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_var(&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_var_event(&bg->nocow_writers, !atomic_read(&bg->nocow_writers));
4003 }
4004
4005 static const char *alloc_name(u64 flags)
4006 {
4007         switch (flags) {
4008         case BTRFS_BLOCK_GROUP_METADATA|BTRFS_BLOCK_GROUP_DATA:
4009                 return "mixed";
4010         case BTRFS_BLOCK_GROUP_METADATA:
4011                 return "metadata";
4012         case BTRFS_BLOCK_GROUP_DATA:
4013                 return "data";
4014         case BTRFS_BLOCK_GROUP_SYSTEM:
4015                 return "system";
4016         default:
4017                 WARN_ON(1);
4018                 return "invalid-combination";
4019         };
4020 }
4021
4022 static int create_space_info(struct btrfs_fs_info *info, u64 flags,
4023                              struct btrfs_space_info **new)
4024 {
4025
4026         struct btrfs_space_info *space_info;
4027         int i;
4028         int ret;
4029
4030         space_info = kzalloc(sizeof(*space_info), GFP_NOFS);
4031         if (!space_info)
4032                 return -ENOMEM;
4033
4034         ret = percpu_counter_init(&space_info->total_bytes_pinned, 0,
4035                                  GFP_KERNEL);
4036         if (ret) {
4037                 kfree(space_info);
4038                 return ret;
4039         }
4040
4041         for (i = 0; i < BTRFS_NR_RAID_TYPES; i++)
4042                 INIT_LIST_HEAD(&space_info->block_groups[i]);
4043         init_rwsem(&space_info->groups_sem);
4044         spin_lock_init(&space_info->lock);
4045         space_info->flags = flags & BTRFS_BLOCK_GROUP_TYPE_MASK;
4046         space_info->force_alloc = CHUNK_ALLOC_NO_FORCE;
4047         init_waitqueue_head(&space_info->wait);
4048         INIT_LIST_HEAD(&space_info->ro_bgs);
4049         INIT_LIST_HEAD(&space_info->tickets);
4050         INIT_LIST_HEAD(&space_info->priority_tickets);
4051
4052         ret = kobject_init_and_add(&space_info->kobj, &space_info_ktype,
4053                                     info->space_info_kobj, "%s",
4054                                     alloc_name(space_info->flags));
4055         if (ret) {
4056                 percpu_counter_destroy(&space_info->total_bytes_pinned);
4057                 kfree(space_info);
4058                 return ret;
4059         }
4060
4061         *new = space_info;
4062         list_add_rcu(&space_info->list, &info->space_info);
4063         if (flags & BTRFS_BLOCK_GROUP_DATA)
4064                 info->data_sinfo = space_info;
4065
4066         return ret;
4067 }
4068
4069 static void update_space_info(struct btrfs_fs_info *info, u64 flags,
4070                              u64 total_bytes, u64 bytes_used,
4071                              u64 bytes_readonly,
4072                              struct btrfs_space_info **space_info)
4073 {
4074         struct btrfs_space_info *found;
4075         int factor;
4076
4077         if (flags & (BTRFS_BLOCK_GROUP_DUP | BTRFS_BLOCK_GROUP_RAID1 |
4078                      BTRFS_BLOCK_GROUP_RAID10))
4079                 factor = 2;
4080         else
4081                 factor = 1;
4082
4083         found = __find_space_info(info, flags);
4084         ASSERT(found);
4085         spin_lock(&found->lock);
4086         found->total_bytes += total_bytes;
4087         found->disk_total += total_bytes * factor;
4088         found->bytes_used += bytes_used;
4089         found->disk_used += bytes_used * factor;
4090         found->bytes_readonly += bytes_readonly;
4091         if (total_bytes > 0)
4092                 found->full = 0;
4093         space_info_add_new_bytes(info, found, total_bytes -
4094                                  bytes_used - bytes_readonly);
4095         spin_unlock(&found->lock);
4096         *space_info = found;
4097 }
4098
4099 static void set_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
4100 {
4101         u64 extra_flags = chunk_to_extended(flags) &
4102                                 BTRFS_EXTENDED_PROFILE_MASK;
4103
4104         write_seqlock(&fs_info->profiles_lock);
4105         if (flags & BTRFS_BLOCK_GROUP_DATA)
4106                 fs_info->avail_data_alloc_bits |= extra_flags;
4107         if (flags & BTRFS_BLOCK_GROUP_METADATA)
4108                 fs_info->avail_metadata_alloc_bits |= extra_flags;
4109         if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
4110                 fs_info->avail_system_alloc_bits |= extra_flags;
4111         write_sequnlock(&fs_info->profiles_lock);
4112 }
4113
4114 /*
4115  * returns target flags in extended format or 0 if restripe for this
4116  * chunk_type is not in progress
4117  *
4118  * should be called with either volume_mutex or balance_lock held
4119  */
4120 static u64 get_restripe_target(struct btrfs_fs_info *fs_info, u64 flags)
4121 {
4122         struct btrfs_balance_control *bctl = fs_info->balance_ctl;
4123         u64 target = 0;
4124
4125         if (!bctl)
4126                 return 0;
4127
4128         if (flags & BTRFS_BLOCK_GROUP_DATA &&
4129             bctl->data.flags & BTRFS_BALANCE_ARGS_CONVERT) {
4130                 target = BTRFS_BLOCK_GROUP_DATA | bctl->data.target;
4131         } else if (flags & BTRFS_BLOCK_GROUP_SYSTEM &&
4132                    bctl->sys.flags & BTRFS_BALANCE_ARGS_CONVERT) {
4133                 target = BTRFS_BLOCK_GROUP_SYSTEM | bctl->sys.target;
4134         } else if (flags & BTRFS_BLOCK_GROUP_METADATA &&
4135                    bctl->meta.flags & BTRFS_BALANCE_ARGS_CONVERT) {
4136                 target = BTRFS_BLOCK_GROUP_METADATA | bctl->meta.target;
4137         }
4138
4139         return target;
4140 }
4141
4142 /*
4143  * @flags: available profiles in extended format (see ctree.h)
4144  *
4145  * Returns reduced profile in chunk format.  If profile changing is in
4146  * progress (either running or paused) picks the target profile (if it's
4147  * already available), otherwise falls back to plain reducing.
4148  */
4149 static u64 btrfs_reduce_alloc_profile(struct btrfs_fs_info *fs_info, u64 flags)
4150 {
4151         u64 num_devices = fs_info->fs_devices->rw_devices;
4152         u64 target;
4153         u64 raid_type;
4154         u64 allowed = 0;
4155
4156         /*
4157          * see if restripe for this chunk_type is in progress, if so
4158          * try to reduce to the target profile
4159          */
4160         spin_lock(&fs_info->balance_lock);
4161         target = get_restripe_target(fs_info, flags);
4162         if (target) {
4163                 /* pick target profile only if it's already available */
4164                 if ((flags & target) & BTRFS_EXTENDED_PROFILE_MASK) {
4165                         spin_unlock(&fs_info->balance_lock);
4166                         return extended_to_chunk(target);
4167                 }
4168         }
4169         spin_unlock(&fs_info->balance_lock);
4170
4171         /* First, mask out the RAID levels which aren't possible */
4172         for (raid_type = 0; raid_type < BTRFS_NR_RAID_TYPES; raid_type++) {
4173                 if (num_devices >= btrfs_raid_array[raid_type].devs_min)
4174                         allowed |= btrfs_raid_group[raid_type];
4175         }
4176         allowed &= flags;
4177
4178         if (allowed & BTRFS_BLOCK_GROUP_RAID6)
4179                 allowed = BTRFS_BLOCK_GROUP_RAID6;
4180         else if (allowed & BTRFS_BLOCK_GROUP_RAID5)
4181                 allowed = BTRFS_BLOCK_GROUP_RAID5;
4182         else if (allowed & BTRFS_BLOCK_GROUP_RAID10)
4183                 allowed = BTRFS_BLOCK_GROUP_RAID10;
4184         else if (allowed & BTRFS_BLOCK_GROUP_RAID1)
4185                 allowed = BTRFS_BLOCK_GROUP_RAID1;
4186         else if (allowed & BTRFS_BLOCK_GROUP_RAID0)
4187                 allowed = BTRFS_BLOCK_GROUP_RAID0;
4188
4189         flags &= ~BTRFS_BLOCK_GROUP_PROFILE_MASK;
4190
4191         return extended_to_chunk(flags | allowed);
4192 }
4193
4194 static u64 get_alloc_profile(struct btrfs_fs_info *fs_info, u64 orig_flags)
4195 {
4196         unsigned seq;
4197         u64 flags;
4198
4199         do {
4200                 flags = orig_flags;
4201                 seq = read_seqbegin(&fs_info->profiles_lock);
4202
4203                 if (flags & BTRFS_BLOCK_GROUP_DATA)
4204                         flags |= fs_info->avail_data_alloc_bits;
4205                 else if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
4206                         flags |= fs_info->avail_system_alloc_bits;
4207                 else if (flags & BTRFS_BLOCK_GROUP_METADATA)
4208                         flags |= fs_info->avail_metadata_alloc_bits;
4209         } while (read_seqretry(&fs_info->profiles_lock, seq));
4210
4211         return btrfs_reduce_alloc_profile(fs_info, flags);
4212 }
4213
4214 static u64 get_alloc_profile_by_root(struct btrfs_root *root, int data)
4215 {
4216         struct btrfs_fs_info *fs_info = root->fs_info;
4217         u64 flags;
4218         u64 ret;
4219
4220         if (data)
4221                 flags = BTRFS_BLOCK_GROUP_DATA;
4222         else if (root == fs_info->chunk_root)
4223                 flags = BTRFS_BLOCK_GROUP_SYSTEM;
4224         else
4225                 flags = BTRFS_BLOCK_GROUP_METADATA;
4226
4227         ret = get_alloc_profile(fs_info, flags);
4228         return ret;
4229 }
4230
4231 u64 btrfs_data_alloc_profile(struct btrfs_fs_info *fs_info)
4232 {
4233         return get_alloc_profile(fs_info, BTRFS_BLOCK_GROUP_DATA);
4234 }
4235
4236 u64 btrfs_metadata_alloc_profile(struct btrfs_fs_info *fs_info)
4237 {
4238         return get_alloc_profile(fs_info, BTRFS_BLOCK_GROUP_METADATA);
4239 }
4240
4241 u64 btrfs_system_alloc_profile(struct btrfs_fs_info *fs_info)
4242 {
4243         return get_alloc_profile(fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
4244 }
4245
4246 static u64 btrfs_space_info_used(struct btrfs_space_info *s_info,
4247                                  bool may_use_included)
4248 {
4249         ASSERT(s_info);
4250         return s_info->bytes_used + s_info->bytes_reserved +
4251                 s_info->bytes_pinned + s_info->bytes_readonly +
4252                 (may_use_included ? s_info->bytes_may_use : 0);
4253 }
4254
4255 int btrfs_alloc_data_chunk_ondemand(struct btrfs_inode *inode, u64 bytes)
4256 {
4257         struct btrfs_root *root = inode->root;
4258         struct btrfs_fs_info *fs_info = root->fs_info;
4259         struct btrfs_space_info *data_sinfo = fs_info->data_sinfo;
4260         u64 used;
4261         int ret = 0;
4262         int need_commit = 2;
4263         int have_pinned_space;
4264
4265         /* make sure bytes are sectorsize aligned */
4266         bytes = ALIGN(bytes, fs_info->sectorsize);
4267
4268         if (btrfs_is_free_space_inode(inode)) {
4269                 need_commit = 0;
4270                 ASSERT(current->journal_info);
4271         }
4272
4273 again:
4274         /* make sure we have enough space to handle the data first */
4275         spin_lock(&data_sinfo->lock);
4276         used = btrfs_space_info_used(data_sinfo, true);
4277
4278         if (used + bytes > data_sinfo->total_bytes) {
4279                 struct btrfs_trans_handle *trans;
4280
4281                 /*
4282                  * if we don't have enough free bytes in this space then we need
4283                  * to alloc a new chunk.
4284                  */
4285                 if (!data_sinfo->full) {
4286                         u64 alloc_target;
4287
4288                         data_sinfo->force_alloc = CHUNK_ALLOC_FORCE;
4289                         spin_unlock(&data_sinfo->lock);
4290
4291                         alloc_target = btrfs_data_alloc_profile(fs_info);
4292                         /*
4293                          * It is ugly that we don't call nolock join
4294                          * transaction for the free space inode case here.
4295                          * But it is safe because we only do the data space
4296                          * reservation for the free space cache in the
4297                          * transaction context, the common join transaction
4298                          * just increase the counter of the current transaction
4299                          * handler, doesn't try to acquire the trans_lock of
4300                          * the fs.
4301                          */
4302                         trans = btrfs_join_transaction(root);
4303                         if (IS_ERR(trans))
4304                                 return PTR_ERR(trans);
4305
4306                         ret = do_chunk_alloc(trans, fs_info, alloc_target,
4307                                              CHUNK_ALLOC_NO_FORCE);
4308                         btrfs_end_transaction(trans);
4309                         if (ret < 0) {
4310                                 if (ret != -ENOSPC)
4311                                         return ret;
4312                                 else {
4313                                         have_pinned_space = 1;
4314                                         goto commit_trans;
4315                                 }
4316                         }
4317
4318                         goto again;
4319                 }
4320
4321                 /*
4322                  * If we don't have enough pinned space to deal with this
4323                  * allocation, and no removed chunk in current transaction,
4324                  * don't bother committing the transaction.
4325                  */
4326                 have_pinned_space = percpu_counter_compare(
4327                         &data_sinfo->total_bytes_pinned,
4328                         used + bytes - data_sinfo->total_bytes);
4329                 spin_unlock(&data_sinfo->lock);
4330
4331                 /* commit the current transaction and try again */
4332 commit_trans:
4333                 if (need_commit) {
4334                         need_commit--;
4335
4336                         if (need_commit > 0) {
4337                                 btrfs_start_delalloc_roots(fs_info, 0, -1);
4338                                 btrfs_wait_ordered_roots(fs_info, U64_MAX, 0,
4339                                                          (u64)-1);
4340                         }
4341
4342                         trans = btrfs_join_transaction(root);
4343                         if (IS_ERR(trans))
4344                                 return PTR_ERR(trans);
4345                         if (have_pinned_space >= 0 ||
4346                             test_bit(BTRFS_TRANS_HAVE_FREE_BGS,
4347                                      &trans->transaction->flags) ||
4348                             need_commit > 0) {
4349                                 ret = btrfs_commit_transaction(trans);
4350                                 if (ret)
4351                                         return ret;
4352                                 /*
4353                                  * The cleaner kthread might still be doing iput
4354                                  * operations. Wait for it to finish so that
4355                                  * more space is released.
4356                                  */
4357                                 mutex_lock(&fs_info->cleaner_delayed_iput_mutex);
4358                                 mutex_unlock(&fs_info->cleaner_delayed_iput_mutex);
4359                                 goto again;
4360                         } else {
4361                                 btrfs_end_transaction(trans);
4362                         }
4363                 }
4364
4365                 trace_btrfs_space_reservation(fs_info,
4366                                               "space_info:enospc",
4367                                               data_sinfo->flags, bytes, 1);
4368                 return -ENOSPC;
4369         }
4370         data_sinfo->bytes_may_use += bytes;
4371         trace_btrfs_space_reservation(fs_info, "space_info",
4372                                       data_sinfo->flags, bytes, 1);
4373         spin_unlock(&data_sinfo->lock);
4374
4375         return ret;
4376 }
4377
4378 int btrfs_check_data_free_space(struct inode *inode,
4379                         struct extent_changeset **reserved, u64 start, u64 len)
4380 {
4381         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
4382         int ret;
4383
4384         /* align the range */
4385         len = round_up(start + len, fs_info->sectorsize) -
4386               round_down(start, fs_info->sectorsize);
4387         start = round_down(start, fs_info->sectorsize);
4388
4389         ret = btrfs_alloc_data_chunk_ondemand(BTRFS_I(inode), len);
4390         if (ret < 0)
4391                 return ret;
4392
4393         /* Use new btrfs_qgroup_reserve_data to reserve precious data space. */
4394         ret = btrfs_qgroup_reserve_data(inode, reserved, start, len);
4395         if (ret < 0)
4396                 btrfs_free_reserved_data_space_noquota(inode, start, len);
4397         else
4398                 ret = 0;
4399         return ret;
4400 }
4401
4402 /*
4403  * Called if we need to clear a data reservation for this inode
4404  * Normally in a error case.
4405  *
4406  * This one will *NOT* use accurate qgroup reserved space API, just for case
4407  * which we can't sleep and is sure it won't affect qgroup reserved space.
4408  * Like clear_bit_hook().
4409  */
4410 void btrfs_free_reserved_data_space_noquota(struct inode *inode, u64 start,
4411                                             u64 len)
4412 {
4413         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
4414         struct btrfs_space_info *data_sinfo;
4415
4416         /* Make sure the range is aligned to sectorsize */
4417         len = round_up(start + len, fs_info->sectorsize) -
4418               round_down(start, fs_info->sectorsize);
4419         start = round_down(start, fs_info->sectorsize);
4420
4421         data_sinfo = fs_info->data_sinfo;
4422         spin_lock(&data_sinfo->lock);
4423         if (WARN_ON(data_sinfo->bytes_may_use < len))
4424                 data_sinfo->bytes_may_use = 0;
4425         else
4426                 data_sinfo->bytes_may_use -= len;
4427         trace_btrfs_space_reservation(fs_info, "space_info",
4428                                       data_sinfo->flags, len, 0);
4429         spin_unlock(&data_sinfo->lock);
4430 }
4431
4432 /*
4433  * Called if we need to clear a data reservation for this inode
4434  * Normally in a error case.
4435  *
4436  * This one will handle the per-inode data rsv map for accurate reserved
4437  * space framework.
4438  */
4439 void btrfs_free_reserved_data_space(struct inode *inode,
4440                         struct extent_changeset *reserved, u64 start, u64 len)
4441 {
4442         struct btrfs_root *root = BTRFS_I(inode)->root;
4443
4444         /* Make sure the range is aligned to sectorsize */
4445         len = round_up(start + len, root->fs_info->sectorsize) -
4446               round_down(start, root->fs_info->sectorsize);
4447         start = round_down(start, root->fs_info->sectorsize);
4448
4449         btrfs_free_reserved_data_space_noquota(inode, start, len);
4450         btrfs_qgroup_free_data(inode, reserved, start, len);
4451 }
4452
4453 static void force_metadata_allocation(struct btrfs_fs_info *info)
4454 {
4455         struct list_head *head = &info->space_info;
4456         struct btrfs_space_info *found;
4457
4458         rcu_read_lock();
4459         list_for_each_entry_rcu(found, head, list) {
4460                 if (found->flags & BTRFS_BLOCK_GROUP_METADATA)
4461                         found->force_alloc = CHUNK_ALLOC_FORCE;
4462         }
4463         rcu_read_unlock();
4464 }
4465
4466 static inline u64 calc_global_rsv_need_space(struct btrfs_block_rsv *global)
4467 {
4468         return (global->size << 1);
4469 }
4470
4471 static int should_alloc_chunk(struct btrfs_fs_info *fs_info,
4472                               struct btrfs_space_info *sinfo, int force)
4473 {
4474         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
4475         u64 bytes_used = btrfs_space_info_used(sinfo, false);
4476         u64 thresh;
4477
4478         if (force == CHUNK_ALLOC_FORCE)
4479                 return 1;
4480
4481         /*
4482          * We need to take into account the global rsv because for all intents
4483          * and purposes it's used space.  Don't worry about locking the
4484          * global_rsv, it doesn't change except when the transaction commits.
4485          */
4486         if (sinfo->flags & BTRFS_BLOCK_GROUP_METADATA)
4487                 bytes_used += calc_global_rsv_need_space(global_rsv);
4488
4489         /*
4490          * in limited mode, we want to have some free space up to
4491          * about 1% of the FS size.
4492          */
4493         if (force == CHUNK_ALLOC_LIMITED) {
4494                 thresh = btrfs_super_total_bytes(fs_info->super_copy);
4495                 thresh = max_t(u64, SZ_64M, div_factor_fine(thresh, 1));
4496
4497                 if (sinfo->total_bytes - bytes_used < thresh)
4498                         return 1;
4499         }
4500
4501         if (bytes_used + SZ_2M < div_factor(sinfo->total_bytes, 8))
4502                 return 0;
4503         return 1;
4504 }
4505
4506 static u64 get_profile_num_devs(struct btrfs_fs_info *fs_info, u64 type)
4507 {
4508         u64 num_dev;
4509
4510         if (type & (BTRFS_BLOCK_GROUP_RAID10 |
4511                     BTRFS_BLOCK_GROUP_RAID0 |
4512                     BTRFS_BLOCK_GROUP_RAID5 |
4513                     BTRFS_BLOCK_GROUP_RAID6))
4514                 num_dev = fs_info->fs_devices->rw_devices;
4515         else if (type & BTRFS_BLOCK_GROUP_RAID1)
4516                 num_dev = 2;
4517         else
4518                 num_dev = 1;    /* DUP or single */
4519
4520         return num_dev;
4521 }
4522
4523 /*
4524  * If @is_allocation is true, reserve space in the system space info necessary
4525  * for allocating a chunk, otherwise if it's false, reserve space necessary for
4526  * removing a chunk.
4527  */
4528 void check_system_chunk(struct btrfs_trans_handle *trans,
4529                         struct btrfs_fs_info *fs_info, u64 type)
4530 {
4531         struct btrfs_space_info *info;
4532         u64 left;
4533         u64 thresh;
4534         int ret = 0;
4535         u64 num_devs;
4536
4537         /*
4538          * Needed because we can end up allocating a system chunk and for an
4539          * atomic and race free space reservation in the chunk block reserve.
4540          */
4541         lockdep_assert_held(&fs_info->chunk_mutex);
4542
4543         info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
4544         spin_lock(&info->lock);
4545         left = info->total_bytes - btrfs_space_info_used(info, true);
4546         spin_unlock(&info->lock);
4547
4548         num_devs = get_profile_num_devs(fs_info, type);
4549
4550         /* num_devs device items to update and 1 chunk item to add or remove */
4551         thresh = btrfs_calc_trunc_metadata_size(fs_info, num_devs) +
4552                 btrfs_calc_trans_metadata_size(fs_info, 1);
4553
4554         if (left < thresh && btrfs_test_opt(fs_info, ENOSPC_DEBUG)) {
4555                 btrfs_info(fs_info, "left=%llu, need=%llu, flags=%llu",
4556                            left, thresh, type);
4557                 dump_space_info(fs_info, info, 0, 0);
4558         }
4559
4560         if (left < thresh) {
4561                 u64 flags = btrfs_system_alloc_profile(fs_info);
4562
4563                 /*
4564                  * Ignore failure to create system chunk. We might end up not
4565                  * needing it, as we might not need to COW all nodes/leafs from
4566                  * the paths we visit in the chunk tree (they were already COWed
4567                  * or created in the current transaction for example).
4568                  */
4569                 ret = btrfs_alloc_chunk(trans, fs_info, flags);
4570         }
4571
4572         if (!ret) {
4573                 ret = btrfs_block_rsv_add(fs_info->chunk_root,
4574                                           &fs_info->chunk_block_rsv,
4575                                           thresh, BTRFS_RESERVE_NO_FLUSH);
4576                 if (!ret)
4577                         trans->chunk_bytes_reserved += thresh;
4578         }
4579 }
4580
4581 /*
4582  * If force is CHUNK_ALLOC_FORCE:
4583  *    - return 1 if it successfully allocates a chunk,
4584  *    - return errors including -ENOSPC otherwise.
4585  * If force is NOT CHUNK_ALLOC_FORCE:
4586  *    - return 0 if it doesn't need to allocate a new chunk,
4587  *    - return 1 if it successfully allocates a chunk,
4588  *    - return errors including -ENOSPC otherwise.
4589  */
4590 static int do_chunk_alloc(struct btrfs_trans_handle *trans,
4591                           struct btrfs_fs_info *fs_info, u64 flags, int force)
4592 {
4593         struct btrfs_space_info *space_info;
4594         int wait_for_alloc = 0;
4595         int ret = 0;
4596
4597         /* Don't re-enter if we're already allocating a chunk */
4598         if (trans->allocating_chunk)
4599                 return -ENOSPC;
4600
4601         space_info = __find_space_info(fs_info, flags);
4602         ASSERT(space_info);
4603
4604 again:
4605         spin_lock(&space_info->lock);
4606         if (force < space_info->force_alloc)
4607                 force = space_info->force_alloc;
4608         if (space_info->full) {
4609                 if (should_alloc_chunk(fs_info, space_info, force))
4610                         ret = -ENOSPC;
4611                 else
4612                         ret = 0;
4613                 spin_unlock(&space_info->lock);
4614                 return ret;
4615         }
4616
4617         if (!should_alloc_chunk(fs_info, space_info, force)) {
4618                 spin_unlock(&space_info->lock);
4619                 return 0;
4620         } else if (space_info->chunk_alloc) {
4621                 wait_for_alloc = 1;
4622         } else {
4623                 space_info->chunk_alloc = 1;
4624         }
4625
4626         spin_unlock(&space_info->lock);
4627
4628         mutex_lock(&fs_info->chunk_mutex);
4629
4630         /*
4631          * The chunk_mutex is held throughout the entirety of a chunk
4632          * allocation, so once we've acquired the chunk_mutex we know that the
4633          * other guy is done and we need to recheck and see if we should
4634          * allocate.
4635          */
4636         if (wait_for_alloc) {
4637                 mutex_unlock(&fs_info->chunk_mutex);
4638                 wait_for_alloc = 0;
4639                 cond_resched();
4640                 goto again;
4641         }
4642
4643         trans->allocating_chunk = true;
4644
4645         /*
4646          * If we have mixed data/metadata chunks we want to make sure we keep
4647          * allocating mixed chunks instead of individual chunks.
4648          */
4649         if (btrfs_mixed_space_info(space_info))
4650                 flags |= (BTRFS_BLOCK_GROUP_DATA | BTRFS_BLOCK_GROUP_METADATA);
4651
4652         /*
4653          * if we're doing a data chunk, go ahead and make sure that
4654          * we keep a reasonable number of metadata chunks allocated in the
4655          * FS as well.
4656          */
4657         if (flags & BTRFS_BLOCK_GROUP_DATA && fs_info->metadata_ratio) {
4658                 fs_info->data_chunk_allocations++;
4659                 if (!(fs_info->data_chunk_allocations %
4660                       fs_info->metadata_ratio))
4661                         force_metadata_allocation(fs_info);
4662         }
4663
4664         /*
4665          * Check if we have enough space in SYSTEM chunk because we may need
4666          * to update devices.
4667          */
4668         check_system_chunk(trans, fs_info, flags);
4669
4670         ret = btrfs_alloc_chunk(trans, fs_info, flags);
4671         trans->allocating_chunk = false;
4672
4673         spin_lock(&space_info->lock);
4674         if (ret < 0 && ret != -ENOSPC)
4675                 goto out;
4676         if (ret)
4677                 space_info->full = 1;
4678         else
4679                 ret = 1;
4680
4681         space_info->force_alloc = CHUNK_ALLOC_NO_FORCE;
4682 out:
4683         space_info->chunk_alloc = 0;
4684         spin_unlock(&space_info->lock);
4685         mutex_unlock(&fs_info->chunk_mutex);
4686         /*
4687          * When we allocate a new chunk we reserve space in the chunk block
4688          * reserve to make sure we can COW nodes/leafs in the chunk tree or
4689          * add new nodes/leafs to it if we end up needing to do it when
4690          * inserting the chunk item and updating device items as part of the
4691          * second phase of chunk allocation, performed by
4692          * btrfs_finish_chunk_alloc(). So make sure we don't accumulate a
4693          * large number of new block groups to create in our transaction
4694          * handle's new_bgs list to avoid exhausting the chunk block reserve
4695          * in extreme cases - like having a single transaction create many new
4696          * block groups when starting to write out the free space caches of all
4697          * the block groups that were made dirty during the lifetime of the
4698          * transaction.
4699          */
4700         if (trans->can_flush_pending_bgs &&
4701             trans->chunk_bytes_reserved >= (u64)SZ_2M) {
4702                 btrfs_create_pending_block_groups(trans);
4703                 btrfs_trans_release_chunk_metadata(trans);
4704         }
4705         return ret;
4706 }
4707
4708 static int can_overcommit(struct btrfs_fs_info *fs_info,
4709                           struct btrfs_space_info *space_info, u64 bytes,
4710                           enum btrfs_reserve_flush_enum flush,
4711                           bool system_chunk)
4712 {
4713         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
4714         u64 profile;
4715         u64 space_size;
4716         u64 avail;
4717         u64 used;
4718
4719         /* Don't overcommit when in mixed mode. */
4720         if (space_info->flags & BTRFS_BLOCK_GROUP_DATA)
4721                 return 0;
4722
4723         if (system_chunk)
4724                 profile = btrfs_system_alloc_profile(fs_info);
4725         else
4726                 profile = btrfs_metadata_alloc_profile(fs_info);
4727
4728         used = btrfs_space_info_used(space_info, false);
4729
4730         /*
4731          * We only want to allow over committing if we have lots of actual space
4732          * free, but if we don't have enough space to handle the global reserve
4733          * space then we could end up having a real enospc problem when trying
4734          * to allocate a chunk or some other such important allocation.
4735          */
4736         spin_lock(&global_rsv->lock);
4737         space_size = calc_global_rsv_need_space(global_rsv);
4738         spin_unlock(&global_rsv->lock);
4739         if (used + space_size >= space_info->total_bytes)
4740                 return 0;
4741
4742         used += space_info->bytes_may_use;
4743
4744         avail = atomic64_read(&fs_info->free_chunk_space);
4745
4746         /*
4747          * If we have dup, raid1 or raid10 then only half of the free
4748          * space is actually useable.  For raid56, the space info used
4749          * doesn't include the parity drive, so we don't have to
4750          * change the math
4751          */
4752         if (profile & (BTRFS_BLOCK_GROUP_DUP |
4753                        BTRFS_BLOCK_GROUP_RAID1 |
4754                        BTRFS_BLOCK_GROUP_RAID10))
4755                 avail >>= 1;
4756
4757         /*
4758          * If we aren't flushing all things, let us overcommit up to
4759          * 1/2th of the space. If we can flush, don't let us overcommit
4760          * too much, let it overcommit up to 1/8 of the space.
4761          */
4762         if (flush == BTRFS_RESERVE_FLUSH_ALL)
4763                 avail >>= 3;
4764         else
4765                 avail >>= 1;
4766
4767         if (used + bytes < space_info->total_bytes + avail)
4768                 return 1;
4769         return 0;
4770 }
4771
4772 static void btrfs_writeback_inodes_sb_nr(struct btrfs_fs_info *fs_info,
4773                                          unsigned long nr_pages, int nr_items)
4774 {
4775         struct super_block *sb = fs_info->sb;
4776
4777         if (down_read_trylock(&sb->s_umount)) {
4778                 writeback_inodes_sb_nr(sb, nr_pages, WB_REASON_FS_FREE_SPACE);
4779                 up_read(&sb->s_umount);
4780         } else {
4781                 /*
4782                  * We needn't worry the filesystem going from r/w to r/o though
4783                  * we don't acquire ->s_umount mutex, because the filesystem
4784                  * should guarantee the delalloc inodes list be empty after
4785                  * the filesystem is readonly(all dirty pages are written to
4786                  * the disk).
4787                  */
4788                 btrfs_start_delalloc_roots(fs_info, 0, nr_items);
4789                 if (!current->journal_info)
4790                         btrfs_wait_ordered_roots(fs_info, nr_items, 0, (u64)-1);
4791         }
4792 }
4793
4794 static inline u64 calc_reclaim_items_nr(struct btrfs_fs_info *fs_info,
4795                                         u64 to_reclaim)
4796 {
4797         u64 bytes;
4798         u64 nr;
4799
4800         bytes = btrfs_calc_trans_metadata_size(fs_info, 1);
4801         nr = div64_u64(to_reclaim, bytes);
4802         if (!nr)
4803                 nr = 1;
4804         return nr;
4805 }
4806
4807 #define EXTENT_SIZE_PER_ITEM    SZ_256K
4808
4809 /*
4810  * shrink metadata reservation for delalloc
4811  */
4812 static void shrink_delalloc(struct btrfs_fs_info *fs_info, u64 to_reclaim,
4813                             u64 orig, bool wait_ordered)
4814 {
4815         struct btrfs_space_info *space_info;
4816         struct btrfs_trans_handle *trans;
4817         u64 delalloc_bytes;
4818         u64 max_reclaim;
4819         u64 items;
4820         long time_left;
4821         unsigned long nr_pages;
4822         int loops;
4823
4824         /* Calc the number of the pages we need flush for space reservation */
4825         items = calc_reclaim_items_nr(fs_info, to_reclaim);
4826         to_reclaim = items * EXTENT_SIZE_PER_ITEM;
4827
4828         trans = (struct btrfs_trans_handle *)current->journal_info;
4829         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
4830
4831         delalloc_bytes = percpu_counter_sum_positive(
4832                                                 &fs_info->delalloc_bytes);
4833         if (delalloc_bytes == 0) {
4834                 if (trans)
4835                         return;
4836                 if (wait_ordered)
4837                         btrfs_wait_ordered_roots(fs_info, items, 0, (u64)-1);
4838                 return;
4839         }
4840
4841         loops = 0;
4842         while (delalloc_bytes && loops < 3) {
4843                 max_reclaim = min(delalloc_bytes, to_reclaim);
4844                 nr_pages = max_reclaim >> PAGE_SHIFT;
4845                 btrfs_writeback_inodes_sb_nr(fs_info, nr_pages, items);
4846                 /*
4847                  * We need to wait for the async pages to actually start before
4848                  * we do anything.
4849                  */
4850                 max_reclaim = atomic_read(&fs_info->async_delalloc_pages);
4851                 if (!max_reclaim)
4852                         goto skip_async;
4853
4854                 if (max_reclaim <= nr_pages)
4855                         max_reclaim = 0;
4856                 else
4857                         max_reclaim -= nr_pages;
4858
4859                 wait_event(fs_info->async_submit_wait,
4860                            atomic_read(&fs_info->async_delalloc_pages) <=
4861                            (int)max_reclaim);
4862 skip_async:
4863                 spin_lock(&space_info->lock);
4864                 if (list_empty(&space_info->tickets) &&
4865                     list_empty(&space_info->priority_tickets)) {
4866                         spin_unlock(&space_info->lock);
4867                         break;
4868                 }
4869                 spin_unlock(&space_info->lock);
4870
4871                 loops++;
4872                 if (wait_ordered && !trans) {
4873                         btrfs_wait_ordered_roots(fs_info, items, 0, (u64)-1);
4874                 } else {
4875                         time_left = schedule_timeout_killable(1);
4876                         if (time_left)
4877                                 break;
4878                 }
4879                 delalloc_bytes = percpu_counter_sum_positive(
4880                                                 &fs_info->delalloc_bytes);
4881         }
4882 }
4883
4884 struct reserve_ticket {
4885         u64 bytes;
4886         int error;
4887         struct list_head list;
4888         wait_queue_head_t wait;
4889 };
4890
4891 /**
4892  * maybe_commit_transaction - possibly commit the transaction if its ok to
4893  * @root - the root we're allocating for
4894  * @bytes - the number of bytes we want to reserve
4895  * @force - force the commit
4896  *
4897  * This will check to make sure that committing the transaction will actually
4898  * get us somewhere and then commit the transaction if it does.  Otherwise it
4899  * will return -ENOSPC.
