6229         return ret;
6230 }
6231
6232 /**
6233  * btrfs_delalloc_release_space - release data and metadata space for delalloc
6234  * @inode: inode we're releasing space for
6235  * @start: start position of the space already reserved
6236  * @len: the len of the space already reserved
6237  * @release_bytes: the len of the space we consumed or didn't use
6238  *
6239  * This function will release the metadata space that was not used and will
6240  * decrement ->delalloc_bytes and remove it from the fs_info delalloc_inodes
6241  * list if there are no delalloc bytes left.
6242  * Also it will handle the qgroup reserved space.
6243  */
6244 void btrfs_delalloc_release_space(struct inode *inode,
6245                                   struct extent_changeset *reserved,
6246                                   u64 start, u64 len, bool qgroup_free)
6247 {
6248         btrfs_delalloc_release_metadata(BTRFS_I(inode), len, qgroup_free);
6249         btrfs_free_reserved_data_space(inode, reserved, start, len);
6250 }
6251
6252 static int update_block_group(struct btrfs_trans_handle *trans,
6253                               struct btrfs_fs_info *info, u64 bytenr,
6254                               u64 num_bytes, int alloc)
6255 {
6256         struct btrfs_block_group_cache *cache = NULL;
6257         u64 total = num_bytes;
6258         u64 old_val;
6259         u64 byte_in_group;
6260         int factor;
6261
6262         /* block accounting for super block */
6263         spin_lock(&info->delalloc_root_lock);
6264         old_val = btrfs_super_bytes_used(info->super_copy);
6265         if (alloc)
6266                 old_val += num_bytes;
6267         else
6268                 old_val -= num_bytes;
6269         btrfs_set_super_bytes_used(info->super_copy, old_val);
6270         spin_unlock(&info->delalloc_root_lock);
6271
6272         while (total) {
6273                 cache = btrfs_lookup_block_group(info, bytenr);
6274                 if (!cache)
6275                         return -ENOENT;
6276                 if (cache->flags & (BTRFS_BLOCK_GROUP_DUP |
6277                                     BTRFS_BLOCK_GROUP_RAID1 |
6278                                     BTRFS_BLOCK_GROUP_RAID10))
6279                         factor = 2;
6280                 else
6281                         factor = 1;
6282                 /*
6283                  * If this block group has free space cache written out, we
6284                  * need to make sure to load it if we are removing space.  This
6285                  * is because we need the unpinning stage to actually add the
6286                  * space back to the block group, otherwise we will leak space.
6287                  */
6288                 if (!alloc && cache->cached == BTRFS_CACHE_NO)
6289                         cache_block_group(cache, 1);
6290
6291                 byte_in_group = bytenr - cache->key.objectid;
6292                 WARN_ON(byte_in_group > cache->key.offset);
6293
6294                 spin_lock(&cache->space_info->lock);
6295                 spin_lock(&cache->lock);
6296
6297                 if (btrfs_test_opt(info, SPACE_CACHE) &&
6298                     cache->disk_cache_state < BTRFS_DC_CLEAR)
6299                         cache->disk_cache_state = BTRFS_DC_CLEAR;
6300
6301                 old_val = btrfs_block_group_used(&cache->item);
6302                 num_bytes = min(total, cache->key.offset - byte_in_group);
6303                 if (alloc) {
6304                         old_val += num_bytes;
6305                         btrfs_set_block_group_used(&cache->item, old_val);
6306                         cache->reserved -= num_bytes;
6307                         cache->space_info->bytes_reserved -= num_bytes;
6308                         cache->space_info->bytes_used += num_bytes;
6309                         cache->space_info->disk_used += num_bytes * factor;
6310                         spin_unlock(&cache->lock);
6311                         spin_unlock(&cache->space_info->lock);
6312                 } else {
6313                         old_val -= num_bytes;
6314                         btrfs_set_block_group_used(&cache->item, old_val);
6315                         cache->pinned += num_bytes;
6316                         cache->space_info->bytes_pinned += num_bytes;
6317                         cache->space_info->bytes_used -= num_bytes;
6318                         cache->space_info->disk_used -= num_bytes * factor;
6319                         spin_unlock(&cache->lock);
6320                         spin_unlock(&cache->space_info->lock);
6321
6322                         trace_btrfs_space_reservation(info, "pinned",
6323                                                       cache->space_info->flags,
6324                                                       num_bytes, 1);
6325                         percpu_counter_add(&cache->space_info->total_bytes_pinned,
6326                                            num_bytes);
6327                         set_extent_dirty(info->pinned_extents,
6328                                          bytenr, bytenr + num_bytes - 1,
6329                                          GFP_NOFS | __GFP_NOFAIL);
6330                 }
6331
6332                 spin_lock(&trans->transaction->dirty_bgs_lock);
6333                 if (list_empty(&cache->dirty_list)) {
6334                         list_add_tail(&cache->dirty_list,
6335                                       &trans->transaction->dirty_bgs);
6336                                 trans->transaction->num_dirty_bgs++;
6337                         btrfs_get_block_group(cache);
6338                 }
6339                 spin_unlock(&trans->transaction->dirty_bgs_lock);
6340
6341                 /*
6342                  * No longer have used bytes in this block group, queue it for
6343                  * deletion. We do this after adding the block group to the
6344                  * dirty list to avoid races between cleaner kthread and space
6345                  * cache writeout.
6346                  */
6347                 if (!alloc && old_val == 0) {
6348                         spin_lock(&info->unused_bgs_lock);
6349                         if (list_empty(&cache->bg_list)) {
6350                                 btrfs_get_block_group(cache);
6351                                 list_add_tail(&cache->bg_list,
6352                                               &info->unused_bgs);
6353                         }
6354                         spin_unlock(&info->unused_bgs_lock);
6355                 }
6356
6357                 btrfs_put_block_group(cache);
6358                 total -= num_bytes;
6359                 bytenr += num_bytes;
6360         }
6361         return 0;
6362 }
6363
6364 static u64 first_logical_byte(struct btrfs_fs_info *fs_info, u64 search_start)
6365 {
6366         struct btrfs_block_group_cache *cache;
6367         u64 bytenr;
6368
6369         spin_lock(&fs_info->block_group_cache_lock);
6370         bytenr = fs_info->first_logical_byte;
6371         spin_unlock(&fs_info->block_group_cache_lock);
6372
6373         if (bytenr < (u64)-1)
6374                 return bytenr;
6375
6376         cache = btrfs_lookup_first_block_group(fs_info, search_start);
6377         if (!cache)
6378                 return 0;
6379
6380         bytenr = cache->key.objectid;
6381         btrfs_put_block_group(cache);
6382
6383         return bytenr;
6384 }
6385
6386 static int pin_down_extent(struct btrfs_fs_info *fs_info,
6387                            struct btrfs_block_group_cache *cache,
6388                            u64 bytenr, u64 num_bytes, int reserved)
6389 {
6390         spin_lock(&cache->space_info->lock);
6391         spin_lock(&cache->lock);
6392         cache->pinned += num_bytes;
6393         cache->space_info->bytes_pinned += num_bytes;
6394         if (reserved) {
6395                 cache->reserved -= num_bytes;
6396                 cache->space_info->bytes_reserved -= num_bytes;
6397         }
6398         spin_unlock(&cache->lock);
6399         spin_unlock(&cache->space_info->lock);
6400
6401         trace_btrfs_space_reservation(fs_info, "pinned",
6402                                       cache->space_info->flags, num_bytes, 1);
6403         percpu_counter_add(&cache->space_info->total_bytes_pinned, num_bytes);
6404         set_extent_dirty(fs_info->pinned_extents, bytenr,
6405                          bytenr + num_bytes - 1, GFP_NOFS | __GFP_NOFAIL);
6406         return 0;
6407 }
6408
6409 /*
6410  * this function must be called within transaction
6411  */
6412 int btrfs_pin_extent(struct btrfs_fs_info *fs_info,
6413                      u64 bytenr, u64 num_bytes, int reserved)
6414 {
6415         struct btrfs_block_group_cache *cache;
6416
6417         cache = btrfs_lookup_block_group(fs_info, bytenr);
6418         BUG_ON(!cache); /* Logic error */
6419
6420         pin_down_extent(fs_info, cache, bytenr, num_bytes, reserved);
6421
6422         btrfs_put_block_group(cache);
6423         return 0;
6424 }
6425
6426 /*
6427  * this function must be called within transaction
6428  */
6429 int btrfs_pin_extent_for_log_replay(struct btrfs_fs_info *fs_info,
6430                                     u64 bytenr, u64 num_bytes)
6431 {
6432         struct btrfs_block_group_cache *cache;
6433         int ret;
6434
6435         cache = btrfs_lookup_block_group(fs_info, bytenr);
6436         if (!cache)
6437                 return -EINVAL;
6438
6439         /*
6440          * pull in the free space cache (if any) so that our pin
6441          * removes the free space from the cache.  We have load_only set
6442          * to one because the slow code to read in the free extents does check
6443          * the pinned extents.
6444          */
6445         cache_block_group(cache, 1);
6446
6447         pin_down_extent(fs_info, cache, bytenr, num_bytes, 0);
6448
6449         /* remove us from the free space cache (if we're there at all) */
6450         ret = btrfs_remove_free_space(cache, bytenr, num_bytes);
6451         btrfs_put_block_group(cache);
6452         return ret;
6453 }
6454
6455 static int __exclude_logged_extent(struct btrfs_fs_info *fs_info,
6456                                    u64 start, u64 num_bytes)
6457 {
6458         int ret;
6459         struct btrfs_block_group_cache *block_group;
6460         struct btrfs_caching_control *caching_ctl;
6461
6462         block_group = btrfs_lookup_block_group(fs_info, start);
6463         if (!block_group)
6464                 return -EINVAL;
6465
6466         cache_block_group(block_group, 0);
6467         caching_ctl = get_caching_control(block_group);
6468
6469         if (!caching_ctl) {
6470                 /* Logic error */
6471                 BUG_ON(!block_group_cache_done(block_group));
6472                 ret = btrfs_remove_free_space(block_group, start, num_bytes);
6473         } else {
6474                 mutex_lock(&caching_ctl->mutex);
6475
6476                 if (start >= caching_ctl->progress) {
6477                         ret = add_excluded_extent(fs_info, start, num_bytes);
6478                 } else if (start + num_bytes <= caching_ctl->progress) {
6479                         ret = btrfs_remove_free_space(block_group,
6480                                                       start, num_bytes);
6481                 } else {
6482                         num_bytes = caching_ctl->progress - start;
6483                         ret = btrfs_remove_free_space(block_group,
6484                                                       start, num_bytes);
6485                         if (ret)
6486                                 goto out_lock;
6487
6488                         num_bytes = (start + num_bytes) -
6489                                 caching_ctl->progress;
6490                         start = caching_ctl->progress;
6491                         ret = add_excluded_extent(fs_info, start, num_bytes);
6492                 }
6493 out_lock:
6494                 mutex_unlock(&caching_ctl->mutex);
6495                 put_caching_control(caching_ctl);
6496         }
6497         btrfs_put_block_group(block_group);
6498         return ret;
6499 }
6500
6501 int btrfs_exclude_logged_extents(struct btrfs_fs_info *fs_info,
6502                                  struct extent_buffer *eb)
6503 {
6504         struct btrfs_file_extent_item *item;
6505         struct btrfs_key key;
6506         int found_type;
6507         int i;
6508
6509         if (!btrfs_fs_incompat(fs_info, MIXED_GROUPS))
6510                 return 0;
6511
6512         for (i = 0; i < btrfs_header_nritems(eb); i++) {
6513                 btrfs_item_key_to_cpu(eb, &key, i);
6514                 if (key.type != BTRFS_EXTENT_DATA_KEY)
6515                         continue;
6516                 item = btrfs_item_ptr(eb, i, struct btrfs_file_extent_item);
6517                 found_type = btrfs_file_extent_type(eb, item);
6518                 if (found_type == BTRFS_FILE_EXTENT_INLINE)
6519                         continue;
6520                 if (btrfs_file_extent_disk_bytenr(eb, item) == 0)
6521                         continue;
6522                 key.objectid = btrfs_file_extent_disk_bytenr(eb, item);
6523                 key.offset = btrfs_file_extent_disk_num_bytes(eb, item);
6524                 __exclude_logged_extent(fs_info, key.objectid, key.offset);
6525         }
6526
6527         return 0;
6528 }
6529
6530 static void
6531 btrfs_inc_block_group_reservations(struct btrfs_block_group_cache *bg)
6532 {
6533         atomic_inc(&bg->reservations);
6534 }
6535
6536 void btrfs_dec_block_group_reservations(struct btrfs_fs_info *fs_info,
6537                                         const u64 start)
6538 {
6539         struct btrfs_block_group_cache *bg;
6540
6541         bg = btrfs_lookup_block_group(fs_info, start);
6542         ASSERT(bg);
6543         if (atomic_dec_and_test(&bg->reservations))
6544                 wake_up_atomic_t(&bg->reservations);
6545         btrfs_put_block_group(bg);
6546 }
6547
6548 void btrfs_wait_block_group_reservations(struct btrfs_block_group_cache *bg)
6549 {
6550         struct btrfs_space_info *space_info = bg->space_info;
6551
6552         ASSERT(bg->ro);
6553
6554         if (!(bg->flags & BTRFS_BLOCK_GROUP_DATA))
6555                 return;
6556
6557         /*
6558          * Our block group is read only but before we set it to read only,
6559          * some task might have had allocated an extent from it already, but it
6560          * has not yet created a respective ordered extent (and added it to a
6561          * root's list of ordered extents).
6562          * Therefore wait for any task currently allocating extents, since the
6563          * block group's reservations counter is incremented while a read lock
6564          * on the groups' semaphore is held and decremented after releasing
6565          * the read access on that semaphore and creating the ordered extent.
6566          */
6567         down_write(&space_info->groups_sem);
6568         up_write(&space_info->groups_sem);
6569
6570         wait_on_atomic_t(&bg->reservations, atomic_t_wait,
6571                          TASK_UNINTERRUPTIBLE);
6572 }
6573
6574 /**
6575  * btrfs_add_reserved_bytes - update the block_group and space info counters
6576  * @cache:      The cache we are manipulating
6577  * @ram_bytes:  The number of bytes of file content, and will be same to
6578  *              @num_bytes except for the compress path.
6579  * @num_bytes:  The number of bytes in question
6580  * @delalloc:   The blocks are allocated for the delalloc write
6581  *
6582  * This is called by the allocator when it reserves space. If this is a
6583  * reservation and the block group has become read only we cannot make the
6584  * reservation and return -EAGAIN, otherwise this function always succeeds.
6585  */
6586 static int btrfs_add_reserved_bytes(struct btrfs_block_group_cache *cache,
6587                                     u64 ram_bytes, u64 num_bytes, int delalloc)
6588 {
6589         struct btrfs_space_info *space_info = cache->space_info;
6590         int ret = 0;
6591
6592         spin_lock(&space_info->lock);
6593         spin_lock(&cache->lock);
6594         if (cache->ro) {
6595                 ret = -EAGAIN;
6596         } else {
6597                 cache->reserved += num_bytes;
6598                 space_info->bytes_reserved += num_bytes;
6599
6600                 trace_btrfs_space_reservation(cache->fs_info,
6601                                 "space_info", space_info->flags,
6602                                 ram_bytes, 0);
6603                 space_info->bytes_may_use -= ram_bytes;
6604                 if (delalloc)
6605                         cache->delalloc_bytes += num_bytes;
6606         }
6607         spin_unlock(&cache->lock);
6608         spin_unlock(&space_info->lock);
6609         return ret;
6610 }
6611
6612 /**
6613  * btrfs_free_reserved_bytes - update the block_group and space info counters
6614  * @cache:      The cache we are manipulating
6615  * @num_bytes:  The number of bytes in question
6616  * @delalloc:   The blocks are allocated for the delalloc write
6617  *
6618  * This is called by somebody who is freeing space that was never actually used
6619  * on disk.  For example if you reserve some space for a new leaf in transaction
6620  * A and before transaction A commits you free that leaf, you call this with
6621  * reserve set to 0 in order to clear the reservation.
6622  */
6623
6624 static int btrfs_free_reserved_bytes(struct btrfs_block_group_cache *cache,
6625                                      u64 num_bytes, int delalloc)
6626 {
6627         struct btrfs_space_info *space_info = cache->space_info;
6628         int ret = 0;
6629
6630         spin_lock(&space_info->lock);
6631         spin_lock(&cache->lock);
6632         if (cache->ro)
6633                 space_info->bytes_readonly += num_bytes;
6634         cache->reserved -= num_bytes;
6635         space_info->bytes_reserved -= num_bytes;
6636
6637         if (delalloc)
6638                 cache->delalloc_bytes -= num_bytes;
6639         spin_unlock(&cache->lock);
6640         spin_unlock(&space_info->lock);
6641         return ret;
6642 }
6643 void btrfs_prepare_extent_commit(struct btrfs_fs_info *fs_info)
6644 {
6645         struct btrfs_caching_control *next;
6646         struct btrfs_caching_control *caching_ctl;
6647         struct btrfs_block_group_cache *cache;
6648
6649         down_write(&fs_info->commit_root_sem);
6650
6651         list_for_each_entry_safe(caching_ctl, next,
6652                                  &fs_info->caching_block_groups, list) {
6653                 cache = caching_ctl->block_group;
6654                 if (block_group_cache_done(cache)) {
6655                         cache->last_byte_to_unpin = (u64)-1;
6656                         list_del_init(&caching_ctl->list);
6657                         put_caching_control(caching_ctl);
6658                 } else {
6659                         cache->last_byte_to_unpin = caching_ctl->progress;
6660                 }
6661         }
6662
6663         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
6664                 fs_info->pinned_extents = &fs_info->freed_extents[1];
6665         else
6666                 fs_info->pinned_extents = &fs_info->freed_extents[0];
6667
6668         up_write(&fs_info->commit_root_sem);
6669
6670         update_global_block_rsv(fs_info);
6671 }
6672
6673 /*
6674  * Returns the free cluster for the given space info and sets empty_cluster to
6675  * what it should be based on the mount options.
6676  */
6677 static struct btrfs_free_cluster *
6678 fetch_cluster_info(struct btrfs_fs_info *fs_info,
6679                    struct btrfs_space_info *space_info, u64 *empty_cluster)
6680 {
6681         struct btrfs_free_cluster *ret = NULL;
6682
6683         *empty_cluster = 0;
6684         if (btrfs_mixed_space_info(space_info))
6685                 return ret;
6686
6687         if (space_info->flags & BTRFS_BLOCK_GROUP_METADATA) {
6688                 ret = &fs_info->meta_alloc_cluster;
6689                 if (btrfs_test_opt(fs_info, SSD))
6690                         *empty_cluster = SZ_2M;
6691                 else
6692                         *empty_cluster = SZ_64K;
6693         } else if ((space_info->flags & BTRFS_BLOCK_GROUP_DATA) &&
6694                    btrfs_test_opt(fs_info, SSD_SPREAD)) {
6695                 *empty_cluster = SZ_2M;
6696                 ret = &fs_info->data_alloc_cluster;
6697         }
6698
6699         return ret;
6700 }
6701
6702 static int unpin_extent_range(struct btrfs_fs_info *fs_info,
6703                               u64 start, u64 end,
6704                               const bool return_free_space)
6705 {
6706         struct btrfs_block_group_cache *cache = NULL;
6707         struct btrfs_space_info *space_info;
6708         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
6709         struct btrfs_free_cluster *cluster = NULL;
6710         u64 len;
6711         u64 total_unpinned = 0;
6712         u64 empty_cluster = 0;
6713         bool readonly;
6714
6715         while (start <= end) {
6716                 readonly = false;
6717                 if (!cache ||
6718                     start >= cache->key.objectid + cache->key.offset) {
6719                         if (cache)
6720                                 btrfs_put_block_group(cache);
6721                         total_unpinned = 0;
6722                         cache = btrfs_lookup_block_group(fs_info, start);
6723                         BUG_ON(!cache); /* Logic error */
6724
6725                         cluster = fetch_cluster_info(fs_info,
6726                                                      cache->space_info,
6727                                                      &empty_cluster);
6728                         empty_cluster <<= 1;
6729                 }
6730
6731                 len = cache->key.objectid + cache->key.offset - start;
6732                 len = min(len, end + 1 - start);
6733
6734                 if (start < cache->last_byte_to_unpin) {
6735                         len = min(len, cache->last_byte_to_unpin - start);
6736                         if (return_free_space)
6737                                 btrfs_add_free_space(cache, start, len);
6738                 }
6739
6740                 start += len;
6741                 total_unpinned += len;
6742                 space_info = cache->space_info;
6743
6744                 /*
6745                  * If this space cluster has been marked as fragmented and we've
6746                  * unpinned enough in this block group to potentially allow a
6747                  * cluster to be created inside of it go ahead and clear the
6748                  * fragmented check.
6749                  */
6750                 if (cluster && cluster->fragmented &&
6751                     total_unpinned > empty_cluster) {
6752                         spin_lock(&cluster->lock);
6753                         cluster->fragmented = 0;
6754                         spin_unlock(&cluster->lock);
6755                 }
6756
6757                 spin_lock(&space_info->lock);
6758                 spin_lock(&cache->lock);
6759                 cache->pinned -= len;
6760                 space_info->bytes_pinned -= len;
6761
6762                 trace_btrfs_space_reservation(fs_info, "pinned",
6763                                               space_info->flags, len, 0);
6764                 space_info->max_extent_size = 0;
6765                 percpu_counter_add(&space_info->total_bytes_pinned, -len);
6766                 if (cache->ro) {
6767                         space_info->bytes_readonly += len;
6768                         readonly = true;
6769                 }
6770                 spin_unlock(&cache->lock);
6771                 if (!readonly && return_free_space &&
6772                     global_rsv->space_info == space_info) {
6773                         u64 to_add = len;
6774
6775                         spin_lock(&global_rsv->lock);
6776                         if (!global_rsv->full) {
6777                                 to_add = min(len, global_rsv->size -
6778                                              global_rsv->reserved);
6779                                 global_rsv->reserved += to_add;
6780                                 space_info->bytes_may_use += to_add;
6781                                 if (global_rsv->reserved >= global_rsv->size)
6782                                         global_rsv->full = 1;
6783                                 trace_btrfs_space_reservation(fs_info,
6784                                                               "space_info",
6785                                                               space_info->flags,
6786                                                               to_add, 1);
6787                                 len -= to_add;
6788                         }
6789                         spin_unlock(&global_rsv->lock);
6790                         /* Add to any tickets we may have */
6791                         if (len)
6792                                 space_info_add_new_bytes(fs_info, space_info,
6793                                                          len);
6794                 }
6795                 spin_unlock(&space_info->lock);
6796         }
6797
6798         if (cache)
6799                 btrfs_put_block_group(cache);
6800         return 0;
6801 }
6802
6803 int btrfs_finish_extent_commit(struct btrfs_trans_handle *trans)
6804 {
6805         struct btrfs_fs_info *fs_info = trans->fs_info;
6806         struct btrfs_block_group_cache *block_group, *tmp;
6807         struct list_head *deleted_bgs;
6808         struct extent_io_tree *unpin;
6809         u64 start;
6810         u64 end;
6811         int ret;
6812
6813         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
6814                 unpin = &fs_info->freed_extents[1];
6815         else
6816                 unpin = &fs_info->freed_extents[0];
6817
6818         while (!trans->aborted) {
6819                 mutex_lock(&fs_info->unused_bg_unpin_mutex);
6820                 ret = find_first_extent_bit(unpin, 0, &start, &end,
6821                                             EXTENT_DIRTY, NULL);
6822                 if (ret) {
6823                         mutex_unlock(&fs_info->unused_bg_unpin_mutex);
6824                         break;
6825                 }
6826
6827                 if (btrfs_test_opt(fs_info, DISCARD))
6828                         ret = btrfs_discard_extent(fs_info, start,
6829                                                    end + 1 - start, NULL);
6830
6831                 clear_extent_dirty(unpin, start, end);
6832                 unpin_extent_range(fs_info, start, end, true);
6833                 mutex_unlock(&fs_info->unused_bg_unpin_mutex);
6834                 cond_resched();
6835         }
6836
6837         /*
6838          * Transaction is finished.  We don't need the lock anymore.  We
6839          * do need to clean up the block groups in case of a transaction
6840          * abort.
