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_var(&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_var_event(&bg->nocow_writers, !atomic_read(&bg->nocow_writers));
4010 }
4011
4012 static const char *alloc_name(u64 flags)
4013 {
4014         switch (flags) {
4015         case BTRFS_BLOCK_GROUP_METADATA|BTRFS_BLOCK_GROUP_DATA:
4016                 return "mixed";
4017         case BTRFS_BLOCK_GROUP_METADATA:
4018                 return "metadata";
4019         case BTRFS_BLOCK_GROUP_DATA:
4020                 return "data";
4021         case BTRFS_BLOCK_GROUP_SYSTEM:
4022                 return "system";
4023         default:
4024                 WARN_ON(1);
4025                 return "invalid-combination";
4026         };
4027 }
4028
4029 static int create_space_info(struct btrfs_fs_info *info, u64 flags,
4030                              struct btrfs_space_info **new)
4031 {
4032
4033         struct btrfs_space_info *space_info;
4034         int i;
4035         int ret;
4036
4037         space_info = kzalloc(sizeof(*space_info), GFP_NOFS);
4038         if (!space_info)
4039                 return -ENOMEM;
4040
4041         ret = percpu_counter_init(&space_info->total_bytes_pinned, 0,
4042                                  GFP_KERNEL);
4043         if (ret) {
4044                 kfree(space_info);
4045                 return ret;
4046         }
4047
4048         for (i = 0; i < BTRFS_NR_RAID_TYPES; i++)
4049                 INIT_LIST_HEAD(&space_info->block_groups[i]);
4050         init_rwsem(&space_info->groups_sem);
4051         spin_lock_init(&space_info->lock);
4052         space_info->flags = flags & BTRFS_BLOCK_GROUP_TYPE_MASK;
4053         space_info->force_alloc = CHUNK_ALLOC_NO_FORCE;
4054         init_waitqueue_head(&space_info->wait);
4055         INIT_LIST_HEAD(&space_info->ro_bgs);
4056         INIT_LIST_HEAD(&space_info->tickets);
4057         INIT_LIST_HEAD(&space_info->priority_tickets);
4058
4059         ret = kobject_init_and_add(&space_info->kobj, &space_info_ktype,
4060                                     info->space_info_kobj, "%s",
4061                                     alloc_name(space_info->flags));
4062         if (ret) {
4063                 percpu_counter_destroy(&space_info->total_bytes_pinned);
4064                 kfree(space_info);
4065                 return ret;
4066         }
4067
4068         *new = space_info;
4069         list_add_rcu(&space_info->list, &info->space_info);
4070         if (flags & BTRFS_BLOCK_GROUP_DATA)
4071                 info->data_sinfo = space_info;
4072
4073         return ret;
4074 }
4075
4076 static void update_space_info(struct btrfs_fs_info *info, u64 flags,
4077                              u64 total_bytes, u64 bytes_used,
4078                              u64 bytes_readonly,
4079                              struct btrfs_space_info **space_info)
4080 {
4081         struct btrfs_space_info *found;
4082         int factor;
4083
4084         if (flags & (BTRFS_BLOCK_GROUP_DUP | BTRFS_BLOCK_GROUP_RAID1 |
4085                      BTRFS_BLOCK_GROUP_RAID10))
4086                 factor = 2;
4087         else
4088                 factor = 1;
4089
4090         found = __find_space_info(info, flags);
4091         ASSERT(found);
4092         spin_lock(&found->lock);
4093         found->total_bytes += total_bytes;
4094         found->disk_total += total_bytes * factor;
4095         found->bytes_used += bytes_used;
4096         found->disk_used += bytes_used * factor;
4097         found->bytes_readonly += bytes_readonly;
4098         if (total_bytes > 0)
4099                 found->full = 0;
4100         space_info_add_new_bytes(info, found, total_bytes -
4101                                  bytes_used - bytes_readonly);
4102         spin_unlock(&found->lock);
4103         *space_info = found;
4104 }
4105
4106 static void set_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
4107 {
4108         u64 extra_flags = chunk_to_extended(flags) &
4109                                 BTRFS_EXTENDED_PROFILE_MASK;
4110
4111         write_seqlock(&fs_info->profiles_lock);
4112         if (flags & BTRFS_BLOCK_GROUP_DATA)
4113                 fs_info->avail_data_alloc_bits |= extra_flags;
4114         if (flags & BTRFS_BLOCK_GROUP_METADATA)
4115                 fs_info->avail_metadata_alloc_bits |= extra_flags;
4116         if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
4117                 fs_info->avail_system_alloc_bits |= extra_flags;
4118         write_sequnlock(&fs_info->profiles_lock);
4119 }
4120
4121 /*
4122  * returns target flags in extended format or 0 if restripe for this
4123  * chunk_type is not in progress
4124  *
4125  * should be called with either volume_mutex or balance_lock held
4126  */
4127 static u64 get_restripe_target(struct btrfs_fs_info *fs_info, u64 flags)
4128 {
4129         struct btrfs_balance_control *bctl = fs_info->balance_ctl;
4130         u64 target = 0;
4131
4132         if (!bctl)
4133                 return 0;
4134
4135         if (flags & BTRFS_BLOCK_GROUP_DATA &&
4136             bctl->data.flags & BTRFS_BALANCE_ARGS_CONVERT) {
4137                 target = BTRFS_BLOCK_GROUP_DATA | bctl->data.target;
4138         } else if (flags & BTRFS_BLOCK_GROUP_SYSTEM &&
4139                    bctl->sys.flags & BTRFS_BALANCE_ARGS_CONVERT) {
4140                 target = BTRFS_BLOCK_GROUP_SYSTEM | bctl->sys.target;
4141         } else if (flags & BTRFS_BLOCK_GROUP_METADATA &&
4142                    bctl->meta.flags & BTRFS_BALANCE_ARGS_CONVERT) {
4143                 target = BTRFS_BLOCK_GROUP_METADATA | bctl->meta.target;
4144         }
4145
4146         return target;
4147 }
4148
4149 /*
4150  * @flags: available profiles in extended format (see ctree.h)
4151  *
4152  * Returns reduced profile in chunk format.  If profile changing is in
4153  * progress (either running or paused) picks the target profile (if it's
4154  * already available), otherwise falls back to plain reducing.
4155  */
4156 static u64 btrfs_reduce_alloc_profile(struct btrfs_fs_info *fs_info, u64 flags)
4157 {
4158         u64 num_devices = fs_info->fs_devices->rw_devices;
4159         u64 target;
4160         u64 raid_type;
4161         u64 allowed = 0;
4162
4163         /*
4164          * see if restripe for this chunk_type is in progress, if so
4165          * try to reduce to the target profile
4166          */
4167         spin_lock(&fs_info->balance_lock);
4168         target = get_restripe_target(fs_info, flags);
4169         if (target) {
4170                 /* pick target profile only if it's already available */
4171                 if ((flags & target) & BTRFS_EXTENDED_PROFILE_MASK) {
4172                         spin_unlock(&fs_info->balance_lock);
4173                         return extended_to_chunk(target);
4174                 }
4175         }
4176         spin_unlock(&fs_info->balance_lock);
4177
4178         /* First, mask out the RAID levels which aren't possible */
4179         for (raid_type = 0; raid_type < BTRFS_NR_RAID_TYPES; raid_type++) {
4180                 if (num_devices >= btrfs_raid_array[raid_type].devs_min)
4181                         allowed |= btrfs_raid_group[raid_type];
4182         }
4183         allowed &= flags;
4184
4185         if (allowed & BTRFS_BLOCK_GROUP_RAID6)
4186                 allowed = BTRFS_BLOCK_GROUP_RAID6;
4187         else if (allowed & BTRFS_BLOCK_GROUP_RAID5)
4188                 allowed = BTRFS_BLOCK_GROUP_RAID5;
4189         else if (allowed & BTRFS_BLOCK_GROUP_RAID10)
4190                 allowed = BTRFS_BLOCK_GROUP_RAID10;
4191         else if (allowed & BTRFS_BLOCK_GROUP_RAID1)
4192                 allowed = BTRFS_BLOCK_GROUP_RAID1;
4193         else if (allowed & BTRFS_BLOCK_GROUP_RAID0)
4194                 allowed = BTRFS_BLOCK_GROUP_RAID0;
4195
4196         flags &= ~BTRFS_BLOCK_GROUP_PROFILE_MASK;
4197
4198         return extended_to_chunk(flags | allowed);
4199 }
4200
4201 static u64 get_alloc_profile(struct btrfs_fs_info *fs_info, u64 orig_flags)
4202 {
4203         unsigned seq;
4204         u64 flags;
4205
4206         do {
4207                 flags = orig_flags;
4208                 seq = read_seqbegin(&fs_info->profiles_lock);
4209
4210                 if (flags & BTRFS_BLOCK_GROUP_DATA)
4211                         flags |= fs_info->avail_data_alloc_bits;
4212                 else if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
4213                         flags |= fs_info->avail_system_alloc_bits;
4214                 else if (flags & BTRFS_BLOCK_GROUP_METADATA)
4215                         flags |= fs_info->avail_metadata_alloc_bits;
4216         } while (read_seqretry(&fs_info->profiles_lock, seq));
4217
4218         return btrfs_reduce_alloc_profile(fs_info, flags);
4219 }
4220
4221 static u64 get_alloc_profile_by_root(struct btrfs_root *root, int data)
4222 {
4223         struct btrfs_fs_info *fs_info = root->fs_info;
4224         u64 flags;
4225         u64 ret;
4226
4227         if (data)
4228                 flags = BTRFS_BLOCK_GROUP_DATA;
4229         else if (root == fs_info->chunk_root)
4230                 flags = BTRFS_BLOCK_GROUP_SYSTEM;
4231         else
4232                 flags = BTRFS_BLOCK_GROUP_METADATA;
4233
4234         ret = get_alloc_profile(fs_info, flags);
4235         return ret;
4236 }
4237
4238 u64 btrfs_data_alloc_profile(struct btrfs_fs_info *fs_info)
4239 {
4240         return get_alloc_profile(fs_info, BTRFS_BLOCK_GROUP_DATA);
4241 }
4242
4243 u64 btrfs_metadata_alloc_profile(struct btrfs_fs_info *fs_info)
4244 {
4245         return get_alloc_profile(fs_info, BTRFS_BLOCK_GROUP_METADATA);
4246 }
4247
4248 u64 btrfs_system_alloc_profile(struct btrfs_fs_info *fs_info)
4249 {
4250         return get_alloc_profile(fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
4251 }
4252
4253 static u64 btrfs_space_info_used(struct btrfs_space_info *s_info,
4254                                  bool may_use_included)
4255 {
4256         ASSERT(s_info);
4257         return s_info->bytes_used + s_info->bytes_reserved +
4258                 s_info->bytes_pinned + s_info->bytes_readonly +
4259                 (may_use_included ? s_info->bytes_may_use : 0);
4260 }
4261
4262 int btrfs_alloc_data_chunk_ondemand(struct btrfs_inode *inode, u64 bytes)
4263 {
4264         struct btrfs_root *root = inode->root;
4265         struct btrfs_fs_info *fs_info = root->fs_info;
4266         struct btrfs_space_info *data_sinfo = fs_info->data_sinfo;
4267         u64 used;
4268         int ret = 0;
4269         int need_commit = 2;
4270         int have_pinned_space;
4271
4272         /* make sure bytes are sectorsize aligned */
4273         bytes = ALIGN(bytes, fs_info->sectorsize);
4274
4275         if (btrfs_is_free_space_inode(inode)) {
4276                 need_commit = 0;
4277                 ASSERT(current->journal_info);
4278         }
4279
4280 again:
4281         /* make sure we have enough space to handle the data first */
4282         spin_lock(&data_sinfo->lock);
4283         used = btrfs_space_info_used(data_sinfo, true);
4284
4285         if (used + bytes > data_sinfo->total_bytes) {
4286                 struct btrfs_trans_handle *trans;
4287
4288                 /*
4289                  * if we don't have enough free bytes in this space then we need
4290                  * to alloc a new chunk.
4291                  */
4292                 if (!data_sinfo->full) {
4293                         u64 alloc_target;
4294
4295                         data_sinfo->force_alloc = CHUNK_ALLOC_FORCE;
4296                         spin_unlock(&data_sinfo->lock);
4297
4298                         alloc_target = btrfs_data_alloc_profile(fs_info);
4299                         /*
4300                          * It is ugly that we don't call nolock join
4301                          * transaction for the free space inode case here.
4302                          * But it is safe because we only do the data space
4303                          * reservation for the free space cache in the
4304                          * transaction context, the common join transaction
4305                          * just increase the counter of the current transaction
4306                          * handler, doesn't try to acquire the trans_lock of
4307                          * the fs.
4308                          */
4309                         trans = btrfs_join_transaction(root);
4310                         if (IS_ERR(trans))
4311                                 return PTR_ERR(trans);
4312
4313                         ret = do_chunk_alloc(trans, fs_info, alloc_target,
4314                                              CHUNK_ALLOC_NO_FORCE);
4315                         btrfs_end_transaction(trans);
4316                         if (ret < 0) {
4317                                 if (ret != -ENOSPC)
4318                                         return ret;
4319                                 else {
4320                                         have_pinned_space = 1;
4321                                         goto commit_trans;
4322                                 }
4323                         }
4324
4325                         goto again;
4326                 }
4327
4328                 /*
4329                  * If we don't have enough pinned space to deal with this
4330                  * allocation, and no removed chunk in current transaction,
4331                  * don't bother committing the transaction.
4332                  */
4333                 have_pinned_space = percpu_counter_compare(
4334                         &data_sinfo->total_bytes_pinned,
4335                         used + bytes - data_sinfo->total_bytes);
4336                 spin_unlock(&data_sinfo->lock);
4337
4338                 /* commit the current transaction and try again */
4339 commit_trans:
4340                 if (need_commit) {
4341                         need_commit--;
4342
4343                         if (need_commit > 0) {
4344                                 btrfs_start_delalloc_roots(fs_info, 0, -1);
4345                                 btrfs_wait_ordered_roots(fs_info, U64_MAX, 0,
4346                                                          (u64)-1);
4347                         }
4348
4349                         trans = btrfs_join_transaction(root);
4350                         if (IS_ERR(trans))
4351                                 return PTR_ERR(trans);
4352                         if (have_pinned_space >= 0 ||
4353                             test_bit(BTRFS_TRANS_HAVE_FREE_BGS,
4354                                      &trans->transaction->flags) ||
4355                             need_commit > 0) {
4356                                 ret = btrfs_commit_transaction(trans);
4357                                 if (ret)
4358                                         return ret;
4359                                 /*
4360                                  * The cleaner kthread might still be doing iput
4361                                  * operations. Wait for it to finish so that
4362                                  * more space is released.
4363                                  */
4364                                 mutex_lock(&fs_info->cleaner_delayed_iput_mutex);
4365                                 mutex_unlock(&fs_info->cleaner_delayed_iput_mutex);
4366                                 goto again;
4367                         } else {
4368                                 btrfs_end_transaction(trans);
4369                         }
4370                 }
4371
4372                 trace_btrfs_space_reservation(fs_info,
4373                                               "space_info:enospc",
4374                                               data_sinfo->flags, bytes, 1);
4375                 return -ENOSPC;
4376         }
4377         data_sinfo->bytes_may_use += bytes;
4378         trace_btrfs_space_reservation(fs_info, "space_info",
4379                                       data_sinfo->flags, bytes, 1);
4380         spin_unlock(&data_sinfo->lock);
4381
4382         return ret;
4383 }
4384
4385 int btrfs_check_data_free_space(struct inode *inode,
4386                         struct extent_changeset **reserved, u64 start, u64 len)
4387 {
4388         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
4389         int ret;
4390
4391         /* align the range */
4392         len = round_up(start + len, fs_info->sectorsize) -
4393               round_down(start, fs_info->sectorsize);
4394         start = round_down(start, fs_info->sectorsize);
4395
4396         ret = btrfs_alloc_data_chunk_ondemand(BTRFS_I(inode), len);
4397         if (ret < 0)
4398                 return ret;
4399
4400         /* Use new btrfs_qgroup_reserve_data to reserve precious data space. */
4401         ret = btrfs_qgroup_reserve_data(inode, reserved, start, len);
4402         if (ret < 0)
4403                 btrfs_free_reserved_data_space_noquota(inode, start, len);
4404         else
4405                 ret = 0;
4406         return ret;
4407 }
4408
4409 /*
4410  * Called if we need to clear a data reservation for this inode
4411  * Normally in a error case.
4412  *
4413  * This one will *NOT* use accurate qgroup reserved space API, just for case
4414  * which we can't sleep and is sure it won't affect qgroup reserved space.
4415  * Like clear_bit_hook().
4416  */
4417 void btrfs_free_reserved_data_space_noquota(struct inode *inode, u64 start,
4418                                             u64 len)
4419 {
4420         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
4421         struct btrfs_space_info *data_sinfo;
4422
4423         /* Make sure the range is aligned to sectorsize */
4424         len = round_up(start + len, fs_info->sectorsize) -
4425               round_down(start, fs_info->sectorsize);
4426         start = round_down(start, fs_info->sectorsize);
4427
4428         data_sinfo = fs_info->data_sinfo;
4429         spin_lock(&data_sinfo->lock);
4430         if (WARN_ON(data_sinfo->bytes_may_use < len))
4431                 data_sinfo->bytes_may_use = 0;
4432         else
4433                 data_sinfo->bytes_may_use -= len;
4434         trace_btrfs_space_reservation(fs_info, "space_info",
4435                                       data_sinfo->flags, len, 0);
4436         spin_unlock(&data_sinfo->lock);
4437 }
4438
4439 /*
4440  * Called if we need to clear a data reservation for this inode
4441  * Normally in a error case.
4442  *
4443  * This one will handle the per-inode data rsv map for accurate reserved
4444  * space framework.
4445  */
4446 void btrfs_free_reserved_data_space(struct inode *inode,
4447                         struct extent_changeset *reserved, u64 start, u64 len)
4448 {
4449         struct btrfs_root *root = BTRFS_I(inode)->root;
4450
4451         /* Make sure the range is aligned to sectorsize */
4452         len = round_up(start + len, root->fs_info->sectorsize) -
4453               round_down(start, root->fs_info->sectorsize);
4454         start = round_down(start, root->fs_info->sectorsize);
4455
4456         btrfs_free_reserved_data_space_noquota(inode, start, len);
4457         btrfs_qgroup_free_data(inode, reserved, start, len);
4458 }
4459
4460 static void force_metadata_allocation(struct btrfs_fs_info *info)
4461 {
4462         struct list_head *head = &info->space_info;
4463         struct btrfs_space_info *found;
4464
4465         rcu_read_lock();
4466         list_for_each_entry_rcu(found, head, list) {
4467                 if (found->flags & BTRFS_BLOCK_GROUP_METADATA)
4468                         found->force_alloc = CHUNK_ALLOC_FORCE;
4469         }
4470         rcu_read_unlock();
4471 }
4472
4473 static inline u64 calc_global_rsv_need_space(struct btrfs_block_rsv *global)
4474 {
4475         return (global->size << 1);
4476 }
4477
4478 static int should_alloc_chunk(struct btrfs_fs_info *fs_info,
4479                               struct btrfs_space_info *sinfo, int force)
4480 {
4481         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
4482         u64 bytes_used = btrfs_space_info_used(sinfo, false);
4483         u64 thresh;
4484
4485         if (force == CHUNK_ALLOC_FORCE)
4486                 return 1;
4487
4488         /*
4489          * We need to take into account the global rsv because for all intents
4490          * and purposes it's used space.  Don't worry about locking the
4491          * global_rsv, it doesn't change except when the transaction commits.
4492          */
4493         if (sinfo->flags & BTRFS_BLOCK_GROUP_METADATA)
4494                 bytes_used += calc_global_rsv_need_space(global_rsv);
4495
4496         /*
4497          * in limited mode, we want to have some free space up to
4498          * about 1% of the FS size.
4499          */
4500         if (force == CHUNK_ALLOC_LIMITED) {
4501                 thresh = btrfs_super_total_bytes(fs_info->super_copy);
4502                 thresh = max_t(u64, SZ_64M, div_factor_fine(thresh, 1));
4503
4504                 if (sinfo->total_bytes - bytes_used < thresh)
4505                         return 1;
4506         }
4507
4508         if (bytes_used + SZ_2M < div_factor(sinfo->total_bytes, 8))
4509                 return 0;
4510         return 1;
4511 }
4512
4513 static u64 get_profile_num_devs(struct btrfs_fs_info *fs_info, u64 type)
4514 {
4515         u64 num_dev;
4516
4517         if (type & (BTRFS_BLOCK_GROUP_RAID10 |
4518                     BTRFS_BLOCK_GROUP_RAID0 |
4519                     BTRFS_BLOCK_GROUP_RAID5 |
4520                     BTRFS_BLOCK_GROUP_RAID6))
4521                 num_dev = fs_info->fs_devices->rw_devices;
4522         else if (type & BTRFS_BLOCK_GROUP_RAID1)
4523                 num_dev = 2;
4524         else
4525                 num_dev = 1;    /* DUP or single */
4526
4527         return num_dev;
4528 }
4529
4530 /*
4531  * If @is_allocation is true, reserve space in the system space info necessary
4532  * for allocating a chunk, otherwise if it's false, reserve space necessary for
4533  * removing a chunk.