4900  */
4901 static int may_commit_transaction(struct btrfs_fs_info *fs_info,
4902                                   struct btrfs_space_info *space_info)
4903 {
4904         struct reserve_ticket *ticket = NULL;
4905         struct btrfs_block_rsv *delayed_rsv = &fs_info->delayed_block_rsv;
4906         struct btrfs_trans_handle *trans;
4907         u64 bytes;
4908
4909         trans = (struct btrfs_trans_handle *)current->journal_info;
4910         if (trans)
4911                 return -EAGAIN;
4912
4913         spin_lock(&space_info->lock);
4914         if (!list_empty(&space_info->priority_tickets))
4915                 ticket = list_first_entry(&space_info->priority_tickets,
4916                                           struct reserve_ticket, list);
4917         else if (!list_empty(&space_info->tickets))
4918                 ticket = list_first_entry(&space_info->tickets,
4919                                           struct reserve_ticket, list);
4920         bytes = (ticket) ? ticket->bytes : 0;
4921         spin_unlock(&space_info->lock);
4922
4923         if (!bytes)
4924                 return 0;
4925
4926         /* See if there is enough pinned space to make this reservation */
4927         if (percpu_counter_compare(&space_info->total_bytes_pinned,
4928                                    bytes) >= 0)
4929                 goto commit;
4930
4931         /*
4932          * See if there is some space in the delayed insertion reservation for
4933          * this reservation.
4934          */
4935         if (space_info != delayed_rsv->space_info)
4936                 return -ENOSPC;
4937
4938         spin_lock(&delayed_rsv->lock);
4939         if (delayed_rsv->size > bytes)
4940                 bytes = 0;
4941         else
4942                 bytes -= delayed_rsv->size;
4943         spin_unlock(&delayed_rsv->lock);
4944
4945         if (percpu_counter_compare(&space_info->total_bytes_pinned,
4946                                    bytes) < 0) {
4947                 return -ENOSPC;
4948         }
4949
4950 commit:
4951         trans = btrfs_join_transaction(fs_info->extent_root);
4952         if (IS_ERR(trans))
4953                 return -ENOSPC;
4954
4955         return btrfs_commit_transaction(trans);
4956 }
4957
4958 /*
4959  * Try to flush some data based on policy set by @state. This is only advisory
4960  * and may fail for various reasons. The caller is supposed to examine the
4961  * state of @space_info to detect the outcome.
4962  */
4963 static void flush_space(struct btrfs_fs_info *fs_info,
4964                        struct btrfs_space_info *space_info, u64 num_bytes,
4965                        int state)
4966 {
4967         struct btrfs_root *root = fs_info->extent_root;
4968         struct btrfs_trans_handle *trans;
4969         int nr;
4970         int ret = 0;
4971
4972         switch (state) {
4973         case FLUSH_DELAYED_ITEMS_NR:
4974         case FLUSH_DELAYED_ITEMS:
4975                 if (state == FLUSH_DELAYED_ITEMS_NR)
4976                         nr = calc_reclaim_items_nr(fs_info, num_bytes) * 2;
4977                 else
4978                         nr = -1;
4979
4980                 trans = btrfs_join_transaction(root);
4981                 if (IS_ERR(trans)) {
4982                         ret = PTR_ERR(trans);
4983                         break;
4984                 }
4985                 ret = btrfs_run_delayed_items_nr(trans, nr);
4986                 btrfs_end_transaction(trans);
4987                 break;
4988         case FLUSH_DELALLOC:
4989         case FLUSH_DELALLOC_WAIT:
4990                 shrink_delalloc(fs_info, num_bytes * 2, num_bytes,
4991                                 state == FLUSH_DELALLOC_WAIT);
4992                 break;
4993         case ALLOC_CHUNK:
4994                 trans = btrfs_join_transaction(root);
4995                 if (IS_ERR(trans)) {
4996                         ret = PTR_ERR(trans);
4997                         break;
4998                 }
4999                 ret = do_chunk_alloc(trans, fs_info,
5000                                      btrfs_metadata_alloc_profile(fs_info),
5001                                      CHUNK_ALLOC_NO_FORCE);
5002                 btrfs_end_transaction(trans);
5003                 if (ret > 0 || ret == -ENOSPC)
5004                         ret = 0;
5005                 break;
5006         case COMMIT_TRANS:
5007                 ret = may_commit_transaction(fs_info, space_info);
5008                 break;
5009         default:
5010                 ret = -ENOSPC;
5011                 break;
5012         }
5013
5014         trace_btrfs_flush_space(fs_info, space_info->flags, num_bytes, state,
5015                                 ret);
5016         return;
5017 }
5018
5019 static inline u64
5020 btrfs_calc_reclaim_metadata_size(struct btrfs_fs_info *fs_info,
5021                                  struct btrfs_space_info *space_info,
5022                                  bool system_chunk)
5023 {
5024         struct reserve_ticket *ticket;
5025         u64 used;
5026         u64 expected;
5027         u64 to_reclaim = 0;
5028
5029         list_for_each_entry(ticket, &space_info->tickets, list)
5030                 to_reclaim += ticket->bytes;
5031         list_for_each_entry(ticket, &space_info->priority_tickets, list)
5032                 to_reclaim += ticket->bytes;
5033         if (to_reclaim)
5034                 return to_reclaim;
5035
5036         to_reclaim = min_t(u64, num_online_cpus() * SZ_1M, SZ_16M);
5037         if (can_overcommit(fs_info, space_info, to_reclaim,
5038                            BTRFS_RESERVE_FLUSH_ALL, system_chunk))
5039                 return 0;
5040
5041         used = btrfs_space_info_used(space_info, true);
5042
5043         if (can_overcommit(fs_info, space_info, SZ_1M,
5044                            BTRFS_RESERVE_FLUSH_ALL, system_chunk))
5045                 expected = div_factor_fine(space_info->total_bytes, 95);
5046         else
5047                 expected = div_factor_fine(space_info->total_bytes, 90);
5048
5049         if (used > expected)
5050                 to_reclaim = used - expected;
5051         else
5052                 to_reclaim = 0;
5053         to_reclaim = min(to_reclaim, space_info->bytes_may_use +
5054                                      space_info->bytes_reserved);
5055         return to_reclaim;
5056 }
5057
5058 static inline int need_do_async_reclaim(struct btrfs_fs_info *fs_info,
5059                                         struct btrfs_space_info *space_info,
5060                                         u64 used, bool system_chunk)
5061 {
5062         u64 thresh = div_factor_fine(space_info->total_bytes, 98);
5063
5064         /* If we're just plain full then async reclaim just slows us down. */
5065         if ((space_info->bytes_used + space_info->bytes_reserved) >= thresh)
5066                 return 0;
5067
5068         if (!btrfs_calc_reclaim_metadata_size(fs_info, space_info,
5069                                               system_chunk))
5070                 return 0;
5071
5072         return (used >= thresh && !btrfs_fs_closing(fs_info) &&
5073                 !test_bit(BTRFS_FS_STATE_REMOUNTING, &fs_info->fs_state));
5074 }
5075
5076 static void wake_all_tickets(struct list_head *head)
5077 {
5078         struct reserve_ticket *ticket;
5079
5080         while (!list_empty(head)) {
5081                 ticket = list_first_entry(head, struct reserve_ticket, list);
5082                 list_del_init(&ticket->list);
5083                 ticket->error = -ENOSPC;
5084                 wake_up(&ticket->wait);
5085         }
5086 }
5087
5088 /*
5089  * This is for normal flushers, we can wait all goddamned day if we want to.  We
5090  * will loop and continuously try to flush as long as we are making progress.
5091  * We count progress as clearing off tickets each time we have to loop.
5092  */
5093 static void btrfs_async_reclaim_metadata_space(struct work_struct *work)
5094 {
5095         struct btrfs_fs_info *fs_info;
5096         struct btrfs_space_info *space_info;
5097         u64 to_reclaim;
5098         int flush_state;
5099         int commit_cycles = 0;
5100         u64 last_tickets_id;
5101
5102         fs_info = container_of(work, struct btrfs_fs_info, async_reclaim_work);
5103         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
5104
5105         spin_lock(&space_info->lock);
5106         to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info, space_info,
5107                                                       false);
5108         if (!to_reclaim) {
5109                 space_info->flush = 0;
5110                 spin_unlock(&space_info->lock);
5111                 return;
5112         }
5113         last_tickets_id = space_info->tickets_id;
5114         spin_unlock(&space_info->lock);
5115
5116         flush_state = FLUSH_DELAYED_ITEMS_NR;
5117         do {
5118                 flush_space(fs_info, space_info, to_reclaim, flush_state);
5119                 spin_lock(&space_info->lock);
5120                 if (list_empty(&space_info->tickets)) {
5121                         space_info->flush = 0;
5122                         spin_unlock(&space_info->lock);
5123                         return;
5124                 }
5125                 to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info,
5126                                                               space_info,
5127                                                               false);
5128                 if (last_tickets_id == space_info->tickets_id) {
5129                         flush_state++;
5130                 } else {
5131                         last_tickets_id = space_info->tickets_id;
5132                         flush_state = FLUSH_DELAYED_ITEMS_NR;
5133                         if (commit_cycles)
5134                                 commit_cycles--;
5135                 }
5136
5137                 if (flush_state > COMMIT_TRANS) {
5138                         commit_cycles++;
5139                         if (commit_cycles > 2) {
5140                                 wake_all_tickets(&space_info->tickets);
5141                                 space_info->flush = 0;
5142                         } else {
5143                                 flush_state = FLUSH_DELAYED_ITEMS_NR;
5144                         }
5145                 }
5146                 spin_unlock(&space_info->lock);
5147         } while (flush_state <= COMMIT_TRANS);
5148 }
5149
5150 void btrfs_init_async_reclaim_work(struct work_struct *work)
5151 {
5152         INIT_WORK(work, btrfs_async_reclaim_metadata_space);
5153 }
5154
5155 static void priority_reclaim_metadata_space(struct btrfs_fs_info *fs_info,
5156                                             struct btrfs_space_info *space_info,
5157                                             struct reserve_ticket *ticket)
5158 {
5159         u64 to_reclaim;
5160         int flush_state = FLUSH_DELAYED_ITEMS_NR;
5161
5162         spin_lock(&space_info->lock);
5163         to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info, space_info,
5164                                                       false);
5165         if (!to_reclaim) {
5166                 spin_unlock(&space_info->lock);
5167                 return;
5168         }
5169         spin_unlock(&space_info->lock);
5170
5171         do {
5172                 flush_space(fs_info, space_info, to_reclaim, flush_state);
5173                 flush_state++;
5174                 spin_lock(&space_info->lock);
5175                 if (ticket->bytes == 0) {
5176                         spin_unlock(&space_info->lock);
5177                         return;
5178                 }
5179                 spin_unlock(&space_info->lock);
5180
5181                 /*
5182                  * Priority flushers can't wait on delalloc without
5183                  * deadlocking.
5184                  */
5185                 if (flush_state == FLUSH_DELALLOC ||
5186                     flush_state == FLUSH_DELALLOC_WAIT)
5187                         flush_state = ALLOC_CHUNK;
5188         } while (flush_state < COMMIT_TRANS);
5189 }
5190
5191 static int wait_reserve_ticket(struct btrfs_fs_info *fs_info,
5192                                struct btrfs_space_info *space_info,
5193                                struct reserve_ticket *ticket, u64 orig_bytes)
5194
5195 {
5196         DEFINE_WAIT(wait);
5197         int ret = 0;
5198
5199         spin_lock(&space_info->lock);
5200         while (ticket->bytes > 0 && ticket->error == 0) {
5201                 ret = prepare_to_wait_event(&ticket->wait, &wait, TASK_KILLABLE);
5202                 if (ret) {
5203                         ret = -EINTR;
5204                         break;
5205                 }
5206                 spin_unlock(&space_info->lock);
5207
5208                 schedule();
5209
5210                 finish_wait(&ticket->wait, &wait);
5211                 spin_lock(&space_info->lock);
5212         }
5213         if (!ret)
5214                 ret = ticket->error;
5215         if (!list_empty(&ticket->list))
5216                 list_del_init(&ticket->list);
5217         if (ticket->bytes && ticket->bytes < orig_bytes) {
5218                 u64 num_bytes = orig_bytes - ticket->bytes;
5219                 space_info->bytes_may_use -= num_bytes;
5220                 trace_btrfs_space_reservation(fs_info, "space_info",
5221                                               space_info->flags, num_bytes, 0);
5222         }
5223         spin_unlock(&space_info->lock);
5224
5225         return ret;
5226 }
5227
5228 /**
5229  * reserve_metadata_bytes - try to reserve bytes from the block_rsv's space
5230  * @root - the root we're allocating for
5231  * @space_info - the space info we want to allocate from
5232  * @orig_bytes - the number of bytes we want
5233  * @flush - whether or not we can flush to make our reservation
5234  *
5235  * This will reserve orig_bytes number of bytes from the space info associated
5236  * with the block_rsv.  If there is not enough space it will make an attempt to
5237  * flush out space to make room.  It will do this by flushing delalloc if
5238  * possible or committing the transaction.  If flush is 0 then no attempts to
5239  * regain reservations will be made and this will fail if there is not enough
5240  * space already.
5241  */
5242 static int __reserve_metadata_bytes(struct btrfs_fs_info *fs_info,
5243                                     struct btrfs_space_info *space_info,
5244                                     u64 orig_bytes,
5245                                     enum btrfs_reserve_flush_enum flush,
5246                                     bool system_chunk)
5247 {
5248         struct reserve_ticket ticket;
5249         u64 used;
5250         int ret = 0;
5251
5252         ASSERT(orig_bytes);
5253         ASSERT(!current->journal_info || flush != BTRFS_RESERVE_FLUSH_ALL);
5254
5255         spin_lock(&space_info->lock);
5256         ret = -ENOSPC;
5257         used = btrfs_space_info_used(space_info, true);
5258
5259         /*
5260          * If we have enough space then hooray, make our reservation and carry
5261          * on.  If not see if we can overcommit, and if we can, hooray carry on.
5262          * If not things get more complicated.
5263          */
5264         if (used + orig_bytes <= space_info->total_bytes) {
5265                 space_info->bytes_may_use += orig_bytes;
5266                 trace_btrfs_space_reservation(fs_info, "space_info",
5267                                               space_info->flags, orig_bytes, 1);
5268                 ret = 0;
5269         } else if (can_overcommit(fs_info, space_info, orig_bytes, flush,
5270                                   system_chunk)) {
5271                 space_info->bytes_may_use += orig_bytes;
5272                 trace_btrfs_space_reservation(fs_info, "space_info",
5273                                               space_info->flags, orig_bytes, 1);
5274                 ret = 0;
5275         }
5276
5277         /*
5278          * If we couldn't make a reservation then setup our reservation ticket
5279          * and kick the async worker if it's not already running.
5280          *
5281          * If we are a priority flusher then we just need to add our ticket to
5282          * the list and we will do our own flushing further down.
5283          */
5284         if (ret && flush != BTRFS_RESERVE_NO_FLUSH) {
5285                 ticket.bytes = orig_bytes;
5286                 ticket.error = 0;
5287                 init_waitqueue_head(&ticket.wait);
5288                 if (flush == BTRFS_RESERVE_FLUSH_ALL) {
5289                         list_add_tail(&ticket.list, &space_info->tickets);
5290                         if (!space_info->flush) {
5291                                 space_info->flush = 1;
5292                                 trace_btrfs_trigger_flush(fs_info,
5293                                                           space_info->flags,
5294                                                           orig_bytes, flush,
5295                                                           "enospc");
5296                                 queue_work(system_unbound_wq,
5297                                            &fs_info->async_reclaim_work);
5298                         }
5299                 } else {
5300                         list_add_tail(&ticket.list,
5301                                       &space_info->priority_tickets);
5302                 }
5303         } else if (!ret && space_info->flags & BTRFS_BLOCK_GROUP_METADATA) {
5304                 used += orig_bytes;
5305                 /*
5306                  * We will do the space reservation dance during log replay,
5307                  * which means we won't have fs_info->fs_root set, so don't do
5308                  * the async reclaim as we will panic.
5309                  */
5310                 if (!test_bit(BTRFS_FS_LOG_RECOVERING, &fs_info->flags) &&
5311                     need_do_async_reclaim(fs_info, space_info,
5312                                           used, system_chunk) &&
5313                     !work_busy(&fs_info->async_reclaim_work)) {
5314                         trace_btrfs_trigger_flush(fs_info, space_info->flags,
5315                                                   orig_bytes, flush, "preempt");
5316                         queue_work(system_unbound_wq,
5317                                    &fs_info->async_reclaim_work);
5318                 }
5319         }
5320         spin_unlock(&space_info->lock);
5321         if (!ret || flush == BTRFS_RESERVE_NO_FLUSH)
5322                 return ret;
5323
5324         if (flush == BTRFS_RESERVE_FLUSH_ALL)
5325                 return wait_reserve_ticket(fs_info, space_info, &ticket,
5326                                            orig_bytes);
5327
5328         ret = 0;
5329         priority_reclaim_metadata_space(fs_info, space_info, &ticket);
5330         spin_lock(&space_info->lock);
5331         if (ticket.bytes) {
5332                 if (ticket.bytes < orig_bytes) {
5333                         u64 num_bytes = orig_bytes - ticket.bytes;
5334                         space_info->bytes_may_use -= num_bytes;
5335                         trace_btrfs_space_reservation(fs_info, "space_info",
5336                                                       space_info->flags,
5337                                                       num_bytes, 0);
5338
5339                 }
5340                 list_del_init(&ticket.list);
5341                 ret = -ENOSPC;
5342         }
5343         spin_unlock(&space_info->lock);
5344         ASSERT(list_empty(&ticket.list));
5345         return ret;
5346 }
5347
5348 /**
5349  * reserve_metadata_bytes - try to reserve bytes from the block_rsv's space
5350  * @root - the root we're allocating for
5351  * @block_rsv - the block_rsv we're allocating for
5352  * @orig_bytes - the number of bytes we want
5353  * @flush - whether or not we can flush to make our reservation
5354  *
5355  * This will reserve orgi_bytes number of bytes from the space info associated
5356  * with the block_rsv.  If there is not enough space it will make an attempt to
5357  * flush out space to make room.  It will do this by flushing delalloc if
5358  * possible or committing the transaction.  If flush is 0 then no attempts to
5359  * regain reservations will be made and this will fail if there is not enough
5360  * space already.
5361  */
5362 static int reserve_metadata_bytes(struct btrfs_root *root,
5363                                   struct btrfs_block_rsv *block_rsv,
5364                                   u64 orig_bytes,
5365                                   enum btrfs_reserve_flush_enum flush)
5366 {
5367         struct btrfs_fs_info *fs_info = root->fs_info;
5368         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
5369         int ret;
5370         bool system_chunk = (root == fs_info->chunk_root);
5371
5372         ret = __reserve_metadata_bytes(fs_info, block_rsv->space_info,
5373                                        orig_bytes, flush, system_chunk);
5374         if (ret == -ENOSPC &&
5375             unlikely(root->orphan_cleanup_state == ORPHAN_CLEANUP_STARTED)) {
5376                 if (block_rsv != global_rsv &&
5377                     !block_rsv_use_bytes(global_rsv, orig_bytes))
5378                         ret = 0;
5379         }
5380         if (ret == -ENOSPC) {
5381                 trace_btrfs_space_reservation(fs_info, "space_info:enospc",
5382                                               block_rsv->space_info->flags,
5383                                               orig_bytes, 1);
5384
5385                 if (btrfs_test_opt(fs_info, ENOSPC_DEBUG))
5386                         dump_space_info(fs_info, block_rsv->space_info,
5387                                         orig_bytes, 0);
5388         }
5389         return ret;
5390 }
5391
5392 static struct btrfs_block_rsv *get_block_rsv(
5393                                         const struct btrfs_trans_handle *trans,
5394                                         const struct btrfs_root *root)
5395 {
5396         struct btrfs_fs_info *fs_info = root->fs_info;
5397         struct btrfs_block_rsv *block_rsv = NULL;
5398
5399         if (test_bit(BTRFS_ROOT_REF_COWS, &root->state) ||
5400             (root == fs_info->csum_root && trans->adding_csums) ||
5401             (root == fs_info->uuid_root))
5402                 block_rsv = trans->block_rsv;
5403
5404         if (!block_rsv)
5405                 block_rsv = root->block_rsv;
5406
5407         if (!block_rsv)
5408                 block_rsv = &fs_info->empty_block_rsv;
5409
5410         return block_rsv;
5411 }
5412
5413 static int block_rsv_use_bytes(struct btrfs_block_rsv *block_rsv,
5414                                u64 num_bytes)
5415 {
5416         int ret = -ENOSPC;
5417         spin_lock(&block_rsv->lock);
5418         if (block_rsv->reserved >= num_bytes) {
5419                 block_rsv->reserved -= num_bytes;
5420                 if (block_rsv->reserved < block_rsv->size)
5421                         block_rsv->full = 0;
5422                 ret = 0;
5423         }
5424         spin_unlock(&block_rsv->lock);
5425         return ret;
5426 }
5427
5428 static void block_rsv_add_bytes(struct btrfs_block_rsv *block_rsv,
5429                                 u64 num_bytes, int update_size)
5430 {
5431         spin_lock(&block_rsv->lock);
5432         block_rsv->reserved += num_bytes;
5433         if (update_size)
5434                 block_rsv->size += num_bytes;
5435         else if (block_rsv->reserved >= block_rsv->size)
5436                 block_rsv->full = 1;
5437         spin_unlock(&block_rsv->lock);
5438 }
5439
5440 int btrfs_cond_migrate_bytes(struct btrfs_fs_info *fs_info,
5441                              struct btrfs_block_rsv *dest, u64 num_bytes,
5442                              int min_factor)
5443 {
5444         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
5445         u64 min_bytes;
5446
5447         if (global_rsv->space_info != dest->space_info)
5448                 return -ENOSPC;
5449
5450         spin_lock(&global_rsv->lock);
5451         min_bytes = div_factor(global_rsv->size, min_factor);
5452         if (global_rsv->reserved < min_bytes + num_bytes) {
5453                 spin_unlock(&global_rsv->lock);
5454                 return -ENOSPC;
5455         }
5456         global_rsv->reserved -= num_bytes;
5457         if (global_rsv->reserved < global_rsv->size)
5458                 global_rsv->full = 0;
5459         spin_unlock(&global_rsv->lock);
5460
5461         block_rsv_add_bytes(dest, num_bytes, 1);
5462         return 0;
5463 }
5464
5465 /*
5466  * This is for space we already have accounted in space_info->bytes_may_use, so
5467  * basically when we're returning space from block_rsv's.
5468  */
5469 static void space_info_add_old_bytes(struct btrfs_fs_info *fs_info,
5470                                      struct btrfs_space_info *space_info,
5471                                      u64 num_bytes)
5472 {
5473         struct reserve_ticket *ticket;
5474         struct list_head *head;
5475         u64 used;
5476         enum btrfs_reserve_flush_enum flush = BTRFS_RESERVE_NO_FLUSH;
5477         bool check_overcommit = false;
5478
5479         spin_lock(&space_info->lock);
5480         head = &space_info->priority_tickets;
5481
5482         /*
5483          * If we are over our limit then we need to check and see if we can
5484          * overcommit, and if we can't then we just need to free up our space
5485          * and not satisfy any requests.
5486          */
5487         used = btrfs_space_info_used(space_info, true);
5488         if (used - num_bytes >= space_info->total_bytes)
5489                 check_overcommit = true;
5490 again:
5491         while (!list_empty(head) && num_bytes) {
5492                 ticket = list_first_entry(head, struct reserve_ticket,
5493                                           list);
5494                 /*
5495                  * We use 0 bytes because this space is already reserved, so
5496                  * adding the ticket space would be a double count.
5497                  */
5498                 if (check_overcommit &&
5499                     !can_overcommit(fs_info, space_info, 0, flush, false))
5500                         break;
5501                 if (num_bytes >= ticket->bytes) {
5502                         list_del_init(&ticket->list);
5503                         num_bytes -= ticket->bytes;
5504                         ticket->bytes = 0;
5505                         space_info->tickets_id++;
5506                         wake_up(&ticket->wait);
5507                 } else {
5508                         ticket->bytes -= num_bytes;
5509                         num_bytes = 0;
5510                 }
5511         }
5512
5513         if (num_bytes && head == &space_info->priority_tickets) {
5514                 head = &space_info->tickets;
5515                 flush = BTRFS_RESERVE_FLUSH_ALL;
5516                 goto again;
5517         }
5518         space_info->bytes_may_use -= num_bytes;
5519         trace_btrfs_space_reservation(fs_info, "space_info",
5520                                       space_info->flags, num_bytes, 0);
5521         spin_unlock(&space_info->lock);
5522 }
5523
5524 /*
5525  * This is for newly allocated space that isn't accounted in
5526  * space_info->bytes_may_use yet.  So if we allocate a chunk or unpin an extent
5527  * we use this helper.
5528  */
5529 static void space_info_add_new_bytes(struct btrfs_fs_info *fs_info,
5530                                      struct btrfs_space_info *space_info,
5531                                      u64 num_bytes)
5532 {
5533         struct reserve_ticket *ticket;
5534         struct list_head *head = &space_info->priority_tickets;
5535
5536 again:
5537         while (!list_empty(head) && num_bytes) {
5538                 ticket = list_first_entry(head, struct reserve_ticket,
5539                                           list);
5540                 if (num_bytes >= ticket->bytes) {
5541                         trace_btrfs_space_reservation(fs_info, "space_info",
5542                                                       space_info->flags,
5543                                                       ticket->bytes, 1);
5544                         list_del_init(&ticket->list);
5545                         num_bytes -= ticket->bytes;
5546                         space_info->bytes_may_use += ticket->bytes;
5547                         ticket->bytes = 0;
5548                         space_info->tickets_id++;
5549                         wake_up(&ticket->wait);
5550                 } else {
5551                         trace_btrfs_space_reservation(fs_info, "space_info",
5552                                                       space_info->flags,
5553                                                       num_bytes, 1);
5554                         space_info->bytes_may_use += num_bytes;
5555                         ticket->bytes -= num_bytes;
5556                         num_bytes = 0;
5557                 }
5558         }
5559
5560         if (num_bytes && head == &space_info->priority_tickets) {
5561                 head = &space_info->tickets;
5562                 goto again;
5563         }
5564 }
5565
5566 static u64 block_rsv_release_bytes(struct btrfs_fs_info *fs_info,
5567                                     struct btrfs_block_rsv *block_rsv,
5568                                     struct btrfs_block_rsv *dest, u64 num_bytes,
5569                                     u64 *qgroup_to_release_ret)
5570 {
5571         struct btrfs_space_info *space_info = block_rsv->space_info;
5572         u64 qgroup_to_release = 0;
5573         u64 ret;
5574
5575         spin_lock(&block_rsv->lock);
5576         if (num_bytes == (u64)-1) {
5577                 num_bytes = block_rsv->size;
5578                 qgroup_to_release = block_rsv->qgroup_rsv_size;
5579         }
5580         block_rsv->size -= num_bytes;
5581         if (block_rsv->reserved >= block_rsv->size) {
5582                 num_bytes = block_rsv->reserved - block_rsv->size;
5583                 block_rsv->reserved = block_rsv->size;
5584                 block_rsv->full = 1;
5585         } else {
5586                 num_bytes = 0;
5587         }
5588         if (block_rsv->qgroup_rsv_reserved >= block_rsv->qgroup_rsv_size) {
5589                 qgroup_to_release = block_rsv->qgroup_rsv_reserved -
5590                                     block_rsv->qgroup_rsv_size;
5591                 block_rsv->qgroup_rsv_reserved = block_rsv->qgroup_rsv_size;
5592         } else {
5593                 qgroup_to_release = 0;
5594         }
5595         spin_unlock(&block_rsv->lock);
5596
5597         ret = num_bytes;
5598         if (num_bytes > 0) {
5599                 if (dest) {
5600                         spin_lock(&dest->lock);
5601                         if (!dest->full) {
5602                                 u64 bytes_to_add;
5603
5604                                 bytes_to_add = dest->size - dest->reserved;
5605                                 bytes_to_add = min(num_bytes, bytes_to_add);
5606                                 dest->reserved += bytes_to_add;
5607                                 if (dest->reserved >= dest->size)
5608                                         dest->full = 1;
5609                                 num_bytes -= bytes_to_add;
5610                         }
5611                         spin_unlock(&dest->lock);
5612                 }
5613                 if (num_bytes)
5614                         space_info_add_old_bytes(fs_info, space_info,
5615                                                  num_bytes);
5616         }
5617         if (qgroup_to_release_ret)
5618                 *qgroup_to_release_ret = qgroup_to_release;
5619         return ret;
5620 }
5621
5622 int btrfs_block_rsv_migrate(struct btrfs_block_rsv *src,
5623                             struct btrfs_block_rsv *dst, u64 num_bytes,
5624                             int update_size)
5625 {
5626         int ret;
5627
5628         ret = block_rsv_use_bytes(src, num_bytes);
5629         if (ret)
5630                 return ret;
5631
5632         block_rsv_add_bytes(dst, num_bytes, update_size);
5633         return 0;
5634 }
5635
5636 void btrfs_init_block_rsv(struct btrfs_block_rsv *rsv, unsigned short type)
5637 {
5638         memset(rsv, 0, sizeof(*rsv));
5639         spin_lock_init(&rsv->lock);
5640         rsv->type = type;
5641 }
5642
5643 void btrfs_init_metadata_block_rsv(struct btrfs_fs_info *fs_info,
5644                                    struct btrfs_block_rsv *rsv,
5645                                    unsigned short type)
5646 {
5647         btrfs_init_block_rsv(rsv, type);
5648         rsv->space_info = __find_space_info(fs_info,
5649                                             BTRFS_BLOCK_GROUP_METADATA);
5650 }
5651
5652 struct btrfs_block_rsv *btrfs_alloc_block_rsv(struct btrfs_fs_info *fs_info,
5653                                               unsigned short type)
5654 {
5655         struct btrfs_block_rsv *block_rsv;
5656
5657         block_rsv = kmalloc(sizeof(*block_rsv), GFP_NOFS);
5658         if (!block_rsv)
5659                 return NULL;
5660
5661         btrfs_init_metadata_block_rsv(fs_info, block_rsv, type);
5662         return block_rsv;
5663 }
5664
5665 void btrfs_free_block_rsv(struct btrfs_fs_info *fs_info,
5666                           struct btrfs_block_rsv *rsv)
5667 {
5668         if (!rsv)
5669                 return;
5670         btrfs_block_rsv_release(fs_info, rsv, (u64)-1);
5671         kfree(rsv);
5672 }
5673
5674 void __btrfs_free_block_rsv(struct btrfs_block_rsv *rsv)
5675 {
5676         kfree(rsv);
5677 }
5678
5679 int btrfs_block_rsv_add(struct btrfs_root *root,
5680                         struct btrfs_block_rsv *block_rsv, u64 num_bytes,
5681                         enum btrfs_reserve_flush_enum flush)
5682 {
5683         int ret;
5684
5685         if (num_bytes == 0)
5686                 return 0;
5687
5688         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
5689         if (!ret) {
5690                 block_rsv_add_bytes(block_rsv, num_bytes, 1);
5691                 return 0;
5692         }
5693
5694         return ret;
5695 }
5696
5697 int btrfs_block_rsv_check(struct btrfs_block_rsv *block_rsv, int min_factor)
5698 {
5699         u64 num_bytes = 0;
5700         int ret = -ENOSPC;
5701
5702         if (!block_rsv)
5703                 return 0;
5704
5705         spin_lock(&block_rsv->lock);
5706         num_bytes = div_factor(block_rsv->size, min_factor);
5707         if (block_rsv->reserved >= num_bytes)
5708                 ret = 0;
5709         spin_unlock(&block_rsv->lock);
5710
5711         return ret;
5712 }
5713
5714 int btrfs_block_rsv_refill(struct btrfs_root *root,
5715                            struct btrfs_block_rsv *block_rsv, u64 min_reserved,
5716                            enum btrfs_reserve_flush_enum flush)
5717 {
5718         u64 num_bytes = 0;
5719         int ret = -ENOSPC;
5720
5721         if (!block_rsv)
5722                 return 0;
5723
5724         spin_lock(&block_rsv->lock);
5725         num_bytes = min_reserved;
5726         if (block_rsv->reserved >= num_bytes)
5727                 ret = 0;
5728         else
5729                 num_bytes -= block_rsv->reserved;
5730         spin_unlock(&block_rsv->lock);
5731
5732         if (!ret)
5733                 return 0;
5734
5735         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
5736         if (!ret) {
5737                 block_rsv_add_bytes(block_rsv, num_bytes, 0);
5738                 return 0;
5739         }
5740
5741         return ret;
5742 }
5743
5744 /**
5745  * btrfs_inode_rsv_refill - refill the inode block rsv.
5746  * @inode - the inode we are refilling.
5747  * @flush - the flusing restriction.
5748  *
5749  * Essentially the same as btrfs_block_rsv_refill, except it uses the
5750  * block_rsv->size as the minimum size.  We'll either refill the missing amount
5751  * or return if we already have enough space.  This will also handle the resreve
5752  * tracepoint for the reserved amount.