6841          */
6842         deleted_bgs = &trans->transaction->deleted_bgs;
6843         list_for_each_entry_safe(block_group, tmp, deleted_bgs, bg_list) {
6844                 u64 trimmed = 0;
6845
6846                 ret = -EROFS;
6847                 if (!trans->aborted)
6848                         ret = btrfs_discard_extent(fs_info,
6849                                                    block_group->key.objectid,
6850                                                    block_group->key.offset,
6851                                                    &trimmed);
6852
6853                 list_del_init(&block_group->bg_list);
6854                 btrfs_put_block_group_trimming(block_group);
6855                 btrfs_put_block_group(block_group);
6856
6857                 if (ret) {
6858                         const char *errstr = btrfs_decode_error(ret);
6859                         btrfs_warn(fs_info,
6860                            "discard failed while removing blockgroup: errno=%d %s",
6861                                    ret, errstr);
6862                 }
6863         }
6864
6865         return 0;
6866 }
6867
6868 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
6869                                 struct btrfs_fs_info *info,
6870                                 struct btrfs_delayed_ref_node *node, u64 parent,
6871                                 u64 root_objectid, u64 owner_objectid,
6872                                 u64 owner_offset, int refs_to_drop,
6873                                 struct btrfs_delayed_extent_op *extent_op)
6874 {
6875         struct btrfs_key key;
6876         struct btrfs_path *path;
6877         struct btrfs_root *extent_root = info->extent_root;
6878         struct extent_buffer *leaf;
6879         struct btrfs_extent_item *ei;
6880         struct btrfs_extent_inline_ref *iref;
6881         int ret;
6882         int is_data;
6883         int extent_slot = 0;
6884         int found_extent = 0;
6885         int num_to_del = 1;
6886         u32 item_size;
6887         u64 refs;
6888         u64 bytenr = node->bytenr;
6889         u64 num_bytes = node->num_bytes;
6890         int last_ref = 0;
6891         bool skinny_metadata = btrfs_fs_incompat(info, SKINNY_METADATA);
6892
6893         path = btrfs_alloc_path();
6894         if (!path)
6895                 return -ENOMEM;
6896
6897         path->reada = READA_FORWARD;
6898         path->leave_spinning = 1;
6899
6900         is_data = owner_objectid >= BTRFS_FIRST_FREE_OBJECTID;
6901         BUG_ON(!is_data && refs_to_drop != 1);
6902
6903         if (is_data)
6904                 skinny_metadata = false;
6905
6906         ret = lookup_extent_backref(trans, info, path, &iref,
6907                                     bytenr, num_bytes, parent,
6908                                     root_objectid, owner_objectid,
6909                                     owner_offset);
6910         if (ret == 0) {
6911                 extent_slot = path->slots[0];
6912                 while (extent_slot >= 0) {
6913                         btrfs_item_key_to_cpu(path->nodes[0], &key,
6914                                               extent_slot);
6915                         if (key.objectid != bytenr)
6916                                 break;
6917                         if (key.type == BTRFS_EXTENT_ITEM_KEY &&
6918                             key.offset == num_bytes) {
6919                                 found_extent = 1;
6920                                 break;
6921                         }
6922                         if (key.type == BTRFS_METADATA_ITEM_KEY &&
6923                             key.offset == owner_objectid) {
6924                                 found_extent = 1;
6925                                 break;
6926                         }
6927                         if (path->slots[0] - extent_slot > 5)
6928                                 break;
6929                         extent_slot--;
6930                 }
6931 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
6932                 item_size = btrfs_item_size_nr(path->nodes[0], extent_slot);
6933                 if (found_extent && item_size < sizeof(*ei))
6934                         found_extent = 0;
6935 #endif
6936                 if (!found_extent) {
6937                         BUG_ON(iref);
6938                         ret = remove_extent_backref(trans, info, path, NULL,
6939                                                     refs_to_drop,
6940                                                     is_data, &last_ref);
6941                         if (ret) {
6942                                 btrfs_abort_transaction(trans, ret);
6943                                 goto out;
6944                         }
6945                         btrfs_release_path(path);
6946                         path->leave_spinning = 1;
6947
6948                         key.objectid = bytenr;
6949                         key.type = BTRFS_EXTENT_ITEM_KEY;
6950                         key.offset = num_bytes;
6951
6952                         if (!is_data && skinny_metadata) {
6953                                 key.type = BTRFS_METADATA_ITEM_KEY;
6954                                 key.offset = owner_objectid;
6955                         }
6956
6957                         ret = btrfs_search_slot(trans, extent_root,
6958                                                 &key, path, -1, 1);
6959                         if (ret > 0 && skinny_metadata && path->slots[0]) {
6960                                 /*
6961                                  * Couldn't find our skinny metadata item,
6962                                  * see if we have ye olde extent item.
6963                                  */
6964                                 path->slots[0]--;
6965                                 btrfs_item_key_to_cpu(path->nodes[0], &key,
6966                                                       path->slots[0]);
6967                                 if (key.objectid == bytenr &&
6968                                     key.type == BTRFS_EXTENT_ITEM_KEY &&
6969                                     key.offset == num_bytes)
6970                                         ret = 0;
6971                         }
6972
6973                         if (ret > 0 && skinny_metadata) {
6974                                 skinny_metadata = false;
6975                                 key.objectid = bytenr;
6976                                 key.type = BTRFS_EXTENT_ITEM_KEY;
6977                                 key.offset = num_bytes;
6978                                 btrfs_release_path(path);
6979                                 ret = btrfs_search_slot(trans, extent_root,
6980                                                         &key, path, -1, 1);
6981                         }
6982
6983                         if (ret) {
6984                                 btrfs_err(info,
6985                                           "umm, got %d back from search, was looking for %llu",
6986                                           ret, bytenr);
6987                                 if (ret > 0)
6988                                         btrfs_print_leaf(path->nodes[0]);
6989                         }
6990                         if (ret < 0) {
6991                                 btrfs_abort_transaction(trans, ret);
6992                                 goto out;
6993                         }
6994                         extent_slot = path->slots[0];
6995                 }
6996         } else if (WARN_ON(ret == -ENOENT)) {
6997                 btrfs_print_leaf(path->nodes[0]);
6998                 btrfs_err(info,
6999                         "unable to find ref byte nr %llu parent %llu root %llu  owner %llu offset %llu",
7000                         bytenr, parent, root_objectid, owner_objectid,
7001                         owner_offset);
7002                 btrfs_abort_transaction(trans, ret);
7003                 goto out;
7004         } else {
7005                 btrfs_abort_transaction(trans, ret);
7006                 goto out;
7007         }
7008
7009         leaf = path->nodes[0];
7010         item_size = btrfs_item_size_nr(leaf, extent_slot);
7011 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
7012         if (item_size < sizeof(*ei)) {
7013                 BUG_ON(found_extent || extent_slot != path->slots[0]);
7014                 ret = convert_extent_item_v0(trans, info, path, owner_objectid,
7015                                              0);
7016                 if (ret < 0) {
7017                         btrfs_abort_transaction(trans, ret);
7018                         goto out;
7019                 }
7020
7021                 btrfs_release_path(path);
7022                 path->leave_spinning = 1;
7023
7024                 key.objectid = bytenr;
7025                 key.type = BTRFS_EXTENT_ITEM_KEY;
7026                 key.offset = num_bytes;
7027
7028                 ret = btrfs_search_slot(trans, extent_root, &key, path,
7029                                         -1, 1);
7030                 if (ret) {
7031                         btrfs_err(info,
7032                                   "umm, got %d back from search, was looking for %llu",
7033                                 ret, bytenr);
7034                         btrfs_print_leaf(path->nodes[0]);
7035                 }
7036                 if (ret < 0) {
7037                         btrfs_abort_transaction(trans, ret);
7038                         goto out;
7039                 }
7040
7041                 extent_slot = path->slots[0];
7042                 leaf = path->nodes[0];
7043                 item_size = btrfs_item_size_nr(leaf, extent_slot);
7044         }
7045 #endif
7046         BUG_ON(item_size < sizeof(*ei));
7047         ei = btrfs_item_ptr(leaf, extent_slot,
7048                             struct btrfs_extent_item);
7049         if (owner_objectid < BTRFS_FIRST_FREE_OBJECTID &&
7050             key.type == BTRFS_EXTENT_ITEM_KEY) {
7051                 struct btrfs_tree_block_info *bi;
7052                 BUG_ON(item_size < sizeof(*ei) + sizeof(*bi));
7053                 bi = (struct btrfs_tree_block_info *)(ei + 1);
7054                 WARN_ON(owner_objectid != btrfs_tree_block_level(leaf, bi));
7055         }
7056
7057         refs = btrfs_extent_refs(leaf, ei);
7058         if (refs < refs_to_drop) {
7059                 btrfs_err(info,
7060                           "trying to drop %d refs but we only have %Lu for bytenr %Lu",
7061                           refs_to_drop, refs, bytenr);
7062                 ret = -EINVAL;
7063                 btrfs_abort_transaction(trans, ret);
7064                 goto out;
7065         }
7066         refs -= refs_to_drop;
7067
7068         if (refs > 0) {
7069                 if (extent_op)
7070                         __run_delayed_extent_op(extent_op, leaf, ei);
7071                 /*
7072                  * In the case of inline back ref, reference count will
7073                  * be updated by remove_extent_backref
7074                  */
7075                 if (iref) {
7076                         BUG_ON(!found_extent);
7077                 } else {
7078                         btrfs_set_extent_refs(leaf, ei, refs);
7079                         btrfs_mark_buffer_dirty(leaf);
7080                 }
7081                 if (found_extent) {
7082                         ret = remove_extent_backref(trans, info, path,
7083                                                     iref, refs_to_drop,
7084                                                     is_data, &last_ref);
7085                         if (ret) {
7086                                 btrfs_abort_transaction(trans, ret);
7087                                 goto out;
7088                         }
7089                 }
7090         } else {
7091                 if (found_extent) {
7092                         BUG_ON(is_data && refs_to_drop !=
7093                                extent_data_ref_count(path, iref));
7094                         if (iref) {
7095                                 BUG_ON(path->slots[0] != extent_slot);
7096                         } else {
7097                                 BUG_ON(path->slots[0] != extent_slot + 1);
7098                                 path->slots[0] = extent_slot;
7099                                 num_to_del = 2;
7100                         }
7101                 }
7102
7103                 last_ref = 1;
7104                 ret = btrfs_del_items(trans, extent_root, path, path->slots[0],
7105                                       num_to_del);
7106                 if (ret) {
7107                         btrfs_abort_transaction(trans, ret);
7108                         goto out;
7109                 }
7110                 btrfs_release_path(path);
7111
7112                 if (is_data) {
7113                         ret = btrfs_del_csums(trans, info, bytenr, num_bytes);
7114                         if (ret) {
7115                                 btrfs_abort_transaction(trans, ret);
7116                                 goto out;
7117                         }
7118                 }
7119
7120                 ret = add_to_free_space_tree(trans, info, bytenr, num_bytes);
7121                 if (ret) {
7122                         btrfs_abort_transaction(trans, ret);
7123                         goto out;
7124                 }
7125
7126                 ret = update_block_group(trans, info, bytenr, num_bytes, 0);
7127                 if (ret) {
7128                         btrfs_abort_transaction(trans, ret);
7129                         goto out;
7130                 }
7131         }
7132         btrfs_release_path(path);
7133
7134 out:
7135         btrfs_free_path(path);
7136         return ret;
7137 }
7138
7139 /*
7140  * when we free an block, it is possible (and likely) that we free the last
7141  * delayed ref for that extent as well.  This searches the delayed ref tree for
7142  * a given extent, and if there are no other delayed refs to be processed, it
7143  * removes it from the tree.
7144  */
7145 static noinline int check_ref_cleanup(struct btrfs_trans_handle *trans,
7146                                       u64 bytenr)
7147 {
7148         struct btrfs_delayed_ref_head *head;
7149         struct btrfs_delayed_ref_root *delayed_refs;
7150         int ret = 0;
7151
7152         delayed_refs = &trans->transaction->delayed_refs;
7153         spin_lock(&delayed_refs->lock);
7154         head = btrfs_find_delayed_ref_head(delayed_refs, bytenr);
7155         if (!head)
7156                 goto out_delayed_unlock;
7157
7158         spin_lock(&head->lock);
7159         if (!RB_EMPTY_ROOT(&head->ref_tree))
7160                 goto out;
7161
7162         if (head->extent_op) {
7163                 if (!head->must_insert_reserved)
7164                         goto out;
7165                 btrfs_free_delayed_extent_op(head->extent_op);
7166                 head->extent_op = NULL;
7167         }
7168
7169         /*
7170          * waiting for the lock here would deadlock.  If someone else has it
7171          * locked they are already in the process of dropping it anyway
7172          */
7173         if (!mutex_trylock(&head->mutex))
7174                 goto out;
7175
7176         /*
7177          * at this point we have a head with no other entries.  Go
7178          * ahead and process it.
7179          */
7180         rb_erase(&head->href_node, &delayed_refs->href_root);
7181         RB_CLEAR_NODE(&head->href_node);
7182         atomic_dec(&delayed_refs->num_entries);
7183
7184         /*
7185          * we don't take a ref on the node because we're removing it from the
7186          * tree, so we just steal the ref the tree was holding.
7187          */
7188         delayed_refs->num_heads--;
7189         if (head->processing == 0)
7190                 delayed_refs->num_heads_ready--;
7191         head->processing = 0;
7192         spin_unlock(&head->lock);
7193         spin_unlock(&delayed_refs->lock);
7194
7195         BUG_ON(head->extent_op);
7196         if (head->must_insert_reserved)
7197                 ret = 1;
7198
7199         mutex_unlock(&head->mutex);
7200         btrfs_put_delayed_ref_head(head);
7201         return ret;
7202 out:
7203         spin_unlock(&head->lock);
7204
7205 out_delayed_unlock:
7206         spin_unlock(&delayed_refs->lock);
7207         return 0;
7208 }
7209
7210 void btrfs_free_tree_block(struct btrfs_trans_handle *trans,
7211                            struct btrfs_root *root,
7212                            struct extent_buffer *buf,
7213                            u64 parent, int last_ref)
7214 {
7215         struct btrfs_fs_info *fs_info = root->fs_info;
7216         int pin = 1;
7217         int ret;
7218
7219         if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
7220                 int old_ref_mod, new_ref_mod;
7221
7222                 btrfs_ref_tree_mod(root, buf->start, buf->len, parent,
7223                                    root->root_key.objectid,
7224                                    btrfs_header_level(buf), 0,
7225                                    BTRFS_DROP_DELAYED_REF);
7226                 ret = btrfs_add_delayed_tree_ref(fs_info, trans, buf->start,
7227                                                  buf->len, parent,
7228                                                  root->root_key.objectid,
7229                                                  btrfs_header_level(buf),
7230                                                  BTRFS_DROP_DELAYED_REF, NULL,
7231                                                  &old_ref_mod, &new_ref_mod);
7232                 BUG_ON(ret); /* -ENOMEM */
7233                 pin = old_ref_mod >= 0 && new_ref_mod < 0;
7234         }
7235
7236         if (last_ref && btrfs_header_generation(buf) == trans->transid) {
7237                 struct btrfs_block_group_cache *cache;
7238
7239                 if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
7240                         ret = check_ref_cleanup(trans, buf->start);
7241                         if (!ret)
7242                                 goto out;
7243                 }
7244
7245                 pin = 0;
7246                 cache = btrfs_lookup_block_group(fs_info, buf->start);
7247
7248                 if (btrfs_header_flag(buf, BTRFS_HEADER_FLAG_WRITTEN)) {
7249                         pin_down_extent(fs_info, cache, buf->start,
7250                                         buf->len, 1);
7251                         btrfs_put_block_group(cache);
7252                         goto out;
7253                 }
7254
7255                 WARN_ON(test_bit(EXTENT_BUFFER_DIRTY, &buf->bflags));
7256
7257                 btrfs_add_free_space(cache, buf->start, buf->len);
7258                 btrfs_free_reserved_bytes(cache, buf->len, 0);
7259                 btrfs_put_block_group(cache);
7260                 trace_btrfs_reserved_extent_free(fs_info, buf->start, buf->len);
7261         }
7262 out:
7263         if (pin)
7264                 add_pinned_bytes(fs_info, buf->len, btrfs_header_level(buf),
7265                                  root->root_key.objectid);
7266
7267         if (last_ref) {
7268                 /*
7269                  * Deleting the buffer, clear the corrupt flag since it doesn't
7270                  * matter anymore.
7271                  */
7272                 clear_bit(EXTENT_BUFFER_CORRUPT, &buf->bflags);
7273         }
7274 }
7275
7276 /* Can return -ENOMEM */
7277 int btrfs_free_extent(struct btrfs_trans_handle *trans,
7278                       struct btrfs_root *root,
7279                       u64 bytenr, u64 num_bytes, u64 parent, u64 root_objectid,
7280                       u64 owner, u64 offset)
7281 {
7282         struct btrfs_fs_info *fs_info = root->fs_info;
7283         int old_ref_mod, new_ref_mod;
7284         int ret;
7285
7286         if (btrfs_is_testing(fs_info))
7287                 return 0;
7288
7289         if (root_objectid != BTRFS_TREE_LOG_OBJECTID)
7290                 btrfs_ref_tree_mod(root, bytenr, num_bytes, parent,
7291                                    root_objectid, owner, offset,
7292                                    BTRFS_DROP_DELAYED_REF);
7293
7294         /*
7295          * tree log blocks never actually go into the extent allocation
7296          * tree, just update pinning info and exit early.
7297          */
7298         if (root_objectid == BTRFS_TREE_LOG_OBJECTID) {
7299                 WARN_ON(owner >= BTRFS_FIRST_FREE_OBJECTID);
7300                 /* unlocks the pinned mutex */
7301                 btrfs_pin_extent(fs_info, bytenr, num_bytes, 1);
7302                 old_ref_mod = new_ref_mod = 0;
7303                 ret = 0;
7304         } else if (owner < BTRFS_FIRST_FREE_OBJECTID) {
7305                 ret = btrfs_add_delayed_tree_ref(fs_info, trans, bytenr,
7306                                                  num_bytes, parent,
7307                                                  root_objectid, (int)owner,
7308                                                  BTRFS_DROP_DELAYED_REF, NULL,
7309                                                  &old_ref_mod, &new_ref_mod);
7310         } else {
7311                 ret = btrfs_add_delayed_data_ref(fs_info, trans, bytenr,
7312                                                  num_bytes, parent,
7313                                                  root_objectid, owner, offset,
7314                                                  0, BTRFS_DROP_DELAYED_REF,
7315                                                  &old_ref_mod, &new_ref_mod);
7316         }
7317
7318         if (ret == 0 && old_ref_mod >= 0 && new_ref_mod < 0)
7319                 add_pinned_bytes(fs_info, num_bytes, owner, root_objectid);
7320
7321         return ret;
7322 }
7323
7324 /*
7325  * when we wait for progress in the block group caching, its because
7326  * our allocation attempt failed at least once.  So, we must sleep
7327  * and let some progress happen before we try again.
7328  *
7329  * This function will sleep at least once waiting for new free space to
7330  * show up, and then it will check the block group free space numbers
7331  * for our min num_bytes.  Another option is to have it go ahead
7332  * and look in the rbtree for a free extent of a given size, but this
7333  * is a good start.
7334  *
7335  * Callers of this must check if cache->cached == BTRFS_CACHE_ERROR before using
7336  * any of the information in this block group.
7337  */
7338 static noinline void
7339 wait_block_group_cache_progress(struct btrfs_block_group_cache *cache,
7340                                 u64 num_bytes)
7341 {
7342         struct btrfs_caching_control *caching_ctl;
7343
7344         caching_ctl = get_caching_control(cache);
7345         if (!caching_ctl)
7346                 return;
7347
7348         wait_event(caching_ctl->wait, block_group_cache_done(cache) ||
7349                    (cache->free_space_ctl->free_space >= num_bytes));
7350
7351         put_caching_control(caching_ctl);
7352 }
7353
7354 static noinline int
7355 wait_block_group_cache_done(struct btrfs_block_group_cache *cache)
7356 {
7357         struct btrfs_caching_control *caching_ctl;
7358         int ret = 0;
7359
7360         caching_ctl = get_caching_control(cache);
7361         if (!caching_ctl)
7362                 return (cache->cached == BTRFS_CACHE_ERROR) ? -EIO : 0;
7363
7364         wait_event(caching_ctl->wait, block_group_cache_done(cache));
7365         if (cache->cached == BTRFS_CACHE_ERROR)
7366                 ret = -EIO;
7367         put_caching_control(caching_ctl);
7368         return ret;
7369 }
7370
7371 static const char *btrfs_raid_type_names[BTRFS_NR_RAID_TYPES] = {
7372         [BTRFS_RAID_RAID10]     = "raid10",
7373         [BTRFS_RAID_RAID1]      = "raid1",
7374         [BTRFS_RAID_DUP]        = "dup",
7375         [BTRFS_RAID_RAID0]      = "raid0",
7376         [BTRFS_RAID_SINGLE]     = "single",
7377         [BTRFS_RAID_RAID5]      = "raid5",
7378         [BTRFS_RAID_RAID6]      = "raid6",
7379 };
7380
7381 static const char *get_raid_name(enum btrfs_raid_types type)
7382 {
7383         if (type >= BTRFS_NR_RAID_TYPES)
7384                 return NULL;
7385
7386         return btrfs_raid_type_names[type];
7387 }
7388
7389 enum btrfs_loop_type {
7390         LOOP_CACHING_NOWAIT = 0,
7391         LOOP_CACHING_WAIT = 1,
7392         LOOP_ALLOC_CHUNK = 2,
7393         LOOP_NO_EMPTY_SIZE = 3,
7394 };
7395
7396 static inline void
7397 btrfs_lock_block_group(struct btrfs_block_group_cache *cache,
7398                        int delalloc)
7399 {
7400         if (delalloc)
7401                 down_read(&cache->data_rwsem);
7402 }
7403
7404 static inline void
7405 btrfs_grab_block_group(struct btrfs_block_group_cache *cache,
7406                        int delalloc)
7407 {
7408         btrfs_get_block_group(cache);
7409         if (delalloc)
7410                 down_read(&cache->data_rwsem);
7411 }
7412
7413 static struct btrfs_block_group_cache *
7414 btrfs_lock_cluster(struct btrfs_block_group_cache *block_group,
7415                    struct btrfs_free_cluster *cluster,
7416                    int delalloc)
7417 {
7418         struct btrfs_block_group_cache *used_bg = NULL;
7419
7420         spin_lock(&cluster->refill_lock);
7421         while (1) {
7422                 used_bg = cluster->block_group;
7423                 if (!used_bg)
7424                         return NULL;
7425
7426                 if (used_bg == block_group)
7427                         return used_bg;
7428
7429                 btrfs_get_block_group(used_bg);
7430
7431                 if (!delalloc)
7432                         return used_bg;
7433
7434                 if (down_read_trylock(&used_bg->data_rwsem))
7435                         return used_bg;
7436
7437                 spin_unlock(&cluster->refill_lock);
7438
7439                 /* We should only have one-level nested. */
7440                 down_read_nested(&used_bg->data_rwsem, SINGLE_DEPTH_NESTING);
7441
7442                 spin_lock(&cluster->refill_lock);
7443                 if (used_bg == cluster->block_group)
7444                         return used_bg;
7445
7446                 up_read(&used_bg->data_rwsem);
7447                 btrfs_put_block_group(used_bg);
7448         }
7449 }
7450
7451 static inline void
7452 btrfs_release_block_group(struct btrfs_block_group_cache *cache,
7453                          int delalloc)
7454 {
7455         if (delalloc)
7456                 up_read(&cache->data_rwsem);
7457         btrfs_put_block_group(cache);
7458 }
7459
7460 /*
7461  * walks the btree of allocated extents and find a hole of a given size.
