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         spin_lock(&root->fs_info->trans_lock);
3146         cur_trans = root->fs_info->running_transaction;
3147         if (cur_trans)
3148                 refcount_inc(&cur_trans->use_count);
3149         spin_unlock(&root->fs_info->trans_lock);
3150         if (!cur_trans)
3151                 return 0;
3152
3153         delayed_refs = &cur_trans->delayed_refs;
3154         spin_lock(&delayed_refs->lock);
3155         head = btrfs_find_delayed_ref_head(delayed_refs, bytenr);
3156         if (!head) {
3157                 spin_unlock(&delayed_refs->lock);
3158                 btrfs_put_transaction(cur_trans);
3159                 return 0;
3160         }
3161
3162         if (!mutex_trylock(&head->mutex)) {
3163                 refcount_inc(&head->refs);
3164                 spin_unlock(&delayed_refs->lock);
3165
3166                 btrfs_release_path(path);
3167
3168                 /*
3169                  * Mutex was contended, block until it's released and let
3170                  * caller try again
3171                  */
3172                 mutex_lock(&head->mutex);
3173                 mutex_unlock(&head->mutex);
3174                 btrfs_put_delayed_ref_head(head);
3175                 btrfs_put_transaction(cur_trans);
3176                 return -EAGAIN;
3177         }
3178         spin_unlock(&delayed_refs->lock);
3179
3180         spin_lock(&head->lock);
3181         /*
3182          * XXX: We should replace this with a proper search function in the
3183          * future.
3184          */
3185         for (node = rb_first(&head->ref_tree); node; node = rb_next(node)) {
3186                 ref = rb_entry(node, struct btrfs_delayed_ref_node, ref_node);
3187                 /* If it's a shared ref we know a cross reference exists */
3188                 if (ref->type != BTRFS_EXTENT_DATA_REF_KEY) {
3189                         ret = 1;
3190                         break;
3191                 }
3192
3193                 data_ref = btrfs_delayed_node_to_data_ref(ref);
3194
3195                 /*
3196                  * If our ref doesn't match the one we're currently looking at
3197                  * then we have a cross reference.
3198                  */
3199                 if (data_ref->root != root->root_key.objectid ||
3200                     data_ref->objectid != objectid ||
3201                     data_ref->offset != offset) {
3202                         ret = 1;
3203                         break;
3204                 }
3205         }
3206         spin_unlock(&head->lock);
3207         mutex_unlock(&head->mutex);
3208         btrfs_put_transaction(cur_trans);
3209         return ret;
3210 }
3211
3212 static noinline int check_committed_ref(struct btrfs_root *root,
3213                                         struct btrfs_path *path,
3214                                         u64 objectid, u64 offset, u64 bytenr)
3215 {
3216         struct btrfs_fs_info *fs_info = root->fs_info;
3217         struct btrfs_root *extent_root = fs_info->extent_root;
3218         struct extent_buffer *leaf;
3219         struct btrfs_extent_data_ref *ref;
3220         struct btrfs_extent_inline_ref *iref;
3221         struct btrfs_extent_item *ei;
3222         struct btrfs_key key;
3223         u32 item_size;
3224         int type;
3225         int ret;
3226
3227         key.objectid = bytenr;
3228         key.offset = (u64)-1;
3229         key.type = BTRFS_EXTENT_ITEM_KEY;
3230
3231         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
3232         if (ret < 0)
3233                 goto out;
3234         BUG_ON(ret == 0); /* Corruption */
3235
3236         ret = -ENOENT;
3237         if (path->slots[0] == 0)
3238                 goto out;
3239
3240         path->slots[0]--;
3241         leaf = path->nodes[0];
3242         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
3243
3244         if (key.objectid != bytenr || key.type != BTRFS_EXTENT_ITEM_KEY)
3245                 goto out;
3246
3247         ret = 1;
3248         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
3249 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
3250         if (item_size < sizeof(*ei)) {
3251                 WARN_ON(item_size != sizeof(struct btrfs_extent_item_v0));
3252                 goto out;
3253         }
3254 #endif
3255         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
3256
3257         if (item_size != sizeof(*ei) +
3258             btrfs_extent_inline_ref_size(BTRFS_EXTENT_DATA_REF_KEY))
3259                 goto out;
3260
3261         if (btrfs_extent_generation(leaf, ei) <=
3262             btrfs_root_last_snapshot(&root->root_item))
3263                 goto out;
3264
3265         iref = (struct btrfs_extent_inline_ref *)(ei + 1);
3266
3267         type = btrfs_get_extent_inline_ref_type(leaf, iref, BTRFS_REF_TYPE_DATA);
3268         if (type != BTRFS_EXTENT_DATA_REF_KEY)
3269                 goto out;
3270
3271         ref = (struct btrfs_extent_data_ref *)(&iref->offset);
3272         if (btrfs_extent_refs(leaf, ei) !=
3273             btrfs_extent_data_ref_count(leaf, ref) ||
3274             btrfs_extent_data_ref_root(leaf, ref) !=
3275             root->root_key.objectid ||
3276             btrfs_extent_data_ref_objectid(leaf, ref) != objectid ||
3277             btrfs_extent_data_ref_offset(leaf, ref) != offset)
3278                 goto out;
3279
3280         ret = 0;
3281 out:
3282         return ret;
3283 }
3284
3285 int btrfs_cross_ref_exist(struct btrfs_root *root, u64 objectid, u64 offset,
3286                           u64 bytenr)
3287 {
3288         struct btrfs_path *path;
3289         int ret;
3290         int ret2;
3291
3292         path = btrfs_alloc_path();
3293         if (!path)
3294                 return -ENOENT;
3295
3296         do {
3297                 ret = check_committed_ref(root, path, objectid,
3298                                           offset, bytenr);
3299                 if (ret && ret != -ENOENT)
3300                         goto out;
3301
3302                 ret2 = check_delayed_ref(root, path, objectid,
3303                                          offset, bytenr);
3304         } while (ret2 == -EAGAIN);
3305
3306         if (ret2 && ret2 != -ENOENT) {
3307                 ret = ret2;
3308                 goto out;
3309         }
3310
3311         if (ret != -ENOENT || ret2 != -ENOENT)
3312                 ret = 0;
3313 out:
3314         btrfs_free_path(path);
3315         if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID)
3316                 WARN_ON(ret > 0);
3317         return ret;
3318 }
3319
3320 static int __btrfs_mod_ref(struct btrfs_trans_handle *trans,
3321                            struct btrfs_root *root,
3322                            struct extent_buffer *buf,
3323                            int full_backref, int inc)
3324 {
3325         struct btrfs_fs_info *fs_info = root->fs_info;
3326         u64 bytenr;
3327         u64 num_bytes;
3328         u64 parent;
3329         u64 ref_root;
3330         u32 nritems;
3331         struct btrfs_key key;
3332         struct btrfs_file_extent_item *fi;
3333         int i;
3334         int level;
3335         int ret = 0;
3336         int (*process_func)(struct btrfs_trans_handle *,
3337                             struct btrfs_root *,
3338                             u64, u64, u64, u64, u64, u64);
3339
3340
3341         if (btrfs_is_testing(fs_info))
3342                 return 0;
3343
3344         ref_root = btrfs_header_owner(buf);
3345         nritems = btrfs_header_nritems(buf);
3346         level = btrfs_header_level(buf);
3347
3348         if (!test_bit(BTRFS_ROOT_REF_COWS, &root->state) && level == 0)
3349                 return 0;
3350
3351         if (inc)
3352                 process_func = btrfs_inc_extent_ref;
3353         else
3354                 process_func = btrfs_free_extent;
3355
3356         if (full_backref)
3357                 parent = buf->start;
3358         else
3359                 parent = 0;
3360
3361         for (i = 0; i < nritems; i++) {
3362                 if (level == 0) {
3363                         btrfs_item_key_to_cpu(buf, &key, i);
3364                         if (key.type != BTRFS_EXTENT_DATA_KEY)
3365                                 continue;
3366                         fi = btrfs_item_ptr(buf, i,
3367                                             struct btrfs_file_extent_item);
3368                         if (btrfs_file_extent_type(buf, fi) ==
3369                             BTRFS_FILE_EXTENT_INLINE)
3370                                 continue;
3371                         bytenr = btrfs_file_extent_disk_bytenr(buf, fi);
3372                         if (bytenr == 0)
3373                                 continue;
3374
3375                         num_bytes = btrfs_file_extent_disk_num_bytes(buf, fi);
3376                         key.offset -= btrfs_file_extent_offset(buf, fi);
3377                         ret = process_func(trans, root, bytenr, num_bytes,
3378                                            parent, ref_root, key.objectid,
3379                                            key.offset);
3380                         if (ret)
3381                                 goto fail;
3382                 } else {
3383                         bytenr = btrfs_node_blockptr(buf, i);
3384                         num_bytes = fs_info->nodesize;
3385                         ret = process_func(trans, root, bytenr, num_bytes,
3386                                            parent, ref_root, level - 1, 0);
3387                         if (ret)
3388                                 goto fail;
3389                 }
3390         }
3391         return 0;
3392 fail:
3393         return ret;
3394 }
3395
3396 int btrfs_inc_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
3397                   struct extent_buffer *buf, int full_backref)
3398 {
3399         return __btrfs_mod_ref(trans, root, buf, full_backref, 1);
3400 }
3401
3402 int btrfs_dec_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
3403                   struct extent_buffer *buf, int full_backref)
3404 {
3405         return __btrfs_mod_ref(trans, root, buf, full_backref, 0);
3406 }
3407
3408 static int write_one_cache_group(struct btrfs_trans_handle *trans,
3409                                  struct btrfs_fs_info *fs_info,
3410                                  struct btrfs_path *path,
3411                                  struct btrfs_block_group_cache *cache)
3412 {
3413         int ret;
3414         struct btrfs_root *extent_root = fs_info->extent_root;
3415         unsigned long bi;
3416         struct extent_buffer *leaf;
3417
3418         ret = btrfs_search_slot(trans, extent_root, &cache->key, path, 0, 1);
3419         if (ret) {
3420                 if (ret > 0)
3421                         ret = -ENOENT;
3422                 goto fail;
3423         }
3424
3425         leaf = path->nodes[0];
3426         bi = btrfs_item_ptr_offset(leaf, path->slots[0]);
3427         write_extent_buffer(leaf, &cache->item, bi, sizeof(cache->item));
3428         btrfs_mark_buffer_dirty(leaf);
3429 fail:
3430         btrfs_release_path(path);
3431         return ret;
3432
3433 }
3434
3435 static struct btrfs_block_group_cache *
3436 next_block_group(struct btrfs_fs_info *fs_info,
3437                  struct btrfs_block_group_cache *cache)
3438 {
3439         struct rb_node *node;
3440
3441         spin_lock(&fs_info->block_group_cache_lock);
3442
3443         /* If our block group was removed, we need a full search. */
3444         if (RB_EMPTY_NODE(&cache->cache_node)) {
3445                 const u64 next_bytenr = cache->key.objectid + cache->key.offset;
3446
3447                 spin_unlock(&fs_info->block_group_cache_lock);
3448                 btrfs_put_block_group(cache);
3449                 cache = btrfs_lookup_first_block_group(fs_info, next_bytenr); return cache;
3450         }
3451         node = rb_next(&cache->cache_node);
3452         btrfs_put_block_group(cache);
3453         if (node) {
3454                 cache = rb_entry(node, struct btrfs_block_group_cache,
3455                                  cache_node);
3456                 btrfs_get_block_group(cache);
3457         } else
3458                 cache = NULL;
3459         spin_unlock(&fs_info->block_group_cache_lock);
3460         return cache;
3461 }
3462
3463 static int cache_save_setup(struct btrfs_block_group_cache *block_group,
3464                             struct btrfs_trans_handle *trans,
3465                             struct btrfs_path *path)
3466 {
3467         struct btrfs_fs_info *fs_info = block_group->fs_info;
3468         struct btrfs_root *root = fs_info->tree_root;
3469         struct inode *inode = NULL;
3470         struct extent_changeset *data_reserved = NULL;
3471         u64 alloc_hint = 0;
3472         int dcs = BTRFS_DC_ERROR;
3473         u64 num_pages = 0;
3474         int retries = 0;
3475         int ret = 0;
3476
3477         /*
3478          * If this block group is smaller than 100 megs don't bother caching the
3479          * block group.
3480          */
3481         if (block_group->key.offset < (100 * SZ_1M)) {
3482                 spin_lock(&block_group->lock);
3483                 block_group->disk_cache_state = BTRFS_DC_WRITTEN;
3484                 spin_unlock(&block_group->lock);
3485                 return 0;
3486         }
3487
3488         if (trans->aborted)
3489                 return 0;
3490 again:
3491         inode = lookup_free_space_inode(fs_info, block_group, path);
3492         if (IS_ERR(inode) && PTR_ERR(inode) != -ENOENT) {
3493                 ret = PTR_ERR(inode);
3494                 btrfs_release_path(path);
3495                 goto out;
3496         }
3497
3498         if (IS_ERR(inode)) {
3499                 BUG_ON(retries);
3500                 retries++;
3501
3502                 if (block_group->ro)
3503                         goto out_free;
3504
3505                 ret = create_free_space_inode(fs_info, trans, block_group,
3506                                               path);
3507                 if (ret)
3508                         goto out_free;
3509                 goto again;
3510         }
3511
3512         /*
3513          * We want to set the generation to 0, that way if anything goes wrong
3514          * from here on out we know not to trust this cache when we load up next
3515          * time.
3516          */
3517         BTRFS_I(inode)->generation = 0;
3518         ret = btrfs_update_inode(trans, root, inode);
3519         if (ret) {
3520                 /*
3521                  * So theoretically we could recover from this, simply set the
3522                  * super cache generation to 0 so we know to invalidate the
3523                  * cache, but then we'd have to keep track of the block groups
3524                  * that fail this way so we know we _have_ to reset this cache
3525                  * before the next commit or risk reading stale cache.  So to
3526                  * limit our exposure to horrible edge cases lets just abort the
3527                  * transaction, this only happens in really bad situations
3528                  * anyway.
3529                  */
3530                 btrfs_abort_transaction(trans, ret);
3531                 goto out_put;
3532         }
3533         WARN_ON(ret);
3534
3535         /* We've already setup this transaction, go ahead and exit */
3536         if (block_group->cache_generation == trans->transid &&
3537             i_size_read(inode)) {
3538                 dcs = BTRFS_DC_SETUP;
3539                 goto out_put;
3540         }
3541
3542         if (i_size_read(inode) > 0) {
3543                 ret = btrfs_check_trunc_cache_free_space(fs_info,
3544                                         &fs_info->global_block_rsv);
3545                 if (ret)
3546                         goto out_put;
3547
3548                 ret = btrfs_truncate_free_space_cache(trans, NULL, inode);
3549                 if (ret)
3550                         goto out_put;
3551         }
3552
3553         spin_lock(&block_group->lock);
3554         if (block_group->cached != BTRFS_CACHE_FINISHED ||
3555             !btrfs_test_opt(fs_info, SPACE_CACHE)) {
3556                 /*
3557                  * don't bother trying to write stuff out _if_
3558                  * a) we're not cached,
3559                  * b) we're with nospace_cache mount option,
3560                  * c) we're with v2 space_cache (FREE_SPACE_TREE).
3561                  */
3562                 dcs = BTRFS_DC_WRITTEN;
3563                 spin_unlock(&block_group->lock);
3564                 goto out_put;
3565         }
3566         spin_unlock(&block_group->lock);
3567
3568         /*
3569          * We hit an ENOSPC when setting up the cache in this transaction, just
3570          * skip doing the setup, we've already cleared the cache so we're safe.
3571          */
3572         if (test_bit(BTRFS_TRANS_CACHE_ENOSPC, &trans->transaction->flags)) {
3573                 ret = -ENOSPC;
3574                 goto out_put;
3575         }
3576
3577         /*
3578          * Try to preallocate enough space based on how big the block group is.
3579          * Keep in mind this has to include any pinned space which could end up
3580          * taking up quite a bit since it's not folded into the other space
3581          * cache.
3582          */
3583         num_pages = div_u64(block_group->key.offset, SZ_256M);
3584         if (!num_pages)
3585                 num_pages = 1;
3586
3587         num_pages *= 16;
3588         num_pages *= PAGE_SIZE;
3589
3590         ret = btrfs_check_data_free_space(inode, &data_reserved, 0, num_pages);
3591         if (ret)
3592                 goto out_put;
3593
3594         ret = btrfs_prealloc_file_range_trans(inode, trans, 0, 0, num_pages,
3595                                               num_pages, num_pages,
3596                                               &alloc_hint);
3597         /*
3598          * Our cache requires contiguous chunks so that we don't modify a bunch
3599          * of metadata or split extents when writing the cache out, which means
3600          * we can enospc if we are heavily fragmented in addition to just normal
3601          * out of space conditions.  So if we hit this just skip setting up any
3602          * other block groups for this transaction, maybe we'll unpin enough
3603          * space the next time around.
3604          */
3605         if (!ret)
3606                 dcs = BTRFS_DC_SETUP;
3607         else if (ret == -ENOSPC)
3608                 set_bit(BTRFS_TRANS_CACHE_ENOSPC, &trans->transaction->flags);
3609
3610 out_put:
3611         iput(inode);
3612 out_free:
3613         btrfs_release_path(path);
3614 out:
3615         spin_lock(&block_group->lock);
3616         if (!ret && dcs == BTRFS_DC_SETUP)
3617                 block_group->cache_generation = trans->transid;
3618         block_group->disk_cache_state = dcs;
3619         spin_unlock(&block_group->lock);
3620
3621         extent_changeset_free(data_reserved);
3622         return ret;
3623 }
3624
3625 int btrfs_setup_space_cache(struct btrfs_trans_handle *trans,
3626                             struct btrfs_fs_info *fs_info)
3627 {
3628         struct btrfs_block_group_cache *cache, *tmp;
3629         struct btrfs_transaction *cur_trans = trans->transaction;
3630         struct btrfs_path *path;
3631
3632         if (list_empty(&cur_trans->dirty_bgs) ||
3633             !btrfs_test_opt(fs_info, SPACE_CACHE))
3634                 return 0;
3635
3636         path = btrfs_alloc_path();
3637         if (!path)
3638                 return -ENOMEM;
3639
3640         /* Could add new block groups, use _safe just in case */
3641         list_for_each_entry_safe(cache, tmp, &cur_trans->dirty_bgs,
3642                                  dirty_list) {
3643                 if (cache->disk_cache_state == BTRFS_DC_CLEAR)
3644                         cache_save_setup(cache, trans, path);
3645         }
3646
3647         btrfs_free_path(path);
3648         return 0;
3649 }
3650
3651 /*
3652  * transaction commit does final block group cache writeback during a
3653  * critical section where nothing is allowed to change the FS.  This is
3654  * required in order for the cache to actually match the block group,
3655  * but can introduce a lot of latency into the commit.
3656  *
3657  * So, btrfs_start_dirty_block_groups is here to kick off block group
3658  * cache IO.  There's a chance we'll have to redo some of it if the
3659  * block group changes again during the commit, but it greatly reduces
3660  * the commit latency by getting rid of the easy block groups while
3661  * we're still allowing others to join the commit.
3662  */
3663 int btrfs_start_dirty_block_groups(struct btrfs_trans_handle *trans)
3664 {
3665         struct btrfs_fs_info *fs_info = trans->fs_info;
3666         struct btrfs_block_group_cache *cache;
3667         struct btrfs_transaction *cur_trans = trans->transaction;
3668         int ret = 0;
3669         int should_put;
3670         struct btrfs_path *path = NULL;
3671         LIST_HEAD(dirty);
3672         struct list_head *io = &cur_trans->io_bgs;
3673         int num_started = 0;
3674         int loops = 0;
3675
3676         spin_lock(&cur_trans->dirty_bgs_lock);
3677         if (list_empty(&cur_trans->dirty_bgs)) {
3678                 spin_unlock(&cur_trans->dirty_bgs_lock);
3679                 return 0;
3680         }
3681         list_splice_init(&cur_trans->dirty_bgs, &dirty);
3682         spin_unlock(&cur_trans->dirty_bgs_lock);
3683
3684 again:
3685         /*
3686          * make sure all the block groups on our dirty list actually
3687          * exist
3688          */
3689         btrfs_create_pending_block_groups(trans);
3690
3691         if (!path) {
3692                 path = btrfs_alloc_path();
3693                 if (!path)
3694                         return -ENOMEM;
3695         }
3696
3697         /*
3698          * cache_write_mutex is here only to save us from balance or automatic
3699          * removal of empty block groups deleting this block group while we are
3700          * writing out the cache
3701          */
3702         mutex_lock(&trans->transaction->cache_write_mutex);
3703         while (!list_empty(&dirty)) {
3704                 cache = list_first_entry(&dirty,
3705                                          struct btrfs_block_group_cache,
3706                                          dirty_list);
3707                 /*
3708                  * this can happen if something re-dirties a block
3709                  * group that is already under IO.  Just wait for it to
3710                  * finish and then do it all again
3711                  */
3712                 if (!list_empty(&cache->io_list)) {
3713                         list_del_init(&cache->io_list);
3714                         btrfs_wait_cache_io(trans, cache, path);
3715                         btrfs_put_block_group(cache);
3716                 }
3717
3718
3719                 /*
3720                  * btrfs_wait_cache_io uses the cache->dirty_list to decide
3721                  * if it should update the cache_state.  Don't delete
3722                  * until after we wait.
3723                  *
3724                  * Since we're not running in the commit critical section
3725                  * we need the dirty_bgs_lock to protect from update_block_group
3726                  */
3727                 spin_lock(&cur_trans->dirty_bgs_lock);
3728                 list_del_init(&cache->dirty_list);
3729                 spin_unlock(&cur_trans->dirty_bgs_lock);
3730
3731                 should_put = 1;
3732
3733                 cache_save_setup(cache, trans, path);
3734
3735                 if (cache->disk_cache_state == BTRFS_DC_SETUP) {
3736                         cache->io_ctl.inode = NULL;
3737                         ret = btrfs_write_out_cache(fs_info, trans,
3738                                                     cache, path);
3739                         if (ret == 0 && cache->io_ctl.inode) {
3740                                 num_started++;
3741                                 should_put = 0;
3742
3743                                 /*
3744                                  * The cache_write_mutex is protecting the
3745                                  * io_list, also refer to the definition of
3746                                  * btrfs_transaction::io_bgs for more details
3747                                  */
3748                                 list_add_tail(&cache->io_list, io);
3749                         } else {
3750                                 /*
3751                                  * if we failed to write the cache, the
3752                                  * generation will be bad and life goes on
3753                                  */
3754                                 ret = 0;
3755                         }
3756                 }
3757                 if (!ret) {
3758                         ret = write_one_cache_group(trans, fs_info,
3759                                                     path, cache);
3760                         /*
3761                          * Our block group might still be attached to the list
3762                          * of new block groups in the transaction handle of some
3763                          * other task (struct btrfs_trans_handle->new_bgs). This
3764                          * means its block group item isn't yet in the extent
3765                          * tree. If this happens ignore the error, as we will
3766                          * try again later in the critical section of the
3767                          * transaction commit.
3768                          */
3769                         if (ret == -ENOENT) {
3770                                 ret = 0;
3771                                 spin_lock(&cur_trans->dirty_bgs_lock);
3772                                 if (list_empty(&cache->dirty_list)) {
3773                                         list_add_tail(&cache->dirty_list,
3774                                                       &cur_trans->dirty_bgs);
3775                                         btrfs_get_block_group(cache);
3776                                 }
3777                                 spin_unlock(&cur_trans->dirty_bgs_lock);
3778                         } else if (ret) {
3779                                 btrfs_abort_transaction(trans, ret);
3780                         }
3781                 }
3782
3783                 /* if its not on the io list, we need to put the block group */
3784                 if (should_put)
3785                         btrfs_put_block_group(cache);
3786
3787                 if (ret)
3788                         break;
3789
3790                 /*
3791                  * Avoid blocking other tasks for too long. It might even save
3792                  * us from writing caches for block groups that are going to be
3793                  * removed.
3794                  */
3795                 mutex_unlock(&trans->transaction->cache_write_mutex);
3796                 mutex_lock(&trans->transaction->cache_write_mutex);
3797         }
3798         mutex_unlock(&trans->transaction->cache_write_mutex);
3799
3800         /*
3801          * go through delayed refs for all the stuff we've just kicked off
3802          * and then loop back (just once)
3803          */
3804         ret = btrfs_run_delayed_refs(trans, 0);
3805         if (!ret && loops == 0) {
3806                 loops++;
3807                 spin_lock(&cur_trans->dirty_bgs_lock);
3808                 list_splice_init(&cur_trans->dirty_bgs, &dirty);
3809                 /*
3810                  * dirty_bgs_lock protects us from concurrent block group
3811                  * deletes too (not just cache_write_mutex).
3812                  */
3813                 if (!list_empty(&dirty)) {
3814                         spin_unlock(&cur_trans->dirty_bgs_lock);
3815                         goto again;
3816                 }
3817                 spin_unlock(&cur_trans->dirty_bgs_lock);
3818         } else if (ret < 0) {
3819                 btrfs_cleanup_dirty_bgs(cur_trans, fs_info);
3820         }
3821
3822         btrfs_free_path(path);
3823         return ret;
3824 }
3825
3826 int btrfs_write_dirty_block_groups(struct btrfs_trans_handle *trans,
3827                                    struct btrfs_fs_info *fs_info)
3828 {
3829         struct btrfs_block_group_cache *cache;
3830         struct btrfs_transaction *cur_trans = trans->transaction;
3831         int ret = 0;
3832         int should_put;
3833         struct btrfs_path *path;
3834         struct list_head *io = &cur_trans->io_bgs;
3835         int num_started = 0;
3836
3837         path = btrfs_alloc_path();
3838         if (!path)
3839                 return -ENOMEM;
3840
3841         /*
3842          * Even though we are in the critical section of the transaction commit,
3843          * we can still have concurrent tasks adding elements to this
3844          * transaction's list of dirty block groups. These tasks correspond to
3845          * endio free space workers started when writeback finishes for a
3846          * space cache, which run inode.c:btrfs_finish_ordered_io(), and can
3847          * allocate new block groups as a result of COWing nodes of the root
3848          * tree when updating the free space inode. The writeback for the space
3849          * caches is triggered by an earlier call to
3850          * btrfs_start_dirty_block_groups() and iterations of the following
3851          * loop.
3852          * Also we want to do the cache_save_setup first and then run the
3853          * delayed refs to make sure we have the best chance at doing this all
3854          * in one shot.
3855          */
3856         spin_lock(&cur_trans->dirty_bgs_lock);
3857         while (!list_empty(&cur_trans->dirty_bgs)) {
3858                 cache = list_first_entry(&cur_trans->dirty_bgs,
3859                                          struct btrfs_block_group_cache,
3860                                          dirty_list);
3861
3862                 /*
3863                  * this can happen if cache_save_setup re-dirties a block
3864                  * group that is already under IO.  Just wait for it to
3865                  * finish and then do it all again
3866                  */
3867                 if (!list_empty(&cache->io_list)) {
3868                         spin_unlock(&cur_trans->dirty_bgs_lock);
3869                         list_del_init(&cache->io_list);
3870                         btrfs_wait_cache_io(trans, cache, path);
3871                         btrfs_put_block_group(cache);
3872                         spin_lock(&cur_trans->dirty_bgs_lock);
3873                 }
3874
3875                 /*
3876                  * don't remove from the dirty list until after we've waited
3877                  * on any pending IO
3878                  */
3879                 list_del_init(&cache->dirty_list);
3880                 spin_unlock(&cur_trans->dirty_bgs_lock);
3881                 should_put = 1;
3882
3883                 cache_save_setup(cache, trans, path);
3884
3885                 if (!ret)
3886                         ret = btrfs_run_delayed_refs(trans,
3887                                                      (unsigned long) -1);
3888
3889                 if (!ret && cache->disk_cache_state == BTRFS_DC_SETUP) {
3890                         cache->io_ctl.inode = NULL;
3891                         ret = btrfs_write_out_cache(fs_info, trans,
3892                                                     cache, path);
3893                         if (ret == 0 && cache->io_ctl.inode) {
3894                                 num_started++;
3895                                 should_put = 0;
3896                                 list_add_tail(&cache->io_list, io);
3897                         } else {
3898                                 /*
3899                                  * if we failed to write the cache, the
3900                                  * generation will be bad and life goes on
3901                                  */
3902                                 ret = 0;
3903                         }
3904                 }
3905                 if (!ret) {
3906                         ret = write_one_cache_group(trans, fs_info,
3907                                                     path, cache);
3908                         /*
3909                          * One of the free space endio workers might have
3910                          * created a new block group while updating a free space
3911                          * cache's inode (at inode.c:btrfs_finish_ordered_io())
3912                          * and hasn't released its transaction handle yet, in
3913                          * which case the new block group is still attached to
3914                          * its transaction handle and its creation has not
3915                          * finished yet (no block group item in the extent tree
3916                          * yet, etc). If this is the case, wait for all free
3917                          * space endio workers to finish and retry. This is a
3918                          * a very rare case so no need for a more efficient and
3919                          * complex approach.
3920                          */
3921                         if (ret == -ENOENT) {
3922                                 wait_event(cur_trans->writer_wait,
3923                                    atomic_read(&cur_trans->num_writers) == 1);
3924                                 ret = write_one_cache_group(trans, fs_info,
3925                                                             path, cache);
3926                         }
3927                         if (ret)
3928                                 btrfs_abort_transaction(trans, ret);
3929                 }
3930
3931                 /* if its not on the io list, we need to put the block group */
3932                 if (should_put)
3933                         btrfs_put_block_group(cache);
3934                 spin_lock(&cur_trans->dirty_bgs_lock);
3935         }
3936         spin_unlock(&cur_trans->dirty_bgs_lock);
3937
3938         /*
3939          * Refer to the definition of io_bgs member for details why it's safe
3940          * to use it without any locking
3941          */
3942         while (!list_empty(io)) {
3943                 cache = list_first_entry(io, struct btrfs_block_group_cache,
3944                                          io_list);
3945                 list_del_init(&cache->io_list);
3946                 btrfs_wait_cache_io(trans, cache, path);
3947                 btrfs_put_block_group(cache);
3948         }
3949
3950         btrfs_free_path(path);
3951         return ret;
3952 }
3953
3954 int btrfs_extent_readonly(struct btrfs_fs_info *fs_info, u64 bytenr)
3955 {
3956         struct btrfs_block_group_cache *block_group;
3957         int readonly = 0;
3958
3959         block_group = btrfs_lookup_block_group(fs_info, bytenr);
3960         if (!block_group || block_group->ro)
3961                 readonly = 1;
3962         if (block_group)
3963                 btrfs_put_block_group(block_group);
3964         return readonly;
3965 }
3966
3967 bool btrfs_inc_nocow_writers(struct btrfs_fs_info *fs_info, u64 bytenr)
3968 {
3969         struct btrfs_block_group_cache *bg;
3970         bool ret = true;
3971
3972         bg = btrfs_lookup_block_group(fs_info, bytenr);
3973         if (!bg)
3974                 return false;
3975
3976         spin_lock(&bg->lock);
3977         if (bg->ro)
3978                 ret = false;
3979         else
3980                 atomic_inc(&bg->nocow_writers);
3981         spin_unlock(&bg->lock);
3982
3983         /* no put on block group, done by btrfs_dec_nocow_writers */
3984         if (!ret)
3985                 btrfs_put_block_group(bg);
3986
3987         return ret;
3988
3989 }
3990
3991 void btrfs_dec_nocow_writers(struct btrfs_fs_info *fs_info, u64 bytenr)
3992 {
3993         struct btrfs_block_group_cache *bg;
3994
3995         bg = btrfs_lookup_block_group(fs_info, bytenr);
3996         ASSERT(bg);
3997         if (atomic_dec_and_test(&bg->nocow_writers))
3998                 wake_up_atomic_t(&bg->nocow_writers);
3999         /*
4000          * Once for our lookup and once for the lookup done by a previous call
4001          * to btrfs_inc_nocow_writers()
4002          */
4003         btrfs_put_block_group(bg);
4004         btrfs_put_block_group(bg);
4005 }
4006
4007 void btrfs_wait_nocow_writers(struct btrfs_block_group_cache *bg)
4008 {
4009         wait_on_atomic_t(&bg->nocow_writers, atomic_t_wait,
4010                          TASK_UNINTERRUPTIBLE);
4011 }
4012
4013 static const char *alloc_name(u64 flags)
4014 {
4015         switch (flags) {
4016         case BTRFS_BLOCK_GROUP_METADATA|BTRFS_BLOCK_GROUP_DATA:
4017                 return "mixed";
4018         case BTRFS_BLOCK_GROUP_METADATA:
4019                 return "metadata";
4020         case BTRFS_BLOCK_GROUP_DATA:
4021                 return "data";
4022         case BTRFS_BLOCK_GROUP_SYSTEM:
4023                 return "system";
4024         default:
4025                 WARN_ON(1);
4026                 return "invalid-combination";
4027         };
4028 }
4029
4030 static int create_space_info(struct btrfs_fs_info *info, u64 flags,
4031                              struct btrfs_space_info **new)
4032 {
4033
4034         struct btrfs_space_info *space_info;
4035         int i;
4036         int ret;
4037
4038         space_info = kzalloc(sizeof(*space_info), GFP_NOFS);
4039         if (!space_info)
4040                 return -ENOMEM;
4041
4042         ret = percpu_counter_init(&space_info->total_bytes_pinned, 0,
4043                                  GFP_KERNEL);
4044         if (ret) {
4045                 kfree(space_info);
4046                 return ret;
4047         }
4048
4049         for (i = 0; i < BTRFS_NR_RAID_TYPES; i++)
4050                 INIT_LIST_HEAD(&space_info->block_groups[i]);
4051         init_rwsem(&space_info->groups_sem);
4052         spin_lock_init(&space_info->lock);
4053         space_info->flags = flags & BTRFS_BLOCK_GROUP_TYPE_MASK;
4054         space_info->force_alloc = CHUNK_ALLOC_NO_FORCE;
4055         init_waitqueue_head(&space_info->wait);
4056         INIT_LIST_HEAD(&space_info->ro_bgs);
4057         INIT_LIST_HEAD(&space_info->tickets);
4058         INIT_LIST_HEAD(&space_info->priority_tickets);
4059
4060         ret = kobject_init_and_add(&space_info->kobj, &space_info_ktype,
4061                                     info->space_info_kobj, "%s",
4062                                     alloc_name(space_info->flags));
4063         if (ret) {
4064                 percpu_counter_destroy(&space_info->total_bytes_pinned);
4065                 kfree(space_info);
4066                 return ret;
4067         }
4068
4069         *new = space_info;
4070         list_add_rcu(&space_info->list, &info->space_info);
4071         if (flags & BTRFS_BLOCK_GROUP_DATA)
4072                 info->data_sinfo = space_info;
4073
4074         return ret;
4075 }
4076
4077 static void update_space_info(struct btrfs_fs_info *info, u64 flags,
4078                              u64 total_bytes, u64 bytes_used,
4079                              u64 bytes_readonly,
4080                              struct btrfs_space_info **space_info)
4081 {
4082         struct btrfs_space_info *found;
4083         int factor;
4084
4085         if (flags & (BTRFS_BLOCK_GROUP_DUP | BTRFS_BLOCK_GROUP_RAID1 |
4086                      BTRFS_BLOCK_GROUP_RAID10))
4087                 factor = 2;
4088         else
4089                 factor = 1;
4090
4091         found = __find_space_info(info, flags);
4092         ASSERT(found);
4093         spin_lock(&found->lock);
4094         found->total_bytes += total_bytes;
4095         found->disk_total += total_bytes * factor;
4096         found->bytes_used += bytes_used;
4097         found->disk_used += bytes_used * factor;
4098         found->bytes_readonly += bytes_readonly;
4099         if (total_bytes > 0)
4100                 found->full = 0;
4101         space_info_add_new_bytes(info, found, total_bytes -
4102                                  bytes_used - bytes_readonly);
4103         spin_unlock(&found->lock);
4104         *space_info = found;
4105 }
4106
4107 static void set_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
4108 {
4109         u64 extra_flags = chunk_to_extended(flags) &
4110                                 BTRFS_EXTENDED_PROFILE_MASK;
4111
4112         write_seqlock(&fs_info->profiles_lock);
4113         if (flags & BTRFS_BLOCK_GROUP_DATA)
4114                 fs_info->avail_data_alloc_bits |= extra_flags;
4115         if (flags & BTRFS_BLOCK_GROUP_METADATA)
4116                 fs_info->avail_metadata_alloc_bits |= extra_flags;
4117         if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
4118                 fs_info->avail_system_alloc_bits |= extra_flags;
4119         write_sequnlock(&fs_info->profiles_lock);
4120 }
4121
4122 /*
4123  * returns target flags in extended format or 0 if restripe for this
4124  * chunk_type is not in progress
4125  *
4126  * should be called with either volume_mutex or balance_lock held
4127  */
4128 static u64 get_restripe_target(struct btrfs_fs_info *fs_info, u64 flags)
4129 {
4130         struct btrfs_balance_control *bctl = fs_info->balance_ctl;
4131         u64 target = 0;
4132
4133         if (!bctl)
4134                 return 0;
4135
4136         if (flags & BTRFS_BLOCK_GROUP_DATA &&
4137             bctl->data.flags & BTRFS_BALANCE_ARGS_CONVERT) {
4138                 target = BTRFS_BLOCK_GROUP_DATA | bctl->data.target;
4139         } else if (flags & BTRFS_BLOCK_GROUP_SYSTEM &&
4140                    bctl->sys.flags & BTRFS_BALANCE_ARGS_CONVERT) {
4141                 target = BTRFS_BLOCK_GROUP_SYSTEM | bctl->sys.target;
4142         } else if (flags & BTRFS_BLOCK_GROUP_METADATA &&
4143                    bctl->meta.flags & BTRFS_BALANCE_ARGS_CONVERT) {
4144                 target = BTRFS_BLOCK_GROUP_METADATA | bctl->meta.target;
4145         }
4146
4147         return target;
4148 }
4149
4150 /*
4151  * @flags: available profiles in extended format (see ctree.h)
4152  *
4153  * Returns reduced profile in chunk format.  If profile changing is in
4154  * progress (either running or paused) picks the target profile (if it's
4155  * already available), otherwise falls back to plain reducing.