4534  */
4535 void check_system_chunk(struct btrfs_trans_handle *trans,
4536                         struct btrfs_fs_info *fs_info, u64 type)
4537 {
4538         struct btrfs_space_info *info;
4539         u64 left;
4540         u64 thresh;
4541         int ret = 0;
4542         u64 num_devs;
4543
4544         /*
4545          * Needed because we can end up allocating a system chunk and for an
4546          * atomic and race free space reservation in the chunk block reserve.
4547          */
4548         lockdep_assert_held(&fs_info->chunk_mutex);
4549
4550         info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
4551         spin_lock(&info->lock);
4552         left = info->total_bytes - btrfs_space_info_used(info, true);
4553         spin_unlock(&info->lock);
4554
4555         num_devs = get_profile_num_devs(fs_info, type);
4556
4557         /* num_devs device items to update and 1 chunk item to add or remove */
4558         thresh = btrfs_calc_trunc_metadata_size(fs_info, num_devs) +
4559                 btrfs_calc_trans_metadata_size(fs_info, 1);
4560
4561         if (left < thresh && btrfs_test_opt(fs_info, ENOSPC_DEBUG)) {
4562                 btrfs_info(fs_info, "left=%llu, need=%llu, flags=%llu",
4563                            left, thresh, type);
4564                 dump_space_info(fs_info, info, 0, 0);
4565         }
4566
4567         if (left < thresh) {
4568                 u64 flags = btrfs_system_alloc_profile(fs_info);
4569
4570                 /*
4571                  * Ignore failure to create system chunk. We might end up not
4572                  * needing it, as we might not need to COW all nodes/leafs from
4573                  * the paths we visit in the chunk tree (they were already COWed
4574                  * or created in the current transaction for example).
4575                  */
4576                 ret = btrfs_alloc_chunk(trans, fs_info, flags);
4577         }
4578
4579         if (!ret) {
4580                 ret = btrfs_block_rsv_add(fs_info->chunk_root,
4581                                           &fs_info->chunk_block_rsv,
4582                                           thresh, BTRFS_RESERVE_NO_FLUSH);
4583                 if (!ret)
4584                         trans->chunk_bytes_reserved += thresh;
4585         }
4586 }
4587
4588 /*
4589  * If force is CHUNK_ALLOC_FORCE:
4590  *    - return 1 if it successfully allocates a chunk,
4591  *    - return errors including -ENOSPC otherwise.
4592  * If force is NOT CHUNK_ALLOC_FORCE:
4593  *    - return 0 if it doesn't need to allocate a new chunk,
4594  *    - return 1 if it successfully allocates a chunk,
4595  *    - return errors including -ENOSPC otherwise.
4596  */
4597 static int do_chunk_alloc(struct btrfs_trans_handle *trans,
4598                           struct btrfs_fs_info *fs_info, u64 flags, int force)
4599 {
4600         struct btrfs_space_info *space_info;
4601         int wait_for_alloc = 0;
4602         int ret = 0;
4603
4604         /* Don't re-enter if we're already allocating a chunk */
4605         if (trans->allocating_chunk)
4606                 return -ENOSPC;
4607
4608         space_info = __find_space_info(fs_info, flags);
4609         ASSERT(space_info);
4610
4611 again:
4612         spin_lock(&space_info->lock);
4613         if (force < space_info->force_alloc)
4614                 force = space_info->force_alloc;
4615         if (space_info->full) {
4616                 if (should_alloc_chunk(fs_info, space_info, force))
4617                         ret = -ENOSPC;
4618                 else
4619                         ret = 0;
4620                 spin_unlock(&space_info->lock);
4621                 return ret;
4622         }
4623
4624         if (!should_alloc_chunk(fs_info, space_info, force)) {
4625                 spin_unlock(&space_info->lock);
4626                 return 0;
4627         } else if (space_info->chunk_alloc) {
4628                 wait_for_alloc = 1;
4629         } else {
4630                 space_info->chunk_alloc = 1;
4631         }
4632
4633         spin_unlock(&space_info->lock);
4634
4635         mutex_lock(&fs_info->chunk_mutex);
4636
4637         /*
4638          * The chunk_mutex is held throughout the entirety of a chunk
4639          * allocation, so once we've acquired the chunk_mutex we know that the
4640          * other guy is done and we need to recheck and see if we should
4641          * allocate.
4642          */
4643         if (wait_for_alloc) {
4644                 mutex_unlock(&fs_info->chunk_mutex);
4645                 wait_for_alloc = 0;
4646                 cond_resched();
4647                 goto again;
4648         }
4649
4650         trans->allocating_chunk = true;
4651
4652         /*
4653          * If we have mixed data/metadata chunks we want to make sure we keep
4654          * allocating mixed chunks instead of individual chunks.
4655          */
4656         if (btrfs_mixed_space_info(space_info))
4657                 flags |= (BTRFS_BLOCK_GROUP_DATA | BTRFS_BLOCK_GROUP_METADATA);
4658
4659         /*
4660          * if we're doing a data chunk, go ahead and make sure that
4661          * we keep a reasonable number of metadata chunks allocated in the
4662          * FS as well.
4663          */
4664         if (flags & BTRFS_BLOCK_GROUP_DATA && fs_info->metadata_ratio) {
4665                 fs_info->data_chunk_allocations++;
4666                 if (!(fs_info->data_chunk_allocations %
4667                       fs_info->metadata_ratio))
4668                         force_metadata_allocation(fs_info);
4669         }
4670
4671         /*
4672          * Check if we have enough space in SYSTEM chunk because we may need
4673          * to update devices.
4674          */
4675         check_system_chunk(trans, fs_info, flags);
4676
4677         ret = btrfs_alloc_chunk(trans, fs_info, flags);
4678         trans->allocating_chunk = false;
4679
4680         spin_lock(&space_info->lock);
4681         if (ret < 0 && ret != -ENOSPC)
4682                 goto out;
4683         if (ret)
4684                 space_info->full = 1;
4685         else
4686                 ret = 1;
4687
4688         space_info->force_alloc = CHUNK_ALLOC_NO_FORCE;
4689 out:
4690         space_info->chunk_alloc = 0;
4691         spin_unlock(&space_info->lock);
4692         mutex_unlock(&fs_info->chunk_mutex);
4693         /*
4694          * When we allocate a new chunk we reserve space in the chunk block
4695          * reserve to make sure we can COW nodes/leafs in the chunk tree or
4696          * add new nodes/leafs to it if we end up needing to do it when
4697          * inserting the chunk item and updating device items as part of the
4698          * second phase of chunk allocation, performed by
4699          * btrfs_finish_chunk_alloc(). So make sure we don't accumulate a
4700          * large number of new block groups to create in our transaction
4701          * handle's new_bgs list to avoid exhausting the chunk block reserve
4702          * in extreme cases - like having a single transaction create many new
4703          * block groups when starting to write out the free space caches of all
4704          * the block groups that were made dirty during the lifetime of the
4705          * transaction.
4706          */
4707         if (trans->can_flush_pending_bgs &&
4708             trans->chunk_bytes_reserved >= (u64)SZ_2M) {
4709                 btrfs_create_pending_block_groups(trans);
4710                 btrfs_trans_release_chunk_metadata(trans);
4711         }
4712         return ret;
4713 }
4714
4715 static int can_overcommit(struct btrfs_fs_info *fs_info,
4716                           struct btrfs_space_info *space_info, u64 bytes,
4717                           enum btrfs_reserve_flush_enum flush,
4718                           bool system_chunk)
4719 {
4720         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
4721         u64 profile;
4722         u64 space_size;
4723         u64 avail;
4724         u64 used;
4725
4726         /* Don't overcommit when in mixed mode. */
4727         if (space_info->flags & BTRFS_BLOCK_GROUP_DATA)
4728                 return 0;
4729
4730         if (system_chunk)
4731                 profile = btrfs_system_alloc_profile(fs_info);
4732         else
4733                 profile = btrfs_metadata_alloc_profile(fs_info);
4734
4735         used = btrfs_space_info_used(space_info, false);
4736
4737         /*
4738          * We only want to allow over committing if we have lots of actual space
4739          * free, but if we don't have enough space to handle the global reserve
4740          * space then we could end up having a real enospc problem when trying
4741          * to allocate a chunk or some other such important allocation.
4742          */
4743         spin_lock(&global_rsv->lock);
4744         space_size = calc_global_rsv_need_space(global_rsv);
4745         spin_unlock(&global_rsv->lock);
4746         if (used + space_size >= space_info->total_bytes)
4747                 return 0;
4748
4749         used += space_info->bytes_may_use;
4750
4751         avail = atomic64_read(&fs_info->free_chunk_space);
4752
4753         /*
4754          * If we have dup, raid1 or raid10 then only half of the free
4755          * space is actually useable.  For raid56, the space info used
4756          * doesn't include the parity drive, so we don't have to
4757          * change the math
4758          */
4759         if (profile & (BTRFS_BLOCK_GROUP_DUP |
4760                        BTRFS_BLOCK_GROUP_RAID1 |
4761                        BTRFS_BLOCK_GROUP_RAID10))
4762                 avail >>= 1;
4763
4764         /*
4765          * If we aren't flushing all things, let us overcommit up to
4766          * 1/2th of the space. If we can flush, don't let us overcommit
4767          * too much, let it overcommit up to 1/8 of the space.
4768          */
4769         if (flush == BTRFS_RESERVE_FLUSH_ALL)
4770                 avail >>= 3;
4771         else
4772                 avail >>= 1;
4773
4774         if (used + bytes < space_info->total_bytes + avail)
4775                 return 1;
4776         return 0;
4777 }
4778
4779 static void btrfs_writeback_inodes_sb_nr(struct btrfs_fs_info *fs_info,
4780                                          unsigned long nr_pages, int nr_items)
4781 {
4782         struct super_block *sb = fs_info->sb;
4783
4784         if (down_read_trylock(&sb->s_umount)) {
4785                 writeback_inodes_sb_nr(sb, nr_pages, WB_REASON_FS_FREE_SPACE);
4786                 up_read(&sb->s_umount);
4787         } else {
4788                 /*
4789                  * We needn't worry the filesystem going from r/w to r/o though
4790                  * we don't acquire ->s_umount mutex, because the filesystem
4791                  * should guarantee the delalloc inodes list be empty after
4792                  * the filesystem is readonly(all dirty pages are written to
4793                  * the disk).
4794                  */
4795                 btrfs_start_delalloc_roots(fs_info, 0, nr_items);
4796                 if (!current->journal_info)
4797                         btrfs_wait_ordered_roots(fs_info, nr_items, 0, (u64)-1);
4798         }
4799 }
4800
4801 static inline u64 calc_reclaim_items_nr(struct btrfs_fs_info *fs_info,
4802                                         u64 to_reclaim)
4803 {
4804         u64 bytes;
4805         u64 nr;
4806
4807         bytes = btrfs_calc_trans_metadata_size(fs_info, 1);
4808         nr = div64_u64(to_reclaim, bytes);
4809         if (!nr)
4810                 nr = 1;
4811         return nr;
4812 }
4813
4814 #define EXTENT_SIZE_PER_ITEM    SZ_256K
4815
4816 /*
4817  * shrink metadata reservation for delalloc
4818  */
4819 static void shrink_delalloc(struct btrfs_fs_info *fs_info, u64 to_reclaim,
4820                             u64 orig, bool wait_ordered)
4821 {
4822         struct btrfs_space_info *space_info;
4823         struct btrfs_trans_handle *trans;
4824         u64 delalloc_bytes;
4825         u64 max_reclaim;
4826         u64 items;
4827         long time_left;
4828         unsigned long nr_pages;
4829         int loops;
4830
4831         /* Calc the number of the pages we need flush for space reservation */
4832         items = calc_reclaim_items_nr(fs_info, to_reclaim);
4833         to_reclaim = items * EXTENT_SIZE_PER_ITEM;
4834
4835         trans = (struct btrfs_trans_handle *)current->journal_info;
4836         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
4837
4838         delalloc_bytes = percpu_counter_sum_positive(
4839                                                 &fs_info->delalloc_bytes);
4840         if (delalloc_bytes == 0) {
4841                 if (trans)
4842                         return;
4843                 if (wait_ordered)
4844                         btrfs_wait_ordered_roots(fs_info, items, 0, (u64)-1);
4845                 return;
4846         }
4847
4848         loops = 0;
4849         while (delalloc_bytes && loops < 3) {
4850                 max_reclaim = min(delalloc_bytes, to_reclaim);
4851                 nr_pages = max_reclaim >> PAGE_SHIFT;
4852                 btrfs_writeback_inodes_sb_nr(fs_info, nr_pages, items);
4853                 /*
4854                  * We need to wait for the async pages to actually start before
4855                  * we do anything.
4856                  */
4857                 max_reclaim = atomic_read(&fs_info->async_delalloc_pages);
4858                 if (!max_reclaim)
4859                         goto skip_async;
4860
4861                 if (max_reclaim <= nr_pages)
4862                         max_reclaim = 0;
4863                 else
4864                         max_reclaim -= nr_pages;
4865
4866                 wait_event(fs_info->async_submit_wait,
4867                            atomic_read(&fs_info->async_delalloc_pages) <=
4868                            (int)max_reclaim);
4869 skip_async:
4870                 spin_lock(&space_info->lock);
4871                 if (list_empty(&space_info->tickets) &&
4872                     list_empty(&space_info->priority_tickets)) {
4873                         spin_unlock(&space_info->lock);
4874                         break;
4875                 }
4876                 spin_unlock(&space_info->lock);
4877
4878                 loops++;
4879                 if (wait_ordered && !trans) {
4880                         btrfs_wait_ordered_roots(fs_info, items, 0, (u64)-1);
4881                 } else {
4882                         time_left = schedule_timeout_killable(1);
4883                         if (time_left)
4884                                 break;
4885                 }
4886                 delalloc_bytes = percpu_counter_sum_positive(
4887                                                 &fs_info->delalloc_bytes);
4888         }
4889 }
4890
4891 struct reserve_ticket {
4892         u64 bytes;
4893         int error;
4894         struct list_head list;
4895         wait_queue_head_t wait;
4896 };
4897
4898 /**
4899  * maybe_commit_transaction - possibly commit the transaction if its ok to
4900  * @root - the root we're allocating for
4901  * @bytes - the number of bytes we want to reserve
4902  * @force - force the commit
4903  *
4904  * This will check to make sure that committing the transaction will actually
4905  * get us somewhere and then commit the transaction if it does.  Otherwise it
4906  * will return -ENOSPC.
4907  */
4908 static int may_commit_transaction(struct btrfs_fs_info *fs_info,
4909                                   struct btrfs_space_info *space_info)
4910 {
4911         struct reserve_ticket *ticket = NULL;
4912         struct btrfs_block_rsv *delayed_rsv = &fs_info->delayed_block_rsv;
4913         struct btrfs_trans_handle *trans;
4914         u64 bytes;
4915
4916         trans = (struct btrfs_trans_handle *)current->journal_info;
4917         if (trans)
4918                 return -EAGAIN;
4919
4920         spin_lock(&space_info->lock);
4921         if (!list_empty(&space_info->priority_tickets))
4922                 ticket = list_first_entry(&space_info->priority_tickets,
4923                                           struct reserve_ticket, list);
4924         else if (!list_empty(&space_info->tickets))
4925                 ticket = list_first_entry(&space_info->tickets,
4926                                           struct reserve_ticket, list);
4927         bytes = (ticket) ? ticket->bytes : 0;
4928         spin_unlock(&space_info->lock);
4929
4930         if (!bytes)
4931                 return 0;
4932
4933         /* See if there is enough pinned space to make this reservation */
4934         if (percpu_counter_compare(&space_info->total_bytes_pinned,
4935                                    bytes) >= 0)
4936                 goto commit;
4937
4938         /*
4939          * See if there is some space in the delayed insertion reservation for
4940          * this reservation.
4941          */
4942         if (space_info != delayed_rsv->space_info)
4943                 return -ENOSPC;
4944
4945         spin_lock(&delayed_rsv->lock);
4946         if (delayed_rsv->size > bytes)
4947                 bytes = 0;
4948         else
4949                 bytes -= delayed_rsv->size;
4950         spin_unlock(&delayed_rsv->lock);
4951
4952         if (percpu_counter_compare(&space_info->total_bytes_pinned,
4953                                    bytes) < 0) {
4954                 return -ENOSPC;
4955         }
4956
4957 commit:
4958         trans = btrfs_join_transaction(fs_info->extent_root);
4959         if (IS_ERR(trans))
4960                 return -ENOSPC;
4961
4962         return btrfs_commit_transaction(trans);
4963 }
4964
4965 /*
4966  * Try to flush some data based on policy set by @state. This is only advisory
4967  * and may fail for various reasons. The caller is supposed to examine the
4968  * state of @space_info to detect the outcome.
4969  */
4970 static void flush_space(struct btrfs_fs_info *fs_info,
4971                        struct btrfs_space_info *space_info, u64 num_bytes,
4972                        int state)
4973 {
4974         struct btrfs_root *root = fs_info->extent_root;
4975         struct btrfs_trans_handle *trans;
4976         int nr;
4977         int ret = 0;
4978
4979         switch (state) {
4980         case FLUSH_DELAYED_ITEMS_NR:
4981         case FLUSH_DELAYED_ITEMS:
4982                 if (state == FLUSH_DELAYED_ITEMS_NR)
4983                         nr = calc_reclaim_items_nr(fs_info, num_bytes) * 2;
4984                 else
4985                         nr = -1;
4986
4987                 trans = btrfs_join_transaction(root);
4988                 if (IS_ERR(trans)) {
4989                         ret = PTR_ERR(trans);
4990                         break;
4991                 }
4992                 ret = btrfs_run_delayed_items_nr(trans, nr);
4993                 btrfs_end_transaction(trans);
4994                 break;
4995         case FLUSH_DELALLOC:
4996         case FLUSH_DELALLOC_WAIT:
4997                 shrink_delalloc(fs_info, num_bytes * 2, num_bytes,
4998                                 state == FLUSH_DELALLOC_WAIT);
4999                 break;
5000         case ALLOC_CHUNK:
5001                 trans = btrfs_join_transaction(root);
5002                 if (IS_ERR(trans)) {
5003                         ret = PTR_ERR(trans);
5004                         break;
5005                 }
5006                 ret = do_chunk_alloc(trans, fs_info,
5007                                      btrfs_metadata_alloc_profile(fs_info),
5008                                      CHUNK_ALLOC_NO_FORCE);
5009                 btrfs_end_transaction(trans);
5010                 if (ret > 0 || ret == -ENOSPC)
5011                         ret = 0;
5012                 break;
5013         case COMMIT_TRANS:
5014                 ret = may_commit_transaction(fs_info, space_info);
5015                 break;
5016         default:
5017                 ret = -ENOSPC;
5018                 break;
5019         }
5020
5021         trace_btrfs_flush_space(fs_info, space_info->flags, num_bytes, state,
5022                                 ret);
5023         return;
5024 }
5025
5026 static inline u64
5027 btrfs_calc_reclaim_metadata_size(struct btrfs_fs_info *fs_info,
5028                                  struct btrfs_space_info *space_info,
5029                                  bool system_chunk)
5030 {
5031         struct reserve_ticket *ticket;
5032         u64 used;
5033         u64 expected;
5034         u64 to_reclaim = 0;
5035
5036         list_for_each_entry(ticket, &space_info->tickets, list)
5037                 to_reclaim += ticket->bytes;
5038         list_for_each_entry(ticket, &space_info->priority_tickets, list)
5039                 to_reclaim += ticket->bytes;
5040         if (to_reclaim)
5041                 return to_reclaim;
5042
5043         to_reclaim = min_t(u64, num_online_cpus() * SZ_1M, SZ_16M);
5044         if (can_overcommit(fs_info, space_info, to_reclaim,
5045                            BTRFS_RESERVE_FLUSH_ALL, system_chunk))
5046                 return 0;
5047
5048         used = btrfs_space_info_used(space_info, true);
5049
5050         if (can_overcommit(fs_info, space_info, SZ_1M,
5051                            BTRFS_RESERVE_FLUSH_ALL, system_chunk))
5052                 expected = div_factor_fine(space_info->total_bytes, 95);
5053         else
5054                 expected = div_factor_fine(space_info->total_bytes, 90);
5055
5056         if (used > expected)
5057                 to_reclaim = used - expected;
5058         else
5059                 to_reclaim = 0;
5060         to_reclaim = min(to_reclaim, space_info->bytes_may_use +
5061                                      space_info->bytes_reserved);
5062         return to_reclaim;
5063 }
5064
5065 static inline int need_do_async_reclaim(struct btrfs_fs_info *fs_info,
5066                                         struct btrfs_space_info *space_info,
5067                                         u64 used, bool system_chunk)
5068 {
5069         u64 thresh = div_factor_fine(space_info->total_bytes, 98);
5070
5071         /* If we're just plain full then async reclaim just slows us down. */
5072         if ((space_info->bytes_used + space_info->bytes_reserved) >= thresh)
5073                 return 0;
5074
5075         if (!btrfs_calc_reclaim_metadata_size(fs_info, space_info,
5076                                               system_chunk))
5077                 return 0;
5078
5079         return (used >= thresh && !btrfs_fs_closing(fs_info) &&
5080                 !test_bit(BTRFS_FS_STATE_REMOUNTING, &fs_info->fs_state));
5081 }
5082
5083 static void wake_all_tickets(struct list_head *head)
5084 {
5085         struct reserve_ticket *ticket;
5086
5087         while (!list_empty(head)) {
5088                 ticket = list_first_entry(head, struct reserve_ticket, list);
5089                 list_del_init(&ticket->list);
5090                 ticket->error = -ENOSPC;
5091                 wake_up(&ticket->wait);
5092         }
5093 }
5094
5095 /*
5096  * This is for normal flushers, we can wait all goddamned day if we want to.  We
5097  * will loop and continuously try to flush as long as we are making progress.