7462  * The key ins is changed to record the hole:
7463  * ins->objectid == start position
7464  * ins->flags = BTRFS_EXTENT_ITEM_KEY
7465  * ins->offset == the size of the hole.
7466  * Any available blocks before search_start are skipped.
7467  *
7468  * If there is no suitable free space, we will record the max size of
7469  * the free space extent currently.
7470  */
7471 static noinline int find_free_extent(struct btrfs_fs_info *fs_info,
7472                                 u64 ram_bytes, u64 num_bytes, u64 empty_size,
7473                                 u64 hint_byte, struct btrfs_key *ins,
7474                                 u64 flags, int delalloc)
7475 {
7476         int ret = 0;
7477         struct btrfs_root *root = fs_info->extent_root;
7478         struct btrfs_free_cluster *last_ptr = NULL;
7479         struct btrfs_block_group_cache *block_group = NULL;
7480         u64 search_start = 0;
7481         u64 max_extent_size = 0;
7482         u64 empty_cluster = 0;
7483         struct btrfs_space_info *space_info;
7484         int loop = 0;
7485         int index = btrfs_bg_flags_to_raid_index(flags);
7486         bool failed_cluster_refill = false;
7487         bool failed_alloc = false;
7488         bool use_cluster = true;
7489         bool have_caching_bg = false;
7490         bool orig_have_caching_bg = false;
7491         bool full_search = false;
7492
7493         WARN_ON(num_bytes < fs_info->sectorsize);
7494         ins->type = BTRFS_EXTENT_ITEM_KEY;
7495         ins->objectid = 0;
7496         ins->offset = 0;
7497
7498         trace_find_free_extent(fs_info, num_bytes, empty_size, flags);
7499
7500         space_info = __find_space_info(fs_info, flags);
7501         if (!space_info) {
7502                 btrfs_err(fs_info, "No space info for %llu", flags);
7503                 return -ENOSPC;
7504         }
7505
7506         /*
7507          * If our free space is heavily fragmented we may not be able to make
7508          * big contiguous allocations, so instead of doing the expensive search
7509          * for free space, simply return ENOSPC with our max_extent_size so we
7510          * can go ahead and search for a more manageable chunk.
7511          *
7512          * If our max_extent_size is large enough for our allocation simply
7513          * disable clustering since we will likely not be able to find enough
7514          * space to create a cluster and induce latency trying.
7515          */
7516         if (unlikely(space_info->max_extent_size)) {
7517                 spin_lock(&space_info->lock);
7518                 if (space_info->max_extent_size &&
7519                     num_bytes > space_info->max_extent_size) {
7520                         ins->offset = space_info->max_extent_size;
7521                         spin_unlock(&space_info->lock);
7522                         return -ENOSPC;
7523                 } else if (space_info->max_extent_size) {
7524                         use_cluster = false;
7525                 }
7526                 spin_unlock(&space_info->lock);
7527         }
7528
7529         last_ptr = fetch_cluster_info(fs_info, space_info, &empty_cluster);
7530         if (last_ptr) {
7531                 spin_lock(&last_ptr->lock);
7532                 if (last_ptr->block_group)
7533                         hint_byte = last_ptr->window_start;
7534                 if (last_ptr->fragmented) {
7535                         /*
7536                          * We still set window_start so we can keep track of the
7537                          * last place we found an allocation to try and save
7538                          * some time.
7539                          */
7540                         hint_byte = last_ptr->window_start;
7541                         use_cluster = false;
7542                 }
7543                 spin_unlock(&last_ptr->lock);
7544         }
7545
7546         search_start = max(search_start, first_logical_byte(fs_info, 0));
7547         search_start = max(search_start, hint_byte);
7548         if (search_start == hint_byte) {
7549                 block_group = btrfs_lookup_block_group(fs_info, search_start);
7550                 /*
7551                  * we don't want to use the block group if it doesn't match our
7552                  * allocation bits, or if its not cached.
7553                  *
7554                  * However if we are re-searching with an ideal block group
7555                  * picked out then we don't care that the block group is cached.
7556                  */
7557                 if (block_group && block_group_bits(block_group, flags) &&
7558                     block_group->cached != BTRFS_CACHE_NO) {
7559                         down_read(&space_info->groups_sem);
7560                         if (list_empty(&block_group->list) ||
7561                             block_group->ro) {
7562                                 /*
7563                                  * someone is removing this block group,
7564                                  * we can't jump into the have_block_group
7565                                  * target because our list pointers are not
7566                                  * valid
7567                                  */
7568                                 btrfs_put_block_group(block_group);
7569                                 up_read(&space_info->groups_sem);
7570                         } else {
7571                                 index = btrfs_bg_flags_to_raid_index(
7572                                                 block_group->flags);
7573                                 btrfs_lock_block_group(block_group, delalloc);
7574                                 goto have_block_group;
7575                         }
7576                 } else if (block_group) {
7577                         btrfs_put_block_group(block_group);
7578                 }
7579         }
7580 search:
7581         have_caching_bg = false;
7582         if (index == 0 || index == btrfs_bg_flags_to_raid_index(flags))
7583                 full_search = true;
7584         down_read(&space_info->groups_sem);
7585         list_for_each_entry(block_group, &space_info->block_groups[index],
7586                             list) {
7587                 u64 offset;
7588                 int cached;
7589
7590                 /* If the block group is read-only, we can skip it entirely. */
7591                 if (unlikely(block_group->ro))
7592                         continue;
7593
7594                 btrfs_grab_block_group(block_group, delalloc);
7595                 search_start = block_group->key.objectid;
7596
7597                 /*
7598                  * this can happen if we end up cycling through all the
7599                  * raid types, but we want to make sure we only allocate
7600                  * for the proper type.
7601                  */
7602                 if (!block_group_bits(block_group, flags)) {
7603                     u64 extra = BTRFS_BLOCK_GROUP_DUP |
7604                                 BTRFS_BLOCK_GROUP_RAID1 |
7605                                 BTRFS_BLOCK_GROUP_RAID5 |
7606                                 BTRFS_BLOCK_GROUP_RAID6 |
7607                                 BTRFS_BLOCK_GROUP_RAID10;
7608
7609                         /*
7610                          * if they asked for extra copies and this block group
7611                          * doesn't provide them, bail.  This does allow us to
7612                          * fill raid0 from raid1.
7613                          */
7614                         if ((flags & extra) && !(block_group->flags & extra))
7615                                 goto loop;
7616                 }
7617
7618 have_block_group:
7619                 cached = block_group_cache_done(block_group);
7620                 if (unlikely(!cached)) {
7621                         have_caching_bg = true;
7622                         ret = cache_block_group(block_group, 0);
7623                         BUG_ON(ret < 0);
7624                         ret = 0;
7625                 }
7626
7627                 if (unlikely(block_group->cached == BTRFS_CACHE_ERROR))
7628                         goto loop;
7629
7630                 /*
7631                  * Ok we want to try and use the cluster allocator, so
7632                  * lets look there
7633                  */
7634                 if (last_ptr && use_cluster) {
7635                         struct btrfs_block_group_cache *used_block_group;
7636                         unsigned long aligned_cluster;
7637                         /*
7638                          * the refill lock keeps out other
7639                          * people trying to start a new cluster
7640                          */
7641                         used_block_group = btrfs_lock_cluster(block_group,
7642                                                               last_ptr,
7643                                                               delalloc);
7644                         if (!used_block_group)
7645                                 goto refill_cluster;
7646
7647                         if (used_block_group != block_group &&
7648                             (used_block_group->ro ||
7649                              !block_group_bits(used_block_group, flags)))
7650                                 goto release_cluster;
7651
7652                         offset = btrfs_alloc_from_cluster(used_block_group,
7653                                                 last_ptr,
7654                                                 num_bytes,
7655                                                 used_block_group->key.objectid,
7656                                                 &max_extent_size);
7657                         if (offset) {
7658                                 /* we have a block, we're done */
7659                                 spin_unlock(&last_ptr->refill_lock);
7660                                 trace_btrfs_reserve_extent_cluster(fs_info,
7661                                                 used_block_group,
7662                                                 search_start, num_bytes);
7663                                 if (used_block_group != block_group) {
7664                                         btrfs_release_block_group(block_group,
7665                                                                   delalloc);
7666                                         block_group = used_block_group;
7667                                 }
7668                                 goto checks;
7669                         }
7670
7671                         WARN_ON(last_ptr->block_group != used_block_group);
7672 release_cluster:
7673                         /* If we are on LOOP_NO_EMPTY_SIZE, we can't
7674                          * set up a new clusters, so lets just skip it
7675                          * and let the allocator find whatever block
7676                          * it can find.  If we reach this point, we
7677                          * will have tried the cluster allocator
7678                          * plenty of times and not have found
7679                          * anything, so we are likely way too
7680                          * fragmented for the clustering stuff to find
7681                          * anything.
7682                          *
7683                          * However, if the cluster is taken from the
7684                          * current block group, release the cluster
7685                          * first, so that we stand a better chance of
7686                          * succeeding in the unclustered
7687                          * allocation.  */
7688                         if (loop >= LOOP_NO_EMPTY_SIZE &&
7689                             used_block_group != block_group) {
7690                                 spin_unlock(&last_ptr->refill_lock);
7691                                 btrfs_release_block_group(used_block_group,
7692                                                           delalloc);
7693                                 goto unclustered_alloc;
7694                         }
7695
7696                         /*
7697                          * this cluster didn't work out, free it and
7698                          * start over
7699                          */
7700                         btrfs_return_cluster_to_free_space(NULL, last_ptr);
7701
7702                         if (used_block_group != block_group)
7703                                 btrfs_release_block_group(used_block_group,
7704                                                           delalloc);
7705 refill_cluster:
7706                         if (loop >= LOOP_NO_EMPTY_SIZE) {
7707                                 spin_unlock(&last_ptr->refill_lock);
7708                                 goto unclustered_alloc;
7709                         }
7710
7711                         aligned_cluster = max_t(unsigned long,
7712                                                 empty_cluster + empty_size,
7713                                               block_group->full_stripe_len);
7714
7715                         /* allocate a cluster in this block group */
7716                         ret = btrfs_find_space_cluster(fs_info, block_group,
7717                                                        last_ptr, search_start,
7718                                                        num_bytes,
7719                                                        aligned_cluster);
7720                         if (ret == 0) {
7721                                 /*
7722                                  * now pull our allocation out of this
7723                                  * cluster
7724                                  */
7725                                 offset = btrfs_alloc_from_cluster(block_group,
7726                                                         last_ptr,
7727                                                         num_bytes,
7728                                                         search_start,
7729                                                         &max_extent_size);
7730                                 if (offset) {
7731                                         /* we found one, proceed */
7732                                         spin_unlock(&last_ptr->refill_lock);
7733                                         trace_btrfs_reserve_extent_cluster(fs_info,
7734                                                 block_group, search_start,
7735                                                 num_bytes);
7736                                         goto checks;
7737                                 }
7738                         } else if (!cached && loop > LOOP_CACHING_NOWAIT
7739                                    && !failed_cluster_refill) {
7740                                 spin_unlock(&last_ptr->refill_lock);
7741
7742                                 failed_cluster_refill = true;
7743                                 wait_block_group_cache_progress(block_group,
7744                                        num_bytes + empty_cluster + empty_size);
7745                                 goto have_block_group;
7746                         }
7747
7748                         /*
7749                          * at this point we either didn't find a cluster
7750                          * or we weren't able to allocate a block from our
7751                          * cluster.  Free the cluster we've been trying
7752                          * to use, and go to the next block group
7753                          */
7754                         btrfs_return_cluster_to_free_space(NULL, last_ptr);
7755                         spin_unlock(&last_ptr->refill_lock);
7756                         goto loop;
7757                 }
7758
7759 unclustered_alloc:
7760                 /*
7761                  * We are doing an unclustered alloc, set the fragmented flag so
7762                  * we don't bother trying to setup a cluster again until we get
7763                  * more space.
7764                  */
7765                 if (unlikely(last_ptr)) {
7766                         spin_lock(&last_ptr->lock);
7767                         last_ptr->fragmented = 1;
7768                         spin_unlock(&last_ptr->lock);
7769                 }
7770                 if (cached) {
7771                         struct btrfs_free_space_ctl *ctl =
7772                                 block_group->free_space_ctl;
7773
7774                         spin_lock(&ctl->tree_lock);
7775                         if (ctl->free_space <
7776                             num_bytes + empty_cluster + empty_size) {
7777                                 if (ctl->free_space > max_extent_size)
7778                                         max_extent_size = ctl->free_space;
7779                                 spin_unlock(&ctl->tree_lock);
7780                                 goto loop;
7781                         }
7782                         spin_unlock(&ctl->tree_lock);
7783                 }
7784
7785                 offset = btrfs_find_space_for_alloc(block_group, search_start,
7786                                                     num_bytes, empty_size,
7787                                                     &max_extent_size);
7788                 /*
7789                  * If we didn't find a chunk, and we haven't failed on this
7790                  * block group before, and this block group is in the middle of
7791                  * caching and we are ok with waiting, then go ahead and wait
7792                  * for progress to be made, and set failed_alloc to true.
7793                  *
7794                  * If failed_alloc is true then we've already waited on this
7795                  * block group once and should move on to the next block group.
7796                  */
7797                 if (!offset && !failed_alloc && !cached &&
7798                     loop > LOOP_CACHING_NOWAIT) {
7799                         wait_block_group_cache_progress(block_group,
7800                                                 num_bytes + empty_size);
7801                         failed_alloc = true;
7802                         goto have_block_group;
7803                 } else if (!offset) {
7804                         goto loop;
7805                 }
7806 checks:
7807                 search_start = ALIGN(offset, fs_info->stripesize);
7808
7809                 /* move on to the next group */
7810                 if (search_start + num_bytes >
7811                     block_group->key.objectid + block_group->key.offset) {
7812                         btrfs_add_free_space(block_group, offset, num_bytes);
7813                         goto loop;
7814                 }
7815
7816                 if (offset < search_start)
7817                         btrfs_add_free_space(block_group, offset,
7818                                              search_start - offset);
7819                 BUG_ON(offset > search_start);
7820
7821                 ret = btrfs_add_reserved_bytes(block_group, ram_bytes,
7822                                 num_bytes, delalloc);
7823                 if (ret == -EAGAIN) {
7824                         btrfs_add_free_space(block_group, offset, num_bytes);
7825                         goto loop;
7826                 }
7827                 btrfs_inc_block_group_reservations(block_group);
7828
7829                 /* we are all good, lets return */
7830                 ins->objectid = search_start;
7831                 ins->offset = num_bytes;
7832
7833                 trace_btrfs_reserve_extent(fs_info, block_group,
7834                                            search_start, num_bytes);
7835                 btrfs_release_block_group(block_group, delalloc);
7836                 break;
7837 loop:
7838                 failed_cluster_refill = false;
7839                 failed_alloc = false;
7840                 BUG_ON(btrfs_bg_flags_to_raid_index(block_group->flags) !=
7841                        index);
7842                 btrfs_release_block_group(block_group, delalloc);
7843                 cond_resched();
7844         }
7845         up_read(&space_info->groups_sem);
7846
7847         if ((loop == LOOP_CACHING_NOWAIT) && have_caching_bg
7848                 && !orig_have_caching_bg)
7849                 orig_have_caching_bg = true;
7850
7851         if (!ins->objectid && loop >= LOOP_CACHING_WAIT && have_caching_bg)
7852                 goto search;
7853
7854         if (!ins->objectid && ++index < BTRFS_NR_RAID_TYPES)
7855                 goto search;
7856
7857         /*
7858          * LOOP_CACHING_NOWAIT, search partially cached block groups, kicking
7859          *                      caching kthreads as we move along
7860          * LOOP_CACHING_WAIT, search everything, and wait if our bg is caching
7861          * LOOP_ALLOC_CHUNK, force a chunk allocation and try again
7862          * LOOP_NO_EMPTY_SIZE, set empty_size and empty_cluster to 0 and try
7863          *                      again
7864          */
7865         if (!ins->objectid && loop < LOOP_NO_EMPTY_SIZE) {
7866                 index = 0;
7867                 if (loop == LOOP_CACHING_NOWAIT) {
7868                         /*
7869                          * We want to skip the LOOP_CACHING_WAIT step if we
7870                          * don't have any uncached bgs and we've already done a
7871                          * full search through.
7872                          */
7873                         if (orig_have_caching_bg || !full_search)
7874                                 loop = LOOP_CACHING_WAIT;
7875                         else
7876                                 loop = LOOP_ALLOC_CHUNK;
7877                 } else {
7878                         loop++;
7879                 }
7880
7881                 if (loop == LOOP_ALLOC_CHUNK) {
7882                         struct btrfs_trans_handle *trans;
7883                         int exist = 0;
7884
7885                         trans = current->journal_info;
7886                         if (trans)
7887                                 exist = 1;
7888                         else
7889                                 trans = btrfs_join_transaction(root);
7890
7891                         if (IS_ERR(trans)) {
7892                                 ret = PTR_ERR(trans);
7893                                 goto out;
7894                         }
7895
7896                         ret = do_chunk_alloc(trans, fs_info, flags,
7897                                              CHUNK_ALLOC_FORCE);
7898
7899                         /*
7900                          * If we can't allocate a new chunk we've already looped
7901                          * through at least once, move on to the NO_EMPTY_SIZE
7902                          * case.
7903                          */
7904                         if (ret == -ENOSPC)
7905                                 loop = LOOP_NO_EMPTY_SIZE;
7906
7907                         /*
7908                          * Do not bail out on ENOSPC since we
7909                          * can do more things.
7910                          */
7911                         if (ret < 0 && ret != -ENOSPC)
7912                                 btrfs_abort_transaction(trans, ret);
7913                         else
7914                                 ret = 0;
7915                         if (!exist)
7916                                 btrfs_end_transaction(trans);
7917                         if (ret)
7918                                 goto out;
7919                 }
7920
7921                 if (loop == LOOP_NO_EMPTY_SIZE) {
7922                         /*
7923                          * Don't loop again if we already have no empty_size and
7924                          * no empty_cluster.
7925                          */
7926                         if (empty_size == 0 &&
7927                             empty_cluster == 0) {
7928                                 ret = -ENOSPC;
7929                                 goto out;
7930                         }
7931                         empty_size = 0;
7932                         empty_cluster = 0;
7933                 }
7934
7935                 goto search;
7936         } else if (!ins->objectid) {
7937                 ret = -ENOSPC;
7938         } else if (ins->objectid) {
7939                 if (!use_cluster && last_ptr) {
7940                         spin_lock(&last_ptr->lock);
7941                         last_ptr->window_start = ins->objectid;
7942                         spin_unlock(&last_ptr->lock);
7943                 }
7944                 ret = 0;
7945         }
7946 out:
7947         if (ret == -ENOSPC) {
7948                 spin_lock(&space_info->lock);
7949                 space_info->max_extent_size = max_extent_size;
7950                 spin_unlock(&space_info->lock);
7951                 ins->offset = max_extent_size;
7952         }
7953         return ret;
7954 }
7955
7956 static void dump_space_info(struct btrfs_fs_info *fs_info,
7957                             struct btrfs_space_info *info, u64 bytes,
7958                             int dump_block_groups)
7959 {
7960         struct btrfs_block_group_cache *cache;
7961         int index = 0;
7962
7963         spin_lock(&info->lock);
7964         btrfs_info(fs_info, "space_info %llu has %llu free, is %sfull",
7965                    info->flags,
7966                    info->total_bytes - btrfs_space_info_used(info, true),
7967                    info->full ? "" : "not ");
7968         btrfs_info(fs_info,
7969                 "space_info total=%llu, used=%llu, pinned=%llu, reserved=%llu, may_use=%llu, readonly=%llu",
7970                 info->total_bytes, info->bytes_used, info->bytes_pinned,
7971                 info->bytes_reserved, info->bytes_may_use,
7972                 info->bytes_readonly);
7973         spin_unlock(&info->lock);
7974
7975         if (!dump_block_groups)
7976                 return;
7977
7978         down_read(&info->groups_sem);
7979 again:
7980         list_for_each_entry(cache, &info->block_groups[index], list) {
7981                 spin_lock(&cache->lock);
7982                 btrfs_info(fs_info,
7983                         "block group %llu has %llu bytes, %llu used %llu pinned %llu reserved %s",
7984                         cache->key.objectid, cache->key.offset,
7985                         btrfs_block_group_used(&cache->item), cache->pinned,
7986                         cache->reserved, cache->ro ? "[readonly]" : "");
7987                 btrfs_dump_free_space(cache, bytes);
7988                 spin_unlock(&cache->lock);
7989         }
7990         if (++index < BTRFS_NR_RAID_TYPES)
7991                 goto again;
7992         up_read(&info->groups_sem);
7993 }
7994
7995 /*
7996  * btrfs_reserve_extent - entry point to the extent allocator. Tries to find a
7997  *                        hole that is at least as big as @num_bytes.
7998  *
7999  * @root           -    The root that will contain this extent
8000  *
8001  * @ram_bytes      -    The amount of space in ram that @num_bytes take. This
8002  *                      is used for accounting purposes. This value differs
8003  *                      from @num_bytes only in the case of compressed extents.
8004  *
8005  * @num_bytes      -    Number of bytes to allocate on-disk.
8006  *
8007  * @min_alloc_size -    Indicates the minimum amount of space that the
8008  *                      allocator should try to satisfy. In some cases
8009  *                      @num_bytes may be larger than what is required and if
8010  *                      the filesystem is fragmented then allocation fails.
8011  *                      However, the presence of @min_alloc_size gives a
8012  *                      chance to try and satisfy the smaller allocation.
8013  *
8014  * @empty_size     -    A hint that you plan on doing more COW. This is the
8015  *                      size in bytes the allocator should try to find free
8016  *                      next to the block it returns.  This is just a hint and
8017  *                      may be ignored by the allocator.
8018  *
8019  * @hint_byte      -    Hint to the allocator to start searching above the byte
8020  *                      address passed. It might be ignored.
8021  *
8022  * @ins            -    This key is modified to record the found hole. It will
8023  *                      have the following values:
8024  *                      ins->objectid == start position
8025  *                      ins->flags = BTRFS_EXTENT_ITEM_KEY
8026  *                      ins->offset == the size of the hole.
8027  *
8028  * @is_data        -    Boolean flag indicating whether an extent is
8029  *                      allocated for data (true) or metadata (false)
8030  *
8031  * @delalloc       -    Boolean flag indicating whether this allocation is for
8032  *                      delalloc or not. If 'true' data_rwsem of block groups
8033  *                      is going to be acquired.
8034  *
8035  *
8036  * Returns 0 when an allocation succeeded or < 0 when an error occurred. In
8037  * case -ENOSPC is returned then @ins->offset will contain the size of the
8038  * largest available hole the allocator managed to find.