4156  */
4157 static u64 btrfs_reduce_alloc_profile(struct btrfs_fs_info *fs_info, u64 flags)
4158 {
4159         u64 num_devices = fs_info->fs_devices->rw_devices;
4160         u64 target;
4161         u64 raid_type;
4162         u64 allowed = 0;
4163
4164         /*
4165          * see if restripe for this chunk_type is in progress, if so
4166          * try to reduce to the target profile
4167          */
4168         spin_lock(&fs_info->balance_lock);
4169         target = get_restripe_target(fs_info, flags);
4170         if (target) {
4171                 /* pick target profile only if it's already available */
4172                 if ((flags & target) & BTRFS_EXTENDED_PROFILE_MASK) {
4173                         spin_unlock(&fs_info->balance_lock);
4174                         return extended_to_chunk(target);
4175                 }
4176         }
4177         spin_unlock(&fs_info->balance_lock);
4178
4179         /* First, mask out the RAID levels which aren't possible */
4180         for (raid_type = 0; raid_type < BTRFS_NR_RAID_TYPES; raid_type++) {
4181                 if (num_devices >= btrfs_raid_array[raid_type].devs_min)
4182                         allowed |= btrfs_raid_group[raid_type];
4183         }
4184         allowed &= flags;
4185
4186         if (allowed & BTRFS_BLOCK_GROUP_RAID6)
4187                 allowed = BTRFS_BLOCK_GROUP_RAID6;
4188         else if (allowed & BTRFS_BLOCK_GROUP_RAID5)
4189                 allowed = BTRFS_BLOCK_GROUP_RAID5;
4190         else if (allowed & BTRFS_BLOCK_GROUP_RAID10)
4191                 allowed = BTRFS_BLOCK_GROUP_RAID10;
4192         else if (allowed & BTRFS_BLOCK_GROUP_RAID1)
4193                 allowed = BTRFS_BLOCK_GROUP_RAID1;
4194         else if (allowed & BTRFS_BLOCK_GROUP_RAID0)
4195                 allowed = BTRFS_BLOCK_GROUP_RAID0;
4196
4197         flags &= ~BTRFS_BLOCK_GROUP_PROFILE_MASK;
4198
4199         return extended_to_chunk(flags | allowed);
4200 }
4201
4202 static u64 get_alloc_profile(struct btrfs_fs_info *fs_info, u64 orig_flags)
4203 {
4204         unsigned seq;
4205         u64 flags;
4206
4207         do {
4208                 flags = orig_flags;
4209                 seq = read_seqbegin(&fs_info->profiles_lock);
4210
4211                 if (flags & BTRFS_BLOCK_GROUP_DATA)
4212                         flags |= fs_info->avail_data_alloc_bits;
4213                 else if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
4214                         flags |= fs_info->avail_system_alloc_bits;
4215                 else if (flags & BTRFS_BLOCK_GROUP_METADATA)
4216                         flags |= fs_info->avail_metadata_alloc_bits;
4217         } while (read_seqretry(&fs_info->profiles_lock, seq));
4218
4219         return btrfs_reduce_alloc_profile(fs_info, flags);
4220 }
4221
4222 static u64 get_alloc_profile_by_root(struct btrfs_root *root, int data)
4223 {
4224         struct btrfs_fs_info *fs_info = root->fs_info;
4225         u64 flags;
4226         u64 ret;
4227
4228         if (data)
4229                 flags = BTRFS_BLOCK_GROUP_DATA;
4230         else if (root == fs_info->chunk_root)
4231                 flags = BTRFS_BLOCK_GROUP_SYSTEM;
4232         else
4233                 flags = BTRFS_BLOCK_GROUP_METADATA;
4234
4235         ret = get_alloc_profile(fs_info, flags);
4236         return ret;
4237 }
4238
4239 u64 btrfs_data_alloc_profile(struct btrfs_fs_info *fs_info)
4240 {
4241         return get_alloc_profile(fs_info, BTRFS_BLOCK_GROUP_DATA);
4242 }
4243
4244 u64 btrfs_metadata_alloc_profile(struct btrfs_fs_info *fs_info)
4245 {
4246         return get_alloc_profile(fs_info, BTRFS_BLOCK_GROUP_METADATA);
4247 }
4248
4249 u64 btrfs_system_alloc_profile(struct btrfs_fs_info *fs_info)
4250 {
4251         return get_alloc_profile(fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
4252 }
4253
4254 static u64 btrfs_space_info_used(struct btrfs_space_info *s_info,
4255                                  bool may_use_included)
4256 {
4257         ASSERT(s_info);
4258         return s_info->bytes_used + s_info->bytes_reserved +
4259                 s_info->bytes_pinned + s_info->bytes_readonly +
4260                 (may_use_included ? s_info->bytes_may_use : 0);
4261 }
4262
4263 int btrfs_alloc_data_chunk_ondemand(struct btrfs_inode *inode, u64 bytes)
4264 {
4265         struct btrfs_root *root = inode->root;
4266         struct btrfs_fs_info *fs_info = root->fs_info;
4267         struct btrfs_space_info *data_sinfo = fs_info->data_sinfo;
4268         u64 used;
4269         int ret = 0;
4270         int need_commit = 2;
4271         int have_pinned_space;
4272
4273         /* make sure bytes are sectorsize aligned */
4274         bytes = ALIGN(bytes, fs_info->sectorsize);
4275
4276         if (btrfs_is_free_space_inode(inode)) {
4277                 need_commit = 0;
4278                 ASSERT(current->journal_info);
4279         }
4280
4281 again:
4282         /* make sure we have enough space to handle the data first */
4283         spin_lock(&data_sinfo->lock);
4284         used = btrfs_space_info_used(data_sinfo, true);
4285
4286         if (used + bytes > data_sinfo->total_bytes) {
4287                 struct btrfs_trans_handle *trans;
4288
4289                 /*
4290                  * if we don't have enough free bytes in this space then we need
4291                  * to alloc a new chunk.
4292                  */
4293                 if (!data_sinfo->full) {
4294                         u64 alloc_target;
4295
4296                         data_sinfo->force_alloc = CHUNK_ALLOC_FORCE;
4297                         spin_unlock(&data_sinfo->lock);
4298
4299                         alloc_target = btrfs_data_alloc_profile(fs_info);
4300                         /*
4301                          * It is ugly that we don't call nolock join
4302                          * transaction for the free space inode case here.
4303                          * But it is safe because we only do the data space
4304                          * reservation for the free space cache in the
4305                          * transaction context, the common join transaction
4306                          * just increase the counter of the current transaction
4307                          * handler, doesn't try to acquire the trans_lock of
4308                          * the fs.
4309                          */
4310                         trans = btrfs_join_transaction(root);
4311                         if (IS_ERR(trans))
4312                                 return PTR_ERR(trans);
4313
4314                         ret = do_chunk_alloc(trans, fs_info, alloc_target,
4315                                              CHUNK_ALLOC_NO_FORCE);
4316                         btrfs_end_transaction(trans);
4317                         if (ret < 0) {
4318                                 if (ret != -ENOSPC)
4319                                         return ret;
4320                                 else {
4321                                         have_pinned_space = 1;
4322                                         goto commit_trans;
4323                                 }
4324                         }
4325
4326                         goto again;
4327                 }
4328
4329                 /*
4330                  * If we don't have enough pinned space to deal with this
4331                  * allocation, and no removed chunk in current transaction,
4332                  * don't bother committing the transaction.
4333                  */
4334                 have_pinned_space = percpu_counter_compare(
4335                         &data_sinfo->total_bytes_pinned,
4336                         used + bytes - data_sinfo->total_bytes);
4337                 spin_unlock(&data_sinfo->lock);
4338
4339                 /* commit the current transaction and try again */
4340 commit_trans:
4341                 if (need_commit) {
4342                         need_commit--;
4343
4344                         if (need_commit > 0) {
4345                                 btrfs_start_delalloc_roots(fs_info, 0, -1);
4346                                 btrfs_wait_ordered_roots(fs_info, U64_MAX, 0,
4347                                                          (u64)-1);
4348                         }
4349
4350                         trans = btrfs_join_transaction(root);
4351                         if (IS_ERR(trans))
4352                                 return PTR_ERR(trans);
4353                         if (have_pinned_space >= 0 ||
4354                             test_bit(BTRFS_TRANS_HAVE_FREE_BGS,
4355                                      &trans->transaction->flags) ||
4356                             need_commit > 0) {
4357                                 ret = btrfs_commit_transaction(trans);
4358                                 if (ret)
4359                                         return ret;
4360                                 /*
4361                                  * The cleaner kthread might still be doing iput
4362                                  * operations. Wait for it to finish so that
4363                                  * more space is released.
4364                                  */
4365                                 mutex_lock(&fs_info->cleaner_delayed_iput_mutex);
4366                                 mutex_unlock(&fs_info->cleaner_delayed_iput_mutex);
4367                                 goto again;
4368                         } else {
4369                                 btrfs_end_transaction(trans);
4370                         }
4371                 }
4372
4373                 trace_btrfs_space_reservation(fs_info,
4374                                               "space_info:enospc",
4375                                               data_sinfo->flags, bytes, 1);
4376                 return -ENOSPC;
4377         }
4378         data_sinfo->bytes_may_use += bytes;
4379         trace_btrfs_space_reservation(fs_info, "space_info",
4380                                       data_sinfo->flags, bytes, 1);
4381         spin_unlock(&data_sinfo->lock);
4382
4383         return ret;
4384 }
4385
4386 int btrfs_check_data_free_space(struct inode *inode,
4387                         struct extent_changeset **reserved, u64 start, u64 len)
4388 {
4389         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
4390         int ret;
4391
4392         /* align the range */
4393         len = round_up(start + len, fs_info->sectorsize) -
4394               round_down(start, fs_info->sectorsize);
4395         start = round_down(start, fs_info->sectorsize);
4396
4397         ret = btrfs_alloc_data_chunk_ondemand(BTRFS_I(inode), len);
4398         if (ret < 0)
4399                 return ret;
4400
4401         /* Use new btrfs_qgroup_reserve_data to reserve precious data space. */
4402         ret = btrfs_qgroup_reserve_data(inode, reserved, start, len);
4403         if (ret < 0)
4404                 btrfs_free_reserved_data_space_noquota(inode, start, len);
4405         else
4406                 ret = 0;
4407         return ret;
4408 }
4409
4410 /*
4411  * Called if we need to clear a data reservation for this inode
4412  * Normally in a error case.
4413  *
4414  * This one will *NOT* use accurate qgroup reserved space API, just for case
4415  * which we can't sleep and is sure it won't affect qgroup reserved space.
4416  * Like clear_bit_hook().
4417  */
4418 void btrfs_free_reserved_data_space_noquota(struct inode *inode, u64 start,
4419                                             u64 len)
4420 {
4421         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
4422         struct btrfs_space_info *data_sinfo;
4423
4424         /* Make sure the range is aligned to sectorsize */
4425         len = round_up(start + len, fs_info->sectorsize) -
4426               round_down(start, fs_info->sectorsize);
4427         start = round_down(start, fs_info->sectorsize);
4428
4429         data_sinfo = fs_info->data_sinfo;
4430         spin_lock(&data_sinfo->lock);
4431         if (WARN_ON(data_sinfo->bytes_may_use < len))
4432                 data_sinfo->bytes_may_use = 0;
4433         else
4434                 data_sinfo->bytes_may_use -= len;
4435         trace_btrfs_space_reservation(fs_info, "space_info",
4436                                       data_sinfo->flags, len, 0);
4437         spin_unlock(&data_sinfo->lock);
4438 }
4439
4440 /*
4441  * Called if we need to clear a data reservation for this inode
4442  * Normally in a error case.
4443  *
4444  * This one will handle the per-inode data rsv map for accurate reserved
4445  * space framework.
4446  */
4447 void btrfs_free_reserved_data_space(struct inode *inode,
4448                         struct extent_changeset *reserved, u64 start, u64 len)
4449 {
4450         struct btrfs_root *root = BTRFS_I(inode)->root;
4451
4452         /* Make sure the range is aligned to sectorsize */
4453         len = round_up(start + len, root->fs_info->sectorsize) -
4454               round_down(start, root->fs_info->sectorsize);
4455         start = round_down(start, root->fs_info->sectorsize);
4456
4457         btrfs_free_reserved_data_space_noquota(inode, start, len);
4458         btrfs_qgroup_free_data(inode, reserved, start, len);
4459 }
4460
4461 static void force_metadata_allocation(struct btrfs_fs_info *info)
4462 {
4463         struct list_head *head = &info->space_info;
4464         struct btrfs_space_info *found;
4465
4466         rcu_read_lock();
4467         list_for_each_entry_rcu(found, head, list) {
4468                 if (found->flags & BTRFS_BLOCK_GROUP_METADATA)
4469                         found->force_alloc = CHUNK_ALLOC_FORCE;
4470         }
4471         rcu_read_unlock();
4472 }
4473
4474 static inline u64 calc_global_rsv_need_space(struct btrfs_block_rsv *global)
4475 {
4476         return (global->size << 1);
4477 }
4478
4479 static int should_alloc_chunk(struct btrfs_fs_info *fs_info,
4480                               struct btrfs_space_info *sinfo, int force)
4481 {
4482         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
4483         u64 bytes_used = btrfs_space_info_used(sinfo, false);
4484         u64 thresh;
4485
4486         if (force == CHUNK_ALLOC_FORCE)
4487                 return 1;
4488
4489         /*
4490          * We need to take into account the global rsv because for all intents
4491          * and purposes it's used space.  Don't worry about locking the
4492          * global_rsv, it doesn't change except when the transaction commits.
4493          */
4494         if (sinfo->flags & BTRFS_BLOCK_GROUP_METADATA)
4495                 bytes_used += calc_global_rsv_need_space(global_rsv);
4496
4497         /*
4498          * in limited mode, we want to have some free space up to
4499          * about 1% of the FS size.
4500          */
4501         if (force == CHUNK_ALLOC_LIMITED) {
4502                 thresh = btrfs_super_total_bytes(fs_info->super_copy);
4503                 thresh = max_t(u64, SZ_64M, div_factor_fine(thresh, 1));
4504
4505                 if (sinfo->total_bytes - bytes_used < thresh)
4506                         return 1;
4507         }
4508
4509         if (bytes_used + SZ_2M < div_factor(sinfo->total_bytes, 8))
4510                 return 0;
4511         return 1;
4512 }
4513
4514 static u64 get_profile_num_devs(struct btrfs_fs_info *fs_info, u64 type)
4515 {
4516         u64 num_dev;
4517
4518         if (type & (BTRFS_BLOCK_GROUP_RAID10 |
4519                     BTRFS_BLOCK_GROUP_RAID0 |
4520                     BTRFS_BLOCK_GROUP_RAID5 |
4521                     BTRFS_BLOCK_GROUP_RAID6))
4522                 num_dev = fs_info->fs_devices->rw_devices;
4523         else if (type & BTRFS_BLOCK_GROUP_RAID1)
4524                 num_dev = 2;
4525         else
4526                 num_dev = 1;    /* DUP or single */
4527
4528         return num_dev;
4529 }
4530
4531 /*
4532  * If @is_allocation is true, reserve space in the system space info necessary
4533  * for allocating a chunk, otherwise if it's false, reserve space necessary for
4534  * removing a chunk.
4535  */
4536 void check_system_chunk(struct btrfs_trans_handle *trans,
4537                         struct btrfs_fs_info *fs_info, u64 type)
4538 {
4539         struct btrfs_space_info *info;
4540         u64 left;
4541         u64 thresh;
4542         int ret = 0;
4543         u64 num_devs;
4544
4545         /*
4546          * Needed because we can end up allocating a system chunk and for an
4547          * atomic and race free space reservation in the chunk block reserve.
4548          */
4549         lockdep_assert_held(&fs_info->chunk_mutex);
4550
4551         info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
4552         spin_lock(&info->lock);
4553         left = info->total_bytes - btrfs_space_info_used(info, true);
4554         spin_unlock(&info->lock);
4555
4556         num_devs = get_profile_num_devs(fs_info, type);
4557
4558         /* num_devs device items to update and 1 chunk item to add or remove */
4559         thresh = btrfs_calc_trunc_metadata_size(fs_info, num_devs) +
4560                 btrfs_calc_trans_metadata_size(fs_info, 1);
4561
4562         if (left < thresh && btrfs_test_opt(fs_info, ENOSPC_DEBUG)) {
4563                 btrfs_info(fs_info, "left=%llu, need=%llu, flags=%llu",
4564                            left, thresh, type);
4565                 dump_space_info(fs_info, info, 0, 0);
4566         }
4567
4568         if (left < thresh) {
4569                 u64 flags = btrfs_system_alloc_profile(fs_info);
4570
4571                 /*
4572                  * Ignore failure to create system chunk. We might end up not
4573                  * needing it, as we might not need to COW all nodes/leafs from
4574                  * the paths we visit in the chunk tree (they were already COWed
4575                  * or created in the current transaction for example).
4576                  */
4577                 ret = btrfs_alloc_chunk(trans, fs_info, flags);
4578         }
4579
4580         if (!ret) {
4581                 ret = btrfs_block_rsv_add(fs_info->chunk_root,
4582                                           &fs_info->chunk_block_rsv,
4583                                           thresh, BTRFS_RESERVE_NO_FLUSH);
4584                 if (!ret)
4585                         trans->chunk_bytes_reserved += thresh;
4586         }
4587 }
4588
4589 /*
4590  * If force is CHUNK_ALLOC_FORCE:
4591  *    - return 1 if it successfully allocates a chunk,
4592  *    - return errors including -ENOSPC otherwise.
4593  * If force is NOT CHUNK_ALLOC_FORCE:
4594  *    - return 0 if it doesn't need to allocate a new chunk,
4595  *    - return 1 if it successfully allocates a chunk,
4596  *    - return errors including -ENOSPC otherwise.
4597  */
4598 static int do_chunk_alloc(struct btrfs_trans_handle *trans,
4599                           struct btrfs_fs_info *fs_info, u64 flags, int force)
4600 {
4601         struct btrfs_space_info *space_info;
4602         int wait_for_alloc = 0;
4603         int ret = 0;
4604
4605         /* Don't re-enter if we're already allocating a chunk */
4606         if (trans->allocating_chunk)
4607                 return -ENOSPC;
4608
4609         space_info = __find_space_info(fs_info, flags);
4610         ASSERT(space_info);
4611
4612 again:
4613         spin_lock(&space_info->lock);
4614         if (force < space_info->force_alloc)
4615                 force = space_info->force_alloc;
4616         if (space_info->full) {
4617                 if (should_alloc_chunk(fs_info, space_info, force))
4618                         ret = -ENOSPC;
4619                 else
4620                         ret = 0;
4621                 spin_unlock(&space_info->lock);
4622                 return ret;
4623         }
4624
4625         if (!should_alloc_chunk(fs_info, space_info, force)) {
4626                 spin_unlock(&space_info->lock);
4627                 return 0;
4628         } else if (space_info->chunk_alloc) {
4629                 wait_for_alloc = 1;
4630         } else {
4631                 space_info->chunk_alloc = 1;
4632         }
4633
4634         spin_unlock(&space_info->lock);
4635
4636         mutex_lock(&fs_info->chunk_mutex);
4637
4638         /*
4639          * The chunk_mutex is held throughout the entirety of a chunk
4640          * allocation, so once we've acquired the chunk_mutex we know that the
4641          * other guy is done and we need to recheck and see if we should
4642          * allocate.
4643          */
4644         if (wait_for_alloc) {
4645                 mutex_unlock(&fs_info->chunk_mutex);
4646                 wait_for_alloc = 0;
4647                 cond_resched();
4648                 goto again;
4649         }
4650
4651         trans->allocating_chunk = true;
4652
4653         /*
4654          * If we have mixed data/metadata chunks we want to make sure we keep
4655          * allocating mixed chunks instead of individual chunks.
4656          */
4657         if (btrfs_mixed_space_info(space_info))
4658                 flags |= (BTRFS_BLOCK_GROUP_DATA | BTRFS_BLOCK_GROUP_METADATA);
4659
4660         /*
4661          * if we're doing a data chunk, go ahead and make sure that
4662          * we keep a reasonable number of metadata chunks allocated in the
4663          * FS as well.
4664          */
4665         if (flags & BTRFS_BLOCK_GROUP_DATA && fs_info->metadata_ratio) {
4666                 fs_info->data_chunk_allocations++;
4667                 if (!(fs_info->data_chunk_allocations %
4668                       fs_info->metadata_ratio))
4669                         force_metadata_allocation(fs_info);
4670         }
4671
4672         /*
4673          * Check if we have enough space in SYSTEM chunk because we may need
4674          * to update devices.
4675          */
4676         check_system_chunk(trans, fs_info, flags);
4677
4678         ret = btrfs_alloc_chunk(trans, fs_info, flags);
4679         trans->allocating_chunk = false;
4680
4681         spin_lock(&space_info->lock);
4682         if (ret < 0 && ret != -ENOSPC)
4683                 goto out;
4684         if (ret)
4685                 space_info->full = 1;
4686         else
4687                 ret = 1;
4688
4689         space_info->force_alloc = CHUNK_ALLOC_NO_FORCE;
4690 out:
4691         space_info->chunk_alloc = 0;
4692         spin_unlock(&space_info->lock);
4693         mutex_unlock(&fs_info->chunk_mutex);
4694         /*
4695          * When we allocate a new chunk we reserve space in the chunk block
4696          * reserve to make sure we can COW nodes/leafs in the chunk tree or
4697          * add new nodes/leafs to it if we end up needing to do it when
4698          * inserting the chunk item and updating device items as part of the
4699          * second phase of chunk allocation, performed by
4700          * btrfs_finish_chunk_alloc(). So make sure we don't accumulate a
4701          * large number of new block groups to create in our transaction
4702          * handle's new_bgs list to avoid exhausting the chunk block reserve
4703          * in extreme cases - like having a single transaction create many new
4704          * block groups when starting to write out the free space caches of all
4705          * the block groups that were made dirty during the lifetime of the
4706          * transaction.
4707          */
4708         if (trans->can_flush_pending_bgs &&
4709             trans->chunk_bytes_reserved >= (u64)SZ_2M) {
4710                 btrfs_create_pending_block_groups(trans);
4711                 btrfs_trans_release_chunk_metadata(trans);
4712         }
4713         return ret;
4714 }
4715
4716 static int can_overcommit(struct btrfs_fs_info *fs_info,
4717                           struct btrfs_space_info *space_info, u64 bytes,
4718                           enum btrfs_reserve_flush_enum flush,
4719                           bool system_chunk)
4720 {
4721         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
4722         u64 profile;
4723         u64 space_size;
4724         u64 avail;
4725         u64 used;
4726
4727         /* Don't overcommit when in mixed mode. */
4728         if (space_info->flags & BTRFS_BLOCK_GROUP_DATA)
4729                 return 0;
4730
4731         if (system_chunk)
4732                 profile = btrfs_system_alloc_profile(fs_info);
4733         else
4734                 profile = btrfs_metadata_alloc_profile(fs_info);
4735
4736         used = btrfs_space_info_used(space_info, false);
4737
4738         /*
4739          * We only want to allow over committing if we have lots of actual space
4740          * free, but if we don't have enough space to handle the global reserve
4741          * space then we could end up having a real enospc problem when trying
4742          * to allocate a chunk or some other such important allocation.
4743          */
4744         spin_lock(&global_rsv->lock);
4745         space_size = calc_global_rsv_need_space(global_rsv);
4746         spin_unlock(&global_rsv->lock);
4747         if (used + space_size >= space_info->total_bytes)
4748                 return 0;
4749
4750         used += space_info->bytes_may_use;
4751
4752         avail = atomic64_read(&fs_info->free_chunk_space);
4753
4754         /*
4755          * If we have dup, raid1 or raid10 then only half of the free
4756          * space is actually useable.  For raid56, the space info used
4757          * doesn't include the parity drive, so we don't have to
4758          * change the math
4759          */
4760         if (profile & (BTRFS_BLOCK_GROUP_DUP |
4761                        BTRFS_BLOCK_GROUP_RAID1 |
4762                        BTRFS_BLOCK_GROUP_RAID10))
4763                 avail >>= 1;
4764
4765         /*
4766          * If we aren't flushing all things, let us overcommit up to
4767          * 1/2th of the space. If we can flush, don't let us overcommit
4768          * too much, let it overcommit up to 1/8 of the space.
4769          */
4770         if (flush == BTRFS_RESERVE_FLUSH_ALL)
4771                 avail >>= 3;
4772         else
4773                 avail >>= 1;
4774
4775         if (used + bytes < space_info->total_bytes + avail)
4776                 return 1;
4777         return 0;
4778 }
4779
4780 static void btrfs_writeback_inodes_sb_nr(struct btrfs_fs_info *fs_info,
4781                                          unsigned long nr_pages, int nr_items)
4782 {
4783         struct super_block *sb = fs_info->sb;
4784
4785         if (down_read_trylock(&sb->s_umount)) {
4786                 writeback_inodes_sb_nr(sb, nr_pages, WB_REASON_FS_FREE_SPACE);
4787                 up_read(&sb->s_umount);
4788         } else {
4789                 /*
4790                  * We needn't worry the filesystem going from r/w to r/o though
4791                  * we don't acquire ->s_umount mutex, because the filesystem
4792                  * should guarantee the delalloc inodes list be empty after
4793                  * the filesystem is readonly(all dirty pages are written to
4794                  * the disk).
4795                  */
4796                 btrfs_start_delalloc_roots(fs_info, 0, nr_items);
4797                 if (!current->journal_info)
4798                         btrfs_wait_ordered_roots(fs_info, nr_items, 0, (u64)-1);
4799         }
4800 }
4801
4802 static inline u64 calc_reclaim_items_nr(struct btrfs_fs_info *fs_info,
4803                                         u64 to_reclaim)
4804 {
4805         u64 bytes;
4806         u64 nr;
4807
4808         bytes = btrfs_calc_trans_metadata_size(fs_info, 1);
4809         nr = div64_u64(to_reclaim, bytes);
4810         if (!nr)
4811                 nr = 1;
4812         return nr;
4813 }
4814
4815 #define EXTENT_SIZE_PER_ITEM    SZ_256K
4816
4817 /*
4818  * shrink metadata reservation for delalloc
4819  */
4820 static void shrink_delalloc(struct btrfs_fs_info *fs_info, u64 to_reclaim,
4821                             u64 orig, bool wait_ordered)
4822 {
4823         struct btrfs_space_info *space_info;
4824         struct btrfs_trans_handle *trans;
4825         u64 delalloc_bytes;
4826         u64 max_reclaim;
4827         u64 items;
4828         long time_left;
4829         unsigned long nr_pages;
4830         int loops;
4831
4832         /* Calc the number of the pages we need flush for space reservation */
4833         items = calc_reclaim_items_nr(fs_info, to_reclaim);
4834         to_reclaim = items * EXTENT_SIZE_PER_ITEM;
4835
4836         trans = (struct btrfs_trans_handle *)current->journal_info;
4837         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
4838
4839         delalloc_bytes = percpu_counter_sum_positive(
4840                                                 &fs_info->delalloc_bytes);
4841         if (delalloc_bytes == 0) {
4842                 if (trans)
4843                         return;
4844                 if (wait_ordered)
4845                         btrfs_wait_ordered_roots(fs_info, items, 0, (u64)-1);
4846                 return;
4847         }
4848
4849         loops = 0;
4850         while (delalloc_bytes && loops < 3) {
4851                 max_reclaim = min(delalloc_bytes, to_reclaim);
4852                 nr_pages = max_reclaim >> PAGE_SHIFT;
4853                 btrfs_writeback_inodes_sb_nr(fs_info, nr_pages, items);
4854                 /*
4855                  * We need to wait for the async pages to actually start before
4856                  * we do anything.
4857                  */
4858                 max_reclaim = atomic_read(&fs_info->async_delalloc_pages);
4859                 if (!max_reclaim)
4860                         goto skip_async;
4861
4862                 if (max_reclaim <= nr_pages)
4863                         max_reclaim = 0;
4864                 else
4865                         max_reclaim -= nr_pages;
4866
4867                 wait_event(fs_info->async_submit_wait,
4868                            atomic_read(&fs_info->async_delalloc_pages) <=
4869                            (int)max_reclaim);
4870 skip_async:
4871                 spin_lock(&space_info->lock);
4872                 if (list_empty(&space_info->tickets) &&
4873                     list_empty(&space_info->priority_tickets)) {
4874                         spin_unlock(&space_info->lock);
4875                         break;
4876                 }
4877                 spin_unlock(&space_info->lock);
4878
4879                 loops++;
4880                 if (wait_ordered && !trans) {
4881                         btrfs_wait_ordered_roots(fs_info, items, 0, (u64)-1);
4882                 } else {
4883                         time_left = schedule_timeout_killable(1);
4884                         if (time_left)
4885                                 break;
4886                 }
4887                 delalloc_bytes = percpu_counter_sum_positive(
4888                                                 &fs_info->delalloc_bytes);
4889         }
4890 }
4891
4892 struct reserve_ticket {
4893         u64 bytes;
4894         int error;
4895         struct list_head list;
4896         wait_queue_head_t wait;
4897 };
4898
4899 /**
4900  * maybe_commit_transaction - possibly commit the transaction if its ok to
4901  * @root - the root we're allocating for
4902  * @bytes - the number of bytes we want to reserve
4903  * @force - force the commit
4904  *
4905  * This will check to make sure that committing the transaction will actually
4906  * get us somewhere and then commit the transaction if it does.  Otherwise it
4907  * will return -ENOSPC.
4908  */
4909 static int may_commit_transaction(struct btrfs_fs_info *fs_info,
4910                                   struct btrfs_space_info *space_info)
4911 {
4912         struct reserve_ticket *ticket = NULL;
4913         struct btrfs_block_rsv *delayed_rsv = &fs_info->delayed_block_rsv;
4914         struct btrfs_trans_handle *trans;
4915         u64 bytes;
4916
4917         trans = (struct btrfs_trans_handle *)current->journal_info;
4918         if (trans)
4919                 return -EAGAIN;
4920
4921         spin_lock(&space_info->lock);
4922         if (!list_empty(&space_info->priority_tickets))
4923                 ticket = list_first_entry(&space_info->priority_tickets,
4924                                           struct reserve_ticket, list);
4925         else if (!list_empty(&space_info->tickets))
4926                 ticket = list_first_entry(&space_info->tickets,
4927                                           struct reserve_ticket, list);
4928         bytes = (ticket) ? ticket->bytes : 0;
4929         spin_unlock(&space_info->lock);
4930
4931         if (!bytes)
4932                 return 0;
4933
4934         /* See if there is enough pinned space to make this reservation */
4935         if (percpu_counter_compare(&space_info->total_bytes_pinned,
4936                                    bytes) >= 0)
4937                 goto commit;
4938
4939         /*
4940          * See if there is some space in the delayed insertion reservation for
4941          * this reservation.
4942          */
4943         if (space_info != delayed_rsv->space_info)
4944                 return -ENOSPC;
4945
4946         spin_lock(&delayed_rsv->lock);
4947         if (delayed_rsv->size > bytes)
4948                 bytes = 0;
4949         else
4950                 bytes -= delayed_rsv->size;
4951         spin_unlock(&delayed_rsv->lock);
4952
4953         if (percpu_counter_compare(&space_info->total_bytes_pinned,
4954                                    bytes) < 0) {
4955                 return -ENOSPC;
4956         }
4957
4958 commit:
4959         trans = btrfs_join_transaction(fs_info->extent_root);
4960         if (IS_ERR(trans))
4961                 return -ENOSPC;
4962
4963         return btrfs_commit_transaction(trans);
4964 }
4965
4966 /*
4967  * Try to flush some data based on policy set by @state. This is only advisory
4968  * and may fail for various reasons. The caller is supposed to examine the
4969  * state of @space_info to detect the outcome.
4970  */
4971 static void flush_space(struct btrfs_fs_info *fs_info,
4972                        struct btrfs_space_info *space_info, u64 num_bytes,
4973                        int state)
4974 {
4975         struct btrfs_root *root = fs_info->extent_root;
4976         struct btrfs_trans_handle *trans;
4977         int nr;
4978         int ret = 0;
4979
4980         switch (state) {
4981         case FLUSH_DELAYED_ITEMS_NR:
4982         case FLUSH_DELAYED_ITEMS:
4983                 if (state == FLUSH_DELAYED_ITEMS_NR)
4984                         nr = calc_reclaim_items_nr(fs_info, num_bytes) * 2;
4985                 else
4986                         nr = -1;
4987
4988                 trans = btrfs_join_transaction(root);
4989                 if (IS_ERR(trans)) {
4990                         ret = PTR_ERR(trans);
4991                         break;
4992                 }
4993                 ret = btrfs_run_delayed_items_nr(trans, nr);
4994                 btrfs_end_transaction(trans);
4995                 break;
4996         case FLUSH_DELALLOC:
4997         case FLUSH_DELALLOC_WAIT:
4998                 shrink_delalloc(fs_info, num_bytes * 2, num_bytes,
4999                                 state == FLUSH_DELALLOC_WAIT);
5000                 break;
5001         case ALLOC_CHUNK:
5002                 trans = btrfs_join_transaction(root);
5003                 if (IS_ERR(trans)) {
5004                         ret = PTR_ERR(trans);
5005                         break;
5006                 }
5007                 ret = do_chunk_alloc(trans, fs_info,
5008                                      btrfs_metadata_alloc_profile(fs_info),
5009                                      CHUNK_ALLOC_NO_FORCE);
5010                 btrfs_end_transaction(trans);
5011                 if (ret > 0 || ret == -ENOSPC)
5012                         ret = 0;
5013                 break;
5014         case COMMIT_TRANS:
5015                 ret = may_commit_transaction(fs_info, space_info);
5016                 break;
5017         default:
5018                 ret = -ENOSPC;
5019                 break;
5020         }
5021
5022         trace_btrfs_flush_space(fs_info, space_info->flags, num_bytes, state,
5023                                 ret);
5024         return;
5025 }
5026
5027 static inline u64
5028 btrfs_calc_reclaim_metadata_size(struct btrfs_fs_info *fs_info,
5029                                  struct btrfs_space_info *space_info,
5030                                  bool system_chunk)
5031 {
5032         struct reserve_ticket *ticket;
5033         u64 used;
5034         u64 expected;
5035         u64 to_reclaim = 0;
5036
5037         list_for_each_entry(ticket, &space_info->tickets, list)
5038                 to_reclaim += ticket->bytes;
5039         list_for_each_entry(ticket, &space_info->priority_tickets, list)
5040                 to_reclaim += ticket->bytes;
5041         if (to_reclaim)
5042                 return to_reclaim;
5043
5044         to_reclaim = min_t(u64, num_online_cpus() * SZ_1M, SZ_16M);
5045         if (can_overcommit(fs_info, space_info, to_reclaim,
5046                            BTRFS_RESERVE_FLUSH_ALL, system_chunk))
5047                 return 0;
5048
5049         used = btrfs_space_info_used(space_info, true);
5050
5051         if (can_overcommit(fs_info, space_info, SZ_1M,
5052                            BTRFS_RESERVE_FLUSH_ALL, system_chunk))
5053                 expected = div_factor_fine(space_info->total_bytes, 95);
5054         else
5055                 expected = div_factor_fine(space_info->total_bytes, 90);
5056
5057         if (used > expected)
5058                 to_reclaim = used - expected;
5059         else
5060                 to_reclaim = 0;
5061         to_reclaim = min(to_reclaim, space_info->bytes_may_use +
5062                                      space_info->bytes_reserved);
5063         return to_reclaim;
5064 }
5065
5066 static inline int need_do_async_reclaim(struct btrfs_fs_info *fs_info,
5067                                         struct btrfs_space_info *space_info,
5068                                         u64 used, bool system_chunk)
5069 {
5070         u64 thresh = div_factor_fine(space_info->total_bytes, 98);
5071
5072         /* If we're just plain full then async reclaim just slows us down. */
5073         if ((space_info->bytes_used + space_info->bytes_reserved) >= thresh)
5074                 return 0;
5075
5076         if (!btrfs_calc_reclaim_metadata_size(fs_info, space_info,
5077                                               system_chunk))
5078                 return 0;
5079
5080         return (used >= thresh && !btrfs_fs_closing(fs_info) &&
5081                 !test_bit(BTRFS_FS_STATE_REMOUNTING, &fs_info->fs_state));
5082 }
5083
5084 static void wake_all_tickets(struct list_head *head)
5085 {
5086         struct reserve_ticket *ticket;
5087
5088         while (!list_empty(head)) {
5089                 ticket = list_first_entry(head, struct reserve_ticket, list);
5090                 list_del_init(&ticket->list);
5091                 ticket->error = -ENOSPC;
5092                 wake_up(&ticket->wait);
5093         }
5094 }
5095
5096 /*
5097  * This is for normal flushers, we can wait all goddamned day if we want to.  We
5098  * will loop and continuously try to flush as long as we are making progress.
5099  * We count progress as clearing off tickets each time we have to loop.
5100  */
5101 static void btrfs_async_reclaim_metadata_space(struct work_struct *work)
5102 {
5103         struct btrfs_fs_info *fs_info;
5104         struct btrfs_space_info *space_info;
5105         u64 to_reclaim;
5106         int flush_state;
5107         int commit_cycles = 0;
5108         u64 last_tickets_id;
5109
5110         fs_info = container_of(work, struct btrfs_fs_info, async_reclaim_work);
5111         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
5112
5113         spin_lock(&space_info->lock);
5114         to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info, space_info,
5115                                                       false);
5116         if (!to_reclaim) {
5117                 space_info->flush = 0;
5118                 spin_unlock(&space_info->lock);
5119                 return;
5120         }
5121         last_tickets_id = space_info->tickets_id;
5122         spin_unlock(&space_info->lock);
5123
5124         flush_state = FLUSH_DELAYED_ITEMS_NR;
5125         do {
5126                 flush_space(fs_info, space_info, to_reclaim, flush_state);
5127                 spin_lock(&space_info->lock);
5128                 if (list_empty(&space_info->tickets)) {
5129                         space_info->flush = 0;
5130                         spin_unlock(&space_info->lock);
5131                         return;
5132                 }
5133                 to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info,
5134                                                               space_info,
5135                                                               false);
5136                 if (last_tickets_id == space_info->tickets_id) {
5137                         flush_state++;
5138                 } else {
5139                         last_tickets_id = space_info->tickets_id;
5140                         flush_state = FLUSH_DELAYED_ITEMS_NR;
5141                         if (commit_cycles)
5142                                 commit_cycles--;
5143                 }
5144
5145                 if (flush_state > COMMIT_TRANS) {
5146                         commit_cycles++;
5147                         if (commit_cycles > 2) {
5148                                 wake_all_tickets(&space_info->tickets);
5149                                 space_info->flush = 0;
5150                         } else {
5151                                 flush_state = FLUSH_DELAYED_ITEMS_NR;
5152                         }
5153                 }
5154                 spin_unlock(&space_info->lock);
5155         } while (flush_state <= COMMIT_TRANS);
5156 }
5157
5158 void btrfs_init_async_reclaim_work(struct work_struct *work)
5159 {
5160         INIT_WORK(work, btrfs_async_reclaim_metadata_space);
5161 }
5162
5163 static void priority_reclaim_metadata_space(struct btrfs_fs_info *fs_info,
5164                                             struct btrfs_space_info *space_info,
5165                                             struct reserve_ticket *ticket)
5166 {
5167         u64 to_reclaim;
5168         int flush_state = FLUSH_DELAYED_ITEMS_NR;
5169
5170         spin_lock(&space_info->lock);
5171         to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info, space_info,
5172                                                       false);
5173         if (!to_reclaim) {
5174                 spin_unlock(&space_info->lock);
5175                 return;
5176         }
5177         spin_unlock(&space_info->lock);
5178
5179         do {
5180                 flush_space(fs_info, space_info, to_reclaim, flush_state);
5181                 flush_state++;
5182                 spin_lock(&space_info->lock);
5183                 if (ticket->bytes == 0) {
5184                         spin_unlock(&space_info->lock);
5185                         return;
5186                 }
5187                 spin_unlock(&space_info->lock);
5188
5189                 /*
5190                  * Priority flushers can't wait on delalloc without
5191                  * deadlocking.