5753  */
5754 static int btrfs_inode_rsv_refill(struct btrfs_inode *inode,
5755                                   enum btrfs_reserve_flush_enum flush)
5756 {
5757         struct btrfs_root *root = inode->root;
5758         struct btrfs_block_rsv *block_rsv = &inode->block_rsv;
5759         u64 num_bytes = 0;
5760         u64 qgroup_num_bytes = 0;
5761         int ret = -ENOSPC;
5762
5763         spin_lock(&block_rsv->lock);
5764         if (block_rsv->reserved < block_rsv->size)
5765                 num_bytes = block_rsv->size - block_rsv->reserved;
5766         if (block_rsv->qgroup_rsv_reserved < block_rsv->qgroup_rsv_size)
5767                 qgroup_num_bytes = block_rsv->qgroup_rsv_size -
5768                                    block_rsv->qgroup_rsv_reserved;
5769         spin_unlock(&block_rsv->lock);
5770
5771         if (num_bytes == 0)
5772                 return 0;
5773
5774         ret = btrfs_qgroup_reserve_meta_prealloc(root, qgroup_num_bytes, true);
5775         if (ret)
5776                 return ret;
5777         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
5778         if (!ret) {
5779                 block_rsv_add_bytes(block_rsv, num_bytes, 0);
5780                 trace_btrfs_space_reservation(root->fs_info, "delalloc",
5781                                               btrfs_ino(inode), num_bytes, 1);
5782
5783                 /* Don't forget to increase qgroup_rsv_reserved */
5784                 spin_lock(&block_rsv->lock);
5785                 block_rsv->qgroup_rsv_reserved += qgroup_num_bytes;
5786                 spin_unlock(&block_rsv->lock);
5787         } else
5788                 btrfs_qgroup_free_meta_prealloc(root, qgroup_num_bytes);
5789         return ret;
5790 }
5791
5792 /**
5793  * btrfs_inode_rsv_release - release any excessive reservation.
5794  * @inode - the inode we need to release from.
5795  * @qgroup_free - free or convert qgroup meta.
5796  *   Unlike normal operation, qgroup meta reservation needs to know if we are
5797  *   freeing qgroup reservation or just converting it into per-trans.  Normally
5798  *   @qgroup_free is true for error handling, and false for normal release.
5799  *
5800  * This is the same as btrfs_block_rsv_release, except that it handles the
5801  * tracepoint for the reservation.
5802  */
5803 static void btrfs_inode_rsv_release(struct btrfs_inode *inode, bool qgroup_free)
5804 {
5805         struct btrfs_fs_info *fs_info = inode->root->fs_info;
5806         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
5807         struct btrfs_block_rsv *block_rsv = &inode->block_rsv;
5808         u64 released = 0;
5809         u64 qgroup_to_release = 0;
5810
5811         /*
5812          * Since we statically set the block_rsv->size we just want to say we
5813          * are releasing 0 bytes, and then we'll just get the reservation over
5814          * the size free'd.
5815          */
5816         released = block_rsv_release_bytes(fs_info, block_rsv, global_rsv, 0,
5817                                            &qgroup_to_release);
5818         if (released > 0)
5819                 trace_btrfs_space_reservation(fs_info, "delalloc",
5820                                               btrfs_ino(inode), released, 0);
5821         if (qgroup_free)
5822                 btrfs_qgroup_free_meta_prealloc(inode->root, qgroup_to_release);
5823         else
5824                 btrfs_qgroup_convert_reserved_meta(inode->root,
5825                                                    qgroup_to_release);
5826 }
5827
5828 void btrfs_block_rsv_release(struct btrfs_fs_info *fs_info,
5829                              struct btrfs_block_rsv *block_rsv,
5830                              u64 num_bytes)
5831 {
5832         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
5833
5834         if (global_rsv == block_rsv ||
5835             block_rsv->space_info != global_rsv->space_info)
5836                 global_rsv = NULL;
5837         block_rsv_release_bytes(fs_info, block_rsv, global_rsv, num_bytes, NULL);
5838 }
5839
5840 static void update_global_block_rsv(struct btrfs_fs_info *fs_info)
5841 {
5842         struct btrfs_block_rsv *block_rsv = &fs_info->global_block_rsv;
5843         struct btrfs_space_info *sinfo = block_rsv->space_info;
5844         u64 num_bytes;
5845
5846         /*
5847          * The global block rsv is based on the size of the extent tree, the
5848          * checksum tree and the root tree.  If the fs is empty we want to set
5849          * it to a minimal amount for safety.
5850          */
5851         num_bytes = btrfs_root_used(&fs_info->extent_root->root_item) +
5852                 btrfs_root_used(&fs_info->csum_root->root_item) +
5853                 btrfs_root_used(&fs_info->tree_root->root_item);
5854         num_bytes = max_t(u64, num_bytes, SZ_16M);
5855
5856         spin_lock(&sinfo->lock);
5857         spin_lock(&block_rsv->lock);
5858
5859         block_rsv->size = min_t(u64, num_bytes, SZ_512M);
5860
5861         if (block_rsv->reserved < block_rsv->size) {
5862                 num_bytes = btrfs_space_info_used(sinfo, true);
5863                 if (sinfo->total_bytes > num_bytes) {
5864                         num_bytes = sinfo->total_bytes - num_bytes;
5865                         num_bytes = min(num_bytes,
5866                                         block_rsv->size - block_rsv->reserved);
5867                         block_rsv->reserved += num_bytes;
5868                         sinfo->bytes_may_use += num_bytes;
5869                         trace_btrfs_space_reservation(fs_info, "space_info",
5870                                                       sinfo->flags, num_bytes,
5871                                                       1);
5872                 }
5873         } else if (block_rsv->reserved > block_rsv->size) {
5874                 num_bytes = block_rsv->reserved - block_rsv->size;
5875                 sinfo->bytes_may_use -= num_bytes;
5876                 trace_btrfs_space_reservation(fs_info, "space_info",
5877                                       sinfo->flags, num_bytes, 0);
5878                 block_rsv->reserved = block_rsv->size;
5879         }
5880
5881         if (block_rsv->reserved == block_rsv->size)
5882                 block_rsv->full = 1;
5883         else
5884                 block_rsv->full = 0;
5885
5886         spin_unlock(&block_rsv->lock);
5887         spin_unlock(&sinfo->lock);
5888 }
5889
5890 static void init_global_block_rsv(struct btrfs_fs_info *fs_info)
5891 {
5892         struct btrfs_space_info *space_info;
5893
5894         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
5895         fs_info->chunk_block_rsv.space_info = space_info;
5896
5897         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
5898         fs_info->global_block_rsv.space_info = space_info;
5899         fs_info->trans_block_rsv.space_info = space_info;
5900         fs_info->empty_block_rsv.space_info = space_info;
5901         fs_info->delayed_block_rsv.space_info = space_info;
5902
5903         fs_info->extent_root->block_rsv = &fs_info->global_block_rsv;
5904         fs_info->csum_root->block_rsv = &fs_info->global_block_rsv;
5905         fs_info->dev_root->block_rsv = &fs_info->global_block_rsv;
5906         fs_info->tree_root->block_rsv = &fs_info->global_block_rsv;
5907         if (fs_info->quota_root)
5908                 fs_info->quota_root->block_rsv = &fs_info->global_block_rsv;
5909         fs_info->chunk_root->block_rsv = &fs_info->chunk_block_rsv;
5910
5911         update_global_block_rsv(fs_info);
5912 }
5913
5914 static void release_global_block_rsv(struct btrfs_fs_info *fs_info)
5915 {
5916         block_rsv_release_bytes(fs_info, &fs_info->global_block_rsv, NULL,
5917                                 (u64)-1, NULL);
5918         WARN_ON(fs_info->trans_block_rsv.size > 0);
5919         WARN_ON(fs_info->trans_block_rsv.reserved > 0);
5920         WARN_ON(fs_info->chunk_block_rsv.size > 0);
5921         WARN_ON(fs_info->chunk_block_rsv.reserved > 0);
5922         WARN_ON(fs_info->delayed_block_rsv.size > 0);
5923         WARN_ON(fs_info->delayed_block_rsv.reserved > 0);
5924 }
5925
5926
5927 /*
5928  * To be called after all the new block groups attached to the transaction
5929  * handle have been created (btrfs_create_pending_block_groups()).
5930  */
5931 void btrfs_trans_release_chunk_metadata(struct btrfs_trans_handle *trans)
5932 {
5933         struct btrfs_fs_info *fs_info = trans->fs_info;
5934
5935         if (!trans->chunk_bytes_reserved)
5936                 return;
5937
5938         WARN_ON_ONCE(!list_empty(&trans->new_bgs));
5939
5940         block_rsv_release_bytes(fs_info, &fs_info->chunk_block_rsv, NULL,
5941                                 trans->chunk_bytes_reserved, NULL);
5942         trans->chunk_bytes_reserved = 0;
5943 }
5944
5945 /* Can only return 0 or -ENOSPC */
5946 int btrfs_orphan_reserve_metadata(struct btrfs_trans_handle *trans,
5947                                   struct btrfs_inode *inode)
5948 {
5949         struct btrfs_fs_info *fs_info = btrfs_sb(inode->vfs_inode.i_sb);
5950         struct btrfs_root *root = inode->root;
5951         /*
5952          * We always use trans->block_rsv here as we will have reserved space
5953          * for our orphan when starting the transaction, using get_block_rsv()
5954          * here will sometimes make us choose the wrong block rsv as we could be
5955          * doing a reloc inode for a non refcounted root.
5956          */
5957         struct btrfs_block_rsv *src_rsv = trans->block_rsv;
5958         struct btrfs_block_rsv *dst_rsv = root->orphan_block_rsv;
5959
5960         /*
5961          * We need to hold space in order to delete our orphan item once we've
5962          * added it, so this takes the reservation so we can release it later
5963          * when we are truly done with the orphan item.
5964          */
5965         u64 num_bytes = btrfs_calc_trans_metadata_size(fs_info, 1);
5966
5967         trace_btrfs_space_reservation(fs_info, "orphan", btrfs_ino(inode),
5968                         num_bytes, 1);
5969         return btrfs_block_rsv_migrate(src_rsv, dst_rsv, num_bytes, 1);
5970 }
5971
5972 void btrfs_orphan_release_metadata(struct btrfs_inode *inode)
5973 {
5974         struct btrfs_fs_info *fs_info = btrfs_sb(inode->vfs_inode.i_sb);
5975         struct btrfs_root *root = inode->root;
5976         u64 num_bytes = btrfs_calc_trans_metadata_size(fs_info, 1);
5977
5978         trace_btrfs_space_reservation(fs_info, "orphan", btrfs_ino(inode),
5979                         num_bytes, 0);
5980         btrfs_block_rsv_release(fs_info, root->orphan_block_rsv, num_bytes);
5981 }
5982
5983 /*
5984  * btrfs_subvolume_reserve_metadata() - reserve space for subvolume operation
5985  * root: the root of the parent directory
5986  * rsv: block reservation
5987  * items: the number of items that we need do reservation
5988  * qgroup_reserved: used to return the reserved size in qgroup
5989  *
5990  * This function is used to reserve the space for snapshot/subvolume
5991  * creation and deletion. Those operations are different with the
5992  * common file/directory operations, they change two fs/file trees
5993  * and root tree, the number of items that the qgroup reserves is
5994  * different with the free space reservation. So we can not use
5995  * the space reservation mechanism in start_transaction().
5996  */
5997 int btrfs_subvolume_reserve_metadata(struct btrfs_root *root,
5998                                      struct btrfs_block_rsv *rsv,
5999                                      int items,
6000                                      u64 *qgroup_reserved,
6001                                      bool use_global_rsv)
6002 {
6003         u64 num_bytes;
6004         int ret;
6005         struct btrfs_fs_info *fs_info = root->fs_info;
6006         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
6007
6008         if (test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags)) {
6009                 /* One for parent inode, two for dir entries */
6010                 num_bytes = 3 * fs_info->nodesize;
6011                 ret = btrfs_qgroup_reserve_meta_prealloc(root, num_bytes, true);
6012                 if (ret)
6013                         return ret;
6014         } else {
6015                 num_bytes = 0;
6016         }
6017
6018         *qgroup_reserved = num_bytes;
6019
6020         num_bytes = btrfs_calc_trans_metadata_size(fs_info, items);
6021         rsv->space_info = __find_space_info(fs_info,
6022                                             BTRFS_BLOCK_GROUP_METADATA);
6023         ret = btrfs_block_rsv_add(root, rsv, num_bytes,
6024                                   BTRFS_RESERVE_FLUSH_ALL);
6025
6026         if (ret == -ENOSPC && use_global_rsv)
6027                 ret = btrfs_block_rsv_migrate(global_rsv, rsv, num_bytes, 1);
6028
6029         if (ret && *qgroup_reserved)
6030                 btrfs_qgroup_free_meta_prealloc(root, *qgroup_reserved);
6031
6032         return ret;
6033 }
6034
6035 void btrfs_subvolume_release_metadata(struct btrfs_fs_info *fs_info,
6036                                       struct btrfs_block_rsv *rsv)
6037 {
6038         btrfs_block_rsv_release(fs_info, rsv, (u64)-1);
6039 }
6040
6041 static void btrfs_calculate_inode_block_rsv_size(struct btrfs_fs_info *fs_info,
6042                                                  struct btrfs_inode *inode)
6043 {
6044         struct btrfs_block_rsv *block_rsv = &inode->block_rsv;
6045         u64 reserve_size = 0;
6046         u64 qgroup_rsv_size = 0;
6047         u64 csum_leaves;
6048         unsigned outstanding_extents;
6049
6050         lockdep_assert_held(&inode->lock);
6051         outstanding_extents = inode->outstanding_extents;
6052         if (outstanding_extents)
6053                 reserve_size = btrfs_calc_trans_metadata_size(fs_info,
6054                                                 outstanding_extents + 1);
6055         csum_leaves = btrfs_csum_bytes_to_leaves(fs_info,
6056                                                  inode->csum_bytes);
6057         reserve_size += btrfs_calc_trans_metadata_size(fs_info,
6058                                                        csum_leaves);
6059         /*
6060          * For qgroup rsv, the calculation is very simple:
6061          * account one nodesize for each outstanding extent
6062          *
6063          * This is overestimating in most cases.
6064          */
6065         qgroup_rsv_size = outstanding_extents * fs_info->nodesize;
6066
6067         spin_lock(&block_rsv->lock);
6068         block_rsv->size = reserve_size;
6069         block_rsv->qgroup_rsv_size = qgroup_rsv_size;
6070         spin_unlock(&block_rsv->lock);
6071 }
6072
6073 int btrfs_delalloc_reserve_metadata(struct btrfs_inode *inode, u64 num_bytes)
6074 {
6075         struct btrfs_fs_info *fs_info = btrfs_sb(inode->vfs_inode.i_sb);
6076         unsigned nr_extents;
6077         enum btrfs_reserve_flush_enum flush = BTRFS_RESERVE_FLUSH_ALL;
6078         int ret = 0;
6079         bool delalloc_lock = true;
6080
6081         /* If we are a free space inode we need to not flush since we will be in
6082          * the middle of a transaction commit.  We also don't need the delalloc
6083          * mutex since we won't race with anybody.  We need this mostly to make
6084          * lockdep shut its filthy mouth.
6085          *
6086          * If we have a transaction open (can happen if we call truncate_block
6087          * from truncate), then we need FLUSH_LIMIT so we don't deadlock.
6088          */
6089         if (btrfs_is_free_space_inode(inode)) {
6090                 flush = BTRFS_RESERVE_NO_FLUSH;
6091                 delalloc_lock = false;
6092         } else {
6093                 if (current->journal_info)
6094                         flush = BTRFS_RESERVE_FLUSH_LIMIT;
6095
6096                 if (btrfs_transaction_in_commit(fs_info))
6097                         schedule_timeout(1);
6098         }
6099
6100         if (delalloc_lock)
6101                 mutex_lock(&inode->delalloc_mutex);
6102
6103         num_bytes = ALIGN(num_bytes, fs_info->sectorsize);
6104
6105         /* Add our new extents and calculate the new rsv size. */
6106         spin_lock(&inode->lock);
6107         nr_extents = count_max_extents(num_bytes);
6108         btrfs_mod_outstanding_extents(inode, nr_extents);
6109         inode->csum_bytes += num_bytes;
6110         btrfs_calculate_inode_block_rsv_size(fs_info, inode);
6111         spin_unlock(&inode->lock);
6112
6113         ret = btrfs_inode_rsv_refill(inode, flush);
6114         if (unlikely(ret))
6115                 goto out_fail;
6116
6117         if (delalloc_lock)
6118                 mutex_unlock(&inode->delalloc_mutex);
6119         return 0;
6120
6121 out_fail:
6122         spin_lock(&inode->lock);
6123         nr_extents = count_max_extents(num_bytes);
6124         btrfs_mod_outstanding_extents(inode, -nr_extents);
6125         inode->csum_bytes -= num_bytes;
6126         btrfs_calculate_inode_block_rsv_size(fs_info, inode);
6127         spin_unlock(&inode->lock);
6128
6129         btrfs_inode_rsv_release(inode, true);
6130         if (delalloc_lock)
6131                 mutex_unlock(&inode->delalloc_mutex);
6132         return ret;
6133 }
6134
6135 /**
6136  * btrfs_delalloc_release_metadata - release a metadata reservation for an inode
6137  * @inode: the inode to release the reservation for.
6138  * @num_bytes: the number of bytes we are releasing.
6139  * @qgroup_free: free qgroup reservation or convert it to per-trans reservation
6140  *
6141  * This will release the metadata reservation for an inode.  This can be called
6142  * once we complete IO for a given set of bytes to release their metadata
6143  * reservations, or on error for the same reason.
6144  */
6145 void btrfs_delalloc_release_metadata(struct btrfs_inode *inode, u64 num_bytes,
6146                                      bool qgroup_free)
6147 {
6148         struct btrfs_fs_info *fs_info = btrfs_sb(inode->vfs_inode.i_sb);
6149
6150         num_bytes = ALIGN(num_bytes, fs_info->sectorsize);
6151         spin_lock(&inode->lock);
6152         inode->csum_bytes -= num_bytes;
6153         btrfs_calculate_inode_block_rsv_size(fs_info, inode);
6154         spin_unlock(&inode->lock);
6155
6156         if (btrfs_is_testing(fs_info))
6157                 return;
6158
6159         btrfs_inode_rsv_release(inode, qgroup_free);
6160 }
6161
6162 /**
6163  * btrfs_delalloc_release_extents - release our outstanding_extents
6164  * @inode: the inode to balance the reservation for.
6165  * @num_bytes: the number of bytes we originally reserved with
6166  * @qgroup_free: do we need to free qgroup meta reservation or convert them.
6167  *
6168  * When we reserve space we increase outstanding_extents for the extents we may
6169  * add.  Once we've set the range as delalloc or created our ordered extents we
6170  * have outstanding_extents to track the real usage, so we use this to free our
6171  * temporarily tracked outstanding_extents.  This _must_ be used in conjunction
6172  * with btrfs_delalloc_reserve_metadata.
6173  */
6174 void btrfs_delalloc_release_extents(struct btrfs_inode *inode, u64 num_bytes,
6175                                     bool qgroup_free)
6176 {
6177         struct btrfs_fs_info *fs_info = btrfs_sb(inode->vfs_inode.i_sb);
6178         unsigned num_extents;
6179
6180         spin_lock(&inode->lock);
6181         num_extents = count_max_extents(num_bytes);
6182         btrfs_mod_outstanding_extents(inode, -num_extents);
6183         btrfs_calculate_inode_block_rsv_size(fs_info, inode);
6184         spin_unlock(&inode->lock);
6185
6186         if (btrfs_is_testing(fs_info))
6187                 return;
6188
6189         btrfs_inode_rsv_release(inode, qgroup_free);
6190 }
6191
6192 /**
6193  * btrfs_delalloc_reserve_space - reserve data and metadata space for
6194  * delalloc
6195  * @inode: inode we're writing to
6196  * @start: start range we are writing to
6197  * @len: how long the range we are writing to
6198  * @reserved: mandatory parameter, record actually reserved qgroup ranges of
6199  *            current reservation.
6200  *
6201  * This will do the following things
6202  *
6203  * o reserve space in data space info for num bytes
6204  *   and reserve precious corresponding qgroup space
6205  *   (Done in check_data_free_space)
6206  *
6207  * o reserve space for metadata space, based on the number of outstanding
6208  *   extents and how much csums will be needed
6209  *   also reserve metadata space in a per root over-reserve method.
6210  * o add to the inodes->delalloc_bytes
6211  * o add it to the fs_info's delalloc inodes list.
6212  *   (Above 3 all done in delalloc_reserve_metadata)
6213  *
6214  * Return 0 for success
6215  * Return <0 for error(-ENOSPC or -EQUOT)
6216  */
6217 int btrfs_delalloc_reserve_space(struct inode *inode,
6218                         struct extent_changeset **reserved, u64 start, u64 len)
6219 {
6220         int ret;
6221
6222         ret = btrfs_check_data_free_space(inode, reserved, start, len);
6223         if (ret < 0)
6224                 return ret;
6225         ret = btrfs_delalloc_reserve_metadata(BTRFS_I(inode), len);
6226         if (ret < 0)
6227                 btrfs_free_reserved_data_space(inode, *reserved, start, len);
6228         return ret;
6229 }
6230
6231 /**
6232  * btrfs_delalloc_release_space - release data and metadata space for delalloc
6233  * @inode: inode we're releasing space for
6234  * @start: start position of the space already reserved
6235  * @len: the len of the space already reserved
6236  * @release_bytes: the len of the space we consumed or didn't use
6237  *
6238  * This function will release the metadata space that was not used and will
6239  * decrement ->delalloc_bytes and remove it from the fs_info delalloc_inodes
6240  * list if there are no delalloc bytes left.
6241  * Also it will handle the qgroup reserved space.
6242  */
6243 void btrfs_delalloc_release_space(struct inode *inode,
6244                                   struct extent_changeset *reserved,
6245                                   u64 start, u64 len, bool qgroup_free)
6246 {
6247         btrfs_delalloc_release_metadata(BTRFS_I(inode), len, qgroup_free);
6248         btrfs_free_reserved_data_space(inode, reserved, start, len);
6249 }
6250
6251 static int update_block_group(struct btrfs_trans_handle *trans,
6252                               struct btrfs_fs_info *info, u64 bytenr,
6253                               u64 num_bytes, int alloc)
6254 {
6255         struct btrfs_block_group_cache *cache = NULL;
6256         u64 total = num_bytes;
6257         u64 old_val;
6258         u64 byte_in_group;
6259         int factor;
6260
6261         /* block accounting for super block */
6262         spin_lock(&info->delalloc_root_lock);
6263         old_val = btrfs_super_bytes_used(info->super_copy);
6264         if (alloc)
6265                 old_val += num_bytes;
6266         else
6267                 old_val -= num_bytes;
6268         btrfs_set_super_bytes_used(info->super_copy, old_val);
6269         spin_unlock(&info->delalloc_root_lock);
6270
6271         while (total) {
6272                 cache = btrfs_lookup_block_group(info, bytenr);
6273                 if (!cache)
6274                         return -ENOENT;
6275                 if (cache->flags & (BTRFS_BLOCK_GROUP_DUP |
6276                                     BTRFS_BLOCK_GROUP_RAID1 |
6277                                     BTRFS_BLOCK_GROUP_RAID10))
6278                         factor = 2;
6279                 else
6280                         factor = 1;
6281                 /*
6282                  * If this block group has free space cache written out, we
6283                  * need to make sure to load it if we are removing space.  This
6284                  * is because we need the unpinning stage to actually add the
6285                  * space back to the block group, otherwise we will leak space.
6286                  */
6287                 if (!alloc && cache->cached == BTRFS_CACHE_NO)
6288                         cache_block_group(cache, 1);
6289
6290                 byte_in_group = bytenr - cache->key.objectid;
6291                 WARN_ON(byte_in_group > cache->key.offset);
6292
6293                 spin_lock(&cache->space_info->lock);
6294                 spin_lock(&cache->lock);
6295
6296                 if (btrfs_test_opt(info, SPACE_CACHE) &&
6297                     cache->disk_cache_state < BTRFS_DC_CLEAR)
6298                         cache->disk_cache_state = BTRFS_DC_CLEAR;
6299
6300                 old_val = btrfs_block_group_used(&cache->item);
6301                 num_bytes = min(total, cache->key.offset - byte_in_group);
6302                 if (alloc) {
6303                         old_val += num_bytes;
6304                         btrfs_set_block_group_used(&cache->item, old_val);
6305                         cache->reserved -= num_bytes;
6306                         cache->space_info->bytes_reserved -= num_bytes;
6307                         cache->space_info->bytes_used += num_bytes;
6308                         cache->space_info->disk_used += num_bytes * factor;
6309                         spin_unlock(&cache->lock);
6310                         spin_unlock(&cache->space_info->lock);
6311                 } else {
6312                         old_val -= num_bytes;
6313                         btrfs_set_block_group_used(&cache->item, old_val);
6314                         cache->pinned += num_bytes;
6315                         cache->space_info->bytes_pinned += num_bytes;
6316                         cache->space_info->bytes_used -= num_bytes;
6317                         cache->space_info->disk_used -= num_bytes * factor;
6318                         spin_unlock(&cache->lock);
6319                         spin_unlock(&cache->space_info->lock);
6320
6321                         trace_btrfs_space_reservation(info, "pinned",
6322                                                       cache->space_info->flags,
6323                                                       num_bytes, 1);
6324                         percpu_counter_add(&cache->space_info->total_bytes_pinned,
6325                                            num_bytes);
6326                         set_extent_dirty(info->pinned_extents,
6327                                          bytenr, bytenr + num_bytes - 1,
6328                                          GFP_NOFS | __GFP_NOFAIL);
6329                 }
6330
6331                 spin_lock(&trans->transaction->dirty_bgs_lock);
6332                 if (list_empty(&cache->dirty_list)) {
6333                         list_add_tail(&cache->dirty_list,
6334                                       &trans->transaction->dirty_bgs);
6335                                 trans->transaction->num_dirty_bgs++;
6336                         btrfs_get_block_group(cache);
6337                 }
6338                 spin_unlock(&trans->transaction->dirty_bgs_lock);
6339
6340                 /*
6341                  * No longer have used bytes in this block group, queue it for
6342                  * deletion. We do this after adding the block group to the
6343                  * dirty list to avoid races between cleaner kthread and space
6344                  * cache writeout.
6345                  */
6346                 if (!alloc && old_val == 0) {
6347                         spin_lock(&info->unused_bgs_lock);
6348                         if (list_empty(&cache->bg_list)) {
6349                                 btrfs_get_block_group(cache);
6350                                 list_add_tail(&cache->bg_list,
6351                                               &info->unused_bgs);
6352                         }
6353                         spin_unlock(&info->unused_bgs_lock);
6354                 }
6355
6356                 btrfs_put_block_group(cache);
6357                 total -= num_bytes;
6358                 bytenr += num_bytes;
6359         }
6360         return 0;
6361 }
6362
6363 static u64 first_logical_byte(struct btrfs_fs_info *fs_info, u64 search_start)
6364 {
6365         struct btrfs_block_group_cache *cache;
6366         u64 bytenr;
6367
6368         spin_lock(&fs_info->block_group_cache_lock);
6369         bytenr = fs_info->first_logical_byte;
6370         spin_unlock(&fs_info->block_group_cache_lock);
6371
6372         if (bytenr < (u64)-1)
6373                 return bytenr;
6374
6375         cache = btrfs_lookup_first_block_group(fs_info, search_start);
6376         if (!cache)
6377                 return 0;
6378
6379         bytenr = cache->key.objectid;
6380         btrfs_put_block_group(cache);
6381
6382         return bytenr;
6383 }
6384
6385 static int pin_down_extent(struct btrfs_fs_info *fs_info,
6386                            struct btrfs_block_group_cache *cache,
6387                            u64 bytenr, u64 num_bytes, int reserved)
6388 {
6389         spin_lock(&cache->space_info->lock);
6390         spin_lock(&cache->lock);
6391         cache->pinned += num_bytes;
6392         cache->space_info->bytes_pinned += num_bytes;
6393         if (reserved) {
6394                 cache->reserved -= num_bytes;
6395                 cache->space_info->bytes_reserved -= num_bytes;
6396         }
6397         spin_unlock(&cache->lock);
6398         spin_unlock(&cache->space_info->lock);
6399
6400         trace_btrfs_space_reservation(fs_info, "pinned",
6401                                       cache->space_info->flags, num_bytes, 1);
6402         percpu_counter_add(&cache->space_info->total_bytes_pinned, num_bytes);
6403         set_extent_dirty(fs_info->pinned_extents, bytenr,
6404                          bytenr + num_bytes - 1, GFP_NOFS | __GFP_NOFAIL);
6405         return 0;
6406 }
6407
6408 /*
6409  * this function must be called within transaction
6410  */
6411 int btrfs_pin_extent(struct btrfs_fs_info *fs_info,
6412                      u64 bytenr, u64 num_bytes, int reserved)
6413 {
6414         struct btrfs_block_group_cache *cache;
6415
6416         cache = btrfs_lookup_block_group(fs_info, bytenr);
6417         BUG_ON(!cache); /* Logic error */
6418
6419         pin_down_extent(fs_info, cache, bytenr, num_bytes, reserved);
6420
6421         btrfs_put_block_group(cache);
6422         return 0;
6423 }
6424
6425 /*
6426  * this function must be called within transaction
6427  */
6428 int btrfs_pin_extent_for_log_replay(struct btrfs_fs_info *fs_info,
6429                                     u64 bytenr, u64 num_bytes)
6430 {
6431         struct btrfs_block_group_cache *cache;
6432         int ret;
6433
6434         cache = btrfs_lookup_block_group(fs_info, bytenr);
6435         if (!cache)
6436                 return -EINVAL;
6437
6438         /*
6439          * pull in the free space cache (if any) so that our pin
6440          * removes the free space from the cache.  We have load_only set
6441          * to one because the slow code to read in the free extents does check
6442          * the pinned extents.
6443          */
6444         cache_block_group(cache, 1);
6445
6446         pin_down_extent(fs_info, cache, bytenr, num_bytes, 0);
6447
6448         /* remove us from the free space cache (if we're there at all) */
6449         ret = btrfs_remove_free_space(cache, bytenr, num_bytes);
6450         btrfs_put_block_group(cache);
6451         return ret;
6452 }
6453
6454 static int __exclude_logged_extent(struct btrfs_fs_info *fs_info,
6455                                    u64 start, u64 num_bytes)
6456 {
6457         int ret;
6458         struct btrfs_block_group_cache *block_group;
6459         struct btrfs_caching_control *caching_ctl;
6460
6461         block_group = btrfs_lookup_block_group(fs_info, start);
6462         if (!block_group)
6463                 return -EINVAL;
6464
6465         cache_block_group(block_group, 0);
6466         caching_ctl = get_caching_control(block_group);
6467
6468         if (!caching_ctl) {
6469                 /* Logic error */
6470                 BUG_ON(!block_group_cache_done(block_group));
6471                 ret = btrfs_remove_free_space(block_group, start, num_bytes);
6472         } else {
6473                 mutex_lock(&caching_ctl->mutex);
6474
6475                 if (start >= caching_ctl->progress) {
6476                         ret = add_excluded_extent(fs_info, start, num_bytes);
6477                 } else if (start + num_bytes <= caching_ctl->progress) {
6478                         ret = btrfs_remove_free_space(block_group,
6479                                                       start, num_bytes);
6480                 } else {
6481                         num_bytes = caching_ctl->progress - start;
6482                         ret = btrfs_remove_free_space(block_group,
6483                                                       start, num_bytes);
6484                         if (ret)
6485                                 goto out_lock;
6486
6487                         num_bytes = (start + num_bytes) -
6488                                 caching_ctl->progress;
6489                         start = caching_ctl->progress;
6490                         ret = add_excluded_extent(fs_info, start, num_bytes);
6491                 }
6492 out_lock:
6493                 mutex_unlock(&caching_ctl->mutex);
6494                 put_caching_control(caching_ctl);
6495         }
6496         btrfs_put_block_group(block_group);
6497         return ret;
6498 }
6499
6500 int btrfs_exclude_logged_extents(struct btrfs_fs_info *fs_info,
6501                                  struct extent_buffer *eb)
6502 {
6503         struct btrfs_file_extent_item *item;
6504         struct btrfs_key key;
6505         int found_type;
6506         int i;
6507
6508         if (!btrfs_fs_incompat(fs_info, MIXED_GROUPS))
6509                 return 0;
6510
6511         for (i = 0; i < btrfs_header_nritems(eb); i++) {
6512                 btrfs_item_key_to_cpu(eb, &key, i);
6513                 if (key.type != BTRFS_EXTENT_DATA_KEY)
6514                         continue;
6515                 item = btrfs_item_ptr(eb, i, struct btrfs_file_extent_item);
6516                 found_type = btrfs_file_extent_type(eb, item);
6517                 if (found_type == BTRFS_FILE_EXTENT_INLINE)
6518                         continue;
6519                 if (btrfs_file_extent_disk_bytenr(eb, item) == 0)
6520                         continue;
6521                 key.objectid = btrfs_file_extent_disk_bytenr(eb, item);
6522                 key.offset = btrfs_file_extent_disk_num_bytes(eb, item);
6523                 __exclude_logged_extent(fs_info, key.objectid, key.offset);
6524         }
6525
6526         return 0;
6527 }
6528
6529 static void
6530 btrfs_inc_block_group_reservations(struct btrfs_block_group_cache *bg)
6531 {
6532         atomic_inc(&bg->reservations);
6533 }
6534
6535 void btrfs_dec_block_group_reservations(struct btrfs_fs_info *fs_info,
6536                                         const u64 start)
6537 {
6538         struct btrfs_block_group_cache *bg;
6539
6540         bg = btrfs_lookup_block_group(fs_info, start);
6541         ASSERT(bg);
6542         if (atomic_dec_and_test(&bg->reservations))
6543                 wake_up_var(&bg->reservations);
6544         btrfs_put_block_group(bg);
6545 }
6546
6547 void btrfs_wait_block_group_reservations(struct btrfs_block_group_cache *bg)
6548 {
6549         struct btrfs_space_info *space_info = bg->space_info;
6550
6551         ASSERT(bg->ro);
6552
6553         if (!(bg->flags & BTRFS_BLOCK_GROUP_DATA))
6554                 return;
6555
6556         /*
6557          * Our block group is read only but before we set it to read only,
6558          * some task might have had allocated an extent from it already, but it
6559          * has not yet created a respective ordered extent (and added it to a
6560          * root's list of ordered extents).
6561          * Therefore wait for any task currently allocating extents, since the
6562          * block group's reservations counter is incremented while a read lock
6563          * on the groups' semaphore is held and decremented after releasing
6564          * the read access on that semaphore and creating the ordered extent.
6565          */
6566         down_write(&space_info->groups_sem);
6567         up_write(&space_info->groups_sem);
6568
6569         wait_var_event(&bg->reservations, !atomic_read(&bg->reservations));
6570 }
6571
6572 /**
6573  * btrfs_add_reserved_bytes - update the block_group and space info counters
6574  * @cache:      The cache we are manipulating
6575  * @ram_bytes:  The number of bytes of file content, and will be same to
6576  *              @num_bytes except for the compress path.
6577  * @num_bytes:  The number of bytes in question
6578  * @delalloc:   The blocks are allocated for the delalloc write
6579  *
6580  * This is called by the allocator when it reserves space. If this is a
6581  * reservation and the block group has become read only we cannot make the
6582  * reservation and return -EAGAIN, otherwise this function always succeeds.