8039  */
8040 int btrfs_reserve_extent(struct btrfs_root *root, u64 ram_bytes,
8041                          u64 num_bytes, u64 min_alloc_size,
8042                          u64 empty_size, u64 hint_byte,
8043                          struct btrfs_key *ins, int is_data, int delalloc)
8044 {
8045         struct btrfs_fs_info *fs_info = root->fs_info;
8046         bool final_tried = num_bytes == min_alloc_size;
8047         u64 flags;
8048         int ret;
8049
8050         flags = get_alloc_profile_by_root(root, is_data);
8051 again:
8052         WARN_ON(num_bytes < fs_info->sectorsize);
8053         ret = find_free_extent(fs_info, ram_bytes, num_bytes, empty_size,
8054                                hint_byte, ins, flags, delalloc);
8055         if (!ret && !is_data) {
8056                 btrfs_dec_block_group_reservations(fs_info, ins->objectid);
8057         } else if (ret == -ENOSPC) {
8058                 if (!final_tried && ins->offset) {
8059                         num_bytes = min(num_bytes >> 1, ins->offset);
8060                         num_bytes = round_down(num_bytes,
8061                                                fs_info->sectorsize);
8062                         num_bytes = max(num_bytes, min_alloc_size);
8063                         ram_bytes = num_bytes;
8064                         if (num_bytes == min_alloc_size)
8065                                 final_tried = true;
8066                         goto again;
8067                 } else if (btrfs_test_opt(fs_info, ENOSPC_DEBUG)) {
8068                         struct btrfs_space_info *sinfo;
8069
8070                         sinfo = __find_space_info(fs_info, flags);
8071                         btrfs_err(fs_info,
8072                                   "allocation failed flags %llu, wanted %llu",
8073                                   flags, num_bytes);
8074                         if (sinfo)
8075                                 dump_space_info(fs_info, sinfo, num_bytes, 1);
8076                 }
8077         }
8078
8079         return ret;
8080 }
8081
8082 static int __btrfs_free_reserved_extent(struct btrfs_fs_info *fs_info,
8083                                         u64 start, u64 len,
8084                                         int pin, int delalloc)
8085 {
8086         struct btrfs_block_group_cache *cache;
8087         int ret = 0;
8088
8089         cache = btrfs_lookup_block_group(fs_info, start);
8090         if (!cache) {
8091                 btrfs_err(fs_info, "Unable to find block group for %llu",
8092                           start);
8093                 return -ENOSPC;
8094         }
8095
8096         if (pin)
8097                 pin_down_extent(fs_info, cache, start, len, 1);
8098         else {
8099                 if (btrfs_test_opt(fs_info, DISCARD))
8100                         ret = btrfs_discard_extent(fs_info, start, len, NULL);
8101                 btrfs_add_free_space(cache, start, len);
8102                 btrfs_free_reserved_bytes(cache, len, delalloc);
8103                 trace_btrfs_reserved_extent_free(fs_info, start, len);
8104         }
8105
8106         btrfs_put_block_group(cache);
8107         return ret;
8108 }
8109
8110 int btrfs_free_reserved_extent(struct btrfs_fs_info *fs_info,
8111                                u64 start, u64 len, int delalloc)
8112 {
8113         return __btrfs_free_reserved_extent(fs_info, start, len, 0, delalloc);
8114 }
8115
8116 int btrfs_free_and_pin_reserved_extent(struct btrfs_fs_info *fs_info,
8117                                        u64 start, u64 len)
8118 {
8119         return __btrfs_free_reserved_extent(fs_info, start, len, 1, 0);
8120 }
8121
8122 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
8123                                       struct btrfs_fs_info *fs_info,
8124                                       u64 parent, u64 root_objectid,
8125                                       u64 flags, u64 owner, u64 offset,
8126                                       struct btrfs_key *ins, int ref_mod)
8127 {
8128         int ret;
8129         struct btrfs_extent_item *extent_item;
8130         struct btrfs_extent_inline_ref *iref;
8131         struct btrfs_path *path;
8132         struct extent_buffer *leaf;
8133         int type;
8134         u32 size;
8135
8136         if (parent > 0)
8137                 type = BTRFS_SHARED_DATA_REF_KEY;
8138         else
8139                 type = BTRFS_EXTENT_DATA_REF_KEY;
8140
8141         size = sizeof(*extent_item) + btrfs_extent_inline_ref_size(type);
8142
8143         path = btrfs_alloc_path();
8144         if (!path)
8145                 return -ENOMEM;
8146
8147         path->leave_spinning = 1;
8148         ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path,
8149                                       ins, size);
8150         if (ret) {
8151                 btrfs_free_path(path);
8152                 return ret;
8153         }
8154
8155         leaf = path->nodes[0];
8156         extent_item = btrfs_item_ptr(leaf, path->slots[0],
8157                                      struct btrfs_extent_item);
8158         btrfs_set_extent_refs(leaf, extent_item, ref_mod);
8159         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
8160         btrfs_set_extent_flags(leaf, extent_item,
8161                                flags | BTRFS_EXTENT_FLAG_DATA);
8162
8163         iref = (struct btrfs_extent_inline_ref *)(extent_item + 1);
8164         btrfs_set_extent_inline_ref_type(leaf, iref, type);
8165         if (parent > 0) {
8166                 struct btrfs_shared_data_ref *ref;
8167                 ref = (struct btrfs_shared_data_ref *)(iref + 1);
8168                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
8169                 btrfs_set_shared_data_ref_count(leaf, ref, ref_mod);
8170         } else {
8171                 struct btrfs_extent_data_ref *ref;
8172                 ref = (struct btrfs_extent_data_ref *)(&iref->offset);
8173                 btrfs_set_extent_data_ref_root(leaf, ref, root_objectid);
8174                 btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
8175                 btrfs_set_extent_data_ref_offset(leaf, ref, offset);
8176                 btrfs_set_extent_data_ref_count(leaf, ref, ref_mod);
8177         }
8178
8179         btrfs_mark_buffer_dirty(path->nodes[0]);
8180         btrfs_free_path(path);
8181
8182         ret = remove_from_free_space_tree(trans, fs_info, ins->objectid,
8183                                           ins->offset);
8184         if (ret)
8185                 return ret;
8186
8187         ret = update_block_group(trans, fs_info, ins->objectid, ins->offset, 1);
8188         if (ret) { /* -ENOENT, logic error */
8189                 btrfs_err(fs_info, "update block group failed for %llu %llu",
8190                         ins->objectid, ins->offset);
8191                 BUG();
8192         }
8193         trace_btrfs_reserved_extent_alloc(fs_info, ins->objectid, ins->offset);
8194         return ret;
8195 }
8196
8197 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
8198                                      struct btrfs_fs_info *fs_info,
8199                                      u64 parent, u64 root_objectid,
8200                                      u64 flags, struct btrfs_disk_key *key,
8201                                      int level, struct btrfs_key *ins)
8202 {
8203         int ret;
8204         struct btrfs_extent_item *extent_item;
8205         struct btrfs_tree_block_info *block_info;
8206         struct btrfs_extent_inline_ref *iref;
8207         struct btrfs_path *path;
8208         struct extent_buffer *leaf;
8209         u32 size = sizeof(*extent_item) + sizeof(*iref);
8210         u64 num_bytes = ins->offset;
8211         bool skinny_metadata = btrfs_fs_incompat(fs_info, SKINNY_METADATA);
8212
8213         if (!skinny_metadata)
8214                 size += sizeof(*block_info);
8215
8216         path = btrfs_alloc_path();
8217         if (!path) {
8218                 btrfs_free_and_pin_reserved_extent(fs_info, ins->objectid,
8219                                                    fs_info->nodesize);
8220                 return -ENOMEM;
8221         }
8222
8223         path->leave_spinning = 1;
8224         ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path,
8225                                       ins, size);
8226         if (ret) {
8227                 btrfs_free_path(path);
8228                 btrfs_free_and_pin_reserved_extent(fs_info, ins->objectid,
8229                                                    fs_info->nodesize);
8230                 return ret;
8231         }
8232
8233         leaf = path->nodes[0];
8234         extent_item = btrfs_item_ptr(leaf, path->slots[0],
8235                                      struct btrfs_extent_item);
8236         btrfs_set_extent_refs(leaf, extent_item, 1);
8237         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
8238         btrfs_set_extent_flags(leaf, extent_item,
8239                                flags | BTRFS_EXTENT_FLAG_TREE_BLOCK);
8240
8241         if (skinny_metadata) {
8242                 iref = (struct btrfs_extent_inline_ref *)(extent_item + 1);
8243                 num_bytes = fs_info->nodesize;
8244         } else {
8245                 block_info = (struct btrfs_tree_block_info *)(extent_item + 1);
8246                 btrfs_set_tree_block_key(leaf, block_info, key);
8247                 btrfs_set_tree_block_level(leaf, block_info, level);
8248                 iref = (struct btrfs_extent_inline_ref *)(block_info + 1);
8249         }
8250
8251         if (parent > 0) {
8252                 BUG_ON(!(flags & BTRFS_BLOCK_FLAG_FULL_BACKREF));
8253                 btrfs_set_extent_inline_ref_type(leaf, iref,
8254                                                  BTRFS_SHARED_BLOCK_REF_KEY);
8255                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
8256         } else {
8257                 btrfs_set_extent_inline_ref_type(leaf, iref,
8258                                                  BTRFS_TREE_BLOCK_REF_KEY);
8259                 btrfs_set_extent_inline_ref_offset(leaf, iref, root_objectid);
8260         }
8261
8262         btrfs_mark_buffer_dirty(leaf);
8263         btrfs_free_path(path);
8264
8265         ret = remove_from_free_space_tree(trans, fs_info, ins->objectid,
8266                                           num_bytes);
8267         if (ret)
8268                 return ret;
8269
8270         ret = update_block_group(trans, fs_info, ins->objectid,
8271                                  fs_info->nodesize, 1);
8272         if (ret) { /* -ENOENT, logic error */
8273                 btrfs_err(fs_info, "update block group failed for %llu %llu",
8274                         ins->objectid, ins->offset);
8275                 BUG();
8276         }
8277
8278         trace_btrfs_reserved_extent_alloc(fs_info, ins->objectid,
8279                                           fs_info->nodesize);
8280         return ret;
8281 }
8282
8283 int btrfs_alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
8284                                      struct btrfs_root *root, u64 owner,
8285                                      u64 offset, u64 ram_bytes,
8286                                      struct btrfs_key *ins)
8287 {
8288         struct btrfs_fs_info *fs_info = root->fs_info;
8289         int ret;
8290
8291         BUG_ON(root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID);
8292
8293         btrfs_ref_tree_mod(root, ins->objectid, ins->offset, 0,
8294                            root->root_key.objectid, owner, offset,
8295                            BTRFS_ADD_DELAYED_EXTENT);
8296
8297         ret = btrfs_add_delayed_data_ref(fs_info, trans, ins->objectid,
8298                                          ins->offset, 0,
8299                                          root->root_key.objectid, owner,
8300                                          offset, ram_bytes,
8301                                          BTRFS_ADD_DELAYED_EXTENT, NULL, NULL);
8302         return ret;
8303 }
8304
8305 /*
8306  * this is used by the tree logging recovery code.  It records that
8307  * an extent has been allocated and makes sure to clear the free
8308  * space cache bits as well
8309  */
8310 int btrfs_alloc_logged_file_extent(struct btrfs_trans_handle *trans,
8311                                    struct btrfs_fs_info *fs_info,
8312                                    u64 root_objectid, u64 owner, u64 offset,
8313                                    struct btrfs_key *ins)
8314 {
8315         int ret;
8316         struct btrfs_block_group_cache *block_group;
8317         struct btrfs_space_info *space_info;
8318
8319         /*
8320          * Mixed block groups will exclude before processing the log so we only
8321          * need to do the exclude dance if this fs isn't mixed.
8322          */
8323         if (!btrfs_fs_incompat(fs_info, MIXED_GROUPS)) {
8324                 ret = __exclude_logged_extent(fs_info, ins->objectid,
8325                                               ins->offset);
8326                 if (ret)
8327                         return ret;
8328         }
8329
8330         block_group = btrfs_lookup_block_group(fs_info, ins->objectid);
8331         if (!block_group)
8332                 return -EINVAL;
8333
8334         space_info = block_group->space_info;
8335         spin_lock(&space_info->lock);
8336         spin_lock(&block_group->lock);
8337         space_info->bytes_reserved += ins->offset;
8338         block_group->reserved += ins->offset;
8339         spin_unlock(&block_group->lock);
8340         spin_unlock(&space_info->lock);
8341
8342         ret = alloc_reserved_file_extent(trans, fs_info, 0, root_objectid,
8343                                          0, owner, offset, ins, 1);
8344         btrfs_put_block_group(block_group);
8345         return ret;
8346 }
8347
8348 static struct extent_buffer *
8349 btrfs_init_new_buffer(struct btrfs_trans_handle *trans, struct btrfs_root *root,
8350                       u64 bytenr, int level)
8351 {
8352         struct btrfs_fs_info *fs_info = root->fs_info;
8353         struct extent_buffer *buf;
8354
8355         buf = btrfs_find_create_tree_block(fs_info, bytenr);
8356         if (IS_ERR(buf))
8357                 return buf;
8358
8359         btrfs_set_header_generation(buf, trans->transid);
8360         btrfs_set_buffer_lockdep_class(root->root_key.objectid, buf, level);
8361         btrfs_tree_lock(buf);
8362         clean_tree_block(fs_info, buf);
8363         clear_bit(EXTENT_BUFFER_STALE, &buf->bflags);
8364
8365         btrfs_set_lock_blocking(buf);
8366         set_extent_buffer_uptodate(buf);
8367
8368         if (root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID) {
8369                 buf->log_index = root->log_transid % 2;
8370                 /*
8371                  * we allow two log transactions at a time, use different
8372                  * EXENT bit to differentiate dirty pages.
8373                  */
8374                 if (buf->log_index == 0)
8375                         set_extent_dirty(&root->dirty_log_pages, buf->start,
8376                                         buf->start + buf->len - 1, GFP_NOFS);
8377                 else
8378                         set_extent_new(&root->dirty_log_pages, buf->start,
8379                                         buf->start + buf->len - 1);
8380         } else {
8381                 buf->log_index = -1;
8382                 set_extent_dirty(&trans->transaction->dirty_pages, buf->start,
8383                          buf->start + buf->len - 1, GFP_NOFS);
8384         }
8385         trans->dirty = true;
8386         /* this returns a buffer locked for blocking */
8387         return buf;
8388 }
8389
8390 static struct btrfs_block_rsv *
8391 use_block_rsv(struct btrfs_trans_handle *trans,
8392               struct btrfs_root *root, u32 blocksize)
8393 {
8394         struct btrfs_fs_info *fs_info = root->fs_info;
8395         struct btrfs_block_rsv *block_rsv;
8396         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
8397         int ret;
8398         bool global_updated = false;
8399
8400         block_rsv = get_block_rsv(trans, root);
8401
8402         if (unlikely(block_rsv->size == 0))
8403                 goto try_reserve;
8404 again:
8405         ret = block_rsv_use_bytes(block_rsv, blocksize);
8406         if (!ret)
8407                 return block_rsv;
8408
8409         if (block_rsv->failfast)
8410                 return ERR_PTR(ret);
8411
8412         if (block_rsv->type == BTRFS_BLOCK_RSV_GLOBAL && !global_updated) {
8413                 global_updated = true;
8414                 update_global_block_rsv(fs_info);
8415                 goto again;
8416         }
8417
8418         if (btrfs_test_opt(fs_info, ENOSPC_DEBUG)) {
8419                 static DEFINE_RATELIMIT_STATE(_rs,
8420                                 DEFAULT_RATELIMIT_INTERVAL * 10,
8421                                 /*DEFAULT_RATELIMIT_BURST*/ 1);
8422                 if (__ratelimit(&_rs))
8423                         WARN(1, KERN_DEBUG
8424                                 "BTRFS: block rsv returned %d\n", ret);
8425         }
8426 try_reserve:
8427         ret = reserve_metadata_bytes(root, block_rsv, blocksize,
8428                                      BTRFS_RESERVE_NO_FLUSH);
8429         if (!ret)
8430                 return block_rsv;
8431         /*
8432          * If we couldn't reserve metadata bytes try and use some from
8433          * the global reserve if its space type is the same as the global
8434          * reservation.
8435          */
8436         if (block_rsv->type != BTRFS_BLOCK_RSV_GLOBAL &&
8437             block_rsv->space_info == global_rsv->space_info) {
8438                 ret = block_rsv_use_bytes(global_rsv, blocksize);
8439                 if (!ret)
8440                         return global_rsv;
8441         }
8442         return ERR_PTR(ret);
8443 }
8444
8445 static void unuse_block_rsv(struct btrfs_fs_info *fs_info,
8446                             struct btrfs_block_rsv *block_rsv, u32 blocksize)
8447 {
8448         block_rsv_add_bytes(block_rsv, blocksize, 0);
8449         block_rsv_release_bytes(fs_info, block_rsv, NULL, 0, NULL);
8450 }
8451
8452 /*
8453  * finds a free extent and does all the dirty work required for allocation
8454  * returns the tree buffer or an ERR_PTR on error.
8455  */
8456 struct extent_buffer *btrfs_alloc_tree_block(struct btrfs_trans_handle *trans,
8457                                              struct btrfs_root *root,
8458                                              u64 parent, u64 root_objectid,
8459                                              const struct btrfs_disk_key *key,
8460                                              int level, u64 hint,
8461                                              u64 empty_size)
8462 {
8463         struct btrfs_fs_info *fs_info = root->fs_info;
8464         struct btrfs_key ins;
8465         struct btrfs_block_rsv *block_rsv;
8466         struct extent_buffer *buf;
8467         struct btrfs_delayed_extent_op *extent_op;
8468         u64 flags = 0;
8469         int ret;
8470         u32 blocksize = fs_info->nodesize;
8471         bool skinny_metadata = btrfs_fs_incompat(fs_info, SKINNY_METADATA);
8472
8473 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
8474         if (btrfs_is_testing(fs_info)) {
8475                 buf = btrfs_init_new_buffer(trans, root, root->alloc_bytenr,
8476                                             level);
8477                 if (!IS_ERR(buf))
8478                         root->alloc_bytenr += blocksize;
8479                 return buf;
8480         }
8481 #endif
8482
8483         block_rsv = use_block_rsv(trans, root, blocksize);
8484         if (IS_ERR(block_rsv))
8485                 return ERR_CAST(block_rsv);
8486
8487         ret = btrfs_reserve_extent(root, blocksize, blocksize, blocksize,
8488                                    empty_size, hint, &ins, 0, 0);
8489         if (ret)
8490                 goto out_unuse;
8491
8492         buf = btrfs_init_new_buffer(trans, root, ins.objectid, level);
8493         if (IS_ERR(buf)) {
8494                 ret = PTR_ERR(buf);
8495                 goto out_free_reserved;
8496         }
8497
8498         if (root_objectid == BTRFS_TREE_RELOC_OBJECTID) {
8499                 if (parent == 0)
8500                         parent = ins.objectid;
8501                 flags |= BTRFS_BLOCK_FLAG_FULL_BACKREF;
8502         } else
8503                 BUG_ON(parent > 0);
8504
8505         if (root_objectid != BTRFS_TREE_LOG_OBJECTID) {
8506                 extent_op = btrfs_alloc_delayed_extent_op();
8507                 if (!extent_op) {
8508                         ret = -ENOMEM;
8509                         goto out_free_buf;
8510                 }
8511                 if (key)
8512                         memcpy(&extent_op->key, key, sizeof(extent_op->key));
8513                 else
8514                         memset(&extent_op->key, 0, sizeof(extent_op->key));
8515                 extent_op->flags_to_set = flags;
8516                 extent_op->update_key = skinny_metadata ? false : true;
8517                 extent_op->update_flags = true;
8518                 extent_op->is_data = false;
8519                 extent_op->level = level;
8520
8521                 btrfs_ref_tree_mod(root, ins.objectid, ins.offset, parent,
8522                                    root_objectid, level, 0,
8523                                    BTRFS_ADD_DELAYED_EXTENT);
8524                 ret = btrfs_add_delayed_tree_ref(fs_info, trans, ins.objectid,
8525                                                  ins.offset, parent,
8526                                                  root_objectid, level,
8527                                                  BTRFS_ADD_DELAYED_EXTENT,
8528                                                  extent_op, NULL, NULL);
8529                 if (ret)
8530                         goto out_free_delayed;
8531         }
8532         return buf;
8533
8534 out_free_delayed:
8535         btrfs_free_delayed_extent_op(extent_op);
8536 out_free_buf:
8537         free_extent_buffer(buf);
8538 out_free_reserved:
8539         btrfs_free_reserved_extent(fs_info, ins.objectid, ins.offset, 0);
8540 out_unuse:
8541         unuse_block_rsv(fs_info, block_rsv, blocksize);
8542         return ERR_PTR(ret);
8543 }
8544
8545 struct walk_control {
8546         u64 refs[BTRFS_MAX_LEVEL];
8547         u64 flags[BTRFS_MAX_LEVEL];
8548         struct btrfs_key update_progress;
8549         int stage;
8550         int level;
8551         int shared_level;
8552         int update_ref;
8553         int keep_locks;
8554         int reada_slot;
8555         int reada_count;
8556         int for_reloc;
8557 };
8558
8559 #define DROP_REFERENCE  1
8560 #define UPDATE_BACKREF  2
8561
8562 static noinline void reada_walk_down(struct btrfs_trans_handle *trans,
8563                                      struct btrfs_root *root,
8564                                      struct walk_control *wc,
8565                                      struct btrfs_path *path)
8566 {
8567         struct btrfs_fs_info *fs_info = root->fs_info;
8568         u64 bytenr;
8569         u64 generation;
8570         u64 refs;
8571         u64 flags;
8572         u32 nritems;
8573         struct btrfs_key key;
8574         struct extent_buffer *eb;
8575         int ret;
8576         int slot;
8577         int nread = 0;
8578
8579         if (path->slots[wc->level] < wc->reada_slot) {
8580                 wc->reada_count = wc->reada_count * 2 / 3;
8581                 wc->reada_count = max(wc->reada_count, 2);
8582         } else {
8583                 wc->reada_count = wc->reada_count * 3 / 2;
8584                 wc->reada_count = min_t(int, wc->reada_count,
8585                                         BTRFS_NODEPTRS_PER_BLOCK(fs_info));
8586         }
8587
8588         eb = path->nodes[wc->level];
8589         nritems = btrfs_header_nritems(eb);
8590
8591         for (slot = path->slots[wc->level]; slot < nritems; slot++) {
8592                 if (nread >= wc->reada_count)
8593                         break;
8594
8595                 cond_resched();
8596                 bytenr = btrfs_node_blockptr(eb, slot);
8597                 generation = btrfs_node_ptr_generation(eb, slot);
8598
8599                 if (slot == path->slots[wc->level])
8600                         goto reada;
8601
8602                 if (wc->stage == UPDATE_BACKREF &&
8603                     generation <= root->root_key.offset)
8604                         continue;
8605
8606                 /* We don't lock the tree block, it's OK to be racy here */
8607                 ret = btrfs_lookup_extent_info(trans, fs_info, bytenr,
8608                                                wc->level - 1, 1, &refs,
8609                                                &flags);
8610                 /* We don't care about errors in readahead. */
8611                 if (ret < 0)
8612                         continue;
8613                 BUG_ON(refs == 0);
8614
8615                 if (wc->stage == DROP_REFERENCE) {
8616                         if (refs == 1)
8617                                 goto reada;
8618
8619                         if (wc->level == 1 &&
8620                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
8621                                 continue;
8622                         if (!wc->update_ref ||
8623                             generation <= root->root_key.offset)
8624                                 continue;
8625                         btrfs_node_key_to_cpu(eb, &key, slot);
8626                         ret = btrfs_comp_cpu_keys(&key,
8627                                                   &wc->update_progress);
8628                         if (ret < 0)
8629                                 continue;
8630                 } else {
8631                         if (wc->level == 1 &&
8632                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
8633                                 continue;
8634                 }
8635 reada:
8636                 readahead_tree_block(fs_info, bytenr);
8637                 nread++;
8638         }
8639         wc->reada_slot = slot;
8640 }
8641
8642 /*
8643  * helper to process tree block while walking down the tree.