5192                  */
5193                 if (flush_state == FLUSH_DELALLOC ||
5194                     flush_state == FLUSH_DELALLOC_WAIT)
5195                         flush_state = ALLOC_CHUNK;
5196         } while (flush_state < COMMIT_TRANS);
5197 }
5198
5199 static int wait_reserve_ticket(struct btrfs_fs_info *fs_info,
5200                                struct btrfs_space_info *space_info,
5201                                struct reserve_ticket *ticket, u64 orig_bytes)
5202
5203 {
5204         DEFINE_WAIT(wait);
5205         int ret = 0;
5206
5207         spin_lock(&space_info->lock);
5208         while (ticket->bytes > 0 && ticket->error == 0) {
5209                 ret = prepare_to_wait_event(&ticket->wait, &wait, TASK_KILLABLE);
5210                 if (ret) {
5211                         ret = -EINTR;
5212                         break;
5213                 }
5214                 spin_unlock(&space_info->lock);
5215
5216                 schedule();
5217
5218                 finish_wait(&ticket->wait, &wait);
5219                 spin_lock(&space_info->lock);
5220         }
5221         if (!ret)
5222                 ret = ticket->error;
5223         if (!list_empty(&ticket->list))
5224                 list_del_init(&ticket->list);
5225         if (ticket->bytes && ticket->bytes < orig_bytes) {
5226                 u64 num_bytes = orig_bytes - ticket->bytes;
5227                 space_info->bytes_may_use -= num_bytes;
5228                 trace_btrfs_space_reservation(fs_info, "space_info",
5229                                               space_info->flags, num_bytes, 0);
5230         }
5231         spin_unlock(&space_info->lock);
5232
5233         return ret;
5234 }
5235
5236 /**
5237  * reserve_metadata_bytes - try to reserve bytes from the block_rsv's space
5238  * @root - the root we're allocating for
5239  * @space_info - the space info we want to allocate from
5240  * @orig_bytes - the number of bytes we want
5241  * @flush - whether or not we can flush to make our reservation
5242  *
5243  * This will reserve orig_bytes number of bytes from the space info associated
5244  * with the block_rsv.  If there is not enough space it will make an attempt to
5245  * flush out space to make room.  It will do this by flushing delalloc if
5246  * possible or committing the transaction.  If flush is 0 then no attempts to
5247  * regain reservations will be made and this will fail if there is not enough
5248  * space already.
5249  */
5250 static int __reserve_metadata_bytes(struct btrfs_fs_info *fs_info,
5251                                     struct btrfs_space_info *space_info,
5252                                     u64 orig_bytes,
5253                                     enum btrfs_reserve_flush_enum flush,
5254                                     bool system_chunk)
5255 {
5256         struct reserve_ticket ticket;
5257         u64 used;
5258         int ret = 0;
5259
5260         ASSERT(orig_bytes);
5261         ASSERT(!current->journal_info || flush != BTRFS_RESERVE_FLUSH_ALL);
5262
5263         spin_lock(&space_info->lock);
5264         ret = -ENOSPC;
5265         used = btrfs_space_info_used(space_info, true);
5266
5267         /*
5268          * If we have enough space then hooray, make our reservation and carry
5269          * on.  If not see if we can overcommit, and if we can, hooray carry on.
5270          * If not things get more complicated.
5271          */
5272         if (used + orig_bytes <= space_info->total_bytes) {
5273                 space_info->bytes_may_use += orig_bytes;
5274                 trace_btrfs_space_reservation(fs_info, "space_info",
5275                                               space_info->flags, orig_bytes, 1);
5276                 ret = 0;
5277         } else if (can_overcommit(fs_info, space_info, orig_bytes, flush,
5278                                   system_chunk)) {
5279                 space_info->bytes_may_use += orig_bytes;
5280                 trace_btrfs_space_reservation(fs_info, "space_info",
5281                                               space_info->flags, orig_bytes, 1);
5282                 ret = 0;
5283         }
5284
5285         /*
5286          * If we couldn't make a reservation then setup our reservation ticket
5287          * and kick the async worker if it's not already running.
5288          *
5289          * If we are a priority flusher then we just need to add our ticket to
5290          * the list and we will do our own flushing further down.
5291          */
5292         if (ret && flush != BTRFS_RESERVE_NO_FLUSH) {
5293                 ticket.bytes = orig_bytes;
5294                 ticket.error = 0;
5295                 init_waitqueue_head(&ticket.wait);
5296                 if (flush == BTRFS_RESERVE_FLUSH_ALL) {
5297                         list_add_tail(&ticket.list, &space_info->tickets);
5298                         if (!space_info->flush) {
5299                                 space_info->flush = 1;
5300                                 trace_btrfs_trigger_flush(fs_info,
5301                                                           space_info->flags,
5302                                                           orig_bytes, flush,
5303                                                           "enospc");
5304                                 queue_work(system_unbound_wq,
5305                                            &fs_info->async_reclaim_work);
5306                         }
5307                 } else {
5308                         list_add_tail(&ticket.list,
5309                                       &space_info->priority_tickets);
5310                 }
5311         } else if (!ret && space_info->flags & BTRFS_BLOCK_GROUP_METADATA) {
5312                 used += orig_bytes;
5313                 /*
5314                  * We will do the space reservation dance during log replay,
5315                  * which means we won't have fs_info->fs_root set, so don't do
5316                  * the async reclaim as we will panic.
5317                  */
5318                 if (!test_bit(BTRFS_FS_LOG_RECOVERING, &fs_info->flags) &&
5319                     need_do_async_reclaim(fs_info, space_info,
5320                                           used, system_chunk) &&
5321                     !work_busy(&fs_info->async_reclaim_work)) {
5322                         trace_btrfs_trigger_flush(fs_info, space_info->flags,
5323                                                   orig_bytes, flush, "preempt");
5324                         queue_work(system_unbound_wq,
5325                                    &fs_info->async_reclaim_work);
5326                 }
5327         }
5328         spin_unlock(&space_info->lock);
5329         if (!ret || flush == BTRFS_RESERVE_NO_FLUSH)
5330                 return ret;
5331
5332         if (flush == BTRFS_RESERVE_FLUSH_ALL)
5333                 return wait_reserve_ticket(fs_info, space_info, &ticket,
5334                                            orig_bytes);
5335
5336         ret = 0;
5337         priority_reclaim_metadata_space(fs_info, space_info, &ticket);
5338         spin_lock(&space_info->lock);
5339         if (ticket.bytes) {
5340                 if (ticket.bytes < orig_bytes) {
5341                         u64 num_bytes = orig_bytes - ticket.bytes;
5342                         space_info->bytes_may_use -= num_bytes;
5343                         trace_btrfs_space_reservation(fs_info, "space_info",
5344                                                       space_info->flags,
5345                                                       num_bytes, 0);
5346
5347                 }
5348                 list_del_init(&ticket.list);
5349                 ret = -ENOSPC;
5350         }
5351         spin_unlock(&space_info->lock);
5352         ASSERT(list_empty(&ticket.list));
5353         return ret;
5354 }
5355
5356 /**
5357  * reserve_metadata_bytes - try to reserve bytes from the block_rsv's space
5358  * @root - the root we're allocating for
5359  * @block_rsv - the block_rsv we're allocating for
5360  * @orig_bytes - the number of bytes we want
5361  * @flush - whether or not we can flush to make our reservation
5362  *
5363  * This will reserve orgi_bytes number of bytes from the space info associated
5364  * with the block_rsv.  If there is not enough space it will make an attempt to
5365  * flush out space to make room.  It will do this by flushing delalloc if
5366  * possible or committing the transaction.  If flush is 0 then no attempts to
5367  * regain reservations will be made and this will fail if there is not enough
5368  * space already.
5369  */
5370 static int reserve_metadata_bytes(struct btrfs_root *root,
5371                                   struct btrfs_block_rsv *block_rsv,
5372                                   u64 orig_bytes,
5373                                   enum btrfs_reserve_flush_enum flush)
5374 {
5375         struct btrfs_fs_info *fs_info = root->fs_info;
5376         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
5377         int ret;
5378         bool system_chunk = (root == fs_info->chunk_root);
5379
5380         ret = __reserve_metadata_bytes(fs_info, block_rsv->space_info,
5381                                        orig_bytes, flush, system_chunk);
5382         if (ret == -ENOSPC &&
5383             unlikely(root->orphan_cleanup_state == ORPHAN_CLEANUP_STARTED)) {
5384                 if (block_rsv != global_rsv &&
5385                     !block_rsv_use_bytes(global_rsv, orig_bytes))
5386                         ret = 0;
5387         }
5388         if (ret == -ENOSPC) {
5389                 trace_btrfs_space_reservation(fs_info, "space_info:enospc",
5390                                               block_rsv->space_info->flags,
5391                                               orig_bytes, 1);
5392
5393                 if (btrfs_test_opt(fs_info, ENOSPC_DEBUG))
5394                         dump_space_info(fs_info, block_rsv->space_info,
5395                                         orig_bytes, 0);
5396         }
5397         return ret;
5398 }
5399
5400 static struct btrfs_block_rsv *get_block_rsv(
5401                                         const struct btrfs_trans_handle *trans,
5402                                         const struct btrfs_root *root)
5403 {
5404         struct btrfs_fs_info *fs_info = root->fs_info;
5405         struct btrfs_block_rsv *block_rsv = NULL;
5406
5407         if (test_bit(BTRFS_ROOT_REF_COWS, &root->state) ||
5408             (root == fs_info->csum_root && trans->adding_csums) ||
5409             (root == fs_info->uuid_root))
5410                 block_rsv = trans->block_rsv;
5411
5412         if (!block_rsv)
5413                 block_rsv = root->block_rsv;
5414
5415         if (!block_rsv)
5416                 block_rsv = &fs_info->empty_block_rsv;
5417
5418         return block_rsv;
5419 }
5420
5421 static int block_rsv_use_bytes(struct btrfs_block_rsv *block_rsv,
5422                                u64 num_bytes)
5423 {
5424         int ret = -ENOSPC;
5425         spin_lock(&block_rsv->lock);
5426         if (block_rsv->reserved >= num_bytes) {
5427                 block_rsv->reserved -= num_bytes;
5428                 if (block_rsv->reserved < block_rsv->size)
5429                         block_rsv->full = 0;
5430                 ret = 0;
5431         }
5432         spin_unlock(&block_rsv->lock);
5433         return ret;
5434 }
5435
5436 static void block_rsv_add_bytes(struct btrfs_block_rsv *block_rsv,
5437                                 u64 num_bytes, int update_size)
5438 {
5439         spin_lock(&block_rsv->lock);
5440         block_rsv->reserved += num_bytes;
5441         if (update_size)
5442                 block_rsv->size += num_bytes;
5443         else if (block_rsv->reserved >= block_rsv->size)
5444                 block_rsv->full = 1;
5445         spin_unlock(&block_rsv->lock);
5446 }
5447
5448 int btrfs_cond_migrate_bytes(struct btrfs_fs_info *fs_info,
5449                              struct btrfs_block_rsv *dest, u64 num_bytes,
5450                              int min_factor)
5451 {
5452         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
5453         u64 min_bytes;
5454
5455         if (global_rsv->space_info != dest->space_info)
5456                 return -ENOSPC;
5457
5458         spin_lock(&global_rsv->lock);
5459         min_bytes = div_factor(global_rsv->size, min_factor);
5460         if (global_rsv->reserved < min_bytes + num_bytes) {
5461                 spin_unlock(&global_rsv->lock);
5462                 return -ENOSPC;
5463         }
5464         global_rsv->reserved -= num_bytes;
5465         if (global_rsv->reserved < global_rsv->size)
5466                 global_rsv->full = 0;
5467         spin_unlock(&global_rsv->lock);
5468
5469         block_rsv_add_bytes(dest, num_bytes, 1);
5470         return 0;
5471 }
5472
5473 /*
5474  * This is for space we already have accounted in space_info->bytes_may_use, so
5475  * basically when we're returning space from block_rsv's.
5476  */
5477 static void space_info_add_old_bytes(struct btrfs_fs_info *fs_info,
5478                                      struct btrfs_space_info *space_info,
5479                                      u64 num_bytes)
5480 {
5481         struct reserve_ticket *ticket;
5482         struct list_head *head;
5483         u64 used;
5484         enum btrfs_reserve_flush_enum flush = BTRFS_RESERVE_NO_FLUSH;
5485         bool check_overcommit = false;
5486
5487         spin_lock(&space_info->lock);
5488         head = &space_info->priority_tickets;
5489
5490         /*
5491          * If we are over our limit then we need to check and see if we can
5492          * overcommit, and if we can't then we just need to free up our space
5493          * and not satisfy any requests.
5494          */
5495         used = btrfs_space_info_used(space_info, true);
5496         if (used - num_bytes >= space_info->total_bytes)
5497                 check_overcommit = true;
5498 again:
5499         while (!list_empty(head) && num_bytes) {
5500                 ticket = list_first_entry(head, struct reserve_ticket,
5501                                           list);
5502                 /*
5503                  * We use 0 bytes because this space is already reserved, so
5504                  * adding the ticket space would be a double count.
5505                  */
5506                 if (check_overcommit &&
5507                     !can_overcommit(fs_info, space_info, 0, flush, false))
5508                         break;
5509                 if (num_bytes >= ticket->bytes) {
5510                         list_del_init(&ticket->list);
5511                         num_bytes -= ticket->bytes;
5512                         ticket->bytes = 0;
5513                         space_info->tickets_id++;
5514                         wake_up(&ticket->wait);
5515                 } else {
5516                         ticket->bytes -= num_bytes;
5517                         num_bytes = 0;
5518                 }
5519         }
5520
5521         if (num_bytes && head == &space_info->priority_tickets) {
5522                 head = &space_info->tickets;
5523                 flush = BTRFS_RESERVE_FLUSH_ALL;
5524                 goto again;
5525         }
5526         space_info->bytes_may_use -= num_bytes;
5527         trace_btrfs_space_reservation(fs_info, "space_info",
5528                                       space_info->flags, num_bytes, 0);
5529         spin_unlock(&space_info->lock);
5530 }
5531
5532 /*
5533  * This is for newly allocated space that isn't accounted in
5534  * space_info->bytes_may_use yet.  So if we allocate a chunk or unpin an extent
5535  * we use this helper.
5536  */
5537 static void space_info_add_new_bytes(struct btrfs_fs_info *fs_info,
5538                                      struct btrfs_space_info *space_info,
5539                                      u64 num_bytes)
5540 {
5541         struct reserve_ticket *ticket;
5542         struct list_head *head = &space_info->priority_tickets;
5543
5544 again:
5545         while (!list_empty(head) && num_bytes) {
5546                 ticket = list_first_entry(head, struct reserve_ticket,
5547                                           list);
5548                 if (num_bytes >= ticket->bytes) {
5549                         trace_btrfs_space_reservation(fs_info, "space_info",
5550                                                       space_info->flags,
5551                                                       ticket->bytes, 1);
5552                         list_del_init(&ticket->list);
5553                         num_bytes -= ticket->bytes;
5554                         space_info->bytes_may_use += ticket->bytes;
5555                         ticket->bytes = 0;
5556                         space_info->tickets_id++;
5557                         wake_up(&ticket->wait);
5558                 } else {
5559                         trace_btrfs_space_reservation(fs_info, "space_info",
5560                                                       space_info->flags,
5561                                                       num_bytes, 1);
5562                         space_info->bytes_may_use += num_bytes;
5563                         ticket->bytes -= num_bytes;
5564                         num_bytes = 0;
5565                 }
5566         }
5567
5568         if (num_bytes && head == &space_info->priority_tickets) {
5569                 head = &space_info->tickets;
5570                 goto again;
5571         }
5572 }
5573
5574 static u64 block_rsv_release_bytes(struct btrfs_fs_info *fs_info,
5575                                     struct btrfs_block_rsv *block_rsv,
5576                                     struct btrfs_block_rsv *dest, u64 num_bytes,
5577                                     u64 *qgroup_to_release_ret)
5578 {
5579         struct btrfs_space_info *space_info = block_rsv->space_info;
5580         u64 qgroup_to_release = 0;
5581         u64 ret;
5582
5583         spin_lock(&block_rsv->lock);
5584         if (num_bytes == (u64)-1) {
5585                 num_bytes = block_rsv->size;
5586                 qgroup_to_release = block_rsv->qgroup_rsv_size;
5587         }
5588         block_rsv->size -= num_bytes;
5589         if (block_rsv->reserved >= block_rsv->size) {
5590                 num_bytes = block_rsv->reserved - block_rsv->size;
5591                 block_rsv->reserved = block_rsv->size;
5592                 block_rsv->full = 1;
5593         } else {
5594                 num_bytes = 0;
5595         }
5596         if (block_rsv->qgroup_rsv_reserved >= block_rsv->qgroup_rsv_size) {
5597                 qgroup_to_release = block_rsv->qgroup_rsv_reserved -
5598                                     block_rsv->qgroup_rsv_size;
5599                 block_rsv->qgroup_rsv_reserved = block_rsv->qgroup_rsv_size;
5600         } else {
5601                 qgroup_to_release = 0;
5602         }
5603         spin_unlock(&block_rsv->lock);
5604
5605         ret = num_bytes;
5606         if (num_bytes > 0) {
5607                 if (dest) {
5608                         spin_lock(&dest->lock);
5609                         if (!dest->full) {
5610                                 u64 bytes_to_add;
5611
5612                                 bytes_to_add = dest->size - dest->reserved;
5613                                 bytes_to_add = min(num_bytes, bytes_to_add);
5614                                 dest->reserved += bytes_to_add;
5615                                 if (dest->reserved >= dest->size)
5616                                         dest->full = 1;
5617                                 num_bytes -= bytes_to_add;
5618                         }
5619                         spin_unlock(&dest->lock);
5620                 }
5621                 if (num_bytes)
5622                         space_info_add_old_bytes(fs_info, space_info,
5623                                                  num_bytes);
5624         }
5625         if (qgroup_to_release_ret)
5626                 *qgroup_to_release_ret = qgroup_to_release;
5627         return ret;
5628 }
5629
5630 int btrfs_block_rsv_migrate(struct btrfs_block_rsv *src,
5631                             struct btrfs_block_rsv *dst, u64 num_bytes,
5632                             int update_size)
5633 {
5634         int ret;
5635
5636         ret = block_rsv_use_bytes(src, num_bytes);
5637         if (ret)
5638                 return ret;
5639
5640         block_rsv_add_bytes(dst, num_bytes, update_size);
5641         return 0;
5642 }
5643
5644 void btrfs_init_block_rsv(struct btrfs_block_rsv *rsv, unsigned short type)
5645 {
5646         memset(rsv, 0, sizeof(*rsv));
5647         spin_lock_init(&rsv->lock);
5648         rsv->type = type;
5649 }
5650
5651 void btrfs_init_metadata_block_rsv(struct btrfs_fs_info *fs_info,
5652                                    struct btrfs_block_rsv *rsv,
5653                                    unsigned short type)
5654 {
5655         btrfs_init_block_rsv(rsv, type);
5656         rsv->space_info = __find_space_info(fs_info,
5657                                             BTRFS_BLOCK_GROUP_METADATA);
5658 }
5659
5660 struct btrfs_block_rsv *btrfs_alloc_block_rsv(struct btrfs_fs_info *fs_info,
5661                                               unsigned short type)
5662 {
5663         struct btrfs_block_rsv *block_rsv;
5664
5665         block_rsv = kmalloc(sizeof(*block_rsv), GFP_NOFS);
5666         if (!block_rsv)
5667                 return NULL;
5668
5669         btrfs_init_metadata_block_rsv(fs_info, block_rsv, type);
5670         return block_rsv;
5671 }
5672
5673 void btrfs_free_block_rsv(struct btrfs_fs_info *fs_info,
5674                           struct btrfs_block_rsv *rsv)
5675 {
5676         if (!rsv)
5677                 return;
5678         btrfs_block_rsv_release(fs_info, rsv, (u64)-1);
5679         kfree(rsv);
5680 }
5681
5682 void __btrfs_free_block_rsv(struct btrfs_block_rsv *rsv)
5683 {
5684         kfree(rsv);
5685 }
5686
5687 int btrfs_block_rsv_add(struct btrfs_root *root,
5688                         struct btrfs_block_rsv *block_rsv, u64 num_bytes,
5689                         enum btrfs_reserve_flush_enum flush)
5690 {
5691         int ret;
5692
5693         if (num_bytes == 0)
5694                 return 0;
5695
5696         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
5697         if (!ret) {
5698                 block_rsv_add_bytes(block_rsv, num_bytes, 1);
5699                 return 0;
5700         }
5701
5702         return ret;
5703 }
5704
5705 int btrfs_block_rsv_check(struct btrfs_block_rsv *block_rsv, int min_factor)
5706 {
5707         u64 num_bytes = 0;
5708         int ret = -ENOSPC;
5709
5710         if (!block_rsv)
5711                 return 0;
5712
5713         spin_lock(&block_rsv->lock);
5714         num_bytes = div_factor(block_rsv->size, min_factor);
5715         if (block_rsv->reserved >= num_bytes)
5716                 ret = 0;
5717         spin_unlock(&block_rsv->lock);
5718
5719         return ret;
5720 }
5721
5722 int btrfs_block_rsv_refill(struct btrfs_root *root,
5723                            struct btrfs_block_rsv *block_rsv, u64 min_reserved,
5724                            enum btrfs_reserve_flush_enum flush)
5725 {
5726         u64 num_bytes = 0;
5727         int ret = -ENOSPC;
5728
5729         if (!block_rsv)
5730                 return 0;
5731
5732         spin_lock(&block_rsv->lock);
5733         num_bytes = min_reserved;
5734         if (block_rsv->reserved >= num_bytes)
5735                 ret = 0;
5736         else
5737                 num_bytes -= block_rsv->reserved;
5738         spin_unlock(&block_rsv->lock);
5739
5740         if (!ret)
5741                 return 0;
5742
5743         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
5744         if (!ret) {
5745                 block_rsv_add_bytes(block_rsv, num_bytes, 0);
5746                 return 0;
5747         }
5748
5749         return ret;
5750 }
5751
5752 /**
5753  * btrfs_inode_rsv_refill - refill the inode block rsv.
5754  * @inode - the inode we are refilling.
5755  * @flush - the flusing restriction.
5756  *
5757  * Essentially the same as btrfs_block_rsv_refill, except it uses the
5758  * block_rsv->size as the minimum size.  We'll either refill the missing amount
5759  * or return if we already have enough space.  This will also handle the resreve
5760  * tracepoint for the reserved amount.
5761  */
5762 static int btrfs_inode_rsv_refill(struct btrfs_inode *inode,
5763                                   enum btrfs_reserve_flush_enum flush)
5764 {
5765         struct btrfs_root *root = inode->root;
5766         struct btrfs_block_rsv *block_rsv = &inode->block_rsv;
5767         u64 num_bytes = 0;
5768         u64 qgroup_num_bytes = 0;
5769         int ret = -ENOSPC;
5770
5771         spin_lock(&block_rsv->lock);
5772         if (block_rsv->reserved < block_rsv->size)
5773                 num_bytes = block_rsv->size - block_rsv->reserved;
5774         if (block_rsv->qgroup_rsv_reserved < block_rsv->qgroup_rsv_size)
5775                 qgroup_num_bytes = block_rsv->qgroup_rsv_size -
5776                                    block_rsv->qgroup_rsv_reserved;
5777         spin_unlock(&block_rsv->lock);
5778
5779         if (num_bytes == 0)
5780                 return 0;
5781
5782         ret = btrfs_qgroup_reserve_meta_prealloc(root, qgroup_num_bytes, true);
5783         if (ret)
5784                 return ret;
5785         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
5786         if (!ret) {
5787                 block_rsv_add_bytes(block_rsv, num_bytes, 0);
5788                 trace_btrfs_space_reservation(root->fs_info, "delalloc",
5789                                               btrfs_ino(inode), num_bytes, 1);
5790
5791                 /* Don't forget to increase qgroup_rsv_reserved */
5792                 spin_lock(&block_rsv->lock);
5793                 block_rsv->qgroup_rsv_reserved += qgroup_num_bytes;
5794                 spin_unlock(&block_rsv->lock);
5795         } else
5796                 btrfs_qgroup_free_meta_prealloc(root, qgroup_num_bytes);
5797         return ret;
5798 }
5799
5800 /**
5801  * btrfs_inode_rsv_release - release any excessive reservation.
5802  * @inode - the inode we need to release from.
5803  * @qgroup_free - free or convert qgroup meta.
5804  *   Unlike normal operation, qgroup meta reservation needs to know if we are
5805  *   freeing qgroup reservation or just converting it into per-trans.  Normally
5806  *   @qgroup_free is true for error handling, and false for normal release.
5807  *
5808  * This is the same as btrfs_block_rsv_release, except that it handles the
5809  * tracepoint for the reservation.
5810  */
5811 static void btrfs_inode_rsv_release(struct btrfs_inode *inode, bool qgroup_free)
5812 {
5813         struct btrfs_fs_info *fs_info = inode->root->fs_info;
5814         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
5815         struct btrfs_block_rsv *block_rsv = &inode->block_rsv;
5816         u64 released = 0;
5817         u64 qgroup_to_release = 0;
5818
5819         /*
5820          * Since we statically set the block_rsv->size we just want to say we
5821          * are releasing 0 bytes, and then we'll just get the reservation over
5822          * the size free'd.
5823          */
5824         released = block_rsv_release_bytes(fs_info, block_rsv, global_rsv, 0,
5825                                            &qgroup_to_release);
5826         if (released > 0)
5827                 trace_btrfs_space_reservation(fs_info, "delalloc",
5828                                               btrfs_ino(inode), released, 0);
5829         if (qgroup_free)
5830                 btrfs_qgroup_free_meta_prealloc(inode->root, qgroup_to_release);
5831         else
5832                 btrfs_qgroup_convert_reserved_meta(inode->root,
5833                                                    qgroup_to_release);
5834 }
5835
5836 void btrfs_block_rsv_release(struct btrfs_fs_info *fs_info,
5837                              struct btrfs_block_rsv *block_rsv,
5838                              u64 num_bytes)
5839 {
5840         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
5841
5842         if (global_rsv == block_rsv ||
5843             block_rsv->space_info != global_rsv->space_info)
5844                 global_rsv = NULL;
5845         block_rsv_release_bytes(fs_info, block_rsv, global_rsv, num_bytes, NULL);
5846 }
5847
5848 static void update_global_block_rsv(struct btrfs_fs_info *fs_info)
5849 {
5850         struct btrfs_block_rsv *block_rsv = &fs_info->global_block_rsv;
5851         struct btrfs_space_info *sinfo = block_rsv->space_info;
5852         u64 num_bytes;
5853
5854         /*
5855          * The global block rsv is based on the size of the extent tree, the
5856          * checksum tree and the root tree.  If the fs is empty we want to set
5857          * it to a minimal amount for safety.
5858          */
5859         num_bytes = btrfs_root_used(&fs_info->extent_root->root_item) +
5860                 btrfs_root_used(&fs_info->csum_root->root_item) +
5861                 btrfs_root_used(&fs_info->tree_root->root_item);
5862         num_bytes = max_t(u64, num_bytes, SZ_16M);
5863
5864         spin_lock(&sinfo->lock);
5865         spin_lock(&block_rsv->lock);
5866
5867         block_rsv->size = min_t(u64, num_bytes, SZ_512M);
5868
5869         if (block_rsv->reserved < block_rsv->size) {
5870                 num_bytes = btrfs_space_info_used(sinfo, true);
5871                 if (sinfo->total_bytes > num_bytes) {
5872                         num_bytes = sinfo->total_bytes - num_bytes;
5873                         num_bytes = min(num_bytes,
5874                                         block_rsv->size - block_rsv->reserved);
5875                         block_rsv->reserved += num_bytes;
5876                         sinfo->bytes_may_use += num_bytes;
5877                         trace_btrfs_space_reservation(fs_info, "space_info",
5878                                                       sinfo->flags, num_bytes,
5879                                                       1);
5880                 }
5881         } else if (block_rsv->reserved > block_rsv->size) {
5882                 num_bytes = block_rsv->reserved - block_rsv->size;
5883                 sinfo->bytes_may_use -= num_bytes;
5884                 trace_btrfs_space_reservation(fs_info, "space_info",
5885                                       sinfo->flags, num_bytes, 0);
5886                 block_rsv->reserved = block_rsv->size;
5887         }
5888
5889         if (block_rsv->reserved == block_rsv->size)
5890                 block_rsv->full = 1;
5891         else
5892                 block_rsv->full = 0;
5893
5894         spin_unlock(&block_rsv->lock);
5895         spin_unlock(&sinfo->lock);
5896 }
5897
5898 static void init_global_block_rsv(struct btrfs_fs_info *fs_info)
5899 {
5900         struct btrfs_space_info *space_info;
5901
5902         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
5903         fs_info->chunk_block_rsv.space_info = space_info;
5904
5905         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
5906         fs_info->global_block_rsv.space_info = space_info;
5907         fs_info->trans_block_rsv.space_info = space_info;
5908         fs_info->empty_block_rsv.space_info = space_info;
5909         fs_info->delayed_block_rsv.space_info = space_info;
5910
5911         fs_info->extent_root->block_rsv = &fs_info->global_block_rsv;
5912         fs_info->csum_root->block_rsv = &fs_info->global_block_rsv;
5913         fs_info->dev_root->block_rsv = &fs_info->global_block_rsv;
5914         fs_info->tree_root->block_rsv = &fs_info->global_block_rsv;
5915         if (fs_info->quota_root)
5916                 fs_info->quota_root->block_rsv = &fs_info->global_block_rsv;
5917         fs_info->chunk_root->block_rsv = &fs_info->chunk_block_rsv;
5918
5919         update_global_block_rsv(fs_info);
5920 }
5921
5922 static void release_global_block_rsv(struct btrfs_fs_info *fs_info)
5923 {
5924         block_rsv_release_bytes(fs_info, &fs_info->global_block_rsv, NULL,
5925                                 (u64)-1, NULL);
5926         WARN_ON(fs_info->trans_block_rsv.size > 0);
5927         WARN_ON(fs_info->trans_block_rsv.reserved > 0);
5928         WARN_ON(fs_info->chunk_block_rsv.size > 0);
5929         WARN_ON(fs_info->chunk_block_rsv.reserved > 0);
5930         WARN_ON(fs_info->delayed_block_rsv.size > 0);
5931         WARN_ON(fs_info->delayed_block_rsv.reserved > 0);
5932 }
5933
5934
5935 /*
5936  * To be called after all the new block groups attached to the transaction
5937  * handle have been created (btrfs_create_pending_block_groups()).
5938  */
5939 void btrfs_trans_release_chunk_metadata(struct btrfs_trans_handle *trans)
5940 {
5941         struct btrfs_fs_info *fs_info = trans->fs_info;
5942
5943         if (!trans->chunk_bytes_reserved)
5944                 return;
5945
5946         WARN_ON_ONCE(!list_empty(&trans->new_bgs));
5947
5948         block_rsv_release_bytes(fs_info, &fs_info->chunk_block_rsv, NULL,
5949                                 trans->chunk_bytes_reserved, NULL);
5950         trans->chunk_bytes_reserved = 0;
5951 }
5952
5953 /* Can only return 0 or -ENOSPC */
5954 int btrfs_orphan_reserve_metadata(struct btrfs_trans_handle *trans,
5955                                   struct btrfs_inode *inode)
5956 {
5957         struct btrfs_fs_info *fs_info = btrfs_sb(inode->vfs_inode.i_sb);
5958         struct btrfs_root *root = inode->root;
5959         /*
5960          * We always use trans->block_rsv here as we will have reserved space
5961          * for our orphan when starting the transaction, using get_block_rsv()
5962          * here will sometimes make us choose the wrong block rsv as we could be
5963          * doing a reloc inode for a non refcounted root.
5964          */
5965         struct btrfs_block_rsv *src_rsv = trans->block_rsv;
5966         struct btrfs_block_rsv *dst_rsv = root->orphan_block_rsv;
5967
5968         /*
5969          * We need to hold space in order to delete our orphan item once we've
5970          * added it, so this takes the reservation so we can release it later
5971          * when we are truly done with the orphan item.
5972          */
5973         u64 num_bytes = btrfs_calc_trans_metadata_size(fs_info, 1);
5974
5975         trace_btrfs_space_reservation(fs_info, "orphan", btrfs_ino(inode),
5976                         num_bytes, 1);
5977         return btrfs_block_rsv_migrate(src_rsv, dst_rsv, num_bytes, 1);
5978 }
5979
5980 void btrfs_orphan_release_metadata(struct btrfs_inode *inode)
5981 {
5982         struct btrfs_fs_info *fs_info = btrfs_sb(inode->vfs_inode.i_sb);
5983         struct btrfs_root *root = inode->root;
5984         u64 num_bytes = btrfs_calc_trans_metadata_size(fs_info, 1);
5985
5986         trace_btrfs_space_reservation(fs_info, "orphan", btrfs_ino(inode),
5987                         num_bytes, 0);
5988         btrfs_block_rsv_release(fs_info, root->orphan_block_rsv, num_bytes);
5989 }
5990
5991 /*
5992  * btrfs_subvolume_reserve_metadata() - reserve space for subvolume operation
5993  * root: the root of the parent directory
5994  * rsv: block reservation
5995  * items: the number of items that we need do reservation
5996  * qgroup_reserved: used to return the reserved size in qgroup
5997  *
5998  * This function is used to reserve the space for snapshot/subvolume
5999  * creation and deletion. Those operations are different with the
6000  * common file/directory operations, they change two fs/file trees
6001  * and root tree, the number of items that the qgroup reserves is
6002  * different with the free space reservation. So we can not use
6003  * the space reservation mechanism in start_transaction().
6004  */
6005 int btrfs_subvolume_reserve_metadata(struct btrfs_root *root,
6006                                      struct btrfs_block_rsv *rsv,
6007                                      int items,
6008                                      u64 *qgroup_reserved,
6009                                      bool use_global_rsv)
6010 {
6011         u64 num_bytes;
6012         int ret;
6013         struct btrfs_fs_info *fs_info = root->fs_info;
6014         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
6015
6016         if (test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags)) {
6017                 /* One for parent inode, two for dir entries */
6018                 num_bytes = 3 * fs_info->nodesize;
6019                 ret = btrfs_qgroup_reserve_meta_prealloc(root, num_bytes, true);
6020                 if (ret)
6021                         return ret;
6022         } else {
6023                 num_bytes = 0;
6024         }
6025
6026         *qgroup_reserved = num_bytes;
6027
6028         num_bytes = btrfs_calc_trans_metadata_size(fs_info, items);
6029         rsv->space_info = __find_space_info(fs_info,
6030                                             BTRFS_BLOCK_GROUP_METADATA);
6031         ret = btrfs_block_rsv_add(root, rsv, num_bytes,
6032                                   BTRFS_RESERVE_FLUSH_ALL);
6033
6034         if (ret == -ENOSPC && use_global_rsv)
6035                 ret = btrfs_block_rsv_migrate(global_rsv, rsv, num_bytes, 1);
6036
6037         if (ret && *qgroup_reserved)
6038                 btrfs_qgroup_free_meta_prealloc(root, *qgroup_reserved);
6039
6040         return ret;
6041 }
6042
6043 void btrfs_subvolume_release_metadata(struct btrfs_fs_info *fs_info,
6044                                       struct btrfs_block_rsv *rsv)
6045 {
6046         btrfs_block_rsv_release(fs_info, rsv, (u64)-1);
6047 }
6048
6049 static void btrfs_calculate_inode_block_rsv_size(struct btrfs_fs_info *fs_info,
6050                                                  struct btrfs_inode *inode)
6051 {
6052         struct btrfs_block_rsv *block_rsv = &inode->block_rsv;
6053         u64 reserve_size = 0;
6054         u64 qgroup_rsv_size = 0;
6055         u64 csum_leaves;
6056         unsigned outstanding_extents;
6057
6058         lockdep_assert_held(&inode->lock);
6059         outstanding_extents = inode->outstanding_extents;
6060         if (outstanding_extents)
6061                 reserve_size = btrfs_calc_trans_metadata_size(fs_info,
6062                                                 outstanding_extents + 1);
6063         csum_leaves = btrfs_csum_bytes_to_leaves(fs_info,
6064                                                  inode->csum_bytes);
6065         reserve_size += btrfs_calc_trans_metadata_size(fs_info,
6066                                                        csum_leaves);
6067         /*
6068          * For qgroup rsv, the calculation is very simple:
6069          * account one nodesize for each outstanding extent
6070          *
6071          * This is overestimating in most cases.
6072          */
6073         qgroup_rsv_size = outstanding_extents * fs_info->nodesize;
6074
6075         spin_lock(&block_rsv->lock);
6076         block_rsv->size = reserve_size;
6077         block_rsv->qgroup_rsv_size = qgroup_rsv_size;
6078         spin_unlock(&block_rsv->lock);
6079 }
6080
6081 int btrfs_delalloc_reserve_metadata(struct btrfs_inode *inode, u64 num_bytes)
6082 {
6083         struct btrfs_fs_info *fs_info = btrfs_sb(inode->vfs_inode.i_sb);
6084         unsigned nr_extents;
6085         enum btrfs_reserve_flush_enum flush = BTRFS_RESERVE_FLUSH_ALL;
6086         int ret = 0;
6087         bool delalloc_lock = true;
6088
6089         /* If we are a free space inode we need to not flush since we will be in
6090          * the middle of a transaction commit.  We also don't need the delalloc
6091          * mutex since we won't race with anybody.  We need this mostly to make
6092          * lockdep shut its filthy mouth.
6093          *
6094          * If we have a transaction open (can happen if we call truncate_block
6095          * from truncate), then we need FLUSH_LIMIT so we don't deadlock.