6583  */
6584 static int btrfs_add_reserved_bytes(struct btrfs_block_group_cache *cache,
6585                                     u64 ram_bytes, u64 num_bytes, int delalloc)
6586 {
6587         struct btrfs_space_info *space_info = cache->space_info;
6588         int ret = 0;
6589
6590         spin_lock(&space_info->lock);
6591         spin_lock(&cache->lock);
6592         if (cache->ro) {
6593                 ret = -EAGAIN;
6594         } else {
6595                 cache->reserved += num_bytes;
6596                 space_info->bytes_reserved += num_bytes;
6597
6598                 trace_btrfs_space_reservation(cache->fs_info,
6599                                 "space_info", space_info->flags,
6600                                 ram_bytes, 0);
6601                 space_info->bytes_may_use -= ram_bytes;
6602                 if (delalloc)
6603                         cache->delalloc_bytes += num_bytes;
6604         }
6605         spin_unlock(&cache->lock);
6606         spin_unlock(&space_info->lock);
6607         return ret;
6608 }
6609
6610 /**
6611  * btrfs_free_reserved_bytes - update the block_group and space info counters
6612  * @cache:      The cache we are manipulating
6613  * @num_bytes:  The number of bytes in question
6614  * @delalloc:   The blocks are allocated for the delalloc write
6615  *
6616  * This is called by somebody who is freeing space that was never actually used
6617  * on disk.  For example if you reserve some space for a new leaf in transaction
6618  * A and before transaction A commits you free that leaf, you call this with
6619  * reserve set to 0 in order to clear the reservation.
6620  */
6621
6622 static int btrfs_free_reserved_bytes(struct btrfs_block_group_cache *cache,
6623                                      u64 num_bytes, int delalloc)
6624 {
6625         struct btrfs_space_info *space_info = cache->space_info;
6626         int ret = 0;
6627
6628         spin_lock(&space_info->lock);
6629         spin_lock(&cache->lock);
6630         if (cache->ro)
6631                 space_info->bytes_readonly += num_bytes;
6632         cache->reserved -= num_bytes;
6633         space_info->bytes_reserved -= num_bytes;
6634
6635         if (delalloc)
6636                 cache->delalloc_bytes -= num_bytes;
6637         spin_unlock(&cache->lock);
6638         spin_unlock(&space_info->lock);
6639         return ret;
6640 }
6641 void btrfs_prepare_extent_commit(struct btrfs_fs_info *fs_info)
6642 {
6643         struct btrfs_caching_control *next;
6644         struct btrfs_caching_control *caching_ctl;
6645         struct btrfs_block_group_cache *cache;
6646
6647         down_write(&fs_info->commit_root_sem);
6648
6649         list_for_each_entry_safe(caching_ctl, next,
6650                                  &fs_info->caching_block_groups, list) {
6651                 cache = caching_ctl->block_group;
6652                 if (block_group_cache_done(cache)) {
6653                         cache->last_byte_to_unpin = (u64)-1;
6654                         list_del_init(&caching_ctl->list);
6655                         put_caching_control(caching_ctl);
6656                 } else {
6657                         cache->last_byte_to_unpin = caching_ctl->progress;
6658                 }
6659         }
6660
6661         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
6662                 fs_info->pinned_extents = &fs_info->freed_extents[1];
6663         else
6664                 fs_info->pinned_extents = &fs_info->freed_extents[0];
6665
6666         up_write(&fs_info->commit_root_sem);
6667
6668         update_global_block_rsv(fs_info);
6669 }
6670
6671 /*
6672  * Returns the free cluster for the given space info and sets empty_cluster to
6673  * what it should be based on the mount options.
6674  */
6675 static struct btrfs_free_cluster *
6676 fetch_cluster_info(struct btrfs_fs_info *fs_info,
6677                    struct btrfs_space_info *space_info, u64 *empty_cluster)
6678 {
6679         struct btrfs_free_cluster *ret = NULL;
6680
6681         *empty_cluster = 0;
6682         if (btrfs_mixed_space_info(space_info))
6683                 return ret;
6684
6685         if (space_info->flags & BTRFS_BLOCK_GROUP_METADATA) {
6686                 ret = &fs_info->meta_alloc_cluster;
6687                 if (btrfs_test_opt(fs_info, SSD))
6688                         *empty_cluster = SZ_2M;
6689                 else
6690                         *empty_cluster = SZ_64K;
6691         } else if ((space_info->flags & BTRFS_BLOCK_GROUP_DATA) &&
6692                    btrfs_test_opt(fs_info, SSD_SPREAD)) {
6693                 *empty_cluster = SZ_2M;
6694                 ret = &fs_info->data_alloc_cluster;
6695         }
6696
6697         return ret;
6698 }
6699
6700 static int unpin_extent_range(struct btrfs_fs_info *fs_info,
6701                               u64 start, u64 end,
6702                               const bool return_free_space)
6703 {
6704         struct btrfs_block_group_cache *cache = NULL;
6705         struct btrfs_space_info *space_info;
6706         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
6707         struct btrfs_free_cluster *cluster = NULL;
6708         u64 len;
6709         u64 total_unpinned = 0;
6710         u64 empty_cluster = 0;
6711         bool readonly;
6712
6713         while (start <= end) {
6714                 readonly = false;
6715                 if (!cache ||
6716                     start >= cache->key.objectid + cache->key.offset) {
6717                         if (cache)
6718                                 btrfs_put_block_group(cache);
6719                         total_unpinned = 0;
6720                         cache = btrfs_lookup_block_group(fs_info, start);
6721                         BUG_ON(!cache); /* Logic error */
6722
6723                         cluster = fetch_cluster_info(fs_info,
6724                                                      cache->space_info,
6725                                                      &empty_cluster);
6726                         empty_cluster <<= 1;
6727                 }
6728
6729                 len = cache->key.objectid + cache->key.offset - start;
6730                 len = min(len, end + 1 - start);
6731
6732                 if (start < cache->last_byte_to_unpin) {
6733                         len = min(len, cache->last_byte_to_unpin - start);
6734                         if (return_free_space)
6735                                 btrfs_add_free_space(cache, start, len);
6736                 }
6737
6738                 start += len;
6739                 total_unpinned += len;
6740                 space_info = cache->space_info;
6741
6742                 /*
6743                  * If this space cluster has been marked as fragmented and we've
6744                  * unpinned enough in this block group to potentially allow a
6745                  * cluster to be created inside of it go ahead and clear the
6746                  * fragmented check.
6747                  */
6748                 if (cluster && cluster->fragmented &&
6749                     total_unpinned > empty_cluster) {
6750                         spin_lock(&cluster->lock);
6751                         cluster->fragmented = 0;
6752                         spin_unlock(&cluster->lock);
6753                 }
6754
6755                 spin_lock(&space_info->lock);
6756                 spin_lock(&cache->lock);
6757                 cache->pinned -= len;
6758                 space_info->bytes_pinned -= len;
6759
6760                 trace_btrfs_space_reservation(fs_info, "pinned",
6761                                               space_info->flags, len, 0);
6762                 space_info->max_extent_size = 0;
6763                 percpu_counter_add(&space_info->total_bytes_pinned, -len);
6764                 if (cache->ro) {
6765                         space_info->bytes_readonly += len;
6766                         readonly = true;
6767                 }
6768                 spin_unlock(&cache->lock);
6769                 if (!readonly && return_free_space &&
6770                     global_rsv->space_info == space_info) {
6771                         u64 to_add = len;
6772
6773                         spin_lock(&global_rsv->lock);
6774                         if (!global_rsv->full) {
6775                                 to_add = min(len, global_rsv->size -
6776                                              global_rsv->reserved);
6777                                 global_rsv->reserved += to_add;
6778                                 space_info->bytes_may_use += to_add;
6779                                 if (global_rsv->reserved >= global_rsv->size)
6780                                         global_rsv->full = 1;
6781                                 trace_btrfs_space_reservation(fs_info,
6782                                                               "space_info",
6783                                                               space_info->flags,
6784                                                               to_add, 1);
6785                                 len -= to_add;
6786                         }
6787                         spin_unlock(&global_rsv->lock);
6788                         /* Add to any tickets we may have */
6789                         if (len)
6790                                 space_info_add_new_bytes(fs_info, space_info,
6791                                                          len);
6792                 }
6793                 spin_unlock(&space_info->lock);
6794         }
6795
6796         if (cache)
6797                 btrfs_put_block_group(cache);
6798         return 0;
6799 }
6800
6801 int btrfs_finish_extent_commit(struct btrfs_trans_handle *trans)
6802 {
6803         struct btrfs_fs_info *fs_info = trans->fs_info;
6804         struct btrfs_block_group_cache *block_group, *tmp;
6805         struct list_head *deleted_bgs;
6806         struct extent_io_tree *unpin;
6807         u64 start;
6808         u64 end;
6809         int ret;
6810
6811         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
6812                 unpin = &fs_info->freed_extents[1];
6813         else
6814                 unpin = &fs_info->freed_extents[0];
6815
6816         while (!trans->aborted) {
6817                 mutex_lock(&fs_info->unused_bg_unpin_mutex);
6818                 ret = find_first_extent_bit(unpin, 0, &start, &end,
6819                                             EXTENT_DIRTY, NULL);
6820                 if (ret) {
6821                         mutex_unlock(&fs_info->unused_bg_unpin_mutex);
6822                         break;
6823                 }
6824
6825                 if (btrfs_test_opt(fs_info, DISCARD))
6826                         ret = btrfs_discard_extent(fs_info, start,
6827                                                    end + 1 - start, NULL);
6828
6829                 clear_extent_dirty(unpin, start, end);
6830                 unpin_extent_range(fs_info, start, end, true);
6831                 mutex_unlock(&fs_info->unused_bg_unpin_mutex);
6832                 cond_resched();
6833         }
6834
6835         /*
6836          * Transaction is finished.  We don't need the lock anymore.  We
6837          * do need to clean up the block groups in case of a transaction
6838          * abort.
6839          */
6840         deleted_bgs = &trans->transaction->deleted_bgs;
6841         list_for_each_entry_safe(block_group, tmp, deleted_bgs, bg_list) {
6842                 u64 trimmed = 0;
6843
6844                 ret = -EROFS;
6845                 if (!trans->aborted)
6846                         ret = btrfs_discard_extent(fs_info,
6847                                                    block_group->key.objectid,
6848                                                    block_group->key.offset,
6849                                                    &trimmed);
6850
6851                 list_del_init(&block_group->bg_list);
6852                 btrfs_put_block_group_trimming(block_group);
6853                 btrfs_put_block_group(block_group);
6854
6855                 if (ret) {
6856                         const char *errstr = btrfs_decode_error(ret);
6857                         btrfs_warn(fs_info,
6858                            "discard failed while removing blockgroup: errno=%d %s",
6859                                    ret, errstr);
6860                 }
6861         }
6862
6863         return 0;
6864 }
6865
6866 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
6867                                 struct btrfs_fs_info *info,
6868                                 struct btrfs_delayed_ref_node *node, u64 parent,
6869                                 u64 root_objectid, u64 owner_objectid,
6870                                 u64 owner_offset, int refs_to_drop,
6871                                 struct btrfs_delayed_extent_op *extent_op)
6872 {
6873         struct btrfs_key key;
6874         struct btrfs_path *path;
6875         struct btrfs_root *extent_root = info->extent_root;
6876         struct extent_buffer *leaf;
6877         struct btrfs_extent_item *ei;
6878         struct btrfs_extent_inline_ref *iref;
6879         int ret;
6880         int is_data;
6881         int extent_slot = 0;
6882         int found_extent = 0;
6883         int num_to_del = 1;
6884         u32 item_size;
6885         u64 refs;
6886         u64 bytenr = node->bytenr;
6887         u64 num_bytes = node->num_bytes;
6888         int last_ref = 0;
6889         bool skinny_metadata = btrfs_fs_incompat(info, SKINNY_METADATA);
6890
6891         path = btrfs_alloc_path();
6892         if (!path)
6893                 return -ENOMEM;
6894
6895         path->reada = READA_FORWARD;
6896         path->leave_spinning = 1;
6897
6898         is_data = owner_objectid >= BTRFS_FIRST_FREE_OBJECTID;
6899         BUG_ON(!is_data && refs_to_drop != 1);
6900
6901         if (is_data)
6902                 skinny_metadata = false;
6903
6904         ret = lookup_extent_backref(trans, info, path, &iref,
6905                                     bytenr, num_bytes, parent,
6906                                     root_objectid, owner_objectid,
6907                                     owner_offset);
6908         if (ret == 0) {
6909                 extent_slot = path->slots[0];
6910                 while (extent_slot >= 0) {
6911                         btrfs_item_key_to_cpu(path->nodes[0], &key,
6912                                               extent_slot);
6913                         if (key.objectid != bytenr)
6914                                 break;
6915                         if (key.type == BTRFS_EXTENT_ITEM_KEY &&
6916                             key.offset == num_bytes) {
6917                                 found_extent = 1;
6918                                 break;
6919                         }
6920                         if (key.type == BTRFS_METADATA_ITEM_KEY &&
6921                             key.offset == owner_objectid) {
6922                                 found_extent = 1;
6923                                 break;
6924                         }
6925                         if (path->slots[0] - extent_slot > 5)
6926                                 break;
6927                         extent_slot--;
6928                 }
6929 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
6930                 item_size = btrfs_item_size_nr(path->nodes[0], extent_slot);
6931                 if (found_extent && item_size < sizeof(*ei))
6932                         found_extent = 0;
6933 #endif
6934                 if (!found_extent) {
6935                         BUG_ON(iref);
6936                         ret = remove_extent_backref(trans, info, path, NULL,
6937                                                     refs_to_drop,
6938                                                     is_data, &last_ref);
6939                         if (ret) {
6940                                 btrfs_abort_transaction(trans, ret);
6941                                 goto out;
6942                         }
6943                         btrfs_release_path(path);
6944                         path->leave_spinning = 1;
6945
6946                         key.objectid = bytenr;
6947                         key.type = BTRFS_EXTENT_ITEM_KEY;
6948                         key.offset = num_bytes;
6949
6950                         if (!is_data && skinny_metadata) {
6951                                 key.type = BTRFS_METADATA_ITEM_KEY;
6952                                 key.offset = owner_objectid;
6953                         }
6954
6955                         ret = btrfs_search_slot(trans, extent_root,
6956                                                 &key, path, -1, 1);
6957                         if (ret > 0 && skinny_metadata && path->slots[0]) {
6958                                 /*
6959                                  * Couldn't find our skinny metadata item,
6960                                  * see if we have ye olde extent item.
6961                                  */
6962                                 path->slots[0]--;
6963                                 btrfs_item_key_to_cpu(path->nodes[0], &key,
6964                                                       path->slots[0]);
6965                                 if (key.objectid == bytenr &&
6966                                     key.type == BTRFS_EXTENT_ITEM_KEY &&
6967                                     key.offset == num_bytes)
6968                                         ret = 0;
6969                         }
6970
6971                         if (ret > 0 && skinny_metadata) {
6972                                 skinny_metadata = false;
6973                                 key.objectid = bytenr;
6974                                 key.type = BTRFS_EXTENT_ITEM_KEY;
6975                                 key.offset = num_bytes;
6976                                 btrfs_release_path(path);
6977                                 ret = btrfs_search_slot(trans, extent_root,
6978                                                         &key, path, -1, 1);
6979                         }
6980
6981                         if (ret) {
6982                                 btrfs_err(info,
6983                                           "umm, got %d back from search, was looking for %llu",
6984                                           ret, bytenr);
6985                                 if (ret > 0)
6986                                         btrfs_print_leaf(path->nodes[0]);
6987                         }
6988                         if (ret < 0) {
6989                                 btrfs_abort_transaction(trans, ret);
6990                                 goto out;
6991                         }
6992                         extent_slot = path->slots[0];
6993                 }
6994         } else if (WARN_ON(ret == -ENOENT)) {
6995                 btrfs_print_leaf(path->nodes[0]);
6996                 btrfs_err(info,
6997                         "unable to find ref byte nr %llu parent %llu root %llu  owner %llu offset %llu",
6998                         bytenr, parent, root_objectid, owner_objectid,
6999                         owner_offset);
7000                 btrfs_abort_transaction(trans, ret);
7001                 goto out;
7002         } else {
7003                 btrfs_abort_transaction(trans, ret);
7004                 goto out;
7005         }
7006
7007         leaf = path->nodes[0];
7008         item_size = btrfs_item_size_nr(leaf, extent_slot);
7009 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
7010         if (item_size < sizeof(*ei)) {
7011                 BUG_ON(found_extent || extent_slot != path->slots[0]);
7012                 ret = convert_extent_item_v0(trans, info, path, owner_objectid,
7013                                              0);
7014                 if (ret < 0) {
7015                         btrfs_abort_transaction(trans, ret);
7016                         goto out;
7017                 }
7018
7019                 btrfs_release_path(path);
7020                 path->leave_spinning = 1;
7021
7022                 key.objectid = bytenr;
7023                 key.type = BTRFS_EXTENT_ITEM_KEY;
7024                 key.offset = num_bytes;
7025
7026                 ret = btrfs_search_slot(trans, extent_root, &key, path,
7027                                         -1, 1);
7028                 if (ret) {
7029                         btrfs_err(info,
7030                                   "umm, got %d back from search, was looking for %llu",
7031                                 ret, bytenr);
7032                         btrfs_print_leaf(path->nodes[0]);
7033                 }
7034                 if (ret < 0) {
7035                         btrfs_abort_transaction(trans, ret);
7036                         goto out;
7037                 }
7038
7039                 extent_slot = path->slots[0];
7040                 leaf = path->nodes[0];
7041                 item_size = btrfs_item_size_nr(leaf, extent_slot);
7042         }
7043 #endif
7044         BUG_ON(item_size < sizeof(*ei));
7045         ei = btrfs_item_ptr(leaf, extent_slot,
7046                             struct btrfs_extent_item);
7047         if (owner_objectid < BTRFS_FIRST_FREE_OBJECTID &&
7048             key.type == BTRFS_EXTENT_ITEM_KEY) {
7049                 struct btrfs_tree_block_info *bi;
7050                 BUG_ON(item_size < sizeof(*ei) + sizeof(*bi));
7051                 bi = (struct btrfs_tree_block_info *)(ei + 1);
7052                 WARN_ON(owner_objectid != btrfs_tree_block_level(leaf, bi));
7053         }
7054
7055         refs = btrfs_extent_refs(leaf, ei);
7056         if (refs < refs_to_drop) {
7057                 btrfs_err(info,
7058                           "trying to drop %d refs but we only have %Lu for bytenr %Lu",
7059                           refs_to_drop, refs, bytenr);
7060                 ret = -EINVAL;
7061                 btrfs_abort_transaction(trans, ret);
7062                 goto out;
7063         }
7064         refs -= refs_to_drop;
7065
7066         if (refs > 0) {
7067                 if (extent_op)
7068                         __run_delayed_extent_op(extent_op, leaf, ei);
7069                 /*
7070                  * In the case of inline back ref, reference count will
7071                  * be updated by remove_extent_backref
7072                  */
7073                 if (iref) {
7074                         BUG_ON(!found_extent);
7075                 } else {
7076                         btrfs_set_extent_refs(leaf, ei, refs);
7077                         btrfs_mark_buffer_dirty(leaf);
7078                 }
7079                 if (found_extent) {
7080                         ret = remove_extent_backref(trans, info, path,
7081                                                     iref, refs_to_drop,
7082                                                     is_data, &last_ref);
7083                         if (ret) {
7084                                 btrfs_abort_transaction(trans, ret);
7085                                 goto out;
7086                         }
7087                 }
7088         } else {
7089                 if (found_extent) {
7090                         BUG_ON(is_data && refs_to_drop !=
7091                                extent_data_ref_count(path, iref));
7092                         if (iref) {
7093                                 BUG_ON(path->slots[0] != extent_slot);
7094                         } else {
7095                                 BUG_ON(path->slots[0] != extent_slot + 1);
7096                                 path->slots[0] = extent_slot;
7097                                 num_to_del = 2;
7098                         }
7099                 }
7100
7101                 last_ref = 1;
7102                 ret = btrfs_del_items(trans, extent_root, path, path->slots[0],
7103                                       num_to_del);
7104                 if (ret) {
7105                         btrfs_abort_transaction(trans, ret);
7106                         goto out;
7107                 }
7108                 btrfs_release_path(path);
7109
7110                 if (is_data) {
7111                         ret = btrfs_del_csums(trans, info, bytenr, num_bytes);
7112                         if (ret) {
7113                                 btrfs_abort_transaction(trans, ret);
7114                                 goto out;
7115                         }
7116                 }
7117
7118                 ret = add_to_free_space_tree(trans, info, bytenr, num_bytes);
7119                 if (ret) {
7120                         btrfs_abort_transaction(trans, ret);
7121                         goto out;
7122                 }
7123
7124                 ret = update_block_group(trans, info, bytenr, num_bytes, 0);
7125                 if (ret) {
7126                         btrfs_abort_transaction(trans, ret);
7127                         goto out;
7128                 }
7129         }
7130         btrfs_release_path(path);
7131
7132 out:
7133         btrfs_free_path(path);
7134         return ret;
7135 }
7136
7137 /*
7138  * when we free an block, it is possible (and likely) that we free the last
7139  * delayed ref for that extent as well.  This searches the delayed ref tree for
7140  * a given extent, and if there are no other delayed refs to be processed, it
7141  * removes it from the tree.
7142  */
7143 static noinline int check_ref_cleanup(struct btrfs_trans_handle *trans,
7144                                       u64 bytenr)
7145 {
7146         struct btrfs_delayed_ref_head *head;
7147         struct btrfs_delayed_ref_root *delayed_refs;
7148         int ret = 0;
7149
7150         delayed_refs = &trans->transaction->delayed_refs;
7151         spin_lock(&delayed_refs->lock);
7152         head = btrfs_find_delayed_ref_head(delayed_refs, bytenr);
7153         if (!head)
7154                 goto out_delayed_unlock;
7155
7156         spin_lock(&head->lock);
7157         if (!RB_EMPTY_ROOT(&head->ref_tree))
7158                 goto out;
7159
7160         if (head->extent_op) {
7161                 if (!head->must_insert_reserved)
7162                         goto out;
7163                 btrfs_free_delayed_extent_op(head->extent_op);
7164                 head->extent_op = NULL;
7165         }
7166
7167         /*
7168          * waiting for the lock here would deadlock.  If someone else has it
7169          * locked they are already in the process of dropping it anyway
7170          */
7171         if (!mutex_trylock(&head->mutex))
7172                 goto out;
7173
7174         /*
7175          * at this point we have a head with no other entries.  Go
7176          * ahead and process it.
7177          */
7178         rb_erase(&head->href_node, &delayed_refs->href_root);
7179         RB_CLEAR_NODE(&head->href_node);
7180         atomic_dec(&delayed_refs->num_entries);
7181
7182         /*
7183          * we don't take a ref on the node because we're removing it from the
7184          * tree, so we just steal the ref the tree was holding.
7185          */
7186         delayed_refs->num_heads--;
7187         if (head->processing == 0)
7188                 delayed_refs->num_heads_ready--;
7189         head->processing = 0;
7190         spin_unlock(&head->lock);
7191         spin_unlock(&delayed_refs->lock);
7192
7193         BUG_ON(head->extent_op);
7194         if (head->must_insert_reserved)
7195                 ret = 1;
7196
7197         mutex_unlock(&head->mutex);
7198         btrfs_put_delayed_ref_head(head);
7199         return ret;
7200 out:
7201         spin_unlock(&head->lock);
7202
7203 out_delayed_unlock:
7204         spin_unlock(&delayed_refs->lock);
7205         return 0;
7206 }
7207
7208 void btrfs_free_tree_block(struct btrfs_trans_handle *trans,
7209                            struct btrfs_root *root,
7210                            struct extent_buffer *buf,
7211                            u64 parent, int last_ref)
7212 {
7213         struct btrfs_fs_info *fs_info = root->fs_info;
7214         int pin = 1;
7215         int ret;
7216
7217         if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
7218                 int old_ref_mod, new_ref_mod;
7219
7220                 btrfs_ref_tree_mod(root, buf->start, buf->len, parent,
7221                                    root->root_key.objectid,
7222                                    btrfs_header_level(buf), 0,
7223                                    BTRFS_DROP_DELAYED_REF);
7224                 ret = btrfs_add_delayed_tree_ref(fs_info, trans, buf->start,
7225                                                  buf->len, parent,
7226                                                  root->root_key.objectid,
7227                                                  btrfs_header_level(buf),
7228                                                  BTRFS_DROP_DELAYED_REF, NULL,
7229                                                  &old_ref_mod, &new_ref_mod);
7230                 BUG_ON(ret); /* -ENOMEM */
7231                 pin = old_ref_mod >= 0 && new_ref_mod < 0;
7232         }
7233
7234         if (last_ref && btrfs_header_generation(buf) == trans->transid) {
7235                 struct btrfs_block_group_cache *cache;
7236
7237                 if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
7238                         ret = check_ref_cleanup(trans, buf->start);
7239                         if (!ret)
7240                                 goto out;
7241                 }
7242
7243                 pin = 0;
7244                 cache = btrfs_lookup_block_group(fs_info, buf->start);
7245
7246                 if (btrfs_header_flag(buf, BTRFS_HEADER_FLAG_WRITTEN)) {
7247                         pin_down_extent(fs_info, cache, buf->start,
7248                                         buf->len, 1);
7249                         btrfs_put_block_group(cache);
7250                         goto out;
7251                 }
7252
7253                 WARN_ON(test_bit(EXTENT_BUFFER_DIRTY, &buf->bflags));
7254
7255                 btrfs_add_free_space(cache, buf->start, buf->len);
7256                 btrfs_free_reserved_bytes(cache, buf->len, 0);
7257                 btrfs_put_block_group(cache);
7258                 trace_btrfs_reserved_extent_free(fs_info, buf->start, buf->len);
7259         }
7260 out:
7261         if (pin)
7262                 add_pinned_bytes(fs_info, buf->len, btrfs_header_level(buf),
7263                                  root->root_key.objectid);
7264
7265         if (last_ref) {
7266                 /*
7267                  * Deleting the buffer, clear the corrupt flag since it doesn't
7268                  * matter anymore.
7269                  */
7270                 clear_bit(EXTENT_BUFFER_CORRUPT, &buf->bflags);
7271         }
7272 }
7273
7274 /* Can return -ENOMEM */
7275 int btrfs_free_extent(struct btrfs_trans_handle *trans,
7276                       struct btrfs_root *root,
7277                       u64 bytenr, u64 num_bytes, u64 parent, u64 root_objectid,
7278                       u64 owner, u64 offset)
7279 {
7280         struct btrfs_fs_info *fs_info = root->fs_info;
7281         int old_ref_mod, new_ref_mod;
7282         int ret;
7283
7284         if (btrfs_is_testing(fs_info))
7285                 return 0;
7286
7287         if (root_objectid != BTRFS_TREE_LOG_OBJECTID)
7288                 btrfs_ref_tree_mod(root, bytenr, num_bytes, parent,
7289                                    root_objectid, owner, offset,
7290                                    BTRFS_DROP_DELAYED_REF);
7291
7292         /*
7293          * tree log blocks never actually go into the extent allocation
7294          * tree, just update pinning info and exit early.
7295          */
7296         if (root_objectid == BTRFS_TREE_LOG_OBJECTID) {
7297                 WARN_ON(owner >= BTRFS_FIRST_FREE_OBJECTID);
7298                 /* unlocks the pinned mutex */
7299                 btrfs_pin_extent(fs_info, bytenr, num_bytes, 1);
7300                 old_ref_mod = new_ref_mod = 0;
7301                 ret = 0;
7302         } else if (owner < BTRFS_FIRST_FREE_OBJECTID) {
7303                 ret = btrfs_add_delayed_tree_ref(fs_info, trans, bytenr,
7304                                                  num_bytes, parent,
7305                                                  root_objectid, (int)owner,
7306                                                  BTRFS_DROP_DELAYED_REF, NULL,
7307                                                  &old_ref_mod, &new_ref_mod);
7308         } else {
7309                 ret = btrfs_add_delayed_data_ref(fs_info, trans, bytenr,
7310                                                  num_bytes, parent,
7311                                                  root_objectid, owner, offset,
7312                                                  0, BTRFS_DROP_DELAYED_REF,
7313                                                  &old_ref_mod, &new_ref_mod);
7314         }
7315
7316         if (ret == 0 && old_ref_mod >= 0 && new_ref_mod < 0)
7317                 add_pinned_bytes(fs_info, num_bytes, owner, root_objectid);
7318
7319         return ret;
7320 }
7321
7322 /*
7323  * when we wait for progress in the block group caching, its because
7324  * our allocation attempt failed at least once.  So, we must sleep
7325  * and let some progress happen before we try again.
7326  *
7327  * This function will sleep at least once waiting for new free space to
7328  * show up, and then it will check the block group free space numbers
7329  * for our min num_bytes.  Another option is to have it go ahead
7330  * and look in the rbtree for a free extent of a given size, but this
7331  * is a good start.
7332  *
7333  * Callers of this must check if cache->cached == BTRFS_CACHE_ERROR before using
7334  * any of the information in this block group.
7335  */
7336 static noinline void
7337 wait_block_group_cache_progress(struct btrfs_block_group_cache *cache,
7338                                 u64 num_bytes)
7339 {
7340         struct btrfs_caching_control *caching_ctl;
7341
7342         caching_ctl = get_caching_control(cache);
7343         if (!caching_ctl)
7344                 return;
7345
7346         wait_event(caching_ctl->wait, block_group_cache_done(cache) ||
7347                    (cache->free_space_ctl->free_space >= num_bytes));
7348
7349         put_caching_control(caching_ctl);
7350 }
7351
7352 static noinline int
7353 wait_block_group_cache_done(struct btrfs_block_group_cache *cache)
7354 {
7355         struct btrfs_caching_control *caching_ctl;
7356         int ret = 0;
7357
7358         caching_ctl = get_caching_control(cache);
7359         if (!caching_ctl)
7360                 return (cache->cached == BTRFS_CACHE_ERROR) ? -EIO : 0;
7361
7362         wait_event(caching_ctl->wait, block_group_cache_done(cache));
7363         if (cache->cached == BTRFS_CACHE_ERROR)
7364                 ret = -EIO;
7365         put_caching_control(caching_ctl);
7366         return ret;
7367 }
7368
7369 static const char *btrfs_raid_type_names[BTRFS_NR_RAID_TYPES] = {
7370         [BTRFS_RAID_RAID10]     = "raid10",
7371         [BTRFS_RAID_RAID1]      = "raid1",
7372         [BTRFS_RAID_DUP]        = "dup",
7373         [BTRFS_RAID_RAID0]      = "raid0",
7374         [BTRFS_RAID_SINGLE]     = "single",
7375         [BTRFS_RAID_RAID5]      = "raid5",
7376         [BTRFS_RAID_RAID6]      = "raid6",
7377 };
7378
7379 static const char *get_raid_name(enum btrfs_raid_types type)
7380 {
7381         if (type >= BTRFS_NR_RAID_TYPES)
7382                 return NULL;
7383
7384         return btrfs_raid_type_names[type];
7385 }
7386
7387 enum btrfs_loop_type {
7388         LOOP_CACHING_NOWAIT = 0,
7389         LOOP_CACHING_WAIT = 1,
7390         LOOP_ALLOC_CHUNK = 2,
7391         LOOP_NO_EMPTY_SIZE = 3,
7392 };
7393
7394 static inline void
7395 btrfs_lock_block_group(struct btrfs_block_group_cache *cache,
7396                        int delalloc)
7397 {
7398         if (delalloc)
7399                 down_read(&cache->data_rwsem);
7400 }
7401
7402 static inline void
7403 btrfs_grab_block_group(struct btrfs_block_group_cache *cache,
7404                        int delalloc)
7405 {
7406         btrfs_get_block_group(cache);
7407         if (delalloc)
7408                 down_read(&cache->data_rwsem);
7409 }
7410
7411 static struct btrfs_block_group_cache *
7412 btrfs_lock_cluster(struct btrfs_block_group_cache *block_group,
7413                    struct btrfs_free_cluster *cluster,
7414                    int delalloc)
7415 {
7416         struct btrfs_block_group_cache *used_bg = NULL;
7417
7418         spin_lock(&cluster->refill_lock);
7419         while (1) {
7420                 used_bg = cluster->block_group;
7421                 if (!used_bg)
7422                         return NULL;
7423
7424                 if (used_bg == block_group)
7425                         return used_bg;
7426
7427                 btrfs_get_block_group(used_bg);
7428
7429                 if (!delalloc)
7430                         return used_bg;
7431
7432                 if (down_read_trylock(&used_bg->data_rwsem))
7433                         return used_bg;
7434
7435                 spin_unlock(&cluster->refill_lock);
7436
7437                 /* We should only have one-level nested. */
7438                 down_read_nested(&used_bg->data_rwsem, SINGLE_DEPTH_NESTING);
7439
7440                 spin_lock(&cluster->refill_lock);
7441                 if (used_bg == cluster->block_group)
7442                         return used_bg;
7443
7444                 up_read(&used_bg->data_rwsem);
7445                 btrfs_put_block_group(used_bg);
7446         }
7447 }
7448
7449 static inline void
7450 btrfs_release_block_group(struct btrfs_block_group_cache *cache,
7451                          int delalloc)
7452 {
7453         if (delalloc)
7454                 up_read(&cache->data_rwsem);
7455         btrfs_put_block_group(cache);
7456 }
7457
7458 /*
7459  * walks the btree of allocated extents and find a hole of a given size.
7460  * The key ins is changed to record the hole:
7461  * ins->objectid == start position
7462  * ins->flags = BTRFS_EXTENT_ITEM_KEY
7463  * ins->offset == the size of the hole.
7464  * Any available blocks before search_start are skipped.
7465  *
7466  * If there is no suitable free space, we will record the max size of
7467  * the free space extent currently.