5098  * We count progress as clearing off tickets each time we have to loop.
5099  */
5100 static void btrfs_async_reclaim_metadata_space(struct work_struct *work)
5101 {
5102         struct btrfs_fs_info *fs_info;
5103         struct btrfs_space_info *space_info;
5104         u64 to_reclaim;
5105         int flush_state;
5106         int commit_cycles = 0;
5107         u64 last_tickets_id;
5108
5109         fs_info = container_of(work, struct btrfs_fs_info, async_reclaim_work);
5110         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
5111
5112         spin_lock(&space_info->lock);
5113         to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info, space_info,
5114                                                       false);
5115         if (!to_reclaim) {
5116                 space_info->flush = 0;
5117                 spin_unlock(&space_info->lock);
5118                 return;
5119         }
5120         last_tickets_id = space_info->tickets_id;
5121         spin_unlock(&space_info->lock);
5122
5123         flush_state = FLUSH_DELAYED_ITEMS_NR;
5124         do {
5125                 flush_space(fs_info, space_info, to_reclaim, flush_state);
5126                 spin_lock(&space_info->lock);
5127                 if (list_empty(&space_info->tickets)) {
5128                         space_info->flush = 0;
5129                         spin_unlock(&space_info->lock);
5130                         return;
5131                 }
5132                 to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info,
5133                                                               space_info,
5134                                                               false);
5135                 if (last_tickets_id == space_info->tickets_id) {
5136                         flush_state++;
5137                 } else {
5138                         last_tickets_id = space_info->tickets_id;
5139                         flush_state = FLUSH_DELAYED_ITEMS_NR;
5140                         if (commit_cycles)
5141                                 commit_cycles--;
5142                 }
5143
5144                 if (flush_state > COMMIT_TRANS) {
5145                         commit_cycles++;
5146                         if (commit_cycles > 2) {
5147                                 wake_all_tickets(&space_info->tickets);
5148                                 space_info->flush = 0;
5149                         } else {
5150                                 flush_state = FLUSH_DELAYED_ITEMS_NR;
5151                         }
5152                 }
5153                 spin_unlock(&space_info->lock);
5154         } while (flush_state <= COMMIT_TRANS);
5155 }
5156
5157 void btrfs_init_async_reclaim_work(struct work_struct *work)
5158 {
5159         INIT_WORK(work, btrfs_async_reclaim_metadata_space);
5160 }
5161
5162 static void priority_reclaim_metadata_space(struct btrfs_fs_info *fs_info,
5163                                             struct btrfs_space_info *space_info,
5164                                             struct reserve_ticket *ticket)
5165 {
5166         u64 to_reclaim;
5167         int flush_state = FLUSH_DELAYED_ITEMS_NR;
5168
5169         spin_lock(&space_info->lock);
5170         to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info, space_info,
5171                                                       false);
5172         if (!to_reclaim) {
5173                 spin_unlock(&space_info->lock);
5174                 return;
5175         }
5176         spin_unlock(&space_info->lock);
5177
5178         do {
5179                 flush_space(fs_info, space_info, to_reclaim, flush_state);
5180                 flush_state++;
5181                 spin_lock(&space_info->lock);
5182                 if (ticket->bytes == 0) {
5183                         spin_unlock(&space_info->lock);
5184                         return;
5185                 }
5186                 spin_unlock(&space_info->lock);
5187
5188                 /*
5189                  * Priority flushers can't wait on delalloc without
5190                  * deadlocking.
5191                  */
5192                 if (flush_state == FLUSH_DELALLOC ||
5193                     flush_state == FLUSH_DELALLOC_WAIT)
5194                         flush_state = ALLOC_CHUNK;
5195         } while (flush_state < COMMIT_TRANS);
5196 }
5197
5198 static int wait_reserve_ticket(struct btrfs_fs_info *fs_info,
5199                                struct btrfs_space_info *space_info,
5200                                struct reserve_ticket *ticket, u64 orig_bytes)
5201
5202 {
5203         DEFINE_WAIT(wait);
5204         int ret = 0;
5205
5206         spin_lock(&space_info->lock);
5207         while (ticket->bytes > 0 && ticket->error == 0) {
5208                 ret = prepare_to_wait_event(&ticket->wait, &wait, TASK_KILLABLE);
5209                 if (ret) {
5210                         ret = -EINTR;
5211                         break;
5212                 }
5213                 spin_unlock(&space_info->lock);
5214
5215                 schedule();
5216
5217                 finish_wait(&ticket->wait, &wait);
5218                 spin_lock(&space_info->lock);
5219         }
5220         if (!ret)
5221                 ret = ticket->error;
5222         if (!list_empty(&ticket->list))
5223                 list_del_init(&ticket->list);
5224         if (ticket->bytes && ticket->bytes < orig_bytes) {
5225                 u64 num_bytes = orig_bytes - ticket->bytes;
5226                 space_info->bytes_may_use -= num_bytes;
5227                 trace_btrfs_space_reservation(fs_info, "space_info",
5228                                               space_info->flags, num_bytes, 0);
5229         }
5230         spin_unlock(&space_info->lock);
5231
5232         return ret;
5233 }
5234
5235 /**
5236  * reserve_metadata_bytes - try to reserve bytes from the block_rsv's space
5237  * @root - the root we're allocating for
5238  * @space_info - the space info we want to allocate from
5239  * @orig_bytes - the number of bytes we want
5240  * @flush - whether or not we can flush to make our reservation
5241  *
5242  * This will reserve orig_bytes number of bytes from the space info associated
5243  * with the block_rsv.  If there is not enough space it will make an attempt to
5244  * flush out space to make room.  It will do this by flushing delalloc if
5245  * possible or committing the transaction.  If flush is 0 then no attempts to
5246  * regain reservations will be made and this will fail if there is not enough
5247  * space already.
5248  */
5249 static int __reserve_metadata_bytes(struct btrfs_fs_info *fs_info,
5250                                     struct btrfs_space_info *space_info,
5251                                     u64 orig_bytes,
5252                                     enum btrfs_reserve_flush_enum flush,
5253                                     bool system_chunk)
5254 {
5255         struct reserve_ticket ticket;
5256         u64 used;
5257         int ret = 0;
5258
5259         ASSERT(orig_bytes);
5260         ASSERT(!current->journal_info || flush != BTRFS_RESERVE_FLUSH_ALL);
5261
5262         spin_lock(&space_info->lock);
5263         ret = -ENOSPC;
5264         used = btrfs_space_info_used(space_info, true);
5265
5266         /*
5267          * If we have enough space then hooray, make our reservation and carry
5268          * on.  If not see if we can overcommit, and if we can, hooray carry on.
5269          * If not things get more complicated.
5270          */
5271         if (used + orig_bytes <= space_info->total_bytes) {
5272                 space_info->bytes_may_use += orig_bytes;
5273                 trace_btrfs_space_reservation(fs_info, "space_info",
5274                                               space_info->flags, orig_bytes, 1);
5275                 ret = 0;
5276         } else if (can_overcommit(fs_info, space_info, orig_bytes, flush,
5277                                   system_chunk)) {
5278                 space_info->bytes_may_use += orig_bytes;
5279                 trace_btrfs_space_reservation(fs_info, "space_info",
5280                                               space_info->flags, orig_bytes, 1);
5281                 ret = 0;
5282         }
5283
5284         /*
5285          * If we couldn't make a reservation then setup our reservation ticket
5286          * and kick the async worker if it's not already running.
5287          *
5288          * If we are a priority flusher then we just need to add our ticket to
5289          * the list and we will do our own flushing further down.
5290          */
5291         if (ret && flush != BTRFS_RESERVE_NO_FLUSH) {
5292                 ticket.bytes = orig_bytes;
5293                 ticket.error = 0;
5294                 init_waitqueue_head(&ticket.wait);
5295                 if (flush == BTRFS_RESERVE_FLUSH_ALL) {
5296                         list_add_tail(&ticket.list, &space_info->tickets);
5297                         if (!space_info->flush) {
5298                                 space_info->flush = 1;
5299                                 trace_btrfs_trigger_flush(fs_info,
5300                                                           space_info->flags,
5301                                                           orig_bytes, flush,
5302                                                           "enospc");
5303                                 queue_work(system_unbound_wq,
5304                                            &fs_info->async_reclaim_work);
5305                         }
5306                 } else {
5307                         list_add_tail(&ticket.list,
5308                                       &space_info->priority_tickets);
5309                 }
5310         } else if (!ret && space_info->flags & BTRFS_BLOCK_GROUP_METADATA) {
5311                 used += orig_bytes;
5312                 /*
5313                  * We will do the space reservation dance during log replay,
5314                  * which means we won't have fs_info->fs_root set, so don't do
5315                  * the async reclaim as we will panic.
5316                  */
5317                 if (!test_bit(BTRFS_FS_LOG_RECOVERING, &fs_info->flags) &&
5318                     need_do_async_reclaim(fs_info, space_info,
5319                                           used, system_chunk) &&
5320                     !work_busy(&fs_info->async_reclaim_work)) {
5321                         trace_btrfs_trigger_flush(fs_info, space_info->flags,
5322                                                   orig_bytes, flush, "preempt");
5323                         queue_work(system_unbound_wq,
5324                                    &fs_info->async_reclaim_work);
5325                 }
5326         }
5327         spin_unlock(&space_info->lock);
5328         if (!ret || flush == BTRFS_RESERVE_NO_FLUSH)
5329                 return ret;
5330
5331         if (flush == BTRFS_RESERVE_FLUSH_ALL)
5332                 return wait_reserve_ticket(fs_info, space_info, &ticket,
5333                                            orig_bytes);
5334
5335         ret = 0;
5336         priority_reclaim_metadata_space(fs_info, space_info, &ticket);
5337         spin_lock(&space_info->lock);
5338         if (ticket.bytes) {
5339                 if (ticket.bytes < orig_bytes) {
5340                         u64 num_bytes = orig_bytes - ticket.bytes;
5341                         space_info->bytes_may_use -= num_bytes;
5342                         trace_btrfs_space_reservation(fs_info, "space_info",
5343                                                       space_info->flags,
5344                                                       num_bytes, 0);
5345
5346                 }
5347                 list_del_init(&ticket.list);
5348                 ret = -ENOSPC;
5349         }
5350         spin_unlock(&space_info->lock);
5351         ASSERT(list_empty(&ticket.list));
5352         return ret;
5353 }
5354
5355 /**
5356  * reserve_metadata_bytes - try to reserve bytes from the block_rsv's space
5357  * @root - the root we're allocating for
5358  * @block_rsv - the block_rsv we're allocating for
5359  * @orig_bytes - the number of bytes we want
5360  * @flush - whether or not we can flush to make our reservation
5361  *
5362  * This will reserve orgi_bytes number of bytes from the space info associated
5363  * with the block_rsv.  If there is not enough space it will make an attempt to
5364  * flush out space to make room.  It will do this by flushing delalloc if
5365  * possible or committing the transaction.  If flush is 0 then no attempts to
5366  * regain reservations will be made and this will fail if there is not enough
5367  * space already.
5368  */
5369 static int reserve_metadata_bytes(struct btrfs_root *root,
5370                                   struct btrfs_block_rsv *block_rsv,
5371                                   u64 orig_bytes,
5372                                   enum btrfs_reserve_flush_enum flush)
5373 {
5374         struct btrfs_fs_info *fs_info = root->fs_info;
5375         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
5376         int ret;
5377         bool system_chunk = (root == fs_info->chunk_root);
5378
5379         ret = __reserve_metadata_bytes(fs_info, block_rsv->space_info,
5380                                        orig_bytes, flush, system_chunk);
5381         if (ret == -ENOSPC &&
5382             unlikely(root->orphan_cleanup_state == ORPHAN_CLEANUP_STARTED)) {
5383                 if (block_rsv != global_rsv &&
5384                     !block_rsv_use_bytes(global_rsv, orig_bytes))
5385                         ret = 0;
5386         }
5387         if (ret == -ENOSPC) {
5388                 trace_btrfs_space_reservation(fs_info, "space_info:enospc",
5389                                               block_rsv->space_info->flags,
5390                                               orig_bytes, 1);
5391
5392                 if (btrfs_test_opt(fs_info, ENOSPC_DEBUG))
5393                         dump_space_info(fs_info, block_rsv->space_info,
5394                                         orig_bytes, 0);
5395         }
5396         return ret;
5397 }
5398
5399 static struct btrfs_block_rsv *get_block_rsv(
5400                                         const struct btrfs_trans_handle *trans,
5401                                         const struct btrfs_root *root)
5402 {
5403         struct btrfs_fs_info *fs_info = root->fs_info;
5404         struct btrfs_block_rsv *block_rsv = NULL;
5405
5406         if (test_bit(BTRFS_ROOT_REF_COWS, &root->state) ||
5407             (root == fs_info->csum_root && trans->adding_csums) ||
5408             (root == fs_info->uuid_root))
5409                 block_rsv = trans->block_rsv;
5410
5411         if (!block_rsv)
5412                 block_rsv = root->block_rsv;
5413
5414         if (!block_rsv)
5415                 block_rsv = &fs_info->empty_block_rsv;
5416
5417         return block_rsv;
5418 }
5419
5420 static int block_rsv_use_bytes(struct btrfs_block_rsv *block_rsv,
5421                                u64 num_bytes)
5422 {
5423         int ret = -ENOSPC;
5424         spin_lock(&block_rsv->lock);
5425         if (block_rsv->reserved >= num_bytes) {
5426                 block_rsv->reserved -= num_bytes;
5427                 if (block_rsv->reserved < block_rsv->size)
5428                         block_rsv->full = 0;
5429                 ret = 0;
5430         }
5431         spin_unlock(&block_rsv->lock);
5432         return ret;
5433 }
5434
5435 static void block_rsv_add_bytes(struct btrfs_block_rsv *block_rsv,
5436                                 u64 num_bytes, int update_size)
5437 {
5438         spin_lock(&block_rsv->lock);
5439         block_rsv->reserved += num_bytes;
5440         if (update_size)
5441                 block_rsv->size += num_bytes;
5442         else if (block_rsv->reserved >= block_rsv->size)
5443                 block_rsv->full = 1;
5444         spin_unlock(&block_rsv->lock);
5445 }
5446
5447 int btrfs_cond_migrate_bytes(struct btrfs_fs_info *fs_info,
5448                              struct btrfs_block_rsv *dest, u64 num_bytes,
5449                              int min_factor)
5450 {
5451         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
5452         u64 min_bytes;
5453
5454         if (global_rsv->space_info != dest->space_info)
5455                 return -ENOSPC;
5456
5457         spin_lock(&global_rsv->lock);
5458         min_bytes = div_factor(global_rsv->size, min_factor);
5459         if (global_rsv->reserved < min_bytes + num_bytes) {
5460                 spin_unlock(&global_rsv->lock);
5461                 return -ENOSPC;
5462         }
5463         global_rsv->reserved -= num_bytes;
5464         if (global_rsv->reserved < global_rsv->size)
5465                 global_rsv->full = 0;
5466         spin_unlock(&global_rsv->lock);
5467
5468         block_rsv_add_bytes(dest, num_bytes, 1);
5469         return 0;
5470 }
5471
5472 /*
5473  * This is for space we already have accounted in space_info->bytes_may_use, so
5474  * basically when we're returning space from block_rsv's.
5475  */
5476 static void space_info_add_old_bytes(struct btrfs_fs_info *fs_info,
5477                                      struct btrfs_space_info *space_info,
5478                                      u64 num_bytes)
5479 {
5480         struct reserve_ticket *ticket;
5481         struct list_head *head;
5482         u64 used;
5483         enum btrfs_reserve_flush_enum flush = BTRFS_RESERVE_NO_FLUSH;
5484         bool check_overcommit = false;
5485
5486         spin_lock(&space_info->lock);
5487         head = &space_info->priority_tickets;
5488
5489         /*
5490          * If we are over our limit then we need to check and see if we can
5491          * overcommit, and if we can't then we just need to free up our space
5492          * and not satisfy any requests.
5493          */
5494         used = btrfs_space_info_used(space_info, true);
5495         if (used - num_bytes >= space_info->total_bytes)
5496                 check_overcommit = true;
5497 again:
5498         while (!list_empty(head) && num_bytes) {
5499                 ticket = list_first_entry(head, struct reserve_ticket,
5500                                           list);
5501                 /*
5502                  * We use 0 bytes because this space is already reserved, so
5503                  * adding the ticket space would be a double count.
5504                  */
5505                 if (check_overcommit &&
5506                     !can_overcommit(fs_info, space_info, 0, flush, false))
5507                         break;
5508                 if (num_bytes >= ticket->bytes) {
5509                         list_del_init(&ticket->list);
5510                         num_bytes -= ticket->bytes;
5511                         ticket->bytes = 0;
5512                         space_info->tickets_id++;
5513                         wake_up(&ticket->wait);
5514                 } else {
5515                         ticket->bytes -= num_bytes;
5516                         num_bytes = 0;
5517                 }
5518         }
5519
5520         if (num_bytes && head == &space_info->priority_tickets) {
5521                 head = &space_info->tickets;
5522                 flush = BTRFS_RESERVE_FLUSH_ALL;
5523                 goto again;
5524         }
5525         space_info->bytes_may_use -= num_bytes;
5526         trace_btrfs_space_reservation(fs_info, "space_info",
5527                                       space_info->flags, num_bytes, 0);
5528         spin_unlock(&space_info->lock);
5529 }
5530
5531 /*
5532  * This is for newly allocated space that isn't accounted in
5533  * space_info->bytes_may_use yet.  So if we allocate a chunk or unpin an extent
5534  * we use this helper.