6096          */
6097         if (btrfs_is_free_space_inode(inode)) {
6098                 flush = BTRFS_RESERVE_NO_FLUSH;
6099                 delalloc_lock = false;
6100         } else {
6101                 if (current->journal_info)
6102                         flush = BTRFS_RESERVE_FLUSH_LIMIT;
6103
6104                 if (btrfs_transaction_in_commit(fs_info))
6105                         schedule_timeout(1);
6106         }
6107
6108         if (delalloc_lock)
6109                 mutex_lock(&inode->delalloc_mutex);
6110
6111         num_bytes = ALIGN(num_bytes, fs_info->sectorsize);
6112
6113         /* Add our new extents and calculate the new rsv size. */
6114         spin_lock(&inode->lock);
6115         nr_extents = count_max_extents(num_bytes);
6116         btrfs_mod_outstanding_extents(inode, nr_extents);
6117         inode->csum_bytes += num_bytes;
6118         btrfs_calculate_inode_block_rsv_size(fs_info, inode);
6119         spin_unlock(&inode->lock);
6120
6121         ret = btrfs_inode_rsv_refill(inode, flush);
6122         if (unlikely(ret))
6123                 goto out_fail;
6124
6125         if (delalloc_lock)
6126                 mutex_unlock(&inode->delalloc_mutex);
6127         return 0;
6128
6129 out_fail:
6130         spin_lock(&inode->lock);
6131         nr_extents = count_max_extents(num_bytes);
6132         btrfs_mod_outstanding_extents(inode, -nr_extents);
6133         inode->csum_bytes -= num_bytes;
6134         btrfs_calculate_inode_block_rsv_size(fs_info, inode);
6135         spin_unlock(&inode->lock);
6136
6137         btrfs_inode_rsv_release(inode, true);
6138         if (delalloc_lock)
6139                 mutex_unlock(&inode->delalloc_mutex);
6140         return ret;
6141 }
6142
6143 /**
6144  * btrfs_delalloc_release_metadata - release a metadata reservation for an inode
6145  * @inode: the inode to release the reservation for.
6146  * @num_bytes: the number of bytes we are releasing.
6147  * @qgroup_free: free qgroup reservation or convert it to per-trans reservation
6148  *
6149  * This will release the metadata reservation for an inode.  This can be called
6150  * once we complete IO for a given set of bytes to release their metadata
6151  * reservations, or on error for the same reason.
6152  */
6153 void btrfs_delalloc_release_metadata(struct btrfs_inode *inode, u64 num_bytes,
6154                                      bool qgroup_free)
6155 {
6156         struct btrfs_fs_info *fs_info = btrfs_sb(inode->vfs_inode.i_sb);
6157
6158         num_bytes = ALIGN(num_bytes, fs_info->sectorsize);
6159         spin_lock(&inode->lock);
6160         inode->csum_bytes -= num_bytes;
6161         btrfs_calculate_inode_block_rsv_size(fs_info, inode);
6162         spin_unlock(&inode->lock);
6163
6164         if (btrfs_is_testing(fs_info))
6165                 return;
6166
6167         btrfs_inode_rsv_release(inode, qgroup_free);
6168 }
6169
6170 /**
6171  * btrfs_delalloc_release_extents - release our outstanding_extents
6172  * @inode: the inode to balance the reservation for.
6173  * @num_bytes: the number of bytes we originally reserved with
6174  * @qgroup_free: do we need to free qgroup meta reservation or convert them.
6175  *
6176  * When we reserve space we increase outstanding_extents for the extents we may
6177  * add.  Once we've set the range as delalloc or created our ordered extents we
6178  * have outstanding_extents to track the real usage, so we use this to free our
6179  * temporarily tracked outstanding_extents.  This _must_ be used in conjunction
6180  * with btrfs_delalloc_reserve_metadata.
6181  */
6182 void btrfs_delalloc_release_extents(struct btrfs_inode *inode, u64 num_bytes,
6183                                     bool qgroup_free)
6184 {
6185         struct btrfs_fs_info *fs_info = btrfs_sb(inode->vfs_inode.i_sb);
6186         unsigned num_extents;
6187
6188         spin_lock(&inode->lock);
6189         num_extents = count_max_extents(num_bytes);
6190         btrfs_mod_outstanding_extents(inode, -num_extents);
6191         btrfs_calculate_inode_block_rsv_size(fs_info, inode);
6192         spin_unlock(&inode->lock);
6193
6194         if (btrfs_is_testing(fs_info))
6195                 return;
6196
6197         btrfs_inode_rsv_release(inode, qgroup_free);
6198 }
6199
6200 /**
6201  * btrfs_delalloc_reserve_space - reserve data and metadata space for
6202  * delalloc
6203  * @inode: inode we're writing to
6204  * @start: start range we are writing to
6205  * @len: how long the range we are writing to
6206  * @reserved: mandatory parameter, record actually reserved qgroup ranges of
6207  *            current reservation.
6208  *
6209  * This will do the following things
6210  *
6211  * o reserve space in data space info for num bytes
6212  *   and reserve precious corresponding qgroup space
6213  *   (Done in check_data_free_space)
6214  *
6215  * o reserve space for metadata space, based on the number of outstanding
6216  *   extents and how much csums will be needed
6217  *   also reserve metadata space in a per root over-reserve method.
6218  * o add to the inodes->delalloc_bytes
6219  * o add it to the fs_info's delalloc inodes list.
6220  *   (Above 3 all done in delalloc_reserve_metadata)
6221  *
6222  * Return 0 for success
6223  * Return <0 for error(-ENOSPC or -EQUOT)
6224  */
6225 int btrfs_delalloc_reserve_space(struct inode *inode,
6226                         struct extent_changeset **reserved, u64 start, u64 len)
6227 {
6228         int ret;
6229
6230         ret = btrfs_check_data_free_space(inode, reserved, start, len);
6231         if (ret < 0)
6232                 return ret;
6233         ret = btrfs_delalloc_reserve_metadata(BTRFS_I(inode), len);
6234         if (ret < 0)
6235                 btrfs_free_reserved_data_space(inode, *reserved, start, len);
6236         return ret;
6237 }
6238
6239 /**
6240  * btrfs_delalloc_release_space - release data and metadata space for delalloc
6241  * @inode: inode we're releasing space for
6242  * @start: start position of the space already reserved
6243  * @len: the len of the space already reserved
6244  * @release_bytes: the len of the space we consumed or didn't use
6245  *
6246  * This function will release the metadata space that was not used and will
6247  * decrement ->delalloc_bytes and remove it from the fs_info delalloc_inodes
6248  * list if there are no delalloc bytes left.
6249  * Also it will handle the qgroup reserved space.
6250  */
6251 void btrfs_delalloc_release_space(struct inode *inode,
6252                                   struct extent_changeset *reserved,
6253                                   u64 start, u64 len, bool qgroup_free)
6254 {
6255         btrfs_delalloc_release_metadata(BTRFS_I(inode), len, qgroup_free);
6256         btrfs_free_reserved_data_space(inode, reserved, start, len);
6257 }
6258
6259 static int update_block_group(struct btrfs_trans_handle *trans,
6260                               struct btrfs_fs_info *info, u64 bytenr,
6261                               u64 num_bytes, int alloc)
6262 {
6263         struct btrfs_block_group_cache *cache = NULL;
6264         u64 total = num_bytes;
6265         u64 old_val;
6266         u64 byte_in_group;
6267         int factor;
6268
6269         /* block accounting for super block */
6270         spin_lock(&info->delalloc_root_lock);
6271         old_val = btrfs_super_bytes_used(info->super_copy);
6272         if (alloc)
6273                 old_val += num_bytes;
6274         else
6275                 old_val -= num_bytes;
6276         btrfs_set_super_bytes_used(info->super_copy, old_val);
6277         spin_unlock(&info->delalloc_root_lock);
6278
6279         while (total) {
6280                 cache = btrfs_lookup_block_group(info, bytenr);
6281                 if (!cache)
6282                         return -ENOENT;
6283                 if (cache->flags & (BTRFS_BLOCK_GROUP_DUP |
6284                                     BTRFS_BLOCK_GROUP_RAID1 |
6285                                     BTRFS_BLOCK_GROUP_RAID10))
6286                         factor = 2;
6287                 else
6288                         factor = 1;
6289                 /*
6290                  * If this block group has free space cache written out, we
6291                  * need to make sure to load it if we are removing space.  This
6292                  * is because we need the unpinning stage to actually add the
6293                  * space back to the block group, otherwise we will leak space.
6294                  */
6295                 if (!alloc && cache->cached == BTRFS_CACHE_NO)
6296                         cache_block_group(cache, 1);
6297
6298                 byte_in_group = bytenr - cache->key.objectid;
6299                 WARN_ON(byte_in_group > cache->key.offset);
6300
6301                 spin_lock(&cache->space_info->lock);
6302                 spin_lock(&cache->lock);
6303
6304                 if (btrfs_test_opt(info, SPACE_CACHE) &&
6305                     cache->disk_cache_state < BTRFS_DC_CLEAR)
6306                         cache->disk_cache_state = BTRFS_DC_CLEAR;
6307
6308                 old_val = btrfs_block_group_used(&cache->item);
6309                 num_bytes = min(total, cache->key.offset - byte_in_group);
6310                 if (alloc) {
6311                         old_val += num_bytes;
6312                         btrfs_set_block_group_used(&cache->item, old_val);
6313                         cache->reserved -= num_bytes;
6314                         cache->space_info->bytes_reserved -= num_bytes;
6315                         cache->space_info->bytes_used += num_bytes;
6316                         cache->space_info->disk_used += num_bytes * factor;
6317                         spin_unlock(&cache->lock);
6318                         spin_unlock(&cache->space_info->lock);
6319                 } else {
6320                         old_val -= num_bytes;
6321                         btrfs_set_block_group_used(&cache->item, old_val);
6322                         cache->pinned += num_bytes;
6323                         cache->space_info->bytes_pinned += num_bytes;
6324                         cache->space_info->bytes_used -= num_bytes;
6325                         cache->space_info->disk_used -= num_bytes * factor;
6326                         spin_unlock(&cache->lock);
6327                         spin_unlock(&cache->space_info->lock);
6328
6329                         trace_btrfs_space_reservation(info, "pinned",
6330                                                       cache->space_info->flags,
6331                                                       num_bytes, 1);
6332                         percpu_counter_add(&cache->space_info->total_bytes_pinned,
6333                                            num_bytes);
6334                         set_extent_dirty(info->pinned_extents,
6335                                          bytenr, bytenr + num_bytes - 1,
6336                                          GFP_NOFS | __GFP_NOFAIL);
6337                 }
6338
6339                 spin_lock(&trans->transaction->dirty_bgs_lock);
6340                 if (list_empty(&cache->dirty_list)) {
6341                         list_add_tail(&cache->dirty_list,
6342                                       &trans->transaction->dirty_bgs);
6343                                 trans->transaction->num_dirty_bgs++;
6344                         btrfs_get_block_group(cache);
6345                 }
6346                 spin_unlock(&trans->transaction->dirty_bgs_lock);
6347
6348                 /*
6349                  * No longer have used bytes in this block group, queue it for
6350                  * deletion. We do this after adding the block group to the
6351                  * dirty list to avoid races between cleaner kthread and space
6352                  * cache writeout.
6353                  */
6354                 if (!alloc && old_val == 0) {
6355                         spin_lock(&info->unused_bgs_lock);
6356                         if (list_empty(&cache->bg_list)) {
6357                                 btrfs_get_block_group(cache);
6358                                 list_add_tail(&cache->bg_list,
6359                                               &info->unused_bgs);
6360                         }
6361                         spin_unlock(&info->unused_bgs_lock);
6362                 }
6363
6364                 btrfs_put_block_group(cache);
6365                 total -= num_bytes;
6366                 bytenr += num_bytes;
6367         }
6368         return 0;
6369 }
6370
6371 static u64 first_logical_byte(struct btrfs_fs_info *fs_info, u64 search_start)
6372 {
6373         struct btrfs_block_group_cache *cache;
6374         u64 bytenr;
6375
6376         spin_lock(&fs_info->block_group_cache_lock);
6377         bytenr = fs_info->first_logical_byte;
6378         spin_unlock(&fs_info->block_group_cache_lock);
6379
6380         if (bytenr < (u64)-1)
6381                 return bytenr;
6382
6383         cache = btrfs_lookup_first_block_group(fs_info, search_start);
6384         if (!cache)
6385                 return 0;
6386
6387         bytenr = cache->key.objectid;
6388         btrfs_put_block_group(cache);
6389
6390         return bytenr;
6391 }
6392
6393 static int pin_down_extent(struct btrfs_fs_info *fs_info,
6394                            struct btrfs_block_group_cache *cache,
6395                            u64 bytenr, u64 num_bytes, int reserved)
6396 {
6397         spin_lock(&cache->space_info->lock);
6398         spin_lock(&cache->lock);
6399         cache->pinned += num_bytes;
6400         cache->space_info->bytes_pinned += num_bytes;
6401         if (reserved) {
6402                 cache->reserved -= num_bytes;
6403                 cache->space_info->bytes_reserved -= num_bytes;
6404         }
6405         spin_unlock(&cache->lock);
6406         spin_unlock(&cache->space_info->lock);
6407
6408         trace_btrfs_space_reservation(fs_info, "pinned",
6409                                       cache->space_info->flags, num_bytes, 1);
6410         percpu_counter_add(&cache->space_info->total_bytes_pinned, num_bytes);
6411         set_extent_dirty(fs_info->pinned_extents, bytenr,
6412                          bytenr + num_bytes - 1, GFP_NOFS | __GFP_NOFAIL);
6413         return 0;
6414 }
6415
6416 /*
6417  * this function must be called within transaction
6418  */
6419 int btrfs_pin_extent(struct btrfs_fs_info *fs_info,
6420                      u64 bytenr, u64 num_bytes, int reserved)
6421 {
6422         struct btrfs_block_group_cache *cache;
6423
6424         cache = btrfs_lookup_block_group(fs_info, bytenr);
6425         BUG_ON(!cache); /* Logic error */
6426
6427         pin_down_extent(fs_info, cache, bytenr, num_bytes, reserved);
6428
6429         btrfs_put_block_group(cache);
6430         return 0;
6431 }
6432
6433 /*
6434  * this function must be called within transaction
6435  */
6436 int btrfs_pin_extent_for_log_replay(struct btrfs_fs_info *fs_info,
6437                                     u64 bytenr, u64 num_bytes)
6438 {
6439         struct btrfs_block_group_cache *cache;
6440         int ret;
6441
6442         cache = btrfs_lookup_block_group(fs_info, bytenr);
6443         if (!cache)
6444                 return -EINVAL;
6445
6446         /*
6447          * pull in the free space cache (if any) so that our pin
6448          * removes the free space from the cache.  We have load_only set
6449          * to one because the slow code to read in the free extents does check
6450          * the pinned extents.
6451          */
6452         cache_block_group(cache, 1);
6453
6454         pin_down_extent(fs_info, cache, bytenr, num_bytes, 0);
6455
6456         /* remove us from the free space cache (if we're there at all) */
6457         ret = btrfs_remove_free_space(cache, bytenr, num_bytes);
6458         btrfs_put_block_group(cache);
6459         return ret;
6460 }
6461
6462 static int __exclude_logged_extent(struct btrfs_fs_info *fs_info,
6463                                    u64 start, u64 num_bytes)
6464 {
6465         int ret;
6466         struct btrfs_block_group_cache *block_group;
6467         struct btrfs_caching_control *caching_ctl;
6468
6469         block_group = btrfs_lookup_block_group(fs_info, start);
6470         if (!block_group)
6471                 return -EINVAL;
6472
6473         cache_block_group(block_group, 0);
6474         caching_ctl = get_caching_control(block_group);
6475
6476         if (!caching_ctl) {
6477                 /* Logic error */
6478                 BUG_ON(!block_group_cache_done(block_group));
6479                 ret = btrfs_remove_free_space(block_group, start, num_bytes);
6480         } else {
6481                 mutex_lock(&caching_ctl->mutex);
6482
6483                 if (start >= caching_ctl->progress) {
6484                         ret = add_excluded_extent(fs_info, start, num_bytes);
6485                 } else if (start + num_bytes <= caching_ctl->progress) {
6486                         ret = btrfs_remove_free_space(block_group,
6487                                                       start, num_bytes);
6488                 } else {
6489                         num_bytes = caching_ctl->progress - start;
6490                         ret = btrfs_remove_free_space(block_group,
6491                                                       start, num_bytes);
6492                         if (ret)
6493                                 goto out_lock;
6494
6495                         num_bytes = (start + num_bytes) -
6496                                 caching_ctl->progress;
6497                         start = caching_ctl->progress;
6498                         ret = add_excluded_extent(fs_info, start, num_bytes);
6499                 }
6500 out_lock:
6501                 mutex_unlock(&caching_ctl->mutex);
6502                 put_caching_control(caching_ctl);
6503         }
6504         btrfs_put_block_group(block_group);
6505         return ret;
6506 }
6507
6508 int btrfs_exclude_logged_extents(struct btrfs_fs_info *fs_info,
6509                                  struct extent_buffer *eb)
6510 {
6511         struct btrfs_file_extent_item *item;
6512         struct btrfs_key key;
6513         int found_type;
6514         int i;
6515
6516         if (!btrfs_fs_incompat(fs_info, MIXED_GROUPS))
6517                 return 0;
6518
6519         for (i = 0; i < btrfs_header_nritems(eb); i++) {
6520                 btrfs_item_key_to_cpu(eb, &key, i);
6521                 if (key.type != BTRFS_EXTENT_DATA_KEY)
6522                         continue;
6523                 item = btrfs_item_ptr(eb, i, struct btrfs_file_extent_item);
6524                 found_type = btrfs_file_extent_type(eb, item);
6525                 if (found_type == BTRFS_FILE_EXTENT_INLINE)
6526                         continue;
6527                 if (btrfs_file_extent_disk_bytenr(eb, item) == 0)
6528                         continue;
6529                 key.objectid = btrfs_file_extent_disk_bytenr(eb, item);
6530                 key.offset = btrfs_file_extent_disk_num_bytes(eb, item);
6531                 __exclude_logged_extent(fs_info, key.objectid, key.offset);
6532         }
6533
6534         return 0;
6535 }
6536
6537 static void
6538 btrfs_inc_block_group_reservations(struct btrfs_block_group_cache *bg)
6539 {
6540         atomic_inc(&bg->reservations);
6541 }
6542
6543 void btrfs_dec_block_group_reservations(struct btrfs_fs_info *fs_info,
6544                                         const u64 start)
6545 {
6546         struct btrfs_block_group_cache *bg;
6547
6548         bg = btrfs_lookup_block_group(fs_info, start);
6549         ASSERT(bg);
6550         if (atomic_dec_and_test(&bg->reservations))
6551                 wake_up_atomic_t(&bg->reservations);
6552         btrfs_put_block_group(bg);
6553 }
6554
6555 void btrfs_wait_block_group_reservations(struct btrfs_block_group_cache *bg)
6556 {
6557         struct btrfs_space_info *space_info = bg->space_info;
6558
6559         ASSERT(bg->ro);
6560
6561         if (!(bg->flags & BTRFS_BLOCK_GROUP_DATA))
6562                 return;
6563
6564         /*
6565          * Our block group is read only but before we set it to read only,
6566          * some task might have had allocated an extent from it already, but it
6567          * has not yet created a respective ordered extent (and added it to a
6568          * root's list of ordered extents).
6569          * Therefore wait for any task currently allocating extents, since the
6570          * block group's reservations counter is incremented while a read lock
6571          * on the groups' semaphore is held and decremented after releasing
6572          * the read access on that semaphore and creating the ordered extent.
6573          */
6574         down_write(&space_info->groups_sem);
6575         up_write(&space_info->groups_sem);
6576
6577         wait_on_atomic_t(&bg->reservations, atomic_t_wait,
6578                          TASK_UNINTERRUPTIBLE);
6579 }
6580
6581 /**
6582  * btrfs_add_reserved_bytes - update the block_group and space info counters
6583  * @cache:      The cache we are manipulating
6584  * @ram_bytes:  The number of bytes of file content, and will be same to
6585  *              @num_bytes except for the compress path.
6586  * @num_bytes:  The number of bytes in question
6587  * @delalloc:   The blocks are allocated for the delalloc write
6588  *
6589  * This is called by the allocator when it reserves space. If this is a
6590  * reservation and the block group has become read only we cannot make the
6591  * reservation and return -EAGAIN, otherwise this function always succeeds.
6592  */
6593 static int btrfs_add_reserved_bytes(struct btrfs_block_group_cache *cache,
6594                                     u64 ram_bytes, u64 num_bytes, int delalloc)
6595 {
6596         struct btrfs_space_info *space_info = cache->space_info;
6597         int ret = 0;
6598
6599         spin_lock(&space_info->lock);
6600         spin_lock(&cache->lock);
6601         if (cache->ro) {
6602                 ret = -EAGAIN;
6603         } else {
6604                 cache->reserved += num_bytes;
6605                 space_info->bytes_reserved += num_bytes;
6606
6607                 trace_btrfs_space_reservation(cache->fs_info,
6608                                 "space_info", space_info->flags,
6609                                 ram_bytes, 0);
6610                 space_info->bytes_may_use -= ram_bytes;
6611                 if (delalloc)
6612                         cache->delalloc_bytes += num_bytes;
6613         }
6614         spin_unlock(&cache->lock);
6615         spin_unlock(&space_info->lock);
6616         return ret;
6617 }
6618
6619 /**
6620  * btrfs_free_reserved_bytes - update the block_group and space info counters
6621  * @cache:      The cache we are manipulating
6622  * @num_bytes:  The number of bytes in question
6623  * @delalloc:   The blocks are allocated for the delalloc write
6624  *
6625  * This is called by somebody who is freeing space that was never actually used
6626  * on disk.  For example if you reserve some space for a new leaf in transaction
6627  * A and before transaction A commits you free that leaf, you call this with
6628  * reserve set to 0 in order to clear the reservation.
6629  */
6630
6631 static int btrfs_free_reserved_bytes(struct btrfs_block_group_cache *cache,
6632                                      u64 num_bytes, int delalloc)
6633 {
6634         struct btrfs_space_info *space_info = cache->space_info;
6635         int ret = 0;
6636
6637         spin_lock(&space_info->lock);
6638         spin_lock(&cache->lock);
6639         if (cache->ro)
6640                 space_info->bytes_readonly += num_bytes;
6641         cache->reserved -= num_bytes;
6642         space_info->bytes_reserved -= num_bytes;
6643
6644         if (delalloc)
6645                 cache->delalloc_bytes -= num_bytes;
6646         spin_unlock(&cache->lock);
6647         spin_unlock(&space_info->lock);
6648         return ret;
6649 }
6650 void btrfs_prepare_extent_commit(struct btrfs_fs_info *fs_info)
6651 {
6652         struct btrfs_caching_control *next;
6653         struct btrfs_caching_control *caching_ctl;
6654         struct btrfs_block_group_cache *cache;
6655
6656         down_write(&fs_info->commit_root_sem);
6657
6658         list_for_each_entry_safe(caching_ctl, next,
6659                                  &fs_info->caching_block_groups, list) {
6660                 cache = caching_ctl->block_group;
6661                 if (block_group_cache_done(cache)) {
6662                         cache->last_byte_to_unpin = (u64)-1;
6663                         list_del_init(&caching_ctl->list);
6664                         put_caching_control(caching_ctl);
6665                 } else {
6666                         cache->last_byte_to_unpin = caching_ctl->progress;
6667                 }
6668         }
6669
6670         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
6671                 fs_info->pinned_extents = &fs_info->freed_extents[1];
6672         else
6673                 fs_info->pinned_extents = &fs_info->freed_extents[0];
6674
6675         up_write(&fs_info->commit_root_sem);
6676
6677         update_global_block_rsv(fs_info);
6678 }
6679
6680 /*
6681  * Returns the free cluster for the given space info and sets empty_cluster to
6682  * what it should be based on the mount options.
6683  */
6684 static struct btrfs_free_cluster *
6685 fetch_cluster_info(struct btrfs_fs_info *fs_info,
6686                    struct btrfs_space_info *space_info, u64 *empty_cluster)
6687 {
6688         struct btrfs_free_cluster *ret = NULL;
6689
6690         *empty_cluster = 0;
6691         if (btrfs_mixed_space_info(space_info))
6692                 return ret;
6693
6694         if (space_info->flags & BTRFS_BLOCK_GROUP_METADATA) {
6695                 ret = &fs_info->meta_alloc_cluster;
6696                 if (btrfs_test_opt(fs_info, SSD))
6697                         *empty_cluster = SZ_2M;
6698                 else
6699                         *empty_cluster = SZ_64K;
6700         } else if ((space_info->flags & BTRFS_BLOCK_GROUP_DATA) &&
6701                    btrfs_test_opt(fs_info, SSD_SPREAD)) {
6702                 *empty_cluster = SZ_2M;
6703                 ret = &fs_info->data_alloc_cluster;
6704         }
6705
6706         return ret;
6707 }
6708
6709 static int unpin_extent_range(struct btrfs_fs_info *fs_info,
6710                               u64 start, u64 end,
6711                               const bool return_free_space)
6712 {
6713         struct btrfs_block_group_cache *cache = NULL;
6714         struct btrfs_space_info *space_info;
6715         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
6716         struct btrfs_free_cluster *cluster = NULL;
6717         u64 len;
6718         u64 total_unpinned = 0;
6719         u64 empty_cluster = 0;
6720         bool readonly;
6721
6722         while (start <= end) {
6723                 readonly = false;
6724                 if (!cache ||
6725                     start >= cache->key.objectid + cache->key.offset) {
6726                         if (cache)
6727                                 btrfs_put_block_group(cache);
6728                         total_unpinned = 0;
6729                         cache = btrfs_lookup_block_group(fs_info, start);
6730                         BUG_ON(!cache); /* Logic error */
6731
6732                         cluster = fetch_cluster_info(fs_info,
6733                                                      cache->space_info,
6734                                                      &empty_cluster);
6735                         empty_cluster <<= 1;
6736                 }
6737
6738                 len = cache->key.objectid + cache->key.offset - start;
6739                 len = min(len, end + 1 - start);
6740
6741                 if (start < cache->last_byte_to_unpin) {
6742                         len = min(len, cache->last_byte_to_unpin - start);
6743                         if (return_free_space)
6744                                 btrfs_add_free_space(cache, start, len);
6745                 }
6746
6747                 start += len;
6748                 total_unpinned += len;
6749                 space_info = cache->space_info;
6750
6751                 /*
6752                  * If this space cluster has been marked as fragmented and we've
6753                  * unpinned enough in this block group to potentially allow a
6754                  * cluster to be created inside of it go ahead and clear the
6755                  * fragmented check.
6756                  */
6757                 if (cluster && cluster->fragmented &&
6758                     total_unpinned > empty_cluster) {
6759                         spin_lock(&cluster->lock);
6760                         cluster->fragmented = 0;
6761                         spin_unlock(&cluster->lock);
6762                 }
6763
6764                 spin_lock(&space_info->lock);
6765                 spin_lock(&cache->lock);
6766                 cache->pinned -= len;
6767                 space_info->bytes_pinned -= len;
6768
6769                 trace_btrfs_space_reservation(fs_info, "pinned",
6770                                               space_info->flags, len, 0);
6771                 space_info->max_extent_size = 0;
6772                 percpu_counter_add(&space_info->total_bytes_pinned, -len);
6773                 if (cache->ro) {
6774                         space_info->bytes_readonly += len;
6775                         readonly = true;
6776                 }
6777                 spin_unlock(&cache->lock);
6778                 if (!readonly && return_free_space &&
6779                     global_rsv->space_info == space_info) {
6780                         u64 to_add = len;
6781
6782                         spin_lock(&global_rsv->lock);
6783                         if (!global_rsv->full) {
6784                                 to_add = min(len, global_rsv->size -
6785                                              global_rsv->reserved);
6786                                 global_rsv->reserved += to_add;
6787                                 space_info->bytes_may_use += to_add;
6788                                 if (global_rsv->reserved >= global_rsv->size)
6789                                         global_rsv->full = 1;
6790                                 trace_btrfs_space_reservation(fs_info,
6791                                                               "space_info",
6792                                                               space_info->flags,
6793                                                               to_add, 1);
6794                                 len -= to_add;
6795                         }
6796                         spin_unlock(&global_rsv->lock);
6797                         /* Add to any tickets we may have */
6798                         if (len)
6799                                 space_info_add_new_bytes(fs_info, space_info,
6800                                                          len);
6801                 }
6802                 spin_unlock(&space_info->lock);
6803         }
6804
6805         if (cache)
6806                 btrfs_put_block_group(cache);
6807         return 0;
6808 }
6809
6810 int btrfs_finish_extent_commit(struct btrfs_trans_handle *trans)
6811 {
6812         struct btrfs_fs_info *fs_info = trans->fs_info;
6813         struct btrfs_block_group_cache *block_group, *tmp;
6814         struct list_head *deleted_bgs;
6815         struct extent_io_tree *unpin;
6816         u64 start;
6817         u64 end;
6818         int ret;
6819
6820         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
6821                 unpin = &fs_info->freed_extents[1];
6822         else
6823                 unpin = &fs_info->freed_extents[0];
6824
6825         while (!trans->aborted) {
6826                 mutex_lock(&fs_info->unused_bg_unpin_mutex);
6827                 ret = find_first_extent_bit(unpin, 0, &start, &end,
6828                                             EXTENT_DIRTY, NULL);
6829                 if (ret) {
6830                         mutex_unlock(&fs_info->unused_bg_unpin_mutex);
6831                         break;
6832                 }
6833
6834                 if (btrfs_test_opt(fs_info, DISCARD))
6835                         ret = btrfs_discard_extent(fs_info, start,
6836                                                    end + 1 - start, NULL);
6837
6838                 clear_extent_dirty(unpin, start, end);
6839                 unpin_extent_range(fs_info, start, end, true);
6840                 mutex_unlock(&fs_info->unused_bg_unpin_mutex);
6841                 cond_resched();
6842         }
6843
6844         /*
6845          * Transaction is finished.  We don't need the lock anymore.  We
6846          * do need to clean up the block groups in case of a transaction
6847          * abort.
6848          */
6849         deleted_bgs = &trans->transaction->deleted_bgs;
6850         list_for_each_entry_safe(block_group, tmp, deleted_bgs, bg_list) {
6851                 u64 trimmed = 0;
6852
6853                 ret = -EROFS;
6854                 if (!trans->aborted)
6855                         ret = btrfs_discard_extent(fs_info,
6856                                                    block_group->key.objectid,
6857                                                    block_group->key.offset,
6858                                                    &trimmed);
6859
6860                 list_del_init(&block_group->bg_list);
6861                 btrfs_put_block_group_trimming(block_group);
6862                 btrfs_put_block_group(block_group);
6863
6864                 if (ret) {
6865                         const char *errstr = btrfs_decode_error(ret);
6866                         btrfs_warn(fs_info,
6867                            "discard failed while removing blockgroup: errno=%d %s",
6868                                    ret, errstr);
6869                 }
6870         }
6871
6872         return 0;
6873 }
6874
6875 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
6876                                 struct btrfs_fs_info *info,
6877                                 struct btrfs_delayed_ref_node *node, u64 parent,
6878                                 u64 root_objectid, u64 owner_objectid,
6879                                 u64 owner_offset, int refs_to_drop,
6880                                 struct btrfs_delayed_extent_op *extent_op)
6881 {
6882         struct btrfs_key key;
6883         struct btrfs_path *path;
6884         struct btrfs_root *extent_root = info->extent_root;
6885         struct extent_buffer *leaf;
6886         struct btrfs_extent_item *ei;
6887         struct btrfs_extent_inline_ref *iref;
6888         int ret;
6889         int is_data;
6890         int extent_slot = 0;
6891         int found_extent = 0;
6892         int num_to_del = 1;
6893         u32 item_size;
6894         u64 refs;
6895         u64 bytenr = node->bytenr;
6896         u64 num_bytes = node->num_bytes;
6897         int last_ref = 0;
6898         bool skinny_metadata = btrfs_fs_incompat(info, SKINNY_METADATA);
6899
6900         path = btrfs_alloc_path();
6901         if (!path)
6902                 return -ENOMEM;
6903
6904         path->reada = READA_FORWARD;
6905         path->leave_spinning = 1;
6906
6907         is_data = owner_objectid >= BTRFS_FIRST_FREE_OBJECTID;
6908         BUG_ON(!is_data && refs_to_drop != 1);
6909
6910         if (is_data)
6911                 skinny_metadata = false;
6912
6913         ret = lookup_extent_backref(trans, info, path, &iref,
6914                                     bytenr, num_bytes, parent,
6915                                     root_objectid, owner_objectid,
6916                                     owner_offset);
6917         if (ret == 0) {
6918                 extent_slot = path->slots[0];
6919                 while (extent_slot >= 0) {
6920                         btrfs_item_key_to_cpu(path->nodes[0], &key,
6921                                               extent_slot);
6922                         if (key.objectid != bytenr)
6923                                 break;
6924                         if (key.type == BTRFS_EXTENT_ITEM_KEY &&
6925                             key.offset == num_bytes) {
6926                                 found_extent = 1;
6927                                 break;
6928                         }
6929                         if (key.type == BTRFS_METADATA_ITEM_KEY &&
6930                             key.offset == owner_objectid) {
6931                                 found_extent = 1;
6932                                 break;
6933                         }
6934                         if (path->slots[0] - extent_slot > 5)
6935                                 break;
6936                         extent_slot--;
6937                 }
6938 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
6939                 item_size = btrfs_item_size_nr(path->nodes[0], extent_slot);
6940                 if (found_extent && item_size < sizeof(*ei))
6941                         found_extent = 0;
6942 #endif
6943                 if (!found_extent) {
6944                         BUG_ON(iref);
6945                         ret = remove_extent_backref(trans, info, path, NULL,
6946                                                     refs_to_drop,
6947                                                     is_data, &last_ref);
6948                         if (ret) {
6949                                 btrfs_abort_transaction(trans, ret);
6950                                 goto out;
6951                         }
6952                         btrfs_release_path(path);
6953                         path->leave_spinning = 1;
6954
6955                         key.objectid = bytenr;
6956                         key.type = BTRFS_EXTENT_ITEM_KEY;
6957                         key.offset = num_bytes;
6958
6959                         if (!is_data && skinny_metadata) {
6960                                 key.type = BTRFS_METADATA_ITEM_KEY;
6961                                 key.offset = owner_objectid;
6962                         }
6963
6964                         ret = btrfs_search_slot(trans, extent_root,
6965                                                 &key, path, -1, 1);
6966                         if (ret > 0 && skinny_metadata && path->slots[0]) {
6967                                 /*
6968                                  * Couldn't find our skinny metadata item,
6969                                  * see if we have ye olde extent item.
6970                                  */
6971                                 path->slots[0]--;
6972                                 btrfs_item_key_to_cpu(path->nodes[0], &key,
6973                                                       path->slots[0]);
6974                                 if (key.objectid == bytenr &&
6975                                     key.type == BTRFS_EXTENT_ITEM_KEY &&
6976                                     key.offset == num_bytes)
6977                                         ret = 0;
6978                         }
6979
6980                         if (ret > 0 && skinny_metadata) {
6981                                 skinny_metadata = false;
6982                                 key.objectid = bytenr;
6983                                 key.type = BTRFS_EXTENT_ITEM_KEY;
6984                                 key.offset = num_bytes;
6985                                 btrfs_release_path(path);
6986                                 ret = btrfs_search_slot(trans, extent_root,
6987                                                         &key, path, -1, 1);
6988                         }
6989
6990                         if (ret) {
6991                                 btrfs_err(info,
6992                                           "umm, got %d back from search, was looking for %llu",
6993                                           ret, bytenr);
6994                                 if (ret > 0)
6995                                         btrfs_print_leaf(path->nodes[0]);
6996                         }
6997                         if (ret < 0) {
6998                                 btrfs_abort_transaction(trans, ret);
6999                                 goto out;
7000                         }
7001                         extent_slot = path->slots[0];
7002                 }
7003         } else if (WARN_ON(ret == -ENOENT)) {
7004                 btrfs_print_leaf(path->nodes[0]);
7005                 btrfs_err(info,
7006                         "unable to find ref byte nr %llu parent %llu root %llu  owner %llu offset %llu",
7007                         bytenr, parent, root_objectid, owner_objectid,
7008                         owner_offset);
7009                 btrfs_abort_transaction(trans, ret);
7010                 goto out;
7011         } else {
7012                 btrfs_abort_transaction(trans, ret);
7013                 goto out;
7014         }
7015
7016         leaf = path->nodes[0];
7017         item_size = btrfs_item_size_nr(leaf, extent_slot);
7018 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
7019         if (item_size < sizeof(*ei)) {
7020                 BUG_ON(found_extent || extent_slot != path->slots[0]);
7021                 ret = convert_extent_item_v0(trans, info, path, owner_objectid,
7022                                              0);
7023                 if (ret < 0) {
7024                         btrfs_abort_transaction(trans, ret);
7025                         goto out;
7026                 }
7027
7028                 btrfs_release_path(path);
7029                 path->leave_spinning = 1;
7030
7031                 key.objectid = bytenr;
7032                 key.type = BTRFS_EXTENT_ITEM_KEY;
7033                 key.offset = num_bytes;
7034
7035                 ret = btrfs_search_slot(trans, extent_root, &key, path,
7036                                         -1, 1);
7037                 if (ret) {
7038                         btrfs_err(info,
7039                                   "umm, got %d back from search, was looking for %llu",
7040                                 ret, bytenr);
7041                         btrfs_print_leaf(path->nodes[0]);
7042                 }
7043                 if (ret < 0) {
7044                         btrfs_abort_transaction(trans, ret);
7045                         goto out;
7046                 }
7047
7048                 extent_slot = path->slots[0];
7049                 leaf = path->nodes[0];
7050                 item_size = btrfs_item_size_nr(leaf, extent_slot);
7051         }
7052 #endif
7053         BUG_ON(item_size < sizeof(*ei));
7054         ei = btrfs_item_ptr(leaf, extent_slot,
7055                             struct btrfs_extent_item);
7056         if (owner_objectid < BTRFS_FIRST_FREE_OBJECTID &&
7057             key.type == BTRFS_EXTENT_ITEM_KEY) {
7058                 struct btrfs_tree_block_info *bi;
7059                 BUG_ON(item_size < sizeof(*ei) + sizeof(*bi));
7060                 bi = (struct btrfs_tree_block_info *)(ei + 1);
7061                 WARN_ON(owner_objectid != btrfs_tree_block_level(leaf, bi));
7062         }
7063
7064         refs = btrfs_extent_refs(leaf, ei);
7065         if (refs < refs_to_drop) {
7066                 btrfs_err(info,
7067                           "trying to drop %d refs but we only have %Lu for bytenr %Lu",
7068                           refs_to_drop, refs, bytenr);
7069                 ret = -EINVAL;
7070                 btrfs_abort_transaction(trans, ret);
7071                 goto out;
7072         }
7073         refs -= refs_to_drop;
7074
7075         if (refs > 0) {
7076                 if (extent_op)
7077                         __run_delayed_extent_op(extent_op, leaf, ei);
7078                 /*
7079                  * In the case of inline back ref, reference count will
7080                  * be updated by remove_extent_backref
7081                  */
7082                 if (iref) {
7083                         BUG_ON(!found_extent);
7084                 } else {
7085                         btrfs_set_extent_refs(leaf, ei, refs);
7086                         btrfs_mark_buffer_dirty(leaf);
7087                 }
7088                 if (found_extent) {
7089                         ret = remove_extent_backref(trans, info, path,
7090                                                     iref, refs_to_drop,
7091                                                     is_data, &last_ref);
7092                         if (ret) {
7093                                 btrfs_abort_transaction(trans, ret);
7094                                 goto out;
7095                         }
7096                 }
7097         } else {
7098                 if (found_extent) {
7099                         BUG_ON(is_data && refs_to_drop !=
7100                                extent_data_ref_count(path, iref));
7101                         if (iref) {
7102                                 BUG_ON(path->slots[0] != extent_slot);
7103                         } else {
7104                                 BUG_ON(path->slots[0] != extent_slot + 1);
7105                                 path->slots[0] = extent_slot;
7106                                 num_to_del = 2;
7107                         }
7108                 }
7109
7110                 last_ref = 1;
7111                 ret = btrfs_del_items(trans, extent_root, path, path->slots[0],
7112                                       num_to_del);
7113                 if (ret) {
7114                         btrfs_abort_transaction(trans, ret);
7115                         goto out;
7116                 }
7117                 btrfs_release_path(path);
7118
7119                 if (is_data) {
7120                         ret = btrfs_del_csums(trans, info, bytenr, num_bytes);
7121                         if (ret) {
7122                                 btrfs_abort_transaction(trans, ret);
7123                                 goto out;
7124                         }
7125                 }
7126
7127                 ret = add_to_free_space_tree(trans, info, bytenr, num_bytes);
7128                 if (ret) {
7129                         btrfs_abort_transaction(trans, ret);
7130                         goto out;
7131                 }
7132
7133                 ret = update_block_group(trans, info, bytenr, num_bytes, 0);
7134                 if (ret) {
7135                         btrfs_abort_transaction(trans, ret);
7136                         goto out;
7137                 }
7138         }
7139         btrfs_release_path(path);
7140
7141 out:
7142         btrfs_free_path(path);
7143         return ret;
7144 }
7145
7146 /*
7147  * when we free an block, it is possible (and likely) that we free the last
7148  * delayed ref for that extent as well.  This searches the delayed ref tree for
7149  * a given extent, and if there are no other delayed refs to be processed, it
7150  * removes it from the tree.