7468  */
7469 static noinline int find_free_extent(struct btrfs_fs_info *fs_info,
7470                                 u64 ram_bytes, u64 num_bytes, u64 empty_size,
7471                                 u64 hint_byte, struct btrfs_key *ins,
7472                                 u64 flags, int delalloc)
7473 {
7474         int ret = 0;
7475         struct btrfs_root *root = fs_info->extent_root;
7476         struct btrfs_free_cluster *last_ptr = NULL;
7477         struct btrfs_block_group_cache *block_group = NULL;
7478         u64 search_start = 0;
7479         u64 max_extent_size = 0;
7480         u64 empty_cluster = 0;
7481         struct btrfs_space_info *space_info;
7482         int loop = 0;
7483         int index = btrfs_bg_flags_to_raid_index(flags);
7484         bool failed_cluster_refill = false;
7485         bool failed_alloc = false;
7486         bool use_cluster = true;
7487         bool have_caching_bg = false;
7488         bool orig_have_caching_bg = false;
7489         bool full_search = false;
7490
7491         WARN_ON(num_bytes < fs_info->sectorsize);
7492         ins->type = BTRFS_EXTENT_ITEM_KEY;
7493         ins->objectid = 0;
7494         ins->offset = 0;
7495
7496         trace_find_free_extent(fs_info, num_bytes, empty_size, flags);
7497
7498         space_info = __find_space_info(fs_info, flags);
7499         if (!space_info) {
7500                 btrfs_err(fs_info, "No space info for %llu", flags);
7501                 return -ENOSPC;
7502         }
7503
7504         /*
7505          * If our free space is heavily fragmented we may not be able to make
7506          * big contiguous allocations, so instead of doing the expensive search
7507          * for free space, simply return ENOSPC with our max_extent_size so we
7508          * can go ahead and search for a more manageable chunk.
7509          *
7510          * If our max_extent_size is large enough for our allocation simply
7511          * disable clustering since we will likely not be able to find enough
7512          * space to create a cluster and induce latency trying.
7513          */
7514         if (unlikely(space_info->max_extent_size)) {
7515                 spin_lock(&space_info->lock);
7516                 if (space_info->max_extent_size &&
7517                     num_bytes > space_info->max_extent_size) {
7518                         ins->offset = space_info->max_extent_size;
7519                         spin_unlock(&space_info->lock);
7520                         return -ENOSPC;
7521                 } else if (space_info->max_extent_size) {
7522                         use_cluster = false;
7523                 }
7524                 spin_unlock(&space_info->lock);
7525         }
7526
7527         last_ptr = fetch_cluster_info(fs_info, space_info, &empty_cluster);
7528         if (last_ptr) {
7529                 spin_lock(&last_ptr->lock);
7530                 if (last_ptr->block_group)
7531                         hint_byte = last_ptr->window_start;
7532                 if (last_ptr->fragmented) {
7533                         /*
7534                          * We still set window_start so we can keep track of the
7535                          * last place we found an allocation to try and save
7536                          * some time.
7537                          */
7538                         hint_byte = last_ptr->window_start;
7539                         use_cluster = false;
7540                 }
7541                 spin_unlock(&last_ptr->lock);
7542         }
7543
7544         search_start = max(search_start, first_logical_byte(fs_info, 0));
7545         search_start = max(search_start, hint_byte);
7546         if (search_start == hint_byte) {
7547                 block_group = btrfs_lookup_block_group(fs_info, search_start);
7548                 /*
7549                  * we don't want to use the block group if it doesn't match our
7550                  * allocation bits, or if its not cached.
7551                  *
7552                  * However if we are re-searching with an ideal block group
7553                  * picked out then we don't care that the block group is cached.
7554                  */
7555                 if (block_group && block_group_bits(block_group, flags) &&
7556                     block_group->cached != BTRFS_CACHE_NO) {
7557                         down_read(&space_info->groups_sem);
7558                         if (list_empty(&block_group->list) ||
7559                             block_group->ro) {
7560                                 /*
7561                                  * someone is removing this block group,
7562                                  * we can't jump into the have_block_group
7563                                  * target because our list pointers are not
7564                                  * valid
7565                                  */
7566                                 btrfs_put_block_group(block_group);
7567                                 up_read(&space_info->groups_sem);
7568                         } else {
7569                                 index = btrfs_bg_flags_to_raid_index(
7570                                                 block_group->flags);
7571                                 btrfs_lock_block_group(block_group, delalloc);
7572                                 goto have_block_group;
7573                         }
7574                 } else if (block_group) {
7575                         btrfs_put_block_group(block_group);
7576                 }
7577         }
7578 search:
7579         have_caching_bg = false;
7580         if (index == 0 || index == btrfs_bg_flags_to_raid_index(flags))
7581                 full_search = true;
7582         down_read(&space_info->groups_sem);
7583         list_for_each_entry(block_group, &space_info->block_groups[index],
7584                             list) {
7585                 u64 offset;
7586                 int cached;
7587
7588                 /* If the block group is read-only, we can skip it entirely. */
7589                 if (unlikely(block_group->ro))
7590                         continue;
7591
7592                 btrfs_grab_block_group(block_group, delalloc);
7593                 search_start = block_group->key.objectid;
7594
7595                 /*
7596                  * this can happen if we end up cycling through all the
7597                  * raid types, but we want to make sure we only allocate
7598                  * for the proper type.
7599                  */
7600                 if (!block_group_bits(block_group, flags)) {
7601                     u64 extra = BTRFS_BLOCK_GROUP_DUP |
7602                                 BTRFS_BLOCK_GROUP_RAID1 |
7603                                 BTRFS_BLOCK_GROUP_RAID5 |
7604                                 BTRFS_BLOCK_GROUP_RAID6 |
7605                                 BTRFS_BLOCK_GROUP_RAID10;
7606
7607                         /*
7608                          * if they asked for extra copies and this block group
7609                          * doesn't provide them, bail.  This does allow us to
7610                          * fill raid0 from raid1.
7611                          */
7612                         if ((flags & extra) && !(block_group->flags & extra))
7613                                 goto loop;
7614                 }
7615
7616 have_block_group:
7617                 cached = block_group_cache_done(block_group);
7618                 if (unlikely(!cached)) {
7619                         have_caching_bg = true;
7620                         ret = cache_block_group(block_group, 0);
7621                         BUG_ON(ret < 0);
7622                         ret = 0;
7623                 }
7624
7625                 if (unlikely(block_group->cached == BTRFS_CACHE_ERROR))
7626                         goto loop;
7627
7628                 /*
7629                  * Ok we want to try and use the cluster allocator, so
7630                  * lets look there
7631                  */
7632                 if (last_ptr && use_cluster) {
7633                         struct btrfs_block_group_cache *used_block_group;
7634                         unsigned long aligned_cluster;
7635                         /*
7636                          * the refill lock keeps out other
7637                          * people trying to start a new cluster
7638                          */
7639                         used_block_group = btrfs_lock_cluster(block_group,
7640                                                               last_ptr,
7641                                                               delalloc);
7642                         if (!used_block_group)
7643                                 goto refill_cluster;
7644
7645                         if (used_block_group != block_group &&
7646                             (used_block_group->ro ||
7647                              !block_group_bits(used_block_group, flags)))
7648                                 goto release_cluster;
7649
7650                         offset = btrfs_alloc_from_cluster(used_block_group,
7651                                                 last_ptr,
7652                                                 num_bytes,
7653                                                 used_block_group->key.objectid,
7654                                                 &max_extent_size);
7655                         if (offset) {
7656                                 /* we have a block, we're done */
7657                                 spin_unlock(&last_ptr->refill_lock);
7658                                 trace_btrfs_reserve_extent_cluster(fs_info,
7659                                                 used_block_group,
7660                                                 search_start, num_bytes);
7661                                 if (used_block_group != block_group) {
7662                                         btrfs_release_block_group(block_group,
7663                                                                   delalloc);
7664                                         block_group = used_block_group;
7665                                 }
7666                                 goto checks;
7667                         }
7668
7669                         WARN_ON(last_ptr->block_group != used_block_group);
7670 release_cluster:
7671                         /* If we are on LOOP_NO_EMPTY_SIZE, we can't
7672                          * set up a new clusters, so lets just skip it
7673                          * and let the allocator find whatever block
7674                          * it can find.  If we reach this point, we
7675                          * will have tried the cluster allocator
7676                          * plenty of times and not have found
7677                          * anything, so we are likely way too
7678                          * fragmented for the clustering stuff to find
7679                          * anything.
7680                          *
7681                          * However, if the cluster is taken from the
7682                          * current block group, release the cluster
7683                          * first, so that we stand a better chance of
7684                          * succeeding in the unclustered
7685                          * allocation.  */
7686                         if (loop >= LOOP_NO_EMPTY_SIZE &&
7687                             used_block_group != block_group) {
7688                                 spin_unlock(&last_ptr->refill_lock);
7689                                 btrfs_release_block_group(used_block_group,
7690                                                           delalloc);
7691                                 goto unclustered_alloc;
7692                         }
7693
7694                         /*
7695                          * this cluster didn't work out, free it and
7696                          * start over
7697                          */
7698                         btrfs_return_cluster_to_free_space(NULL, last_ptr);
7699
7700                         if (used_block_group != block_group)
7701                                 btrfs_release_block_group(used_block_group,
7702                                                           delalloc);
7703 refill_cluster:
7704                         if (loop >= LOOP_NO_EMPTY_SIZE) {
7705                                 spin_unlock(&last_ptr->refill_lock);
7706                                 goto unclustered_alloc;
7707                         }
7708
7709                         aligned_cluster = max_t(unsigned long,
7710                                                 empty_cluster + empty_size,
7711                                               block_group->full_stripe_len);
7712
7713                         /* allocate a cluster in this block group */
7714                         ret = btrfs_find_space_cluster(fs_info, block_group,
7715                                                        last_ptr, search_start,
7716                                                        num_bytes,
7717                                                        aligned_cluster);
7718                         if (ret == 0) {
7719                                 /*
7720                                  * now pull our allocation out of this
7721                                  * cluster
7722                                  */
7723                                 offset = btrfs_alloc_from_cluster(block_group,
7724                                                         last_ptr,
7725                                                         num_bytes,
7726                                                         search_start,
7727                                                         &max_extent_size);
7728                                 if (offset) {
7729                                         /* we found one, proceed */
7730                                         spin_unlock(&last_ptr->refill_lock);
7731                                         trace_btrfs_reserve_extent_cluster(fs_info,
7732                                                 block_group, search_start,
7733                                                 num_bytes);
7734                                         goto checks;
7735                                 }
7736                         } else if (!cached && loop > LOOP_CACHING_NOWAIT
7737                                    && !failed_cluster_refill) {
7738                                 spin_unlock(&last_ptr->refill_lock);
7739
7740                                 failed_cluster_refill = true;
7741                                 wait_block_group_cache_progress(block_group,
7742                                        num_bytes + empty_cluster + empty_size);
7743                                 goto have_block_group;
7744                         }
7745
7746                         /*
7747                          * at this point we either didn't find a cluster
7748                          * or we weren't able to allocate a block from our
7749                          * cluster.  Free the cluster we've been trying
7750                          * to use, and go to the next block group
7751                          */
7752                         btrfs_return_cluster_to_free_space(NULL, last_ptr);
7753                         spin_unlock(&last_ptr->refill_lock);
7754                         goto loop;
7755                 }
7756
7757 unclustered_alloc:
7758                 /*
7759                  * We are doing an unclustered alloc, set the fragmented flag so
7760                  * we don't bother trying to setup a cluster again until we get
7761                  * more space.
7762                  */
7763                 if (unlikely(last_ptr)) {
7764                         spin_lock(&last_ptr->lock);
7765                         last_ptr->fragmented = 1;
7766                         spin_unlock(&last_ptr->lock);
7767                 }
7768                 if (cached) {
7769                         struct btrfs_free_space_ctl *ctl =
7770                                 block_group->free_space_ctl;
7771
7772                         spin_lock(&ctl->tree_lock);
7773                         if (ctl->free_space <
7774                             num_bytes + empty_cluster + empty_size) {
7775                                 if (ctl->free_space > max_extent_size)
7776                                         max_extent_size = ctl->free_space;
7777                                 spin_unlock(&ctl->tree_lock);
7778                                 goto loop;
7779                         }
7780                         spin_unlock(&ctl->tree_lock);
7781                 }
7782
7783                 offset = btrfs_find_space_for_alloc(block_group, search_start,
7784                                                     num_bytes, empty_size,
7785                                                     &max_extent_size);
7786                 /*
7787                  * If we didn't find a chunk, and we haven't failed on this
7788                  * block group before, and this block group is in the middle of
7789                  * caching and we are ok with waiting, then go ahead and wait
7790                  * for progress to be made, and set failed_alloc to true.
7791                  *
7792                  * If failed_alloc is true then we've already waited on this
7793                  * block group once and should move on to the next block group.
7794                  */
7795                 if (!offset && !failed_alloc && !cached &&
7796                     loop > LOOP_CACHING_NOWAIT) {
7797                         wait_block_group_cache_progress(block_group,
7798                                                 num_bytes + empty_size);
7799                         failed_alloc = true;
7800                         goto have_block_group;
7801                 } else if (!offset) {
7802                         goto loop;
7803                 }
7804 checks:
7805                 search_start = ALIGN(offset, fs_info->stripesize);
7806
7807                 /* move on to the next group */
7808                 if (search_start + num_bytes >
7809                     block_group->key.objectid + block_group->key.offset) {
7810                         btrfs_add_free_space(block_group, offset, num_bytes);
7811                         goto loop;
7812                 }
7813
7814                 if (offset < search_start)
7815                         btrfs_add_free_space(block_group, offset,
7816                                              search_start - offset);
7817                 BUG_ON(offset > search_start);
7818
7819                 ret = btrfs_add_reserved_bytes(block_group, ram_bytes,
7820                                 num_bytes, delalloc);
7821                 if (ret == -EAGAIN) {
7822                         btrfs_add_free_space(block_group, offset, num_bytes);
7823                         goto loop;
7824                 }
7825                 btrfs_inc_block_group_reservations(block_group);
7826
7827                 /* we are all good, lets return */
7828                 ins->objectid = search_start;
7829                 ins->offset = num_bytes;
7830
7831                 trace_btrfs_reserve_extent(fs_info, block_group,
7832                                            search_start, num_bytes);
7833                 btrfs_release_block_group(block_group, delalloc);
7834                 break;
7835 loop:
7836                 failed_cluster_refill = false;
7837                 failed_alloc = false;
7838                 BUG_ON(btrfs_bg_flags_to_raid_index(block_group->flags) !=
7839                        index);
7840                 btrfs_release_block_group(block_group, delalloc);
7841                 cond_resched();
7842         }
7843         up_read(&space_info->groups_sem);
7844
7845         if ((loop == LOOP_CACHING_NOWAIT) && have_caching_bg
7846                 && !orig_have_caching_bg)
7847                 orig_have_caching_bg = true;
7848
7849         if (!ins->objectid && loop >= LOOP_CACHING_WAIT && have_caching_bg)
7850                 goto search;
7851
7852         if (!ins->objectid && ++index < BTRFS_NR_RAID_TYPES)
7853                 goto search;
7854
7855         /*
7856          * LOOP_CACHING_NOWAIT, search partially cached block groups, kicking
7857          *                      caching kthreads as we move along
7858          * LOOP_CACHING_WAIT, search everything, and wait if our bg is caching
7859          * LOOP_ALLOC_CHUNK, force a chunk allocation and try again
7860          * LOOP_NO_EMPTY_SIZE, set empty_size and empty_cluster to 0 and try
7861          *                      again
7862          */
7863         if (!ins->objectid && loop < LOOP_NO_EMPTY_SIZE) {
7864                 index = 0;
7865                 if (loop == LOOP_CACHING_NOWAIT) {
7866                         /*
7867                          * We want to skip the LOOP_CACHING_WAIT step if we
7868                          * don't have any uncached bgs and we've already done a
7869                          * full search through.
7870                          */
7871                         if (orig_have_caching_bg || !full_search)
7872                                 loop = LOOP_CACHING_WAIT;
7873                         else
7874                                 loop = LOOP_ALLOC_CHUNK;
7875                 } else {
7876                         loop++;
7877                 }
7878
7879                 if (loop == LOOP_ALLOC_CHUNK) {
7880                         struct btrfs_trans_handle *trans;
7881                         int exist = 0;
7882
7883                         trans = current->journal_info;
7884                         if (trans)
7885                                 exist = 1;
7886                         else
7887                                 trans = btrfs_join_transaction(root);
7888
7889                         if (IS_ERR(trans)) {
7890                                 ret = PTR_ERR(trans);
7891                                 goto out;
7892                         }
7893
7894                         ret = do_chunk_alloc(trans, fs_info, flags,
7895                                              CHUNK_ALLOC_FORCE);
7896
7897                         /*
7898                          * If we can't allocate a new chunk we've already looped
7899                          * through at least once, move on to the NO_EMPTY_SIZE
7900                          * case.
7901                          */
7902                         if (ret == -ENOSPC)
7903                                 loop = LOOP_NO_EMPTY_SIZE;
7904
7905                         /*
7906                          * Do not bail out on ENOSPC since we
7907                          * can do more things.
7908                          */
7909                         if (ret < 0 && ret != -ENOSPC)
7910                                 btrfs_abort_transaction(trans, ret);
7911                         else
7912                                 ret = 0;
7913                         if (!exist)
7914                                 btrfs_end_transaction(trans);
7915                         if (ret)
7916                                 goto out;
7917                 }
7918
7919                 if (loop == LOOP_NO_EMPTY_SIZE) {
7920                         /*
7921                          * Don't loop again if we already have no empty_size and
7922                          * no empty_cluster.
7923                          */
7924                         if (empty_size == 0 &&
7925                             empty_cluster == 0) {
7926                                 ret = -ENOSPC;
7927                                 goto out;
7928                         }
7929                         empty_size = 0;
7930                         empty_cluster = 0;
7931                 }
7932
7933                 goto search;
7934         } else if (!ins->objectid) {
7935                 ret = -ENOSPC;
7936         } else if (ins->objectid) {
7937                 if (!use_cluster && last_ptr) {
7938                         spin_lock(&last_ptr->lock);
7939                         last_ptr->window_start = ins->objectid;
7940                         spin_unlock(&last_ptr->lock);
7941                 }
7942                 ret = 0;
7943         }
7944 out:
7945         if (ret == -ENOSPC) {
7946                 spin_lock(&space_info->lock);
7947                 space_info->max_extent_size = max_extent_size;
7948                 spin_unlock(&space_info->lock);
7949                 ins->offset = max_extent_size;
7950         }
7951         return ret;
7952 }
7953
7954 static void dump_space_info(struct btrfs_fs_info *fs_info,
7955                             struct btrfs_space_info *info, u64 bytes,
7956                             int dump_block_groups)
7957 {
7958         struct btrfs_block_group_cache *cache;
7959         int index = 0;
7960
7961         spin_lock(&info->lock);
7962         btrfs_info(fs_info, "space_info %llu has %llu free, is %sfull",
7963                    info->flags,
7964                    info->total_bytes - btrfs_space_info_used(info, true),
7965                    info->full ? "" : "not ");
7966         btrfs_info(fs_info,
7967                 "space_info total=%llu, used=%llu, pinned=%llu, reserved=%llu, may_use=%llu, readonly=%llu",
7968                 info->total_bytes, info->bytes_used, info->bytes_pinned,
7969                 info->bytes_reserved, info->bytes_may_use,
7970                 info->bytes_readonly);
7971         spin_unlock(&info->lock);
7972
7973         if (!dump_block_groups)
7974                 return;
7975
7976         down_read(&info->groups_sem);
7977 again:
7978         list_for_each_entry(cache, &info->block_groups[index], list) {
7979                 spin_lock(&cache->lock);
7980                 btrfs_info(fs_info,
7981                         "block group %llu has %llu bytes, %llu used %llu pinned %llu reserved %s",
7982                         cache->key.objectid, cache->key.offset,
7983                         btrfs_block_group_used(&cache->item), cache->pinned,
7984                         cache->reserved, cache->ro ? "[readonly]" : "");
7985                 btrfs_dump_free_space(cache, bytes);
7986                 spin_unlock(&cache->lock);
7987         }
7988         if (++index < BTRFS_NR_RAID_TYPES)
7989                 goto again;
7990         up_read(&info->groups_sem);
7991 }
7992
7993 /*
7994  * btrfs_reserve_extent - entry point to the extent allocator. Tries to find a
7995  *                        hole that is at least as big as @num_bytes.
7996  *
7997  * @root           -    The root that will contain this extent
7998  *
7999  * @ram_bytes      -    The amount of space in ram that @num_bytes take. This
8000  *                      is used for accounting purposes. This value differs
8001  *                      from @num_bytes only in the case of compressed extents.
8002  *
8003  * @num_bytes      -    Number of bytes to allocate on-disk.
8004  *
8005  * @min_alloc_size -    Indicates the minimum amount of space that the
8006  *                      allocator should try to satisfy. In some cases
8007  *                      @num_bytes may be larger than what is required and if
8008  *                      the filesystem is fragmented then allocation fails.
8009  *                      However, the presence of @min_alloc_size gives a
8010  *                      chance to try and satisfy the smaller allocation.
8011  *
8012  * @empty_size     -    A hint that you plan on doing more COW. This is the
8013  *                      size in bytes the allocator should try to find free
8014  *                      next to the block it returns.  This is just a hint and
8015  *                      may be ignored by the allocator.
8016  *
8017  * @hint_byte      -    Hint to the allocator to start searching above the byte
8018  *                      address passed. It might be ignored.
8019  *
8020  * @ins            -    This key is modified to record the found hole. It will
8021  *                      have the following values:
8022  *                      ins->objectid == start position
8023  *                      ins->flags = BTRFS_EXTENT_ITEM_KEY
8024  *                      ins->offset == the size of the hole.
8025  *
8026  * @is_data        -    Boolean flag indicating whether an extent is
8027  *                      allocated for data (true) or metadata (false)
8028  *
8029  * @delalloc       -    Boolean flag indicating whether this allocation is for
8030  *                      delalloc or not. If 'true' data_rwsem of block groups
8031  *                      is going to be acquired.
8032  *
8033  *
8034  * Returns 0 when an allocation succeeded or < 0 when an error occurred. In
8035  * case -ENOSPC is returned then @ins->offset will contain the size of the
8036  * largest available hole the allocator managed to find.
8037  */
8038 int btrfs_reserve_extent(struct btrfs_root *root, u64 ram_bytes,
8039                          u64 num_bytes, u64 min_alloc_size,
8040                          u64 empty_size, u64 hint_byte,
8041                          struct btrfs_key *ins, int is_data, int delalloc)
8042 {
8043         struct btrfs_fs_info *fs_info = root->fs_info;
8044         bool final_tried = num_bytes == min_alloc_size;
8045         u64 flags;
8046         int ret;
8047
8048         flags = get_alloc_profile_by_root(root, is_data);
8049 again:
8050         WARN_ON(num_bytes < fs_info->sectorsize);
8051         ret = find_free_extent(fs_info, ram_bytes, num_bytes, empty_size,
8052                                hint_byte, ins, flags, delalloc);
8053         if (!ret && !is_data) {
8054                 btrfs_dec_block_group_reservations(fs_info, ins->objectid);
8055         } else if (ret == -ENOSPC) {
8056                 if (!final_tried && ins->offset) {
8057                         num_bytes = min(num_bytes >> 1, ins->offset);
8058                         num_bytes = round_down(num_bytes,
8059                                                fs_info->sectorsize);
8060                         num_bytes = max(num_bytes, min_alloc_size);
8061                         ram_bytes = num_bytes;
8062                         if (num_bytes == min_alloc_size)
8063                                 final_tried = true;
8064                         goto again;
8065                 } else if (btrfs_test_opt(fs_info, ENOSPC_DEBUG)) {
8066                         struct btrfs_space_info *sinfo;
8067
8068                         sinfo = __find_space_info(fs_info, flags);
8069                         btrfs_err(fs_info,
8070                                   "allocation failed flags %llu, wanted %llu",
8071                                   flags, num_bytes);
8072                         if (sinfo)
8073                                 dump_space_info(fs_info, sinfo, num_bytes, 1);
8074                 }
8075         }
8076
8077         return ret;
8078 }
8079
8080 static int __btrfs_free_reserved_extent(struct btrfs_fs_info *fs_info,
8081                                         u64 start, u64 len,
8082                                         int pin, int delalloc)
8083 {
8084         struct btrfs_block_group_cache *cache;
8085         int ret = 0;
8086
8087         cache = btrfs_lookup_block_group(fs_info, start);
8088         if (!cache) {
8089                 btrfs_err(fs_info, "Unable to find block group for %llu",
8090                           start);
8091                 return -ENOSPC;
8092         }
8093
8094         if (pin)
8095                 pin_down_extent(fs_info, cache, start, len, 1);
8096         else {
8097                 if (btrfs_test_opt(fs_info, DISCARD))
8098                         ret = btrfs_discard_extent(fs_info, start, len, NULL);
8099                 btrfs_add_free_space(cache, start, len);
8100                 btrfs_free_reserved_bytes(cache, len, delalloc);
8101                 trace_btrfs_reserved_extent_free(fs_info, start, len);
8102         }
8103
8104         btrfs_put_block_group(cache);
8105         return ret;
8106 }
8107
8108 int btrfs_free_reserved_extent(struct btrfs_fs_info *fs_info,
8109                                u64 start, u64 len, int delalloc)
8110 {
8111         return __btrfs_free_reserved_extent(fs_info, start, len, 0, delalloc);
8112 }
8113
8114 int btrfs_free_and_pin_reserved_extent(struct btrfs_fs_info *fs_info,
8115                                        u64 start, u64 len)
8116 {
8117         return __btrfs_free_reserved_extent(fs_info, start, len, 1, 0);
8118 }
8119
8120 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
8121                                       struct btrfs_fs_info *fs_info,
8122                                       u64 parent, u64 root_objectid,
8123                                       u64 flags, u64 owner, u64 offset,
8124                                       struct btrfs_key *ins, int ref_mod)
8125 {
8126         int ret;
8127         struct btrfs_extent_item *extent_item;
8128         struct btrfs_extent_inline_ref *iref;
8129         struct btrfs_path *path;
8130         struct extent_buffer *leaf;
8131         int type;
8132         u32 size;
8133
8134         if (parent > 0)
8135                 type = BTRFS_SHARED_DATA_REF_KEY;
8136         else
8137                 type = BTRFS_EXTENT_DATA_REF_KEY;
8138
8139         size = sizeof(*extent_item) + btrfs_extent_inline_ref_size(type);
8140
8141         path = btrfs_alloc_path();
8142         if (!path)
8143                 return -ENOMEM;
8144
8145         path->leave_spinning = 1;
8146         ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path,
8147                                       ins, size);
8148         if (ret) {
8149                 btrfs_free_path(path);
8150                 return ret;
8151         }
8152
8153         leaf = path->nodes[0];
8154         extent_item = btrfs_item_ptr(leaf, path->slots[0],
8155                                      struct btrfs_extent_item);
8156         btrfs_set_extent_refs(leaf, extent_item, ref_mod);
8157         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
8158         btrfs_set_extent_flags(leaf, extent_item,
8159                                flags | BTRFS_EXTENT_FLAG_DATA);
8160
8161         iref = (struct btrfs_extent_inline_ref *)(extent_item + 1);
8162         btrfs_set_extent_inline_ref_type(leaf, iref, type);
8163         if (parent > 0) {
8164                 struct btrfs_shared_data_ref *ref;
8165                 ref = (struct btrfs_shared_data_ref *)(iref + 1);
8166                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
8167                 btrfs_set_shared_data_ref_count(leaf, ref, ref_mod);
8168         } else {
8169                 struct btrfs_extent_data_ref *ref;
8170                 ref = (struct btrfs_extent_data_ref *)(&iref->offset);
8171                 btrfs_set_extent_data_ref_root(leaf, ref, root_objectid);
8172                 btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
8173                 btrfs_set_extent_data_ref_offset(leaf, ref, offset);
8174                 btrfs_set_extent_data_ref_count(leaf, ref, ref_mod);
8175         }
8176
8177         btrfs_mark_buffer_dirty(path->nodes[0]);
8178         btrfs_free_path(path);
8179
8180         ret = remove_from_free_space_tree(trans, fs_info, ins->objectid,
8181                                           ins->offset);
8182         if (ret)
8183                 return ret;
8184
8185         ret = update_block_group(trans, fs_info, ins->objectid, ins->offset, 1);
8186         if (ret) { /* -ENOENT, logic error */
8187                 btrfs_err(fs_info, "update block group failed for %llu %llu",
8188                         ins->objectid, ins->offset);
8189                 BUG();
8190         }
8191         trace_btrfs_reserved_extent_alloc(fs_info, ins->objectid, ins->offset);
8192         return ret;
8193 }
8194
8195 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
8196                                      struct btrfs_fs_info *fs_info,
8197                                      u64 parent, u64 root_objectid,
8198                                      u64 flags, struct btrfs_disk_key *key,
8199                                      int level, struct btrfs_key *ins)
8200 {
8201         int ret;
8202         struct btrfs_extent_item *extent_item;
8203         struct btrfs_tree_block_info *block_info;
8204         struct btrfs_extent_inline_ref *iref;
8205         struct btrfs_path *path;
8206         struct extent_buffer *leaf;
8207         u32 size = sizeof(*extent_item) + sizeof(*iref);
8208         u64 num_bytes = ins->offset;
8209         bool skinny_metadata = btrfs_fs_incompat(fs_info, SKINNY_METADATA);
8210
8211         if (!skinny_metadata)
8212                 size += sizeof(*block_info);
8213
8214         path = btrfs_alloc_path();
8215         if (!path) {
8216                 btrfs_free_and_pin_reserved_extent(fs_info, ins->objectid,
8217                                                    fs_info->nodesize);
8218                 return -ENOMEM;
8219         }
8220
8221         path->leave_spinning = 1;
8222         ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path,
8223                                       ins, size);
8224         if (ret) {
8225                 btrfs_free_path(path);
8226                 btrfs_free_and_pin_reserved_extent(fs_info, ins->objectid,
8227                                                    fs_info->nodesize);
8228                 return ret;
8229         }
8230
8231         leaf = path->nodes[0];
8232         extent_item = btrfs_item_ptr(leaf, path->slots[0],
8233                                      struct btrfs_extent_item);
8234         btrfs_set_extent_refs(leaf, extent_item, 1);
8235         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
8236         btrfs_set_extent_flags(leaf, extent_item,
8237                                flags | BTRFS_EXTENT_FLAG_TREE_BLOCK);
8238
8239         if (skinny_metadata) {
8240                 iref = (struct btrfs_extent_inline_ref *)(extent_item + 1);
8241                 num_bytes = fs_info->nodesize;
8242         } else {
8243                 block_info = (struct btrfs_tree_block_info *)(extent_item + 1);
8244                 btrfs_set_tree_block_key(leaf, block_info, key);
8245                 btrfs_set_tree_block_level(leaf, block_info, level);
8246                 iref = (struct btrfs_extent_inline_ref *)(block_info + 1);
8247         }
8248
8249         if (parent > 0) {
8250                 BUG_ON(!(flags & BTRFS_BLOCK_FLAG_FULL_BACKREF));
8251                 btrfs_set_extent_inline_ref_type(leaf, iref,
8252                                                  BTRFS_SHARED_BLOCK_REF_KEY);
8253                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
8254         } else {
8255                 btrfs_set_extent_inline_ref_type(leaf, iref,
8256                                                  BTRFS_TREE_BLOCK_REF_KEY);
8257                 btrfs_set_extent_inline_ref_offset(leaf, iref, root_objectid);
8258         }
8259
8260         btrfs_mark_buffer_dirty(leaf);
8261         btrfs_free_path(path);
8262
8263         ret = remove_from_free_space_tree(trans, fs_info, ins->objectid,
8264                                           num_bytes);
8265         if (ret)
8266                 return ret;
8267
8268         ret = update_block_group(trans, fs_info, ins->objectid,
8269                                  fs_info->nodesize, 1);
8270         if (ret) { /* -ENOENT, logic error */
8271                 btrfs_err(fs_info, "update block group failed for %llu %llu",
8272                         ins->objectid, ins->offset);
8273                 BUG();
8274         }
8275
8276         trace_btrfs_reserved_extent_alloc(fs_info, ins->objectid,
8277                                           fs_info->nodesize);
8278         return ret;
8279 }
8280
8281 int btrfs_alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
8282                                      struct btrfs_root *root, u64 owner,
8283                                      u64 offset, u64 ram_bytes,
8284                                      struct btrfs_key *ins)
8285 {
8286         struct btrfs_fs_info *fs_info = root->fs_info;
8287         int ret;
8288
8289         BUG_ON(root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID);
8290
8291         btrfs_ref_tree_mod(root, ins->objectid, ins->offset, 0,
8292                            root->root_key.objectid, owner, offset,
8293                            BTRFS_ADD_DELAYED_EXTENT);
8294
8295         ret = btrfs_add_delayed_data_ref(fs_info, trans, ins->objectid,
8296                                          ins->offset, 0,
8297                                          root->root_key.objectid, owner,
8298                                          offset, ram_bytes,
8299                                          BTRFS_ADD_DELAYED_EXTENT, NULL, NULL);
8300         return ret;
8301 }
8302
8303 /*
8304  * this is used by the tree logging recovery code.  It records that
8305  * an extent has been allocated and makes sure to clear the free
8306  * space cache bits as well
8307  */
8308 int btrfs_alloc_logged_file_extent(struct btrfs_trans_handle *trans,
8309                                    struct btrfs_fs_info *fs_info,
8310                                    u64 root_objectid, u64 owner, u64 offset,
8311                                    struct btrfs_key *ins)
8312 {
8313         int ret;
8314         struct btrfs_block_group_cache *block_group;
8315         struct btrfs_space_info *space_info;
8316
8317         /*
8318          * Mixed block groups will exclude before processing the log so we only
8319          * need to do the exclude dance if this fs isn't mixed.
8320          */
8321         if (!btrfs_fs_incompat(fs_info, MIXED_GROUPS)) {
8322                 ret = __exclude_logged_extent(fs_info, ins->objectid,
8323                                               ins->offset);
8324                 if (ret)
8325                         return ret;
8326         }
8327
8328         block_group = btrfs_lookup_block_group(fs_info, ins->objectid);
8329         if (!block_group)
8330                 return -EINVAL;
8331
8332         space_info = block_group->space_info;
8333         spin_lock(&space_info->lock);
8334         spin_lock(&block_group->lock);
8335         space_info->bytes_reserved += ins->offset;
8336         block_group->reserved += ins->offset;
8337         spin_unlock(&block_group->lock);
8338         spin_unlock(&space_info->lock);
8339
8340         ret = alloc_reserved_file_extent(trans, fs_info, 0, root_objectid,
8341                                          0, owner, offset, ins, 1);
8342         btrfs_put_block_group(block_group);
8343         return ret;
8344 }
8345
8346 static struct extent_buffer *
8347 btrfs_init_new_buffer(struct btrfs_trans_handle *trans, struct btrfs_root *root,
8348                       u64 bytenr, int level)
8349 {
8350         struct btrfs_fs_info *fs_info = root->fs_info;
8351         struct extent_buffer *buf;
8352
8353         buf = btrfs_find_create_tree_block(fs_info, bytenr);
8354         if (IS_ERR(buf))
8355                 return buf;
8356
8357         btrfs_set_header_generation(buf, trans->transid);
8358         btrfs_set_buffer_lockdep_class(root->root_key.objectid, buf, level);
8359         btrfs_tree_lock(buf);
8360         clean_tree_block(fs_info, buf);
8361         clear_bit(EXTENT_BUFFER_STALE, &buf->bflags);
8362
8363         btrfs_set_lock_blocking(buf);
8364         set_extent_buffer_uptodate(buf);
8365
8366         if (root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID) {
8367                 buf->log_index = root->log_transid % 2;
8368                 /*
8369                  * we allow two log transactions at a time, use different
8370                  * EXENT bit to differentiate dirty pages.