5535  */
5536 static void space_info_add_new_bytes(struct btrfs_fs_info *fs_info,
5537                                      struct btrfs_space_info *space_info,
5538                                      u64 num_bytes)
5539 {
5540         struct reserve_ticket *ticket;
5541         struct list_head *head = &space_info->priority_tickets;
5542
5543 again:
5544         while (!list_empty(head) && num_bytes) {
5545                 ticket = list_first_entry(head, struct reserve_ticket,
5546                                           list);
5547                 if (num_bytes >= ticket->bytes) {
5548                         trace_btrfs_space_reservation(fs_info, "space_info",
5549                                                       space_info->flags,
5550                                                       ticket->bytes, 1);
5551                         list_del_init(&ticket->list);
5552                         num_bytes -= ticket->bytes;
5553                         space_info->bytes_may_use += ticket->bytes;
5554                         ticket->bytes = 0;
5555                         space_info->tickets_id++;
5556                         wake_up(&ticket->wait);
5557                 } else {
5558                         trace_btrfs_space_reservation(fs_info, "space_info",
5559                                                       space_info->flags,
5560                                                       num_bytes, 1);
5561                         space_info->bytes_may_use += num_bytes;
5562                         ticket->bytes -= num_bytes;
5563                         num_bytes = 0;
5564                 }
5565         }
5566
5567         if (num_bytes && head == &space_info->priority_tickets) {
5568                 head = &space_info->tickets;
5569                 goto again;
5570         }
5571 }
5572
5573 static u64 block_rsv_release_bytes(struct btrfs_fs_info *fs_info,
5574                                     struct btrfs_block_rsv *block_rsv,
5575                                     struct btrfs_block_rsv *dest, u64 num_bytes,
5576                                     u64 *qgroup_to_release_ret)
5577 {
5578         struct btrfs_space_info *space_info = block_rsv->space_info;
5579         u64 qgroup_to_release = 0;
5580         u64 ret;
5581
5582         spin_lock(&block_rsv->lock);
5583         if (num_bytes == (u64)-1) {
5584                 num_bytes = block_rsv->size;
5585                 qgroup_to_release = block_rsv->qgroup_rsv_size;
5586         }
5587         block_rsv->size -= num_bytes;
5588         if (block_rsv->reserved >= block_rsv->size) {
5589                 num_bytes = block_rsv->reserved - block_rsv->size;
5590                 block_rsv->reserved = block_rsv->size;
5591                 block_rsv->full = 1;
5592         } else {
5593                 num_bytes = 0;
5594         }
5595         if (block_rsv->qgroup_rsv_reserved >= block_rsv->qgroup_rsv_size) {
5596                 qgroup_to_release = block_rsv->qgroup_rsv_reserved -
5597                                     block_rsv->qgroup_rsv_size;
5598                 block_rsv->qgroup_rsv_reserved = block_rsv->qgroup_rsv_size;
5599         } else {
5600                 qgroup_to_release = 0;
5601         }
5602         spin_unlock(&block_rsv->lock);
5603
5604         ret = num_bytes;
5605         if (num_bytes > 0) {
5606                 if (dest) {
5607                         spin_lock(&dest->lock);
5608                         if (!dest->full) {
5609                                 u64 bytes_to_add;
5610
5611                                 bytes_to_add = dest->size - dest->reserved;
5612                                 bytes_to_add = min(num_bytes, bytes_to_add);
5613                                 dest->reserved += bytes_to_add;
5614                                 if (dest->reserved >= dest->size)
5615                                         dest->full = 1;
5616                                 num_bytes -= bytes_to_add;
5617                         }
5618                         spin_unlock(&dest->lock);
5619                 }
5620                 if (num_bytes)
5621                         space_info_add_old_bytes(fs_info, space_info,
5622                                                  num_bytes);
5623         }
5624         if (qgroup_to_release_ret)
5625                 *qgroup_to_release_ret = qgroup_to_release;
5626         return ret;
5627 }
5628
5629 int btrfs_block_rsv_migrate(struct btrfs_block_rsv *src,
5630                             struct btrfs_block_rsv *dst, u64 num_bytes,
5631                             int update_size)
5632 {
5633         int ret;
5634
5635         ret = block_rsv_use_bytes(src, num_bytes);
5636         if (ret)
5637                 return ret;
5638
5639         block_rsv_add_bytes(dst, num_bytes, update_size);
5640         return 0;
5641 }
5642
5643 void btrfs_init_block_rsv(struct btrfs_block_rsv *rsv, unsigned short type)
5644 {
5645         memset(rsv, 0, sizeof(*rsv));
5646         spin_lock_init(&rsv->lock);
5647         rsv->type = type;
5648 }
5649
5650 void btrfs_init_metadata_block_rsv(struct btrfs_fs_info *fs_info,
5651                                    struct btrfs_block_rsv *rsv,
5652                                    unsigned short type)
5653 {
5654         btrfs_init_block_rsv(rsv, type);
5655         rsv->space_info = __find_space_info(fs_info,
5656                                             BTRFS_BLOCK_GROUP_METADATA);
5657 }
5658
5659 struct btrfs_block_rsv *btrfs_alloc_block_rsv(struct btrfs_fs_info *fs_info,
5660                                               unsigned short type)
5661 {
5662         struct btrfs_block_rsv *block_rsv;
5663
5664         block_rsv = kmalloc(sizeof(*block_rsv), GFP_NOFS);
5665         if (!block_rsv)
5666                 return NULL;
5667
5668         btrfs_init_metadata_block_rsv(fs_info, block_rsv, type);
5669         return block_rsv;
5670 }
5671
5672 void btrfs_free_block_rsv(struct btrfs_fs_info *fs_info,
5673                           struct btrfs_block_rsv *rsv)
5674 {
5675         if (!rsv)
5676                 return;
5677         btrfs_block_rsv_release(fs_info, rsv, (u64)-1);
5678         kfree(rsv);
5679 }
5680
5681 void __btrfs_free_block_rsv(struct btrfs_block_rsv *rsv)
5682 {
5683         kfree(rsv);
5684 }
5685
5686 int btrfs_block_rsv_add(struct btrfs_root *root,
5687                         struct btrfs_block_rsv *block_rsv, u64 num_bytes,
5688                         enum btrfs_reserve_flush_enum flush)
5689 {
5690         int ret;
5691
5692         if (num_bytes == 0)
5693                 return 0;
5694
5695         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
5696         if (!ret) {
5697                 block_rsv_add_bytes(block_rsv, num_bytes, 1);
5698                 return 0;
5699         }
5700
5701         return ret;
5702 }
5703
5704 int btrfs_block_rsv_check(struct btrfs_block_rsv *block_rsv, int min_factor)
5705 {
5706         u64 num_bytes = 0;
5707         int ret = -ENOSPC;
5708
5709         if (!block_rsv)
5710                 return 0;
5711
5712         spin_lock(&block_rsv->lock);
5713         num_bytes = div_factor(block_rsv->size, min_factor);
5714         if (block_rsv->reserved >= num_bytes)
5715                 ret = 0;
5716         spin_unlock(&block_rsv->lock);
5717
5718         return ret;
5719 }
5720
5721 int btrfs_block_rsv_refill(struct btrfs_root *root,
5722                            struct btrfs_block_rsv *block_rsv, u64 min_reserved,
5723                            enum btrfs_reserve_flush_enum flush)
5724 {
5725         u64 num_bytes = 0;
5726         int ret = -ENOSPC;
5727
5728         if (!block_rsv)
5729                 return 0;
5730
5731         spin_lock(&block_rsv->lock);
5732         num_bytes = min_reserved;
5733         if (block_rsv->reserved >= num_bytes)
5734                 ret = 0;
5735         else
5736                 num_bytes -= block_rsv->reserved;
5737         spin_unlock(&block_rsv->lock);
5738
5739         if (!ret)
5740                 return 0;
5741
5742         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
5743         if (!ret) {
5744                 block_rsv_add_bytes(block_rsv, num_bytes, 0);
5745                 return 0;
5746         }
5747
5748         return ret;
5749 }
5750
5751 /**
5752  * btrfs_inode_rsv_refill - refill the inode block rsv.
5753  * @inode - the inode we are refilling.
5754  * @flush - the flusing restriction.
5755  *
5756  * Essentially the same as btrfs_block_rsv_refill, except it uses the
5757  * block_rsv->size as the minimum size.  We'll either refill the missing amount
5758  * or return if we already have enough space.  This will also handle the resreve
5759  * tracepoint for the reserved amount.
5760  */
5761 static int btrfs_inode_rsv_refill(struct btrfs_inode *inode,
5762                                   enum btrfs_reserve_flush_enum flush)
5763 {
5764         struct btrfs_root *root = inode->root;
5765         struct btrfs_block_rsv *block_rsv = &inode->block_rsv;
5766         u64 num_bytes = 0;
5767         u64 qgroup_num_bytes = 0;
5768         int ret = -ENOSPC;
5769
5770         spin_lock(&block_rsv->lock);
5771         if (block_rsv->reserved < block_rsv->size)
5772                 num_bytes = block_rsv->size - block_rsv->reserved;
5773         if (block_rsv->qgroup_rsv_reserved < block_rsv->qgroup_rsv_size)
5774                 qgroup_num_bytes = block_rsv->qgroup_rsv_size -
5775                                    block_rsv->qgroup_rsv_reserved;
5776         spin_unlock(&block_rsv->lock);
5777
5778         if (num_bytes == 0)
5779                 return 0;
5780
5781         ret = btrfs_qgroup_reserve_meta_prealloc(root, qgroup_num_bytes, true);
5782         if (ret)
5783                 return ret;
5784         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
5785         if (!ret) {
5786                 block_rsv_add_bytes(block_rsv, num_bytes, 0);
5787                 trace_btrfs_space_reservation(root->fs_info, "delalloc",
5788                                               btrfs_ino(inode), num_bytes, 1);
5789
5790                 /* Don't forget to increase qgroup_rsv_reserved */
5791                 spin_lock(&block_rsv->lock);
5792                 block_rsv->qgroup_rsv_reserved += qgroup_num_bytes;
5793                 spin_unlock(&block_rsv->lock);
5794         } else
5795                 btrfs_qgroup_free_meta_prealloc(root, qgroup_num_bytes);
5796         return ret;
5797 }
5798
5799 /**
5800  * btrfs_inode_rsv_release - release any excessive reservation.
5801  * @inode - the inode we need to release from.
5802  * @qgroup_free - free or convert qgroup meta.
5803  *   Unlike normal operation, qgroup meta reservation needs to know if we are
5804  *   freeing qgroup reservation or just converting it into per-trans.  Normally
5805  *   @qgroup_free is true for error handling, and false for normal release.
5806  *
5807  * This is the same as btrfs_block_rsv_release, except that it handles the
5808  * tracepoint for the reservation.
5809  */
5810 static void btrfs_inode_rsv_release(struct btrfs_inode *inode, bool qgroup_free)
5811 {
5812         struct btrfs_fs_info *fs_info = inode->root->fs_info;
5813         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
5814         struct btrfs_block_rsv *block_rsv = &inode->block_rsv;
5815         u64 released = 0;
5816         u64 qgroup_to_release = 0;
5817
5818         /*
5819          * Since we statically set the block_rsv->size we just want to say we
5820          * are releasing 0 bytes, and then we'll just get the reservation over
5821          * the size free'd.
5822          */
5823         released = block_rsv_release_bytes(fs_info, block_rsv, global_rsv, 0,
5824                                            &qgroup_to_release);
5825         if (released > 0)
5826                 trace_btrfs_space_reservation(fs_info, "delalloc",
5827                                               btrfs_ino(inode), released, 0);
5828         if (qgroup_free)
5829                 btrfs_qgroup_free_meta_prealloc(inode->root, qgroup_to_release);
5830         else
5831                 btrfs_qgroup_convert_reserved_meta(inode->root,
5832                                                    qgroup_to_release);
5833 }
5834
5835 void btrfs_block_rsv_release(struct btrfs_fs_info *fs_info,
5836                              struct btrfs_block_rsv *block_rsv,
5837                              u64 num_bytes)
5838 {
5839         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
5840
5841         if (global_rsv == block_rsv ||
5842             block_rsv->space_info != global_rsv->space_info)
5843                 global_rsv = NULL;
5844         block_rsv_release_bytes(fs_info, block_rsv, global_rsv, num_bytes, NULL);
5845 }
5846
5847 static void update_global_block_rsv(struct btrfs_fs_info *fs_info)
5848 {
5849         struct btrfs_block_rsv *block_rsv = &fs_info->global_block_rsv;
5850         struct btrfs_space_info *sinfo = block_rsv->space_info;
5851         u64 num_bytes;
5852
5853         /*
5854          * The global block rsv is based on the size of the extent tree, the
5855          * checksum tree and the root tree.  If the fs is empty we want to set
5856          * it to a minimal amount for safety.
5857          */
5858         num_bytes = btrfs_root_used(&fs_info->extent_root->root_item) +
5859                 btrfs_root_used(&fs_info->csum_root->root_item) +
5860                 btrfs_root_used(&fs_info->tree_root->root_item);
5861         num_bytes = max_t(u64, num_bytes, SZ_16M);
5862
5863         spin_lock(&sinfo->lock);
5864         spin_lock(&block_rsv->lock);
5865
5866         block_rsv->size = min_t(u64, num_bytes, SZ_512M);
5867
5868         if (block_rsv->reserved < block_rsv->size) {
5869                 num_bytes = btrfs_space_info_used(sinfo, true);
5870                 if (sinfo->total_bytes > num_bytes) {
5871                         num_bytes = sinfo->total_bytes - num_bytes;
5872                         num_bytes = min(num_bytes,
5873                                         block_rsv->size - block_rsv->reserved);
5874                         block_rsv->reserved += num_bytes;
5875                         sinfo->bytes_may_use += num_bytes;
5876                         trace_btrfs_space_reservation(fs_info, "space_info",
5877                                                       sinfo->flags, num_bytes,
5878                                                       1);
5879                 }
5880         } else if (block_rsv->reserved > block_rsv->size) {
5881                 num_bytes = block_rsv->reserved - block_rsv->size;
5882                 sinfo->bytes_may_use -= num_bytes;
5883                 trace_btrfs_space_reservation(fs_info, "space_info",
5884                                       sinfo->flags, num_bytes, 0);
5885                 block_rsv->reserved = block_rsv->size;
5886         }
5887
5888         if (block_rsv->reserved == block_rsv->size)
5889                 block_rsv->full = 1;
5890         else
5891                 block_rsv->full = 0;
5892
5893         spin_unlock(&block_rsv->lock);
5894         spin_unlock(&sinfo->lock);
5895 }
5896
5897 static void init_global_block_rsv(struct btrfs_fs_info *fs_info)
5898 {
5899         struct btrfs_space_info *space_info;
5900
5901         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
5902         fs_info->chunk_block_rsv.space_info = space_info;
5903
5904         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
5905         fs_info->global_block_rsv.space_info = space_info;
5906         fs_info->trans_block_rsv.space_info = space_info;
5907         fs_info->empty_block_rsv.space_info = space_info;
5908         fs_info->delayed_block_rsv.space_info = space_info;
5909
5910         fs_info->extent_root->block_rsv = &fs_info->global_block_rsv;
5911         fs_info->csum_root->block_rsv = &fs_info->global_block_rsv;
5912         fs_info->dev_root->block_rsv = &fs_info->global_block_rsv;
5913         fs_info->tree_root->block_rsv = &fs_info->global_block_rsv;
5914         if (fs_info->quota_root)
5915                 fs_info->quota_root->block_rsv = &fs_info->global_block_rsv;
5916         fs_info->chunk_root->block_rsv = &fs_info->chunk_block_rsv;
5917
5918         update_global_block_rsv(fs_info);
5919 }
5920
5921 static void release_global_block_rsv(struct btrfs_fs_info *fs_info)
5922 {
5923         block_rsv_release_bytes(fs_info, &fs_info->global_block_rsv, NULL,
5924                                 (u64)-1, NULL);
5925         WARN_ON(fs_info->trans_block_rsv.size > 0);
5926         WARN_ON(fs_info->trans_block_rsv.reserved > 0);
5927         WARN_ON(fs_info->chunk_block_rsv.size > 0);
5928         WARN_ON(fs_info->chunk_block_rsv.reserved > 0);
5929         WARN_ON(fs_info->delayed_block_rsv.size > 0);
5930         WARN_ON(fs_info->delayed_block_rsv.reserved > 0);
5931 }
5932
5933
5934 /*
5935  * To be called after all the new block groups attached to the transaction
5936  * handle have been created (btrfs_create_pending_block_groups()).
5937  */
5938 void btrfs_trans_release_chunk_metadata(struct btrfs_trans_handle *trans)
5939 {
5940         struct btrfs_fs_info *fs_info = trans->fs_info;
5941
5942         if (!trans->chunk_bytes_reserved)
5943                 return;
5944
5945         WARN_ON_ONCE(!list_empty(&trans->new_bgs));
5946
5947         block_rsv_release_bytes(fs_info, &fs_info->chunk_block_rsv, NULL,
5948                                 trans->chunk_bytes_reserved, NULL);
5949         trans->chunk_bytes_reserved = 0;
5950 }
5951
5952 /* Can only return 0 or -ENOSPC */
5953 int btrfs_orphan_reserve_metadata(struct btrfs_trans_handle *trans,
5954                                   struct btrfs_inode *inode)
5955 {
5956         struct btrfs_fs_info *fs_info = btrfs_sb(inode->vfs_inode.i_sb);
5957         struct btrfs_root *root = inode->root;
5958         /*
5959          * We always use trans->block_rsv here as we will have reserved space
5960          * for our orphan when starting the transaction, using get_block_rsv()
5961          * here will sometimes make us choose the wrong block rsv as we could be
5962          * doing a reloc inode for a non refcounted root.
5963          */
5964         struct btrfs_block_rsv *src_rsv = trans->block_rsv;
5965         struct btrfs_block_rsv *dst_rsv = root->orphan_block_rsv;
5966
5967         /*
5968          * We need to hold space in order to delete our orphan item once we've
5969          * added it, so this takes the reservation so we can release it later
5970          * when we are truly done with the orphan item.
5971          */
5972         u64 num_bytes = btrfs_calc_trans_metadata_size(fs_info, 1);
5973
5974         trace_btrfs_space_reservation(fs_info, "orphan", btrfs_ino(inode),
5975                         num_bytes, 1);
5976         return btrfs_block_rsv_migrate(src_rsv, dst_rsv, num_bytes, 1);
5977 }
5978
5979 void btrfs_orphan_release_metadata(struct btrfs_inode *inode)
5980 {
5981         struct btrfs_fs_info *fs_info = btrfs_sb(inode->vfs_inode.i_sb);
5982         struct btrfs_root *root = inode->root;
5983         u64 num_bytes = btrfs_calc_trans_metadata_size(fs_info, 1);
5984
5985         trace_btrfs_space_reservation(fs_info, "orphan", btrfs_ino(inode),
5986                         num_bytes, 0);
5987         btrfs_block_rsv_release(fs_info, root->orphan_block_rsv, num_bytes);
5988 }
5989
5990 /*
5991  * btrfs_subvolume_reserve_metadata() - reserve space for subvolume operation
5992  * root: the root of the parent directory
5993  * rsv: block reservation
5994  * items: the number of items that we need do reservation
5995  * qgroup_reserved: used to return the reserved size in qgroup
5996  *
5997  * This function is used to reserve the space for snapshot/subvolume
5998  * creation and deletion. Those operations are different with the
5999  * common file/directory operations, they change two fs/file trees
6000  * and root tree, the number of items that the qgroup reserves is
6001  * different with the free space reservation. So we can not use
6002  * the space reservation mechanism in start_transaction().
6003  */
6004 int btrfs_subvolume_reserve_metadata(struct btrfs_root *root,
6005                                      struct btrfs_block_rsv *rsv,
6006                                      int items,
6007                                      u64 *qgroup_reserved,
6008                                      bool use_global_rsv)
6009 {
6010         u64 num_bytes;
6011         int ret;
6012         struct btrfs_fs_info *fs_info = root->fs_info;
6013         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
6014
6015         if (test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags)) {
6016                 /* One for parent inode, two for dir entries */
6017                 num_bytes = 3 * fs_info->nodesize;
6018                 ret = btrfs_qgroup_reserve_meta_prealloc(root, num_bytes, true);
6019                 if (ret)
6020                         return ret;
6021         } else {
6022                 num_bytes = 0;
6023         }
6024
6025         *qgroup_reserved = num_bytes;
6026
6027         num_bytes = btrfs_calc_trans_metadata_size(fs_info, items);
6028         rsv->space_info = __find_space_info(fs_info,
6029                                             BTRFS_BLOCK_GROUP_METADATA);
6030         ret = btrfs_block_rsv_add(root, rsv, num_bytes,
6031                                   BTRFS_RESERVE_FLUSH_ALL);
6032
6033         if (ret == -ENOSPC && use_global_rsv)
6034                 ret = btrfs_block_rsv_migrate(global_rsv, rsv, num_bytes, 1);
6035
6036         if (ret && *qgroup_reserved)
6037                 btrfs_qgroup_free_meta_prealloc(root, *qgroup_reserved);
6038
6039         return ret;
6040 }
6041
6042 void btrfs_subvolume_release_metadata(struct btrfs_fs_info *fs_info,
6043                                       struct btrfs_block_rsv *rsv)
6044 {
6045         btrfs_block_rsv_release(fs_info, rsv, (u64)-1);
6046 }
6047
6048 static void btrfs_calculate_inode_block_rsv_size(struct btrfs_fs_info *fs_info,
6049                                                  struct btrfs_inode *inode)
6050 {
6051         struct btrfs_block_rsv *block_rsv = &inode->block_rsv;
6052         u64 reserve_size = 0;
6053         u64 qgroup_rsv_size = 0;
6054         u64 csum_leaves;
6055         unsigned outstanding_extents;
6056
6057         lockdep_assert_held(&inode->lock);
6058         outstanding_extents = inode->outstanding_extents;
6059         if (outstanding_extents)
6060                 reserve_size = btrfs_calc_trans_metadata_size(fs_info,
6061                                                 outstanding_extents + 1);
6062         csum_leaves = btrfs_csum_bytes_to_leaves(fs_info,
6063                                                  inode->csum_bytes);
6064         reserve_size += btrfs_calc_trans_metadata_size(fs_info,
6065                                                        csum_leaves);
6066         /*
6067          * For qgroup rsv, the calculation is very simple:
6068          * account one nodesize for each outstanding extent
6069          *
6070          * This is overestimating in most cases.