7151  */
7152 static noinline int check_ref_cleanup(struct btrfs_trans_handle *trans,
7153                                       u64 bytenr)
7154 {
7155         struct btrfs_delayed_ref_head *head;
7156         struct btrfs_delayed_ref_root *delayed_refs;
7157         int ret = 0;
7158
7159         delayed_refs = &trans->transaction->delayed_refs;
7160         spin_lock(&delayed_refs->lock);
7161         head = btrfs_find_delayed_ref_head(delayed_refs, bytenr);
7162         if (!head)
7163                 goto out_delayed_unlock;
7164
7165         spin_lock(&head->lock);
7166         if (!RB_EMPTY_ROOT(&head->ref_tree))
7167                 goto out;
7168
7169         if (head->extent_op) {
7170                 if (!head->must_insert_reserved)
7171                         goto out;
7172                 btrfs_free_delayed_extent_op(head->extent_op);
7173                 head->extent_op = NULL;
7174         }
7175
7176         /*
7177          * waiting for the lock here would deadlock.  If someone else has it
7178          * locked they are already in the process of dropping it anyway
7179          */
7180         if (!mutex_trylock(&head->mutex))
7181                 goto out;
7182
7183         /*
7184          * at this point we have a head with no other entries.  Go
7185          * ahead and process it.
7186          */
7187         rb_erase(&head->href_node, &delayed_refs->href_root);
7188         RB_CLEAR_NODE(&head->href_node);
7189         atomic_dec(&delayed_refs->num_entries);
7190
7191         /*
7192          * we don't take a ref on the node because we're removing it from the
7193          * tree, so we just steal the ref the tree was holding.
7194          */
7195         delayed_refs->num_heads--;
7196         if (head->processing == 0)
7197                 delayed_refs->num_heads_ready--;
7198         head->processing = 0;
7199         spin_unlock(&head->lock);
7200         spin_unlock(&delayed_refs->lock);
7201
7202         BUG_ON(head->extent_op);
7203         if (head->must_insert_reserved)
7204                 ret = 1;
7205
7206         mutex_unlock(&head->mutex);
7207         btrfs_put_delayed_ref_head(head);
7208         return ret;
7209 out:
7210         spin_unlock(&head->lock);
7211
7212 out_delayed_unlock:
7213         spin_unlock(&delayed_refs->lock);
7214         return 0;
7215 }
7216
7217 void btrfs_free_tree_block(struct btrfs_trans_handle *trans,
7218                            struct btrfs_root *root,
7219                            struct extent_buffer *buf,
7220                            u64 parent, int last_ref)
7221 {
7222         struct btrfs_fs_info *fs_info = root->fs_info;
7223         int pin = 1;
7224         int ret;
7225
7226         if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
7227                 int old_ref_mod, new_ref_mod;
7228
7229                 btrfs_ref_tree_mod(root, buf->start, buf->len, parent,
7230                                    root->root_key.objectid,
7231                                    btrfs_header_level(buf), 0,
7232                                    BTRFS_DROP_DELAYED_REF);
7233                 ret = btrfs_add_delayed_tree_ref(fs_info, trans, buf->start,
7234                                                  buf->len, parent,
7235                                                  root->root_key.objectid,
7236                                                  btrfs_header_level(buf),
7237                                                  BTRFS_DROP_DELAYED_REF, NULL,
7238                                                  &old_ref_mod, &new_ref_mod);
7239                 BUG_ON(ret); /* -ENOMEM */
7240                 pin = old_ref_mod >= 0 && new_ref_mod < 0;
7241         }
7242
7243         if (last_ref && btrfs_header_generation(buf) == trans->transid) {
7244                 struct btrfs_block_group_cache *cache;
7245
7246                 if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
7247                         ret = check_ref_cleanup(trans, buf->start);
7248                         if (!ret)
7249                                 goto out;
7250                 }
7251
7252                 pin = 0;
7253                 cache = btrfs_lookup_block_group(fs_info, buf->start);
7254
7255                 if (btrfs_header_flag(buf, BTRFS_HEADER_FLAG_WRITTEN)) {
7256                         pin_down_extent(fs_info, cache, buf->start,
7257                                         buf->len, 1);
7258                         btrfs_put_block_group(cache);
7259                         goto out;
7260                 }
7261
7262                 WARN_ON(test_bit(EXTENT_BUFFER_DIRTY, &buf->bflags));
7263
7264                 btrfs_add_free_space(cache, buf->start, buf->len);
7265                 btrfs_free_reserved_bytes(cache, buf->len, 0);
7266                 btrfs_put_block_group(cache);
7267                 trace_btrfs_reserved_extent_free(fs_info, buf->start, buf->len);
7268         }
7269 out:
7270         if (pin)
7271                 add_pinned_bytes(fs_info, buf->len, btrfs_header_level(buf),
7272                                  root->root_key.objectid);
7273
7274         if (last_ref) {
7275                 /*
7276                  * Deleting the buffer, clear the corrupt flag since it doesn't
7277                  * matter anymore.
7278                  */
7279                 clear_bit(EXTENT_BUFFER_CORRUPT, &buf->bflags);
7280         }
7281 }
7282
7283 /* Can return -ENOMEM */
7284 int btrfs_free_extent(struct btrfs_trans_handle *trans,
7285                       struct btrfs_root *root,
7286                       u64 bytenr, u64 num_bytes, u64 parent, u64 root_objectid,
7287                       u64 owner, u64 offset)
7288 {
7289         struct btrfs_fs_info *fs_info = root->fs_info;
7290         int old_ref_mod, new_ref_mod;
7291         int ret;
7292
7293         if (btrfs_is_testing(fs_info))
7294                 return 0;
7295
7296         if (root_objectid != BTRFS_TREE_LOG_OBJECTID)
7297                 btrfs_ref_tree_mod(root, bytenr, num_bytes, parent,
7298                                    root_objectid, owner, offset,
7299                                    BTRFS_DROP_DELAYED_REF);
7300
7301         /*
7302          * tree log blocks never actually go into the extent allocation
7303          * tree, just update pinning info and exit early.
7304          */
7305         if (root_objectid == BTRFS_TREE_LOG_OBJECTID) {
7306                 WARN_ON(owner >= BTRFS_FIRST_FREE_OBJECTID);
7307                 /* unlocks the pinned mutex */
7308                 btrfs_pin_extent(fs_info, bytenr, num_bytes, 1);
7309                 old_ref_mod = new_ref_mod = 0;
7310                 ret = 0;
7311         } else if (owner < BTRFS_FIRST_FREE_OBJECTID) {
7312                 ret = btrfs_add_delayed_tree_ref(fs_info, trans, bytenr,
7313                                                  num_bytes, parent,
7314                                                  root_objectid, (int)owner,
7315                                                  BTRFS_DROP_DELAYED_REF, NULL,
7316                                                  &old_ref_mod, &new_ref_mod);
7317         } else {
7318                 ret = btrfs_add_delayed_data_ref(fs_info, trans, bytenr,
7319                                                  num_bytes, parent,
7320                                                  root_objectid, owner, offset,
7321                                                  0, BTRFS_DROP_DELAYED_REF,
7322                                                  &old_ref_mod, &new_ref_mod);
7323         }
7324
7325         if (ret == 0 && old_ref_mod >= 0 && new_ref_mod < 0)
7326                 add_pinned_bytes(fs_info, num_bytes, owner, root_objectid);
7327
7328         return ret;
7329 }
7330
7331 /*
7332  * when we wait for progress in the block group caching, its because
7333  * our allocation attempt failed at least once.  So, we must sleep
7334  * and let some progress happen before we try again.
7335  *
7336  * This function will sleep at least once waiting for new free space to
7337  * show up, and then it will check the block group free space numbers
7338  * for our min num_bytes.  Another option is to have it go ahead
7339  * and look in the rbtree for a free extent of a given size, but this
7340  * is a good start.
7341  *
7342  * Callers of this must check if cache->cached == BTRFS_CACHE_ERROR before using
7343  * any of the information in this block group.
7344  */
7345 static noinline void
7346 wait_block_group_cache_progress(struct btrfs_block_group_cache *cache,
7347                                 u64 num_bytes)
7348 {
7349         struct btrfs_caching_control *caching_ctl;
7350
7351         caching_ctl = get_caching_control(cache);
7352         if (!caching_ctl)
7353                 return;
7354
7355         wait_event(caching_ctl->wait, block_group_cache_done(cache) ||
7356                    (cache->free_space_ctl->free_space >= num_bytes));
7357
7358         put_caching_control(caching_ctl);
7359 }
7360
7361 static noinline int
7362 wait_block_group_cache_done(struct btrfs_block_group_cache *cache)
7363 {
7364         struct btrfs_caching_control *caching_ctl;
7365         int ret = 0;
7366
7367         caching_ctl = get_caching_control(cache);
7368         if (!caching_ctl)
7369                 return (cache->cached == BTRFS_CACHE_ERROR) ? -EIO : 0;
7370
7371         wait_event(caching_ctl->wait, block_group_cache_done(cache));
7372         if (cache->cached == BTRFS_CACHE_ERROR)
7373                 ret = -EIO;
7374         put_caching_control(caching_ctl);
7375         return ret;
7376 }
7377
7378 static const char *btrfs_raid_type_names[BTRFS_NR_RAID_TYPES] = {
7379         [BTRFS_RAID_RAID10]     = "raid10",
7380         [BTRFS_RAID_RAID1]      = "raid1",
7381         [BTRFS_RAID_DUP]        = "dup",
7382         [BTRFS_RAID_RAID0]      = "raid0",
7383         [BTRFS_RAID_SINGLE]     = "single",
7384         [BTRFS_RAID_RAID5]      = "raid5",
7385         [BTRFS_RAID_RAID6]      = "raid6",
7386 };
7387
7388 static const char *get_raid_name(enum btrfs_raid_types type)
7389 {
7390         if (type >= BTRFS_NR_RAID_TYPES)
7391                 return NULL;
7392
7393         return btrfs_raid_type_names[type];
7394 }
7395
7396 enum btrfs_loop_type {
7397         LOOP_CACHING_NOWAIT = 0,
7398         LOOP_CACHING_WAIT = 1,
7399         LOOP_ALLOC_CHUNK = 2,
7400         LOOP_NO_EMPTY_SIZE = 3,
7401 };
7402
7403 static inline void
7404 btrfs_lock_block_group(struct btrfs_block_group_cache *cache,
7405                        int delalloc)
7406 {
7407         if (delalloc)
7408                 down_read(&cache->data_rwsem);
7409 }
7410
7411 static inline void
7412 btrfs_grab_block_group(struct btrfs_block_group_cache *cache,
7413                        int delalloc)
7414 {
7415         btrfs_get_block_group(cache);
7416         if (delalloc)
7417                 down_read(&cache->data_rwsem);
7418 }
7419
7420 static struct btrfs_block_group_cache *
7421 btrfs_lock_cluster(struct btrfs_block_group_cache *block_group,
7422                    struct btrfs_free_cluster *cluster,
7423                    int delalloc)
7424 {
7425         struct btrfs_block_group_cache *used_bg = NULL;
7426
7427         spin_lock(&cluster->refill_lock);
7428         while (1) {
7429                 used_bg = cluster->block_group;
7430                 if (!used_bg)
7431                         return NULL;
7432
7433                 if (used_bg == block_group)
7434                         return used_bg;
7435
7436                 btrfs_get_block_group(used_bg);
7437
7438                 if (!delalloc)
7439                         return used_bg;
7440
7441                 if (down_read_trylock(&used_bg->data_rwsem))
7442                         return used_bg;
7443
7444                 spin_unlock(&cluster->refill_lock);
7445
7446                 /* We should only have one-level nested. */
7447                 down_read_nested(&used_bg->data_rwsem, SINGLE_DEPTH_NESTING);
7448
7449                 spin_lock(&cluster->refill_lock);
7450                 if (used_bg == cluster->block_group)
7451                         return used_bg;
7452
7453                 up_read(&used_bg->data_rwsem);
7454                 btrfs_put_block_group(used_bg);
7455         }
7456 }
7457
7458 static inline void
7459 btrfs_release_block_group(struct btrfs_block_group_cache *cache,
7460                          int delalloc)
7461 {
7462         if (delalloc)
7463                 up_read(&cache->data_rwsem);
7464         btrfs_put_block_group(cache);
7465 }
7466
7467 /*
7468  * walks the btree of allocated extents and find a hole of a given size.
7469  * The key ins is changed to record the hole:
7470  * ins->objectid == start position
7471  * ins->flags = BTRFS_EXTENT_ITEM_KEY
7472  * ins->offset == the size of the hole.
7473  * Any available blocks before search_start are skipped.
7474  *
7475  * If there is no suitable free space, we will record the max size of
7476  * the free space extent currently.
7477  */
7478 static noinline int find_free_extent(struct btrfs_fs_info *fs_info,
7479                                 u64 ram_bytes, u64 num_bytes, u64 empty_size,
7480                                 u64 hint_byte, struct btrfs_key *ins,
7481                                 u64 flags, int delalloc)
7482 {
7483         int ret = 0;
7484         struct btrfs_root *root = fs_info->extent_root;
7485         struct btrfs_free_cluster *last_ptr = NULL;
7486         struct btrfs_block_group_cache *block_group = NULL;
7487         u64 search_start = 0;
7488         u64 max_extent_size = 0;
7489         u64 empty_cluster = 0;
7490         struct btrfs_space_info *space_info;
7491         int loop = 0;
7492         int index = btrfs_bg_flags_to_raid_index(flags);
7493         bool failed_cluster_refill = false;
7494         bool failed_alloc = false;
7495         bool use_cluster = true;
7496         bool have_caching_bg = false;
7497         bool orig_have_caching_bg = false;
7498         bool full_search = false;
7499
7500         WARN_ON(num_bytes < fs_info->sectorsize);
7501         ins->type = BTRFS_EXTENT_ITEM_KEY;
7502         ins->objectid = 0;
7503         ins->offset = 0;
7504
7505         trace_find_free_extent(fs_info, num_bytes, empty_size, flags);
7506
7507         space_info = __find_space_info(fs_info, flags);
7508         if (!space_info) {
7509                 btrfs_err(fs_info, "No space info for %llu", flags);
7510                 return -ENOSPC;
7511         }
7512
7513         /*
7514          * If our free space is heavily fragmented we may not be able to make
7515          * big contiguous allocations, so instead of doing the expensive search
7516          * for free space, simply return ENOSPC with our max_extent_size so we
7517          * can go ahead and search for a more manageable chunk.
7518          *
7519          * If our max_extent_size is large enough for our allocation simply
7520          * disable clustering since we will likely not be able to find enough
7521          * space to create a cluster and induce latency trying.
7522          */
7523         if (unlikely(space_info->max_extent_size)) {
7524                 spin_lock(&space_info->lock);
7525                 if (space_info->max_extent_size &&
7526                     num_bytes > space_info->max_extent_size) {
7527                         ins->offset = space_info->max_extent_size;
7528                         spin_unlock(&space_info->lock);
7529                         return -ENOSPC;
7530                 } else if (space_info->max_extent_size) {
7531                         use_cluster = false;
7532                 }
7533                 spin_unlock(&space_info->lock);
7534         }
7535
7536         last_ptr = fetch_cluster_info(fs_info, space_info, &empty_cluster);
7537         if (last_ptr) {
7538                 spin_lock(&last_ptr->lock);
7539                 if (last_ptr->block_group)
7540                         hint_byte = last_ptr->window_start;
7541                 if (last_ptr->fragmented) {
7542                         /*
7543                          * We still set window_start so we can keep track of the
7544                          * last place we found an allocation to try and save
7545                          * some time.
7546                          */
7547                         hint_byte = last_ptr->window_start;
7548                         use_cluster = false;
7549                 }
7550                 spin_unlock(&last_ptr->lock);
7551         }
7552
7553         search_start = max(search_start, first_logical_byte(fs_info, 0));
7554         search_start = max(search_start, hint_byte);
7555         if (search_start == hint_byte) {
7556                 block_group = btrfs_lookup_block_group(fs_info, search_start);
7557                 /*
7558                  * we don't want to use the block group if it doesn't match our
7559                  * allocation bits, or if its not cached.
7560                  *
7561                  * However if we are re-searching with an ideal block group
7562                  * picked out then we don't care that the block group is cached.
7563                  */
7564                 if (block_group && block_group_bits(block_group, flags) &&
7565                     block_group->cached != BTRFS_CACHE_NO) {
7566                         down_read(&space_info->groups_sem);
7567                         if (list_empty(&block_group->list) ||
7568                             block_group->ro) {
7569                                 /*
7570                                  * someone is removing this block group,
7571                                  * we can't jump into the have_block_group
7572                                  * target because our list pointers are not
7573                                  * valid
7574                                  */
7575                                 btrfs_put_block_group(block_group);
7576                                 up_read(&space_info->groups_sem);
7577                         } else {
7578                                 index = btrfs_bg_flags_to_raid_index(
7579                                                 block_group->flags);
7580                                 btrfs_lock_block_group(block_group, delalloc);
7581                                 goto have_block_group;
7582                         }
7583                 } else if (block_group) {
7584                         btrfs_put_block_group(block_group);
7585                 }
7586         }
7587 search:
7588         have_caching_bg = false;
7589         if (index == 0 || index == btrfs_bg_flags_to_raid_index(flags))
7590                 full_search = true;
7591         down_read(&space_info->groups_sem);
7592         list_for_each_entry(block_group, &space_info->block_groups[index],
7593                             list) {
7594                 u64 offset;
7595                 int cached;
7596
7597                 /* If the block group is read-only, we can skip it entirely. */
7598                 if (unlikely(block_group->ro))
7599                         continue;
7600
7601                 btrfs_grab_block_group(block_group, delalloc);
7602                 search_start = block_group->key.objectid;
7603
7604                 /*
7605                  * this can happen if we end up cycling through all the
7606                  * raid types, but we want to make sure we only allocate
7607                  * for the proper type.
7608                  */
7609                 if (!block_group_bits(block_group, flags)) {
7610                     u64 extra = BTRFS_BLOCK_GROUP_DUP |
7611                                 BTRFS_BLOCK_GROUP_RAID1 |
7612                                 BTRFS_BLOCK_GROUP_RAID5 |
7613                                 BTRFS_BLOCK_GROUP_RAID6 |
7614                                 BTRFS_BLOCK_GROUP_RAID10;
7615
7616                         /*
7617                          * if they asked for extra copies and this block group
7618                          * doesn't provide them, bail.  This does allow us to
7619                          * fill raid0 from raid1.
7620                          */
7621                         if ((flags & extra) && !(block_group->flags & extra))
7622                                 goto loop;
7623                 }
7624
7625 have_block_group:
7626                 cached = block_group_cache_done(block_group);
7627                 if (unlikely(!cached)) {
7628                         have_caching_bg = true;
7629                         ret = cache_block_group(block_group, 0);
7630                         BUG_ON(ret < 0);
7631                         ret = 0;
7632                 }
7633
7634                 if (unlikely(block_group->cached == BTRFS_CACHE_ERROR))
7635                         goto loop;
7636
7637                 /*
7638                  * Ok we want to try and use the cluster allocator, so
7639                  * lets look there
7640                  */
7641                 if (last_ptr && use_cluster) {
7642                         struct btrfs_block_group_cache *used_block_group;
7643                         unsigned long aligned_cluster;
7644                         /*
7645                          * the refill lock keeps out other
7646                          * people trying to start a new cluster
7647                          */
7648                         used_block_group = btrfs_lock_cluster(block_group,
7649                                                               last_ptr,
7650                                                               delalloc);
7651                         if (!used_block_group)
7652                                 goto refill_cluster;
7653
7654                         if (used_block_group != block_group &&
7655                             (used_block_group->ro ||
7656                              !block_group_bits(used_block_group, flags)))
7657                                 goto release_cluster;
7658
7659                         offset = btrfs_alloc_from_cluster(used_block_group,
7660                                                 last_ptr,
7661                                                 num_bytes,
7662                                                 used_block_group->key.objectid,
7663                                                 &max_extent_size);
7664                         if (offset) {
7665                                 /* we have a block, we're done */
7666                                 spin_unlock(&last_ptr->refill_lock);
7667                                 trace_btrfs_reserve_extent_cluster(fs_info,
7668                                                 used_block_group,
7669                                                 search_start, num_bytes);
7670                                 if (used_block_group != block_group) {
7671                                         btrfs_release_block_group(block_group,
7672                                                                   delalloc);
7673                                         block_group = used_block_group;
7674                                 }
7675                                 goto checks;
7676                         }
7677
7678                         WARN_ON(last_ptr->block_group != used_block_group);
7679 release_cluster:
7680                         /* If we are on LOOP_NO_EMPTY_SIZE, we can't
7681                          * set up a new clusters, so lets just skip it
7682                          * and let the allocator find whatever block
7683                          * it can find.  If we reach this point, we
7684                          * will have tried the cluster allocator
7685                          * plenty of times and not have found
7686                          * anything, so we are likely way too
7687                          * fragmented for the clustering stuff to find
7688                          * anything.
7689                          *
7690                          * However, if the cluster is taken from the
7691                          * current block group, release the cluster
7692                          * first, so that we stand a better chance of
7693                          * succeeding in the unclustered
7694                          * allocation.  */
7695                         if (loop >= LOOP_NO_EMPTY_SIZE &&
7696                             used_block_group != block_group) {
7697                                 spin_unlock(&last_ptr->refill_lock);
7698                                 btrfs_release_block_group(used_block_group,
7699                                                           delalloc);
7700                                 goto unclustered_alloc;
7701                         }
7702
7703                         /*
7704                          * this cluster didn't work out, free it and
7705                          * start over
7706                          */
7707                         btrfs_return_cluster_to_free_space(NULL, last_ptr);
7708
7709                         if (used_block_group != block_group)
7710                                 btrfs_release_block_group(used_block_group,
7711                                                           delalloc);
7712 refill_cluster:
7713                         if (loop >= LOOP_NO_EMPTY_SIZE) {
7714                                 spin_unlock(&last_ptr->refill_lock);
7715                                 goto unclustered_alloc;
7716                         }
7717
7718                         aligned_cluster = max_t(unsigned long,
7719                                                 empty_cluster + empty_size,
7720                                               block_group->full_stripe_len);
7721
7722                         /* allocate a cluster in this block group */
7723                         ret = btrfs_find_space_cluster(fs_info, block_group,
7724                                                        last_ptr, search_start,
7725                                                        num_bytes,
7726                                                        aligned_cluster);
7727                         if (ret == 0) {
7728                                 /*
7729                                  * now pull our allocation out of this
7730                                  * cluster
7731                                  */
7732                                 offset = btrfs_alloc_from_cluster(block_group,
7733                                                         last_ptr,
7734                                                         num_bytes,
7735                                                         search_start,
7736                                                         &max_extent_size);
7737                                 if (offset) {
7738                                         /* we found one, proceed */
7739                                         spin_unlock(&last_ptr->refill_lock);
7740                                         trace_btrfs_reserve_extent_cluster(fs_info,
7741                                                 block_group, search_start,
7742                                                 num_bytes);
7743                                         goto checks;
7744                                 }
7745                         } else if (!cached && loop > LOOP_CACHING_NOWAIT
7746                                    && !failed_cluster_refill) {
7747                                 spin_unlock(&last_ptr->refill_lock);
7748
7749                                 failed_cluster_refill = true;
7750                                 wait_block_group_cache_progress(block_group,
7751                                        num_bytes + empty_cluster + empty_size);
7752                                 goto have_block_group;
7753                         }
7754
7755                         /*
7756                          * at this point we either didn't find a cluster
7757                          * or we weren't able to allocate a block from our
7758                          * cluster.  Free the cluster we've been trying
7759                          * to use, and go to the next block group
7760                          */
7761                         btrfs_return_cluster_to_free_space(NULL, last_ptr);
7762                         spin_unlock(&last_ptr->refill_lock);
7763                         goto loop;
7764                 }
7765
7766 unclustered_alloc:
7767                 /*
7768                  * We are doing an unclustered alloc, set the fragmented flag so
7769                  * we don't bother trying to setup a cluster again until we get
7770                  * more space.
7771                  */
7772                 if (unlikely(last_ptr)) {
7773                         spin_lock(&last_ptr->lock);
7774                         last_ptr->fragmented = 1;
7775                         spin_unlock(&last_ptr->lock);
7776                 }
7777                 if (cached) {
7778                         struct btrfs_free_space_ctl *ctl =
7779                                 block_group->free_space_ctl;
7780
7781                         spin_lock(&ctl->tree_lock);
7782                         if (ctl->free_space <
7783                             num_bytes + empty_cluster + empty_size) {
7784                                 if (ctl->free_space > max_extent_size)
7785                                         max_extent_size = ctl->free_space;
7786                                 spin_unlock(&ctl->tree_lock);
7787                                 goto loop;
7788                         }
7789                         spin_unlock(&ctl->tree_lock);
7790                 }
7791
7792                 offset = btrfs_find_space_for_alloc(block_group, search_start,
7793                                                     num_bytes, empty_size,
7794                                                     &max_extent_size);
7795                 /*
7796                  * If we didn't find a chunk, and we haven't failed on this
7797                  * block group before, and this block group is in the middle of
7798                  * caching and we are ok with waiting, then go ahead and wait
7799                  * for progress to be made, and set failed_alloc to true.
7800                  *
7801                  * If failed_alloc is true then we've already waited on this
7802                  * block group once and should move on to the next block group.
7803                  */
7804                 if (!offset && !failed_alloc && !cached &&
7805                     loop > LOOP_CACHING_NOWAIT) {
7806                         wait_block_group_cache_progress(block_group,
7807                                                 num_bytes + empty_size);
7808                         failed_alloc = true;
7809                         goto have_block_group;
7810                 } else if (!offset) {
7811                         goto loop;
7812                 }
7813 checks:
7814                 search_start = ALIGN(offset, fs_info->stripesize);
7815
7816                 /* move on to the next group */
7817                 if (search_start + num_bytes >
7818                     block_group->key.objectid + block_group->key.offset) {
7819                         btrfs_add_free_space(block_group, offset, num_bytes);
7820                         goto loop;
7821                 }
7822
7823                 if (offset < search_start)
7824                         btrfs_add_free_space(block_group, offset,
7825                                              search_start - offset);
7826                 BUG_ON(offset > search_start);
7827
7828                 ret = btrfs_add_reserved_bytes(block_group, ram_bytes,
7829                                 num_bytes, delalloc);
7830                 if (ret == -EAGAIN) {
7831                         btrfs_add_free_space(block_group, offset, num_bytes);
7832                         goto loop;
7833                 }
7834                 btrfs_inc_block_group_reservations(block_group);
7835
7836                 /* we are all good, lets return */
7837                 ins->objectid = search_start;
7838                 ins->offset = num_bytes;
7839
7840                 trace_btrfs_reserve_extent(fs_info, block_group,
7841                                            search_start, num_bytes);
7842                 btrfs_release_block_group(block_group, delalloc);
7843                 break;
7844 loop:
7845                 failed_cluster_refill = false;
7846                 failed_alloc = false;
7847                 BUG_ON(btrfs_bg_flags_to_raid_index(block_group->flags) !=
7848                        index);
7849                 btrfs_release_block_group(block_group, delalloc);
7850                 cond_resched();
7851         }
7852         up_read(&space_info->groups_sem);
7853
7854         if ((loop == LOOP_CACHING_NOWAIT) && have_caching_bg
7855                 && !orig_have_caching_bg)
7856                 orig_have_caching_bg = true;
7857
7858         if (!ins->objectid && loop >= LOOP_CACHING_WAIT && have_caching_bg)
7859                 goto search;
7860
7861         if (!ins->objectid && ++index < BTRFS_NR_RAID_TYPES)
7862                 goto search;
7863
7864         /*
7865          * LOOP_CACHING_NOWAIT, search partially cached block groups, kicking
7866          *                      caching kthreads as we move along
7867          * LOOP_CACHING_WAIT, search everything, and wait if our bg is caching
7868          * LOOP_ALLOC_CHUNK, force a chunk allocation and try again
7869          * LOOP_NO_EMPTY_SIZE, set empty_size and empty_cluster to 0 and try
7870          *                      again
7871          */
7872         if (!ins->objectid && loop < LOOP_NO_EMPTY_SIZE) {
7873                 index = 0;
7874                 if (loop == LOOP_CACHING_NOWAIT) {
7875                         /*
7876                          * We want to skip the LOOP_CACHING_WAIT step if we
7877                          * don't have any uncached bgs and we've already done a
7878                          * full search through.
7879                          */
7880                         if (orig_have_caching_bg || !full_search)
7881                                 loop = LOOP_CACHING_WAIT;
7882                         else
7883                                 loop = LOOP_ALLOC_CHUNK;
7884                 } else {
7885                         loop++;
7886                 }
7887
7888                 if (loop == LOOP_ALLOC_CHUNK) {
7889                         struct btrfs_trans_handle *trans;
7890                         int exist = 0;
7891
7892                         trans = current->journal_info;
7893                         if (trans)
7894                                 exist = 1;
7895                         else
7896                                 trans = btrfs_join_transaction(root);
7897
7898                         if (IS_ERR(trans)) {
7899                                 ret = PTR_ERR(trans);
7900                                 goto out;
7901                         }
7902
7903                         ret = do_chunk_alloc(trans, fs_info, flags,
7904                                              CHUNK_ALLOC_FORCE);
7905
7906                         /*
7907                          * If we can't allocate a new chunk we've already looped
7908                          * through at least once, move on to the NO_EMPTY_SIZE
7909                          * case.
7910                          */
7911                         if (ret == -ENOSPC)
7912                                 loop = LOOP_NO_EMPTY_SIZE;
7913
7914                         /*
7915                          * Do not bail out on ENOSPC since we
7916                          * can do more things.
7917                          */
7918                         if (ret < 0 && ret != -ENOSPC)
7919                                 btrfs_abort_transaction(trans, ret);
7920                         else
7921                                 ret = 0;
7922                         if (!exist)
7923                                 btrfs_end_transaction(trans);
7924                         if (ret)
7925                                 goto out;
7926                 }
7927
7928                 if (loop == LOOP_NO_EMPTY_SIZE) {
7929                         /*
7930                          * Don't loop again if we already have no empty_size and
7931                          * no empty_cluster.
7932                          */
7933                         if (empty_size == 0 &&
7934                             empty_cluster == 0) {
7935                                 ret = -ENOSPC;
7936                                 goto out;
7937                         }
7938                         empty_size = 0;
7939                         empty_cluster = 0;
7940                 }
7941
7942                 goto search;
7943         } else if (!ins->objectid) {
7944                 ret = -ENOSPC;
7945         } else if (ins->objectid) {
7946                 if (!use_cluster && last_ptr) {
7947                         spin_lock(&last_ptr->lock);
7948                         last_ptr->window_start = ins->objectid;
7949                         spin_unlock(&last_ptr->lock);
7950                 }
7951                 ret = 0;
7952         }
7953 out:
7954         if (ret == -ENOSPC) {
7955                 spin_lock(&space_info->lock);
7956                 space_info->max_extent_size = max_extent_size;
7957                 spin_unlock(&space_info->lock);
7958                 ins->offset = max_extent_size;
7959         }
7960         return ret;
7961 }
7962
7963 static void dump_space_info(struct btrfs_fs_info *fs_info,
7964                             struct btrfs_space_info *info, u64 bytes,
7965                             int dump_block_groups)
7966 {
7967         struct btrfs_block_group_cache *cache;
7968         int index = 0;
7969
7970         spin_lock(&info->lock);
7971         btrfs_info(fs_info, "space_info %llu has %llu free, is %sfull",
7972                    info->flags,
7973                    info->total_bytes - btrfs_space_info_used(info, true),
7974                    info->full ? "" : "not ");
7975         btrfs_info(fs_info,
7976                 "space_info total=%llu, used=%llu, pinned=%llu, reserved=%llu, may_use=%llu, readonly=%llu",
7977                 info->total_bytes, info->bytes_used, info->bytes_pinned,
7978                 info->bytes_reserved, info->bytes_may_use,
7979                 info->bytes_readonly);
7980         spin_unlock(&info->lock);
7981
7982         if (!dump_block_groups)
7983                 return;
7984
7985         down_read(&info->groups_sem);
7986 again:
7987         list_for_each_entry(cache, &info->block_groups[index], list) {
7988                 spin_lock(&cache->lock);
7989                 btrfs_info(fs_info,
7990                         "block group %llu has %llu bytes, %llu used %llu pinned %llu reserved %s",
7991                         cache->key.objectid, cache->key.offset,
7992                         btrfs_block_group_used(&cache->item), cache->pinned,
7993                         cache->reserved, cache->ro ? "[readonly]" : "");
7994                 btrfs_dump_free_space(cache, bytes);
7995                 spin_unlock(&cache->lock);
7996         }
7997         if (++index < BTRFS_NR_RAID_TYPES)
7998                 goto again;
7999         up_read(&info->groups_sem);
8000 }
8001
8002 /*
8003  * btrfs_reserve_extent - entry point to the extent allocator. Tries to find a
8004  *                        hole that is at least as big as @num_bytes.
8005  *
8006  * @root           -    The root that will contain this extent
8007  *
8008  * @ram_bytes      -    The amount of space in ram that @num_bytes take. This
8009  *                      is used for accounting purposes. This value differs
8010  *                      from @num_bytes only in the case of compressed extents.
8011  *
8012  * @num_bytes      -    Number of bytes to allocate on-disk.
8013  *
8014  * @min_alloc_size -    Indicates the minimum amount of space that the
8015  *                      allocator should try to satisfy. In some cases
8016  *                      @num_bytes may be larger than what is required and if
8017  *                      the filesystem is fragmented then allocation fails.
8018  *                      However, the presence of @min_alloc_size gives a
8019  *                      chance to try and satisfy the smaller allocation.
8020  *
8021  * @empty_size     -    A hint that you plan on doing more COW. This is the
8022  *                      size in bytes the allocator should try to find free
8023  *                      next to the block it returns.  This is just a hint and
8024  *                      may be ignored by the allocator.
8025  *
8026  * @hint_byte      -    Hint to the allocator to start searching above the byte
8027  *                      address passed. It might be ignored.
8028  *
8029  * @ins            -    This key is modified to record the found hole. It will
8030  *                      have the following values:
8031  *                      ins->objectid == start position
8032  *                      ins->flags = BTRFS_EXTENT_ITEM_KEY
8033  *                      ins->offset == the size of the hole.
8034  *
8035  * @is_data        -    Boolean flag indicating whether an extent is
8036  *                      allocated for data (true) or metadata (false)
8037  *
8038  * @delalloc       -    Boolean flag indicating whether this allocation is for
8039  *                      delalloc or not. If 'true' data_rwsem of block groups
8040  *                      is going to be acquired.
8041  *
8042  *
8043  * Returns 0 when an allocation succeeded or < 0 when an error occurred. In
8044  * case -ENOSPC is returned then @ins->offset will contain the size of the
8045  * largest available hole the allocator managed to find.