8371                  */
8372                 if (buf->log_index == 0)
8373                         set_extent_dirty(&root->dirty_log_pages, buf->start,
8374                                         buf->start + buf->len - 1, GFP_NOFS);
8375                 else
8376                         set_extent_new(&root->dirty_log_pages, buf->start,
8377                                         buf->start + buf->len - 1);
8378         } else {
8379                 buf->log_index = -1;
8380                 set_extent_dirty(&trans->transaction->dirty_pages, buf->start,
8381                          buf->start + buf->len - 1, GFP_NOFS);
8382         }
8383         trans->dirty = true;
8384         /* this returns a buffer locked for blocking */
8385         return buf;
8386 }
8387
8388 static struct btrfs_block_rsv *
8389 use_block_rsv(struct btrfs_trans_handle *trans,
8390               struct btrfs_root *root, u32 blocksize)
8391 {
8392         struct btrfs_fs_info *fs_info = root->fs_info;
8393         struct btrfs_block_rsv *block_rsv;
8394         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
8395         int ret;
8396         bool global_updated = false;
8397
8398         block_rsv = get_block_rsv(trans, root);
8399
8400         if (unlikely(block_rsv->size == 0))
8401                 goto try_reserve;
8402 again:
8403         ret = block_rsv_use_bytes(block_rsv, blocksize);
8404         if (!ret)
8405                 return block_rsv;
8406
8407         if (block_rsv->failfast)
8408                 return ERR_PTR(ret);
8409
8410         if (block_rsv->type == BTRFS_BLOCK_RSV_GLOBAL && !global_updated) {
8411                 global_updated = true;
8412                 update_global_block_rsv(fs_info);
8413                 goto again;
8414         }
8415
8416         if (btrfs_test_opt(fs_info, ENOSPC_DEBUG)) {
8417                 static DEFINE_RATELIMIT_STATE(_rs,
8418                                 DEFAULT_RATELIMIT_INTERVAL * 10,
8419                                 /*DEFAULT_RATELIMIT_BURST*/ 1);
8420                 if (__ratelimit(&_rs))
8421                         WARN(1, KERN_DEBUG
8422                                 "BTRFS: block rsv returned %d\n", ret);
8423         }
8424 try_reserve:
8425         ret = reserve_metadata_bytes(root, block_rsv, blocksize,
8426                                      BTRFS_RESERVE_NO_FLUSH);
8427         if (!ret)
8428                 return block_rsv;
8429         /*
8430          * If we couldn't reserve metadata bytes try and use some from
8431          * the global reserve if its space type is the same as the global
8432          * reservation.
8433          */
8434         if (block_rsv->type != BTRFS_BLOCK_RSV_GLOBAL &&
8435             block_rsv->space_info == global_rsv->space_info) {
8436                 ret = block_rsv_use_bytes(global_rsv, blocksize);
8437                 if (!ret)
8438                         return global_rsv;
8439         }
8440         return ERR_PTR(ret);
8441 }
8442
8443 static void unuse_block_rsv(struct btrfs_fs_info *fs_info,
8444                             struct btrfs_block_rsv *block_rsv, u32 blocksize)
8445 {
8446         block_rsv_add_bytes(block_rsv, blocksize, 0);
8447         block_rsv_release_bytes(fs_info, block_rsv, NULL, 0, NULL);
8448 }
8449
8450 /*
8451  * finds a free extent and does all the dirty work required for allocation
8452  * returns the tree buffer or an ERR_PTR on error.
8453  */
8454 struct extent_buffer *btrfs_alloc_tree_block(struct btrfs_trans_handle *trans,
8455                                              struct btrfs_root *root,
8456                                              u64 parent, u64 root_objectid,
8457                                              const struct btrfs_disk_key *key,
8458                                              int level, u64 hint,
8459                                              u64 empty_size)
8460 {
8461         struct btrfs_fs_info *fs_info = root->fs_info;
8462         struct btrfs_key ins;
8463         struct btrfs_block_rsv *block_rsv;
8464         struct extent_buffer *buf;
8465         struct btrfs_delayed_extent_op *extent_op;
8466         u64 flags = 0;
8467         int ret;
8468         u32 blocksize = fs_info->nodesize;
8469         bool skinny_metadata = btrfs_fs_incompat(fs_info, SKINNY_METADATA);
8470
8471 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
8472         if (btrfs_is_testing(fs_info)) {
8473                 buf = btrfs_init_new_buffer(trans, root, root->alloc_bytenr,
8474                                             level);
8475                 if (!IS_ERR(buf))
8476                         root->alloc_bytenr += blocksize;
8477                 return buf;
8478         }
8479 #endif
8480
8481         block_rsv = use_block_rsv(trans, root, blocksize);
8482         if (IS_ERR(block_rsv))
8483                 return ERR_CAST(block_rsv);
8484
8485         ret = btrfs_reserve_extent(root, blocksize, blocksize, blocksize,
8486                                    empty_size, hint, &ins, 0, 0);
8487         if (ret)
8488                 goto out_unuse;
8489
8490         buf = btrfs_init_new_buffer(trans, root, ins.objectid, level);
8491         if (IS_ERR(buf)) {
8492                 ret = PTR_ERR(buf);
8493                 goto out_free_reserved;
8494         }
8495
8496         if (root_objectid == BTRFS_TREE_RELOC_OBJECTID) {
8497                 if (parent == 0)
8498                         parent = ins.objectid;
8499                 flags |= BTRFS_BLOCK_FLAG_FULL_BACKREF;
8500         } else
8501                 BUG_ON(parent > 0);
8502
8503         if (root_objectid != BTRFS_TREE_LOG_OBJECTID) {
8504                 extent_op = btrfs_alloc_delayed_extent_op();
8505                 if (!extent_op) {
8506                         ret = -ENOMEM;
8507                         goto out_free_buf;
8508                 }
8509                 if (key)
8510                         memcpy(&extent_op->key, key, sizeof(extent_op->key));
8511                 else
8512                         memset(&extent_op->key, 0, sizeof(extent_op->key));
8513                 extent_op->flags_to_set = flags;
8514                 extent_op->update_key = skinny_metadata ? false : true;
8515                 extent_op->update_flags = true;
8516                 extent_op->is_data = false;
8517                 extent_op->level = level;
8518
8519                 btrfs_ref_tree_mod(root, ins.objectid, ins.offset, parent,
8520                                    root_objectid, level, 0,
8521                                    BTRFS_ADD_DELAYED_EXTENT);
8522                 ret = btrfs_add_delayed_tree_ref(fs_info, trans, ins.objectid,
8523                                                  ins.offset, parent,
8524                                                  root_objectid, level,
8525                                                  BTRFS_ADD_DELAYED_EXTENT,
8526                                                  extent_op, NULL, NULL);
8527                 if (ret)
8528                         goto out_free_delayed;
8529         }
8530         return buf;
8531
8532 out_free_delayed:
8533         btrfs_free_delayed_extent_op(extent_op);
8534 out_free_buf:
8535         free_extent_buffer(buf);
8536 out_free_reserved:
8537         btrfs_free_reserved_extent(fs_info, ins.objectid, ins.offset, 0);
8538 out_unuse:
8539         unuse_block_rsv(fs_info, block_rsv, blocksize);
8540         return ERR_PTR(ret);
8541 }
8542
8543 struct walk_control {
8544         u64 refs[BTRFS_MAX_LEVEL];
8545         u64 flags[BTRFS_MAX_LEVEL];
8546         struct btrfs_key update_progress;
8547         int stage;
8548         int level;
8549         int shared_level;
8550         int update_ref;
8551         int keep_locks;
8552         int reada_slot;
8553         int reada_count;
8554         int for_reloc;
8555 };
8556
8557 #define DROP_REFERENCE  1
8558 #define UPDATE_BACKREF  2
8559
8560 static noinline void reada_walk_down(struct btrfs_trans_handle *trans,
8561                                      struct btrfs_root *root,
8562                                      struct walk_control *wc,
8563                                      struct btrfs_path *path)
8564 {
8565         struct btrfs_fs_info *fs_info = root->fs_info;
8566         u64 bytenr;
8567         u64 generation;
8568         u64 refs;
8569         u64 flags;
8570         u32 nritems;
8571         struct btrfs_key key;
8572         struct extent_buffer *eb;
8573         int ret;
8574         int slot;
8575         int nread = 0;
8576
8577         if (path->slots[wc->level] < wc->reada_slot) {
8578                 wc->reada_count = wc->reada_count * 2 / 3;
8579                 wc->reada_count = max(wc->reada_count, 2);
8580         } else {
8581                 wc->reada_count = wc->reada_count * 3 / 2;
8582                 wc->reada_count = min_t(int, wc->reada_count,
8583                                         BTRFS_NODEPTRS_PER_BLOCK(fs_info));
8584         }
8585
8586         eb = path->nodes[wc->level];
8587         nritems = btrfs_header_nritems(eb);
8588
8589         for (slot = path->slots[wc->level]; slot < nritems; slot++) {
8590                 if (nread >= wc->reada_count)
8591                         break;
8592
8593                 cond_resched();
8594                 bytenr = btrfs_node_blockptr(eb, slot);
8595                 generation = btrfs_node_ptr_generation(eb, slot);
8596
8597                 if (slot == path->slots[wc->level])
8598                         goto reada;
8599
8600                 if (wc->stage == UPDATE_BACKREF &&
8601                     generation <= root->root_key.offset)
8602                         continue;
8603
8604                 /* We don't lock the tree block, it's OK to be racy here */
8605                 ret = btrfs_lookup_extent_info(trans, fs_info, bytenr,
8606                                                wc->level - 1, 1, &refs,
8607                                                &flags);
8608                 /* We don't care about errors in readahead. */
8609                 if (ret < 0)
8610                         continue;
8611                 BUG_ON(refs == 0);
8612
8613                 if (wc->stage == DROP_REFERENCE) {
8614                         if (refs == 1)
8615                                 goto reada;
8616
8617                         if (wc->level == 1 &&
8618                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
8619                                 continue;
8620                         if (!wc->update_ref ||
8621                             generation <= root->root_key.offset)
8622                                 continue;
8623                         btrfs_node_key_to_cpu(eb, &key, slot);
8624                         ret = btrfs_comp_cpu_keys(&key,
8625                                                   &wc->update_progress);
8626                         if (ret < 0)
8627                                 continue;
8628                 } else {
8629                         if (wc->level == 1 &&
8630                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
8631                                 continue;
8632                 }
8633 reada:
8634                 readahead_tree_block(fs_info, bytenr);
8635                 nread++;
8636         }
8637         wc->reada_slot = slot;
8638 }
8639
8640 /*
8641  * helper to process tree block while walking down the tree.
8642  *
8643  * when wc->stage == UPDATE_BACKREF, this function updates
8644  * back refs for pointers in the block.
8645  *
8646  * NOTE: return value 1 means we should stop walking down.
8647  */
8648 static noinline int walk_down_proc(struct btrfs_trans_handle *trans,
8649                                    struct btrfs_root *root,
8650                                    struct btrfs_path *path,
8651                                    struct walk_control *wc, int lookup_info)
8652 {
8653         struct btrfs_fs_info *fs_info = root->fs_info;
8654         int level = wc->level;
8655         struct extent_buffer *eb = path->nodes[level];
8656         u64 flag = BTRFS_BLOCK_FLAG_FULL_BACKREF;
8657         int ret;
8658
8659         if (wc->stage == UPDATE_BACKREF &&
8660             btrfs_header_owner(eb) != root->root_key.objectid)
8661                 return 1;
8662
8663         /*
8664          * when reference count of tree block is 1, it won't increase
8665          * again. once full backref flag is set, we never clear it.
8666          */
8667         if (lookup_info &&
8668             ((wc->stage == DROP_REFERENCE && wc->refs[level] != 1) ||
8669              (wc->stage == UPDATE_BACKREF && !(wc->flags[level] & flag)))) {
8670                 BUG_ON(!path->locks[level]);
8671                 ret = btrfs_lookup_extent_info(trans, fs_info,
8672                                                eb->start, level, 1,
8673                                                &wc->refs[level],
8674                                                &wc->flags[level]);
8675                 BUG_ON(ret == -ENOMEM);
8676                 if (ret)
8677                         return ret;
8678                 BUG_ON(wc->refs[level] == 0);
8679         }
8680
8681         if (wc->stage == DROP_REFERENCE) {
8682                 if (wc->refs[level] > 1)
8683                         return 1;
8684
8685                 if (path->locks[level] && !wc->keep_locks) {
8686                         btrfs_tree_unlock_rw(eb, path->locks[level]);
8687                         path->locks[level] = 0;
8688                 }
8689                 return 0;
8690         }
8691
8692         /* wc->stage == UPDATE_BACKREF */
8693         if (!(wc->flags[level] & flag)) {
8694                 BUG_ON(!path->locks[level]);
8695                 ret = btrfs_inc_ref(trans, root, eb, 1);
8696                 BUG_ON(ret); /* -ENOMEM */
8697                 ret = btrfs_dec_ref(trans, root, eb, 0);
8698                 BUG_ON(ret); /* -ENOMEM */
8699                 ret = btrfs_set_disk_extent_flags(trans, fs_info, eb->start,
8700                                                   eb->len, flag,
8701                                                   btrfs_header_level(eb), 0);
8702                 BUG_ON(ret); /* -ENOMEM */
8703                 wc->flags[level] |= flag;
8704         }
8705
8706         /*
8707          * the block is shared by multiple trees, so it's not good to
8708          * keep the tree lock
8709          */
8710         if (path->locks[level] && level > 0) {
8711                 btrfs_tree_unlock_rw(eb, path->locks[level]);
8712                 path->locks[level] = 0;
8713         }
8714         return 0;
8715 }
8716
8717 /*
8718  * helper to process tree block pointer.
8719  *
8720  * when wc->stage == DROP_REFERENCE, this function checks
8721  * reference count of the block pointed to. if the block
8722  * is shared and we need update back refs for the subtree
8723  * rooted at the block, this function changes wc->stage to
8724  * UPDATE_BACKREF. if the block is shared and there is no
8725  * need to update back, this function drops the reference
8726  * to the block.
8727  *
8728  * NOTE: return value 1 means we should stop walking down.
8729  */
8730 static noinline int do_walk_down(struct btrfs_trans_handle *trans,
8731                                  struct btrfs_root *root,
8732                                  struct btrfs_path *path,
8733                                  struct walk_control *wc, int *lookup_info)
8734 {
8735         struct btrfs_fs_info *fs_info = root->fs_info;
8736         u64 bytenr;
8737         u64 generation;
8738         u64 parent;
8739         u32 blocksize;
8740         struct btrfs_key key;
8741         struct btrfs_key first_key;
8742         struct extent_buffer *next;
8743         int level = wc->level;
8744         int reada = 0;
8745         int ret = 0;
8746         bool need_account = false;
8747
8748         generation = btrfs_node_ptr_generation(path->nodes[level],
8749                                                path->slots[level]);
8750         /*
8751          * if the lower level block was created before the snapshot
8752          * was created, we know there is no need to update back refs
8753          * for the subtree
8754          */
8755         if (wc->stage == UPDATE_BACKREF &&
8756             generation <= root->root_key.offset) {
8757                 *lookup_info = 1;
8758                 return 1;
8759         }
8760
8761         bytenr = btrfs_node_blockptr(path->nodes[level], path->slots[level]);
8762         btrfs_node_key_to_cpu(path->nodes[level], &first_key,
8763                               path->slots[level]);
8764         blocksize = fs_info->nodesize;
8765
8766         next = find_extent_buffer(fs_info, bytenr);
8767         if (!next) {
8768                 next = btrfs_find_create_tree_block(fs_info, bytenr);
8769                 if (IS_ERR(next))
8770                         return PTR_ERR(next);
8771
8772                 btrfs_set_buffer_lockdep_class(root->root_key.objectid, next,
8773                                                level - 1);
8774                 reada = 1;
8775         }
8776         btrfs_tree_lock(next);
8777         btrfs_set_lock_blocking(next);
8778
8779         ret = btrfs_lookup_extent_info(trans, fs_info, bytenr, level - 1, 1,
8780                                        &wc->refs[level - 1],
8781                                        &wc->flags[level - 1]);
8782         if (ret < 0)
8783                 goto out_unlock;
8784
8785         if (unlikely(wc->refs[level - 1] == 0)) {
8786                 btrfs_err(fs_info, "Missing references.");
8787                 ret = -EIO;
8788                 goto out_unlock;
8789         }
8790         *lookup_info = 0;
8791
8792         if (wc->stage == DROP_REFERENCE) {
8793                 if (wc->refs[level - 1] > 1) {
8794                         need_account = true;
8795                         if (level == 1 &&
8796                             (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
8797                                 goto skip;
8798
8799                         if (!wc->update_ref ||
8800                             generation <= root->root_key.offset)
8801                                 goto skip;
8802
8803                         btrfs_node_key_to_cpu(path->nodes[level], &key,
8804                                               path->slots[level]);
8805                         ret = btrfs_comp_cpu_keys(&key, &wc->update_progress);
8806                         if (ret < 0)
8807                                 goto skip;
8808
8809                         wc->stage = UPDATE_BACKREF;
8810                         wc->shared_level = level - 1;
8811                 }
8812         } else {
8813                 if (level == 1 &&
8814                     (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
8815                         goto skip;
8816         }
8817
8818         if (!btrfs_buffer_uptodate(next, generation, 0)) {
8819                 btrfs_tree_unlock(next);
8820                 free_extent_buffer(next);
8821                 next = NULL;
8822                 *lookup_info = 1;
8823         }
8824
8825         if (!next) {
8826                 if (reada && level == 1)
8827                         reada_walk_down(trans, root, wc, path);
8828                 next = read_tree_block(fs_info, bytenr, generation, level - 1,
8829                                        &first_key);
8830                 if (IS_ERR(next)) {
8831                         return PTR_ERR(next);
8832                 } else if (!extent_buffer_uptodate(next)) {
8833                         free_extent_buffer(next);
8834                         return -EIO;
8835                 }
8836                 btrfs_tree_lock(next);
8837                 btrfs_set_lock_blocking(next);
8838         }
8839
8840         level--;
8841         ASSERT(level == btrfs_header_level(next));
8842         if (level != btrfs_header_level(next)) {
8843                 btrfs_err(root->fs_info, "mismatched level");
8844                 ret = -EIO;
8845                 goto out_unlock;
8846         }
8847         path->nodes[level] = next;
8848         path->slots[level] = 0;
8849         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
8850         wc->level = level;
8851         if (wc->level == 1)
8852                 wc->reada_slot = 0;
8853         return 0;
8854 skip:
8855         wc->refs[level - 1] = 0;
8856         wc->flags[level - 1] = 0;
8857         if (wc->stage == DROP_REFERENCE) {
8858                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF) {
8859                         parent = path->nodes[level]->start;
8860                 } else {
8861                         ASSERT(root->root_key.objectid ==
8862                                btrfs_header_owner(path->nodes[level]));
8863                         if (root->root_key.objectid !=
8864                             btrfs_header_owner(path->nodes[level])) {
8865                                 btrfs_err(root->fs_info,
8866                                                 "mismatched block owner");
8867                                 ret = -EIO;
8868                                 goto out_unlock;
8869                         }
8870                         parent = 0;
8871                 }
8872
8873                 if (need_account) {
8874                         ret = btrfs_qgroup_trace_subtree(trans, root, next,
8875                                                          generation, level - 1);
8876                         if (ret) {
8877                                 btrfs_err_rl(fs_info,
8878                                              "Error %d accounting shared subtree. Quota is out of sync, rescan required.",
8879                                              ret);
8880                         }
8881                 }
8882                 ret = btrfs_free_extent(trans, root, bytenr, blocksize,
8883                                         parent, root->root_key.objectid,
8884                                         level - 1, 0);
8885                 if (ret)
8886                         goto out_unlock;
8887         }
8888
8889         *lookup_info = 1;
8890         ret = 1;
8891
8892 out_unlock:
8893         btrfs_tree_unlock(next);
8894         free_extent_buffer(next);
8895
8896         return ret;
8897 }
8898
8899 /*
8900  * helper to process tree block while walking up the tree.
8901  *
8902  * when wc->stage == DROP_REFERENCE, this function drops
8903  * reference count on the block.
8904  *
8905  * when wc->stage == UPDATE_BACKREF, this function changes
8906  * wc->stage back to DROP_REFERENCE if we changed wc->stage
8907  * to UPDATE_BACKREF previously while processing the block.
8908  *
8909  * NOTE: return value 1 means we should stop walking up.
8910  */
8911 static noinline int walk_up_proc(struct btrfs_trans_handle *trans,
8912                                  struct btrfs_root *root,
8913                                  struct btrfs_path *path,
8914                                  struct walk_control *wc)
8915 {
8916         struct btrfs_fs_info *fs_info = root->fs_info;
8917         int ret;
8918         int level = wc->level;
8919         struct extent_buffer *eb = path->nodes[level];
8920         u64 parent = 0;
8921
8922         if (wc->stage == UPDATE_BACKREF) {
8923                 BUG_ON(wc->shared_level < level);
8924                 if (level < wc->shared_level)
8925                         goto out;
8926
8927                 ret = find_next_key(path, level + 1, &wc->update_progress);
8928                 if (ret > 0)
8929                         wc->update_ref = 0;
8930
8931                 wc->stage = DROP_REFERENCE;
8932                 wc->shared_level = -1;
8933                 path->slots[level] = 0;
8934
8935                 /*
8936                  * check reference count again if the block isn't locked.
8937                  * we should start walking down the tree again if reference
8938                  * count is one.
8939                  */
8940                 if (!path->locks[level]) {
8941                         BUG_ON(level == 0);
8942                         btrfs_tree_lock(eb);
8943                         btrfs_set_lock_blocking(eb);
8944                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
8945
8946                         ret = btrfs_lookup_extent_info(trans, fs_info,
8947                                                        eb->start, level, 1,
8948                                                        &wc->refs[level],
8949                                                        &wc->flags[level]);
8950                         if (ret < 0) {
8951                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
8952                                 path->locks[level] = 0;
8953                                 return ret;
8954                         }
8955                         BUG_ON(wc->refs[level] == 0);
8956                         if (wc->refs[level] == 1) {
8957                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
8958                                 path->locks[level] = 0;
8959                                 return 1;
8960                         }
8961                 }
8962         }
8963
8964         /* wc->stage == DROP_REFERENCE */
8965         BUG_ON(wc->refs[level] > 1 && !path->locks[level]);
8966
8967         if (wc->refs[level] == 1) {
8968                 if (level == 0) {
8969                         if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
8970                                 ret = btrfs_dec_ref(trans, root, eb, 1);
8971                         else
8972                                 ret = btrfs_dec_ref(trans, root, eb, 0);
8973                         BUG_ON(ret); /* -ENOMEM */
8974                         ret = btrfs_qgroup_trace_leaf_items(trans, fs_info, eb);
8975                         if (ret) {
8976                                 btrfs_err_rl(fs_info,
8977                                              "error %d accounting leaf items. Quota is out of sync, rescan required.",
8978                                              ret);
8979                         }
8980                 }
8981                 /* make block locked assertion in clean_tree_block happy */
8982                 if (!path->locks[level] &&
8983                     btrfs_header_generation(eb) == trans->transid) {
8984                         btrfs_tree_lock(eb);
8985                         btrfs_set_lock_blocking(eb);
8986                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
8987                 }
8988                 clean_tree_block(fs_info, eb);
8989         }
8990
8991         if (eb == root->node) {
8992                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
8993                         parent = eb->start;
8994                 else
8995                         BUG_ON(root->root_key.objectid !=
8996                                btrfs_header_owner(eb));
8997         } else {
8998                 if (wc->flags[level + 1] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
8999                         parent = path->nodes[level + 1]->start;
9000                 else
9001                         BUG_ON(root->root_key.objectid !=
9002                                btrfs_header_owner(path->nodes[level + 1]));
9003         }
9004
9005         btrfs_free_tree_block(trans, root, eb, parent, wc->refs[level] == 1);
9006 out:
9007         wc->refs[level] = 0;
9008         wc->flags[level] = 0;
9009         return 0;
9010 }
9011
9012 static noinline int walk_down_tree(struct btrfs_trans_handle *trans,
9013                                    struct btrfs_root *root,
9014                                    struct btrfs_path *path,
9015                                    struct walk_control *wc)
9016 {
9017         int level = wc->level;
9018         int lookup_info = 1;
9019         int ret;
9020
9021         while (level >= 0) {
9022                 ret = walk_down_proc(trans, root, path, wc, lookup_info);
9023                 if (ret > 0)
9024                         break;
9025
9026                 if (level == 0)
9027                         break;
9028
9029                 if (path->slots[level] >=
9030                     btrfs_header_nritems(path->nodes[level]))
9031                         break;
9032
9033                 ret = do_walk_down(trans, root, path, wc, &lookup_info);
9034                 if (ret > 0) {
9035                         path->slots[level]++;
9036                         continue;
9037                 } else if (ret < 0)
9038                         return ret;
9039                 level = wc->level;
9040         }
9041         return 0;
9042 }
9043
9044 static noinline int walk_up_tree(struct btrfs_trans_handle *trans,
9045                                  struct btrfs_root *root,
9046                                  struct btrfs_path *path,
9047                                  struct walk_control *wc, int max_level)
9048 {
9049         int level = wc->level;
9050         int ret;
9051
9052         path->slots[level] = btrfs_header_nritems(path->nodes[level]);
9053         while (level < max_level && path->nodes[level]) {
9054                 wc->level = level;
9055                 if (path->slots[level] + 1 <
9056                     btrfs_header_nritems(path->nodes[level])) {
9057                         path->slots[level]++;
9058                         return 0;
9059                 } else {
9060                         ret = walk_up_proc(trans, root, path, wc);
9061                         if (ret > 0)
9062                                 return 0;
9063
9064                         if (path->locks[level]) {
9065                                 btrfs_tree_unlock_rw(path->nodes[level],
9066                                                      path->locks[level]);
9067                                 path->locks[level] = 0;
9068                         }
9069                         free_extent_buffer(path->nodes[level]);
9070                         path->nodes[level] = NULL;
9071                         level++;
9072                 }
9073         }
9074         return 1;
9075 }
9076
9077 /*
9078  * drop a subvolume tree.
9079  *
9080  * this function traverses the tree freeing any blocks that only
9081  * referenced by the tree.
9082  *
9083  * when a shared tree block is found. this function decreases its
9084  * reference count by one. if update_ref is true, this function
9085  * also make sure backrefs for the shared block and all lower level
9086  * blocks are properly updated.
9087  *
9088  * If called with for_reloc == 0, may exit early with -EAGAIN
9089  */
9090 int btrfs_drop_snapshot(struct btrfs_root *root,
9091                          struct btrfs_block_rsv *block_rsv, int update_ref,
9092                          int for_reloc)
9093 {
9094         struct btrfs_fs_info *fs_info = root->fs_info;
9095         struct btrfs_path *path;
9096         struct btrfs_trans_handle *trans;
9097         struct btrfs_root *tree_root = fs_info->tree_root;
9098         struct btrfs_root_item *root_item = &root->root_item;
9099         struct walk_control *wc;
9100         struct btrfs_key key;
9101         int err = 0;
9102         int ret;
9103         int level;
9104         bool root_dropped = false;
9105
9106         btrfs_debug(fs_info, "Drop subvolume %llu", root->objectid);
9107
9108         path = btrfs_alloc_path();
9109         if (!path) {
9110                 err = -ENOMEM;
9111                 goto out;
9112         }
9113
9114         wc = kzalloc(sizeof(*wc), GFP_NOFS);
9115         if (!wc) {
9116                 btrfs_free_path(path);
9117                 err = -ENOMEM;
9118                 goto out;
9119         }
9120
9121         trans = btrfs_start_transaction(tree_root, 0);
9122         if (IS_ERR(trans)) {
9123                 err = PTR_ERR(trans);
9124                 goto out_free;
9125         }
9126
9127         if (block_rsv)
9128                 trans->block_rsv = block_rsv;
9129
9130         if (btrfs_disk_key_objectid(&root_item->drop_progress) == 0) {
9131                 level = btrfs_header_level(root->node);
9132                 path->nodes[level] = btrfs_lock_root_node(root);
9133                 btrfs_set_lock_blocking(path->nodes[level]);
9134                 path->slots[level] = 0;
9135                 path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
9136                 memset(&wc->update_progress, 0,
9137                        sizeof(wc->update_progress));
9138         } else {
9139                 btrfs_disk_key_to_cpu(&key, &root_item->drop_progress);
9140                 memcpy(&wc->update_progress, &key,
9141                        sizeof(wc->update_progress));
9142
9143                 level = root_item->drop_level;
9144                 BUG_ON(level == 0);
9145                 path->lowest_level = level;
9146                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
9147                 path->lowest_level = 0;
9148                 if (ret < 0) {
9149                         err = ret;
9150                         goto out_end_trans;
9151                 }
9152                 WARN_ON(ret > 0);
9153
9154                 /*
9155                  * unlock our path, this is safe because only this
9156                  * function is allowed to delete this snapshot
9157                  */
9158                 btrfs_unlock_up_safe(path, 0);
9159
9160                 level = btrfs_header_level(root->node);
9161                 while (1) {
9162                         btrfs_tree_lock(path->nodes[level]);
9163                         btrfs_set_lock_blocking(path->nodes[level]);
9164                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
9165
9166                         ret = btrfs_lookup_extent_info(trans, fs_info,
9167                                                 path->nodes[level]->start,
9168                                                 level, 1, &wc->refs[level],
9169                                                 &wc->flags[level]);
9170                         if (ret < 0) {
9171                                 err = ret;
9172                                 goto out_end_trans;
9173                         }
9174                         BUG_ON(wc->refs[level] == 0);
9175
9176                         if (level == root_item->drop_level)
9177                                 break;
9178
9179                         btrfs_tree_unlock(path->nodes[level]);
9180                         path->locks[level] = 0;
9181                         WARN_ON(wc->refs[level] != 1);
9182                         level--;
9183                 }
9184         }
9185
9186         wc->level = level;
9187         wc->shared_level = -1;
9188         wc->stage = DROP_REFERENCE;
9189         wc->update_ref = update_ref;
9190         wc->keep_locks = 0;
9191         wc->for_reloc = for_reloc;
9192         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(fs_info);
9193
9194         while (1) {
9195
9196                 ret = walk_down_tree(trans, root, path, wc);
9197                 if (ret < 0) {
9198                         err = ret;
9199                         break;
9200                 }
9201
9202                 ret = walk_up_tree(trans, root, path, wc, BTRFS_MAX_LEVEL);
9203                 if (ret < 0) {
9204                         err = ret;
9205                         break;
9206                 }
9207
9208                 if (ret > 0) {
9209                         BUG_ON(wc->stage != DROP_REFERENCE);
9210                         break;
9211                 }
9212
9213                 if (wc->stage == DROP_REFERENCE) {
9214                         level = wc->level;
9215                         btrfs_node_key(path->nodes[level],
9216                                        &root_item->drop_progress,
9217                                        path->slots[level]);
9218                         root_item->drop_level = level;
9219                 }
9220
9221                 BUG_ON(wc->level == 0);
9222                 if (btrfs_should_end_transaction(trans) ||
9223                     (!for_reloc && btrfs_need_cleaner_sleep(fs_info))) {
9224                         ret = btrfs_update_root(trans, tree_root,
9225                                                 &root->root_key,
9226                                                 root_item);
9227                         if (ret) {
9228                                 btrfs_abort_transaction(trans, ret);
9229                                 err = ret;
9230                                 goto out_end_trans;
9231                         }
9232
9233                         btrfs_end_transaction_throttle(trans);
9234                         if (!for_reloc && btrfs_need_cleaner_sleep(fs_info)) {
9235                                 btrfs_debug(fs_info,
9236                                             "drop snapshot early exit");
9237                                 err = -EAGAIN;
9238                                 goto out_free;
9239                         }
9240
9241                         trans = btrfs_start_transaction(tree_root, 0);
9242                         if (IS_ERR(trans)) {
9243                                 err = PTR_ERR(trans);
9244                                 goto out_free;
9245                         }
9246                         if (block_rsv)
9247                                 trans->block_rsv = block_rsv;
9248                 }
9249         }
9250         btrfs_release_path(path);
9251         if (err)
9252                 goto out_end_trans;
9253
9254         ret = btrfs_del_root(trans, fs_info, &root->root_key);
9255         if (ret) {
9256                 btrfs_abort_transaction(trans, ret);
9257                 err = ret;
9258                 goto out_end_trans;
9259         }
9260
9261         if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) {
9262                 ret = btrfs_find_root(tree_root, &root->root_key, path,
9263                                       NULL, NULL);
9264                 if (ret < 0) {
9265                         btrfs_abort_transaction(trans, ret);
9266                         err = ret;
9267                         goto out_end_trans;
9268                 } else if (ret > 0) {
9269                         /* if we fail to delete the orphan item this time
9270                          * around, it'll get picked up the next time.
9271                          *
9272                          * The most common failure here is just -ENOENT.
9273                          */
9274                         btrfs_del_orphan_item(trans, tree_root,
9275                                               root->root_key.objectid);
9276                 }
9277         }
9278
9279         if (test_bit(BTRFS_ROOT_IN_RADIX, &root->state)) {
9280                 btrfs_add_dropped_root(trans, root);
9281         } else {
9282                 free_extent_buffer(root->node);
9283                 free_extent_buffer(root->commit_root);
9284                 btrfs_put_fs_root(root);
9285         }
9286         root_dropped = true;
9287 out_end_trans:
9288         btrfs_end_transaction_throttle(trans);
9289 out_free:
9290         kfree(wc);
9291         btrfs_free_path(path);
9292 out:
9293         /*
9294          * So if we need to stop dropping the snapshot for whatever reason we
9295          * need to make sure to add it back to the dead root list so that we
9296          * keep trying to do the work later.  This also cleans up roots if we
9297          * don't have it in the radix (like when we recover after a power fail
9298          * or unmount) so we don't leak memory.
9299          */
9300         if (!for_reloc && !root_dropped)
9301                 btrfs_add_dead_root(root);
9302         if (err && err != -EAGAIN)
9303                 btrfs_handle_fs_error(fs_info, err, NULL);
9304         return err;
9305 }
9306
9307 /*
9308  * drop subtree rooted at tree block 'node'.