6071          */
6072         qgroup_rsv_size = outstanding_extents * fs_info->nodesize;
6073
6074         spin_lock(&block_rsv->lock);
6075         block_rsv->size = reserve_size;
6076         block_rsv->qgroup_rsv_size = qgroup_rsv_size;
6077         spin_unlock(&block_rsv->lock);
6078 }
6079
6080 int btrfs_delalloc_reserve_metadata(struct btrfs_inode *inode, u64 num_bytes)
6081 {
6082         struct btrfs_fs_info *fs_info = btrfs_sb(inode->vfs_inode.i_sb);
6083         unsigned nr_extents;
6084         enum btrfs_reserve_flush_enum flush = BTRFS_RESERVE_FLUSH_ALL;
6085         int ret = 0;
6086         bool delalloc_lock = true;
6087
6088         /* If we are a free space inode we need to not flush since we will be in
6089          * the middle of a transaction commit.  We also don't need the delalloc
6090          * mutex since we won't race with anybody.  We need this mostly to make
6091          * lockdep shut its filthy mouth.
6092          *
6093          * If we have a transaction open (can happen if we call truncate_block
6094          * from truncate), then we need FLUSH_LIMIT so we don't deadlock.
6095          */
6096         if (btrfs_is_free_space_inode(inode)) {
6097                 flush = BTRFS_RESERVE_NO_FLUSH;
6098                 delalloc_lock = false;
6099         } else {
6100                 if (current->journal_info)
6101                         flush = BTRFS_RESERVE_FLUSH_LIMIT;
6102
6103                 if (btrfs_transaction_in_commit(fs_info))
6104                         schedule_timeout(1);
6105         }
6106
6107         if (delalloc_lock)
6108                 mutex_lock(&inode->delalloc_mutex);
6109
6110         num_bytes = ALIGN(num_bytes, fs_info->sectorsize);
6111
6112         /* Add our new extents and calculate the new rsv size. */
6113         spin_lock(&inode->lock);
6114         nr_extents = count_max_extents(num_bytes);
6115         btrfs_mod_outstanding_extents(inode, nr_extents);
6116         inode->csum_bytes += num_bytes;
6117         btrfs_calculate_inode_block_rsv_size(fs_info, inode);
6118         spin_unlock(&inode->lock);
6119
6120         ret = btrfs_inode_rsv_refill(inode, flush);
6121         if (unlikely(ret))
6122                 goto out_fail;
6123
6124         if (delalloc_lock)
6125                 mutex_unlock(&inode->delalloc_mutex);
6126         return 0;
6127
6128 out_fail:
6129         spin_lock(&inode->lock);
6130         nr_extents = count_max_extents(num_bytes);
6131         btrfs_mod_outstanding_extents(inode, -nr_extents);
6132         inode->csum_bytes -= num_bytes;
6133         btrfs_calculate_inode_block_rsv_size(fs_info, inode);
6134         spin_unlock(&inode->lock);
6135
6136         btrfs_inode_rsv_release(inode, true);
6137         if (delalloc_lock)
6138                 mutex_unlock(&inode->delalloc_mutex);
6139         return ret;
6140 }
6141
6142 /**
6143  * btrfs_delalloc_release_metadata - release a metadata reservation for an inode
6144  * @inode: the inode to release the reservation for.
6145  * @num_bytes: the number of bytes we are releasing.
6146  * @qgroup_free: free qgroup reservation or convert it to per-trans reservation
6147  *
6148  * This will release the metadata reservation for an inode.  This can be called
6149  * once we complete IO for a given set of bytes to release their metadata
6150  * reservations, or on error for the same reason.
6151  */
6152 void btrfs_delalloc_release_metadata(struct btrfs_inode *inode, u64 num_bytes,
6153                                      bool qgroup_free)
6154 {
6155         struct btrfs_fs_info *fs_info = btrfs_sb(inode->vfs_inode.i_sb);
6156
6157         num_bytes = ALIGN(num_bytes, fs_info->sectorsize);
6158         spin_lock(&inode->lock);
6159         inode->csum_bytes -= num_bytes;
6160         btrfs_calculate_inode_block_rsv_size(fs_info, inode);
6161         spin_unlock(&inode->lock);
6162
6163         if (btrfs_is_testing(fs_info))
6164                 return;
6165
6166         btrfs_inode_rsv_release(inode, qgroup_free);
6167 }
6168
6169 /**
6170  * btrfs_delalloc_release_extents - release our outstanding_extents
6171  * @inode: the inode to balance the reservation for.
6172  * @num_bytes: the number of bytes we originally reserved with
6173  * @qgroup_free: do we need to free qgroup meta reservation or convert them.
6174  *
6175  * When we reserve space we increase outstanding_extents for the extents we may
6176  * add.  Once we've set the range as delalloc or created our ordered extents we
6177  * have outstanding_extents to track the real usage, so we use this to free our
6178  * temporarily tracked outstanding_extents.  This _must_ be used in conjunction
6179  * with btrfs_delalloc_reserve_metadata.
6180  */
6181 void btrfs_delalloc_release_extents(struct btrfs_inode *inode, u64 num_bytes,
6182                                     bool qgroup_free)
6183 {
6184         struct btrfs_fs_info *fs_info = btrfs_sb(inode->vfs_inode.i_sb);
6185         unsigned num_extents;
6186
6187         spin_lock(&inode->lock);
6188         num_extents = count_max_extents(num_bytes);
6189         btrfs_mod_outstanding_extents(inode, -num_extents);
6190         btrfs_calculate_inode_block_rsv_size(fs_info, inode);
6191         spin_unlock(&inode->lock);
6192
6193         if (btrfs_is_testing(fs_info))
6194                 return;
6195
6196         btrfs_inode_rsv_release(inode, qgroup_free);
6197 }
6198
6199 /**
6200  * btrfs_delalloc_reserve_space - reserve data and metadata space for
6201  * delalloc
6202  * @inode: inode we're writing to
6203  * @start: start range we are writing to
6204  * @len: how long the range we are writing to
6205  * @reserved: mandatory parameter, record actually reserved qgroup ranges of
6206  *            current reservation.
6207  *
6208  * This will do the following things
6209  *
6210  * o reserve space in data space info for num bytes
6211  *   and reserve precious corresponding qgroup space
6212  *   (Done in check_data_free_space)
6213  *
6214  * o reserve space for metadata space, based on the number of outstanding
6215  *   extents and how much csums will be needed
6216  *   also reserve metadata space in a per root over-reserve method.
6217  * o add to the inodes->delalloc_bytes
6218  * o add it to the fs_info's delalloc inodes list.
6219  *   (Above 3 all done in delalloc_reserve_metadata)
6220  *
6221  * Return 0 for success
6222  * Return <0 for error(-ENOSPC or -EQUOT)
6223  */
6224 int btrfs_delalloc_reserve_space(struct inode *inode,
6225                         struct extent_changeset **reserved, u64 start, u64 len)
6226 {
6227         int ret;
6228
6229         ret = btrfs_check_data_free_space(inode, reserved, start, len);
6230         if (ret < 0)
6231                 return ret;
6232         ret = btrfs_delalloc_reserve_metadata(BTRFS_I(inode), len);
6233         if (ret < 0)
6234                 btrfs_free_reserved_data_space(inode, *reserved, start, len);
6235         return ret;
6236 }
6237
6238 /**
6239  * btrfs_delalloc_release_space - release data and metadata space for delalloc
6240  * @inode: inode we're releasing space for
6241  * @start: start position of the space already reserved
6242  * @len: the len of the space already reserved
6243  * @release_bytes: the len of the space we consumed or didn't use
6244  *
6245  * This function will release the metadata space that was not used and will
6246  * decrement ->delalloc_bytes and remove it from the fs_info delalloc_inodes
6247  * list if there are no delalloc bytes left.
6248  * Also it will handle the qgroup reserved space.
6249  */
6250 void btrfs_delalloc_release_space(struct inode *inode,
6251                                   struct extent_changeset *reserved,
6252                                   u64 start, u64 len, bool qgroup_free)
6253 {
6254         btrfs_delalloc_release_metadata(BTRFS_I(inode), len, qgroup_free);
6255         btrfs_free_reserved_data_space(inode, reserved, start, len);
6256 }
6257
6258 static int update_block_group(struct btrfs_trans_handle *trans,
6259                               struct btrfs_fs_info *info, u64 bytenr,
6260                               u64 num_bytes, int alloc)
6261 {
6262         struct btrfs_block_group_cache *cache = NULL;
6263         u64 total = num_bytes;
6264         u64 old_val;
6265         u64 byte_in_group;
6266         int factor;
6267
6268         /* block accounting for super block */
6269         spin_lock(&info->delalloc_root_lock);
6270         old_val = btrfs_super_bytes_used(info->super_copy);
6271         if (alloc)
6272                 old_val += num_bytes;
6273         else
6274                 old_val -= num_bytes;
6275         btrfs_set_super_bytes_used(info->super_copy, old_val);
6276         spin_unlock(&info->delalloc_root_lock);
6277
6278         while (total) {
6279                 cache = btrfs_lookup_block_group(info, bytenr);
6280                 if (!cache)
6281                         return -ENOENT;
6282                 if (cache->flags & (BTRFS_BLOCK_GROUP_DUP |
6283                                     BTRFS_BLOCK_GROUP_RAID1 |
6284                                     BTRFS_BLOCK_GROUP_RAID10))
6285                         factor = 2;
6286                 else
6287                         factor = 1;
6288                 /*
6289                  * If this block group has free space cache written out, we
6290                  * need to make sure to load it if we are removing space.  This
6291                  * is because we need the unpinning stage to actually add the
6292                  * space back to the block group, otherwise we will leak space.
6293                  */
6294                 if (!alloc && cache->cached == BTRFS_CACHE_NO)
6295                         cache_block_group(cache, 1);
6296
6297                 byte_in_group = bytenr - cache->key.objectid;
6298                 WARN_ON(byte_in_group > cache->key.offset);
6299
6300                 spin_lock(&cache->space_info->lock);
6301                 spin_lock(&cache->lock);
6302
6303                 if (btrfs_test_opt(info, SPACE_CACHE) &&
6304                     cache->disk_cache_state < BTRFS_DC_CLEAR)
6305                         cache->disk_cache_state = BTRFS_DC_CLEAR;
6306
6307                 old_val = btrfs_block_group_used(&cache->item);
6308                 num_bytes = min(total, cache->key.offset - byte_in_group);
6309                 if (alloc) {
6310                         old_val += num_bytes;
6311                         btrfs_set_block_group_used(&cache->item, old_val);
6312                         cache->reserved -= num_bytes;
6313                         cache->space_info->bytes_reserved -= num_bytes;
6314                         cache->space_info->bytes_used += num_bytes;
6315                         cache->space_info->disk_used += num_bytes * factor;
6316                         spin_unlock(&cache->lock);
6317                         spin_unlock(&cache->space_info->lock);
6318                 } else {
6319                         old_val -= num_bytes;
6320                         btrfs_set_block_group_used(&cache->item, old_val);
6321                         cache->pinned += num_bytes;
6322                         cache->space_info->bytes_pinned += num_bytes;
6323                         cache->space_info->bytes_used -= num_bytes;
6324                         cache->space_info->disk_used -= num_bytes * factor;
6325                         spin_unlock(&cache->lock);
6326                         spin_unlock(&cache->space_info->lock);
6327
6328                         trace_btrfs_space_reservation(info, "pinned",
6329                                                       cache->space_info->flags,
6330                                                       num_bytes, 1);
6331                         percpu_counter_add(&cache->space_info->total_bytes_pinned,
6332                                            num_bytes);
6333                         set_extent_dirty(info->pinned_extents,
6334                                          bytenr, bytenr + num_bytes - 1,
6335                                          GFP_NOFS | __GFP_NOFAIL);
6336                 }
6337
6338                 spin_lock(&trans->transaction->dirty_bgs_lock);
6339                 if (list_empty(&cache->dirty_list)) {
6340                         list_add_tail(&cache->dirty_list,
6341                                       &trans->transaction->dirty_bgs);
6342                                 trans->transaction->num_dirty_bgs++;
6343                         btrfs_get_block_group(cache);
6344                 }
6345                 spin_unlock(&trans->transaction->dirty_bgs_lock);
6346
6347                 /*
6348                  * No longer have used bytes in this block group, queue it for
6349                  * deletion. We do this after adding the block group to the
6350                  * dirty list to avoid races between cleaner kthread and space
6351                  * cache writeout.
6352                  */
6353                 if (!alloc && old_val == 0) {
6354                         spin_lock(&info->unused_bgs_lock);
6355                         if (list_empty(&cache->bg_list)) {
6356                                 btrfs_get_block_group(cache);
6357                                 list_add_tail(&cache->bg_list,
6358                                               &info->unused_bgs);
6359                         }
6360                         spin_unlock(&info->unused_bgs_lock);
6361                 }
6362
6363                 btrfs_put_block_group(cache);
6364                 total -= num_bytes;
6365                 bytenr += num_bytes;
6366         }
6367         return 0;
6368 }
6369
6370 static u64 first_logical_byte(struct btrfs_fs_info *fs_info, u64 search_start)
6371 {
6372         struct btrfs_block_group_cache *cache;
6373         u64 bytenr;
6374
6375         spin_lock(&fs_info->block_group_cache_lock);
6376         bytenr = fs_info->first_logical_byte;
6377         spin_unlock(&fs_info->block_group_cache_lock);
6378
6379         if (bytenr < (u64)-1)
6380                 return bytenr;
6381
6382         cache = btrfs_lookup_first_block_group(fs_info, search_start);
6383         if (!cache)
6384                 return 0;
6385
6386         bytenr = cache->key.objectid;
6387         btrfs_put_block_group(cache);
6388
6389         return bytenr;
6390 }
6391
6392 static int pin_down_extent(struct btrfs_fs_info *fs_info,
6393                            struct btrfs_block_group_cache *cache,
6394                            u64 bytenr, u64 num_bytes, int reserved)
6395 {
6396         spin_lock(&cache->space_info->lock);
6397         spin_lock(&cache->lock);
6398         cache->pinned += num_bytes;
6399         cache->space_info->bytes_pinned += num_bytes;
6400         if (reserved) {
6401                 cache->reserved -= num_bytes;
6402                 cache->space_info->bytes_reserved -= num_bytes;
6403         }
6404         spin_unlock(&cache->lock);
6405         spin_unlock(&cache->space_info->lock);
6406
6407         trace_btrfs_space_reservation(fs_info, "pinned",
6408                                       cache->space_info->flags, num_bytes, 1);
6409         percpu_counter_add(&cache->space_info->total_bytes_pinned, num_bytes);
6410         set_extent_dirty(fs_info->pinned_extents, bytenr,
6411                          bytenr + num_bytes - 1, GFP_NOFS | __GFP_NOFAIL);
6412         return 0;
6413 }
6414
6415 /*
6416  * this function must be called within transaction
6417  */
6418 int btrfs_pin_extent(struct btrfs_fs_info *fs_info,
6419                      u64 bytenr, u64 num_bytes, int reserved)
6420 {
6421         struct btrfs_block_group_cache *cache;
6422
6423         cache = btrfs_lookup_block_group(fs_info, bytenr);
6424         BUG_ON(!cache); /* Logic error */
6425
6426         pin_down_extent(fs_info, cache, bytenr, num_bytes, reserved);
6427
6428         btrfs_put_block_group(cache);
6429         return 0;
6430 }
6431
6432 /*
6433  * this function must be called within transaction
6434  */
6435 int btrfs_pin_extent_for_log_replay(struct btrfs_fs_info *fs_info,
6436                                     u64 bytenr, u64 num_bytes)
6437 {
6438         struct btrfs_block_group_cache *cache;
6439         int ret;
6440
6441         cache = btrfs_lookup_block_group(fs_info, bytenr);
6442         if (!cache)
6443                 return -EINVAL;
6444
6445         /*
6446          * pull in the free space cache (if any) so that our pin
6447          * removes the free space from the cache.  We have load_only set
6448          * to one because the slow code to read in the free extents does check
6449          * the pinned extents.
6450          */
6451         cache_block_group(cache, 1);
6452
6453         pin_down_extent(fs_info, cache, bytenr, num_bytes, 0);
6454
6455         /* remove us from the free space cache (if we're there at all) */
6456         ret = btrfs_remove_free_space(cache, bytenr, num_bytes);
6457         btrfs_put_block_group(cache);
6458         return ret;
6459 }
6460
6461 static int __exclude_logged_extent(struct btrfs_fs_info *fs_info,
6462                                    u64 start, u64 num_bytes)
6463 {
6464         int ret;
6465         struct btrfs_block_group_cache *block_group;
6466         struct btrfs_caching_control *caching_ctl;
6467
6468         block_group = btrfs_lookup_block_group(fs_info, start);
6469         if (!block_group)
6470                 return -EINVAL;
6471
6472         cache_block_group(block_group, 0);
6473         caching_ctl = get_caching_control(block_group);
6474
6475         if (!caching_ctl) {
6476                 /* Logic error */
6477                 BUG_ON(!block_group_cache_done(block_group));
6478                 ret = btrfs_remove_free_space(block_group, start, num_bytes);
6479         } else {
6480                 mutex_lock(&caching_ctl->mutex);
6481
6482                 if (start >= caching_ctl->progress) {
6483                         ret = add_excluded_extent(fs_info, start, num_bytes);
6484                 } else if (start + num_bytes <= caching_ctl->progress) {
6485                         ret = btrfs_remove_free_space(block_group,
6486                                                       start, num_bytes);
6487                 } else {
6488                         num_bytes = caching_ctl->progress - start;
6489                         ret = btrfs_remove_free_space(block_group,
6490                                                       start, num_bytes);
6491                         if (ret)
6492                                 goto out_lock;
6493
6494                         num_bytes = (start + num_bytes) -
6495                                 caching_ctl->progress;
6496                         start = caching_ctl->progress;
6497                         ret = add_excluded_extent(fs_info, start, num_bytes);
6498                 }
6499 out_lock:
6500                 mutex_unlock(&caching_ctl->mutex);
6501                 put_caching_control(caching_ctl);
6502         }
6503         btrfs_put_block_group(block_group);
6504         return ret;
6505 }
6506
6507 int btrfs_exclude_logged_extents(struct btrfs_fs_info *fs_info,
6508                                  struct extent_buffer *eb)
6509 {
6510         struct btrfs_file_extent_item *item;
6511         struct btrfs_key key;
6512         int found_type;
6513         int i;
6514
6515         if (!btrfs_fs_incompat(fs_info, MIXED_GROUPS))
6516                 return 0;
6517
6518         for (i = 0; i < btrfs_header_nritems(eb); i++) {
6519                 btrfs_item_key_to_cpu(eb, &key, i);
6520                 if (key.type != BTRFS_EXTENT_DATA_KEY)
6521                         continue;
6522                 item = btrfs_item_ptr(eb, i, struct btrfs_file_extent_item);
6523                 found_type = btrfs_file_extent_type(eb, item);
6524                 if (found_type == BTRFS_FILE_EXTENT_INLINE)
6525                         continue;
6526                 if (btrfs_file_extent_disk_bytenr(eb, item) == 0)
6527                         continue;
6528                 key.objectid = btrfs_file_extent_disk_bytenr(eb, item);
6529                 key.offset = btrfs_file_extent_disk_num_bytes(eb, item);
6530                 __exclude_logged_extent(fs_info, key.objectid, key.offset);
6531         }
6532
6533         return 0;
6534 }
6535
6536 static void
6537 btrfs_inc_block_group_reservations(struct btrfs_block_group_cache *bg)
6538 {
6539         atomic_inc(&bg->reservations);
6540 }
6541
6542 void btrfs_dec_block_group_reservations(struct btrfs_fs_info *fs_info,
6543                                         const u64 start)
6544 {
6545         struct btrfs_block_group_cache *bg;
6546
6547         bg = btrfs_lookup_block_group(fs_info, start);
6548         ASSERT(bg);
6549         if (atomic_dec_and_test(&bg->reservations))
6550                 wake_up_var(&bg->reservations);
6551         btrfs_put_block_group(bg);
6552 }
6553
6554 void btrfs_wait_block_group_reservations(struct btrfs_block_group_cache *bg)
6555 {
6556         struct btrfs_space_info *space_info = bg->space_info;
6557
6558         ASSERT(bg->ro);
6559
6560         if (!(bg->flags & BTRFS_BLOCK_GROUP_DATA))
6561                 return;
6562
6563         /*
6564          * Our block group is read only but before we set it to read only,
6565          * some task might have had allocated an extent from it already, but it
6566          * has not yet created a respective ordered extent (and added it to a
6567          * root's list of ordered extents).