8046  */
8047 int btrfs_reserve_extent(struct btrfs_root *root, u64 ram_bytes,
8048                          u64 num_bytes, u64 min_alloc_size,
8049                          u64 empty_size, u64 hint_byte,
8050                          struct btrfs_key *ins, int is_data, int delalloc)
8051 {
8052         struct btrfs_fs_info *fs_info = root->fs_info;
8053         bool final_tried = num_bytes == min_alloc_size;
8054         u64 flags;
8055         int ret;
8056
8057         flags = get_alloc_profile_by_root(root, is_data);
8058 again:
8059         WARN_ON(num_bytes < fs_info->sectorsize);
8060         ret = find_free_extent(fs_info, ram_bytes, num_bytes, empty_size,
8061                                hint_byte, ins, flags, delalloc);
8062         if (!ret && !is_data) {
8063                 btrfs_dec_block_group_reservations(fs_info, ins->objectid);
8064         } else if (ret == -ENOSPC) {
8065                 if (!final_tried && ins->offset) {
8066                         num_bytes = min(num_bytes >> 1, ins->offset);
8067                         num_bytes = round_down(num_bytes,
8068                                                fs_info->sectorsize);
8069                         num_bytes = max(num_bytes, min_alloc_size);
8070                         ram_bytes = num_bytes;
8071                         if (num_bytes == min_alloc_size)
8072                                 final_tried = true;
8073                         goto again;
8074                 } else if (btrfs_test_opt(fs_info, ENOSPC_DEBUG)) {
8075                         struct btrfs_space_info *sinfo;
8076
8077                         sinfo = __find_space_info(fs_info, flags);
8078                         btrfs_err(fs_info,
8079                                   "allocation failed flags %llu, wanted %llu",
8080                                   flags, num_bytes);
8081                         if (sinfo)
8082                                 dump_space_info(fs_info, sinfo, num_bytes, 1);
8083                 }
8084         }
8085
8086         return ret;
8087 }
8088
8089 static int __btrfs_free_reserved_extent(struct btrfs_fs_info *fs_info,
8090                                         u64 start, u64 len,
8091                                         int pin, int delalloc)
8092 {
8093         struct btrfs_block_group_cache *cache;
8094         int ret = 0;
8095
8096         cache = btrfs_lookup_block_group(fs_info, start);
8097         if (!cache) {
8098                 btrfs_err(fs_info, "Unable to find block group for %llu",
8099                           start);
8100                 return -ENOSPC;
8101         }
8102
8103         if (pin)
8104                 pin_down_extent(fs_info, cache, start, len, 1);
8105         else {
8106                 if (btrfs_test_opt(fs_info, DISCARD))
8107                         ret = btrfs_discard_extent(fs_info, start, len, NULL);
8108                 btrfs_add_free_space(cache, start, len);
8109                 btrfs_free_reserved_bytes(cache, len, delalloc);
8110                 trace_btrfs_reserved_extent_free(fs_info, start, len);
8111         }
8112
8113         btrfs_put_block_group(cache);
8114         return ret;
8115 }
8116
8117 int btrfs_free_reserved_extent(struct btrfs_fs_info *fs_info,
8118                                u64 start, u64 len, int delalloc)
8119 {
8120         return __btrfs_free_reserved_extent(fs_info, start, len, 0, delalloc);
8121 }
8122
8123 int btrfs_free_and_pin_reserved_extent(struct btrfs_fs_info *fs_info,
8124                                        u64 start, u64 len)
8125 {
8126         return __btrfs_free_reserved_extent(fs_info, start, len, 1, 0);
8127 }
8128
8129 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
8130                                       struct btrfs_fs_info *fs_info,
8131                                       u64 parent, u64 root_objectid,
8132                                       u64 flags, u64 owner, u64 offset,
8133                                       struct btrfs_key *ins, int ref_mod)
8134 {
8135         int ret;
8136         struct btrfs_extent_item *extent_item;
8137         struct btrfs_extent_inline_ref *iref;
8138         struct btrfs_path *path;
8139         struct extent_buffer *leaf;
8140         int type;
8141         u32 size;
8142
8143         if (parent > 0)
8144                 type = BTRFS_SHARED_DATA_REF_KEY;
8145         else
8146                 type = BTRFS_EXTENT_DATA_REF_KEY;
8147
8148         size = sizeof(*extent_item) + btrfs_extent_inline_ref_size(type);
8149
8150         path = btrfs_alloc_path();
8151         if (!path)
8152                 return -ENOMEM;
8153
8154         path->leave_spinning = 1;
8155         ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path,
8156                                       ins, size);
8157         if (ret) {
8158                 btrfs_free_path(path);
8159                 return ret;
8160         }
8161
8162         leaf = path->nodes[0];
8163         extent_item = btrfs_item_ptr(leaf, path->slots[0],
8164                                      struct btrfs_extent_item);
8165         btrfs_set_extent_refs(leaf, extent_item, ref_mod);
8166         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
8167         btrfs_set_extent_flags(leaf, extent_item,
8168                                flags | BTRFS_EXTENT_FLAG_DATA);
8169
8170         iref = (struct btrfs_extent_inline_ref *)(extent_item + 1);
8171         btrfs_set_extent_inline_ref_type(leaf, iref, type);
8172         if (parent > 0) {
8173                 struct btrfs_shared_data_ref *ref;
8174                 ref = (struct btrfs_shared_data_ref *)(iref + 1);
8175                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
8176                 btrfs_set_shared_data_ref_count(leaf, ref, ref_mod);
8177         } else {
8178                 struct btrfs_extent_data_ref *ref;
8179                 ref = (struct btrfs_extent_data_ref *)(&iref->offset);
8180                 btrfs_set_extent_data_ref_root(leaf, ref, root_objectid);
8181                 btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
8182                 btrfs_set_extent_data_ref_offset(leaf, ref, offset);
8183                 btrfs_set_extent_data_ref_count(leaf, ref, ref_mod);
8184         }
8185
8186         btrfs_mark_buffer_dirty(path->nodes[0]);
8187         btrfs_free_path(path);
8188
8189         ret = remove_from_free_space_tree(trans, fs_info, ins->objectid,
8190                                           ins->offset);
8191         if (ret)
8192                 return ret;
8193
8194         ret = update_block_group(trans, fs_info, ins->objectid, ins->offset, 1);
8195         if (ret) { /* -ENOENT, logic error */
8196                 btrfs_err(fs_info, "update block group failed for %llu %llu",
8197                         ins->objectid, ins->offset);
8198                 BUG();
8199         }
8200         trace_btrfs_reserved_extent_alloc(fs_info, ins->objectid, ins->offset);
8201         return ret;
8202 }
8203
8204 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
8205                                      struct btrfs_fs_info *fs_info,
8206                                      u64 parent, u64 root_objectid,
8207                                      u64 flags, struct btrfs_disk_key *key,
8208                                      int level, struct btrfs_key *ins)
8209 {
8210         int ret;
8211         struct btrfs_extent_item *extent_item;
8212         struct btrfs_tree_block_info *block_info;
8213         struct btrfs_extent_inline_ref *iref;
8214         struct btrfs_path *path;
8215         struct extent_buffer *leaf;
8216         u32 size = sizeof(*extent_item) + sizeof(*iref);
8217         u64 num_bytes = ins->offset;
8218         bool skinny_metadata = btrfs_fs_incompat(fs_info, SKINNY_METADATA);
8219
8220         if (!skinny_metadata)
8221                 size += sizeof(*block_info);
8222
8223         path = btrfs_alloc_path();
8224         if (!path) {
8225                 btrfs_free_and_pin_reserved_extent(fs_info, ins->objectid,
8226                                                    fs_info->nodesize);
8227                 return -ENOMEM;
8228         }
8229
8230         path->leave_spinning = 1;
8231         ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path,
8232                                       ins, size);
8233         if (ret) {
8234                 btrfs_free_path(path);
8235                 btrfs_free_and_pin_reserved_extent(fs_info, ins->objectid,
8236                                                    fs_info->nodesize);
8237                 return ret;
8238         }
8239
8240         leaf = path->nodes[0];
8241         extent_item = btrfs_item_ptr(leaf, path->slots[0],
8242                                      struct btrfs_extent_item);
8243         btrfs_set_extent_refs(leaf, extent_item, 1);
8244         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
8245         btrfs_set_extent_flags(leaf, extent_item,
8246                                flags | BTRFS_EXTENT_FLAG_TREE_BLOCK);
8247
8248         if (skinny_metadata) {
8249                 iref = (struct btrfs_extent_inline_ref *)(extent_item + 1);
8250                 num_bytes = fs_info->nodesize;
8251         } else {
8252                 block_info = (struct btrfs_tree_block_info *)(extent_item + 1);
8253                 btrfs_set_tree_block_key(leaf, block_info, key);
8254                 btrfs_set_tree_block_level(leaf, block_info, level);
8255                 iref = (struct btrfs_extent_inline_ref *)(block_info + 1);
8256         }
8257
8258         if (parent > 0) {
8259                 BUG_ON(!(flags & BTRFS_BLOCK_FLAG_FULL_BACKREF));
8260                 btrfs_set_extent_inline_ref_type(leaf, iref,
8261                                                  BTRFS_SHARED_BLOCK_REF_KEY);
8262                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
8263         } else {
8264                 btrfs_set_extent_inline_ref_type(leaf, iref,
8265                                                  BTRFS_TREE_BLOCK_REF_KEY);
8266                 btrfs_set_extent_inline_ref_offset(leaf, iref, root_objectid);
8267         }
8268
8269         btrfs_mark_buffer_dirty(leaf);
8270         btrfs_free_path(path);
8271
8272         ret = remove_from_free_space_tree(trans, fs_info, ins->objectid,
8273                                           num_bytes);
8274         if (ret)
8275                 return ret;
8276
8277         ret = update_block_group(trans, fs_info, ins->objectid,
8278                                  fs_info->nodesize, 1);
8279         if (ret) { /* -ENOENT, logic error */
8280                 btrfs_err(fs_info, "update block group failed for %llu %llu",
8281                         ins->objectid, ins->offset);
8282                 BUG();
8283         }
8284
8285         trace_btrfs_reserved_extent_alloc(fs_info, ins->objectid,
8286                                           fs_info->nodesize);
8287         return ret;
8288 }
8289
8290 int btrfs_alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
8291                                      struct btrfs_root *root, u64 owner,
8292                                      u64 offset, u64 ram_bytes,
8293                                      struct btrfs_key *ins)
8294 {
8295         struct btrfs_fs_info *fs_info = root->fs_info;
8296         int ret;
8297
8298         BUG_ON(root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID);
8299
8300         btrfs_ref_tree_mod(root, ins->objectid, ins->offset, 0,
8301                            root->root_key.objectid, owner, offset,
8302                            BTRFS_ADD_DELAYED_EXTENT);
8303
8304         ret = btrfs_add_delayed_data_ref(fs_info, trans, ins->objectid,
8305                                          ins->offset, 0,
8306                                          root->root_key.objectid, owner,
8307                                          offset, ram_bytes,
8308                                          BTRFS_ADD_DELAYED_EXTENT, NULL, NULL);
8309         return ret;
8310 }
8311
8312 /*
8313  * this is used by the tree logging recovery code.  It records that
8314  * an extent has been allocated and makes sure to clear the free
8315  * space cache bits as well
8316  */
8317 int btrfs_alloc_logged_file_extent(struct btrfs_trans_handle *trans,
8318                                    struct btrfs_fs_info *fs_info,
8319                                    u64 root_objectid, u64 owner, u64 offset,
8320                                    struct btrfs_key *ins)
8321 {
8322         int ret;
8323         struct btrfs_block_group_cache *block_group;
8324         struct btrfs_space_info *space_info;
8325
8326         /*
8327          * Mixed block groups will exclude before processing the log so we only
8328          * need to do the exclude dance if this fs isn't mixed.
8329          */
8330         if (!btrfs_fs_incompat(fs_info, MIXED_GROUPS)) {
8331                 ret = __exclude_logged_extent(fs_info, ins->objectid,
8332                                               ins->offset);
8333                 if (ret)
8334                         return ret;
8335         }
8336
8337         block_group = btrfs_lookup_block_group(fs_info, ins->objectid);
8338         if (!block_group)
8339                 return -EINVAL;
8340
8341         space_info = block_group->space_info;
8342         spin_lock(&space_info->lock);
8343         spin_lock(&block_group->lock);
8344         space_info->bytes_reserved += ins->offset;
8345         block_group->reserved += ins->offset;
8346         spin_unlock(&block_group->lock);
8347         spin_unlock(&space_info->lock);
8348
8349         ret = alloc_reserved_file_extent(trans, fs_info, 0, root_objectid,
8350                                          0, owner, offset, ins, 1);
8351         btrfs_put_block_group(block_group);
8352         return ret;
8353 }
8354
8355 static struct extent_buffer *
8356 btrfs_init_new_buffer(struct btrfs_trans_handle *trans, struct btrfs_root *root,
8357                       u64 bytenr, int level)
8358 {
8359         struct btrfs_fs_info *fs_info = root->fs_info;
8360         struct extent_buffer *buf;
8361
8362         buf = btrfs_find_create_tree_block(fs_info, bytenr);
8363         if (IS_ERR(buf))
8364                 return buf;
8365
8366         btrfs_set_header_generation(buf, trans->transid);
8367         btrfs_set_buffer_lockdep_class(root->root_key.objectid, buf, level);
8368         btrfs_tree_lock(buf);
8369         clean_tree_block(fs_info, buf);
8370         clear_bit(EXTENT_BUFFER_STALE, &buf->bflags);
8371
8372         btrfs_set_lock_blocking(buf);
8373         set_extent_buffer_uptodate(buf);
8374
8375         if (root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID) {
8376                 buf->log_index = root->log_transid % 2;
8377                 /*
8378                  * we allow two log transactions at a time, use different
8379                  * EXENT bit to differentiate dirty pages.
8380                  */
8381                 if (buf->log_index == 0)
8382                         set_extent_dirty(&root->dirty_log_pages, buf->start,
8383                                         buf->start + buf->len - 1, GFP_NOFS);
8384                 else
8385                         set_extent_new(&root->dirty_log_pages, buf->start,
8386                                         buf->start + buf->len - 1);
8387         } else {
8388                 buf->log_index = -1;
8389                 set_extent_dirty(&trans->transaction->dirty_pages, buf->start,
8390                          buf->start + buf->len - 1, GFP_NOFS);
8391         }
8392         trans->dirty = true;
8393         /* this returns a buffer locked for blocking */
8394         return buf;
8395 }
8396
8397 static struct btrfs_block_rsv *
8398 use_block_rsv(struct btrfs_trans_handle *trans,
8399               struct btrfs_root *root, u32 blocksize)
8400 {
8401         struct btrfs_fs_info *fs_info = root->fs_info;
8402         struct btrfs_block_rsv *block_rsv;
8403         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
8404         int ret;
8405         bool global_updated = false;
8406
8407         block_rsv = get_block_rsv(trans, root);
8408
8409         if (unlikely(block_rsv->size == 0))
8410                 goto try_reserve;
8411 again:
8412         ret = block_rsv_use_bytes(block_rsv, blocksize);
8413         if (!ret)
8414                 return block_rsv;
8415
8416         if (block_rsv->failfast)
8417                 return ERR_PTR(ret);
8418
8419         if (block_rsv->type == BTRFS_BLOCK_RSV_GLOBAL && !global_updated) {
8420                 global_updated = true;
8421                 update_global_block_rsv(fs_info);
8422                 goto again;
8423         }
8424
8425         if (btrfs_test_opt(fs_info, ENOSPC_DEBUG)) {
8426                 static DEFINE_RATELIMIT_STATE(_rs,
8427                                 DEFAULT_RATELIMIT_INTERVAL * 10,
8428                                 /*DEFAULT_RATELIMIT_BURST*/ 1);
8429                 if (__ratelimit(&_rs))
8430                         WARN(1, KERN_DEBUG
8431                                 "BTRFS: block rsv returned %d\n", ret);
8432         }
8433 try_reserve:
8434         ret = reserve_metadata_bytes(root, block_rsv, blocksize,
8435                                      BTRFS_RESERVE_NO_FLUSH);
8436         if (!ret)
8437                 return block_rsv;
8438         /*
8439          * If we couldn't reserve metadata bytes try and use some from
8440          * the global reserve if its space type is the same as the global
8441          * reservation.
8442          */
8443         if (block_rsv->type != BTRFS_BLOCK_RSV_GLOBAL &&
8444             block_rsv->space_info == global_rsv->space_info) {
8445                 ret = block_rsv_use_bytes(global_rsv, blocksize);
8446                 if (!ret)
8447                         return global_rsv;
8448         }
8449         return ERR_PTR(ret);
8450 }
8451
8452 static void unuse_block_rsv(struct btrfs_fs_info *fs_info,
8453                             struct btrfs_block_rsv *block_rsv, u32 blocksize)
8454 {
8455         block_rsv_add_bytes(block_rsv, blocksize, 0);
8456         block_rsv_release_bytes(fs_info, block_rsv, NULL, 0, NULL);
8457 }
8458
8459 /*
8460  * finds a free extent and does all the dirty work required for allocation
8461  * returns the tree buffer or an ERR_PTR on error.
8462  */
8463 struct extent_buffer *btrfs_alloc_tree_block(struct btrfs_trans_handle *trans,
8464                                              struct btrfs_root *root,
8465                                              u64 parent, u64 root_objectid,
8466                                              const struct btrfs_disk_key *key,
8467                                              int level, u64 hint,
8468                                              u64 empty_size)
8469 {
8470         struct btrfs_fs_info *fs_info = root->fs_info;
8471         struct btrfs_key ins;
8472         struct btrfs_block_rsv *block_rsv;
8473         struct extent_buffer *buf;
8474         struct btrfs_delayed_extent_op *extent_op;
8475         u64 flags = 0;
8476         int ret;
8477         u32 blocksize = fs_info->nodesize;
8478         bool skinny_metadata = btrfs_fs_incompat(fs_info, SKINNY_METADATA);
8479
8480 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
8481         if (btrfs_is_testing(fs_info)) {
8482                 buf = btrfs_init_new_buffer(trans, root, root->alloc_bytenr,
8483                                             level);
8484                 if (!IS_ERR(buf))
8485                         root->alloc_bytenr += blocksize;
8486                 return buf;
8487         }
8488 #endif
8489
8490         block_rsv = use_block_rsv(trans, root, blocksize);
8491         if (IS_ERR(block_rsv))
8492                 return ERR_CAST(block_rsv);
8493
8494         ret = btrfs_reserve_extent(root, blocksize, blocksize, blocksize,
8495                                    empty_size, hint, &ins, 0, 0);
8496         if (ret)
8497                 goto out_unuse;
8498
8499         buf = btrfs_init_new_buffer(trans, root, ins.objectid, level);
8500         if (IS_ERR(buf)) {
8501                 ret = PTR_ERR(buf);
8502                 goto out_free_reserved;
8503         }
8504
8505         if (root_objectid == BTRFS_TREE_RELOC_OBJECTID) {
8506                 if (parent == 0)
8507                         parent = ins.objectid;
8508                 flags |= BTRFS_BLOCK_FLAG_FULL_BACKREF;
8509         } else
8510                 BUG_ON(parent > 0);
8511
8512         if (root_objectid != BTRFS_TREE_LOG_OBJECTID) {
8513                 extent_op = btrfs_alloc_delayed_extent_op();
8514                 if (!extent_op) {
8515                         ret = -ENOMEM;
8516                         goto out_free_buf;
8517                 }
8518                 if (key)
8519                         memcpy(&extent_op->key, key, sizeof(extent_op->key));
8520                 else
8521                         memset(&extent_op->key, 0, sizeof(extent_op->key));
8522                 extent_op->flags_to_set = flags;
8523                 extent_op->update_key = skinny_metadata ? false : true;
8524                 extent_op->update_flags = true;
8525                 extent_op->is_data = false;
8526                 extent_op->level = level;
8527
8528                 btrfs_ref_tree_mod(root, ins.objectid, ins.offset, parent,
8529                                    root_objectid, level, 0,
8530                                    BTRFS_ADD_DELAYED_EXTENT);
8531                 ret = btrfs_add_delayed_tree_ref(fs_info, trans, ins.objectid,
8532                                                  ins.offset, parent,
8533                                                  root_objectid, level,
8534                                                  BTRFS_ADD_DELAYED_EXTENT,
8535                                                  extent_op, NULL, NULL);
8536                 if (ret)
8537                         goto out_free_delayed;
8538         }
8539         return buf;
8540
8541 out_free_delayed:
8542         btrfs_free_delayed_extent_op(extent_op);
8543 out_free_buf:
8544         free_extent_buffer(buf);
8545 out_free_reserved:
8546         btrfs_free_reserved_extent(fs_info, ins.objectid, ins.offset, 0);
8547 out_unuse:
8548         unuse_block_rsv(fs_info, block_rsv, blocksize);
8549         return ERR_PTR(ret);
8550 }
8551
8552 struct walk_control {
8553         u64 refs[BTRFS_MAX_LEVEL];
8554         u64 flags[BTRFS_MAX_LEVEL];
8555         struct btrfs_key update_progress;
8556         int stage;
8557         int level;
8558         int shared_level;
8559         int update_ref;
8560         int keep_locks;
8561         int reada_slot;
8562         int reada_count;
8563         int for_reloc;
8564 };
8565
8566 #define DROP_REFERENCE  1
8567 #define UPDATE_BACKREF  2
8568
8569 static noinline void reada_walk_down(struct btrfs_trans_handle *trans,
8570                                      struct btrfs_root *root,
8571                                      struct walk_control *wc,
8572                                      struct btrfs_path *path)
8573 {
8574         struct btrfs_fs_info *fs_info = root->fs_info;
8575         u64 bytenr;
8576         u64 generation;
8577         u64 refs;
8578         u64 flags;
8579         u32 nritems;
8580         struct btrfs_key key;
8581         struct extent_buffer *eb;
8582         int ret;
8583         int slot;
8584         int nread = 0;
8585
8586         if (path->slots[wc->level] < wc->reada_slot) {
8587                 wc->reada_count = wc->reada_count * 2 / 3;
8588                 wc->reada_count = max(wc->reada_count, 2);
8589         } else {
8590                 wc->reada_count = wc->reada_count * 3 / 2;
8591                 wc->reada_count = min_t(int, wc->reada_count,
8592                                         BTRFS_NODEPTRS_PER_BLOCK(fs_info));
8593         }
8594
8595         eb = path->nodes[wc->level];
8596         nritems = btrfs_header_nritems(eb);
8597
8598         for (slot = path->slots[wc->level]; slot < nritems; slot++) {
8599                 if (nread >= wc->reada_count)
8600                         break;
8601
8602                 cond_resched();
8603                 bytenr = btrfs_node_blockptr(eb, slot);
8604                 generation = btrfs_node_ptr_generation(eb, slot);
8605
8606                 if (slot == path->slots[wc->level])
8607                         goto reada;
8608
8609                 if (wc->stage == UPDATE_BACKREF &&
8610                     generation <= root->root_key.offset)
8611                         continue;
8612
8613                 /* We don't lock the tree block, it's OK to be racy here */
8614                 ret = btrfs_lookup_extent_info(trans, fs_info, bytenr,
8615                                                wc->level - 1, 1, &refs,
8616                                                &flags);
8617                 /* We don't care about errors in readahead. */
8618                 if (ret < 0)
8619                         continue;
8620                 BUG_ON(refs == 0);
8621
8622                 if (wc->stage == DROP_REFERENCE) {
8623                         if (refs == 1)
8624                                 goto reada;
8625
8626                         if (wc->level == 1 &&
8627                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
8628                                 continue;
8629                         if (!wc->update_ref ||
8630                             generation <= root->root_key.offset)
8631                                 continue;
8632                         btrfs_node_key_to_cpu(eb, &key, slot);
8633                         ret = btrfs_comp_cpu_keys(&key,
8634                                                   &wc->update_progress);
8635                         if (ret < 0)
8636                                 continue;
8637                 } else {
8638                         if (wc->level == 1 &&
8639                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
8640                                 continue;
8641                 }
8642 reada:
8643                 readahead_tree_block(fs_info, bytenr);
8644                 nread++;
8645         }
8646         wc->reada_slot = slot;
8647 }
8648
8649 /*
8650  * helper to process tree block while walking down the tree.
8651  *
8652  * when wc->stage == UPDATE_BACKREF, this function updates
8653  * back refs for pointers in the block.
8654  *
8655  * NOTE: return value 1 means we should stop walking down.
8656  */
8657 static noinline int walk_down_proc(struct btrfs_trans_handle *trans,
8658                                    struct btrfs_root *root,
8659                                    struct btrfs_path *path,
8660                                    struct walk_control *wc, int lookup_info)
8661 {
8662         struct btrfs_fs_info *fs_info = root->fs_info;
8663         int level = wc->level;
8664         struct extent_buffer *eb = path->nodes[level];
8665         u64 flag = BTRFS_BLOCK_FLAG_FULL_BACKREF;
8666         int ret;
8667
8668         if (wc->stage == UPDATE_BACKREF &&
8669             btrfs_header_owner(eb) != root->root_key.objectid)
8670                 return 1;
8671
8672         /*
8673          * when reference count of tree block is 1, it won't increase
8674          * again. once full backref flag is set, we never clear it.
8675          */
8676         if (lookup_info &&
8677             ((wc->stage == DROP_REFERENCE && wc->refs[level] != 1) ||
8678              (wc->stage == UPDATE_BACKREF && !(wc->flags[level] & flag)))) {
8679                 BUG_ON(!path->locks[level]);
8680                 ret = btrfs_lookup_extent_info(trans, fs_info,
8681                                                eb->start, level, 1,
8682                                                &wc->refs[level],
8683                                                &wc->flags[level]);
8684                 BUG_ON(ret == -ENOMEM);
8685                 if (ret)
8686                         return ret;
8687                 BUG_ON(wc->refs[level] == 0);
8688         }
8689
8690         if (wc->stage == DROP_REFERENCE) {
8691                 if (wc->refs[level] > 1)
8692                         return 1;
8693
8694                 if (path->locks[level] && !wc->keep_locks) {
8695                         btrfs_tree_unlock_rw(eb, path->locks[level]);
8696                         path->locks[level] = 0;
8697                 }
8698                 return 0;
8699         }
8700
8701         /* wc->stage == UPDATE_BACKREF */
8702         if (!(wc->flags[level] & flag)) {
8703                 BUG_ON(!path->locks[level]);
8704                 ret = btrfs_inc_ref(trans, root, eb, 1);
8705                 BUG_ON(ret); /* -ENOMEM */
8706                 ret = btrfs_dec_ref(trans, root, eb, 0);
8707                 BUG_ON(ret); /* -ENOMEM */
8708                 ret = btrfs_set_disk_extent_flags(trans, fs_info, eb->start,
8709                                                   eb->len, flag,
8710                                                   btrfs_header_level(eb), 0);
8711                 BUG_ON(ret); /* -ENOMEM */
8712                 wc->flags[level] |= flag;
8713         }
8714
8715         /*
8716          * the block is shared by multiple trees, so it's not good to
8717          * keep the tree lock
8718          */
8719         if (path->locks[level] && level > 0) {
8720                 btrfs_tree_unlock_rw(eb, path->locks[level]);
8721                 path->locks[level] = 0;
8722         }
8723         return 0;
8724 }
8725
8726 /*
8727  * helper to process tree block pointer.
8728  *
8729  * when wc->stage == DROP_REFERENCE, this function checks
8730  * reference count of the block pointed to. if the block
8731  * is shared and we need update back refs for the subtree
8732  * rooted at the block, this function changes wc->stage to
8733  * UPDATE_BACKREF. if the block is shared and there is no
8734  * need to update back, this function drops the reference
8735  * to the block.
8736  *
8737  * NOTE: return value 1 means we should stop walking down.
8738  */
8739 static noinline int do_walk_down(struct btrfs_trans_handle *trans,
8740                                  struct btrfs_root *root,
8741                                  struct btrfs_path *path,
8742                                  struct walk_control *wc, int *lookup_info)
8743 {
8744         struct btrfs_fs_info *fs_info = root->fs_info;
8745         u64 bytenr;
8746         u64 generation;
8747         u64 parent;
8748         u32 blocksize;
8749         struct btrfs_key key;
8750         struct btrfs_key first_key;
8751         struct extent_buffer *next;
8752         int level = wc->level;
8753         int reada = 0;
8754         int ret = 0;
8755         bool need_account = false;
8756
8757         generation = btrfs_node_ptr_generation(path->nodes[level],
8758                                                path->slots[level]);
8759         /*
8760          * if the lower level block was created before the snapshot
8761          * was created, we know there is no need to update back refs
8762          * for the subtree
8763          */
8764         if (wc->stage == UPDATE_BACKREF &&
8765             generation <= root->root_key.offset) {
8766                 *lookup_info = 1;
8767                 return 1;
8768         }
8769
8770         bytenr = btrfs_node_blockptr(path->nodes[level], path->slots[level]);
8771         btrfs_node_key_to_cpu(path->nodes[level], &first_key,
8772                               path->slots[level]);
8773         blocksize = fs_info->nodesize;
8774
8775         next = find_extent_buffer(fs_info, bytenr);
8776         if (!next) {
8777                 next = btrfs_find_create_tree_block(fs_info, bytenr);
8778                 if (IS_ERR(next))
8779                         return PTR_ERR(next);
8780
8781                 btrfs_set_buffer_lockdep_class(root->root_key.objectid, next,
8782                                                level - 1);
8783                 reada = 1;
8784         }
8785         btrfs_tree_lock(next);
8786         btrfs_set_lock_blocking(next);
8787
8788         ret = btrfs_lookup_extent_info(trans, fs_info, bytenr, level - 1, 1,
8789                                        &wc->refs[level - 1],
8790                                        &wc->flags[level - 1]);
8791         if (ret < 0)
8792                 goto out_unlock;
8793
8794         if (unlikely(wc->refs[level - 1] == 0)) {
8795                 btrfs_err(fs_info, "Missing references.");
8796                 ret = -EIO;
8797                 goto out_unlock;
8798         }
8799         *lookup_info = 0;
8800
8801         if (wc->stage == DROP_REFERENCE) {
8802                 if (wc->refs[level - 1] > 1) {
8803                         need_account = true;
8804                         if (level == 1 &&
8805                             (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
8806                                 goto skip;
8807
8808                         if (!wc->update_ref ||
8809                             generation <= root->root_key.offset)
8810                                 goto skip;
8811
8812                         btrfs_node_key_to_cpu(path->nodes[level], &key,
8813                                               path->slots[level]);
8814                         ret = btrfs_comp_cpu_keys(&key, &wc->update_progress);
8815                         if (ret < 0)
8816                                 goto skip;
8817
8818                         wc->stage = UPDATE_BACKREF;
8819                         wc->shared_level = level - 1;
8820                 }
8821         } else {
8822                 if (level == 1 &&
8823                     (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
8824                         goto skip;
8825         }
8826
8827         if (!btrfs_buffer_uptodate(next, generation, 0)) {
8828                 btrfs_tree_unlock(next);
8829                 free_extent_buffer(next);
8830                 next = NULL;
8831                 *lookup_info = 1;
8832         }
8833
8834         if (!next) {
8835                 if (reada && level == 1)
8836                         reada_walk_down(trans, root, wc, path);
8837                 next = read_tree_block(fs_info, bytenr, generation, level - 1,
8838                                        &first_key);
8839                 if (IS_ERR(next)) {
8840                         return PTR_ERR(next);
8841                 } else if (!extent_buffer_uptodate(next)) {
8842                         free_extent_buffer(next);
8843                         return -EIO;
8844                 }
8845                 btrfs_tree_lock(next);
8846                 btrfs_set_lock_blocking(next);
8847         }
8848
8849         level--;
8850         ASSERT(level == btrfs_header_level(next));
8851         if (level != btrfs_header_level(next)) {
8852                 btrfs_err(root->fs_info, "mismatched level");
8853                 ret = -EIO;
8854                 goto out_unlock;
8855         }
8856         path->nodes[level] = next;
8857         path->slots[level] = 0;
8858         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
8859         wc->level = level;
8860         if (wc->level == 1)
8861                 wc->reada_slot = 0;
8862         return 0;
8863 skip:
8864         wc->refs[level - 1] = 0;
8865         wc->flags[level - 1] = 0;
8866         if (wc->stage == DROP_REFERENCE) {
8867                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF) {
8868                         parent = path->nodes[level]->start;
8869                 } else {
8870                         ASSERT(root->root_key.objectid ==
8871                                btrfs_header_owner(path->nodes[level]));
8872                         if (root->root_key.objectid !=
8873                             btrfs_header_owner(path->nodes[level])) {
8874                                 btrfs_err(root->fs_info,
8875                                                 "mismatched block owner");
8876                                 ret = -EIO;
8877                                 goto out_unlock;
8878                         }
8879                         parent = 0;
8880                 }
8881
8882                 if (need_account) {
8883                         ret = btrfs_qgroup_trace_subtree(trans, root, next,
8884                                                          generation, level - 1);
8885                         if (ret) {
8886                                 btrfs_err_rl(fs_info,
8887                                              "Error %d accounting shared subtree. Quota is out of sync, rescan required.",
8888                                              ret);
8889                         }
8890                 }
8891                 ret = btrfs_free_extent(trans, root, bytenr, blocksize,
8892                                         parent, root->root_key.objectid,
8893                                         level - 1, 0);
8894                 if (ret)
8895                         goto out_unlock;
8896         }
8897
8898         *lookup_info = 1;
8899         ret = 1;
8900
8901 out_unlock:
8902         btrfs_tree_unlock(next);
8903         free_extent_buffer(next);
8904
8905         return ret;
8906 }
8907
8908 /*
8909  * helper to process tree block while walking up the tree.
8910  *
8911  * when wc->stage == DROP_REFERENCE, this function drops
8912  * reference count on the block.
8913  *
8914  * when wc->stage == UPDATE_BACKREF, this function changes
8915  * wc->stage back to DROP_REFERENCE if we changed wc->stage
8916  * to UPDATE_BACKREF previously while processing the block.
8917  *
8918  * NOTE: return value 1 means we should stop walking up.
8919  */
8920 static noinline int walk_up_proc(struct btrfs_trans_handle *trans,
8921                                  struct btrfs_root *root,
8922                                  struct btrfs_path *path,
8923                                  struct walk_control *wc)
8924 {
8925         struct btrfs_fs_info *fs_info = root->fs_info;
8926         int ret;
8927         int level = wc->level;
8928         struct extent_buffer *eb = path->nodes[level];
8929         u64 parent = 0;
8930
8931         if (wc->stage == UPDATE_BACKREF) {
8932                 BUG_ON(wc->shared_level < level);
8933                 if (level < wc->shared_level)
8934                         goto out;
8935
8936                 ret = find_next_key(path, level + 1, &wc->update_progress);
8937                 if (ret > 0)
8938                         wc->update_ref = 0;
8939
8940                 wc->stage = DROP_REFERENCE;
8941                 wc->shared_level = -1;
8942                 path->slots[level] = 0;
8943
8944                 /*
8945                  * check reference count again if the block isn't locked.
8946                  * we should start walking down the tree again if reference
8947                  * count is one.
8948                  */
8949                 if (!path->locks[level]) {
8950                         BUG_ON(level == 0);
8951                         btrfs_tree_lock(eb);
8952                         btrfs_set_lock_blocking(eb);
8953                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
8954
8955                         ret = btrfs_lookup_extent_info(trans, fs_info,
8956                                                        eb->start, level, 1,
8957                                                        &wc->refs[level],
8958                                                        &wc->flags[level]);
8959                         if (ret < 0) {
8960                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
8961                                 path->locks[level] = 0;
8962                                 return ret;
8963                         }
8964                         BUG_ON(wc->refs[level] == 0);
8965                         if (wc->refs[level] == 1) {
8966                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
8967                                 path->locks[level] = 0;
8968                                 return 1;
8969                         }
8970                 }
8971         }
8972
8973         /* wc->stage == DROP_REFERENCE */
8974         BUG_ON(wc->refs[level] > 1 && !path->locks[level]);
8975
8976         if (wc->refs[level] == 1) {
8977                 if (level == 0) {
8978                         if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
8979                                 ret = btrfs_dec_ref(trans, root, eb, 1);
8980                         else
8981                                 ret = btrfs_dec_ref(trans, root, eb, 0);
8982                         BUG_ON(ret); /* -ENOMEM */
8983                         ret = btrfs_qgroup_trace_leaf_items(trans, fs_info, eb);
8984                         if (ret) {
8985                                 btrfs_err_rl(fs_info,
8986                                              "error %d accounting leaf items. Quota is out of sync, rescan required.",
8987                                              ret);
8988                         }
8989                 }
8990                 /* make block locked assertion in clean_tree_block happy */
8991                 if (!path->locks[level] &&
8992                     btrfs_header_generation(eb) == trans->transid) {
8993                         btrfs_tree_lock(eb);
8994                         btrfs_set_lock_blocking(eb);
8995                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
8996                 }
8997                 clean_tree_block(fs_info, eb);
8998         }
8999
9000         if (eb == root->node) {
9001                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
9002                         parent = eb->start;
9003                 else
9004                         BUG_ON(root->root_key.objectid !=
9005                                btrfs_header_owner(eb));
9006         } else {
9007                 if (wc->flags[level + 1] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
9008                         parent = path->nodes[level + 1]->start;
9009                 else
9010                         BUG_ON(root->root_key.objectid !=
9011                                btrfs_header_owner(path->nodes[level + 1]));
9012         }
9013
9014         btrfs_free_tree_block(trans, root, eb, parent, wc->refs[level] == 1);
9015 out:
9016         wc->refs[level] = 0;
9017         wc->flags[level] = 0;
9018         return 0;
9019 }
9020
9021 static noinline int walk_down_tree(struct btrfs_trans_handle *trans,
9022                                    struct btrfs_root *root,
9023                                    struct btrfs_path *path,
9024                                    struct walk_control *wc)
9025 {
9026         int level = wc->level;
9027         int lookup_info = 1;
9028         int ret;
9029
9030         while (level >= 0) {
9031                 ret = walk_down_proc(trans, root, path, wc, lookup_info);
9032                 if (ret > 0)
9033                         break;
9034
9035                 if (level == 0)
9036                         break;
9037
9038                 if (path->slots[level] >=
9039                     btrfs_header_nritems(path->nodes[level]))
9040                         break;
9041
9042                 ret = do_walk_down(trans, root, path, wc, &lookup_info);
9043                 if (ret > 0) {
9044                         path->slots[level]++;
9045                         continue;
9046                 } else if (ret < 0)
9047                         return ret;
9048                 level = wc->level;
9049         }
9050         return 0;
9051 }
9052
9053 static noinline int walk_up_tree(struct btrfs_trans_handle *trans,
9054                                  struct btrfs_root *root,
9055                                  struct btrfs_path *path,
9056                                  struct walk_control *wc, int max_level)
9057 {
9058         int level = wc->level;
9059         int ret;
9060
9061         path->slots[level] = btrfs_header_nritems(path->nodes[level]);
9062         while (level < max_level && path->nodes[level]) {
9063                 wc->level = level;
9064                 if (path->slots[level] + 1 <
9065                     btrfs_header_nritems(path->nodes[level])) {
9066                         path->slots[level]++;
9067                         return 0;
9068                 } else {
9069                         ret = walk_up_proc(trans, root, path, wc);
9070                         if (ret > 0)
9071                                 return 0;
9072
9073                         if (path->locks[level]) {
9074                                 btrfs_tree_unlock_rw(path->nodes[level],
9075                                                      path->locks[level]);
9076                                 path->locks[level] = 0;
9077                         }
9078                         free_extent_buffer(path->nodes[level]);
9079                         path->nodes[level] = NULL;
9080                         level++;
9081                 }
9082         }
9083         return 1;
9084 }
9085
9086 /*
9087  * drop a subvolume tree.