9309  *
9310  * NOTE: this function will unlock and release tree block 'node'
9311  * only used by relocation code
9312  */
9313 int btrfs_drop_subtree(struct btrfs_trans_handle *trans,
9314                         struct btrfs_root *root,
9315                         struct extent_buffer *node,
9316                         struct extent_buffer *parent)
9317 {
9318         struct btrfs_fs_info *fs_info = root->fs_info;
9319         struct btrfs_path *path;
9320         struct walk_control *wc;
9321         int level;
9322         int parent_level;
9323         int ret = 0;
9324         int wret;
9325
9326         BUG_ON(root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID);
9327
9328         path = btrfs_alloc_path();
9329         if (!path)
9330                 return -ENOMEM;
9331
9332         wc = kzalloc(sizeof(*wc), GFP_NOFS);
9333         if (!wc) {
9334                 btrfs_free_path(path);
9335                 return -ENOMEM;
9336         }
9337
9338         btrfs_assert_tree_locked(parent);
9339         parent_level = btrfs_header_level(parent);
9340         extent_buffer_get(parent);
9341         path->nodes[parent_level] = parent;
9342         path->slots[parent_level] = btrfs_header_nritems(parent);
9343
9344         btrfs_assert_tree_locked(node);
9345         level = btrfs_header_level(node);
9346         path->nodes[level] = node;
9347         path->slots[level] = 0;
9348         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
9349
9350         wc->refs[parent_level] = 1;
9351         wc->flags[parent_level] = BTRFS_BLOCK_FLAG_FULL_BACKREF;
9352         wc->level = level;
9353         wc->shared_level = -1;
9354         wc->stage = DROP_REFERENCE;
9355         wc->update_ref = 0;
9356         wc->keep_locks = 1;
9357         wc->for_reloc = 1;
9358         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(fs_info);
9359
9360         while (1) {
9361                 wret = walk_down_tree(trans, root, path, wc);
9362                 if (wret < 0) {
9363                         ret = wret;
9364                         break;
9365                 }
9366
9367                 wret = walk_up_tree(trans, root, path, wc, parent_level);
9368                 if (wret < 0)
9369                         ret = wret;
9370                 if (wret != 0)
9371                         break;
9372         }
9373
9374         kfree(wc);
9375         btrfs_free_path(path);
9376         return ret;
9377 }
9378
9379 static u64 update_block_group_flags(struct btrfs_fs_info *fs_info, u64 flags)
9380 {
9381         u64 num_devices;
9382         u64 stripped;
9383
9384         /*
9385          * if restripe for this chunk_type is on pick target profile and
9386          * return, otherwise do the usual balance
9387          */
9388         stripped = get_restripe_target(fs_info, flags);
9389         if (stripped)
9390                 return extended_to_chunk(stripped);
9391
9392         num_devices = fs_info->fs_devices->rw_devices;
9393
9394         stripped = BTRFS_BLOCK_GROUP_RAID0 |
9395                 BTRFS_BLOCK_GROUP_RAID5 | BTRFS_BLOCK_GROUP_RAID6 |
9396                 BTRFS_BLOCK_GROUP_RAID1 | BTRFS_BLOCK_GROUP_RAID10;
9397
9398         if (num_devices == 1) {
9399                 stripped |= BTRFS_BLOCK_GROUP_DUP;
9400                 stripped = flags & ~stripped;
9401
9402                 /* turn raid0 into single device chunks */
9403                 if (flags & BTRFS_BLOCK_GROUP_RAID0)
9404                         return stripped;
9405
9406                 /* turn mirroring into duplication */
9407                 if (flags & (BTRFS_BLOCK_GROUP_RAID1 |
9408                              BTRFS_BLOCK_GROUP_RAID10))
9409                         return stripped | BTRFS_BLOCK_GROUP_DUP;
9410         } else {
9411                 /* they already had raid on here, just return */
9412                 if (flags & stripped)
9413                         return flags;
9414
9415                 stripped |= BTRFS_BLOCK_GROUP_DUP;
9416                 stripped = flags & ~stripped;
9417
9418                 /* switch duplicated blocks with raid1 */
9419                 if (flags & BTRFS_BLOCK_GROUP_DUP)
9420                         return stripped | BTRFS_BLOCK_GROUP_RAID1;
9421
9422                 /* this is drive concat, leave it alone */
9423         }
9424
9425         return flags;
9426 }
9427
9428 static int inc_block_group_ro(struct btrfs_block_group_cache *cache, int force)
9429 {
9430         struct btrfs_space_info *sinfo = cache->space_info;
9431         u64 num_bytes;
9432         u64 min_allocable_bytes;
9433         int ret = -ENOSPC;
9434
9435         /*
9436          * We need some metadata space and system metadata space for
9437          * allocating chunks in some corner cases until we force to set
9438          * it to be readonly.
9439          */
9440         if ((sinfo->flags &
9441              (BTRFS_BLOCK_GROUP_SYSTEM | BTRFS_BLOCK_GROUP_METADATA)) &&
9442             !force)
9443                 min_allocable_bytes = SZ_1M;
9444         else
9445                 min_allocable_bytes = 0;
9446
9447         spin_lock(&sinfo->lock);
9448         spin_lock(&cache->lock);
9449
9450         if (cache->ro) {
9451                 cache->ro++;
9452                 ret = 0;
9453                 goto out;
9454         }
9455
9456         num_bytes = cache->key.offset - cache->reserved - cache->pinned -
9457                     cache->bytes_super - btrfs_block_group_used(&cache->item);
9458
9459         if (btrfs_space_info_used(sinfo, true) + num_bytes +
9460             min_allocable_bytes <= sinfo->total_bytes) {
9461                 sinfo->bytes_readonly += num_bytes;
9462                 cache->ro++;
9463                 list_add_tail(&cache->ro_list, &sinfo->ro_bgs);
9464                 ret = 0;
9465         }
9466 out:
9467         spin_unlock(&cache->lock);
9468         spin_unlock(&sinfo->lock);
9469         return ret;
9470 }
9471
9472 int btrfs_inc_block_group_ro(struct btrfs_fs_info *fs_info,
9473                              struct btrfs_block_group_cache *cache)
9474
9475 {
9476         struct btrfs_trans_handle *trans;
9477         u64 alloc_flags;
9478         int ret;
9479
9480 again:
9481         trans = btrfs_join_transaction(fs_info->extent_root);
9482         if (IS_ERR(trans))
9483                 return PTR_ERR(trans);
9484
9485         /*
9486          * we're not allowed to set block groups readonly after the dirty
9487          * block groups cache has started writing.  If it already started,
9488          * back off and let this transaction commit
9489          */
9490         mutex_lock(&fs_info->ro_block_group_mutex);
9491         if (test_bit(BTRFS_TRANS_DIRTY_BG_RUN, &trans->transaction->flags)) {
9492                 u64 transid = trans->transid;
9493
9494                 mutex_unlock(&fs_info->ro_block_group_mutex);
9495                 btrfs_end_transaction(trans);
9496
9497                 ret = btrfs_wait_for_commit(fs_info, transid);
9498                 if (ret)
9499                         return ret;
9500                 goto again;
9501         }
9502
9503         /*
9504          * if we are changing raid levels, try to allocate a corresponding
9505          * block group with the new raid level.
9506          */
9507         alloc_flags = update_block_group_flags(fs_info, cache->flags);
9508         if (alloc_flags != cache->flags) {
9509                 ret = do_chunk_alloc(trans, fs_info, alloc_flags,
9510                                      CHUNK_ALLOC_FORCE);
9511                 /*
9512                  * ENOSPC is allowed here, we may have enough space
9513                  * already allocated at the new raid level to
9514                  * carry on
9515                  */
9516                 if (ret == -ENOSPC)
9517                         ret = 0;
9518                 if (ret < 0)
9519                         goto out;
9520         }
9521
9522         ret = inc_block_group_ro(cache, 0);
9523         if (!ret)
9524                 goto out;
9525         alloc_flags = get_alloc_profile(fs_info, cache->space_info->flags);
9526         ret = do_chunk_alloc(trans, fs_info, alloc_flags,
9527                              CHUNK_ALLOC_FORCE);
9528         if (ret < 0)
9529                 goto out;
9530         ret = inc_block_group_ro(cache, 0);
9531 out:
9532         if (cache->flags & BTRFS_BLOCK_GROUP_SYSTEM) {
9533                 alloc_flags = update_block_group_flags(fs_info, cache->flags);
9534                 mutex_lock(&fs_info->chunk_mutex);
9535                 check_system_chunk(trans, fs_info, alloc_flags);
9536                 mutex_unlock(&fs_info->chunk_mutex);
9537         }
9538         mutex_unlock(&fs_info->ro_block_group_mutex);
9539
9540         btrfs_end_transaction(trans);
9541         return ret;
9542 }
9543
9544 int btrfs_force_chunk_alloc(struct btrfs_trans_handle *trans,
9545                             struct btrfs_fs_info *fs_info, u64 type)
9546 {
9547         u64 alloc_flags = get_alloc_profile(fs_info, type);
9548
9549         return do_chunk_alloc(trans, fs_info, alloc_flags, CHUNK_ALLOC_FORCE);
9550 }
9551
9552 /*
9553  * helper to account the unused space of all the readonly block group in the
9554  * space_info. takes mirrors into account.
9555  */
9556 u64 btrfs_account_ro_block_groups_free_space(struct btrfs_space_info *sinfo)
9557 {
9558         struct btrfs_block_group_cache *block_group;
9559         u64 free_bytes = 0;
9560         int factor;
9561
9562         /* It's df, we don't care if it's racy */
9563         if (list_empty(&sinfo->ro_bgs))
9564                 return 0;
9565
9566         spin_lock(&sinfo->lock);
9567         list_for_each_entry(block_group, &sinfo->ro_bgs, ro_list) {
9568                 spin_lock(&block_group->lock);
9569
9570                 if (!block_group->ro) {
9571                         spin_unlock(&block_group->lock);
9572                         continue;
9573                 }
9574
9575                 if (block_group->flags & (BTRFS_BLOCK_GROUP_RAID1 |
9576                                           BTRFS_BLOCK_GROUP_RAID10 |
9577                                           BTRFS_BLOCK_GROUP_DUP))
9578                         factor = 2;
9579                 else
9580                         factor = 1;
9581
9582                 free_bytes += (block_group->key.offset -
9583                                btrfs_block_group_used(&block_group->item)) *
9584                                factor;
9585
9586                 spin_unlock(&block_group->lock);
9587         }
9588         spin_unlock(&sinfo->lock);
9589
9590         return free_bytes;
9591 }
9592
9593 void btrfs_dec_block_group_ro(struct btrfs_block_group_cache *cache)
9594 {
9595         struct btrfs_space_info *sinfo = cache->space_info;
9596         u64 num_bytes;
9597
9598         BUG_ON(!cache->ro);
9599
9600         spin_lock(&sinfo->lock);
9601         spin_lock(&cache->lock);
9602         if (!--cache->ro) {
9603                 num_bytes = cache->key.offset - cache->reserved -
9604                             cache->pinned - cache->bytes_super -
9605                             btrfs_block_group_used(&cache->item);
9606                 sinfo->bytes_readonly -= num_bytes;
9607                 list_del_init(&cache->ro_list);
9608         }
9609         spin_unlock(&cache->lock);
9610         spin_unlock(&sinfo->lock);
9611 }
9612
9613 /*
9614  * checks to see if its even possible to relocate this block group.
9615  *
9616  * @return - -1 if it's not a good idea to relocate this block group, 0 if its
9617  * ok to go ahead and try.
9618  */
9619 int btrfs_can_relocate(struct btrfs_fs_info *fs_info, u64 bytenr)
9620 {
9621         struct btrfs_root *root = fs_info->extent_root;
9622         struct btrfs_block_group_cache *block_group;
9623         struct btrfs_space_info *space_info;
9624         struct btrfs_fs_devices *fs_devices = fs_info->fs_devices;
9625         struct btrfs_device *device;
9626         struct btrfs_trans_handle *trans;
9627         u64 min_free;
9628         u64 dev_min = 1;
9629         u64 dev_nr = 0;
9630         u64 target;
9631         int debug;
9632         int index;
9633         int full = 0;
9634         int ret = 0;
9635
9636         debug = btrfs_test_opt(fs_info, ENOSPC_DEBUG);
9637
9638         block_group = btrfs_lookup_block_group(fs_info, bytenr);
9639
9640         /* odd, couldn't find the block group, leave it alone */
9641         if (!block_group) {
9642                 if (debug)
9643                         btrfs_warn(fs_info,
9644                                    "can't find block group for bytenr %llu",
9645                                    bytenr);
9646                 return -1;
9647         }
9648
9649         min_free = btrfs_block_group_used(&block_group->item);
9650
9651         /* no bytes used, we're good */
9652         if (!min_free)
9653                 goto out;
9654
9655         space_info = block_group->space_info;
9656         spin_lock(&space_info->lock);
9657
9658         full = space_info->full;
9659
9660         /*
9661          * if this is the last block group we have in this space, we can't
9662          * relocate it unless we're able to allocate a new chunk below.
9663          *
9664          * Otherwise, we need to make sure we have room in the space to handle
9665          * all of the extents from this block group.  If we can, we're good
9666          */
9667         if ((space_info->total_bytes != block_group->key.offset) &&
9668             (btrfs_space_info_used(space_info, false) + min_free <
9669              space_info->total_bytes)) {
9670                 spin_unlock(&space_info->lock);
9671                 goto out;
9672         }
9673         spin_unlock(&space_info->lock);
9674
9675         /*
9676          * ok we don't have enough space, but maybe we have free space on our
9677          * devices to allocate new chunks for relocation, so loop through our
9678          * alloc devices and guess if we have enough space.  if this block
9679          * group is going to be restriped, run checks against the target
9680          * profile instead of the current one.
9681          */
9682         ret = -1;
9683
9684         /*
9685          * index:
9686          *      0: raid10
9687          *      1: raid1
9688          *      2: dup
9689          *      3: raid0
9690          *      4: single
9691          */
9692         target = get_restripe_target(fs_info, block_group->flags);
9693         if (target) {
9694                 index = btrfs_bg_flags_to_raid_index(extended_to_chunk(target));
9695         } else {
9696                 /*
9697                  * this is just a balance, so if we were marked as full
9698                  * we know there is no space for a new chunk
9699                  */
9700                 if (full) {
9701                         if (debug)
9702                                 btrfs_warn(fs_info,
9703                                            "no space to alloc new chunk for block group %llu",
9704                                            block_group->key.objectid);
9705                         goto out;
9706                 }
9707
9708                 index = btrfs_bg_flags_to_raid_index(block_group->flags);
9709         }
9710
9711         if (index == BTRFS_RAID_RAID10) {
9712                 dev_min = 4;
9713                 /* Divide by 2 */
9714                 min_free >>= 1;
9715         } else if (index == BTRFS_RAID_RAID1) {
9716                 dev_min = 2;
9717         } else if (index == BTRFS_RAID_DUP) {
9718                 /* Multiply by 2 */
9719                 min_free <<= 1;
9720         } else if (index == BTRFS_RAID_RAID0) {
9721                 dev_min = fs_devices->rw_devices;
9722                 min_free = div64_u64(min_free, dev_min);
9723         }
9724
9725         /* We need to do this so that we can look at pending chunks */
9726         trans = btrfs_join_transaction(root);
9727         if (IS_ERR(trans)) {
9728                 ret = PTR_ERR(trans);
9729                 goto out;
9730         }
9731
9732         mutex_lock(&fs_info->chunk_mutex);
9733         list_for_each_entry(device, &fs_devices->alloc_list, dev_alloc_list) {
9734                 u64 dev_offset;
9735
9736                 /*
9737                  * check to make sure we can actually find a chunk with enough
9738                  * space to fit our block group in.
9739                  */
9740                 if (device->total_bytes > device->bytes_used + min_free &&
9741                     !test_bit(BTRFS_DEV_STATE_REPLACE_TGT, &device->dev_state)) {
9742                         ret = find_free_dev_extent(trans, device, min_free,
9743                                                    &dev_offset, NULL);
9744                         if (!ret)
9745                                 dev_nr++;
9746
9747                         if (dev_nr >= dev_min)
9748                                 break;
9749
9750                         ret = -1;
9751                 }
9752         }
9753         if (debug && ret == -1)
9754                 btrfs_warn(fs_info,
9755                            "no space to allocate a new chunk for block group %llu",
9756                            block_group->key.objectid);
9757         mutex_unlock(&fs_info->chunk_mutex);
9758         btrfs_end_transaction(trans);
9759 out:
9760         btrfs_put_block_group(block_group);
9761         return ret;
9762 }
9763
9764 static int find_first_block_group(struct btrfs_fs_info *fs_info,
9765                                   struct btrfs_path *path,
9766                                   struct btrfs_key *key)
9767 {
9768         struct btrfs_root *root = fs_info->extent_root;
9769         int ret = 0;
9770         struct btrfs_key found_key;
9771         struct extent_buffer *leaf;
9772         int slot;
9773
9774         ret = btrfs_search_slot(NULL, root, key, path, 0, 0);
9775         if (ret < 0)
9776                 goto out;
9777
9778         while (1) {
9779                 slot = path->slots[0];
9780                 leaf = path->nodes[0];
9781                 if (slot >= btrfs_header_nritems(leaf)) {
9782                         ret = btrfs_next_leaf(root, path);
9783                         if (ret == 0)
9784                                 continue;
9785                         if (ret < 0)
9786                                 goto out;
9787                         break;
9788                 }
9789                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
9790
9791                 if (found_key.objectid >= key->objectid &&
9792                     found_key.type == BTRFS_BLOCK_GROUP_ITEM_KEY) {
9793                         struct extent_map_tree *em_tree;
9794                         struct extent_map *em;
9795
9796                         em_tree = &root->fs_info->mapping_tree.map_tree;
9797                         read_lock(&em_tree->lock);
9798                         em = lookup_extent_mapping(em_tree, found_key.objectid,
9799                                                    found_key.offset);
9800                         read_unlock(&em_tree->lock);
9801                         if (!em) {
9802                                 btrfs_err(fs_info,
9803                         "logical %llu len %llu found bg but no related chunk",
9804                                           found_key.objectid, found_key.offset);
9805                                 ret = -ENOENT;
9806                         } else {
9807                                 ret = 0;
9808                         }
9809                         free_extent_map(em);
9810                         goto out;
9811                 }
9812                 path->slots[0]++;
9813         }
9814 out:
9815         return ret;
9816 }
9817
9818 void btrfs_put_block_group_cache(struct btrfs_fs_info *info)
9819 {
9820         struct btrfs_block_group_cache *block_group;
9821         u64 last = 0;
9822
9823         while (1) {
9824                 struct inode *inode;
9825
9826                 block_group = btrfs_lookup_first_block_group(info, last);
9827                 while (block_group) {
9828                         spin_lock(&block_group->lock);
9829                         if (block_group->iref)
9830                                 break;
9831                         spin_unlock(&block_group->lock);
9832                         block_group = next_block_group(info, block_group);
9833                 }
9834                 if (!block_group) {
9835                         if (last == 0)
9836                                 break;
9837                         last = 0;
9838                         continue;
9839                 }
9840
9841                 inode = block_group->inode;
9842                 block_group->iref = 0;
9843                 block_group->inode = NULL;
9844                 spin_unlock(&block_group->lock);
9845                 ASSERT(block_group->io_ctl.inode == NULL);
9846                 iput(inode);
9847                 last = block_group->key.objectid + block_group->key.offset;
9848                 btrfs_put_block_group(block_group);
9849         }
9850 }
9851
9852 /*
9853  * Must be called only after stopping all workers, since we could have block
9854  * group caching kthreads running, and therefore they could race with us if we
9855  * freed the block groups before stopping them.
9856  */
9857 int btrfs_free_block_groups(struct btrfs_fs_info *info)
9858 {
9859         struct btrfs_block_group_cache *block_group;
9860         struct btrfs_space_info *space_info;
9861         struct btrfs_caching_control *caching_ctl;
9862         struct rb_node *n;
9863
9864         down_write(&info->commit_root_sem);
9865         while (!list_empty(&info->caching_block_groups)) {
9866                 caching_ctl = list_entry(info->caching_block_groups.next,
9867                                          struct btrfs_caching_control, list);
9868                 list_del(&caching_ctl->list);
9869                 put_caching_control(caching_ctl);
9870         }
9871         up_write(&info->commit_root_sem);
9872
9873         spin_lock(&info->unused_bgs_lock);
9874         while (!list_empty(&info->unused_bgs)) {
9875                 block_group = list_first_entry(&info->unused_bgs,
9876                                                struct btrfs_block_group_cache,
9877                                                bg_list);
9878                 list_del_init(&block_group->bg_list);
9879                 btrfs_put_block_group(block_group);
9880         }
9881         spin_unlock(&info->unused_bgs_lock);
9882
9883         spin_lock(&info->block_group_cache_lock);
9884         while ((n = rb_last(&info->block_group_cache_tree)) != NULL) {
9885                 block_group = rb_entry(n, struct btrfs_block_group_cache,
9886                                        cache_node);
9887                 rb_erase(&block_group->cache_node,
9888                          &info->block_group_cache_tree);
9889                 RB_CLEAR_NODE(&block_group->cache_node);
9890                 spin_unlock(&info->block_group_cache_lock);
9891
9892                 down_write(&block_group->space_info->groups_sem);
9893                 list_del(&block_group->list);
9894                 up_write(&block_group->space_info->groups_sem);
9895
9896                 /*
9897                  * We haven't cached this block group, which means we could
9898                  * possibly have excluded extents on this block group.
9899                  */
9900                 if (block_group->cached == BTRFS_CACHE_NO ||
9901                     block_group->cached == BTRFS_CACHE_ERROR)
9902                         free_excluded_extents(info, block_group);
9903
9904                 btrfs_remove_free_space_cache(block_group);
9905                 ASSERT(block_group->cached != BTRFS_CACHE_STARTED);
9906                 ASSERT(list_empty(&block_group->dirty_list));
9907                 ASSERT(list_empty(&block_group->io_list));
9908                 ASSERT(list_empty(&block_group->bg_list));
9909                 ASSERT(atomic_read(&block_group->count) == 1);
9910                 btrfs_put_block_group(block_group);
9911
9912                 spin_lock(&info->block_group_cache_lock);
9913         }
9914         spin_unlock(&info->block_group_cache_lock);
9915
9916         /* now that all the block groups are freed, go through and
9917          * free all the space_info structs.  This is only called during
9918          * the final stages of unmount, and so we know nobody is
9919          * using them.  We call synchronize_rcu() once before we start,
9920          * just to be on the safe side.
9921          */
9922         synchronize_rcu();
9923
9924         release_global_block_rsv(info);
9925
9926         while (!list_empty(&info->space_info)) {
9927                 int i;
9928
9929                 space_info = list_entry(info->space_info.next,
9930                                         struct btrfs_space_info,
9931                                         list);
9932
9933                 /*
9934                  * Do not hide this behind enospc_debug, this is actually
9935                  * important and indicates a real bug if this happens.
9936                  */
9937                 if (WARN_ON(space_info->bytes_pinned > 0 ||
9938                             space_info->bytes_reserved > 0 ||
9939                             space_info->bytes_may_use > 0))
9940                         dump_space_info(info, space_info, 0, 0);
9941                 list_del(&space_info->list);
9942                 for (i = 0; i < BTRFS_NR_RAID_TYPES; i++) {
9943                         struct kobject *kobj;
9944                         kobj = space_info->block_group_kobjs[i];
9945                         space_info->block_group_kobjs[i] = NULL;
9946                         if (kobj) {
9947                                 kobject_del(kobj);
9948                                 kobject_put(kobj);
9949                         }
9950                 }
9951                 kobject_del(&space_info->kobj);
9952                 kobject_put(&space_info->kobj);
9953         }
9954         return 0;
9955 }
9956
9957 /* link_block_group will queue up kobjects to add when we're reclaim-safe */
9958 void btrfs_add_raid_kobjects(struct btrfs_fs_info *fs_info)
9959 {
9960         struct btrfs_space_info *space_info;
9961         struct raid_kobject *rkobj;
9962         LIST_HEAD(list);
9963         int index;
9964         int ret = 0;
9965
9966         spin_lock(&fs_info->pending_raid_kobjs_lock);
9967         list_splice_init(&fs_info->pending_raid_kobjs, &list);
9968         spin_unlock(&fs_info->pending_raid_kobjs_lock);
9969
9970         list_for_each_entry(rkobj, &list, list) {
9971                 space_info = __find_space_info(fs_info, rkobj->flags);
9972                 index = btrfs_bg_flags_to_raid_index(rkobj->flags);
9973
9974                 ret = kobject_add(&rkobj->kobj, &space_info->kobj,
9975                                   "%s", get_raid_name(index));
9976                 if (ret) {
9977                         kobject_put(&rkobj->kobj);
9978                         break;
9979                 }
9980         }
9981         if (ret)
9982                 btrfs_warn(fs_info,
9983                            "failed to add kobject for block cache, ignoring");
9984 }
9985
9986 static void link_block_group(struct btrfs_block_group_cache *cache)
9987 {
9988         struct btrfs_space_info *space_info = cache->space_info;
9989         struct btrfs_fs_info *fs_info = cache->fs_info;
9990         int index = btrfs_bg_flags_to_raid_index(cache->flags);
9991         bool first = false;
9992
9993         down_write(&space_info->groups_sem);
9994         if (list_empty(&space_info->block_groups[index]))
9995                 first = true;
9996         list_add_tail(&cache->list, &space_info->block_groups[index]);
9997         up_write(&space_info->groups_sem);
9998
9999         if (first) {
10000                 struct raid_kobject *rkobj = kzalloc(sizeof(*rkobj), GFP_NOFS);
10001                 if (!rkobj) {
10002                         btrfs_warn(cache->fs_info,
10003                                 "couldn't alloc memory for raid level kobject");
10004                         return;
10005                 }
10006                 rkobj->flags = cache->flags;
10007                 kobject_init(&rkobj->kobj, &btrfs_raid_ktype);
10008
10009                 spin_lock(&fs_info->pending_raid_kobjs_lock);
10010                 list_add_tail(&rkobj->list, &fs_info->pending_raid_kobjs);
10011                 spin_unlock(&fs_info->pending_raid_kobjs_lock);
10012                 space_info->block_group_kobjs[index] = &rkobj->kobj;
10013         }
10014 }
10015
10016 static struct btrfs_block_group_cache *
10017 btrfs_create_block_group_cache(struct btrfs_fs_info *fs_info,
10018                                u64 start, u64 size)
10019 {
10020         struct btrfs_block_group_cache *cache;
10021
10022         cache = kzalloc(sizeof(*cache), GFP_NOFS);
10023         if (!cache)
10024                 return NULL;
10025
10026         cache->free_space_ctl = kzalloc(sizeof(*cache->free_space_ctl),
10027                                         GFP_NOFS);
10028         if (!cache->free_space_ctl) {
10029                 kfree(cache);
10030                 return NULL;
10031         }
10032
10033         cache->key.objectid = start;
10034         cache->key.offset = size;
10035         cache->key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
10036
10037         cache->fs_info = fs_info;
10038         cache->full_stripe_len = btrfs_full_stripe_len(fs_info, start);
10039         set_free_space_tree_thresholds(cache);
10040
10041         atomic_set(&cache->count, 1);
10042         spin_lock_init(&cache->lock);
10043         init_rwsem(&cache->data_rwsem);
10044         INIT_LIST_HEAD(&cache->list);
10045         INIT_LIST_HEAD(&cache->cluster_list);
10046         INIT_LIST_HEAD(&cache->bg_list);
10047         INIT_LIST_HEAD(&cache->ro_list);
10048         INIT_LIST_HEAD(&cache->dirty_list);
10049         INIT_LIST_HEAD(&cache->io_list);
10050         btrfs_init_free_space_ctl(cache);
10051         atomic_set(&cache->trimming, 0);
10052         mutex_init(&cache->free_space_lock);
10053         btrfs_init_full_stripe_locks_tree(&cache->full_stripe_locks_root);
10054
10055         return cache;
10056 }
10057
10058 int btrfs_read_block_groups(struct btrfs_fs_info *info)
10059 {
10060         struct btrfs_path *path;
10061         int ret;
10062         struct btrfs_block_group_cache *cache;
10063         struct btrfs_space_info *space_info;
10064         struct btrfs_key key;
10065         struct btrfs_key found_key;
10066         struct extent_buffer *leaf;
10067         int need_clear = 0;
10068         u64 cache_gen;
10069         u64 feature;
10070         int mixed;
10071
10072         feature = btrfs_super_incompat_flags(info->super_copy);
10073         mixed = !!(feature & BTRFS_FEATURE_INCOMPAT_MIXED_GROUPS);
10074
10075         key.objectid = 0;
10076         key.offset = 0;
10077         key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
10078         path = btrfs_alloc_path();
10079         if (!path)
10080                 return -ENOMEM;
10081         path->reada = READA_FORWARD;
10082
10083         cache_gen = btrfs_super_cache_generation(info->super_copy);
10084         if (btrfs_test_opt(info, SPACE_CACHE) &&
10085             btrfs_super_generation(info->super_copy) != cache_gen)
10086                 need_clear = 1;
10087         if (btrfs_test_opt(info, CLEAR_CACHE))
10088                 need_clear = 1;
10089
10090         while (1) {
10091                 ret = find_first_block_group(info, path, &key);
10092                 if (ret > 0)
10093                         break;
10094                 if (ret != 0)
10095                         goto error;
10096
10097                 leaf = path->nodes[0];
10098                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
10099
10100                 cache = btrfs_create_block_group_cache(info, found_key.objectid,
10101                                                        found_key.offset);
10102                 if (!cache) {
10103                         ret = -ENOMEM;
10104                         goto error;
10105                 }
10106
10107                 if (need_clear) {
10108                         /*
10109                          * When we mount with old space cache, we need to
10110                          * set BTRFS_DC_CLEAR and set dirty flag.
10111                          *
10112                          * a) Setting 'BTRFS_DC_CLEAR' makes sure that we
10113                          *    truncate the old free space cache inode and
10114                          *    setup a new one.
10115                          * b) Setting 'dirty flag' makes sure that we flush
10116                          *    the new space cache info onto disk.
10117                          */
10118                         if (btrfs_test_opt(info, SPACE_CACHE))
10119                                 cache->disk_cache_state = BTRFS_DC_CLEAR;
10120                 }
10121
10122                 read_extent_buffer(leaf, &cache->item,
10123                                    btrfs_item_ptr_offset(leaf, path->slots[0]),
10124                                    sizeof(cache->item));
10125                 cache->flags = btrfs_block_group_flags(&cache->item);
10126                 if (!mixed &&
10127                     ((cache->flags & BTRFS_BLOCK_GROUP_METADATA) &&
10128                     (cache->flags & BTRFS_BLOCK_GROUP_DATA))) {
10129                         btrfs_err(info,
10130 "bg %llu is a mixed block group but filesystem hasn't enabled mixed block groups",
10131                                   cache->key.objectid);
10132                         ret = -EINVAL;
10133                         goto error;
10134                 }
10135
10136                 key.objectid = found_key.objectid + found_key.offset;
10137                 btrfs_release_path(path);
10138
10139                 /*
10140                  * We need to exclude the super stripes now so that the space
10141                  * info has super bytes accounted for, otherwise we'll think
10142                  * we have more space than we actually do.
10143                  */
10144                 ret = exclude_super_stripes(info, cache);
10145                 if (ret) {
10146                         /*
10147                          * We may have excluded something, so call this just in
10148                          * case.
10149                          */
10150                         free_excluded_extents(info, cache);
10151                         btrfs_put_block_group(cache);
10152                         goto error;
10153                 }
10154
10155                 /*
10156                  * check for two cases, either we are full, and therefore
10157                  * don't need to bother with the caching work since we won't
10158                  * find any space, or we are empty, and we can just add all
10159                  * the space in and be done with it.  This saves us _alot_ of
10160                  * time, particularly in the full case.
10161                  */
10162                 if (found_key.offset == btrfs_block_group_used(&cache->item)) {
10163                         cache->last_byte_to_unpin = (u64)-1;
10164                         cache->cached = BTRFS_CACHE_FINISHED;
10165                         free_excluded_extents(info, cache);
10166                 } else if (btrfs_block_group_used(&cache->item) == 0) {
10167                         cache->last_byte_to_unpin = (u64)-1;
10168                         cache->cached = BTRFS_CACHE_FINISHED;
10169                         add_new_free_space(cache, info,
10170                                            found_key.objectid,
10171                                            found_key.objectid +
10172                                            found_key.offset);
10173                         free_excluded_extents(info, cache);
10174                 }
10175
10176                 ret = btrfs_add_block_group_cache(info, cache);
10177                 if (ret) {
10178                         btrfs_remove_free_space_cache(cache);
10179                         btrfs_put_block_group(cache);
10180                         goto error;
10181                 }
10182
10183                 trace_btrfs_add_block_group(info, cache, 0);
10184                 update_space_info(info, cache->flags, found_key.offset,
10185                                   btrfs_block_group_used(&cache->item),
10186                                   cache->bytes_super, &space_info);
10187
10188                 cache->space_info = space_info;
10189
10190                 link_block_group(cache);
10191
10192                 set_avail_alloc_bits(info, cache->flags);
10193                 if (btrfs_chunk_readonly(info, cache->key.objectid)) {
10194                         inc_block_group_ro(cache, 1);
10195                 } else if (btrfs_block_group_used(&cache->item) == 0) {
10196                         spin_lock(&info->unused_bgs_lock);
10197                         /* Should always be true but just in case. */
10198                         if (list_empty(&cache->bg_list)) {
10199                                 btrfs_get_block_group(cache);
10200                                 list_add_tail(&cache->bg_list,
10201                                               &info->unused_bgs);
10202                         }
10203                         spin_unlock(&info->unused_bgs_lock);
10204                 }
10205         }
10206
10207         list_for_each_entry_rcu(space_info, &info->space_info, list) {
10208                 if (!(get_alloc_profile(info, space_info->flags) &
10209                       (BTRFS_BLOCK_GROUP_RAID10 |
10210                        BTRFS_BLOCK_GROUP_RAID1 |
10211                        BTRFS_BLOCK_GROUP_RAID5 |
10212                        BTRFS_BLOCK_GROUP_RAID6 |
10213                        BTRFS_BLOCK_GROUP_DUP)))
10214                         continue;
10215                 /*
10216                  * avoid allocating from un-mirrored block group if there are
10217                  * mirrored block groups.