6568          * Therefore wait for any task currently allocating extents, since the
6569          * block group's reservations counter is incremented while a read lock
6570          * on the groups' semaphore is held and decremented after releasing
6571          * the read access on that semaphore and creating the ordered extent.
6572          */
6573         down_write(&space_info->groups_sem);
6574         up_write(&space_info->groups_sem);
6575
6576         wait_var_event(&bg->reservations, !atomic_read(&bg->reservations));
6577 }
6578
6579 /**
6580  * btrfs_add_reserved_bytes - update the block_group and space info counters
6581  * @cache:      The cache we are manipulating
6582  * @ram_bytes:  The number of bytes of file content, and will be same to
6583  *              @num_bytes except for the compress path.
6584  * @num_bytes:  The number of bytes in question
6585  * @delalloc:   The blocks are allocated for the delalloc write
6586  *
6587  * This is called by the allocator when it reserves space. If this is a
6588  * reservation and the block group has become read only we cannot make the
6589  * reservation and return -EAGAIN, otherwise this function always succeeds.
6590  */
6591 static int btrfs_add_reserved_bytes(struct btrfs_block_group_cache *cache,
6592                                     u64 ram_bytes, u64 num_bytes, int delalloc)
6593 {
6594         struct btrfs_space_info *space_info = cache->space_info;
6595         int ret = 0;
6596
6597         spin_lock(&space_info->lock);
6598         spin_lock(&cache->lock);
6599         if (cache->ro) {
6600                 ret = -EAGAIN;
6601         } else {
6602                 cache->reserved += num_bytes;
6603                 space_info->bytes_reserved += num_bytes;
6604
6605                 trace_btrfs_space_reservation(cache->fs_info,
6606                                 "space_info", space_info->flags,
6607                                 ram_bytes, 0);
6608                 space_info->bytes_may_use -= ram_bytes;
6609                 if (delalloc)
6610                         cache->delalloc_bytes += num_bytes;
6611         }
6612         spin_unlock(&cache->lock);
6613         spin_unlock(&space_info->lock);
6614         return ret;
6615 }
6616
6617 /**
6618  * btrfs_free_reserved_bytes - update the block_group and space info counters
6619  * @cache:      The cache we are manipulating
6620  * @num_bytes:  The number of bytes in question
6621  * @delalloc:   The blocks are allocated for the delalloc write
6622  *
6623  * This is called by somebody who is freeing space that was never actually used
6624  * on disk.  For example if you reserve some space for a new leaf in transaction
6625  * A and before transaction A commits you free that leaf, you call this with
6626  * reserve set to 0 in order to clear the reservation.
6627  */
6628
6629 static int btrfs_free_reserved_bytes(struct btrfs_block_group_cache *cache,
6630                                      u64 num_bytes, int delalloc)
6631 {
6632         struct btrfs_space_info *space_info = cache->space_info;
6633         int ret = 0;
6634
6635         spin_lock(&space_info->lock);
6636         spin_lock(&cache->lock);
6637         if (cache->ro)
6638                 space_info->bytes_readonly += num_bytes;
6639         cache->reserved -= num_bytes;
6640         space_info->bytes_reserved -= num_bytes;
6641
6642         if (delalloc)
6643                 cache->delalloc_bytes -= num_bytes;
6644         spin_unlock(&cache->lock);
6645         spin_unlock(&space_info->lock);
6646         return ret;
6647 }
6648 void btrfs_prepare_extent_commit(struct btrfs_fs_info *fs_info)
6649 {
6650         struct btrfs_caching_control *next;
6651         struct btrfs_caching_control *caching_ctl;
6652         struct btrfs_block_group_cache *cache;
6653
6654         down_write(&fs_info->commit_root_sem);
6655
6656         list_for_each_entry_safe(caching_ctl, next,
6657                                  &fs_info->caching_block_groups, list) {
6658                 cache = caching_ctl->block_group;
6659                 if (block_group_cache_done(cache)) {
6660                         cache->last_byte_to_unpin = (u64)-1;
6661                         list_del_init(&caching_ctl->list);
6662                         put_caching_control(caching_ctl);
6663                 } else {
6664                         cache->last_byte_to_unpin = caching_ctl->progress;
6665                 }
6666         }
6667
6668         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
6669                 fs_info->pinned_extents = &fs_info->freed_extents[1];
6670         else
6671                 fs_info->pinned_extents = &fs_info->freed_extents[0];
6672
6673         up_write(&fs_info->commit_root_sem);
6674
6675         update_global_block_rsv(fs_info);
6676 }
6677
6678 /*
6679  * Returns the free cluster for the given space info and sets empty_cluster to
6680  * what it should be based on the mount options.
6681  */
6682 static struct btrfs_free_cluster *
6683 fetch_cluster_info(struct btrfs_fs_info *fs_info,
6684                    struct btrfs_space_info *space_info, u64 *empty_cluster)
6685 {
6686         struct btrfs_free_cluster *ret = NULL;
6687
6688         *empty_cluster = 0;
6689         if (btrfs_mixed_space_info(space_info))
6690                 return ret;
6691
6692         if (space_info->flags & BTRFS_BLOCK_GROUP_METADATA) {
6693                 ret = &fs_info->meta_alloc_cluster;
6694                 if (btrfs_test_opt(fs_info, SSD))
6695                         *empty_cluster = SZ_2M;
6696                 else
6697                         *empty_cluster = SZ_64K;
6698         } else if ((space_info->flags & BTRFS_BLOCK_GROUP_DATA) &&
6699                    btrfs_test_opt(fs_info, SSD_SPREAD)) {
6700                 *empty_cluster = SZ_2M;
6701                 ret = &fs_info->data_alloc_cluster;
6702         }
6703
6704         return ret;
6705 }
6706
6707 static int unpin_extent_range(struct btrfs_fs_info *fs_info,
6708                               u64 start, u64 end,
6709                               const bool return_free_space)
6710 {
6711         struct btrfs_block_group_cache *cache = NULL;
6712         struct btrfs_space_info *space_info;
6713         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
6714         struct btrfs_free_cluster *cluster = NULL;
6715         u64 len;
6716         u64 total_unpinned = 0;
6717         u64 empty_cluster = 0;
6718         bool readonly;
6719
6720         while (start <= end) {
6721                 readonly = false;
6722                 if (!cache ||
6723                     start >= cache->key.objectid + cache->key.offset) {
6724                         if (cache)
6725                                 btrfs_put_block_group(cache);
6726                         total_unpinned = 0;
6727                         cache = btrfs_lookup_block_group(fs_info, start);
6728                         BUG_ON(!cache); /* Logic error */
6729
6730                         cluster = fetch_cluster_info(fs_info,
6731                                                      cache->space_info,
6732                                                      &empty_cluster);
6733                         empty_cluster <<= 1;
6734                 }
6735
6736                 len = cache->key.objectid + cache->key.offset - start;
6737                 len = min(len, end + 1 - start);
6738
6739                 if (start < cache->last_byte_to_unpin) {
6740                         len = min(len, cache->last_byte_to_unpin - start);
6741                         if (return_free_space)
6742                                 btrfs_add_free_space(cache, start, len);
6743                 }
6744
6745                 start += len;
6746                 total_unpinned += len;
6747                 space_info = cache->space_info;
6748
6749                 /*
6750                  * If this space cluster has been marked as fragmented and we've
6751                  * unpinned enough in this block group to potentially allow a
6752                  * cluster to be created inside of it go ahead and clear the
6753                  * fragmented check.
6754                  */
6755                 if (cluster && cluster->fragmented &&
6756                     total_unpinned > empty_cluster) {
6757                         spin_lock(&cluster->lock);
6758                         cluster->fragmented = 0;
6759                         spin_unlock(&cluster->lock);
6760                 }
6761
6762                 spin_lock(&space_info->lock);
6763                 spin_lock(&cache->lock);
6764                 cache->pinned -= len;
6765                 space_info->bytes_pinned -= len;
6766
6767                 trace_btrfs_space_reservation(fs_info, "pinned",
6768                                               space_info->flags, len, 0);
6769                 space_info->max_extent_size = 0;
6770                 percpu_counter_add(&space_info->total_bytes_pinned, -len);
6771                 if (cache->ro) {
6772                         space_info->bytes_readonly += len;
6773                         readonly = true;
6774                 }
6775                 spin_unlock(&cache->lock);
6776                 if (!readonly && return_free_space &&
6777                     global_rsv->space_info == space_info) {
6778                         u64 to_add = len;
6779
6780                         spin_lock(&global_rsv->lock);
6781                         if (!global_rsv->full) {
6782                                 to_add = min(len, global_rsv->size -
6783                                              global_rsv->reserved);
6784                                 global_rsv->reserved += to_add;
6785                                 space_info->bytes_may_use += to_add;
6786                                 if (global_rsv->reserved >= global_rsv->size)
6787                                         global_rsv->full = 1;
6788                                 trace_btrfs_space_reservation(fs_info,
6789                                                               "space_info",
6790                                                               space_info->flags,
6791                                                               to_add, 1);
6792                                 len -= to_add;
6793                         }
6794                         spin_unlock(&global_rsv->lock);
6795                         /* Add to any tickets we may have */
6796                         if (len)
6797                                 space_info_add_new_bytes(fs_info, space_info,
6798                                                          len);
6799                 }
6800                 spin_unlock(&space_info->lock);
6801         }
6802
6803         if (cache)
6804                 btrfs_put_block_group(cache);
6805         return 0;
6806 }
6807
6808 int btrfs_finish_extent_commit(struct btrfs_trans_handle *trans)
6809 {
6810         struct btrfs_fs_info *fs_info = trans->fs_info;
6811         struct btrfs_block_group_cache *block_group, *tmp;
6812         struct list_head *deleted_bgs;
6813         struct extent_io_tree *unpin;
6814         u64 start;
6815         u64 end;
6816         int ret;
6817
6818         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
6819                 unpin = &fs_info->freed_extents[1];
6820         else
6821                 unpin = &fs_info->freed_extents[0];
6822
6823         while (!trans->aborted) {
6824                 mutex_lock(&fs_info->unused_bg_unpin_mutex);
6825                 ret = find_first_extent_bit(unpin, 0, &start, &end,
6826                                             EXTENT_DIRTY, NULL);
6827                 if (ret) {
6828                         mutex_unlock(&fs_info->unused_bg_unpin_mutex);
6829                         break;
6830                 }
6831
6832                 if (btrfs_test_opt(fs_info, DISCARD))
6833                         ret = btrfs_discard_extent(fs_info, start,
6834                                                    end + 1 - start, NULL);
6835
6836                 clear_extent_dirty(unpin, start, end);
6837                 unpin_extent_range(fs_info, start, end, true);
6838                 mutex_unlock(&fs_info->unused_bg_unpin_mutex);
6839                 cond_resched();
6840         }
6841
6842         /*
6843          * Transaction is finished.  We don't need the lock anymore.  We
6844          * do need to clean up the block groups in case of a transaction
6845          * abort.
6846          */
6847         deleted_bgs = &trans->transaction->deleted_bgs;
6848         list_for_each_entry_safe(block_group, tmp, deleted_bgs, bg_list) {
6849                 u64 trimmed = 0;
6850
6851                 ret = -EROFS;
6852                 if (!trans->aborted)
6853                         ret = btrfs_discard_extent(fs_info,
6854                                                    block_group->key.objectid,
6855                                                    block_group->key.offset,
6856                                                    &trimmed);
6857
6858                 list_del_init(&block_group->bg_list);
6859                 btrfs_put_block_group_trimming(block_group);
6860                 btrfs_put_block_group(block_group);
6861
6862                 if (ret) {
6863                         const char *errstr = btrfs_decode_error(ret);
6864                         btrfs_warn(fs_info,
6865                            "discard failed while removing blockgroup: errno=%d %s",
6866                                    ret, errstr);
6867                 }
6868         }
6869
6870         return 0;
6871 }
6872
6873 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
6874                                 struct btrfs_fs_info *info,
6875                                 struct btrfs_delayed_ref_node *node, u64 parent,
6876                                 u64 root_objectid, u64 owner_objectid,
6877                                 u64 owner_offset, int refs_to_drop,
6878                                 struct btrfs_delayed_extent_op *extent_op)
6879 {
6880         struct btrfs_key key;
6881         struct btrfs_path *path;
6882         struct btrfs_root *extent_root = info->extent_root;
6883         struct extent_buffer *leaf;
6884         struct btrfs_extent_item *ei;
6885         struct btrfs_extent_inline_ref *iref;
6886         int ret;
6887         int is_data;
6888         int extent_slot = 0;
6889         int found_extent = 0;
6890         int num_to_del = 1;
6891         u32 item_size;
6892         u64 refs;
6893         u64 bytenr = node->bytenr;
6894         u64 num_bytes = node->num_bytes;
6895         int last_ref = 0;
6896         bool skinny_metadata = btrfs_fs_incompat(info, SKINNY_METADATA);
6897
6898         path = btrfs_alloc_path();
6899         if (!path)
6900                 return -ENOMEM;
6901
6902         path->reada = READA_FORWARD;
6903         path->leave_spinning = 1;
6904
6905         is_data = owner_objectid >= BTRFS_FIRST_FREE_OBJECTID;
6906         BUG_ON(!is_data && refs_to_drop != 1);
6907
6908         if (is_data)
6909                 skinny_metadata = false;
6910
6911         ret = lookup_extent_backref(trans, info, path, &iref,
6912                                     bytenr, num_bytes, parent,
6913                                     root_objectid, owner_objectid,
6914                                     owner_offset);
6915         if (ret == 0) {
6916                 extent_slot = path->slots[0];
6917                 while (extent_slot >= 0) {
6918                         btrfs_item_key_to_cpu(path->nodes[0], &key,
6919                                               extent_slot);
6920                         if (key.objectid != bytenr)
6921                                 break;
6922                         if (key.type == BTRFS_EXTENT_ITEM_KEY &&
6923                             key.offset == num_bytes) {
6924                                 found_extent = 1;
6925                                 break;
6926                         }
6927                         if (key.type == BTRFS_METADATA_ITEM_KEY &&
6928                             key.offset == owner_objectid) {
6929                                 found_extent = 1;
6930                                 break;
6931                         }
6932                         if (path->slots[0] - extent_slot > 5)
6933                                 break;
6934                         extent_slot--;
6935                 }
6936 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
6937                 item_size = btrfs_item_size_nr(path->nodes[0], extent_slot);
6938                 if (found_extent && item_size < sizeof(*ei))
6939                         found_extent = 0;
6940 #endif
6941                 if (!found_extent) {
6942                         BUG_ON(iref);
6943                         ret = remove_extent_backref(trans, info, path, NULL,
6944                                                     refs_to_drop,
6945                                                     is_data, &last_ref);
6946                         if (ret) {
6947                                 btrfs_abort_transaction(trans, ret);
6948                                 goto out;
6949                         }
6950                         btrfs_release_path(path);
6951                         path->leave_spinning = 1;
6952
6953                         key.objectid = bytenr;
6954                         key.type = BTRFS_EXTENT_ITEM_KEY;
6955                         key.offset = num_bytes;
6956
6957                         if (!is_data && skinny_metadata) {
6958                                 key.type = BTRFS_METADATA_ITEM_KEY;
6959                                 key.offset = owner_objectid;
6960                         }
6961
6962                         ret = btrfs_search_slot(trans, extent_root,
6963                                                 &key, path, -1, 1);
6964                         if (ret > 0 && skinny_metadata && path->slots[0]) {
6965                                 /*
6966                                  * Couldn't find our skinny metadata item,
6967                                  * see if we have ye olde extent item.
6968                                  */
6969                                 path->slots[0]--;
6970                                 btrfs_item_key_to_cpu(path->nodes[0], &key,
6971                                                       path->slots[0]);
6972                                 if (key.objectid == bytenr &&
6973                                     key.type == BTRFS_EXTENT_ITEM_KEY &&
6974                                     key.offset == num_bytes)
6975                                         ret = 0;
6976                         }
6977
6978                         if (ret > 0 && skinny_metadata) {
6979                                 skinny_metadata = false;
6980                                 key.objectid = bytenr;
6981                                 key.type = BTRFS_EXTENT_ITEM_KEY;
6982                                 key.offset = num_bytes;
6983                                 btrfs_release_path(path);
6984                                 ret = btrfs_search_slot(trans, extent_root,
6985                                                         &key, path, -1, 1);
6986                         }
6987
6988                         if (ret) {
6989                                 btrfs_err(info,
6990                                           "umm, got %d back from search, was looking for %llu",
6991                                           ret, bytenr);
6992                                 if (ret > 0)
6993                                         btrfs_print_leaf(path->nodes[0]);
6994                         }
6995                         if (ret < 0) {
6996                                 btrfs_abort_transaction(trans, ret);
6997                                 goto out;
6998                         }
6999                         extent_slot = path->slots[0];
7000                 }
7001         } else if (WARN_ON(ret == -ENOENT)) {
7002                 btrfs_print_leaf(path->nodes[0]);
7003                 btrfs_err(info,
7004                         "unable to find ref byte nr %llu parent %llu root %llu  owner %llu offset %llu",
7005                         bytenr, parent, root_objectid, owner_objectid,
7006                         owner_offset);
7007                 btrfs_abort_transaction(trans, ret);
7008                 goto out;
7009         } else {
7010                 btrfs_abort_transaction(trans, ret);
7011                 goto out;
7012         }
7013
7014         leaf = path->nodes[0];
7015         item_size = btrfs_item_size_nr(leaf, extent_slot);
7016 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
7017         if (item_size < sizeof(*ei)) {
7018                 BUG_ON(found_extent || extent_slot != path->slots[0]);
7019                 ret = convert_extent_item_v0(trans, info, path, owner_objectid,
7020                                              0);
7021                 if (ret < 0) {
7022                         btrfs_abort_transaction(trans, ret);
7023                         goto out;
7024                 }
7025
7026                 btrfs_release_path(path);
7027                 path->leave_spinning = 1;
7028
7029                 key.objectid = bytenr;
7030                 key.type = BTRFS_EXTENT_ITEM_KEY;
7031                 key.offset = num_bytes;
7032
7033                 ret = btrfs_search_slot(trans, extent_root, &key, path,
7034                                         -1, 1);
7035                 if (ret) {
7036                         btrfs_err(info,
7037                                   "umm, got %d back from search, was looking for %llu",
7038                                 ret, bytenr);
7039                         btrfs_print_leaf(path->nodes[0]);
7040                 }
7041                 if (ret < 0) {
7042                         btrfs_abort_transaction(trans, ret);
7043                         goto out;
7044                 }
7045
7046                 extent_slot = path->slots[0];
7047                 leaf = path->nodes[0];
7048                 item_size = btrfs_item_size_nr(leaf, extent_slot);
7049         }
7050 #endif
7051         BUG_ON(item_size < sizeof(*ei));
7052         ei = btrfs_item_ptr(leaf, extent_slot,
7053                             struct btrfs_extent_item);
7054         if (owner_objectid < BTRFS_FIRST_FREE_OBJECTID &&
7055             key.type == BTRFS_EXTENT_ITEM_KEY) {
7056                 struct btrfs_tree_block_info *bi;
7057                 BUG_ON(item_size < sizeof(*ei) + sizeof(*bi));
7058                 bi = (struct btrfs_tree_block_info *)(ei + 1);
7059                 WARN_ON(owner_objectid != btrfs_tree_block_level(leaf, bi));
7060         }
7061
7062         refs = btrfs_extent_refs(leaf, ei);
7063         if (refs < refs_to_drop) {
7064                 btrfs_err(info,
7065                           "trying to drop %d refs but we only have %Lu for bytenr %Lu",
7066                           refs_to_drop, refs, bytenr);
7067                 ret = -EINVAL;
7068                 btrfs_abort_transaction(trans, ret);
7069                 goto out;
7070         }
7071         refs -= refs_to_drop;
7072
7073         if (refs > 0) {
7074                 if (extent_op)
7075                         __run_delayed_extent_op(extent_op, leaf, ei);
7076                 /*
7077                  * In the case of inline back ref, reference count will
7078                  * be updated by remove_extent_backref
7079                  */
7080                 if (iref) {
7081                         BUG_ON(!found_extent);
7082                 } else {
7083                         btrfs_set_extent_refs(leaf, ei, refs);
7084                         btrfs_mark_buffer_dirty(leaf);
7085                 }
7086                 if (found_extent) {
7087                         ret = remove_extent_backref(trans, info, path,
7088                                                     iref, refs_to_drop,
7089                                                     is_data, &last_ref);
7090                         if (ret) {
7091                                 btrfs_abort_transaction(trans, ret);
7092                                 goto out;
7093                         }
7094                 }
7095         } else {
7096                 if (found_extent) {
7097                         BUG_ON(is_data && refs_to_drop !=
7098                                extent_data_ref_count(path, iref));
7099                         if (iref) {
7100                                 BUG_ON(path->slots[0] != extent_slot);
7101                         } else {
7102                                 BUG_ON(path->slots[0] != extent_slot + 1);
7103                                 path->slots[0] = extent_slot;
7104                                 num_to_del = 2;
7105                         }
7106                 }
7107
7108                 last_ref = 1;
7109                 ret = btrfs_del_items(trans, extent_root, path, path->slots[0],
7110                                       num_to_del);
7111                 if (ret) {
7112                         btrfs_abort_transaction(trans, ret);
7113                         goto out;
7114                 }
7115                 btrfs_release_path(path);
7116
7117                 if (is_data) {
7118                         ret = btrfs_del_csums(trans, info, bytenr, num_bytes);
7119                         if (ret) {
7120                                 btrfs_abort_transaction(trans, ret);
7121                                 goto out;
7122                         }
7123                 }
7124
7125                 ret = add_to_free_space_tree(trans, info, bytenr, num_bytes);
7126                 if (ret) {
7127                         btrfs_abort_transaction(trans, ret);
7128                         goto out;
7129                 }
7130
7131                 ret = update_block_group(trans, info, bytenr, num_bytes, 0);
7132                 if (ret) {
7133                         btrfs_abort_transaction(trans, ret);
7134                         goto out;
7135                 }
7136         }
7137         btrfs_release_path(path);
7138
7139 out:
7140         btrfs_free_path(path);
7141         return ret;
7142 }
7143
7144 /*
7145  * when we free an block, it is possible (and likely) that we free the last
7146  * delayed ref for that extent as well.  This searches the delayed ref tree for
7147  * a given extent, and if there are no other delayed refs to be processed, it
7148  * removes it from the tree.