9088  *
9089  * this function traverses the tree freeing any blocks that only
9090  * referenced by the tree.
9091  *
9092  * when a shared tree block is found. this function decreases its
9093  * reference count by one. if update_ref is true, this function
9094  * also make sure backrefs for the shared block and all lower level
9095  * blocks are properly updated.
9096  *
9097  * If called with for_reloc == 0, may exit early with -EAGAIN
9098  */
9099 int btrfs_drop_snapshot(struct btrfs_root *root,
9100                          struct btrfs_block_rsv *block_rsv, int update_ref,
9101                          int for_reloc)
9102 {
9103         struct btrfs_fs_info *fs_info = root->fs_info;
9104         struct btrfs_path *path;
9105         struct btrfs_trans_handle *trans;
9106         struct btrfs_root *tree_root = fs_info->tree_root;
9107         struct btrfs_root_item *root_item = &root->root_item;
9108         struct walk_control *wc;
9109         struct btrfs_key key;
9110         int err = 0;
9111         int ret;
9112         int level;
9113         bool root_dropped = false;
9114
9115         btrfs_debug(fs_info, "Drop subvolume %llu", root->objectid);
9116
9117         path = btrfs_alloc_path();
9118         if (!path) {
9119                 err = -ENOMEM;
9120                 goto out;
9121         }
9122
9123         wc = kzalloc(sizeof(*wc), GFP_NOFS);
9124         if (!wc) {
9125                 btrfs_free_path(path);
9126                 err = -ENOMEM;
9127                 goto out;
9128         }
9129
9130         trans = btrfs_start_transaction(tree_root, 0);
9131         if (IS_ERR(trans)) {
9132                 err = PTR_ERR(trans);
9133                 goto out_free;
9134         }
9135
9136         if (block_rsv)
9137                 trans->block_rsv = block_rsv;
9138
9139         if (btrfs_disk_key_objectid(&root_item->drop_progress) == 0) {
9140                 level = btrfs_header_level(root->node);
9141                 path->nodes[level] = btrfs_lock_root_node(root);
9142                 btrfs_set_lock_blocking(path->nodes[level]);
9143                 path->slots[level] = 0;
9144                 path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
9145                 memset(&wc->update_progress, 0,
9146                        sizeof(wc->update_progress));
9147         } else {
9148                 btrfs_disk_key_to_cpu(&key, &root_item->drop_progress);
9149                 memcpy(&wc->update_progress, &key,
9150                        sizeof(wc->update_progress));
9151
9152                 level = root_item->drop_level;
9153                 BUG_ON(level == 0);
9154                 path->lowest_level = level;
9155                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
9156                 path->lowest_level = 0;
9157                 if (ret < 0) {
9158                         err = ret;
9159                         goto out_end_trans;
9160                 }
9161                 WARN_ON(ret > 0);
9162
9163                 /*
9164                  * unlock our path, this is safe because only this
9165                  * function is allowed to delete this snapshot
9166                  */
9167                 btrfs_unlock_up_safe(path, 0);
9168
9169                 level = btrfs_header_level(root->node);
9170                 while (1) {
9171                         btrfs_tree_lock(path->nodes[level]);
9172                         btrfs_set_lock_blocking(path->nodes[level]);
9173                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
9174
9175                         ret = btrfs_lookup_extent_info(trans, fs_info,
9176                                                 path->nodes[level]->start,
9177                                                 level, 1, &wc->refs[level],
9178                                                 &wc->flags[level]);
9179                         if (ret < 0) {
9180                                 err = ret;
9181                                 goto out_end_trans;
9182                         }
9183                         BUG_ON(wc->refs[level] == 0);
9184
9185                         if (level == root_item->drop_level)
9186                                 break;
9187
9188                         btrfs_tree_unlock(path->nodes[level]);
9189                         path->locks[level] = 0;
9190                         WARN_ON(wc->refs[level] != 1);
9191                         level--;
9192                 }
9193         }
9194
9195         wc->level = level;
9196         wc->shared_level = -1;
9197         wc->stage = DROP_REFERENCE;
9198         wc->update_ref = update_ref;
9199         wc->keep_locks = 0;
9200         wc->for_reloc = for_reloc;
9201         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(fs_info);
9202
9203         while (1) {
9204
9205                 ret = walk_down_tree(trans, root, path, wc);
9206                 if (ret < 0) {
9207                         err = ret;
9208                         break;
9209                 }
9210
9211                 ret = walk_up_tree(trans, root, path, wc, BTRFS_MAX_LEVEL);
9212                 if (ret < 0) {
9213                         err = ret;
9214                         break;
9215                 }
9216
9217                 if (ret > 0) {
9218                         BUG_ON(wc->stage != DROP_REFERENCE);
9219                         break;
9220                 }
9221
9222                 if (wc->stage == DROP_REFERENCE) {
9223                         level = wc->level;
9224                         btrfs_node_key(path->nodes[level],
9225                                        &root_item->drop_progress,
9226                                        path->slots[level]);
9227                         root_item->drop_level = level;
9228                 }
9229
9230                 BUG_ON(wc->level == 0);
9231                 if (btrfs_should_end_transaction(trans) ||
9232                     (!for_reloc && btrfs_need_cleaner_sleep(fs_info))) {
9233                         ret = btrfs_update_root(trans, tree_root,
9234                                                 &root->root_key,
9235                                                 root_item);
9236                         if (ret) {
9237                                 btrfs_abort_transaction(trans, ret);
9238                                 err = ret;
9239                                 goto out_end_trans;
9240                         }
9241
9242                         btrfs_end_transaction_throttle(trans);
9243                         if (!for_reloc && btrfs_need_cleaner_sleep(fs_info)) {
9244                                 btrfs_debug(fs_info,
9245                                             "drop snapshot early exit");
9246                                 err = -EAGAIN;
9247                                 goto out_free;
9248                         }
9249
9250                         trans = btrfs_start_transaction(tree_root, 0);
9251                         if (IS_ERR(trans)) {
9252                                 err = PTR_ERR(trans);
9253                                 goto out_free;
9254                         }
9255                         if (block_rsv)
9256                                 trans->block_rsv = block_rsv;
9257                 }
9258         }
9259         btrfs_release_path(path);
9260         if (err)
9261                 goto out_end_trans;
9262
9263         ret = btrfs_del_root(trans, fs_info, &root->root_key);
9264         if (ret) {
9265                 btrfs_abort_transaction(trans, ret);
9266                 err = ret;
9267                 goto out_end_trans;
9268         }
9269
9270         if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) {
9271                 ret = btrfs_find_root(tree_root, &root->root_key, path,
9272                                       NULL, NULL);
9273                 if (ret < 0) {
9274                         btrfs_abort_transaction(trans, ret);
9275                         err = ret;
9276                         goto out_end_trans;
9277                 } else if (ret > 0) {
9278                         /* if we fail to delete the orphan item this time
9279                          * around, it'll get picked up the next time.
9280                          *
9281                          * The most common failure here is just -ENOENT.
9282                          */
9283                         btrfs_del_orphan_item(trans, tree_root,
9284                                               root->root_key.objectid);
9285                 }
9286         }
9287
9288         if (test_bit(BTRFS_ROOT_IN_RADIX, &root->state)) {
9289                 btrfs_add_dropped_root(trans, root);
9290         } else {
9291                 free_extent_buffer(root->node);
9292                 free_extent_buffer(root->commit_root);
9293                 btrfs_put_fs_root(root);
9294         }
9295         root_dropped = true;
9296 out_end_trans:
9297         btrfs_end_transaction_throttle(trans);
9298 out_free:
9299         kfree(wc);
9300         btrfs_free_path(path);
9301 out:
9302         /*
9303          * So if we need to stop dropping the snapshot for whatever reason we
9304          * need to make sure to add it back to the dead root list so that we
9305          * keep trying to do the work later.  This also cleans up roots if we
9306          * don't have it in the radix (like when we recover after a power fail
9307          * or unmount) so we don't leak memory.
9308          */
9309         if (!for_reloc && !root_dropped)
9310                 btrfs_add_dead_root(root);
9311         if (err && err != -EAGAIN)
9312                 btrfs_handle_fs_error(fs_info, err, NULL);
9313         return err;
9314 }
9315
9316 /*
9317  * drop subtree rooted at tree block 'node'.
9318  *
9319  * NOTE: this function will unlock and release tree block 'node'
9320  * only used by relocation code
9321  */
9322 int btrfs_drop_subtree(struct btrfs_trans_handle *trans,
9323                         struct btrfs_root *root,
9324                         struct extent_buffer *node,
9325                         struct extent_buffer *parent)
9326 {
9327         struct btrfs_fs_info *fs_info = root->fs_info;
9328         struct btrfs_path *path;
9329         struct walk_control *wc;
9330         int level;
9331         int parent_level;
9332         int ret = 0;
9333         int wret;
9334
9335         BUG_ON(root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID);
9336
9337         path = btrfs_alloc_path();
9338         if (!path)
9339                 return -ENOMEM;
9340
9341         wc = kzalloc(sizeof(*wc), GFP_NOFS);
9342         if (!wc) {
9343                 btrfs_free_path(path);
9344                 return -ENOMEM;
9345         }
9346
9347         btrfs_assert_tree_locked(parent);
9348         parent_level = btrfs_header_level(parent);
9349         extent_buffer_get(parent);
9350         path->nodes[parent_level] = parent;
9351         path->slots[parent_level] = btrfs_header_nritems(parent);
9352
9353         btrfs_assert_tree_locked(node);
9354         level = btrfs_header_level(node);
9355         path->nodes[level] = node;
9356         path->slots[level] = 0;
9357         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
9358
9359         wc->refs[parent_level] = 1;
9360         wc->flags[parent_level] = BTRFS_BLOCK_FLAG_FULL_BACKREF;
9361         wc->level = level;
9362         wc->shared_level = -1;
9363         wc->stage = DROP_REFERENCE;
9364         wc->update_ref = 0;
9365         wc->keep_locks = 1;
9366         wc->for_reloc = 1;
9367         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(fs_info);
9368
9369         while (1) {
9370                 wret = walk_down_tree(trans, root, path, wc);
9371                 if (wret < 0) {
9372                         ret = wret;
9373                         break;
9374                 }
9375
9376                 wret = walk_up_tree(trans, root, path, wc, parent_level);
9377                 if (wret < 0)
9378                         ret = wret;
9379                 if (wret != 0)
9380                         break;
9381         }
9382
9383         kfree(wc);
9384         btrfs_free_path(path);
9385         return ret;
9386 }
9387
9388 static u64 update_block_group_flags(struct btrfs_fs_info *fs_info, u64 flags)
9389 {
9390         u64 num_devices;
9391         u64 stripped;
9392
9393         /*
9394          * if restripe for this chunk_type is on pick target profile and
9395          * return, otherwise do the usual balance
9396          */
9397         stripped = get_restripe_target(fs_info, flags);
9398         if (stripped)
9399                 return extended_to_chunk(stripped);
9400
9401         num_devices = fs_info->fs_devices->rw_devices;
9402
9403         stripped = BTRFS_BLOCK_GROUP_RAID0 |
9404                 BTRFS_BLOCK_GROUP_RAID5 | BTRFS_BLOCK_GROUP_RAID6 |
9405                 BTRFS_BLOCK_GROUP_RAID1 | BTRFS_BLOCK_GROUP_RAID10;
9406
9407         if (num_devices == 1) {
9408                 stripped |= BTRFS_BLOCK_GROUP_DUP;
9409                 stripped = flags & ~stripped;
9410
9411                 /* turn raid0 into single device chunks */
9412                 if (flags & BTRFS_BLOCK_GROUP_RAID0)
9413                         return stripped;
9414
9415                 /* turn mirroring into duplication */
9416                 if (flags & (BTRFS_BLOCK_GROUP_RAID1 |
9417                              BTRFS_BLOCK_GROUP_RAID10))
9418                         return stripped | BTRFS_BLOCK_GROUP_DUP;
9419         } else {
9420                 /* they already had raid on here, just return */
9421                 if (flags & stripped)
9422                         return flags;
9423
9424                 stripped |= BTRFS_BLOCK_GROUP_DUP;
9425                 stripped = flags & ~stripped;
9426
9427                 /* switch duplicated blocks with raid1 */
9428                 if (flags & BTRFS_BLOCK_GROUP_DUP)
9429                         return stripped | BTRFS_BLOCK_GROUP_RAID1;
9430
9431                 /* this is drive concat, leave it alone */
9432         }
9433
9434         return flags;
9435 }
9436
9437 static int inc_block_group_ro(struct btrfs_block_group_cache *cache, int force)
9438 {
9439         struct btrfs_space_info *sinfo = cache->space_info;
9440         u64 num_bytes;
9441         u64 min_allocable_bytes;
9442         int ret = -ENOSPC;
9443
9444         /*
9445          * We need some metadata space and system metadata space for
9446          * allocating chunks in some corner cases until we force to set
9447          * it to be readonly.
9448          */
9449         if ((sinfo->flags &
9450              (BTRFS_BLOCK_GROUP_SYSTEM | BTRFS_BLOCK_GROUP_METADATA)) &&
9451             !force)
9452                 min_allocable_bytes = SZ_1M;
9453         else
9454                 min_allocable_bytes = 0;
9455
9456         spin_lock(&sinfo->lock);
9457         spin_lock(&cache->lock);
9458
9459         if (cache->ro) {
9460                 cache->ro++;
9461                 ret = 0;
9462                 goto out;
9463         }
9464
9465         num_bytes = cache->key.offset - cache->reserved - cache->pinned -
9466                     cache->bytes_super - btrfs_block_group_used(&cache->item);
9467
9468         if (btrfs_space_info_used(sinfo, true) + num_bytes +
9469             min_allocable_bytes <= sinfo->total_bytes) {
9470                 sinfo->bytes_readonly += num_bytes;
9471                 cache->ro++;
9472                 list_add_tail(&cache->ro_list, &sinfo->ro_bgs);
9473                 ret = 0;
9474         }
9475 out:
9476         spin_unlock(&cache->lock);
9477         spin_unlock(&sinfo->lock);
9478         return ret;
9479 }
9480
9481 int btrfs_inc_block_group_ro(struct btrfs_fs_info *fs_info,
9482                              struct btrfs_block_group_cache *cache)
9483
9484 {
9485         struct btrfs_trans_handle *trans;
9486         u64 alloc_flags;
9487         int ret;
9488
9489 again:
9490         trans = btrfs_join_transaction(fs_info->extent_root);
9491         if (IS_ERR(trans))
9492                 return PTR_ERR(trans);
9493
9494         /*
9495          * we're not allowed to set block groups readonly after the dirty
9496          * block groups cache has started writing.  If it already started,
9497          * back off and let this transaction commit
9498          */
9499         mutex_lock(&fs_info->ro_block_group_mutex);
9500         if (test_bit(BTRFS_TRANS_DIRTY_BG_RUN, &trans->transaction->flags)) {
9501                 u64 transid = trans->transid;
9502
9503                 mutex_unlock(&fs_info->ro_block_group_mutex);
9504                 btrfs_end_transaction(trans);
9505
9506                 ret = btrfs_wait_for_commit(fs_info, transid);
9507                 if (ret)
9508                         return ret;
9509                 goto again;
9510         }
9511
9512         /*
9513          * if we are changing raid levels, try to allocate a corresponding
9514          * block group with the new raid level.
9515          */
9516         alloc_flags = update_block_group_flags(fs_info, cache->flags);
9517         if (alloc_flags != cache->flags) {
9518                 ret = do_chunk_alloc(trans, fs_info, alloc_flags,
9519                                      CHUNK_ALLOC_FORCE);
9520                 /*
9521                  * ENOSPC is allowed here, we may have enough space
9522                  * already allocated at the new raid level to
9523                  * carry on
9524                  */
9525                 if (ret == -ENOSPC)
9526                         ret = 0;
9527                 if (ret < 0)
9528                         goto out;
9529         }
9530
9531         ret = inc_block_group_ro(cache, 0);
9532         if (!ret)
9533                 goto out;
9534         alloc_flags = get_alloc_profile(fs_info, cache->space_info->flags);
9535         ret = do_chunk_alloc(trans, fs_info, alloc_flags,
9536                              CHUNK_ALLOC_FORCE);
9537         if (ret < 0)
9538                 goto out;
9539         ret = inc_block_group_ro(cache, 0);
9540 out:
9541         if (cache->flags & BTRFS_BLOCK_GROUP_SYSTEM) {
9542                 alloc_flags = update_block_group_flags(fs_info, cache->flags);
9543                 mutex_lock(&fs_info->chunk_mutex);
9544                 check_system_chunk(trans, fs_info, alloc_flags);
9545                 mutex_unlock(&fs_info->chunk_mutex);
9546         }
9547         mutex_unlock(&fs_info->ro_block_group_mutex);
9548
9549         btrfs_end_transaction(trans);
9550         return ret;
9551 }
9552
9553 int btrfs_force_chunk_alloc(struct btrfs_trans_handle *trans,
9554                             struct btrfs_fs_info *fs_info, u64 type)
9555 {
9556         u64 alloc_flags = get_alloc_profile(fs_info, type);
9557
9558         return do_chunk_alloc(trans, fs_info, alloc_flags, CHUNK_ALLOC_FORCE);
9559 }
9560
9561 /*
9562  * helper to account the unused space of all the readonly block group in the
9563  * space_info. takes mirrors into account.
9564  */
9565 u64 btrfs_account_ro_block_groups_free_space(struct btrfs_space_info *sinfo)
9566 {
9567         struct btrfs_block_group_cache *block_group;
9568         u64 free_bytes = 0;
9569         int factor;
9570
9571         /* It's df, we don't care if it's racy */
9572         if (list_empty(&sinfo->ro_bgs))
9573                 return 0;
9574
9575         spin_lock(&sinfo->lock);
9576         list_for_each_entry(block_group, &sinfo->ro_bgs, ro_list) {
9577                 spin_lock(&block_group->lock);
9578
9579                 if (!block_group->ro) {
9580                         spin_unlock(&block_group->lock);
9581                         continue;
9582                 }
9583
9584                 if (block_group->flags & (BTRFS_BLOCK_GROUP_RAID1 |
9585                                           BTRFS_BLOCK_GROUP_RAID10 |
9586                                           BTRFS_BLOCK_GROUP_DUP))
9587                         factor = 2;
9588                 else
9589                         factor = 1;
9590
9591                 free_bytes += (block_group->key.offset -
9592                                btrfs_block_group_used(&block_group->item)) *
9593                                factor;
9594
9595                 spin_unlock(&block_group->lock);
9596         }
9597         spin_unlock(&sinfo->lock);
9598
9599         return free_bytes;
9600 }
9601
9602 void btrfs_dec_block_group_ro(struct btrfs_block_group_cache *cache)
9603 {
9604         struct btrfs_space_info *sinfo = cache->space_info;
9605         u64 num_bytes;
9606
9607         BUG_ON(!cache->ro);
9608
9609         spin_lock(&sinfo->lock);
9610         spin_lock(&cache->lock);
9611         if (!--cache->ro) {
9612                 num_bytes = cache->key.offset - cache->reserved -
9613                             cache->pinned - cache->bytes_super -
9614                             btrfs_block_group_used(&cache->item);
9615                 sinfo->bytes_readonly -= num_bytes;
9616                 list_del_init(&cache->ro_list);
9617         }
9618         spin_unlock(&cache->lock);
9619         spin_unlock(&sinfo->lock);
9620 }
9621
9622 /*
9623  * checks to see if its even possible to relocate this block group.
9624  *
9625  * @return - -1 if it's not a good idea to relocate this block group, 0 if its
9626  * ok to go ahead and try.
9627  */
9628 int btrfs_can_relocate(struct btrfs_fs_info *fs_info, u64 bytenr)
9629 {
9630         struct btrfs_root *root = fs_info->extent_root;
9631         struct btrfs_block_group_cache *block_group;
9632         struct btrfs_space_info *space_info;
9633         struct btrfs_fs_devices *fs_devices = fs_info->fs_devices;
9634         struct btrfs_device *device;
9635         struct btrfs_trans_handle *trans;
9636         u64 min_free;
9637         u64 dev_min = 1;
9638         u64 dev_nr = 0;
9639         u64 target;
9640         int debug;
9641         int index;
9642         int full = 0;
9643         int ret = 0;
9644
9645         debug = btrfs_test_opt(fs_info, ENOSPC_DEBUG);
9646
9647         block_group = btrfs_lookup_block_group(fs_info, bytenr);
9648
9649         /* odd, couldn't find the block group, leave it alone */
9650         if (!block_group) {
9651                 if (debug)
9652                         btrfs_warn(fs_info,
9653                                    "can't find block group for bytenr %llu",
9654                                    bytenr);
9655                 return -1;
9656         }
9657
9658         min_free = btrfs_block_group_used(&block_group->item);
9659
9660         /* no bytes used, we're good */
9661         if (!min_free)
9662                 goto out;
9663
9664         space_info = block_group->space_info;
9665         spin_lock(&space_info->lock);
9666
9667         full = space_info->full;
9668
9669         /*
9670          * if this is the last block group we have in this space, we can't
9671          * relocate it unless we're able to allocate a new chunk below.
9672          *
9673          * Otherwise, we need to make sure we have room in the space to handle
9674          * all of the extents from this block group.  If we can, we're good
9675          */
9676         if ((space_info->total_bytes != block_group->key.offset) &&
9677             (btrfs_space_info_used(space_info, false) + min_free <
9678              space_info->total_bytes)) {
9679                 spin_unlock(&space_info->lock);
9680                 goto out;
9681         }
9682         spin_unlock(&space_info->lock);
9683
9684         /*
9685          * ok we don't have enough space, but maybe we have free space on our
9686          * devices to allocate new chunks for relocation, so loop through our
9687          * alloc devices and guess if we have enough space.  if this block
9688          * group is going to be restriped, run checks against the target
9689          * profile instead of the current one.
9690          */
9691         ret = -1;
9692
9693         /*
9694          * index:
9695          *      0: raid10
9696          *      1: raid1
9697          *      2: dup
9698          *      3: raid0
9699          *      4: single
9700          */
9701         target = get_restripe_target(fs_info, block_group->flags);
9702         if (target) {
9703                 index = btrfs_bg_flags_to_raid_index(extended_to_chunk(target));
9704         } else {
9705                 /*
9706                  * this is just a balance, so if we were marked as full
9707                  * we know there is no space for a new chunk
9708                  */
9709                 if (full) {
9710                         if (debug)
9711                                 btrfs_warn(fs_info,
9712                                            "no space to alloc new chunk for block group %llu",
9713                                            block_group->key.objectid);
9714                         goto out;
9715                 }
9716
9717                 index = btrfs_bg_flags_to_raid_index(block_group->flags);
9718         }
9719
9720         if (index == BTRFS_RAID_RAID10) {
9721                 dev_min = 4;
9722                 /* Divide by 2 */
9723                 min_free >>= 1;
9724         } else if (index == BTRFS_RAID_RAID1) {
9725                 dev_min = 2;
9726         } else if (index == BTRFS_RAID_DUP) {
9727                 /* Multiply by 2 */
9728                 min_free <<= 1;
9729         } else if (index == BTRFS_RAID_RAID0) {
9730                 dev_min = fs_devices->rw_devices;
9731                 min_free = div64_u64(min_free, dev_min);
9732         }
9733
9734         /* We need to do this so that we can look at pending chunks */
9735         trans = btrfs_join_transaction(root);
9736         if (IS_ERR(trans)) {
9737                 ret = PTR_ERR(trans);
9738                 goto out;
9739         }
9740
9741         mutex_lock(&fs_info->chunk_mutex);
9742         list_for_each_entry(device, &fs_devices->alloc_list, dev_alloc_list) {
9743                 u64 dev_offset;
9744
9745                 /*
9746                  * check to make sure we can actually find a chunk with enough
9747                  * space to fit our block group in.
9748                  */
9749                 if (device->total_bytes > device->bytes_used + min_free &&
9750                     !test_bit(BTRFS_DEV_STATE_REPLACE_TGT, &device->dev_state)) {
9751                         ret = find_free_dev_extent(trans, device, min_free,
9752                                                    &dev_offset, NULL);
9753                         if (!ret)
9754                                 dev_nr++;
9755
9756                         if (dev_nr >= dev_min)
9757                                 break;
9758
9759                         ret = -1;
9760                 }
9761         }
9762         if (debug && ret == -1)
9763                 btrfs_warn(fs_info,
9764                            "no space to allocate a new chunk for block group %llu",
9765                            block_group->key.objectid);
9766         mutex_unlock(&fs_info->chunk_mutex);
9767         btrfs_end_transaction(trans);
9768 out:
9769         btrfs_put_block_group(block_group);
9770         return ret;
9771 }
9772
9773 static int find_first_block_group(struct btrfs_fs_info *fs_info,
9774                                   struct btrfs_path *path,
9775                                   struct btrfs_key *key)
9776 {
9777         struct btrfs_root *root = fs_info->extent_root;
9778         int ret = 0;
9779         struct btrfs_key found_key;
9780         struct extent_buffer *leaf;
9781         int slot;
9782
9783         ret = btrfs_search_slot(NULL, root, key, path, 0, 0);
9784         if (ret < 0)
9785                 goto out;
9786
9787         while (1) {
9788                 slot = path->slots[0];
9789                 leaf = path->nodes[0];
9790                 if (slot >= btrfs_header_nritems(leaf)) {
9791                         ret = btrfs_next_leaf(root, path);
9792                         if (ret == 0)
9793                                 continue;
9794                         if (ret < 0)
9795                                 goto out;
9796                         break;
9797                 }
9798                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
9799
9800                 if (found_key.objectid >= key->objectid &&
9801                     found_key.type == BTRFS_BLOCK_GROUP_ITEM_KEY) {
9802                         struct extent_map_tree *em_tree;
9803                         struct extent_map *em;
9804
9805                         em_tree = &root->fs_info->mapping_tree.map_tree;
9806                         read_lock(&em_tree->lock);
9807                         em = lookup_extent_mapping(em_tree, found_key.objectid,
9808                                                    found_key.offset);
9809                         read_unlock(&em_tree->lock);
9810                         if (!em) {
9811                                 btrfs_err(fs_info,
9812                         "logical %llu len %llu found bg but no related chunk",
9813                                           found_key.objectid, found_key.offset);
9814                                 ret = -ENOENT;
9815                         } else {
9816                                 ret = 0;
9817                         }
9818                         free_extent_map(em);
9819                         goto out;
9820                 }
9821                 path->slots[0]++;
9822         }
9823 out:
9824         return ret;
9825 }
9826
9827 void btrfs_put_block_group_cache(struct btrfs_fs_info *info)
9828 {
9829         struct btrfs_block_group_cache *block_group;
9830         u64 last = 0;
9831
9832         while (1) {
9833                 struct inode *inode;
9834
9835                 block_group = btrfs_lookup_first_block_group(info, last);
9836                 while (block_group) {
9837                         spin_lock(&block_group->lock);
9838                         if (block_group->iref)
9839                                 break;
9840                         spin_unlock(&block_group->lock);
9841                         block_group = next_block_group(info, block_group);
9842                 }
9843                 if (!block_group) {
9844                         if (last == 0)
9845                                 break;
9846                         last = 0;
9847                         continue;
9848                 }
9849
9850                 inode = block_group->inode;
9851                 block_group->iref = 0;
9852                 block_group->inode = NULL;
9853                 spin_unlock(&block_group->lock);
9854                 ASSERT(block_group->io_ctl.inode == NULL);
9855                 iput(inode);
9856                 last = block_group->key.objectid + block_group->key.offset;
9857                 btrfs_put_block_group(block_group);
9858         }
9859 }
9860
9861 /*
9862  * Must be called only after stopping all workers, since we could have block
9863  * group caching kthreads running, and therefore they could race with us if we
9864  * freed the block groups before stopping them.
9865  */
9866 int btrfs_free_block_groups(struct btrfs_fs_info *info)
9867 {
9868         struct btrfs_block_group_cache *block_group;
9869         struct btrfs_space_info *space_info;
9870         struct btrfs_caching_control *caching_ctl;
9871         struct rb_node *n;
9872
9873         down_write(&info->commit_root_sem);
9874         while (!list_empty(&info->caching_block_groups)) {
9875                 caching_ctl = list_entry(info->caching_block_groups.next,
9876                                          struct btrfs_caching_control, list);
9877                 list_del(&caching_ctl->list);
9878                 put_caching_control(caching_ctl);
9879         }
9880         up_write(&info->commit_root_sem);
9881
9882         spin_lock(&info->unused_bgs_lock);
9883         while (!list_empty(&info->unused_bgs)) {
9884                 block_group = list_first_entry(&info->unused_bgs,
9885                                                struct btrfs_block_group_cache,
9886                                                bg_list);
9887                 list_del_init(&block_group->bg_list);
9888                 btrfs_put_block_group(block_group);
9889         }
9890         spin_unlock(&info->unused_bgs_lock);
9891
9892         spin_lock(&info->block_group_cache_lock);
9893         while ((n = rb_last(&info->block_group_cache_tree)) != NULL) {
9894                 block_group = rb_entry(n, struct btrfs_block_group_cache,
9895                                        cache_node);
9896                 rb_erase(&block_group->cache_node,
9897                          &info->block_group_cache_tree);
9898                 RB_CLEAR_NODE(&block_group->cache_node);
9899                 spin_unlock(&info->block_group_cache_lock);
9900
9901                 down_write(&block_group->space_info->groups_sem);
9902                 list_del(&block_group->list);
9903                 up_write(&block_group->space_info->groups_sem);
9904
9905                 /*
9906                  * We haven't cached this block group, which means we could
9907                  * possibly have excluded extents on this block group.
9908                  */
9909                 if (block_group->cached == BTRFS_CACHE_NO ||
9910                     block_group->cached == BTRFS_CACHE_ERROR)
9911                         free_excluded_extents(info, block_group);
9912
9913                 btrfs_remove_free_space_cache(block_group);
9914                 ASSERT(block_group->cached != BTRFS_CACHE_STARTED);
9915                 ASSERT(list_empty(&block_group->dirty_list));
9916                 ASSERT(list_empty(&block_group->io_list));
9917                 ASSERT(list_empty(&block_group->bg_list));
9918                 ASSERT(atomic_read(&block_group->count) == 1);
9919                 btrfs_put_block_group(block_group);
9920
9921                 spin_lock(&info->block_group_cache_lock);
9922         }
9923         spin_unlock(&info->block_group_cache_lock);
9924
9925         /* now that all the block groups are freed, go through and
9926          * free all the space_info structs.  This is only called during
9927          * the final stages of unmount, and so we know nobody is
9928          * using them.  We call synchronize_rcu() once before we start,
9929          * just to be on the safe side.
9930          */
9931         synchronize_rcu();
9932
9933         release_global_block_rsv(info);
9934
9935         while (!list_empty(&info->space_info)) {
9936                 int i;
9937
9938                 space_info = list_entry(info->space_info.next,
9939                                         struct btrfs_space_info,
9940                                         list);
9941
9942                 /*
9943                  * Do not hide this behind enospc_debug, this is actually
9944                  * important and indicates a real bug if this happens.
9945                  */
9946                 if (WARN_ON(space_info->bytes_pinned > 0 ||
9947                             space_info->bytes_reserved > 0 ||
9948                             space_info->bytes_may_use > 0))
9949                         dump_space_info(info, space_info, 0, 0);
9950                 list_del(&space_info->list);
9951                 for (i = 0; i < BTRFS_NR_RAID_TYPES; i++) {
9952                         struct kobject *kobj;
9953                         kobj = space_info->block_group_kobjs[i];
9954                         space_info->block_group_kobjs[i] = NULL;
9955                         if (kobj) {
9956                                 kobject_del(kobj);
9957                                 kobject_put(kobj);
9958                         }
9959                 }
9960                 kobject_del(&space_info->kobj);
9961                 kobject_put(&space_info->kobj);
9962         }
9963         return 0;
9964 }
9965
9966 /* link_block_group will queue up kobjects to add when we're reclaim-safe */
9967 void btrfs_add_raid_kobjects(struct btrfs_fs_info *fs_info)
9968 {
9969         struct btrfs_space_info *space_info;
9970         struct raid_kobject *rkobj;
9971         LIST_HEAD(list);
9972         int index;
9973         int ret = 0;
9974
9975         spin_lock(&fs_info->pending_raid_kobjs_lock);
9976         list_splice_init(&fs_info->pending_raid_kobjs, &list);
9977         spin_unlock(&fs_info->pending_raid_kobjs_lock);
9978
9979         list_for_each_entry(rkobj, &list, list) {
9980                 space_info = __find_space_info(fs_info, rkobj->flags);
9981                 index = btrfs_bg_flags_to_raid_index(rkobj->flags);
9982
9983                 ret = kobject_add(&rkobj->kobj, &space_info->kobj,
9984                                   "%s", get_raid_name(index));
9985                 if (ret) {
9986                         kobject_put(&rkobj->kobj);
9987                         break;
9988                 }
9989         }
9990         if (ret)
9991                 btrfs_warn(fs_info,
9992                            "failed to add kobject for block cache, ignoring");
9993 }
9994
9995 static void link_block_group(struct btrfs_block_group_cache *cache)
9996 {
9997         struct btrfs_space_info *space_info = cache->space_info;
9998         struct btrfs_fs_info *fs_info = cache->fs_info;
9999         int index = btrfs_bg_flags_to_raid_index(cache->flags);
10000         bool first = false;
10001
10002         down_write(&space_info->groups_sem);
10003         if (list_empty(&space_info->block_groups[index]))
10004                 first = true;
10005         list_add_tail(&cache->list, &space_info->block_groups[index]);
10006         up_write(&space_info->groups_sem);
10007
10008         if (first) {
10009                 struct raid_kobject *rkobj = kzalloc(sizeof(*rkobj), GFP_NOFS);
10010                 if (!rkobj) {
10011                         btrfs_warn(cache->fs_info,
10012                                 "couldn't alloc memory for raid level kobject");
10013                         return;
10014                 }
10015                 rkobj->flags = cache->flags;
10016                 kobject_init(&rkobj->kobj, &btrfs_raid_ktype);
10017
10018                 spin_lock(&fs_info->pending_raid_kobjs_lock);
10019                 list_add_tail(&rkobj->list, &fs_info->pending_raid_kobjs);
10020                 spin_unlock(&fs_info->pending_raid_kobjs_lock);
10021                 space_info->block_group_kobjs[index] = &rkobj->kobj;
10022         }
10023 }
10024
10025 static struct btrfs_block_group_cache *
10026 btrfs_create_block_group_cache(struct btrfs_fs_info *fs_info,
10027                                u64 start, u64 size)
10028 {
10029         struct btrfs_block_group_cache *cache;
10030
10031         cache = kzalloc(sizeof(*cache), GFP_NOFS);
10032         if (!cache)
10033                 return NULL;
10034
10035         cache->free_space_ctl = kzalloc(sizeof(*cache->free_space_ctl),
10036                                         GFP_NOFS);
10037         if (!cache->free_space_ctl) {
10038                 kfree(cache);
10039                 return NULL;
10040         }
10041
10042         cache->key.objectid = start;
10043         cache->key.offset = size;
10044         cache->key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
10045
10046         cache->fs_info = fs_info;
10047         cache->full_stripe_len = btrfs_full_stripe_len(fs_info, start);
10048         set_free_space_tree_thresholds(cache);
10049
10050         atomic_set(&cache->count, 1);
10051         spin_lock_init(&cache->lock);
10052         init_rwsem(&cache->data_rwsem);
10053         INIT_LIST_HEAD(&cache->list);
10054         INIT_LIST_HEAD(&cache->cluster_list);
10055         INIT_LIST_HEAD(&cache->bg_list);
10056         INIT_LIST_HEAD(&cache->ro_list);
10057         INIT_LIST_HEAD(&cache->dirty_list);
10058         INIT_LIST_HEAD(&cache->io_list);
10059         btrfs_init_free_space_ctl(cache);
10060         atomic_set(&cache->trimming, 0);
10061         mutex_init(&cache->free_space_lock);
10062         btrfs_init_full_stripe_locks_tree(&cache->full_stripe_locks_root);
10063
10064         return cache;
10065 }
10066
10067 int btrfs_read_block_groups(struct btrfs_fs_info *info)
10068 {
10069         struct btrfs_path *path;
10070         int ret;
10071         struct btrfs_block_group_cache *cache;
10072         struct btrfs_space_info *space_info;
10073         struct btrfs_key key;
10074         struct btrfs_key found_key;
10075         struct extent_buffer *leaf;
10076         int need_clear = 0;
10077         u64 cache_gen;
10078         u64 feature;
10079         int mixed;
10080
10081         feature = btrfs_super_incompat_flags(info->super_copy);
10082         mixed = !!(feature & BTRFS_FEATURE_INCOMPAT_MIXED_GROUPS);
10083
10084         key.objectid = 0;
10085         key.offset = 0;
10086         key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
10087         path = btrfs_alloc_path();
10088         if (!path)
10089                 return -ENOMEM;
10090         path->reada = READA_FORWARD;
10091
10092         cache_gen = btrfs_super_cache_generation(info->super_copy);
10093         if (btrfs_test_opt(info, SPACE_CACHE) &&
10094             btrfs_super_generation(info->super_copy) != cache_gen)
10095                 need_clear = 1;
10096         if (btrfs_test_opt(info, CLEAR_CACHE))
10097                 need_clear = 1;
10098
10099         while (1) {
10100                 ret = find_first_block_group(info, path, &key);
10101                 if (ret > 0)
10102                         break;
10103                 if (ret != 0)
10104                         goto error;
10105
10106                 leaf = path->nodes[0];
10107                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
10108
10109                 cache = btrfs_create_block_group_cache(info, found_key.objectid,
10110                                                        found_key.offset);
10111                 if (!cache) {
10112                         ret = -ENOMEM;
10113                         goto error;
10114                 }
10115
10116                 if (need_clear) {
10117                         /*
10118                          * When we mount with old space cache, we need to
10119                          * set BTRFS_DC_CLEAR and set dirty flag.
10120                          *
10121                          * a) Setting 'BTRFS_DC_CLEAR' makes sure that we
10122                          *    truncate the old free space cache inode and
10123                          *    setup a new one.
10124                          * b) Setting 'dirty flag' makes sure that we flush
10125                          *    the new space cache info onto disk.