10218                  */
10219                 list_for_each_entry(cache,
10220                                 &space_info->block_groups[BTRFS_RAID_RAID0],
10221                                 list)
10222                         inc_block_group_ro(cache, 1);
10223                 list_for_each_entry(cache,
10224                                 &space_info->block_groups[BTRFS_RAID_SINGLE],
10225                                 list)
10226                         inc_block_group_ro(cache, 1);
10227         }
10228
10229         btrfs_add_raid_kobjects(info);
10230         init_global_block_rsv(info);
10231         ret = 0;
10232 error:
10233         btrfs_free_path(path);
10234         return ret;
10235 }
10236
10237 void btrfs_create_pending_block_groups(struct btrfs_trans_handle *trans)
10238 {
10239         struct btrfs_fs_info *fs_info = trans->fs_info;
10240         struct btrfs_block_group_cache *block_group, *tmp;
10241         struct btrfs_root *extent_root = fs_info->extent_root;
10242         struct btrfs_block_group_item item;
10243         struct btrfs_key key;
10244         int ret = 0;
10245         bool can_flush_pending_bgs = trans->can_flush_pending_bgs;
10246
10247         trans->can_flush_pending_bgs = false;
10248         list_for_each_entry_safe(block_group, tmp, &trans->new_bgs, bg_list) {
10249                 if (ret)
10250                         goto next;
10251
10252                 spin_lock(&block_group->lock);
10253                 memcpy(&item, &block_group->item, sizeof(item));
10254                 memcpy(&key, &block_group->key, sizeof(key));
10255                 spin_unlock(&block_group->lock);
10256
10257                 ret = btrfs_insert_item(trans, extent_root, &key, &item,
10258                                         sizeof(item));
10259                 if (ret)
10260                         btrfs_abort_transaction(trans, ret);
10261                 ret = btrfs_finish_chunk_alloc(trans, fs_info, key.objectid,
10262                                                key.offset);
10263                 if (ret)
10264                         btrfs_abort_transaction(trans, ret);
10265                 add_block_group_free_space(trans, fs_info, block_group);
10266                 /* already aborted the transaction if it failed. */
10267 next:
10268                 list_del_init(&block_group->bg_list);
10269         }
10270         trans->can_flush_pending_bgs = can_flush_pending_bgs;
10271 }
10272
10273 int btrfs_make_block_group(struct btrfs_trans_handle *trans,
10274                            struct btrfs_fs_info *fs_info, u64 bytes_used,
10275                            u64 type, u64 chunk_offset, u64 size)
10276 {
10277         struct btrfs_block_group_cache *cache;
10278         int ret;
10279
10280         btrfs_set_log_full_commit(fs_info, trans);
10281
10282         cache = btrfs_create_block_group_cache(fs_info, chunk_offset, size);
10283         if (!cache)
10284                 return -ENOMEM;
10285
10286         btrfs_set_block_group_used(&cache->item, bytes_used);
10287         btrfs_set_block_group_chunk_objectid(&cache->item,
10288                                              BTRFS_FIRST_CHUNK_TREE_OBJECTID);
10289         btrfs_set_block_group_flags(&cache->item, type);
10290
10291         cache->flags = type;
10292         cache->last_byte_to_unpin = (u64)-1;
10293         cache->cached = BTRFS_CACHE_FINISHED;
10294         cache->needs_free_space = 1;
10295         ret = exclude_super_stripes(fs_info, cache);
10296         if (ret) {
10297                 /*
10298                  * We may have excluded something, so call this just in
10299                  * case.
10300                  */
10301                 free_excluded_extents(fs_info, cache);
10302                 btrfs_put_block_group(cache);
10303                 return ret;
10304         }
10305
10306         add_new_free_space(cache, fs_info, chunk_offset, chunk_offset + size);
10307
10308         free_excluded_extents(fs_info, cache);
10309
10310 #ifdef CONFIG_BTRFS_DEBUG
10311         if (btrfs_should_fragment_free_space(cache)) {
10312                 u64 new_bytes_used = size - bytes_used;
10313
10314                 bytes_used += new_bytes_used >> 1;
10315                 fragment_free_space(cache);
10316         }
10317 #endif
10318         /*
10319          * Ensure the corresponding space_info object is created and
10320          * assigned to our block group. We want our bg to be added to the rbtree
10321          * with its ->space_info set.
10322          */
10323         cache->space_info = __find_space_info(fs_info, cache->flags);
10324         ASSERT(cache->space_info);
10325
10326         ret = btrfs_add_block_group_cache(fs_info, cache);
10327         if (ret) {
10328                 btrfs_remove_free_space_cache(cache);
10329                 btrfs_put_block_group(cache);
10330                 return ret;
10331         }
10332
10333         /*
10334          * Now that our block group has its ->space_info set and is inserted in
10335          * the rbtree, update the space info's counters.
10336          */
10337         trace_btrfs_add_block_group(fs_info, cache, 1);
10338         update_space_info(fs_info, cache->flags, size, bytes_used,
10339                                 cache->bytes_super, &cache->space_info);
10340         update_global_block_rsv(fs_info);
10341
10342         link_block_group(cache);
10343
10344         list_add_tail(&cache->bg_list, &trans->new_bgs);
10345
10346         set_avail_alloc_bits(fs_info, type);
10347         return 0;
10348 }
10349
10350 static void clear_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
10351 {
10352         u64 extra_flags = chunk_to_extended(flags) &
10353                                 BTRFS_EXTENDED_PROFILE_MASK;
10354
10355         write_seqlock(&fs_info->profiles_lock);
10356         if (flags & BTRFS_BLOCK_GROUP_DATA)
10357                 fs_info->avail_data_alloc_bits &= ~extra_flags;
10358         if (flags & BTRFS_BLOCK_GROUP_METADATA)
10359                 fs_info->avail_metadata_alloc_bits &= ~extra_flags;
10360         if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
10361                 fs_info->avail_system_alloc_bits &= ~extra_flags;
10362         write_sequnlock(&fs_info->profiles_lock);
10363 }
10364
10365 int btrfs_remove_block_group(struct btrfs_trans_handle *trans,
10366                              struct btrfs_fs_info *fs_info, u64 group_start,
10367                              struct extent_map *em)
10368 {
10369         struct btrfs_root *root = fs_info->extent_root;
10370         struct btrfs_path *path;
10371         struct btrfs_block_group_cache *block_group;
10372         struct btrfs_free_cluster *cluster;
10373         struct btrfs_root *tree_root = fs_info->tree_root;
10374         struct btrfs_key key;
10375         struct inode *inode;
10376         struct kobject *kobj = NULL;
10377         int ret;
10378         int index;
10379         int factor;
10380         struct btrfs_caching_control *caching_ctl = NULL;
10381         bool remove_em;
10382
10383         block_group = btrfs_lookup_block_group(fs_info, group_start);
10384         BUG_ON(!block_group);
10385         BUG_ON(!block_group->ro);
10386
10387         /*
10388          * Free the reserved super bytes from this block group before
10389          * remove it.
10390          */
10391         free_excluded_extents(fs_info, block_group);
10392         btrfs_free_ref_tree_range(fs_info, block_group->key.objectid,
10393                                   block_group->key.offset);
10394
10395         memcpy(&key, &block_group->key, sizeof(key));
10396         index = btrfs_bg_flags_to_raid_index(block_group->flags);
10397         if (block_group->flags & (BTRFS_BLOCK_GROUP_DUP |
10398                                   BTRFS_BLOCK_GROUP_RAID1 |
10399                                   BTRFS_BLOCK_GROUP_RAID10))
10400                 factor = 2;
10401         else
10402                 factor = 1;
10403
10404         /* make sure this block group isn't part of an allocation cluster */
10405         cluster = &fs_info->data_alloc_cluster;
10406         spin_lock(&cluster->refill_lock);
10407         btrfs_return_cluster_to_free_space(block_group, cluster);
10408         spin_unlock(&cluster->refill_lock);
10409
10410         /*
10411          * make sure this block group isn't part of a metadata
10412          * allocation cluster
10413          */
10414         cluster = &fs_info->meta_alloc_cluster;
10415         spin_lock(&cluster->refill_lock);
10416         btrfs_return_cluster_to_free_space(block_group, cluster);
10417         spin_unlock(&cluster->refill_lock);
10418
10419         path = btrfs_alloc_path();
10420         if (!path) {
10421                 ret = -ENOMEM;
10422                 goto out;
10423         }
10424
10425         /*
10426          * get the inode first so any iput calls done for the io_list
10427          * aren't the final iput (no unlinks allowed now)
10428          */
10429         inode = lookup_free_space_inode(fs_info, block_group, path);
10430
10431         mutex_lock(&trans->transaction->cache_write_mutex);
10432         /*
10433          * make sure our free spache cache IO is done before remove the
10434          * free space inode
10435          */
10436         spin_lock(&trans->transaction->dirty_bgs_lock);
10437         if (!list_empty(&block_group->io_list)) {
10438                 list_del_init(&block_group->io_list);
10439
10440                 WARN_ON(!IS_ERR(inode) && inode != block_group->io_ctl.inode);
10441
10442                 spin_unlock(&trans->transaction->dirty_bgs_lock);
10443                 btrfs_wait_cache_io(trans, block_group, path);
10444                 btrfs_put_block_group(block_group);
10445                 spin_lock(&trans->transaction->dirty_bgs_lock);
10446         }
10447
10448         if (!list_empty(&block_group->dirty_list)) {
10449                 list_del_init(&block_group->dirty_list);
10450                 btrfs_put_block_group(block_group);
10451         }
10452         spin_unlock(&trans->transaction->dirty_bgs_lock);
10453         mutex_unlock(&trans->transaction->cache_write_mutex);
10454
10455         if (!IS_ERR(inode)) {
10456                 ret = btrfs_orphan_add(trans, BTRFS_I(inode));
10457                 if (ret) {
10458                         btrfs_add_delayed_iput(inode);
10459                         goto out;
10460                 }
10461                 clear_nlink(inode);
10462                 /* One for the block groups ref */
10463                 spin_lock(&block_group->lock);
10464                 if (block_group->iref) {
10465                         block_group->iref = 0;
10466                         block_group->inode = NULL;
10467                         spin_unlock(&block_group->lock);
10468                         iput(inode);
10469                 } else {
10470                         spin_unlock(&block_group->lock);
10471                 }
10472                 /* One for our lookup ref */
10473                 btrfs_add_delayed_iput(inode);
10474         }
10475
10476         key.objectid = BTRFS_FREE_SPACE_OBJECTID;
10477         key.offset = block_group->key.objectid;
10478         key.type = 0;
10479
10480         ret = btrfs_search_slot(trans, tree_root, &key, path, -1, 1);
10481         if (ret < 0)
10482                 goto out;
10483         if (ret > 0)
10484                 btrfs_release_path(path);
10485         if (ret == 0) {
10486                 ret = btrfs_del_item(trans, tree_root, path);
10487                 if (ret)
10488                         goto out;
10489                 btrfs_release_path(path);
10490         }
10491
10492         spin_lock(&fs_info->block_group_cache_lock);
10493         rb_erase(&block_group->cache_node,
10494                  &fs_info->block_group_cache_tree);
10495         RB_CLEAR_NODE(&block_group->cache_node);
10496
10497         if (fs_info->first_logical_byte == block_group->key.objectid)
10498                 fs_info->first_logical_byte = (u64)-1;
10499         spin_unlock(&fs_info->block_group_cache_lock);
10500
10501         down_write(&block_group->space_info->groups_sem);
10502         /*
10503          * we must use list_del_init so people can check to see if they
10504          * are still on the list after taking the semaphore
10505          */
10506         list_del_init(&block_group->list);
10507         if (list_empty(&block_group->space_info->block_groups[index])) {
10508                 kobj = block_group->space_info->block_group_kobjs[index];
10509                 block_group->space_info->block_group_kobjs[index] = NULL;
10510                 clear_avail_alloc_bits(fs_info, block_group->flags);
10511         }
10512         up_write(&block_group->space_info->groups_sem);
10513         if (kobj) {
10514                 kobject_del(kobj);
10515                 kobject_put(kobj);
10516         }
10517
10518         if (block_group->has_caching_ctl)
10519                 caching_ctl = get_caching_control(block_group);
10520         if (block_group->cached == BTRFS_CACHE_STARTED)
10521                 wait_block_group_cache_done(block_group);
10522         if (block_group->has_caching_ctl) {
10523                 down_write(&fs_info->commit_root_sem);
10524                 if (!caching_ctl) {
10525                         struct btrfs_caching_control *ctl;
10526
10527                         list_for_each_entry(ctl,
10528                                     &fs_info->caching_block_groups, list)
10529                                 if (ctl->block_group == block_group) {
10530                                         caching_ctl = ctl;
10531                                         refcount_inc(&caching_ctl->count);
10532                                         break;
10533                                 }
10534                 }
10535                 if (caching_ctl)
10536                         list_del_init(&caching_ctl->list);
10537                 up_write(&fs_info->commit_root_sem);
10538                 if (caching_ctl) {
10539                         /* Once for the caching bgs list and once for us. */
10540                         put_caching_control(caching_ctl);
10541                         put_caching_control(caching_ctl);
10542                 }
10543         }
10544
10545         spin_lock(&trans->transaction->dirty_bgs_lock);
10546         if (!list_empty(&block_group->dirty_list)) {
10547                 WARN_ON(1);
10548         }
10549         if (!list_empty(&block_group->io_list)) {
10550                 WARN_ON(1);
10551         }
10552         spin_unlock(&trans->transaction->dirty_bgs_lock);
10553         btrfs_remove_free_space_cache(block_group);
10554
10555         spin_lock(&block_group->space_info->lock);
10556         list_del_init(&block_group->ro_list);
10557
10558         if (btrfs_test_opt(fs_info, ENOSPC_DEBUG)) {
10559                 WARN_ON(block_group->space_info->total_bytes
10560                         < block_group->key.offset);
10561                 WARN_ON(block_group->space_info->bytes_readonly
10562                         < block_group->key.offset);
10563                 WARN_ON(block_group->space_info->disk_total
10564                         < block_group->key.offset * factor);
10565         }
10566         block_group->space_info->total_bytes -= block_group->key.offset;
10567         block_group->space_info->bytes_readonly -= block_group->key.offset;
10568         block_group->space_info->disk_total -= block_group->key.offset * factor;
10569
10570         spin_unlock(&block_group->space_info->lock);
10571
10572         memcpy(&key, &block_group->key, sizeof(key));
10573
10574         mutex_lock(&fs_info->chunk_mutex);
10575         if (!list_empty(&em->list)) {
10576                 /* We're in the transaction->pending_chunks list. */
10577                 free_extent_map(em);
10578         }
10579         spin_lock(&block_group->lock);
10580         block_group->removed = 1;
10581         /*
10582          * At this point trimming can't start on this block group, because we
10583          * removed the block group from the tree fs_info->block_group_cache_tree
10584          * so no one can't find it anymore and even if someone already got this
10585          * block group before we removed it from the rbtree, they have already
10586          * incremented block_group->trimming - if they didn't, they won't find
10587          * any free space entries because we already removed them all when we
10588          * called btrfs_remove_free_space_cache().
10589          *
10590          * And we must not remove the extent map from the fs_info->mapping_tree
10591          * to prevent the same logical address range and physical device space
10592          * ranges from being reused for a new block group. This is because our
10593          * fs trim operation (btrfs_trim_fs() / btrfs_ioctl_fitrim()) is
10594          * completely transactionless, so while it is trimming a range the
10595          * currently running transaction might finish and a new one start,
10596          * allowing for new block groups to be created that can reuse the same
10597          * physical device locations unless we take this special care.
10598          *
10599          * There may also be an implicit trim operation if the file system
10600          * is mounted with -odiscard. The same protections must remain
10601          * in place until the extents have been discarded completely when
10602          * the transaction commit has completed.
10603          */
10604         remove_em = (atomic_read(&block_group->trimming) == 0);
10605         /*
10606          * Make sure a trimmer task always sees the em in the pinned_chunks list
10607          * if it sees block_group->removed == 1 (needs to lock block_group->lock
10608          * before checking block_group->removed).
10609          */
10610         if (!remove_em) {
10611                 /*
10612                  * Our em might be in trans->transaction->pending_chunks which
10613                  * is protected by fs_info->chunk_mutex ([lock|unlock]_chunks),
10614                  * and so is the fs_info->pinned_chunks list.
10615                  *
10616                  * So at this point we must be holding the chunk_mutex to avoid
10617                  * any races with chunk allocation (more specifically at
10618                  * volumes.c:contains_pending_extent()), to ensure it always
10619                  * sees the em, either in the pending_chunks list or in the
10620                  * pinned_chunks list.
10621                  */
10622                 list_move_tail(&em->list, &fs_info->pinned_chunks);
10623         }
10624         spin_unlock(&block_group->lock);
10625
10626         if (remove_em) {
10627                 struct extent_map_tree *em_tree;
10628
10629                 em_tree = &fs_info->mapping_tree.map_tree;
10630                 write_lock(&em_tree->lock);
10631                 /*
10632                  * The em might be in the pending_chunks list, so make sure the
10633                  * chunk mutex is locked, since remove_extent_mapping() will
10634                  * delete us from that list.
10635                  */
10636                 remove_extent_mapping(em_tree, em);
10637                 write_unlock(&em_tree->lock);
10638                 /* once for the tree */
10639                 free_extent_map(em);
10640         }
10641
10642         mutex_unlock(&fs_info->chunk_mutex);
10643
10644         ret = remove_block_group_free_space(trans, fs_info, block_group);
10645         if (ret)
10646                 goto out;
10647
10648         btrfs_put_block_group(block_group);
10649         btrfs_put_block_group(block_group);
10650
10651         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
10652         if (ret > 0)
10653                 ret = -EIO;
10654         if (ret < 0)
10655                 goto out;
10656
10657         ret = btrfs_del_item(trans, root, path);
10658 out:
10659         btrfs_free_path(path);
10660         return ret;
10661 }
10662
10663 struct btrfs_trans_handle *
10664 btrfs_start_trans_remove_block_group(struct btrfs_fs_info *fs_info,
10665                                      const u64 chunk_offset)
10666 {
10667         struct extent_map_tree *em_tree = &fs_info->mapping_tree.map_tree;
10668         struct extent_map *em;
10669         struct map_lookup *map;
10670         unsigned int num_items;
10671
10672         read_lock(&em_tree->lock);
10673         em = lookup_extent_mapping(em_tree, chunk_offset, 1);
10674         read_unlock(&em_tree->lock);
10675         ASSERT(em && em->start == chunk_offset);
10676
10677         /*
10678          * We need to reserve 3 + N units from the metadata space info in order
10679          * to remove a block group (done at btrfs_remove_chunk() and at
10680          * btrfs_remove_block_group()), which are used for:
10681          *
10682          * 1 unit for adding the free space inode's orphan (located in the tree
10683          * of tree roots).
10684          * 1 unit for deleting the block group item (located in the extent
10685          * tree).
10686          * 1 unit for deleting the free space item (located in tree of tree
10687          * roots).
10688          * N units for deleting N device extent items corresponding to each
10689          * stripe (located in the device tree).
10690          *
10691          * In order to remove a block group we also need to reserve units in the
10692          * system space info in order to update the chunk tree (update one or
10693          * more device items and remove one chunk item), but this is done at
10694          * btrfs_remove_chunk() through a call to check_system_chunk().
10695          */
10696         map = em->map_lookup;
10697         num_items = 3 + map->num_stripes;
10698         free_extent_map(em);
10699
10700         return btrfs_start_transaction_fallback_global_rsv(fs_info->extent_root,
10701                                                            num_items, 1);
10702 }
10703
10704 /*
10705  * Process the unused_bgs list and remove any that don't have any allocated
10706  * space inside of them.
10707  */
10708 void btrfs_delete_unused_bgs(struct btrfs_fs_info *fs_info)
10709 {
10710         struct btrfs_block_group_cache *block_group;
10711         struct btrfs_space_info *space_info;
10712         struct btrfs_trans_handle *trans;
10713         int ret = 0;
10714
10715         if (!test_bit(BTRFS_FS_OPEN, &fs_info->flags))
10716                 return;
10717
10718         spin_lock(&fs_info->unused_bgs_lock);
10719         while (!list_empty(&fs_info->unused_bgs)) {
10720                 u64 start, end;
10721                 int trimming;
10722
10723                 block_group = list_first_entry(&fs_info->unused_bgs,
10724                                                struct btrfs_block_group_cache,
10725                                                bg_list);
10726                 list_del_init(&block_group->bg_list);
10727
10728                 space_info = block_group->space_info;
10729
10730                 if (ret || btrfs_mixed_space_info(space_info)) {
10731                         btrfs_put_block_group(block_group);
10732                         continue;
10733                 }
10734                 spin_unlock(&fs_info->unused_bgs_lock);
10735
10736                 mutex_lock(&fs_info->delete_unused_bgs_mutex);
10737
10738                 /* Don't want to race with allocators so take the groups_sem */
10739                 down_write(&space_info->groups_sem);
10740                 spin_lock(&block_group->lock);
10741                 if (block_group->reserved ||
10742                     btrfs_block_group_used(&block_group->item) ||
10743                     block_group->ro ||
10744                     list_is_singular(&block_group->list)) {
10745                         /*
10746                          * We want to bail if we made new allocations or have
10747                          * outstanding allocations in this block group.  We do
10748                          * the ro check in case balance is currently acting on
10749                          * this block group.
10750                          */
10751                         spin_unlock(&block_group->lock);
10752                         up_write(&space_info->groups_sem);
10753                         goto next;
10754                 }
10755                 spin_unlock(&block_group->lock);
10756
10757                 /* We don't want to force the issue, only flip if it's ok. */
10758                 ret = inc_block_group_ro(block_group, 0);
10759                 up_write(&space_info->groups_sem);
10760                 if (ret < 0) {
10761                         ret = 0;
10762                         goto next;
10763                 }
10764
10765                 /*
10766                  * Want to do this before we do anything else so we can recover
10767                  * properly if we fail to join the transaction.
10768                  */
10769                 trans = btrfs_start_trans_remove_block_group(fs_info,
10770                                                      block_group->key.objectid);
10771                 if (IS_ERR(trans)) {
10772                         btrfs_dec_block_group_ro(block_group);
10773                         ret = PTR_ERR(trans);
10774                         goto next;
10775                 }
10776
10777                 /*
10778                  * We could have pending pinned extents for this block group,
10779                  * just delete them, we don't care about them anymore.
10780                  */
10781                 start = block_group->key.objectid;
10782                 end = start + block_group->key.offset - 1;
10783                 /*
10784                  * Hold the unused_bg_unpin_mutex lock to avoid racing with
10785                  * btrfs_finish_extent_commit(). If we are at transaction N,
10786                  * another task might be running finish_extent_commit() for the
10787                  * previous transaction N - 1, and have seen a range belonging
10788                  * to the block group in freed_extents[] before we were able to
10789                  * clear the whole block group range from freed_extents[]. This
10790                  * means that task can lookup for the block group after we
10791                  * unpinned it from freed_extents[] and removed it, leading to
10792                  * a BUG_ON() at btrfs_unpin_extent_range().
10793                  */
10794                 mutex_lock(&fs_info->unused_bg_unpin_mutex);
10795                 ret = clear_extent_bits(&fs_info->freed_extents[0], start, end,
10796                                   EXTENT_DIRTY);
10797                 if (ret) {
10798                         mutex_unlock(&fs_info->unused_bg_unpin_mutex);
10799                         btrfs_dec_block_group_ro(block_group);
10800                         goto end_trans;
10801                 }
10802                 ret = clear_extent_bits(&fs_info->freed_extents[1], start, end,
10803                                   EXTENT_DIRTY);
10804                 if (ret) {
10805                         mutex_unlock(&fs_info->unused_bg_unpin_mutex);
10806                         btrfs_dec_block_group_ro(block_group);
10807                         goto end_trans;
10808                 }
10809                 mutex_unlock(&fs_info->unused_bg_unpin_mutex);
10810
10811                 /* Reset pinned so btrfs_put_block_group doesn't complain */
10812                 spin_lock(&space_info->lock);
10813                 spin_lock(&block_group->lock);
10814
10815                 space_info->bytes_pinned -= block_group->pinned;
10816                 space_info->bytes_readonly += block_group->pinned;
10817                 percpu_counter_add(&space_info->total_bytes_pinned,
10818                                    -block_group->pinned);
10819                 block_group->pinned = 0;
10820
10821                 spin_unlock(&block_group->lock);
10822                 spin_unlock(&space_info->lock);
10823
10824                 /* DISCARD can flip during remount */
10825                 trimming = btrfs_test_opt(fs_info, DISCARD);
10826
10827                 /* Implicit trim during transaction commit. */
10828                 if (trimming)
10829                         btrfs_get_block_group_trimming(block_group);
10830
10831                 /*
10832                  * Btrfs_remove_chunk will abort the transaction if things go
10833                  * horribly wrong.
10834                  */
10835                 ret = btrfs_remove_chunk(trans, fs_info,
10836                                          block_group->key.objectid);
10837
10838                 if (ret) {
10839                         if (trimming)
10840                                 btrfs_put_block_group_trimming(block_group);
10841                         goto end_trans;
10842                 }
10843
10844                 /*
10845                  * If we're not mounted with -odiscard, we can just forget
10846                  * about this block group. Otherwise we'll need to wait
10847                  * until transaction commit to do the actual discard.
10848                  */
10849                 if (trimming) {
10850                         spin_lock(&fs_info->unused_bgs_lock);
10851                         /*
10852                          * A concurrent scrub might have added us to the list
10853                          * fs_info->unused_bgs, so use a list_move operation
10854                          * to add the block group to the deleted_bgs list.
10855                          */
10856                         list_move(&block_group->bg_list,
10857                                   &trans->transaction->deleted_bgs);
10858                         spin_unlock(&fs_info->unused_bgs_lock);
10859                         btrfs_get_block_group(block_group);
10860                 }
10861 end_trans:
10862                 btrfs_end_transaction(trans);
10863 next:
10864                 mutex_unlock(&fs_info->delete_unused_bgs_mutex);
10865                 btrfs_put_block_group(block_group);
10866                 spin_lock(&fs_info->unused_bgs_lock);
10867         }
10868         spin_unlock(&fs_info->unused_bgs_lock);
10869 }
10870
10871 int btrfs_init_space_info(struct btrfs_fs_info *fs_info)
10872 {
10873         struct btrfs_space_info *space_info;
10874         struct btrfs_super_block *disk_super;
10875         u64 features;
10876         u64 flags;
10877         int mixed = 0;
10878         int ret;
10879
10880         disk_super = fs_info->super_copy;
10881         if (!btrfs_super_root(disk_super))
10882                 return -EINVAL;
10883
10884         features = btrfs_super_incompat_flags(disk_super);
10885         if (features & BTRFS_FEATURE_INCOMPAT_MIXED_GROUPS)
10886                 mixed = 1;
10887
10888         flags = BTRFS_BLOCK_GROUP_SYSTEM;
10889         ret = create_space_info(fs_info, flags, &space_info);
10890         if (ret)
10891                 goto out;
10892
10893         if (mixed) {
10894                 flags = BTRFS_BLOCK_GROUP_METADATA | BTRFS_BLOCK_GROUP_DATA;
10895                 ret = create_space_info(fs_info, flags, &space_info);
10896         } else {
10897                 flags = BTRFS_BLOCK_GROUP_METADATA;
10898                 ret = create_space_info(fs_info, flags, &space_info);
10899                 if (ret)
10900                         goto out;
10901
10902                 flags = BTRFS_BLOCK_GROUP_DATA;
10903                 ret = create_space_info(fs_info, flags, &space_info);
10904         }
10905 out:
10906         return ret;
10907 }
10908
10909 int btrfs_error_unpin_extent_range(struct btrfs_fs_info *fs_info,
10910                                    u64 start, u64 end)
10911 {
10912         return unpin_extent_range(fs_info, start, end, false);
10913 }
10914
10915 /*
10916  * It used to be that old block groups would be left around forever.
10917  * Iterating over them would be enough to trim unused space.  Since we
10918  * now automatically remove them, we also need to iterate over unallocated
10919  * space.
10920  *
10921  * We don't want a transaction for this since the discard may take a
10922  * substantial amount of time.  We don't require that a transaction be
10923  * running, but we do need to take a running transaction into account
10924  * to ensure that we're not discarding chunks that were released in
10925  * the current transaction.
10926  *
10927  * Holding the chunks lock will prevent other threads from allocating
10928  * or releasing chunks, but it won't prevent a running transaction
10929  * from committing and releasing the memory that the pending chunks
10930  * list head uses.  For that, we need to take a reference to the
10931  * transaction.
10932  */
10933 static int btrfs_trim_free_extents(struct btrfs_device *device,
10934                                    u64 minlen, u64 *trimmed)
10935 {
10936         u64 start = 0, len = 0;
10937         int ret;
10938
10939         *trimmed = 0;
10940
10941         /* Not writeable = nothing to do. */
10942         if (!test_bit(BTRFS_DEV_STATE_WRITEABLE, &device->dev_state))
10943                 return 0;
10944
10945         /* No free space = nothing to do. */
10946         if (device->total_bytes <= device->bytes_used)
10947                 return 0;
10948
10949         ret = 0;
10950
10951         while (1) {
10952                 struct btrfs_fs_info *fs_info = device->fs_info;
10953                 struct btrfs_transaction *trans;
10954                 u64 bytes;
10955
10956                 ret = mutex_lock_interruptible(&fs_info->chunk_mutex);
10957                 if (ret)
10958                         return ret;
10959
10960                 down_read(&fs_info->commit_root_sem);
10961
10962                 spin_lock(&fs_info->trans_lock);
10963                 trans = fs_info->running_transaction;
10964                 if (trans)
10965                         refcount_inc(&trans->use_count);
10966                 spin_unlock(&fs_info->trans_lock);
10967
10968                 ret = find_free_dev_extent_start(trans, device, minlen, start,
10969                                                  &start, &len);
10970                 if (trans)
10971                         btrfs_put_transaction(trans);
10972
10973                 if (ret) {
10974                         up_read(&fs_info->commit_root_sem);
10975                         mutex_unlock(&fs_info->chunk_mutex);
10976                         if (ret == -ENOSPC)
10977                                 ret = 0;
10978                         break;
10979                 }
10980
10981                 ret = btrfs_issue_discard(device->bdev, start, len, &bytes);
10982                 up_read(&fs_info->commit_root_sem);
10983                 mutex_unlock(&fs_info->chunk_mutex);
10984
10985                 if (ret)
10986                         break;
10987
10988                 start += len;
10989                 *trimmed += bytes;
10990
10991                 if (fatal_signal_pending(current)) {
10992                         ret = -ERESTARTSYS;
10993                         break;
10994                 }
10995
10996                 cond_resched();
10997         }
10998
10999         return ret;
11000 }
11001
11002 int btrfs_trim_fs(struct btrfs_fs_info *fs_info, struct fstrim_range *range)
11003 {
11004         struct btrfs_block_group_cache *cache = NULL;
11005         struct btrfs_device *device;
11006         struct list_head *devices;
11007         u64 group_trimmed;
11008         u64 start;
11009         u64 end;
11010         u64 trimmed = 0;
11011         u64 total_bytes = btrfs_super_total_bytes(fs_info->super_copy);
11012         int ret = 0;
11013
11014         /*
11015          * try to trim all FS space, our block group may start from non-zero.
11016          */
11017         if (range->len == total_bytes)
11018                 cache = btrfs_lookup_first_block_group(fs_info, range->start);
11019         else
11020                 cache = btrfs_lookup_block_group(fs_info, range->start);
11021
11022         while (cache) {
11023                 if (cache->key.objectid >= (range->start + range->len)) {
11024                         btrfs_put_block_group(cache);
11025                         break;
11026                 }
11027
11028                 start = max(range->start, cache->key.objectid);
11029                 end = min(range->start + range->len,
11030                                 cache->key.objectid + cache->key.offset);
11031
11032                 if (end - start >= range->minlen) {
11033                         if (!block_group_cache_done(cache)) {
11034                                 ret = cache_block_group(cache, 0);
11035                                 if (ret) {
11036                                         btrfs_put_block_group(cache);
11037                                         break;
11038                                 }
11039                                 ret = wait_block_group_cache_done(cache);
11040                                 if (ret) {
11041                                         btrfs_put_block_group(cache);
11042                                         break;
11043                                 }
11044                         }
11045                         ret = btrfs_trim_block_group(cache,
11046                                                      &group_trimmed,
11047                                                      start,
11048                                                      end,
11049                                                      range->minlen);
11050
11051                         trimmed += group_trimmed;
11052                         if (ret) {
11053                                 btrfs_put_block_group(cache);
11054                                 break;
11055                         }
11056                 }
11057
11058                 cache = next_block_group(fs_info, cache);
11059         }
11060
11061         mutex_lock(&fs_info->fs_devices->device_list_mutex);
11062         devices = &fs_info->fs_devices->alloc_list;
11063         list_for_each_entry(device, devices, dev_alloc_list) {
11064                 ret = btrfs_trim_free_extents(device, range->minlen,
11065                                               &group_trimmed);
11066                 if (ret)
11067                         break;
11068
11069                 trimmed += group_trimmed;
11070         }
11071         mutex_unlock(&fs_info->fs_devices->device_list_mutex);
11072
11073         range->len = trimmed;
11074         return ret;
11075 }
11076
11077 /*
11078  * btrfs_{start,end}_write_no_snapshotting() are similar to
11079  * mnt_{want,drop}_write(), they are used to prevent some tasks from writing
11080  * data into the page cache through nocow before the subvolume is snapshoted,
11081  * but flush the data into disk after the snapshot creation, or to prevent
11082  * operations while snapshotting is ongoing and that cause the snapshot to be
11083  * inconsistent (writes followed by expanding truncates for example).
11084  */
11085 void btrfs_end_write_no_snapshotting(struct btrfs_root *root)
11086 {
11087         percpu_counter_dec(&root->subv_writers->counter);
11088         /*
11089          * Make sure counter is updated before we wake up waiters.
11090          */
11091         smp_mb();
11092         if (waitqueue_active(&root->subv_writers->wait))
11093                 wake_up(&root->subv_writers->wait);
11094 }
11095
11096 int btrfs_start_write_no_snapshotting(struct btrfs_root *root)
11097 {
11098         if (atomic_read(&root->will_be_snapshotted))
11099                 return 0;
11100
11101         percpu_counter_inc(&root->subv_writers->counter);
11102         /*
11103          * Make sure counter is updated before we check for snapshot creation.
11104          */
11105         smp_mb();
11106         if (atomic_read(&root->will_be_snapshotted)) {
11107                 btrfs_end_write_no_snapshotting(root);
11108                 return 0;
11109         }
11110         return 1;
11111 }
11112
11113 void btrfs_wait_for_snapshot_creation(struct btrfs_root *root)
11114 {
11115         while (true) {
11116                 int ret;
11117
11118                 ret = btrfs_start_write_no_snapshotting(root);
11119                 if (ret)
11120                         break;
11121                 wait_var_event(&root->will_be_snapshotted,
11122                                !atomic_read(&root->will_be_snapshotted));
11123         }
11124 }