8644  *
8645  * when wc->stage == UPDATE_BACKREF, this function updates
8646  * back refs for pointers in the block.
8647  *
8648  * NOTE: return value 1 means we should stop walking down.
8649  */
8650 static noinline int walk_down_proc(struct btrfs_trans_handle *trans,
8651                                    struct btrfs_root *root,
8652                                    struct btrfs_path *path,
8653                                    struct walk_control *wc, int lookup_info)
8654 {
8655         struct btrfs_fs_info *fs_info = root->fs_info;
8656         int level = wc->level;
8657         struct extent_buffer *eb = path->nodes[level];
8658         u64 flag = BTRFS_BLOCK_FLAG_FULL_BACKREF;
8659         int ret;
8660
8661         if (wc->stage == UPDATE_BACKREF &&
8662             btrfs_header_owner(eb) != root->root_key.objectid)
8663                 return 1;
8664
8665         /*
8666          * when reference count of tree block is 1, it won't increase
8667          * again. once full backref flag is set, we never clear it.
8668          */
8669         if (lookup_info &&
8670             ((wc->stage == DROP_REFERENCE && wc->refs[level] != 1) ||
8671              (wc->stage == UPDATE_BACKREF && !(wc->flags[level] & flag)))) {
8672                 BUG_ON(!path->locks[level]);
8673                 ret = btrfs_lookup_extent_info(trans, fs_info,
8674                                                eb->start, level, 1,
8675                                                &wc->refs[level],
8676                                                &wc->flags[level]);
8677                 BUG_ON(ret == -ENOMEM);
8678                 if (ret)
8679                         return ret;
8680                 BUG_ON(wc->refs[level] == 0);
8681         }
8682
8683         if (wc->stage == DROP_REFERENCE) {
8684                 if (wc->refs[level] > 1)
8685                         return 1;
8686
8687                 if (path->locks[level] && !wc->keep_locks) {
8688                         btrfs_tree_unlock_rw(eb, path->locks[level]);
8689                         path->locks[level] = 0;
8690                 }
8691                 return 0;
8692         }
8693
8694         /* wc->stage == UPDATE_BACKREF */
8695         if (!(wc->flags[level] & flag)) {
8696                 BUG_ON(!path->locks[level]);
8697                 ret = btrfs_inc_ref(trans, root, eb, 1);
8698                 BUG_ON(ret); /* -ENOMEM */
8699                 ret = btrfs_dec_ref(trans, root, eb, 0);
8700                 BUG_ON(ret); /* -ENOMEM */
8701                 ret = btrfs_set_disk_extent_flags(trans, fs_info, eb->start,
8702                                                   eb->len, flag,
8703                                                   btrfs_header_level(eb), 0);
8704                 BUG_ON(ret); /* -ENOMEM */
8705                 wc->flags[level] |= flag;
8706         }
8707
8708         /*
8709          * the block is shared by multiple trees, so it's not good to
8710          * keep the tree lock
8711          */
8712         if (path->locks[level] && level > 0) {
8713                 btrfs_tree_unlock_rw(eb, path->locks[level]);
8714                 path->locks[level] = 0;
8715         }
8716         return 0;
8717 }
8718
8719 /*
8720  * helper to process tree block pointer.
8721  *
8722  * when wc->stage == DROP_REFERENCE, this function checks
8723  * reference count of the block pointed to. if the block
8724  * is shared and we need update back refs for the subtree
8725  * rooted at the block, this function changes wc->stage to
8726  * UPDATE_BACKREF. if the block is shared and there is no
8727  * need to update back, this function drops the reference
8728  * to the block.
8729  *
8730  * NOTE: return value 1 means we should stop walking down.
8731  */
8732 static noinline int do_walk_down(struct btrfs_trans_handle *trans,
8733                                  struct btrfs_root *root,
8734                                  struct btrfs_path *path,
8735                                  struct walk_control *wc, int *lookup_info)
8736 {
8737         struct btrfs_fs_info *fs_info = root->fs_info;
8738         u64 bytenr;
8739         u64 generation;
8740         u64 parent;
8741         u32 blocksize;
8742         struct btrfs_key key;
8743         struct btrfs_key first_key;
8744         struct extent_buffer *next;
8745         int level = wc->level;
8746         int reada = 0;
8747         int ret = 0;
8748         bool need_account = false;
8749
8750         generation = btrfs_node_ptr_generation(path->nodes[level],
8751                                                path->slots[level]);
8752         /*
8753          * if the lower level block was created before the snapshot
8754          * was created, we know there is no need to update back refs
8755          * for the subtree
8756          */
8757         if (wc->stage == UPDATE_BACKREF &&
8758             generation <= root->root_key.offset) {
8759                 *lookup_info = 1;
8760                 return 1;
8761         }
8762
8763         bytenr = btrfs_node_blockptr(path->nodes[level], path->slots[level]);
8764         btrfs_node_key_to_cpu(path->nodes[level], &first_key,
8765                               path->slots[level]);
8766         blocksize = fs_info->nodesize;
8767
8768         next = find_extent_buffer(fs_info, bytenr);
8769         if (!next) {
8770                 next = btrfs_find_create_tree_block(fs_info, bytenr);
8771                 if (IS_ERR(next))
8772                         return PTR_ERR(next);
8773
8774                 btrfs_set_buffer_lockdep_class(root->root_key.objectid, next,
8775                                                level - 1);
8776                 reada = 1;
8777         }
8778         btrfs_tree_lock(next);
8779         btrfs_set_lock_blocking(next);
8780
8781         ret = btrfs_lookup_extent_info(trans, fs_info, bytenr, level - 1, 1,
8782                                        &wc->refs[level - 1],
8783                                        &wc->flags[level - 1]);
8784         if (ret < 0)
8785                 goto out_unlock;
8786
8787         if (unlikely(wc->refs[level - 1] == 0)) {
8788                 btrfs_err(fs_info, "Missing references.");
8789                 ret = -EIO;
8790                 goto out_unlock;
8791         }
8792         *lookup_info = 0;
8793
8794         if (wc->stage == DROP_REFERENCE) {
8795                 if (wc->refs[level - 1] > 1) {
8796                         need_account = true;
8797                         if (level == 1 &&
8798                             (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
8799                                 goto skip;
8800
8801                         if (!wc->update_ref ||
8802                             generation <= root->root_key.offset)
8803                                 goto skip;
8804
8805                         btrfs_node_key_to_cpu(path->nodes[level], &key,
8806                                               path->slots[level]);
8807                         ret = btrfs_comp_cpu_keys(&key, &wc->update_progress);
8808                         if (ret < 0)
8809                                 goto skip;
8810
8811                         wc->stage = UPDATE_BACKREF;
8812                         wc->shared_level = level - 1;
8813                 }
8814         } else {
8815                 if (level == 1 &&
8816                     (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
8817                         goto skip;
8818         }
8819
8820         if (!btrfs_buffer_uptodate(next, generation, 0)) {
8821                 btrfs_tree_unlock(next);
8822                 free_extent_buffer(next);
8823                 next = NULL;
8824                 *lookup_info = 1;
8825         }
8826
8827         if (!next) {
8828                 if (reada && level == 1)
8829                         reada_walk_down(trans, root, wc, path);
8830                 next = read_tree_block(fs_info, bytenr, generation, level - 1,
8831                                        &first_key);
8832                 if (IS_ERR(next)) {
8833                         return PTR_ERR(next);
8834                 } else if (!extent_buffer_uptodate(next)) {
8835                         free_extent_buffer(next);
8836                         return -EIO;
8837                 }
8838                 btrfs_tree_lock(next);
8839                 btrfs_set_lock_blocking(next);
8840         }
8841
8842         level--;
8843         ASSERT(level == btrfs_header_level(next));
8844         if (level != btrfs_header_level(next)) {
8845                 btrfs_err(root->fs_info, "mismatched level");
8846                 ret = -EIO;
8847                 goto out_unlock;
8848         }
8849         path->nodes[level] = next;
8850         path->slots[level] = 0;
8851         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
8852         wc->level = level;
8853         if (wc->level == 1)
8854                 wc->reada_slot = 0;
8855         return 0;
8856 skip:
8857         wc->refs[level - 1] = 0;
8858         wc->flags[level - 1] = 0;
8859         if (wc->stage == DROP_REFERENCE) {
8860                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF) {
8861                         parent = path->nodes[level]->start;
8862                 } else {
8863                         ASSERT(root->root_key.objectid ==
8864                                btrfs_header_owner(path->nodes[level]));
8865                         if (root->root_key.objectid !=
8866                             btrfs_header_owner(path->nodes[level])) {
8867                                 btrfs_err(root->fs_info,
8868                                                 "mismatched block owner");
8869                                 ret = -EIO;
8870                                 goto out_unlock;
8871                         }
8872                         parent = 0;
8873                 }
8874
8875                 if (need_account) {
8876                         ret = btrfs_qgroup_trace_subtree(trans, root, next,
8877                                                          generation, level - 1);
8878                         if (ret) {
8879                                 btrfs_err_rl(fs_info,
8880                                              "Error %d accounting shared subtree. Quota is out of sync, rescan required.",
8881                                              ret);
8882                         }
8883                 }
8884                 ret = btrfs_free_extent(trans, root, bytenr, blocksize,
8885                                         parent, root->root_key.objectid,
8886                                         level - 1, 0);
8887                 if (ret)
8888                         goto out_unlock;
8889         }
8890
8891         *lookup_info = 1;
8892         ret = 1;
8893
8894 out_unlock:
8895         btrfs_tree_unlock(next);
8896         free_extent_buffer(next);
8897
8898         return ret;
8899 }
8900
8901 /*
8902  * helper to process tree block while walking up the tree.
8903  *
8904  * when wc->stage == DROP_REFERENCE, this function drops
8905  * reference count on the block.
8906  *
8907  * when wc->stage == UPDATE_BACKREF, this function changes
8908  * wc->stage back to DROP_REFERENCE if we changed wc->stage
8909  * to UPDATE_BACKREF previously while processing the block.
8910  *
8911  * NOTE: return value 1 means we should stop walking up.
8912  */
8913 static noinline int walk_up_proc(struct btrfs_trans_handle *trans,
8914                                  struct btrfs_root *root,
8915                                  struct btrfs_path *path,
8916                                  struct walk_control *wc)
8917 {
8918         struct btrfs_fs_info *fs_info = root->fs_info;
8919         int ret;
8920         int level = wc->level;
8921         struct extent_buffer *eb = path->nodes[level];
8922         u64 parent = 0;
8923
8924         if (wc->stage == UPDATE_BACKREF) {
8925                 BUG_ON(wc->shared_level < level);
8926                 if (level < wc->shared_level)
8927                         goto out;
8928
8929                 ret = find_next_key(path, level + 1, &wc->update_progress);
8930                 if (ret > 0)
8931                         wc->update_ref = 0;
8932
8933                 wc->stage = DROP_REFERENCE;
8934                 wc->shared_level = -1;
8935                 path->slots[level] = 0;
8936
8937                 /*
8938                  * check reference count again if the block isn't locked.
8939                  * we should start walking down the tree again if reference
8940                  * count is one.
8941                  */
8942                 if (!path->locks[level]) {
8943                         BUG_ON(level == 0);
8944                         btrfs_tree_lock(eb);
8945                         btrfs_set_lock_blocking(eb);
8946                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
8947
8948                         ret = btrfs_lookup_extent_info(trans, fs_info,
8949                                                        eb->start, level, 1,
8950                                                        &wc->refs[level],
8951                                                        &wc->flags[level]);
8952                         if (ret < 0) {
8953                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
8954                                 path->locks[level] = 0;
8955                                 return ret;
8956                         }
8957                         BUG_ON(wc->refs[level] == 0);
8958                         if (wc->refs[level] == 1) {
8959                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
8960                                 path->locks[level] = 0;
8961                                 return 1;
8962                         }
8963                 }
8964         }
8965
8966         /* wc->stage == DROP_REFERENCE */
8967         BUG_ON(wc->refs[level] > 1 && !path->locks[level]);
8968
8969         if (wc->refs[level] == 1) {
8970                 if (level == 0) {
8971                         if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
8972                                 ret = btrfs_dec_ref(trans, root, eb, 1);
8973                         else
8974                                 ret = btrfs_dec_ref(trans, root, eb, 0);
8975                         BUG_ON(ret); /* -ENOMEM */
8976                         ret = btrfs_qgroup_trace_leaf_items(trans, fs_info, eb);
8977                         if (ret) {
8978                                 btrfs_err_rl(fs_info,
8979                                              "error %d accounting leaf items. Quota is out of sync, rescan required.",
8980                                              ret);
8981                         }
8982                 }
8983                 /* make block locked assertion in clean_tree_block happy */
8984                 if (!path->locks[level] &&
8985                     btrfs_header_generation(eb) == trans->transid) {
8986                         btrfs_tree_lock(eb);
8987                         btrfs_set_lock_blocking(eb);
8988                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
8989                 }
8990                 clean_tree_block(fs_info, eb);
8991         }
8992
8993         if (eb == root->node) {
8994                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
8995                         parent = eb->start;
8996                 else
8997                         BUG_ON(root->root_key.objectid !=
8998                                btrfs_header_owner(eb));
8999         } else {
9000                 if (wc->flags[level + 1] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
9001                         parent = path->nodes[level + 1]->start;
9002                 else
9003                         BUG_ON(root->root_key.objectid !=
9004                                btrfs_header_owner(path->nodes[level + 1]));
9005         }
9006
9007         btrfs_free_tree_block(trans, root, eb, parent, wc->refs[level] == 1);
9008 out:
9009         wc->refs[level] = 0;
9010         wc->flags[level] = 0;
9011         return 0;
9012 }
9013
9014 static noinline int walk_down_tree(struct btrfs_trans_handle *trans,
9015                                    struct btrfs_root *root,
9016                                    struct btrfs_path *path,
9017                                    struct walk_control *wc)
9018 {
9019         int level = wc->level;
9020         int lookup_info = 1;
9021         int ret;
9022
9023         while (level >= 0) {
9024                 ret = walk_down_proc(trans, root, path, wc, lookup_info);
9025                 if (ret > 0)
9026                         break;
9027
9028                 if (level == 0)
9029                         break;
9030
9031                 if (path->slots[level] >=
9032                     btrfs_header_nritems(path->nodes[level]))
9033                         break;
9034
9035                 ret = do_walk_down(trans, root, path, wc, &lookup_info);
9036                 if (ret > 0) {
9037                         path->slots[level]++;
9038                         continue;
9039                 } else if (ret < 0)
9040                         return ret;
9041                 level = wc->level;
9042         }
9043         return 0;
9044 }
9045
9046 static noinline int walk_up_tree(struct btrfs_trans_handle *trans,
9047                                  struct btrfs_root *root,
9048                                  struct btrfs_path *path,
9049                                  struct walk_control *wc, int max_level)
9050 {
9051         int level = wc->level;
9052         int ret;
9053
9054         path->slots[level] = btrfs_header_nritems(path->nodes[level]);
9055         while (level < max_level && path->nodes[level]) {
9056                 wc->level = level;
9057                 if (path->slots[level] + 1 <
9058                     btrfs_header_nritems(path->nodes[level])) {
9059                         path->slots[level]++;
9060                         return 0;
9061                 } else {
9062                         ret = walk_up_proc(trans, root, path, wc);
9063                         if (ret > 0)
9064                                 return 0;
9065
9066                         if (path->locks[level]) {
9067                                 btrfs_tree_unlock_rw(path->nodes[level],
9068                                                      path->locks[level]);
9069                                 path->locks[level] = 0;
9070                         }
9071                         free_extent_buffer(path->nodes[level]);
9072                         path->nodes[level] = NULL;
9073                         level++;
9074                 }
9075         }
9076         return 1;
9077 }
9078
9079 /*
9080  * drop a subvolume tree.
9081  *
9082  * this function traverses the tree freeing any blocks that only
9083  * referenced by the tree.
9084  *
9085  * when a shared tree block is found. this function decreases its
9086  * reference count by one. if update_ref is true, this function
9087  * also make sure backrefs for the shared block and all lower level
9088  * blocks are properly updated.
9089  *
9090  * If called with for_reloc == 0, may exit early with -EAGAIN
9091  */
9092 int btrfs_drop_snapshot(struct btrfs_root *root,
9093                          struct btrfs_block_rsv *block_rsv, int update_ref,
9094                          int for_reloc)
9095 {
9096         struct btrfs_fs_info *fs_info = root->fs_info;
9097         struct btrfs_path *path;
9098         struct btrfs_trans_handle *trans;
9099         struct btrfs_root *tree_root = fs_info->tree_root;
9100         struct btrfs_root_item *root_item = &root->root_item;
9101         struct walk_control *wc;
9102         struct btrfs_key key;
9103         int err = 0;
9104         int ret;
9105         int level;
9106         bool root_dropped = false;
9107
9108         btrfs_debug(fs_info, "Drop subvolume %llu", root->objectid);
9109
9110         path = btrfs_alloc_path();
9111         if (!path) {
9112                 err = -ENOMEM;
9113                 goto out;
9114         }
9115
9116         wc = kzalloc(sizeof(*wc), GFP_NOFS);
9117         if (!wc) {
9118                 btrfs_free_path(path);
9119                 err = -ENOMEM;
9120                 goto out;
9121         }
9122
9123         trans = btrfs_start_transaction(tree_root, 0);
9124         if (IS_ERR(trans)) {
9125                 err = PTR_ERR(trans);
9126                 goto out_free;
9127         }
9128
9129         if (block_rsv)
9130                 trans->block_rsv = block_rsv;
9131
9132         if (btrfs_disk_key_objectid(&root_item->drop_progress) == 0) {
9133                 level = btrfs_header_level(root->node);
9134                 path->nodes[level] = btrfs_lock_root_node(root);
9135                 btrfs_set_lock_blocking(path->nodes[level]);
9136                 path->slots[level] = 0;
9137                 path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
9138                 memset(&wc->update_progress, 0,
9139                        sizeof(wc->update_progress));
9140         } else {
9141                 btrfs_disk_key_to_cpu(&key, &root_item->drop_progress);
9142                 memcpy(&wc->update_progress, &key,
9143                        sizeof(wc->update_progress));
9144
9145                 level = root_item->drop_level;
9146                 BUG_ON(level == 0);
9147                 path->lowest_level = level;
9148                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
9149                 path->lowest_level = 0;
9150                 if (ret < 0) {
9151                         err = ret;
9152                         goto out_end_trans;
9153                 }
9154                 WARN_ON(ret > 0);
9155
9156                 /*
9157                  * unlock our path, this is safe because only this
9158                  * function is allowed to delete this snapshot
9159                  */
9160                 btrfs_unlock_up_safe(path, 0);
9161
9162                 level = btrfs_header_level(root->node);
9163                 while (1) {
9164                         btrfs_tree_lock(path->nodes[level]);
9165                         btrfs_set_lock_blocking(path->nodes[level]);
9166                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
9167
9168                         ret = btrfs_lookup_extent_info(trans, fs_info,
9169                                                 path->nodes[level]->start,
9170                                                 level, 1, &wc->refs[level],
9171                                                 &wc->flags[level]);
9172                         if (ret < 0) {
9173                                 err = ret;
9174                                 goto out_end_trans;
9175                         }
9176                         BUG_ON(wc->refs[level] == 0);
9177
9178                         if (level == root_item->drop_level)
9179                                 break;
9180
9181                         btrfs_tree_unlock(path->nodes[level]);
9182                         path->locks[level] = 0;
9183                         WARN_ON(wc->refs[level] != 1);
9184                         level--;
9185                 }
9186         }
9187
9188         wc->level = level;
9189         wc->shared_level = -1;
9190         wc->stage = DROP_REFERENCE;
9191         wc->update_ref = update_ref;
9192         wc->keep_locks = 0;
9193         wc->for_reloc = for_reloc;
9194         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(fs_info);
9195
9196         while (1) {
9197
9198                 ret = walk_down_tree(trans, root, path, wc);
9199                 if (ret < 0) {
9200                         err = ret;
9201                         break;
9202                 }
9203
9204                 ret = walk_up_tree(trans, root, path, wc, BTRFS_MAX_LEVEL);
9205                 if (ret < 0) {
9206                         err = ret;
9207                         break;
9208                 }
9209
9210                 if (ret > 0) {
9211                         BUG_ON(wc->stage != DROP_REFERENCE);
9212                         break;
9213                 }
9214
9215                 if (wc->stage == DROP_REFERENCE) {
9216                         level = wc->level;
9217                         btrfs_node_key(path->nodes[level],
9218                                        &root_item->drop_progress,
9219                                        path->slots[level]);
9220                         root_item->drop_level = level;
9221                 }
9222
9223                 BUG_ON(wc->level == 0);
9224                 if (btrfs_should_end_transaction(trans) ||
9225                     (!for_reloc && btrfs_need_cleaner_sleep(fs_info))) {
9226                         ret = btrfs_update_root(trans, tree_root,
9227                                                 &root->root_key,
9228                                                 root_item);
9229                         if (ret) {
9230                                 btrfs_abort_transaction(trans, ret);
9231                                 err = ret;
9232                                 goto out_end_trans;
9233                         }
9234
9235                         btrfs_end_transaction_throttle(trans);
9236                         if (!for_reloc && btrfs_need_cleaner_sleep(fs_info)) {
9237                                 btrfs_debug(fs_info,
9238                                             "drop snapshot early exit");
9239                                 err = -EAGAIN;
9240                                 goto out_free;
9241                         }
9242
9243                         trans = btrfs_start_transaction(tree_root, 0);
9244                         if (IS_ERR(trans)) {
9245                                 err = PTR_ERR(trans);
9246                                 goto out_free;
9247                         }
9248                         if (block_rsv)
9249                                 trans->block_rsv = block_rsv;
9250                 }
9251         }
9252         btrfs_release_path(path);
9253         if (err)
9254                 goto out_end_trans;
9255
9256         ret = btrfs_del_root(trans, fs_info, &root->root_key);
9257         if (ret) {
9258                 btrfs_abort_transaction(trans, ret);
9259                 err = ret;
9260                 goto out_end_trans;
9261         }
9262
9263         if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) {
9264                 ret = btrfs_find_root(tree_root, &root->root_key, path,
9265                                       NULL, NULL);
9266                 if (ret < 0) {
9267                         btrfs_abort_transaction(trans, ret);
9268                         err = ret;
9269                         goto out_end_trans;
9270                 } else if (ret > 0) {
9271                         /* if we fail to delete the orphan item this time
9272                          * around, it'll get picked up the next time.
9273                          *
9274                          * The most common failure here is just -ENOENT.