10126                          */
10127                         if (btrfs_test_opt(info, SPACE_CACHE))
10128                                 cache->disk_cache_state = BTRFS_DC_CLEAR;
10129                 }
10130
10131                 read_extent_buffer(leaf, &cache->item,
10132                                    btrfs_item_ptr_offset(leaf, path->slots[0]),
10133                                    sizeof(cache->item));
10134                 cache->flags = btrfs_block_group_flags(&cache->item);
10135                 if (!mixed &&
10136                     ((cache->flags & BTRFS_BLOCK_GROUP_METADATA) &&
10137                     (cache->flags & BTRFS_BLOCK_GROUP_DATA))) {
10138                         btrfs_err(info,
10139 "bg %llu is a mixed block group but filesystem hasn't enabled mixed block groups",
10140                                   cache->key.objectid);
10141                         ret = -EINVAL;
10142                         goto error;
10143                 }
10144
10145                 key.objectid = found_key.objectid + found_key.offset;
10146                 btrfs_release_path(path);
10147
10148                 /*
10149                  * We need to exclude the super stripes now so that the space
10150                  * info has super bytes accounted for, otherwise we'll think
10151                  * we have more space than we actually do.
10152                  */
10153                 ret = exclude_super_stripes(info, cache);
10154                 if (ret) {
10155                         /*
10156                          * We may have excluded something, so call this just in
10157                          * case.
10158                          */
10159                         free_excluded_extents(info, cache);
10160                         btrfs_put_block_group(cache);
10161                         goto error;
10162                 }
10163
10164                 /*
10165                  * check for two cases, either we are full, and therefore
10166                  * don't need to bother with the caching work since we won't
10167                  * find any space, or we are empty, and we can just add all
10168                  * the space in and be done with it.  This saves us _alot_ of
10169                  * time, particularly in the full case.
10170                  */
10171                 if (found_key.offset == btrfs_block_group_used(&cache->item)) {
10172                         cache->last_byte_to_unpin = (u64)-1;
10173                         cache->cached = BTRFS_CACHE_FINISHED;
10174                         free_excluded_extents(info, cache);
10175                 } else if (btrfs_block_group_used(&cache->item) == 0) {
10176                         cache->last_byte_to_unpin = (u64)-1;
10177                         cache->cached = BTRFS_CACHE_FINISHED;
10178                         add_new_free_space(cache, info,
10179                                            found_key.objectid,
10180                                            found_key.objectid +
10181                                            found_key.offset);
10182                         free_excluded_extents(info, cache);
10183                 }
10184
10185                 ret = btrfs_add_block_group_cache(info, cache);
10186                 if (ret) {
10187                         btrfs_remove_free_space_cache(cache);
10188                         btrfs_put_block_group(cache);
10189                         goto error;
10190                 }
10191
10192                 trace_btrfs_add_block_group(info, cache, 0);
10193                 update_space_info(info, cache->flags, found_key.offset,
10194                                   btrfs_block_group_used(&cache->item),
10195                                   cache->bytes_super, &space_info);
10196
10197                 cache->space_info = space_info;
10198
10199                 link_block_group(cache);
10200
10201                 set_avail_alloc_bits(info, cache->flags);
10202                 if (btrfs_chunk_readonly(info, cache->key.objectid)) {
10203                         inc_block_group_ro(cache, 1);
10204                 } else if (btrfs_block_group_used(&cache->item) == 0) {
10205                         spin_lock(&info->unused_bgs_lock);
10206                         /* Should always be true but just in case. */
10207                         if (list_empty(&cache->bg_list)) {
10208                                 btrfs_get_block_group(cache);
10209                                 list_add_tail(&cache->bg_list,
10210                                               &info->unused_bgs);
10211                         }
10212                         spin_unlock(&info->unused_bgs_lock);
10213                 }
10214         }
10215
10216         list_for_each_entry_rcu(space_info, &info->space_info, list) {
10217                 if (!(get_alloc_profile(info, space_info->flags) &
10218                       (BTRFS_BLOCK_GROUP_RAID10 |
10219                        BTRFS_BLOCK_GROUP_RAID1 |
10220                        BTRFS_BLOCK_GROUP_RAID5 |
10221                        BTRFS_BLOCK_GROUP_RAID6 |
10222                        BTRFS_BLOCK_GROUP_DUP)))
10223                         continue;
10224                 /*
10225                  * avoid allocating from un-mirrored block group if there are
10226                  * mirrored block groups.
10227                  */
10228                 list_for_each_entry(cache,
10229                                 &space_info->block_groups[BTRFS_RAID_RAID0],
10230                                 list)
10231                         inc_block_group_ro(cache, 1);
10232                 list_for_each_entry(cache,
10233                                 &space_info->block_groups[BTRFS_RAID_SINGLE],
10234                                 list)
10235                         inc_block_group_ro(cache, 1);
10236         }
10237
10238         btrfs_add_raid_kobjects(info);
10239         init_global_block_rsv(info);
10240         ret = 0;
10241 error:
10242         btrfs_free_path(path);
10243         return ret;
10244 }
10245
10246 void btrfs_create_pending_block_groups(struct btrfs_trans_handle *trans)
10247 {
10248         struct btrfs_fs_info *fs_info = trans->fs_info;
10249         struct btrfs_block_group_cache *block_group, *tmp;
10250         struct btrfs_root *extent_root = fs_info->extent_root;
10251         struct btrfs_block_group_item item;
10252         struct btrfs_key key;
10253         int ret = 0;
10254         bool can_flush_pending_bgs = trans->can_flush_pending_bgs;
10255
10256         trans->can_flush_pending_bgs = false;
10257         list_for_each_entry_safe(block_group, tmp, &trans->new_bgs, bg_list) {
10258                 if (ret)
10259                         goto next;
10260
10261                 spin_lock(&block_group->lock);
10262                 memcpy(&item, &block_group->item, sizeof(item));
10263                 memcpy(&key, &block_group->key, sizeof(key));
10264                 spin_unlock(&block_group->lock);
10265
10266                 ret = btrfs_insert_item(trans, extent_root, &key, &item,
10267                                         sizeof(item));
10268                 if (ret)
10269                         btrfs_abort_transaction(trans, ret);
10270                 ret = btrfs_finish_chunk_alloc(trans, fs_info, key.objectid,
10271                                                key.offset);
10272                 if (ret)
10273                         btrfs_abort_transaction(trans, ret);
10274                 add_block_group_free_space(trans, fs_info, block_group);
10275                 /* already aborted the transaction if it failed. */
10276 next:
10277                 list_del_init(&block_group->bg_list);
10278         }
10279         trans->can_flush_pending_bgs = can_flush_pending_bgs;
10280 }
10281
10282 int btrfs_make_block_group(struct btrfs_trans_handle *trans,
10283                            struct btrfs_fs_info *fs_info, u64 bytes_used,
10284                            u64 type, u64 chunk_offset, u64 size)
10285 {
10286         struct btrfs_block_group_cache *cache;
10287         int ret;
10288
10289         btrfs_set_log_full_commit(fs_info, trans);
10290
10291         cache = btrfs_create_block_group_cache(fs_info, chunk_offset, size);
10292         if (!cache)
10293                 return -ENOMEM;
10294
10295         btrfs_set_block_group_used(&cache->item, bytes_used);
10296         btrfs_set_block_group_chunk_objectid(&cache->item,
10297                                              BTRFS_FIRST_CHUNK_TREE_OBJECTID);
10298         btrfs_set_block_group_flags(&cache->item, type);
10299
10300         cache->flags = type;
10301         cache->last_byte_to_unpin = (u64)-1;
10302         cache->cached = BTRFS_CACHE_FINISHED;
10303         cache->needs_free_space = 1;
10304         ret = exclude_super_stripes(fs_info, cache);
10305         if (ret) {
10306                 /*
10307                  * We may have excluded something, so call this just in
10308                  * case.
10309                  */
10310                 free_excluded_extents(fs_info, cache);
10311                 btrfs_put_block_group(cache);
10312                 return ret;
10313         }
10314
10315         add_new_free_space(cache, fs_info, chunk_offset, chunk_offset + size);
10316
10317         free_excluded_extents(fs_info, cache);
10318
10319 #ifdef CONFIG_BTRFS_DEBUG
10320         if (btrfs_should_fragment_free_space(cache)) {
10321                 u64 new_bytes_used = size - bytes_used;
10322
10323                 bytes_used += new_bytes_used >> 1;
10324                 fragment_free_space(cache);
10325         }
10326 #endif
10327         /*
10328          * Ensure the corresponding space_info object is created and
10329          * assigned to our block group. We want our bg to be added to the rbtree
10330          * with its ->space_info set.
10331          */
10332         cache->space_info = __find_space_info(fs_info, cache->flags);
10333         ASSERT(cache->space_info);
10334
10335         ret = btrfs_add_block_group_cache(fs_info, cache);
10336         if (ret) {
10337                 btrfs_remove_free_space_cache(cache);
10338                 btrfs_put_block_group(cache);
10339                 return ret;
10340         }
10341
10342         /*
10343          * Now that our block group has its ->space_info set and is inserted in
10344          * the rbtree, update the space info's counters.
10345          */
10346         trace_btrfs_add_block_group(fs_info, cache, 1);
10347         update_space_info(fs_info, cache->flags, size, bytes_used,
10348                                 cache->bytes_super, &cache->space_info);
10349         update_global_block_rsv(fs_info);
10350
10351         link_block_group(cache);
10352
10353         list_add_tail(&cache->bg_list, &trans->new_bgs);
10354
10355         set_avail_alloc_bits(fs_info, type);
10356         return 0;
10357 }
10358
10359 static void clear_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
10360 {
10361         u64 extra_flags = chunk_to_extended(flags) &
10362                                 BTRFS_EXTENDED_PROFILE_MASK;
10363
10364         write_seqlock(&fs_info->profiles_lock);
10365         if (flags & BTRFS_BLOCK_GROUP_DATA)
10366                 fs_info->avail_data_alloc_bits &= ~extra_flags;
10367         if (flags & BTRFS_BLOCK_GROUP_METADATA)
10368                 fs_info->avail_metadata_alloc_bits &= ~extra_flags;
10369         if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
10370                 fs_info->avail_system_alloc_bits &= ~extra_flags;
10371         write_sequnlock(&fs_info->profiles_lock);
10372 }
10373
10374 int btrfs_remove_block_group(struct btrfs_trans_handle *trans,
10375                              struct btrfs_fs_info *fs_info, u64 group_start,
10376                              struct extent_map *em)
10377 {
10378         struct btrfs_root *root = fs_info->extent_root;
10379         struct btrfs_path *path;
10380         struct btrfs_block_group_cache *block_group;
10381         struct btrfs_free_cluster *cluster;
10382         struct btrfs_root *tree_root = fs_info->tree_root;
10383         struct btrfs_key key;
10384         struct inode *inode;
10385         struct kobject *kobj = NULL;
10386         int ret;
10387         int index;
10388         int factor;
10389         struct btrfs_caching_control *caching_ctl = NULL;
10390         bool remove_em;
10391
10392         block_group = btrfs_lookup_block_group(fs_info, group_start);
10393         BUG_ON(!block_group);
10394         BUG_ON(!block_group->ro);
10395
10396         /*
10397          * Free the reserved super bytes from this block group before
10398          * remove it.
10399          */
10400         free_excluded_extents(fs_info, block_group);
10401         btrfs_free_ref_tree_range(fs_info, block_group->key.objectid,
10402                                   block_group->key.offset);
10403
10404         memcpy(&key, &block_group->key, sizeof(key));
10405         index = btrfs_bg_flags_to_raid_index(block_group->flags);
10406         if (block_group->flags & (BTRFS_BLOCK_GROUP_DUP |
10407                                   BTRFS_BLOCK_GROUP_RAID1 |
10408                                   BTRFS_BLOCK_GROUP_RAID10))
10409                 factor = 2;
10410         else
10411                 factor = 1;
10412
10413         /* make sure this block group isn't part of an allocation cluster */
10414         cluster = &fs_info->data_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         /*
10420          * make sure this block group isn't part of a metadata
10421          * allocation cluster
10422          */
10423         cluster = &fs_info->meta_alloc_cluster;
10424         spin_lock(&cluster->refill_lock);
10425         btrfs_return_cluster_to_free_space(block_group, cluster);
10426         spin_unlock(&cluster->refill_lock);
10427
10428         path = btrfs_alloc_path();
10429         if (!path) {
10430                 ret = -ENOMEM;
10431                 goto out;
10432         }
10433
10434         /*
10435          * get the inode first so any iput calls done for the io_list
10436          * aren't the final iput (no unlinks allowed now)
10437          */
10438         inode = lookup_free_space_inode(fs_info, block_group, path);
10439
10440         mutex_lock(&trans->transaction->cache_write_mutex);
10441         /*
10442          * make sure our free spache cache IO is done before remove the
10443          * free space inode
10444          */
10445         spin_lock(&trans->transaction->dirty_bgs_lock);
10446         if (!list_empty(&block_group->io_list)) {
10447                 list_del_init(&block_group->io_list);
10448
10449                 WARN_ON(!IS_ERR(inode) && inode != block_group->io_ctl.inode);
10450
10451                 spin_unlock(&trans->transaction->dirty_bgs_lock);
10452                 btrfs_wait_cache_io(trans, block_group, path);
10453                 btrfs_put_block_group(block_group);
10454                 spin_lock(&trans->transaction->dirty_bgs_lock);
10455         }
10456
10457         if (!list_empty(&block_group->dirty_list)) {
10458                 list_del_init(&block_group->dirty_list);
10459                 btrfs_put_block_group(block_group);
10460         }
10461         spin_unlock(&trans->transaction->dirty_bgs_lock);
10462         mutex_unlock(&trans->transaction->cache_write_mutex);
10463
10464         if (!IS_ERR(inode)) {
10465                 ret = btrfs_orphan_add(trans, BTRFS_I(inode));
10466                 if (ret) {
10467                         btrfs_add_delayed_iput(inode);
10468                         goto out;
10469                 }
10470                 clear_nlink(inode);
10471                 /* One for the block groups ref */
10472                 spin_lock(&block_group->lock);
10473                 if (block_group->iref) {
10474                         block_group->iref = 0;
10475                         block_group->inode = NULL;
10476                         spin_unlock(&block_group->lock);
10477                         iput(inode);
10478                 } else {
10479                         spin_unlock(&block_group->lock);
10480                 }
10481                 /* One for our lookup ref */
10482                 btrfs_add_delayed_iput(inode);
10483         }
10484
10485         key.objectid = BTRFS_FREE_SPACE_OBJECTID;
10486         key.offset = block_group->key.objectid;
10487         key.type = 0;
10488
10489         ret = btrfs_search_slot(trans, tree_root, &key, path, -1, 1);
10490         if (ret < 0)
10491                 goto out;
10492         if (ret > 0)
10493                 btrfs_release_path(path);
10494         if (ret == 0) {
10495                 ret = btrfs_del_item(trans, tree_root, path);
10496                 if (ret)
10497                         goto out;
10498                 btrfs_release_path(path);
10499         }
10500
10501         spin_lock(&fs_info->block_group_cache_lock);
10502         rb_erase(&block_group->cache_node,
10503                  &fs_info->block_group_cache_tree);
10504         RB_CLEAR_NODE(&block_group->cache_node);
10505
10506         if (fs_info->first_logical_byte == block_group->key.objectid)
10507                 fs_info->first_logical_byte = (u64)-1;
10508         spin_unlock(&fs_info->block_group_cache_lock);
10509
10510         down_write(&block_group->space_info->groups_sem);
10511         /*
10512          * we must use list_del_init so people can check to see if they
10513          * are still on the list after taking the semaphore
10514          */
10515         list_del_init(&block_group->list);
10516         if (list_empty(&block_group->space_info->block_groups[index])) {
10517                 kobj = block_group->space_info->block_group_kobjs[index];
10518                 block_group->space_info->block_group_kobjs[index] = NULL;
10519                 clear_avail_alloc_bits(fs_info, block_group->flags);
10520         }
10521         up_write(&block_group->space_info->groups_sem);
10522         if (kobj) {
10523                 kobject_del(kobj);
10524                 kobject_put(kobj);
10525         }
10526
10527         if (block_group->has_caching_ctl)
10528                 caching_ctl = get_caching_control(block_group);
10529         if (block_group->cached == BTRFS_CACHE_STARTED)
10530                 wait_block_group_cache_done(block_group);
10531         if (block_group->has_caching_ctl) {
10532                 down_write(&fs_info->commit_root_sem);
10533                 if (!caching_ctl) {
10534                         struct btrfs_caching_control *ctl;
10535
10536                         list_for_each_entry(ctl,
10537                                     &fs_info->caching_block_groups, list)
10538                                 if (ctl->block_group == block_group) {
10539                                         caching_ctl = ctl;
10540                                         refcount_inc(&caching_ctl->count);
10541                                         break;
10542                                 }
10543                 }
10544                 if (caching_ctl)
10545                         list_del_init(&caching_ctl->list);
10546                 up_write(&fs_info->commit_root_sem);
10547                 if (caching_ctl) {
10548                         /* Once for the caching bgs list and once for us. */
10549                         put_caching_control(caching_ctl);
10550                         put_caching_control(caching_ctl);
10551                 }
10552         }
10553
10554         spin_lock(&trans->transaction->dirty_bgs_lock);
10555         if (!list_empty(&block_group->dirty_list)) {
10556                 WARN_ON(1);
10557         }
10558         if (!list_empty(&block_group->io_list)) {
10559                 WARN_ON(1);
10560         }
10561         spin_unlock(&trans->transaction->dirty_bgs_lock);
10562         btrfs_remove_free_space_cache(block_group);
10563
10564         spin_lock(&block_group->space_info->lock);
10565         list_del_init(&block_group->ro_list);
10566
10567         if (btrfs_test_opt(fs_info, ENOSPC_DEBUG)) {
10568                 WARN_ON(block_group->space_info->total_bytes
10569                         < block_group->key.offset);
10570                 WARN_ON(block_group->space_info->bytes_readonly
10571                         < block_group->key.offset);
10572                 WARN_ON(block_group->space_info->disk_total
10573                         < block_group->key.offset * factor);
10574         }
10575         block_group->space_info->total_bytes -= block_group->key.offset;
10576         block_group->space_info->bytes_readonly -= block_group->key.offset;
10577         block_group->space_info->disk_total -= block_group->key.offset * factor;
10578
10579         spin_unlock(&block_group->space_info->lock);
10580
10581         memcpy(&key, &block_group->key, sizeof(key));
10582
10583         mutex_lock(&fs_info->chunk_mutex);
10584         if (!list_empty(&em->list)) {
10585                 /* We're in the transaction->pending_chunks list. */
10586                 free_extent_map(em);
10587         }
10588         spin_lock(&block_group->lock);
10589         block_group->removed = 1;
10590         /*
10591          * At this point trimming can't start on this block group, because we
10592          * removed the block group from the tree fs_info->block_group_cache_tree
10593          * so no one can't find it anymore and even if someone already got this
10594          * block group before we removed it from the rbtree, they have already
10595          * incremented block_group->trimming - if they didn't, they won't find
10596          * any free space entries because we already removed them all when we
10597          * called btrfs_remove_free_space_cache().
10598          *
10599          * And we must not remove the extent map from the fs_info->mapping_tree
10600          * to prevent the same logical address range and physical device space
10601          * ranges from being reused for a new block group. This is because our
10602          * fs trim operation (btrfs_trim_fs() / btrfs_ioctl_fitrim()) is
10603          * completely transactionless, so while it is trimming a range the
10604          * currently running transaction might finish and a new one start,
10605          * allowing for new block groups to be created that can reuse the same
10606          * physical device locations unless we take this special care.
10607          *
10608          * There may also be an implicit trim operation if the file system
10609          * is mounted with -odiscard. The same protections must remain
10610          * in place until the extents have been discarded completely when
10611          * the transaction commit has completed.
10612          */
10613         remove_em = (atomic_read(&block_group->trimming) == 0);
10614         /*
10615          * Make sure a trimmer task always sees the em in the pinned_chunks list
10616          * if it sees block_group->removed == 1 (needs to lock block_group->lock
10617          * before checking block_group->removed).
10618          */
10619         if (!remove_em) {
10620                 /*
10621                  * Our em might be in trans->transaction->pending_chunks which
10622                  * is protected by fs_info->chunk_mutex ([lock|unlock]_chunks),
10623                  * and so is the fs_info->pinned_chunks list.
10624                  *
10625                  * So at this point we must be holding the chunk_mutex to avoid
10626                  * any races with chunk allocation (more specifically at
10627                  * volumes.c:contains_pending_extent()), to ensure it always
10628                  * sees the em, either in the pending_chunks list or in the
10629                  * pinned_chunks list.
10630                  */
10631                 list_move_tail(&em->list, &fs_info->pinned_chunks);
10632         }
10633         spin_unlock(&block_group->lock);
10634
10635         if (remove_em) {
10636                 struct extent_map_tree *em_tree;
10637
10638                 em_tree = &fs_info->mapping_tree.map_tree;
10639                 write_lock(&em_tree->lock);
10640                 /*
10641                  * The em might be in the pending_chunks list, so make sure the
10642                  * chunk mutex is locked, since remove_extent_mapping() will
10643                  * delete us from that list.
10644                  */
10645                 remove_extent_mapping(em_tree, em);
10646                 write_unlock(&em_tree->lock);
10647                 /* once for the tree */
10648                 free_extent_map(em);
10649         }
10650
10651         mutex_unlock(&fs_info->chunk_mutex);
10652
10653         ret = remove_block_group_free_space(trans, fs_info, block_group);
10654         if (ret)
10655                 goto out;
10656
10657         btrfs_put_block_group(block_group);
10658         btrfs_put_block_group(block_group);
10659
10660         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
10661         if (ret > 0)
10662                 ret = -EIO;
10663         if (ret < 0)
10664                 goto out;
10665
10666         ret = btrfs_del_item(trans, root, path);
10667 out:
10668         btrfs_free_path(path);
10669         return ret;
10670 }
10671
10672 struct btrfs_trans_handle *
10673 btrfs_start_trans_remove_block_group(struct btrfs_fs_info *fs_info,
10674                                      const u64 chunk_offset)
10675 {
10676         struct extent_map_tree *em_tree = &fs_info->mapping_tree.map_tree;
10677         struct extent_map *em;
10678         struct map_lookup *map;
10679         unsigned int num_items;
10680
10681         read_lock(&em_tree->lock);
10682         em = lookup_extent_mapping(em_tree, chunk_offset, 1);
10683         read_unlock(&em_tree->lock);
10684         ASSERT(em && em->start == chunk_offset);
10685
10686         /*
10687          * We need to reserve 3 + N units from the metadata space info in order
10688          * to remove a block group (done at btrfs_remove_chunk() and at
10689          * btrfs_remove_block_group()), which are used for:
10690          *
10691          * 1 unit for adding the free space inode's orphan (located in the tree
10692          * of tree roots).
10693          * 1 unit for deleting the block group item (located in the extent
10694          * tree).
10695          * 1 unit for deleting the free space item (located in tree of tree
10696          * roots).
10697          * N units for deleting N device extent items corresponding to each
10698          * stripe (located in the device tree).
10699          *
10700          * In order to remove a block group we also need to reserve units in the
10701          * system space info in order to update the chunk tree (update one or
10702          * more device items and remove one chunk item), but this is done at
10703          * btrfs_remove_chunk() through a call to check_system_chunk().
10704          */
10705         map = em->map_lookup;
10706         num_items = 3 + map->num_stripes;
10707         free_extent_map(em);
10708
10709         return btrfs_start_transaction_fallback_global_rsv(fs_info->extent_root,
10710                                                            num_items, 1);
10711 }
10712
10713 /*
10714  * Process the unused_bgs list and remove any that don't have any allocated
10715  * space inside of them.
10716  */
10717 void btrfs_delete_unused_bgs(struct btrfs_fs_info *fs_info)
10718 {
10719         struct btrfs_block_group_cache *block_group;
10720         struct btrfs_space_info *space_info;
10721         struct btrfs_trans_handle *trans;
10722         int ret = 0;
10723
10724         if (!test_bit(BTRFS_FS_OPEN, &fs_info->flags))
10725                 return;
10726
10727         spin_lock(&fs_info->unused_bgs_lock);
10728         while (!list_empty(&fs_info->unused_bgs)) {
10729                 u64 start, end;
10730                 int trimming;
10731
10732                 block_group = list_first_entry(&fs_info->unused_bgs,
10733                                                struct btrfs_block_group_cache,
10734                                                bg_list);
10735                 list_del_init(&block_group->bg_list);
10736
10737                 space_info = block_group->space_info;
10738
10739                 if (ret || btrfs_mixed_space_info(space_info)) {
10740                         btrfs_put_block_group(block_group);
10741                         continue;
10742                 }
10743                 spin_unlock(&fs_info->unused_bgs_lock);
10744
10745                 mutex_lock(&fs_info->delete_unused_bgs_mutex);
10746
10747                 /* Don't want to race with allocators so take the groups_sem */
10748                 down_write(&space_info->groups_sem);
10749                 spin_lock(&block_group->lock);
10750                 if (block_group->reserved ||
10751                     btrfs_block_group_used(&block_group->item) ||
10752                     block_group->ro ||
10753                     list_is_singular(&block_group->list)) {
10754                         /*
10755                          * We want to bail if we made new allocations or have
10756                          * outstanding allocations in this block group.  We do
10757                          * the ro check in case balance is currently acting on
10758                          * this block group.
10759                          */
10760                         spin_unlock(&block_group->lock);
10761                         up_write(&space_info->groups_sem);
10762                         goto next;
10763                 }
10764                 spin_unlock(&block_group->lock);
10765
10766                 /* We don't want to force the issue, only flip if it's ok. */
10767                 ret = inc_block_group_ro(block_group, 0);
10768                 up_write(&space_info->groups_sem);
10769                 if (ret < 0) {
10770                         ret = 0;
10771                         goto next;
10772                 }
10773
10774                 /*
10775                  * Want to do this before we do anything else so we can recover
10776                  * properly if we fail to join the transaction.
10777                  */
10778                 trans = btrfs_start_trans_remove_block_group(fs_info,
10779                                                      block_group->key.objectid);
10780                 if (IS_ERR(trans)) {
10781                         btrfs_dec_block_group_ro(block_group);
10782                         ret = PTR_ERR(trans);
10783                         goto next;
10784                 }
10785
10786                 /*
10787                  * We could have pending pinned extents for this block group,
10788                  * just delete them, we don't care about them anymore.
10789                  */
10790                 start = block_group->key.objectid;
10791                 end = start + block_group->key.offset - 1;
10792                 /*
10793                  * Hold the unused_bg_unpin_mutex lock to avoid racing with
10794                  * btrfs_finish_extent_commit(). If we are at transaction N,
10795                  * another task might be running finish_extent_commit() for the
10796                  * previous transaction N - 1, and have seen a range belonging
10797                  * to the block group in freed_extents[] before we were able to
10798                  * clear the whole block group range from freed_extents[]. This
10799                  * means that task can lookup for the block group after we
10800                  * unpinned it from freed_extents[] and removed it, leading to
10801                  * a BUG_ON() at btrfs_unpin_extent_range().
10802                  */
10803                 mutex_lock(&fs_info->unused_bg_unpin_mutex);
10804                 ret = clear_extent_bits(&fs_info->freed_extents[0], start, end,
10805                                   EXTENT_DIRTY);
10806                 if (ret) {
10807                         mutex_unlock(&fs_info->unused_bg_unpin_mutex);
10808                         btrfs_dec_block_group_ro(block_group);
10809                         goto end_trans;
10810                 }
10811                 ret = clear_extent_bits(&fs_info->freed_extents[1], start, end,
10812                                   EXTENT_DIRTY);
10813                 if (ret) {
10814                         mutex_unlock(&fs_info->unused_bg_unpin_mutex);
10815                         btrfs_dec_block_group_ro(block_group);
10816                         goto end_trans;
10817                 }
10818                 mutex_unlock(&fs_info->unused_bg_unpin_mutex);
10819
10820                 /* Reset pinned so btrfs_put_block_group doesn't complain */
10821                 spin_lock(&space_info->lock);
10822                 spin_lock(&block_group->lock);
10823
10824                 space_info->bytes_pinned -= block_group->pinned;
10825                 space_info->bytes_readonly += block_group->pinned;
10826                 percpu_counter_add(&space_info->total_bytes_pinned,
10827                                    -block_group->pinned);
10828                 block_group->pinned = 0;
10829
10830                 spin_unlock(&block_group->lock);
10831                 spin_unlock(&space_info->lock);
10832
10833                 /* DISCARD can flip during remount */
10834                 trimming = btrfs_test_opt(fs_info, DISCARD);
10835
10836                 /* Implicit trim during transaction commit. */
10837                 if (trimming)
10838                         btrfs_get_block_group_trimming(block_group);
10839
10840                 /*
10841                  * Btrfs_remove_chunk will abort the transaction if things go
10842                  * horribly wrong.
10843                  */
10844                 ret = btrfs_remove_chunk(trans, fs_info,
10845                                          block_group->key.objectid);
10846
10847                 if (ret) {
10848                         if (trimming)
10849                                 btrfs_put_block_group_trimming(block_group);
10850                         goto end_trans;
10851                 }
10852
10853                 /*
10854                  * If we're not mounted with -odiscard, we can just forget
10855                  * about this block group. Otherwise we'll need to wait
10856                  * until transaction commit to do the actual discard.
10857                  */
10858                 if (trimming) {
10859                         spin_lock(&fs_info->unused_bgs_lock);
10860                         /*
10861                          * A concurrent scrub might have added us to the list
10862                          * fs_info->unused_bgs, so use a list_move operation
10863                          * to add the block group to the deleted_bgs list.
10864                          */
10865                         list_move(&block_group->bg_list,
10866                                   &trans->transaction->deleted_bgs);
10867                         spin_unlock(&fs_info->unused_bgs_lock);
10868                         btrfs_get_block_group(block_group);
10869                 }
10870 end_trans:
10871                 btrfs_end_transaction(trans);
10872 next:
10873                 mutex_unlock(&fs_info->delete_unused_bgs_mutex);
10874                 btrfs_put_block_group(block_group);
10875                 spin_lock(&fs_info->unused_bgs_lock);
10876         }
10877         spin_unlock(&fs_info->unused_bgs_lock);
10878 }
10879
10880 int btrfs_init_space_info(struct btrfs_fs_info *fs_info)
10881 {
10882         struct btrfs_space_info *space_info;
10883         struct btrfs_super_block *disk_super;
10884         u64 features;
10885         u64 flags;
10886         int mixed = 0;
10887         int ret;
10888
10889         disk_super = fs_info->super_copy;
10890         if (!btrfs_super_root(disk_super))
10891                 return -EINVAL;
10892
10893         features = btrfs_super_incompat_flags(disk_super);
10894         if (features & BTRFS_FEATURE_INCOMPAT_MIXED_GROUPS)
10895                 mixed = 1;
10896
10897         flags = BTRFS_BLOCK_GROUP_SYSTEM;
10898         ret = create_space_info(fs_info, flags, &space_info);
10899         if (ret)
10900                 goto out;
10901
10902         if (mixed) {
10903                 flags = BTRFS_BLOCK_GROUP_METADATA | BTRFS_BLOCK_GROUP_DATA;
10904                 ret = create_space_info(fs_info, flags, &space_info);
10905         } else {
10906                 flags = BTRFS_BLOCK_GROUP_METADATA;
10907                 ret = create_space_info(fs_info, flags, &space_info);
10908                 if (ret)
10909                         goto out;
10910
10911                 flags = BTRFS_BLOCK_GROUP_DATA;
10912                 ret = create_space_info(fs_info, flags, &space_info);
10913         }
10914 out:
10915         return ret;
10916 }
10917
10918 int btrfs_error_unpin_extent_range(struct btrfs_fs_info *fs_info,
10919                                    u64 start, u64 end)
10920 {
10921         return unpin_extent_range(fs_info, start, end, false);
10922 }
10923
10924 /*
10925  * It used to be that old block groups would be left around forever.
10926  * Iterating over them would be enough to trim unused space.  Since we
10927  * now automatically remove them, we also need to iterate over unallocated
10928  * space.
10929  *
10930  * We don't want a transaction for this since the discard may take a
10931  * substantial amount of time.  We don't require that a transaction be
10932  * running, but we do need to take a running transaction into account
10933  * to ensure that we're not discarding chunks that were released in
10934  * the current transaction.
10935  *
10936  * Holding the chunks lock will prevent other threads from allocating
10937  * or releasing chunks, but it won't prevent a running transaction
10938  * from committing and releasing the memory that the pending chunks
10939  * list head uses.  For that, we need to take a reference to the
10940  * transaction.
10941  */
10942 static int btrfs_trim_free_extents(struct btrfs_device *device,
10943                                    u64 minlen, u64 *trimmed)
10944 {
10945         u64 start = 0, len = 0;
10946         int ret;
10947
10948         *trimmed = 0;
10949
10950         /* Not writeable = nothing to do. */
10951         if (!test_bit(BTRFS_DEV_STATE_WRITEABLE, &device->dev_state))
10952                 return 0;
10953
10954         /* No free space = nothing to do. */
10955         if (device->total_bytes <= device->bytes_used)
10956                 return 0;
10957
10958         ret = 0;
10959
10960         while (1) {
10961                 struct btrfs_fs_info *fs_info = device->fs_info;
10962                 struct btrfs_transaction *trans;
10963                 u64 bytes;
10964
10965                 ret = mutex_lock_interruptible(&fs_info->chunk_mutex);
10966                 if (ret)
10967                         return ret;
10968
10969                 down_read(&fs_info->commit_root_sem);
10970
10971                 spin_lock(&fs_info->trans_lock);
10972                 trans = fs_info->running_transaction;
10973                 if (trans)
10974                         refcount_inc(&trans->use_count);
10975                 spin_unlock(&fs_info->trans_lock);
10976
10977                 ret = find_free_dev_extent_start(trans, device, minlen, start,
10978                                                  &start, &len);
10979                 if (trans)
10980                         btrfs_put_transaction(trans);
10981
10982                 if (ret) {
10983                         up_read(&fs_info->commit_root_sem);
10984                         mutex_unlock(&fs_info->chunk_mutex);
10985                         if (ret == -ENOSPC)
10986                                 ret = 0;
10987                         break;
10988                 }
10989
10990                 ret = btrfs_issue_discard(device->bdev, start, len, &bytes);
10991                 up_read(&fs_info->commit_root_sem);
10992                 mutex_unlock(&fs_info->chunk_mutex);
10993
10994                 if (ret)
10995                         break;
10996
10997                 start += len;
10998                 *trimmed += bytes;
10999
11000                 if (fatal_signal_pending(current)) {
11001                         ret = -ERESTARTSYS;
11002                         break;
11003                 }
11004
11005                 cond_resched();
11006         }
11007
11008         return ret;
11009 }
11010
11011 int btrfs_trim_fs(struct btrfs_fs_info *fs_info, struct fstrim_range *range)
11012 {
11013         struct btrfs_block_group_cache *cache = NULL;
11014         struct btrfs_device *device;
11015         struct list_head *devices;
11016         u64 group_trimmed;
11017         u64 start;
11018         u64 end;
11019         u64 trimmed = 0;
11020         u64 total_bytes = btrfs_super_total_bytes(fs_info->super_copy);
11021         int ret = 0;
11022
11023         /*
11024          * try to trim all FS space, our block group may start from non-zero.
11025          */
11026         if (range->len == total_bytes)
11027                 cache = btrfs_lookup_first_block_group(fs_info, range->start);
11028         else
11029                 cache = btrfs_lookup_block_group(fs_info, range->start);
11030
11031         while (cache) {
11032                 if (cache->key.objectid >= (range->start + range->len)) {
11033                         btrfs_put_block_group(cache);
11034                         break;
11035                 }
11036
11037                 start = max(range->start, cache->key.objectid);
11038                 end = min(range->start + range->len,
11039                                 cache->key.objectid + cache->key.offset);
11040
11041                 if (end - start >= range->minlen) {
11042                         if (!block_group_cache_done(cache)) {
11043                                 ret = cache_block_group(cache, 0);
11044                                 if (ret) {
11045                                         btrfs_put_block_group(cache);
11046                                         break;
11047                                 }
11048                                 ret = wait_block_group_cache_done(cache);
11049                                 if (ret) {
11050                                         btrfs_put_block_group(cache);
11051                                         break;
11052                                 }
11053                         }
11054                         ret = btrfs_trim_block_group(cache,
11055                                                      &group_trimmed,
11056                                                      start,
11057                                                      end,
11058                                                      range->minlen);
11059
11060                         trimmed += group_trimmed;
11061                         if (ret) {
11062                                 btrfs_put_block_group(cache);
11063                                 break;
11064                         }
11065                 }
11066
11067                 cache = next_block_group(fs_info, cache);
11068         }
11069
11070         mutex_lock(&fs_info->fs_devices->device_list_mutex);
11071         devices = &fs_info->fs_devices->alloc_list;
11072         list_for_each_entry(device, devices, dev_alloc_list) {
11073                 ret = btrfs_trim_free_extents(device, range->minlen,
11074                                               &group_trimmed);
11075                 if (ret)
11076                         break;
11077
11078                 trimmed += group_trimmed;
11079         }
11080         mutex_unlock(&fs_info->fs_devices->device_list_mutex);
11081
11082         range->len = trimmed;
11083         return ret;
11084 }
11085
11086 /*
11087  * btrfs_{start,end}_write_no_snapshotting() are similar to
11088  * mnt_{want,drop}_write(), they are used to prevent some tasks from writing
11089  * data into the page cache through nocow before the subvolume is snapshoted,
11090  * but flush the data into disk after the snapshot creation, or to prevent
11091  * operations while snapshotting is ongoing and that cause the snapshot to be
11092  * inconsistent (writes followed by expanding truncates for example).
11093  */
11094 void btrfs_end_write_no_snapshotting(struct btrfs_root *root)
11095 {
11096         percpu_counter_dec(&root->subv_writers->counter);
11097         /*
11098          * Make sure counter is updated before we wake up waiters.
11099          */
11100         smp_mb();
11101         if (waitqueue_active(&root->subv_writers->wait))
11102                 wake_up(&root->subv_writers->wait);
11103 }
11104
11105 int btrfs_start_write_no_snapshotting(struct btrfs_root *root)
11106 {
11107         if (atomic_read(&root->will_be_snapshotted))
11108                 return 0;
11109
11110         percpu_counter_inc(&root->subv_writers->counter);
11111         /*
11112          * Make sure counter is updated before we check for snapshot creation.
11113          */
11114         smp_mb();
11115         if (atomic_read(&root->will_be_snapshotted)) {
11116                 btrfs_end_write_no_snapshotting(root);
11117                 return 0;
11118         }
11119         return 1;
11120 }
11121
11122 void btrfs_wait_for_snapshot_creation(struct btrfs_root *root)
11123 {
11124         while (true) {
11125                 int ret;
11126
11127                 ret = btrfs_start_write_no_snapshotting(root);
11128                 if (ret)
11129                         break;
11130                 wait_on_atomic_t(&root->will_be_snapshotted, atomic_t_wait,
11131                                  TASK_UNINTERRUPTIBLE);
11132         }
11133 }