7149  */
7150 static noinline int check_ref_cleanup(struct btrfs_trans_handle *trans,
7151                                       u64 bytenr)
7152 {
7153         struct btrfs_delayed_ref_head *head;
7154         struct btrfs_delayed_ref_root *delayed_refs;
7155         int ret = 0;
7156
7157         delayed_refs = &trans->transaction->delayed_refs;
7158         spin_lock(&delayed_refs->lock);
7159         head = btrfs_find_delayed_ref_head(delayed_refs, bytenr);
7160         if (!head)
7161                 goto out_delayed_unlock;
7162
7163         spin_lock(&head->lock);
7164         if (!RB_EMPTY_ROOT(&head->ref_tree))
7165                 goto out;
7166
7167         if (head->extent_op) {
7168                 if (!head->must_insert_reserved)
7169                         goto out;
7170                 btrfs_free_delayed_extent_op(head->extent_op);
7171                 head->extent_op = NULL;
7172         }
7173
7174         /*
7175          * waiting for the lock here would deadlock.  If someone else has it
7176          * locked they are already in the process of dropping it anyway
7177          */
7178         if (!mutex_trylock(&head->mutex))
7179                 goto out;
7180
7181         /*
7182          * at this point we have a head with no other entries.  Go
7183          * ahead and process it.
7184          */
7185         rb_erase(&head->href_node, &delayed_refs->href_root);
7186         RB_CLEAR_NODE(&head->href_node);
7187         atomic_dec(&delayed_refs->num_entries);
7188
7189         /*
7190          * we don't take a ref on the node because we're removing it from the
7191          * tree, so we just steal the ref the tree was holding.
7192          */
7193         delayed_refs->num_heads--;
7194         if (head->processing == 0)
7195                 delayed_refs->num_heads_ready--;
7196         head->processing = 0;
7197         spin_unlock(&head->lock);
7198         spin_unlock(&delayed_refs->lock);
7199
7200         BUG_ON(head->extent_op);
7201         if (head->must_insert_reserved)
7202                 ret = 1;
7203
7204         mutex_unlock(&head->mutex);
7205         btrfs_put_delayed_ref_head(head);
7206         return ret;
7207 out:
7208         spin_unlock(&head->lock);
7209
7210 out_delayed_unlock:
7211         spin_unlock(&delayed_refs->lock);
7212         return 0;
7213 }
7214
7215 void btrfs_free_tree_block(struct btrfs_trans_handle *trans,
7216                            struct btrfs_root *root,
7217                            struct extent_buffer *buf,
7218                            u64 parent, int last_ref)
7219 {
7220         struct btrfs_fs_info *fs_info = root->fs_info;
7221         int pin = 1;
7222         int ret;
7223
7224         if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
7225                 int old_ref_mod, new_ref_mod;
7226
7227                 btrfs_ref_tree_mod(root, buf->start, buf->len, parent,
7228                                    root->root_key.objectid,
7229                                    btrfs_header_level(buf), 0,
7230                                    BTRFS_DROP_DELAYED_REF);
7231                 ret = btrfs_add_delayed_tree_ref(fs_info, trans, buf->start,
7232                                                  buf->len, parent,
7233                                                  root->root_key.objectid,
7234                                                  btrfs_header_level(buf),
7235                                                  BTRFS_DROP_DELAYED_REF, NULL,
7236                                                  &old_ref_mod, &new_ref_mod);
7237                 BUG_ON(ret); /* -ENOMEM */
7238                 pin = old_ref_mod >= 0 && new_ref_mod < 0;
7239         }
7240
7241         if (last_ref && btrfs_header_generation(buf) == trans->transid) {
7242                 struct btrfs_block_group_cache *cache;
7243
7244                 if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
7245                         ret = check_ref_cleanup(trans, buf->start);
7246                         if (!ret)
7247                                 goto out;
7248                 }
7249
7250                 pin = 0;
7251                 cache = btrfs_lookup_block_group(fs_info, buf->start);
7252
7253                 if (btrfs_header_flag(buf, BTRFS_HEADER_FLAG_WRITTEN)) {
7254                         pin_down_extent(fs_info, cache, buf->start,
7255                                         buf->len, 1);
7256                         btrfs_put_block_group(cache);
7257                         goto out;
7258                 }
7259
7260                 WARN_ON(test_bit(EXTENT_BUFFER_DIRTY, &buf->bflags));
7261
7262                 btrfs_add_free_space(cache, buf->start, buf->len);
7263                 btrfs_free_reserved_bytes(cache, buf->len, 0);
7264                 btrfs_put_block_group(cache);
7265                 trace_btrfs_reserved_extent_free(fs_info, buf->start, buf->len);
7266         }
7267 out:
7268         if (pin)
7269                 add_pinned_bytes(fs_info, buf->len, btrfs_header_level(buf),
7270                                  root->root_key.objectid);
7271
7272         if (last_ref) {
7273                 /*
7274                  * Deleting the buffer, clear the corrupt flag since it doesn't
7275                  * matter anymore.
7276                  */
7277                 clear_bit(EXTENT_BUFFER_CORRUPT, &buf->bflags);
7278         }
7279 }
7280
7281 /* Can return -ENOMEM */
7282 int btrfs_free_extent(struct btrfs_trans_handle *trans,
7283                       struct btrfs_root *root,
7284                       u64 bytenr, u64 num_bytes, u64 parent, u64 root_objectid,
7285                       u64 owner, u64 offset)
7286 {
7287         struct btrfs_fs_info *fs_info = root->fs_info;
7288         int old_ref_mod, new_ref_mod;
7289         int ret;
7290
7291         if (btrfs_is_testing(fs_info))
7292                 return 0;
7293
7294         if (root_objectid != BTRFS_TREE_LOG_OBJECTID)
7295                 btrfs_ref_tree_mod(root, bytenr, num_bytes, parent,
7296                                    root_objectid, owner, offset,
7297                                    BTRFS_DROP_DELAYED_REF);
7298
7299         /*
7300          * tree log blocks never actually go into the extent allocation
7301          * tree, just update pinning info and exit early.
7302          */
7303         if (root_objectid == BTRFS_TREE_LOG_OBJECTID) {
7304                 WARN_ON(owner >= BTRFS_FIRST_FREE_OBJECTID);
7305                 /* unlocks the pinned mutex */
7306                 btrfs_pin_extent(fs_info, bytenr, num_bytes, 1);
7307                 old_ref_mod = new_ref_mod = 0;
7308                 ret = 0;
7309         } else if (owner < BTRFS_FIRST_FREE_OBJECTID) {
7310                 ret = btrfs_add_delayed_tree_ref(fs_info, trans, bytenr,
7311                                                  num_bytes, parent,
7312                                                  root_objectid, (int)owner,
7313                                                  BTRFS_DROP_DELAYED_REF, NULL,
7314                                                  &old_ref_mod, &new_ref_mod);
7315         } else {
7316                 ret = btrfs_add_delayed_data_ref(fs_info, trans, bytenr,
7317                                                  num_bytes, parent,
7318                                                  root_objectid, owner, offset,
7319                                                  0, BTRFS_DROP_DELAYED_REF,
7320                                                  &old_ref_mod, &new_ref_mod);
7321         }
7322
7323         if (ret == 0 && old_ref_mod >= 0 && new_ref_mod < 0)
7324                 add_pinned_bytes(fs_info, num_bytes, owner, root_objectid);
7325
7326         return ret;
7327 }
7328
7329 /*
7330  * when we wait for progress in the block group caching, its because
7331  * our allocation attempt failed at least once.  So, we must sleep
7332  * and let some progress happen before we try again.
7333  *
7334  * This function will sleep at least once waiting for new free space to
7335  * show up, and then it will check the block group free space numbers
7336  * for our min num_bytes.  Another option is to have it go ahead
7337  * and look in the rbtree for a free extent of a given size, but this
7338  * is a good start.
7339  *
7340  * Callers of this must check if cache->cached == BTRFS_CACHE_ERROR before using
7341  * any of the information in this block group.
7342  */
7343 static noinline void
7344 wait_block_group_cache_progress(struct btrfs_block_group_cache *cache,
7345                                 u64 num_bytes)
7346 {
7347         struct btrfs_caching_control *caching_ctl;
7348
7349         caching_ctl = get_caching_control(cache);
7350         if (!caching_ctl)
7351                 return;
7352
7353         wait_event(caching_ctl->wait, block_group_cache_done(cache) ||
7354                    (cache->free_space_ctl->free_space >= num_bytes));
7355
7356         put_caching_control(caching_ctl);
7357 }
7358
7359 static noinline int
7360 wait_block_group_cache_done(struct btrfs_block_group_cache *cache)
7361 {
7362         struct btrfs_caching_control *caching_ctl;
7363         int ret = 0;
7364
7365         caching_ctl = get_caching_control(cache);
7366         if (!caching_ctl)
7367                 return (cache->cached == BTRFS_CACHE_ERROR) ? -EIO : 0;
7368
7369         wait_event(caching_ctl->wait, block_group_cache_done(cache));
7370         if (cache->cached == BTRFS_CACHE_ERROR)
7371                 ret = -EIO;
7372         put_caching_control(caching_ctl);
7373         return ret;
7374 }
7375
7376 static const char *btrfs_raid_type_names[BTRFS_NR_RAID_TYPES] = {
7377         [BTRFS_RAID_RAID10]     = "raid10",
7378         [BTRFS_RAID_RAID1]      = "raid1",
7379         [BTRFS_RAID_DUP]        = "dup",
7380         [BTRFS_RAID_RAID0]      = "raid0",
7381         [BTRFS_RAID_SINGLE]     = "single",
7382         [BTRFS_RAID_RAID5]      = "raid5",
7383         [BTRFS_RAID_RAID6]      = "raid6",
7384 };
7385
7386 static const char *get_raid_name(enum btrfs_raid_types type)
7387 {
7388         if (type >= BTRFS_NR_RAID_TYPES)
7389                 return NULL;
7390
7391         return btrfs_raid_type_names[type];
7392 }
7393
7394 enum btrfs_loop_type {
7395         LOOP_CACHING_NOWAIT = 0,
7396         LOOP_CACHING_WAIT = 1,
7397         LOOP_ALLOC_CHUNK = 2,
7398         LOOP_NO_EMPTY_SIZE = 3,
7399 };
7400
7401 static inline void
7402 btrfs_lock_block_group(struct btrfs_block_group_cache *cache,
7403                        int delalloc)
7404 {
7405         if (delalloc)
7406                 down_read(&cache->data_rwsem);
7407 }
7408
7409 static inline void
7410 btrfs_grab_block_group(struct btrfs_block_group_cache *cache,
7411                        int delalloc)
7412 {
7413         btrfs_get_block_group(cache);
7414         if (delalloc)
7415                 down_read(&cache->data_rwsem);
7416 }
7417
7418 static struct btrfs_block_group_cache *
7419 btrfs_lock_cluster(struct btrfs_block_group_cache *block_group,
7420                    struct btrfs_free_cluster *cluster,
7421                    int delalloc)
7422 {
7423         struct btrfs_block_group_cache *used_bg = NULL;
7424
7425         spin_lock(&cluster->refill_lock);
7426         while (1) {
7427                 used_bg = cluster->block_group;
7428                 if (!used_bg)
7429                         return NULL;
7430
7431                 if (used_bg == block_group)
7432                         return used_bg;
7433
7434                 btrfs_get_block_group(used_bg);
7435
7436                 if (!delalloc)
7437                         return used_bg;
7438
7439                 if (down_read_trylock(&used_bg->data_rwsem))
7440                         return used_bg;
7441
7442                 spin_unlock(&cluster->refill_lock);
7443
7444                 /* We should only have one-level nested. */
7445                 down_read_nested(&used_bg->data_rwsem, SINGLE_DEPTH_NESTING);
7446
7447                 spin_lock(&cluster->refill_lock);
7448                 if (used_bg == cluster->block_group)
7449                         return used_bg;
7450
7451                 up_read(&used_bg->data_rwsem);
7452                 btrfs_put_block_group(used_bg);
7453         }
7454 }
7455
7456 static inline void
7457 btrfs_release_block_group(struct btrfs_block_group_cache *cache,
7458                          int delalloc)
7459 {
7460         if (delalloc)
7461                 up_read(&cache->data_rwsem);
7462         btrfs_put_block_group(cache);
7463 }
7464
7465 /*
7466  * walks the btree of allocated extents and find a hole of a given size.
7467  * The key ins is changed to record the hole:
7468  * ins->objectid == start position
7469  * ins->flags = BTRFS_EXTENT_ITEM_KEY
7470  * ins->offset == the size of the hole.
7471  * Any available blocks before search_start are skipped.
7472  *
7473  * If there is no suitable free space, we will record the max size of
7474  * the free space extent currently.
7475  */
7476 static noinline int find_free_extent(struct btrfs_fs_info *fs_info,
7477                                 u64 ram_bytes, u64 num_bytes, u64 empty_size,
7478                                 u64 hint_byte, struct btrfs_key *ins,
7479                                 u64 flags, int delalloc)
7480 {
7481         int ret = 0;
7482         struct btrfs_root *root = fs_info->extent_root;
7483         struct btrfs_free_cluster *last_ptr = NULL;
7484         struct btrfs_block_group_cache *block_group = NULL;
7485         u64 search_start = 0;
7486         u64 max_extent_size = 0;
7487         u64 empty_cluster = 0;
7488         struct btrfs_space_info *space_info;
7489         int loop = 0;
7490         int index = btrfs_bg_flags_to_raid_index(flags);
7491         bool failed_cluster_refill = false;
7492         bool failed_alloc = false;
7493         bool use_cluster = true;
7494         bool have_caching_bg = false;
7495         bool orig_have_caching_bg = false;
7496         bool full_search = false;
7497
7498         WARN_ON(num_bytes < fs_info->sectorsize);
7499         ins->type = BTRFS_EXTENT_ITEM_KEY;
7500         ins->objectid = 0;
7501         ins->offset = 0;
7502
7503         trace_find_free_extent(fs_info, num_bytes, empty_size, flags);
7504
7505         space_info = __find_space_info(fs_info, flags);
7506         if (!space_info) {
7507                 btrfs_err(fs_info, "No space info for %llu", flags);
7508                 return -ENOSPC;
7509         }
7510
7511         /*
7512          * If our free space is heavily fragmented we may not be able to make
7513          * big contiguous allocations, so instead of doing the expensive search
7514          * for free space, simply return ENOSPC with our max_extent_size so we
7515          * can go ahead and search for a more manageable chunk.
7516          *
7517          * If our max_extent_size is large enough for our allocation simply
7518          * disable clustering since we will likely not be able to find enough
7519          * space to create a cluster and induce latency trying.
7520          */
7521         if (unlikely(space_info->max_extent_size)) {
7522                 spin_lock(&space_info->lock);
7523                 if (space_info->max_extent_size &&
7524                     num_bytes > space_info->max_extent_size) {
7525                         ins->offset = space_info->max_extent_size;
7526                         spin_unlock(&space_info->lock);
7527                         return -ENOSPC;
7528                 } else if (space_info->max_extent_size) {
7529                         use_cluster = false;
7530                 }
7531                 spin_unlock(&space_info->lock);
7532         }
7533
7534         last_ptr = fetch_cluster_info(fs_info, space_info, &empty_cluster);
7535         if (last_ptr) {
7536                 spin_lock(&last_ptr->lock);
7537                 if (last_ptr->block_group)
7538                         hint_byte = last_ptr->window_start;
7539                 if (last_ptr->fragmented) {
7540                         /*
7541                          * We still set window_start so we can keep track of the
7542                          * last place we found an allocation to try and save
7543                          * some time.
7544                          */
7545                         hint_byte = last_ptr->window_start;
7546                         use_cluster = false;
7547                 }
7548                 spin_unlock(&last_ptr->lock);
7549         }
7550
7551         search_start = max(search_start, first_logical_byte(fs_info, 0));
7552         search_start = max(search_start, hint_byte);
7553         if (search_start == hint_byte) {
7554                 block_group = btrfs_lookup_block_group(fs_info, search_start);
7555                 /*
7556                  * we don't want to use the block group if it doesn't match our
7557                  * allocation bits, or if its not cached.
7558                  *
7559                  * However if we are re-searching with an ideal block group
7560                  * picked out then we don't care that the block group is cached.
7561                  */
7562                 if (block_group && block_group_bits(block_group, flags) &&
7563                     block_group->cached != BTRFS_CACHE_NO) {
7564                         down_read(&space_info->groups_sem);
7565                         if (list_empty(&block_group->list) ||
7566                             block_group->ro) {
7567                                 /*
7568                                  * someone is removing this block group,
7569                                  * we can't jump into the have_block_group
7570                                  * target because our list pointers are not
7571                                  * valid
7572                                  */
7573                                 btrfs_put_block_group(block_group);
7574                                 up_read(&space_info->groups_sem);
7575                         } else {
7576                                 index = btrfs_bg_flags_to_raid_index(
7577                                                 block_group->flags);
7578                                 btrfs_lock_block_group(block_group, delalloc);
7579                                 goto have_block_group;
7580                         }
7581                 } else if (block_group) {
7582                         btrfs_put_block_group(block_group);
7583                 }
7584         }
7585 search:
7586         have_caching_bg = false;
7587         if (index == 0 || index == btrfs_bg_flags_to_raid_index(flags))
7588                 full_search = true;
7589         down_read(&space_info->groups_sem);
7590         list_for_each_entry(block_group, &space_info->block_groups[index],
7591                             list) {
7592                 u64 offset;
7593                 int cached;
7594
7595                 /* If the block group is read-only, we can skip it entirely. */
7596                 if (unlikely(block_group->ro))
7597                         continue;
7598
7599                 btrfs_grab_block_group(block_group, delalloc);
7600                 search_start = block_group->key.objectid;
7601
7602                 /*
7603                  * this can happen if we end up cycling through all the
7604                  * raid types, but we want to make sure we only allocate
7605                  * for the proper type.
7606                  */
7607                 if (!block_group_bits(block_group, flags)) {
7608                     u64 extra = BTRFS_BLOCK_GROUP_DUP |
7609                                 BTRFS_BLOCK_GROUP_RAID1 |
7610                                 BTRFS_BLOCK_GROUP_RAID5 |
7611                                 BTRFS_BLOCK_GROUP_RAID6 |
7612                                 BTRFS_BLOCK_GROUP_RAID10;
7613
7614                         /*
7615                          * if they asked for extra copies and this block group
7616                          * doesn't provide them, bail.  This does allow us to
7617                          * fill raid0 from raid1.