9275                          */
9276                         btrfs_del_orphan_item(trans, tree_root,
9277                                               root->root_key.objectid);
9278                 }
9279         }
9280
9281         if (test_bit(BTRFS_ROOT_IN_RADIX, &root->state)) {
9282                 btrfs_add_dropped_root(trans, root);
9283         } else {
9284                 free_extent_buffer(root->node);
9285                 free_extent_buffer(root->commit_root);
9286                 btrfs_put_fs_root(root);
9287         }
9288         root_dropped = true;
9289 out_end_trans:
9290         btrfs_end_transaction_throttle(trans);
9291 out_free:
9292         kfree(wc);
9293         btrfs_free_path(path);
9294 out:
9295         /*
9296          * So if we need to stop dropping the snapshot for whatever reason we
9297          * need to make sure to add it back to the dead root list so that we
9298          * keep trying to do the work later.  This also cleans up roots if we
9299          * don't have it in the radix (like when we recover after a power fail
9300          * or unmount) so we don't leak memory.
9301          */
9302         if (!for_reloc && !root_dropped)
9303                 btrfs_add_dead_root(root);
9304         if (err && err != -EAGAIN)
9305                 btrfs_handle_fs_error(fs_info, err, NULL);
9306         return err;
9307 }
9308
9309 /*
9310  * drop subtree rooted at tree block 'node'.
9311  *
9312  * NOTE: this function will unlock and release tree block 'node'
9313  * only used by relocation code
9314  */
9315 int btrfs_drop_subtree(struct btrfs_trans_handle *trans,
9316                         struct btrfs_root *root,
9317                         struct extent_buffer *node,
9318                         struct extent_buffer *parent)
9319 {
9320         struct btrfs_fs_info *fs_info = root->fs_info;
9321         struct btrfs_path *path;
9322         struct walk_control *wc;
9323         int level;
9324         int parent_level;
9325         int ret = 0;
9326         int wret;
9327
9328         BUG_ON(root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID);
9329
9330         path = btrfs_alloc_path();
9331         if (!path)
9332                 return -ENOMEM;
9333
9334         wc = kzalloc(sizeof(*wc), GFP_NOFS);
9335         if (!wc) {
9336                 btrfs_free_path(path);
9337                 return -ENOMEM;
9338         }
9339
9340         btrfs_assert_tree_locked(parent);
9341         parent_level = btrfs_header_level(parent);
9342         extent_buffer_get(parent);
9343         path->nodes[parent_level] = parent;
9344         path->slots[parent_level] = btrfs_header_nritems(parent);
9345
9346         btrfs_assert_tree_locked(node);
9347         level = btrfs_header_level(node);
9348         path->nodes[level] = node;
9349         path->slots[level] = 0;
9350         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
9351
9352         wc->refs[parent_level] = 1;
9353         wc->flags[parent_level] = BTRFS_BLOCK_FLAG_FULL_BACKREF;
9354         wc->level = level;
9355         wc->shared_level = -1;
9356         wc->stage = DROP_REFERENCE;
9357         wc->update_ref = 0;
9358         wc->keep_locks = 1;
9359         wc->for_reloc = 1;
9360         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(fs_info);
9361
9362         while (1) {
9363                 wret = walk_down_tree(trans, root, path, wc);
9364                 if (wret < 0) {
9365                         ret = wret;
9366                         break;
9367                 }
9368
9369                 wret = walk_up_tree(trans, root, path, wc, parent_level);
9370                 if (wret < 0)
9371                         ret = wret;
9372                 if (wret != 0)
9373                         break;
9374         }
9375
9376         kfree(wc);
9377         btrfs_free_path(path);
9378         return ret;
9379 }
9380
9381 static u64 update_block_group_flags(struct btrfs_fs_info *fs_info, u64 flags)
9382 {
9383         u64 num_devices;
9384         u64 stripped;
9385
9386         /*
9387          * if restripe for this chunk_type is on pick target profile and
9388          * return, otherwise do the usual balance
9389          */
9390         stripped = get_restripe_target(fs_info, flags);
9391         if (stripped)
9392                 return extended_to_chunk(stripped);
9393
9394         num_devices = fs_info->fs_devices->rw_devices;
9395
9396         stripped = BTRFS_BLOCK_GROUP_RAID0 |
9397                 BTRFS_BLOCK_GROUP_RAID5 | BTRFS_BLOCK_GROUP_RAID6 |
9398                 BTRFS_BLOCK_GROUP_RAID1 | BTRFS_BLOCK_GROUP_RAID10;
9399
9400         if (num_devices == 1) {
9401                 stripped |= BTRFS_BLOCK_GROUP_DUP;
9402                 stripped = flags & ~stripped;
9403
9404                 /* turn raid0 into single device chunks */
9405                 if (flags & BTRFS_BLOCK_GROUP_RAID0)
9406                         return stripped;
9407
9408                 /* turn mirroring into duplication */
9409                 if (flags & (BTRFS_BLOCK_GROUP_RAID1 |
9410                              BTRFS_BLOCK_GROUP_RAID10))
9411                         return stripped | BTRFS_BLOCK_GROUP_DUP;
9412         } else {
9413                 /* they already had raid on here, just return */
9414                 if (flags & stripped)
9415                         return flags;
9416
9417                 stripped |= BTRFS_BLOCK_GROUP_DUP;
9418                 stripped = flags & ~stripped;
9419
9420                 /* switch duplicated blocks with raid1 */
9421                 if (flags & BTRFS_BLOCK_GROUP_DUP)
9422                         return stripped | BTRFS_BLOCK_GROUP_RAID1;
9423
9424                 /* this is drive concat, leave it alone */
9425         }
9426
9427         return flags;
9428 }
9429
9430 static int inc_block_group_ro(struct btrfs_block_group_cache *cache, int force)
9431 {
9432         struct btrfs_space_info *sinfo = cache->space_info;
9433         u64 num_bytes;
9434         u64 min_allocable_bytes;
9435         int ret = -ENOSPC;
9436
9437         /*
9438          * We need some metadata space and system metadata space for
9439          * allocating chunks in some corner cases until we force to set
9440          * it to be readonly.
9441          */
9442         if ((sinfo->flags &
9443              (BTRFS_BLOCK_GROUP_SYSTEM | BTRFS_BLOCK_GROUP_METADATA)) &&
9444             !force)
9445                 min_allocable_bytes = SZ_1M;
9446         else
9447                 min_allocable_bytes = 0;
9448
9449         spin_lock(&sinfo->lock);
9450         spin_lock(&cache->lock);
9451
9452         if (cache->ro) {
9453                 cache->ro++;
9454                 ret = 0;
9455                 goto out;
9456         }
9457
9458         num_bytes = cache->key.offset - cache->reserved - cache->pinned -
9459                     cache->bytes_super - btrfs_block_group_used(&cache->item);
9460
9461         if (btrfs_space_info_used(sinfo, true) + num_bytes +
9462             min_allocable_bytes <= sinfo->total_bytes) {
9463                 sinfo->bytes_readonly += num_bytes;
9464                 cache->ro++;
9465                 list_add_tail(&cache->ro_list, &sinfo->ro_bgs);
9466                 ret = 0;
9467         }
9468 out:
9469         spin_unlock(&cache->lock);
9470         spin_unlock(&sinfo->lock);
9471         return ret;
9472 }
9473
9474 int btrfs_inc_block_group_ro(struct btrfs_fs_info *fs_info,
9475                              struct btrfs_block_group_cache *cache)
9476
9477 {
9478         struct btrfs_trans_handle *trans;
9479         u64 alloc_flags;
9480         int ret;
9481
9482 again:
9483         trans = btrfs_join_transaction(fs_info->extent_root);
9484         if (IS_ERR(trans))
9485                 return PTR_ERR(trans);
9486
9487         /*
9488          * we're not allowed to set block groups readonly after the dirty
9489          * block groups cache has started writing.  If it already started,
9490          * back off and let this transaction commit
9491          */
9492         mutex_lock(&fs_info->ro_block_group_mutex);
9493         if (test_bit(BTRFS_TRANS_DIRTY_BG_RUN, &trans->transaction->flags)) {
9494                 u64 transid = trans->transid;
9495
9496                 mutex_unlock(&fs_info->ro_block_group_mutex);
9497                 btrfs_end_transaction(trans);
9498
9499                 ret = btrfs_wait_for_commit(fs_info, transid);
9500                 if (ret)
9501                         return ret;
9502                 goto again;
9503         }
9504
9505         /*
9506          * if we are changing raid levels, try to allocate a corresponding
9507          * block group with the new raid level.
9508          */
9509         alloc_flags = update_block_group_flags(fs_info, cache->flags);
9510         if (alloc_flags != cache->flags) {
9511                 ret = do_chunk_alloc(trans, fs_info, alloc_flags,
9512                                      CHUNK_ALLOC_FORCE);
9513                 /*
9514                  * ENOSPC is allowed here, we may have enough space
9515                  * already allocated at the new raid level to
9516                  * carry on
9517                  */
9518                 if (ret == -ENOSPC)
9519                         ret = 0;
9520                 if (ret < 0)
9521                         goto out;
9522         }
9523
9524         ret = inc_block_group_ro(cache, 0);
9525         if (!ret)
9526                 goto out;
9527         alloc_flags = get_alloc_profile(fs_info, cache->space_info->flags);
9528         ret = do_chunk_alloc(trans, fs_info, alloc_flags,
9529                              CHUNK_ALLOC_FORCE);
9530         if (ret < 0)
9531                 goto out;
9532         ret = inc_block_group_ro(cache, 0);
9533 out:
9534         if (cache->flags & BTRFS_BLOCK_GROUP_SYSTEM) {
9535                 alloc_flags = update_block_group_flags(fs_info, cache->flags);
9536                 mutex_lock(&fs_info->chunk_mutex);
9537                 check_system_chunk(trans, fs_info, alloc_flags);
9538                 mutex_unlock(&fs_info->chunk_mutex);
9539         }
9540         mutex_unlock(&fs_info->ro_block_group_mutex);
9541
9542         btrfs_end_transaction(trans);
9543         return ret;
9544 }
9545
9546 int btrfs_force_chunk_alloc(struct btrfs_trans_handle *trans,
9547                             struct btrfs_fs_info *fs_info, u64 type)
9548 {
9549         u64 alloc_flags = get_alloc_profile(fs_info, type);
9550
9551         return do_chunk_alloc(trans, fs_info, alloc_flags, CHUNK_ALLOC_FORCE);
9552 }
9553
9554 /*
9555  * helper to account the unused space of all the readonly block group in the
9556  * space_info. takes mirrors into account.
9557  */
9558 u64 btrfs_account_ro_block_groups_free_space(struct btrfs_space_info *sinfo)
9559 {
9560         struct btrfs_block_group_cache *block_group;
9561         u64 free_bytes = 0;
9562         int factor;
9563
9564         /* It's df, we don't care if it's racy */
9565         if (list_empty(&sinfo->ro_bgs))
9566                 return 0;
9567
9568         spin_lock(&sinfo->lock);
9569         list_for_each_entry(block_group, &sinfo->ro_bgs, ro_list) {
9570                 spin_lock(&block_group->lock);
9571
9572                 if (!block_group->ro) {
9573                         spin_unlock(&block_group->lock);
9574                         continue;
9575                 }
9576
9577                 if (block_group->flags & (BTRFS_BLOCK_GROUP_RAID1 |
9578                                           BTRFS_BLOCK_GROUP_RAID10 |
9579                                           BTRFS_BLOCK_GROUP_DUP))
9580                         factor = 2;
9581                 else
9582                         factor = 1;
9583
9584                 free_bytes += (block_group->key.offset -
9585                                btrfs_block_group_used(&block_group->item)) *
9586                                factor;
9587
9588                 spin_unlock(&block_group->lock);
9589         }
9590         spin_unlock(&sinfo->lock);
9591
9592         return free_bytes;
9593 }
9594
9595 void btrfs_dec_block_group_ro(struct btrfs_block_group_cache *cache)
9596 {
9597         struct btrfs_space_info *sinfo = cache->space_info;
9598         u64 num_bytes;
9599
9600         BUG_ON(!cache->ro);
9601
9602         spin_lock(&sinfo->lock);
9603         spin_lock(&cache->lock);
9604         if (!--cache->ro) {
9605                 num_bytes = cache->key.offset - cache->reserved -
9606                             cache->pinned - cache->bytes_super -
9607                             btrfs_block_group_used(&cache->item);
9608                 sinfo->bytes_readonly -= num_bytes;
9609                 list_del_init(&cache->ro_list);
9610         }
9611         spin_unlock(&cache->lock);
9612         spin_unlock(&sinfo->lock);
9613 }
9614
9615 /*
9616  * checks to see if its even possible to relocate this block group.
9617  *
9618  * @return - -1 if it's not a good idea to relocate this block group, 0 if its
9619  * ok to go ahead and try.
9620  */
9621 int btrfs_can_relocate(struct btrfs_fs_info *fs_info, u64 bytenr)
9622 {
9623         struct btrfs_root *root = fs_info->extent_root;
9624         struct btrfs_block_group_cache *block_group;
9625         struct btrfs_space_info *space_info;
9626         struct btrfs_fs_devices *fs_devices = fs_info->fs_devices;
9627         struct btrfs_device *device;
9628         struct btrfs_trans_handle *trans;
9629         u64 min_free;
9630         u64 dev_min = 1;
9631         u64 dev_nr = 0;
9632         u64 target;
9633         int debug;
9634         int index;
9635         int full = 0;
9636         int ret = 0;
9637
9638         debug = btrfs_test_opt(fs_info, ENOSPC_DEBUG);
9639
9640         block_group = btrfs_lookup_block_group(fs_info, bytenr);
9641
9642         /* odd, couldn't find the block group, leave it alone */
9643         if (!block_group) {
9644                 if (debug)
9645                         btrfs_warn(fs_info,
9646                                    "can't find block group for bytenr %llu",
9647                                    bytenr);
9648                 return -1;
9649         }
9650
9651         min_free = btrfs_block_group_used(&block_group->item);
9652
9653         /* no bytes used, we're good */
9654         if (!min_free)
9655                 goto out;
9656
9657         space_info = block_group->space_info;
9658         spin_lock(&space_info->lock);
9659
9660         full = space_info->full;
9661
9662         /*
9663          * if this is the last block group we have in this space, we can't
9664          * relocate it unless we're able to allocate a new chunk below.
9665          *
9666          * Otherwise, we need to make sure we have room in the space to handle
9667          * all of the extents from this block group.  If we can, we're good
9668          */
9669         if ((space_info->total_bytes != block_group->key.offset) &&
9670             (btrfs_space_info_used(space_info, false) + min_free <
9671              space_info->total_bytes)) {
9672                 spin_unlock(&space_info->lock);
9673                 goto out;
9674         }
9675         spin_unlock(&space_info->lock);
9676
9677         /*
9678          * ok we don't have enough space, but maybe we have free space on our
9679          * devices to allocate new chunks for relocation, so loop through our
9680          * alloc devices and guess if we have enough space.  if this block
9681          * group is going to be restriped, run checks against the target
9682          * profile instead of the current one.
9683          */
9684         ret = -1;
9685
9686         /*
9687          * index:
9688          *      0: raid10
9689          *      1: raid1
9690          *      2: dup
9691          *      3: raid0
9692          *      4: single
9693          */
9694         target = get_restripe_target(fs_info, block_group->flags);
9695         if (target) {
9696                 index = btrfs_bg_flags_to_raid_index(extended_to_chunk(target));
9697         } else {
9698                 /*
9699                  * this is just a balance, so if we were marked as full
9700                  * we know there is no space for a new chunk
9701                  */
9702                 if (full) {
9703                         if (debug)
9704                                 btrfs_warn(fs_info,
9705                                            "no space to alloc new chunk for block group %llu",
9706                                            block_group->key.objectid);
9707                         goto out;
9708                 }
9709
9710                 index = btrfs_bg_flags_to_raid_index(block_group->flags);
9711         }
9712
9713         if (index == BTRFS_RAID_RAID10) {
9714                 dev_min = 4;
9715                 /* Divide by 2 */
9716                 min_free >>= 1;
9717         } else if (index == BTRFS_RAID_RAID1) {
9718                 dev_min = 2;
9719         } else if (index == BTRFS_RAID_DUP) {
9720                 /* Multiply by 2 */
9721                 min_free <<= 1;
9722         } else if (index == BTRFS_RAID_RAID0) {
9723                 dev_min = fs_devices->rw_devices;
9724                 min_free = div64_u64(min_free, dev_min);
9725         }
9726
9727         /* We need to do this so that we can look at pending chunks */
9728         trans = btrfs_join_transaction(root);
9729         if (IS_ERR(trans)) {
9730                 ret = PTR_ERR(trans);
9731                 goto out;
9732         }
9733
9734         mutex_lock(&fs_info->chunk_mutex);
9735         list_for_each_entry(device, &fs_devices->alloc_list, dev_alloc_list) {
9736                 u64 dev_offset;
9737
9738                 /*
9739                  * check to make sure we can actually find a chunk with enough
9740                  * space to fit our block group in.
9741                  */
9742                 if (device->total_bytes > device->bytes_used + min_free &&
9743                     !test_bit(BTRFS_DEV_STATE_REPLACE_TGT, &device->dev_state)) {
9744                         ret = find_free_dev_extent(trans, device, min_free,
9745                                                    &dev_offset, NULL);
9746                         if (!ret)
9747                                 dev_nr++;
9748
9749                         if (dev_nr >= dev_min)
9750                                 break;
9751
9752                         ret = -1;
9753                 }
9754         }
9755         if (debug && ret == -1)
9756                 btrfs_warn(fs_info,
9757                            "no space to allocate a new chunk for block group %llu",
9758                            block_group->key.objectid);
9759         mutex_unlock(&fs_info->chunk_mutex);
9760         btrfs_end_transaction(trans);
9761 out:
9762         btrfs_put_block_group(block_group);
9763         return ret;
9764 }
9765
9766 static int find_first_block_group(struct btrfs_fs_info *fs_info,
9767                                   struct btrfs_path *path,
9768                                   struct btrfs_key *key)
9769 {
9770         struct btrfs_root *root = fs_info->extent_root;
9771         int ret = 0;
9772         struct btrfs_key found_key;
9773         struct extent_buffer *leaf;
9774         int slot;
9775
9776         ret = btrfs_search_slot(NULL, root, key, path, 0, 0);
9777         if (ret < 0)
9778                 goto out;
9779
9780         while (1) {
9781                 slot = path->slots[0];
9782                 leaf = path->nodes[0];
9783                 if (slot >= btrfs_header_nritems(leaf)) {
9784                         ret = btrfs_next_leaf(root, path);
9785                         if (ret == 0)
9786                                 continue;
9787                         if (ret < 0)
9788                                 goto out;
9789                         break;
9790                 }
9791                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
9792
9793                 if (found_key.objectid >= key->objectid &&
9794                     found_key.type == BTRFS_BLOCK_GROUP_ITEM_KEY) {
9795                         struct extent_map_tree *em_tree;
9796                         struct extent_map *em;
9797
9798                         em_tree = &root->fs_info->mapping_tree.map_tree;
9799                         read_lock(&em_tree->lock);
9800                         em = lookup_extent_mapping(em_tree, found_key.objectid,
9801                                                    found_key.offset);
9802                         read_unlock(&em_tree->lock);
9803                         if (!em) {
9804                                 btrfs_err(fs_info,
9805                         "logical %llu len %llu found bg but no related chunk",
9806                                           found_key.objectid, found_key.offset);
9807                                 ret = -ENOENT;
9808                         } else {
9809                                 ret = 0;
9810                         }
9811                         free_extent_map(em);
9812                         goto out;
9813                 }
9814                 path->slots[0]++;
9815         }
9816 out:
9817         return ret;
9818 }
9819
9820 void btrfs_put_block_group_cache(struct btrfs_fs_info *info)
9821 {
9822         struct btrfs_block_group_cache *block_group;
9823         u64 last = 0;
9824
9825         while (1) {
9826                 struct inode *inode;
9827
9828                 block_group = btrfs_lookup_first_block_group(info, last);
9829                 while (block_group) {
9830                         spin_lock(&block_group->lock);
9831                         if (block_group->iref)
9832                                 break;
9833                         spin_unlock(&block_group->lock);
9834                         block_group = next_block_group(info, block_group);
9835                 }
9836                 if (!block_group) {
9837                         if (last == 0)
9838                                 break;
9839                         last = 0;
9840                         continue;
9841                 }
9842
9843                 inode = block_group->inode;
9844                 block_group->iref = 0;
9845                 block_group->inode = NULL;
9846                 spin_unlock(&block_group->lock);
9847                 ASSERT(block_group->io_ctl.inode == NULL);
9848                 iput(inode);
9849                 last = block_group->key.objectid + block_group->key.offset;
9850                 btrfs_put_block_group(block_group);
9851         }
9852 }
9853
9854 /*
9855  * Must be called only after stopping all workers, since we could have block
9856  * group caching kthreads running, and therefore they could race with us if we
9857  * freed the block groups before stopping them.
9858  */
9859 int btrfs_free_block_groups(struct btrfs_fs_info *info)
9860 {
9861         struct btrfs_block_group_cache *block_group;
9862         struct btrfs_space_info *space_info;
9863         struct btrfs_caching_control *caching_ctl;
9864         struct rb_node *n;
9865
9866         down_write(&info->commit_root_sem);
9867         while (!list_empty(&info->caching_block_groups)) {
9868                 caching_ctl = list_entry(info->caching_block_groups.next,
9869                                          struct btrfs_caching_control, list);
9870                 list_del(&caching_ctl->list);
9871                 put_caching_control(caching_ctl);
9872         }
9873         up_write(&info->commit_root_sem);
9874
9875         spin_lock(&info->unused_bgs_lock);
9876         while (!list_empty(&info->unused_bgs)) {
9877                 block_group = list_first_entry(&info->unused_bgs,
9878                                                struct btrfs_block_group_cache,
9879                                                bg_list);
9880                 list_del_init(&block_group->bg_list);
9881                 btrfs_put_block_group(block_group);
9882         }
9883         spin_unlock(&info->unused_bgs_lock);
9884
9885         spin_lock(&info->block_group_cache_lock);
9886         while ((n = rb_last(&info->block_group_cache_tree)) != NULL) {
9887                 block_group = rb_entry(n, struct btrfs_block_group_cache,
9888                                        cache_node);
9889                 rb_erase(&block_group->cache_node,
9890                          &info->block_group_cache_tree);
9891                 RB_CLEAR_NODE(&block_group->cache_node);
9892                 spin_unlock(&info->block_group_cache_lock);
9893
9894                 down_write(&block_group->space_info->groups_sem);
9895                 list_del(&block_group->list);
9896                 up_write(&block_group->space_info->groups_sem);
9897
9898                 /*
9899                  * We haven't cached this block group, which means we could
9900                  * possibly have excluded extents on this block group.
9901                  */
9902                 if (block_group->cached == BTRFS_CACHE_NO ||
9903                     block_group->cached == BTRFS_CACHE_ERROR)
9904                         free_excluded_extents(info, block_group);
9905
9906                 btrfs_remove_free_space_cache(block_group);
9907                 ASSERT(block_group->cached != BTRFS_CACHE_STARTED);
9908                 ASSERT(list_empty(&block_group->dirty_list));
9909                 ASSERT(list_empty(&block_group->io_list));
9910                 ASSERT(list_empty(&block_group->bg_list));
9911                 ASSERT(atomic_read(&block_group->count) == 1);
9912                 btrfs_put_block_group(block_group);
9913
9914                 spin_lock(&info->block_group_cache_lock);
9915         }
9916         spin_unlock(&info->block_group_cache_lock);
9917
9918         /* now that all the block groups are freed, go through and
9919          * free all the space_info structs.  This is only called during
9920          * the final stages of unmount, and so we know nobody is
9921          * using them.  We call synchronize_rcu() once before we start,
9922          * just to be on the safe side.
9923          */
9924         synchronize_rcu();
9925
9926         release_global_block_rsv(info);
9927
9928         while (!list_empty(&info->space_info)) {
9929                 int i;
9930
9931                 space_info = list_entry(info->space_info.next,
9932                                         struct btrfs_space_info,
9933                                         list);
9934
9935                 /*
9936                  * Do not hide this behind enospc_debug, this is actually
9937                  * important and indicates a real bug if this happens.
9938                  */
9939                 if (WARN_ON(space_info->bytes_pinned > 0 ||
9940                             space_info->bytes_reserved > 0 ||
9941                             space_info->bytes_may_use > 0))
9942                         dump_space_info(info, space_info, 0, 0);
9943                 list_del(&space_info->list);
9944                 for (i = 0; i < BTRFS_NR_RAID_TYPES; i++) {
9945                         struct kobject *kobj;
9946                         kobj = space_info->block_group_kobjs[i];
9947                         space_info->block_group_kobjs[i] = NULL;
9948                         if (kobj) {
9949                                 kobject_del(kobj);
9950                                 kobject_put(kobj);
9951                         }
9952                 }
9953                 kobject_del(&space_info->kobj);
9954                 kobject_put(&space_info->kobj);
9955         }
9956         return 0;
9957 }
9958
9959 /* link_block_group will queue up kobjects to add when we're reclaim-safe */
9960 void btrfs_add_raid_kobjects(struct btrfs_fs_info *fs_info)
9961 {
9962         struct btrfs_space_info *space_info;
9963         struct raid_kobject *rkobj;
9964         LIST_HEAD(list);
9965         int index;
9966         int ret = 0;
9967
9968         spin_lock(&fs_info->pending_raid_kobjs_lock);
9969         list_splice_init(&fs_info->pending_raid_kobjs, &list);
9970         spin_unlock(&fs_info->pending_raid_kobjs_lock);
9971
9972         list_for_each_entry(rkobj, &list, list) {
9973                 space_info = __find_space_info(fs_info, rkobj->flags);
9974                 index = btrfs_bg_flags_to_raid_index(rkobj->flags);
9975
9976                 ret = kobject_add(&rkobj->kobj, &space_info->kobj,
9977                                   "%s", get_raid_name(index));
9978                 if (ret) {
9979                         kobject_put(&rkobj->kobj);
9980                         break;
9981                 }
9982         }
9983         if (ret)
9984                 btrfs_warn(fs_info,
9985                            "failed to add kobject for block cache, ignoring");
9986 }
9987
9988 static void link_block_group(struct btrfs_block_group_cache *cache)
9989 {
9990         struct btrfs_space_info *space_info = cache->space_info;
9991         struct btrfs_fs_info *fs_info = cache->fs_info;
9992         int index = btrfs_bg_flags_to_raid_index(cache->flags);
9993         bool first = false;
9994
9995         down_write(&space_info->groups_sem);
9996         if (list_empty(&space_info->block_groups[index]))
9997                 first = true;
9998         list_add_tail(&cache->list, &space_info->block_groups[index]);
9999         up_write(&space_info->groups_sem);
10000
10001         if (first) {
10002                 struct raid_kobject *rkobj = kzalloc(sizeof(*rkobj), GFP_NOFS);
10003                 if (!rkobj) {
10004                         btrfs_warn(cache->fs_info,
10005                                 "couldn't alloc memory for raid level kobject");
10006                         return;
10007                 }
10008                 rkobj->flags = cache->flags;
10009                 kobject_init(&rkobj->kobj, &btrfs_raid_ktype);
10010
10011                 spin_lock(&fs_info->pending_raid_kobjs_lock);
10012                 list_add_tail(&rkobj->list, &fs_info->pending_raid_kobjs);
10013                 spin_unlock(&fs_info->pending_raid_kobjs_lock);
10014                 space_info->block_group_kobjs[index] = &rkobj->kobj;
10015         }
10016 }
10017
10018 static struct btrfs_block_group_cache *
10019 btrfs_create_block_group_cache(struct btrfs_fs_info *fs_info,
10020                                u64 start, u64 size)
10021 {
10022         struct btrfs_block_group_cache *cache;
10023
10024         cache = kzalloc(sizeof(*cache), GFP_NOFS);
10025         if (!cache)
10026                 return NULL;
10027
10028         cache->free_space_ctl = kzalloc(sizeof(*cache->free_space_ctl),
10029                                         GFP_NOFS);
10030         if (!cache->free_space_ctl) {
10031                 kfree(cache);
10032                 return NULL;
10033         }
10034
10035         cache->key.objectid = start;
10036         cache->key.offset = size;
10037         cache->key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
10038
10039         cache->fs_info = fs_info;
10040         cache->full_stripe_len = btrfs_full_stripe_len(fs_info, start);
10041         set_free_space_tree_thresholds(cache);
10042
10043         atomic_set(&cache->count, 1);
10044         spin_lock_init(&cache->lock);
10045         init_rwsem(&cache->data_rwsem);
10046         INIT_LIST_HEAD(&cache->list);
10047         INIT_LIST_HEAD(&cache->cluster_list);
10048         INIT_LIST_HEAD(&cache->bg_list);
10049         INIT_LIST_HEAD(&cache->ro_list);
10050         INIT_LIST_HEAD(&cache->dirty_list);
10051         INIT_LIST_HEAD(&cache->io_list);
10052         btrfs_init_free_space_ctl(cache);
10053         atomic_set(&cache->trimming, 0);
10054         mutex_init(&cache->free_space_lock);
10055         btrfs_init_full_stripe_locks_tree(&cache->full_stripe_locks_root);
10056
10057         return cache;
10058 }
10059
10060 int btrfs_read_block_groups(struct btrfs_fs_info *info)
10061 {
10062         struct btrfs_path *path;
10063         int ret;
10064         struct btrfs_block_group_cache *cache;
10065         struct btrfs_space_info *space_info;
10066         struct btrfs_key key;
10067         struct btrfs_key found_key;
10068         struct extent_buffer *leaf;
10069         int need_clear = 0;
10070         u64 cache_gen;
10071         u64 feature;
10072         int mixed;
10073
10074         feature = btrfs_super_incompat_flags(info->super_copy);
10075         mixed = !!(feature & BTRFS_FEATURE_INCOMPAT_MIXED_GROUPS);
10076
10077         key.objectid = 0;
10078         key.offset = 0;
10079         key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
10080         path = btrfs_alloc_path();
10081         if (!path)
10082                 return -ENOMEM;
10083         path->reada = READA_FORWARD;
10084
10085         cache_gen = btrfs_super_cache_generation(info->super_copy);
10086         if (btrfs_test_opt(info, SPACE_CACHE) &&
10087             btrfs_super_generation(info->super_copy) != cache_gen)
10088                 need_clear = 1;
10089         if (btrfs_test_opt(info, CLEAR_CACHE))
10090                 need_clear = 1;
10091
10092         while (1) {
10093                 ret = find_first_block_group(info, path, &key);
10094                 if (ret > 0)
10095                         break;
10096                 if (ret != 0)
10097                         goto error;
10098
10099                 leaf = path->nodes[0];
10100                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
10101
10102                 cache = btrfs_create_block_group_cache(info, found_key.objectid,
10103                                                        found_key.offset);
10104                 if (!cache) {
10105                         ret = -ENOMEM;
10106                         goto error;
10107                 }
10108
10109                 if (need_clear) {
10110                         /*
10111                          * When we mount with old space cache, we need to
10112                          * set BTRFS_DC_CLEAR and set dirty flag.
10113                          *
10114                          * a) Setting 'BTRFS_DC_CLEAR' makes sure that we
10115                          *    truncate the old free space cache inode and
10116                          *    setup a new one.
10117                          * b) Setting 'dirty flag' makes sure that we flush
10118                          *    the new space cache info onto disk.
10119                          */
10120                         if (btrfs_test_opt(info, SPACE_CACHE))
10121                                 cache->disk_cache_state = BTRFS_DC_CLEAR;
10122                 }
10123
10124                 read_extent_buffer(leaf, &cache->item,
10125                                    btrfs_item_ptr_offset(leaf, path->slots[0]),
10126                                    sizeof(cache->item));
10127                 cache->flags = btrfs_block_group_flags(&cache->item);
10128                 if (!mixed &&
10129                     ((cache->flags & BTRFS_BLOCK_GROUP_METADATA) &&
10130                     (cache->flags & BTRFS_BLOCK_GROUP_DATA))) {
10131                         btrfs_err(info,
10132 "bg %llu is a mixed block group but filesystem hasn't enabled mixed block groups",
10133                                   cache->key.objectid);
10134                         ret = -EINVAL;
10135                         goto error;
10136                 }
10137
10138                 key.objectid = found_key.objectid + found_key.offset;
10139                 btrfs_release_path(path);
10140
10141                 /*
10142                  * We need to exclude the super stripes now so that the space
10143                  * info has super bytes accounted for, otherwise we'll think
10144                  * we have more space than we actually do.
10145                  */
10146                 ret = exclude_super_stripes(info, cache);
10147                 if (ret) {
10148                         /*
10149                          * We may have excluded something, so call this just in
10150                          * case.
10151                          */
10152                         free_excluded_extents(info, cache);
10153                         btrfs_put_block_group(cache);
10154                         goto error;
10155                 }
10156
10157                 /*
10158                  * check for two cases, either we are full, and therefore
10159                  * don't need to bother with the caching work since we won't
10160                  * find any space, or we are empty, and we can just add all
10161                  * the space in and be done with it.  This saves us _alot_ of
10162                  * time, particularly in the full case.
10163                  */
10164                 if (found_key.offset == btrfs_block_group_used(&cache->item)) {
10165                         cache->last_byte_to_unpin = (u64)-1;
10166                         cache->cached = BTRFS_CACHE_FINISHED;
10167                         free_excluded_extents(info, cache);
10168                 } else if (btrfs_block_group_used(&cache->item) == 0) {
10169                         cache->last_byte_to_unpin = (u64)-1;
10170                         cache->cached = BTRFS_CACHE_FINISHED;
10171                         add_new_free_space(cache, info,
10172                                            found_key.objectid,
10173                                            found_key.objectid +
10174                                            found_key.offset);
10175                         free_excluded_extents(info, cache);
10176                 }
10177
10178                 ret = btrfs_add_block_group_cache(info, cache);
10179                 if (ret) {
10180                         btrfs_remove_free_space_cache(cache);
10181                         btrfs_put_block_group(cache);
10182                         goto error;
10183                 }
10184
10185                 trace_btrfs_add_block_group(info, cache, 0);
10186                 update_space_info(info, cache->flags, found_key.offset,
10187                                   btrfs_block_group_used(&cache->item),
10188                                   cache->bytes_super, &space_info);
10189
10190                 cache->space_info = space_info;
10191
10192                 link_block_group(cache);
10193
10194                 set_avail_alloc_bits(info, cache->flags);
10195                 if (btrfs_chunk_readonly(info, cache->key.objectid)) {
10196                         inc_block_group_ro(cache, 1);
10197                 } else if (btrfs_block_group_used(&cache->item) == 0) {
10198                         spin_lock(&info->unused_bgs_lock);
10199                         /* Should always be true but just in case. */
10200                         if (list_empty(&cache->bg_list)) {
10201                                 btrfs_get_block_group(cache);
10202                                 list_add_tail(&cache->bg_list,
10203                                               &info->unused_bgs);
10204                         }
10205                         spin_unlock(&info->unused_bgs_lock);
10206                 }
10207         }
10208
10209         list_for_each_entry_rcu(space_info, &info->space_info, list) {
10210                 if (!(get_alloc_profile(info, space_info->flags) &
10211                       (BTRFS_BLOCK_GROUP_RAID10 |
10212                        BTRFS_BLOCK_GROUP_RAID1 |
10213                        BTRFS_BLOCK_GROUP_RAID5 |
10214                        BTRFS_BLOCK_GROUP_RAID6 |
10215                        BTRFS_BLOCK_GROUP_DUP)))
10216                         continue;
10217                 /*
10218                  * avoid allocating from un-mirrored block group if there are
10219                  * mirrored block groups.
10220                  */
10221                 list_for_each_entry(cache,
10222                                 &space_info->block_groups[BTRFS_RAID_RAID0],
10223                                 list)
10224                         inc_block_group_ro(cache, 1);
10225                 list_for_each_entry(cache,
10226                                 &space_info->block_groups[BTRFS_RAID_SINGLE],
10227                                 list)
10228                         inc_block_group_ro(cache, 1);
10229         }
10230
10231         btrfs_add_raid_kobjects(info);
10232         init_global_block_rsv(info);
10233         ret = 0;
10234 error:
10235         btrfs_free_path(path);
10236         return ret;
10237 }
10238
10239 void btrfs_create_pending_block_groups(struct btrfs_trans_handle *trans)
10240 {
10241         struct btrfs_fs_info *fs_info = trans->fs_info;
10242         struct btrfs_block_group_cache *block_group, *tmp;
10243         struct btrfs_root *extent_root = fs_info->extent_root;
10244         struct btrfs_block_group_item item;
10245         struct btrfs_key key;
10246         int ret = 0;
10247         bool can_flush_pending_bgs = trans->can_flush_pending_bgs;
10248
10249         trans->can_flush_pending_bgs = false;
10250         list_for_each_entry_safe(block_group, tmp, &trans->new_bgs, bg_list) {
10251                 if (ret)
10252                         goto next;
10253
10254                 spin_lock(&block_group->lock);
10255                 memcpy(&item, &block_group->item, sizeof(item));
10256                 memcpy(&key, &block_group->key, sizeof(key));
10257                 spin_unlock(&block_group->lock);
10258
10259                 ret = btrfs_insert_item(trans, extent_root, &key, &item,
10260                                         sizeof(item));
10261                 if (ret)
10262                         btrfs_abort_transaction(trans, ret);
10263                 ret = btrfs_finish_chunk_alloc(trans, fs_info, key.objectid,
10264                                                key.offset);
10265                 if (ret)
10266                         btrfs_abort_transaction(trans, ret);
10267                 add_block_group_free_space(trans, fs_info, block_group);
10268                 /* already aborted the transaction if it failed. */
10269 next:
10270                 list_del_init(&block_group->bg_list);
10271         }
10272         trans->can_flush_pending_bgs = can_flush_pending_bgs;
10273 }
10274
10275 int btrfs_make_block_group(struct btrfs_trans_handle *trans,
10276                            struct btrfs_fs_info *fs_info, u64 bytes_used,
10277                            u64 type, u64 chunk_offset, u64 size)
10278 {
10279         struct btrfs_block_group_cache *cache;
10280         int ret;
10281
10282         btrfs_set_log_full_commit(fs_info, trans);
10283
10284         cache = btrfs_create_block_group_cache(fs_info, chunk_offset, size);
10285         if (!cache)
10286                 return -ENOMEM;
10287
10288         btrfs_set_block_group_used(&cache->item, bytes_used);
10289         btrfs_set_block_group_chunk_objectid(&cache->item,
10290                                              BTRFS_FIRST_CHUNK_TREE_OBJECTID);
10291         btrfs_set_block_group_flags(&cache->item, type);
10292
10293         cache->flags = type;
10294         cache->last_byte_to_unpin = (u64)-1;
10295         cache->cached = BTRFS_CACHE_FINISHED;
10296         cache->needs_free_space = 1;
10297         ret = exclude_super_stripes(fs_info, cache);
10298         if (ret) {
10299                 /*
10300                  * We may have excluded something, so call this just in
10301                  * case.
10302                  */
10303                 free_excluded_extents(fs_info, cache);
10304                 btrfs_put_block_group(cache);
10305                 return ret;
10306         }
10307
10308         add_new_free_space(cache, fs_info, chunk_offset, chunk_offset + size);
10309
10310         free_excluded_extents(fs_info, cache);
10311
10312 #ifdef CONFIG_BTRFS_DEBUG
10313         if (btrfs_should_fragment_free_space(cache)) {
10314                 u64 new_bytes_used = size - bytes_used;
10315
10316                 bytes_used += new_bytes_used >> 1;
10317                 fragment_free_space(cache);
10318         }
10319 #endif
10320         /*
10321          * Ensure the corresponding space_info object is created and
10322          * assigned to our block group. We want our bg to be added to the rbtree
10323          * with its ->space_info set.
10324          */
10325         cache->space_info = __find_space_info(fs_info, cache->flags);
10326         ASSERT(cache->space_info);
10327
10328         ret = btrfs_add_block_group_cache(fs_info, cache);
10329         if (ret) {
10330                 btrfs_remove_free_space_cache(cache);
10331                 btrfs_put_block_group(cache);
10332                 return ret;
10333         }
10334
10335         /*
10336          * Now that our block group has its ->space_info set and is inserted in
10337          * the rbtree, update the space info's counters.
10338          */
10339         trace_btrfs_add_block_group(fs_info, cache, 1);
10340         update_space_info(fs_info, cache->flags, size, bytes_used,
10341                                 cache->bytes_super, &cache->space_info);
10342         update_global_block_rsv(fs_info);
10343
10344         link_block_group(cache);
10345
10346         list_add_tail(&cache->bg_list, &trans->new_bgs);
10347
10348         set_avail_alloc_bits(fs_info, type);
10349         return 0;
10350 }
10351
10352 static void clear_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
10353 {
10354         u64 extra_flags = chunk_to_extended(flags) &
10355                                 BTRFS_EXTENDED_PROFILE_MASK;
10356
10357         write_seqlock(&fs_info->profiles_lock);
10358         if (flags & BTRFS_BLOCK_GROUP_DATA)
10359                 fs_info->avail_data_alloc_bits &= ~extra_flags;
10360         if (flags & BTRFS_BLOCK_GROUP_METADATA)
10361                 fs_info->avail_metadata_alloc_bits &= ~extra_flags;
10362         if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
10363                 fs_info->avail_system_alloc_bits &= ~extra_flags;
10364         write_sequnlock(&fs_info->profiles_lock);
10365 }
10366
10367 int btrfs_remove_block_group(struct btrfs_trans_handle *trans,
10368                              struct btrfs_fs_info *fs_info, u64 group_start,
10369                              struct extent_map *em)
10370 {
10371         struct btrfs_root *root = fs_info->extent_root;
10372         struct btrfs_path *path;
10373         struct btrfs_block_group_cache *block_group;
10374         struct btrfs_free_cluster *cluster;
10375         struct btrfs_root *tree_root = fs_info->tree_root;
10376         struct btrfs_key key;
10377         struct inode *inode;
10378         struct kobject *kobj = NULL;
10379         int ret;
10380         int index;
10381         int factor;
10382         struct btrfs_caching_control *caching_ctl = NULL;
10383         bool remove_em;
10384
10385         block_group = btrfs_lookup_block_group(fs_info, group_start);
10386         BUG_ON(!block_group);
10387         BUG_ON(!block_group->ro);
10388
10389         /*
10390          * Free the reserved super bytes from this block group before
10391          * remove it.
10392          */
10393         free_excluded_extents(fs_info, block_group);
10394         btrfs_free_ref_tree_range(fs_info, block_group->key.objectid,
10395                                   block_group->key.offset);
10396
10397         memcpy(&key, &block_group->key, sizeof(key));
10398         index = btrfs_bg_flags_to_raid_index(block_group->flags);
10399         if (block_group->flags & (BTRFS_BLOCK_GROUP_DUP |
10400                                   BTRFS_BLOCK_GROUP_RAID1 |
10401                                   BTRFS_BLOCK_GROUP_RAID10))
10402                 factor = 2;
10403         else
10404                 factor = 1;
10405
10406         /* make sure this block group isn't part of an allocation cluster */
10407         cluster = &fs_info->data_alloc_cluster;
10408         spin_lock(&cluster->refill_lock);
10409         btrfs_return_cluster_to_free_space(block_group, cluster);
10410         spin_unlock(&cluster->refill_lock);
10411
10412         /*
10413          * make sure this block group isn't part of a metadata
10414          * allocation cluster
10415          */
10416         cluster = &fs_info->meta_alloc_cluster;
10417         spin_lock(&cluster->refill_lock);
10418         btrfs_return_cluster_to_free_space(block_group, cluster);
10419         spin_unlock(&cluster->refill_lock);
10420
10421         path = btrfs_alloc_path();
10422         if (!path) {
10423                 ret = -ENOMEM;
10424                 goto out;
10425         }
10426
10427         /*
10428          * get the inode first so any iput calls done for the io_list
10429          * aren't the final iput (no unlinks allowed now)
10430          */
10431         inode = lookup_free_space_inode(fs_info, block_group, path);
10432
10433         mutex_lock(&trans->transaction->cache_write_mutex);
10434         /*
10435          * make sure our free spache cache IO is done before remove the
10436          * free space inode
10437          */
10438         spin_lock(&trans->transaction->dirty_bgs_lock);
10439         if (!list_empty(&block_group->io_list)) {
10440                 list_del_init(&block_group->io_list);
10441
10442                 WARN_ON(!IS_ERR(inode) && inode != block_group->io_ctl.inode);
10443
10444                 spin_unlock(&trans->transaction->dirty_bgs_lock);
10445                 btrfs_wait_cache_io(trans, block_group, path);
10446                 btrfs_put_block_group(block_group);
10447                 spin_lock(&trans->transaction->dirty_bgs_lock);
10448         }
10449
10450         if (!list_empty(&block_group->dirty_list)) {
10451                 list_del_init(&block_group->dirty_list);
10452                 btrfs_put_block_group(block_group);
10453         }
10454         spin_unlock(&trans->transaction->dirty_bgs_lock);
10455         mutex_unlock(&trans->transaction->cache_write_mutex);
10456
10457         if (!IS_ERR(inode)) {
10458                 ret = btrfs_orphan_add(trans, BTRFS_I(inode));
10459                 if (ret) {
10460                         btrfs_add_delayed_iput(inode);
10461                         goto out;
10462                 }
10463                 clear_nlink(inode);
10464                 /* One for the block groups ref */
10465                 spin_lock(&block_group->lock);
10466                 if (block_group->iref) {
10467                         block_group->iref = 0;
10468                         block_group->inode = NULL;
10469                         spin_unlock(&block_group->lock);
10470                         iput(inode);
10471                 } else {
10472                         spin_unlock(&block_group->lock);
10473                 }
10474                 /* One for our lookup ref */
10475                 btrfs_add_delayed_iput(inode);
10476         }
10477
10478         key.objectid = BTRFS_FREE_SPACE_OBJECTID;
10479         key.offset = block_group->key.objectid;
10480         key.type = 0;
10481
10482         ret = btrfs_search_slot(trans, tree_root, &key, path, -1, 1);
10483         if (ret < 0)
10484                 goto out;
10485         if (ret > 0)
10486                 btrfs_release_path(path);
10487         if (ret == 0) {
10488                 ret = btrfs_del_item(trans, tree_root, path);
10489                 if (ret)
10490                         goto out;
10491                 btrfs_release_path(path);
10492         }
10493
10494         spin_lock(&fs_info->block_group_cache_lock);
10495         rb_erase(&block_group->cache_node,
10496                  &fs_info->block_group_cache_tree);
10497         RB_CLEAR_NODE(&block_group->cache_node);
10498
10499         if (fs_info->first_logical_byte == block_group->key.objectid)
10500                 fs_info->first_logical_byte = (u64)-1;
10501         spin_unlock(&fs_info->block_group_cache_lock);
10502
10503         down_write(&block_group->space_info->groups_sem);
10504         /*
10505          * we must use list_del_init so people can check to see if they
10506          * are still on the list after taking the semaphore
10507          */
10508         list_del_init(&block_group->list);
10509         if (list_empty(&block_group->space_info->block_groups[index])) {
10510                 kobj = block_group->space_info->block_group_kobjs[index];
10511                 block_group->space_info->block_group_kobjs[index] = NULL;
10512                 clear_avail_alloc_bits(fs_info, block_group->flags);
10513         }
10514         up_write(&block_group->space_info->groups_sem);
10515         if (kobj) {
10516                 kobject_del(kobj);
10517                 kobject_put(kobj);
10518         }
10519
10520         if (block_group->has_caching_ctl)
10521                 caching_ctl = get_caching_control(block_group);
10522         if (block_group->cached == BTRFS_CACHE_STARTED)
10523                 wait_block_group_cache_done(block_group);
10524         if (block_group->has_caching_ctl) {
10525                 down_write(&fs_info->commit_root_sem);
10526                 if (!caching_ctl) {
10527                         struct btrfs_caching_control *ctl;
10528
10529                         list_for_each_entry(ctl,
10530                                     &fs_info->caching_block_groups, list)
10531                                 if (ctl->block_group == block_group) {
10532                                         caching_ctl = ctl;
10533                                         refcount_inc(&caching_ctl->count);
10534                                         break;
10535                                 }
10536                 }
10537                 if (caching_ctl)
10538                         list_del_init(&caching_ctl->list);
10539                 up_write(&fs_info->commit_root_sem);
10540                 if (caching_ctl) {
10541                         /* Once for the caching bgs list and once for us. */
10542                         put_caching_control(caching_ctl);
10543                         put_caching_control(caching_ctl);
10544                 }
10545         }
10546
10547         spin_lock(&trans->transaction->dirty_bgs_lock);
10548         if (!list_empty(&block_group->dirty_list)) {
10549                 WARN_ON(1);
10550         }
10551         if (!list_empty(&block_group->io_list)) {
10552                 WARN_ON(1);
10553         }
10554         spin_unlock(&trans->transaction->dirty_bgs_lock);
10555         btrfs_remove_free_space_cache(block_group);
10556
10557         spin_lock(&block_group->space_info->lock);
10558         list_del_init(&block_group->ro_list);
10559
10560         if (btrfs_test_opt(fs_info, ENOSPC_DEBUG)) {
10561                 WARN_ON(block_group->space_info->total_bytes
10562                         < block_group->key.offset);
10563                 WARN_ON(block_group->space_info->bytes_readonly
10564                         < block_group->key.offset);
10565                 WARN_ON(block_group->space_info->disk_total
10566                         < block_group->key.offset * factor);
10567         }
10568         block_group->space_info->total_bytes -= block_group->key.offset;
10569         block_group->space_info->bytes_readonly -= block_group->key.offset;
10570         block_group->space_info->disk_total -= block_group->key.offset * factor;
10571
10572         spin_unlock(&block_group->space_info->lock);
10573
10574         memcpy(&key, &block_group->key, sizeof(key));
10575
10576         mutex_lock(&fs_info->chunk_mutex);
10577         if (!list_empty(&em->list)) {
10578                 /* We're in the transaction->pending_chunks list. */
10579                 free_extent_map(em);
10580         }
10581         spin_lock(&block_group->lock);
10582         block_group->removed = 1;
10583         /*
10584          * At this point trimming can't start on this block group, because we
10585          * removed the block group from the tree fs_info->block_group_cache_tree
10586          * so no one can't find it anymore and even if someone already got this
10587          * block group before we removed it from the rbtree, they have already
10588          * incremented block_group->trimming - if they didn't, they won't find
10589          * any free space entries because we already removed them all when we
10590          * called btrfs_remove_free_space_cache().
10591          *
10592          * And we must not remove the extent map from the fs_info->mapping_tree
10593          * to prevent the same logical address range and physical device space
10594          * ranges from being reused for a new block group. This is because our
10595          * fs trim operation (btrfs_trim_fs() / btrfs_ioctl_fitrim()) is
10596          * completely transactionless, so while it is trimming a range the
10597          * currently running transaction might finish and a new one start,
10598          * allowing for new block groups to be created that can reuse the same
10599          * physical device locations unless we take this special care.
10600          *
10601          * There may also be an implicit trim operation if the file system
10602          * is mounted with -odiscard. The same protections must remain
10603          * in place until the extents have been discarded completely when
10604          * the transaction commit has completed.
10605          */
10606         remove_em = (atomic_read(&block_group->trimming) == 0);
10607         /*
10608          * Make sure a trimmer task always sees the em in the pinned_chunks list
10609          * if it sees block_group->removed == 1 (needs to lock block_group->lock
10610          * before checking block_group->removed).
10611          */
10612         if (!remove_em) {
10613                 /*
10614                  * Our em might be in trans->transaction->pending_chunks which
10615                  * is protected by fs_info->chunk_mutex ([lock|unlock]_chunks),
10616                  * and so is the fs_info->pinned_chunks list.
10617                  *
10618                  * So at this point we must be holding the chunk_mutex to avoid
10619                  * any races with chunk allocation (more specifically at
10620                  * volumes.c:contains_pending_extent()), to ensure it always
10621                  * sees the em, either in the pending_chunks list or in the
10622                  * pinned_chunks list.
10623                  */
10624                 list_move_tail(&em->list, &fs_info->pinned_chunks);
10625         }
10626         spin_unlock(&block_group->lock);
10627
10628         if (remove_em) {
10629                 struct extent_map_tree *em_tree;
10630
10631                 em_tree = &fs_info->mapping_tree.map_tree;
10632                 write_lock(&em_tree->lock);
10633                 /*
10634                  * The em might be in the pending_chunks list, so make sure the
10635                  * chunk mutex is locked, since remove_extent_mapping() will
10636                  * delete us from that list.
10637                  */
10638                 remove_extent_mapping(em_tree, em);
10639                 write_unlock(&em_tree->lock);
10640                 /* once for the tree */
10641                 free_extent_map(em);
10642         }
10643
10644         mutex_unlock(&fs_info->chunk_mutex);
10645
10646         ret = remove_block_group_free_space(trans, fs_info, block_group);
10647         if (ret)
10648                 goto out;
10649
10650         btrfs_put_block_group(block_group);
10651         btrfs_put_block_group(block_group);
10652
10653         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
10654         if (ret > 0)
10655                 ret = -EIO;
10656         if (ret < 0)
10657                 goto out;
10658
10659         ret = btrfs_del_item(trans, root, path);
10660 out:
10661         btrfs_free_path(path);
10662         return ret;
10663 }
10664
10665 struct btrfs_trans_handle *
10666 btrfs_start_trans_remove_block_group(struct btrfs_fs_info *fs_info,
10667                                      const u64 chunk_offset)
10668 {
10669         struct extent_map_tree *em_tree = &fs_info->mapping_tree.map_tree;
10670         struct extent_map *em;
10671         struct map_lookup *map;
10672         unsigned int num_items;
10673
10674         read_lock(&em_tree->lock);
10675         em = lookup_extent_mapping(em_tree, chunk_offset, 1);
10676         read_unlock(&em_tree->lock);
10677         ASSERT(em && em->start == chunk_offset);
10678
10679         /*
10680          * We need to reserve 3 + N units from the metadata space info in order
10681          * to remove a block group (done at btrfs_remove_chunk() and at
10682          * btrfs_remove_block_group()), which are used for:
10683          *
10684          * 1 unit for adding the free space inode's orphan (located in the tree
10685          * of tree roots).
10686          * 1 unit for deleting the block group item (located in the extent
10687          * tree).
10688          * 1 unit for deleting the free space item (located in tree of tree
10689          * roots).
10690          * N units for deleting N device extent items corresponding to each
10691          * stripe (located in the device tree).
10692          *
10693          * In order to remove a block group we also need to reserve units in the
10694          * system space info in order to update the chunk tree (update one or
10695          * more device items and remove one chunk item), but this is done at
10696          * btrfs_remove_chunk() through a call to check_system_chunk().
10697          */
10698         map = em->map_lookup;
10699         num_items = 3 + map->num_stripes;
10700         free_extent_map(em);
10701
10702         return btrfs_start_transaction_fallback_global_rsv(fs_info->extent_root,
10703                                                            num_items, 1);
10704 }
10705
10706 /*
10707  * Process the unused_bgs list and remove any that don't have any allocated
10708  * space inside of them.
10709  */
10710 void btrfs_delete_unused_bgs(struct btrfs_fs_info *fs_info)
10711 {
10712         struct btrfs_block_group_cache *block_group;
10713         struct btrfs_space_info *space_info;
10714         struct btrfs_trans_handle *trans;
10715         int ret = 0;
10716
10717         if (!test_bit(BTRFS_FS_OPEN, &fs_info->flags))
10718                 return;
10719
10720         spin_lock(&fs_info->unused_bgs_lock);
10721         while (!list_empty(&fs_info->unused_bgs)) {
10722                 u64 start, end;
10723                 int trimming;
10724
10725                 block_group = list_first_entry(&fs_info->unused_bgs,
10726                                                struct btrfs_block_group_cache,
10727                                                bg_list);
10728                 list_del_init(&block_group->bg_list);
10729
10730                 space_info = block_group->space_info;
10731
10732                 if (ret || btrfs_mixed_space_info(space_info)) {
10733                         btrfs_put_block_group(block_group);
10734                         continue;
10735                 }
10736                 spin_unlock(&fs_info->unused_bgs_lock);
10737
10738                 mutex_lock(&fs_info->delete_unused_bgs_mutex);
10739
10740                 /* Don't want to race with allocators so take the groups_sem */
10741                 down_write(&space_info->groups_sem);
10742                 spin_lock(&block_group->lock);
10743                 if (block_group->reserved ||
10744                     btrfs_block_group_used(&block_group->item) ||
10745                     block_group->ro ||
10746                     list_is_singular(&block_group->list)) {
10747                         /*
10748                          * We want to bail if we made new allocations or have
10749                          * outstanding allocations in this block group.  We do
10750                          * the ro check in case balance is currently acting on
10751                          * this block group.
10752                          */
10753                         spin_unlock(&block_group->lock);
10754                         up_write(&space_info->groups_sem);
10755                         goto next;
10756                 }
10757                 spin_unlock(&block_group->lock);
10758
10759                 /* We don't want to force the issue, only flip if it's ok. */
10760                 ret = inc_block_group_ro(block_group, 0);
10761                 up_write(&space_info->groups_sem);
10762                 if (ret < 0) {
10763                         ret = 0;
10764                         goto next;
10765                 }
10766
10767                 /*
10768                  * Want to do this before we do anything else so we can recover
10769                  * properly if we fail to join the transaction.
10770                  */
10771                 trans = btrfs_start_trans_remove_block_group(fs_info,
10772                                                      block_group->key.objectid);
10773                 if (IS_ERR(trans)) {
10774                         btrfs_dec_block_group_ro(block_group);
10775                         ret = PTR_ERR(trans);
10776                         goto next;
10777                 }
10778
10779                 /*
10780                  * We could have pending pinned extents for this block group,
10781                  * just delete them, we don't care about them anymore.
10782                  */
10783                 start = block_group->key.objectid;
10784                 end = start + block_group->key.offset - 1;
10785                 /*
10786                  * Hold the unused_bg_unpin_mutex lock to avoid racing with
10787                  * btrfs_finish_extent_commit(). If we are at transaction N,
10788                  * another task might be running finish_extent_commit() for the
10789                  * previous transaction N - 1, and have seen a range belonging
10790                  * to the block group in freed_extents[] before we were able to
10791                  * clear the whole block group range from freed_extents[]. This
10792                  * means that task can lookup for the block group after we
10793                  * unpinned it from freed_extents[] and removed it, leading to
10794                  * a BUG_ON() at btrfs_unpin_extent_range().
10795                  */
10796                 mutex_lock(&fs_info->unused_bg_unpin_mutex);
10797                 ret = clear_extent_bits(&fs_info->freed_extents[0], start, end,
10798                                   EXTENT_DIRTY);
10799                 if (ret) {
10800                         mutex_unlock(&fs_info->unused_bg_unpin_mutex);
10801                         btrfs_dec_block_group_ro(block_group);
10802                         goto end_trans;
10803                 }
10804                 ret = clear_extent_bits(&fs_info->freed_extents[1], start, end,
10805                                   EXTENT_DIRTY);
10806                 if (ret) {
10807                         mutex_unlock(&fs_info->unused_bg_unpin_mutex);
10808                         btrfs_dec_block_group_ro(block_group);
10809                         goto end_trans;
10810                 }
10811                 mutex_unlock(&fs_info->unused_bg_unpin_mutex);
10812
10813                 /* Reset pinned so btrfs_put_block_group doesn't complain */
10814                 spin_lock(&space_info->lock);
10815                 spin_lock(&block_group->lock);
10816
10817                 space_info->bytes_pinned -= block_group->pinned;
10818                 space_info->bytes_readonly += block_group->pinned;
10819                 percpu_counter_add(&space_info->total_bytes_pinned,
10820                                    -block_group->pinned);
10821                 block_group->pinned = 0;
10822
10823                 spin_unlock(&block_group->lock);
10824                 spin_unlock(&space_info->lock);
10825
10826                 /* DISCARD can flip during remount */
10827                 trimming = btrfs_test_opt(fs_info, DISCARD);
10828
10829                 /* Implicit trim during transaction commit. */
10830                 if (trimming)
10831                         btrfs_get_block_group_trimming(block_group);
10832
10833                 /*
10834                  * Btrfs_remove_chunk will abort the transaction if things go
10835                  * horribly wrong.
10836                  */
10837                 ret = btrfs_remove_chunk(trans, fs_info,
10838                                          block_group->key.objectid);
10839
10840                 if (ret) {
10841                         if (trimming)
10842                                 btrfs_put_block_group_trimming(block_group);
10843                         goto end_trans;
10844                 }
10845
10846                 /*
10847                  * If we're not mounted with -odiscard, we can just forget
10848                  * about this block group. Otherwise we'll need to wait
10849                  * until transaction commit to do the actual discard.
10850                  */
10851                 if (trimming) {
10852                         spin_lock(&fs_info->unused_bgs_lock);
10853                         /*
10854                          * A concurrent scrub might have added us to the list
10855                          * fs_info->unused_bgs, so use a list_move operation
10856                          * to add the block group to the deleted_bgs list.
10857                          */
10858                         list_move(&block_group->bg_list,
10859                                   &trans->transaction->deleted_bgs);
10860                         spin_unlock(&fs_info->unused_bgs_lock);
10861                         btrfs_get_block_group(block_group);
10862                 }
10863 end_trans:
10864                 btrfs_end_transaction(trans);
10865 next:
10866                 mutex_unlock(&fs_info->delete_unused_bgs_mutex);
10867                 btrfs_put_block_group(block_group);
10868                 spin_lock(&fs_info->unused_bgs_lock);
10869         }
10870         spin_unlock(&fs_info->unused_bgs_lock);
10871 }
10872
10873 int btrfs_init_space_info(struct btrfs_fs_info *fs_info)
10874 {
10875         struct btrfs_space_info *space_info;
10876         struct btrfs_super_block *disk_super;
10877         u64 features;
10878         u64 flags;
10879         int mixed = 0;
10880         int ret;
10881
10882         disk_super = fs_info->super_copy;
10883         if (!btrfs_super_root(disk_super))
10884                 return -EINVAL;
10885
10886         features = btrfs_super_incompat_flags(disk_super);
10887         if (features & BTRFS_FEATURE_INCOMPAT_MIXED_GROUPS)
10888                 mixed = 1;
10889
10890         flags = BTRFS_BLOCK_GROUP_SYSTEM;
10891         ret = create_space_info(fs_info, flags, &space_info);
10892         if (ret)
10893                 goto out;
10894
10895         if (mixed) {
10896                 flags = BTRFS_BLOCK_GROUP_METADATA | BTRFS_BLOCK_GROUP_DATA;
10897                 ret = create_space_info(fs_info, flags, &space_info);
10898         } else {
10899                 flags = BTRFS_BLOCK_GROUP_METADATA;
10900                 ret = create_space_info(fs_info, flags, &space_info);
10901                 if (ret)
10902                         goto out;
10903
10904                 flags = BTRFS_BLOCK_GROUP_DATA;
10905                 ret = create_space_info(fs_info, flags, &space_info);
10906         }
10907 out:
10908         return ret;
10909 }
10910
10911 int btrfs_error_unpin_extent_range(struct btrfs_fs_info *fs_info,
10912                                    u64 start, u64 end)
10913 {
10914         return unpin_extent_range(fs_info, start, end, false);
10915 }
10916
10917 /*
10918  * It used to be that old block groups would be left around forever.
10919  * Iterating over them would be enough to trim unused space.  Since we
10920  * now automatically remove them, we also need to iterate over unallocated
10921  * space.
10922  *
10923  * We don't want a transaction for this since the discard may take a
10924  * substantial amount of time.  We don't require that a transaction be
10925  * running, but we do need to take a running transaction into account
10926  * to ensure that we're not discarding chunks that were released in
10927  * the current transaction.
10928  *
10929  * Holding the chunks lock will prevent other threads from allocating
10930  * or releasing chunks, but it won't prevent a running transaction
10931  * from committing and releasing the memory that the pending chunks
10932  * list head uses.  For that, we need to take a reference to the
10933  * transaction.
10934  */
10935 static int btrfs_trim_free_extents(struct btrfs_device *device,
10936                                    u64 minlen, u64 *trimmed)
10937 {
10938         u64 start = 0, len = 0;
10939         int ret;
10940
10941         *trimmed = 0;
10942
10943         /* Not writeable = nothing to do. */
10944         if (!test_bit(BTRFS_DEV_STATE_WRITEABLE, &device->dev_state))
10945                 return 0;
10946
10947         /* No free space = nothing to do. */
10948         if (device->total_bytes <= device->bytes_used)
10949                 return 0;
10950
10951         ret = 0;
10952
10953         while (1) {
10954                 struct btrfs_fs_info *fs_info = device->fs_info;
10955                 struct btrfs_transaction *trans;
10956                 u64 bytes;
10957
10958                 ret = mutex_lock_interruptible(&fs_info->chunk_mutex);
10959                 if (ret)
10960                         return ret;
10961
10962                 down_read(&fs_info->commit_root_sem);
10963
10964                 spin_lock(&fs_info->trans_lock);
10965                 trans = fs_info->running_transaction;
10966                 if (trans)
10967                         refcount_inc(&trans->use_count);
10968                 spin_unlock(&fs_info->trans_lock);
10969
10970                 ret = find_free_dev_extent_start(trans, device, minlen, start,
10971                                                  &start, &len);
10972                 if (trans)
10973                         btrfs_put_transaction(trans);
10974
10975                 if (ret) {
10976                         up_read(&fs_info->commit_root_sem);
10977                         mutex_unlock(&fs_info->chunk_mutex);
10978                         if (ret == -ENOSPC)
10979                                 ret = 0;
10980                         break;
10981                 }
10982
10983                 ret = btrfs_issue_discard(device->bdev, start, len, &bytes);
10984                 up_read(&fs_info->commit_root_sem);
10985                 mutex_unlock(&fs_info->chunk_mutex);
10986
10987                 if (ret)
10988                         break;
10989
10990                 start += len;
10991                 *trimmed += bytes;
10992
10993                 if (fatal_signal_pending(current)) {
10994                         ret = -ERESTARTSYS;
10995                         break;
10996                 }
10997
10998                 cond_resched();
10999         }
11000
11001         return ret;
11002 }
11003
11004 int btrfs_trim_fs(struct btrfs_fs_info *fs_info, struct fstrim_range *range)
11005 {
11006         struct btrfs_block_group_cache *cache = NULL;
11007         struct btrfs_device *device;
11008         struct list_head *devices;
11009         u64 group_trimmed;
11010         u64 start;
11011         u64 end;
11012         u64 trimmed = 0;
11013         u64 total_bytes = btrfs_super_total_bytes(fs_info->super_copy);
11014         int ret = 0;
11015
11016         /*
11017          * try to trim all FS space, our block group may start from non-zero.
11018          */
11019         if (range->len == total_bytes)
11020                 cache = btrfs_lookup_first_block_group(fs_info, range->start);
11021         else
11022                 cache = btrfs_lookup_block_group(fs_info, range->start);
11023
11024         while (cache) {
11025                 if (cache->key.objectid >= (range->start + range->len)) {
11026                         btrfs_put_block_group(cache);
11027                         break;
11028                 }
11029
11030                 start = max(range->start, cache->key.objectid);
11031                 end = min(range->start + range->len,
11032                                 cache->key.objectid + cache->key.offset);
11033
11034                 if (end - start >= range->minlen) {
11035                         if (!block_group_cache_done(cache)) {
11036                                 ret = cache_block_group(cache, 0);
11037                                 if (ret) {
11038                                         btrfs_put_block_group(cache);
11039                                         break;
11040                                 }
11041                                 ret = wait_block_group_cache_done(cache);
11042                                 if (ret) {
11043                                         btrfs_put_block_group(cache);
11044                                         break;
11045                                 }
11046                         }
11047                         ret = btrfs_trim_block_group(cache,
11048                                                      &group_trimmed,
11049                                                      start,
11050                                                      end,
11051                                                      range->minlen);
11052
11053                         trimmed += group_trimmed;
11054                         if (ret) {
11055                                 btrfs_put_block_group(cache);
11056                                 break;
11057                         }
11058                 }
11059
11060                 cache = next_block_group(fs_info, cache);
11061         }
11062
11063         mutex_lock(&fs_info->fs_devices->device_list_mutex);
11064         devices = &fs_info->fs_devices->alloc_list;
11065         list_for_each_entry(device, devices, dev_alloc_list) {
11066                 ret = btrfs_trim_free_extents(device, range->minlen,
11067                                               &group_trimmed);
11068                 if (ret)
11069                         break;
11070
11071                 trimmed += group_trimmed;
11072         }
11073         mutex_unlock(&fs_info->fs_devices->device_list_mutex);
11074
11075         range->len = trimmed;
11076         return ret;
11077 }
11078
11079 /*
11080  * btrfs_{start,end}_write_no_snapshotting() are similar to
11081  * mnt_{want,drop}_write(), they are used to prevent some tasks from writing
11082  * data into the page cache through nocow before the subvolume is snapshoted,
11083  * but flush the data into disk after the snapshot creation, or to prevent
11084  * operations while snapshotting is ongoing and that cause the snapshot to be
11085  * inconsistent (writes followed by expanding truncates for example).
11086  */
11087 void btrfs_end_write_no_snapshotting(struct btrfs_root *root)
11088 {
11089         percpu_counter_dec(&root->subv_writers->counter);
11090         /*
11091          * Make sure counter is updated before we wake up waiters.
11092          */
11093         smp_mb();
11094         if (waitqueue_active(&root->subv_writers->wait))
11095                 wake_up(&root->subv_writers->wait);
11096 }
11097
11098 int btrfs_start_write_no_snapshotting(struct btrfs_root *root)
11099 {
11100         if (atomic_read(&root->will_be_snapshotted))
11101                 return 0;
11102
11103         percpu_counter_inc(&root->subv_writers->counter);
11104         /*
11105          * Make sure counter is updated before we check for snapshot creation.
11106          */
11107         smp_mb();
11108         if (atomic_read(&root->will_be_snapshotted)) {
11109                 btrfs_end_write_no_snapshotting(root);
11110                 return 0;
11111         }
11112         return 1;
11113 }
11114
11115 void btrfs_wait_for_snapshot_creation(struct btrfs_root *root)
11116 {
11117         while (true) {
11118                 int ret;
11119
11120                 ret = btrfs_start_write_no_snapshotting(root);
11121                 if (ret)
11122                         break;
11123                 wait_on_atomic_t(&root->will_be_snapshotted, atomic_t_wait,
11124                                  TASK_UNINTERRUPTIBLE);
11125         }
11126 }