7618                          */
7619                         if ((flags & extra) && !(block_group->flags & extra))
7620                                 goto loop;
7621                 }
7622
7623 have_block_group:
7624                 cached = block_group_cache_done(block_group);
7625                 if (unlikely(!cached)) {
7626                         have_caching_bg = true;
7627                         ret = cache_block_group(block_group, 0);
7628                         BUG_ON(ret < 0);
7629                         ret = 0;
7630                 }
7631
7632                 if (unlikely(block_group->cached == BTRFS_CACHE_ERROR))
7633                         goto loop;
7634
7635                 /*
7636                  * Ok we want to try and use the cluster allocator, so
7637                  * lets look there
7638                  */
7639                 if (last_ptr && use_cluster) {
7640                         struct btrfs_block_group_cache *used_block_group;
7641                         unsigned long aligned_cluster;
7642                         /*
7643                          * the refill lock keeps out other
7644                          * people trying to start a new cluster
7645                          */
7646                         used_block_group = btrfs_lock_cluster(block_group,
7647                                                               last_ptr,
7648                                                               delalloc);
7649                         if (!used_block_group)
7650                                 goto refill_cluster;
7651
7652                         if (used_block_group != block_group &&
7653                             (used_block_group->ro ||
7654                              !block_group_bits(used_block_group, flags)))
7655                                 goto release_cluster;
7656
7657                         offset = btrfs_alloc_from_cluster(used_block_group,
7658                                                 last_ptr,
7659                                                 num_bytes,
7660                                                 used_block_group->key.objectid,
7661                                                 &max_extent_size);
7662                         if (offset) {
7663                                 /* we have a block, we're done */
7664                                 spin_unlock(&last_ptr->refill_lock);
7665                                 trace_btrfs_reserve_extent_cluster(fs_info,
7666                                                 used_block_group,
7667                                                 search_start, num_bytes);
7668                                 if (used_block_group != block_group) {
7669                                         btrfs_release_block_group(block_group,
7670                                                                   delalloc);
7671                                         block_group = used_block_group;
7672                                 }
7673                                 goto checks;
7674                         }
7675
7676                         WARN_ON(last_ptr->block_group != used_block_group);
7677 release_cluster:
7678                         /* If we are on LOOP_NO_EMPTY_SIZE, we can't
7679                          * set up a new clusters, so lets just skip it
7680                          * and let the allocator find whatever block
7681                          * it can find.  If we reach this point, we
7682                          * will have tried the cluster allocator
7683                          * plenty of times and not have found
7684                          * anything, so we are likely way too
7685                          * fragmented for the clustering stuff to find
7686                          * anything.
7687                          *
7688                          * However, if the cluster is taken from the
7689                          * current block group, release the cluster
7690                          * first, so that we stand a better chance of
7691                          * succeeding in the unclustered
7692                          * allocation.  */
7693                         if (loop >= LOOP_NO_EMPTY_SIZE &&
7694                             used_block_group != block_group) {
7695                                 spin_unlock(&last_ptr->refill_lock);
7696                                 btrfs_release_block_group(used_block_group,
7697                                                           delalloc);
7698                                 goto unclustered_alloc;
7699                         }
7700
7701                         /*
7702                          * this cluster didn't work out, free it and
7703                          * start over
7704                          */
7705                         btrfs_return_cluster_to_free_space(NULL, last_ptr);
7706
7707                         if (used_block_group != block_group)
7708                                 btrfs_release_block_group(used_block_group,
7709                                                           delalloc);
7710 refill_cluster:
7711                         if (loop >= LOOP_NO_EMPTY_SIZE) {
7712                                 spin_unlock(&last_ptr->refill_lock);
7713                                 goto unclustered_alloc;
7714                         }
7715
7716                         aligned_cluster = max_t(unsigned long,
7717                                                 empty_cluster + empty_size,
7718                                               block_group->full_stripe_len);
7719
7720                         /* allocate a cluster in this block group */
7721                         ret = btrfs_find_space_cluster(fs_info, block_group,
7722                                                        last_ptr, search_start,
7723                                                        num_bytes,
7724                                                        aligned_cluster);
7725                         if (ret == 0) {
7726                                 /*
7727                                  * now pull our allocation out of this
7728                                  * cluster
7729                                  */
7730                                 offset = btrfs_alloc_from_cluster(block_group,
7731                                                         last_ptr,
7732                                                         num_bytes,
7733                                                         search_start,
7734                                                         &max_extent_size);
7735                                 if (offset) {
7736                                         /* we found one, proceed */
7737                                         spin_unlock(&last_ptr->refill_lock);
7738                                         trace_btrfs_reserve_extent_cluster(fs_info,
7739                                                 block_group, search_start,
7740                                                 num_bytes);
7741                                         goto checks;
7742                                 }
7743                         } else if (!cached && loop > LOOP_CACHING_NOWAIT
7744                                    && !failed_cluster_refill) {
7745                                 spin_unlock(&last_ptr->refill_lock);
7746
7747                                 failed_cluster_refill = true;
7748                                 wait_block_group_cache_progress(block_group,
7749                                        num_bytes + empty_cluster + empty_size);
7750                                 goto have_block_group;
7751                         }
7752
7753                         /*
7754                          * at this point we either didn't find a cluster
7755                          * or we weren't able to allocate a block from our
7756                          * cluster.  Free the cluster we've been trying
7757                          * to use, and go to the next block group
7758                          */
7759                         btrfs_return_cluster_to_free_space(NULL, last_ptr);
7760                         spin_unlock(&last_ptr->refill_lock);
7761                         goto loop;
7762                 }
7763
7764 unclustered_alloc:
7765                 /*
7766                  * We are doing an unclustered alloc, set the fragmented flag so
7767                  * we don't bother trying to setup a cluster again until we get
7768                  * more space.
7769                  */
7770                 if (unlikely(last_ptr)) {
7771                         spin_lock(&last_ptr->lock);
7772                         last_ptr->fragmented = 1;
7773                         spin_unlock(&last_ptr->lock);
7774                 }
7775                 if (cached) {
7776                         struct btrfs_free_space_ctl *ctl =
7777                                 block_group->free_space_ctl;
7778
7779                         spin_lock(&ctl->tree_lock);
7780                         if (ctl->free_space <
7781                             num_bytes + empty_cluster + empty_size) {
7782                                 if (ctl->free_space > max_extent_size)
7783                                         max_extent_size = ctl->free_space;
7784                                 spin_unlock(&ctl->tree_lock);
7785                                 goto loop;
7786                         }
7787                         spin_unlock(&ctl->tree_lock);
7788                 }
7789
7790                 offset = btrfs_find_space_for_alloc(block_group, search_start,
7791                                                     num_bytes, empty_size,
7792                                                     &max_extent_size);
7793                 /*
7794                  * If we didn't find a chunk, and we haven't failed on this
7795                  * block group before, and this block group is in the middle of
7796                  * caching and we are ok with waiting, then go ahead and wait
7797                  * for progress to be made, and set failed_alloc to true.
7798                  *
7799                  * If failed_alloc is true then we've already waited on this
7800                  * block group once and should move on to the next block group.
7801                  */
7802                 if (!offset && !failed_alloc && !cached &&
7803                     loop > LOOP_CACHING_NOWAIT) {
7804                         wait_block_group_cache_progress(block_group,
7805                                                 num_bytes + empty_size);
7806                         failed_alloc = true;
7807                         goto have_block_group;
7808                 } else if (!offset) {
7809                         goto loop;
7810                 }
7811 checks:
7812                 search_start = ALIGN(offset, fs_info->stripesize);
7813
7814                 /* move on to the next group */
7815                 if (search_start + num_bytes >
7816                     block_group->key.objectid + block_group->key.offset) {
7817                         btrfs_add_free_space(block_group, offset, num_bytes);
7818                         goto loop;
7819                 }
7820
7821                 if (offset < search_start)
7822                         btrfs_add_free_space(block_group, offset,
7823                                              search_start - offset);
7824                 BUG_ON(offset > search_start);
7825
7826                 ret = btrfs_add_reserved_bytes(block_group, ram_bytes,
7827                                 num_bytes, delalloc);
7828                 if (ret == -EAGAIN) {
7829                         btrfs_add_free_space(block_group, offset, num_bytes);
7830                         goto loop;
7831                 }
7832                 btrfs_inc_block_group_reservations(block_group);
7833
7834                 /* we are all good, lets return */
7835                 ins->objectid = search_start;
7836                 ins->offset = num_bytes;
7837
7838                 trace_btrfs_reserve_extent(fs_info, block_group,
7839                                            search_start, num_bytes);
7840                 btrfs_release_block_group(block_group, delalloc);
7841                 break;
7842 loop:
7843                 failed_cluster_refill = false;
7844                 failed_alloc = false;
7845                 BUG_ON(btrfs_bg_flags_to_raid_index(block_group->flags) !=
7846                        index);
7847                 btrfs_release_block_group(block_group, delalloc);
7848                 cond_resched();
7849         }
7850         up_read(&space_info->groups_sem);
7851
7852         if ((loop == LOOP_CACHING_NOWAIT) && have_caching_bg
7853                 && !orig_have_caching_bg)
7854                 orig_have_caching_bg = true;
7855
7856         if (!ins->objectid && loop >= LOOP_CACHING_WAIT && have_caching_bg)
7857                 goto search;
7858
7859         if (!ins->objectid && ++index < BTRFS_NR_RAID_TYPES)
7860                 goto search;
7861
7862         /*
7863          * LOOP_CACHING_NOWAIT, search partially cached block groups, kicking
7864          *                      caching kthreads as we move along
7865          * LOOP_CACHING_WAIT, search everything, and wait if our bg is caching
7866          * LOOP_ALLOC_CHUNK, force a chunk allocation and try again
7867          * LOOP_NO_EMPTY_SIZE, set empty_size and empty_cluster to 0 and try
7868          *                      again
7869          */
7870         if (!ins->objectid && loop < LOOP_NO_EMPTY_SIZE) {
7871                 index = 0;
7872                 if (loop == LOOP_CACHING_NOWAIT) {
7873                         /*
7874                          * We want to skip the LOOP_CACHING_WAIT step if we
7875                          * don't have any uncached bgs and we've already done a
7876                          * full search through.
7877                          */
7878                         if (orig_have_caching_bg || !full_search)
7879                                 loop = LOOP_CACHING_WAIT;
7880                         else
7881                                 loop = LOOP_ALLOC_CHUNK;
7882                 } else {
7883                         loop++;
7884                 }
7885
7886                 if (loop == LOOP_ALLOC_CHUNK) {
7887                         struct btrfs_trans_handle *trans;
7888                         int exist = 0;
7889
7890                         trans = current->journal_info;
7891                         if (trans)
7892                                 exist = 1;
7893                         else
7894                                 trans = btrfs_join_transaction(root);
7895
7896                         if (IS_ERR(trans)) {
7897                                 ret = PTR_ERR(trans);
7898                                 goto out;
7899                         }
7900
7901                         ret = do_chunk_alloc(trans, fs_info, flags,
7902                                              CHUNK_ALLOC_FORCE);
7903
7904                         /*
7905                          * If we can't allocate a new chunk we've already looped
7906                          * through at least once, move on to the NO_EMPTY_SIZE
7907                          * case.
7908                          */
7909                         if (ret == -ENOSPC)
7910                                 loop = LOOP_NO_EMPTY_SIZE;
7911
7912                         /*
7913                          * Do not bail out on ENOSPC since we
7914                          * can do more things.
7915                          */
7916                         if (ret < 0 && ret != -ENOSPC)
7917                                 btrfs_abort_transaction(trans, ret);
7918                         else
7919                                 ret = 0;
7920                         if (!exist)
7921                                 btrfs_end_transaction(trans);
7922                         if (ret)
7923                                 goto out;
7924                 }
7925
7926                 if (loop == LOOP_NO_EMPTY_SIZE) {
7927                         /*
7928                          * Don't loop again if we already have no empty_size and
7929                          * no empty_cluster.
7930                          */
7931                         if (empty_size == 0 &&
7932                             empty_cluster == 0) {
7933                                 ret = -ENOSPC;
7934                                 goto out;
7935                         }
7936                         empty_size = 0;
7937                         empty_cluster = 0;
7938                 }
7939
7940                 goto search;
7941         } else if (!ins->objectid) {
7942                 ret = -ENOSPC;
7943         } else if (ins->objectid) {
7944                 if (!use_cluster && last_ptr) {
7945                         spin_lock(&last_ptr->lock);
7946                         last_ptr->window_start = ins->objectid;
7947                         spin_unlock(&last_ptr->lock);
7948                 }
7949                 ret = 0;
7950         }
7951 out:
7952         if (ret == -ENOSPC) {
7953                 spin_lock(&space_info->lock);
7954                 space_info->max_extent_size = max_extent_size;
7955                 spin_unlock(&space_info->lock);
7956                 ins->offset = max_extent_size;
7957         }
7958         return ret;
7959 }
7960
7961 static void dump_space_info(struct btrfs_fs_info *fs_info,
7962                             struct btrfs_space_info *info, u64 bytes,
7963                             int dump_block_groups)
7964 {
7965         struct btrfs_block_group_cache *cache;
7966         int index = 0;
7967
7968         spin_lock(&info->lock);
7969         btrfs_info(fs_info, "space_info %llu has %llu free, is %sfull",
7970                    info->flags,
7971                    info->total_bytes - btrfs_space_info_used(info, true),
7972                    info->full ? "" : "not ");
7973         btrfs_info(fs_info,
7974                 "space_info total=%llu, used=%llu, pinned=%llu, reserved=%llu, may_use=%llu, readonly=%llu",
7975                 info->total_bytes, info->bytes_used, info->bytes_pinned,
7976                 info->bytes_reserved, info->bytes_may_use,
7977                 info->bytes_readonly);
7978         spin_unlock(&info->lock);
7979
7980         if (!dump_block_groups)
7981                 return;
7982
7983         down_read(&info->groups_sem);
7984 again:
7985         list_for_each_entry(cache, &info->block_groups[index], list) {
7986                 spin_lock(&cache->lock);
7987                 btrfs_info(fs_info,
7988                         "block group %llu has %llu bytes, %llu used %llu pinned %llu reserved %s",
7989                         cache->key.objectid, cache->key.offset,
7990                         btrfs_block_group_used(&cache->item), cache->pinned,
7991                         cache->reserved, cache->ro ? "[readonly]" : "");
7992                 btrfs_dump_free_space(cache, bytes);
7993                 spin_unlock(&cache->lock);
7994         }
7995         if (++index < BTRFS_NR_RAID_TYPES)
7996                 goto again;
7997         up_read(&info->groups_sem);
7998 }
7999
8000 /*
8001  * btrfs_reserve_extent - entry point to the extent allocator. Tries to find a
8002  *                        hole that is at least as big as @num_bytes.
8003  *
8004  * @root           -    The root that will contain this extent
8005  *
8006  * @ram_bytes      -    The amount of space in ram that @num_bytes take. This
8007  *                      is used for accounting purposes. This value differs
8008  *                      from @num_bytes only in the case of compressed extents.
8009  *
8010  * @num_bytes      -    Number of bytes to allocate on-disk.
8011  *
8012  * @min_alloc_size -    Indicates the minimum amount of space that the
8013  *                      allocator should try to satisfy. In some cases
8014  *                      @num_bytes may be larger than what is required and if
8015  *                      the filesystem is fragmented then allocation fails.
8016  *                      However, the presence of @min_alloc_size gives a
8017  *                      chance to try and satisfy the smaller allocation.
8018  *
8019  * @empty_size     -    A hint that you plan on doing more COW. This is the
8020  *                      size in bytes the allocator should try to find free
8021  *                      next to the block it returns.  This is just a hint and
8022  *                      may be ignored by the allocator.
8023  *
8024  * @hint_byte      -    Hint to the allocator to start searching above the byte
8025  *                      address passed. It might be ignored.
8026  *
8027  * @ins            -    This key is modified to record the found hole. It will
8028  *                      have the following values:
8029  *                      ins->objectid == start position
8030  *                      ins->flags = BTRFS_EXTENT_ITEM_KEY
8031  *                      ins->offset == the size of the hole.
8032  *
8033  * @is_data        -    Boolean flag indicating whether an extent is
8034  *                      allocated for data (true) or metadata (false)
8035  *
8036  * @delalloc       -    Boolean flag indicating whether this allocation is for
8037  *                      delalloc or not. If 'true' data_rwsem of block groups
8038  *                      is going to be acquired.
8039  *
8040  *
8041  * Returns 0 when an allocation succeeded or < 0 when an error occurred. In
8042  * case -ENOSPC is returned then @ins->offset will contain the size of the
8043  * largest available hole the allocator managed to find.
8044  */
8045 int btrfs_reserve_extent(struct btrfs_root *root, u64 ram_bytes,
8046                          u64 num_bytes, u64 min_alloc_size,
8047                          u64 empty_size, u64 hint_byte,
8048                          struct btrfs_key *ins, int is_data, int delalloc)
8049 {
8050         struct btrfs_fs_info *fs_info = root->fs_info;
8051         bool final_tried = num_bytes == min_alloc_size;
8052         u64 flags;
8053         int ret;
8054
8055         flags = get_alloc_profile_by_root(root, is_data);
8056 again:
8057         WARN_ON(num_bytes < fs_info->sectorsize);
8058         ret = find_free_extent(fs_info, ram_bytes, num_bytes, empty_size,
8059                                hint_byte, ins, flags, delalloc);
8060         if (!ret && !is_data) {
8061                 btrfs_dec_block_group_reservations(fs_info, ins->objectid);
8062         } else if (ret == -ENOSPC) {
8063                 if (!final_tried && ins->offset) {
8064                         num_bytes = min(num_bytes >> 1, ins->offset);
8065                         num_bytes = round_down(num_bytes,
8066                                                fs_info->sectorsize);
8067                         num_bytes = max(num_bytes, min_alloc_size);
8068                         ram_bytes = num_bytes;
8069                         if (num_bytes == min_alloc_size)
8070                                 final_tried = true;
8071                         goto again;
8072                 } else if (btrfs_test_opt(fs_info, ENOSPC_DEBUG)) {
8073                         struct btrfs_space_info *sinfo;
8074
8075                         sinfo = __find_space_info(fs_info, flags);
8076                         btrfs_err(fs_info,
8077                                   "allocation failed flags %llu, wanted %llu",
8078                                   flags, num_bytes);
8079                         if (sinfo)
8080                                 dump_space_info(fs_info, sinfo, num_bytes, 1);
8081                 }
8082         }
8083
8084         return ret;
8085 }
8086
8087 static int __btrfs_free_reserved_extent(struct btrfs_fs_info *fs_info,
8088                                         u64 start, u64 len,
8089                                         int pin, int delalloc)
8090 {
8091         struct btrfs_block_group_cache *cache;
8092         int ret = 0;
8093
8094         cache = btrfs_lookup_block_group(fs_info, start);
8095         if (!cache) {
8096                 btrfs_err(fs_info, "Unable to find block group for %llu",
8097                           start);
8098                 return -ENOSPC;
8099         }
8100
8101         if (pin)
8102                 pin_down_extent(fs_info, cache, start, len, 1);
8103         else {
8104                 if (btrfs_test_opt(fs_info, DISCARD))
8105                         ret = btrfs_discard_extent(fs_info, start, len, NULL);
8106                 btrfs_add_free_space(cache, start, len);
8107                 btrfs_free_reserved_bytes(cache, len, delalloc);
8108                 trace_btrfs_reserved_extent_free(fs_info, start, len);
8109         }
8110
8111         btrfs_put_block_group(cache);
8112         return ret;
8113 }
8114
8115 int btrfs_free_reserved_extent(struct btrfs_fs_info *fs_info,
8116                                u64 start, u64 len, int delalloc)
8117 {
8118         return __btrfs_free_reserved_extent(fs_info, start, len, 0, delalloc);
8119 }
8120
8121 int btrfs_free_and_pin_reserved_extent(struct btrfs_fs_info *fs_info,
8122                                        u64 start, u64 len)
8123 {
8124         return __btrfs_free_reserved_extent(fs_info, start, len, 1, 0);
8125 }
8126
8127 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
8128                                       struct btrfs_fs_info *fs_info,
8129                                       u64 parent, u64 root_objectid,
8130                                       u64 flags, u64 owner, u64 offset,
8131                                       struct btrfs_key *ins, int ref_mod)
8132 {
8133         int ret;
8134         struct btrfs_extent_item *extent_item;
8135         struct btrfs_extent_inline_ref *iref;
8136         struct btrfs_path *path;
8137         struct extent_buffer *leaf;
8138         int type;
8139         u32 size;
8140
8141         if (parent > 0)
8142                 type = BTRFS_SHARED_DATA_REF_KEY;
8143         else
8144                 type = BTRFS_EXTENT_DATA_REF_KEY;
8145
8146         size = sizeof(*extent_item) + btrfs_extent_inline_ref_size(type);
8147
8148         path = btrfs_alloc_path();
8149         if (!path)
8150                 return -ENOMEM;
8151
8152         path->leave_spinning = 1;
8153         ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path,
8154                                       ins, size);
8155         if (ret) {
8156                 btrfs_free_path(path);
8157                 return ret;
8158         }
8159
8160         leaf = path->nodes[0];
8161         extent_item = btrfs_item_ptr(leaf, path->slots[0],
8162                                      struct btrfs_extent_item);
8163         btrfs_set_extent_refs(leaf, extent_item, ref_mod);
8164         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
8165         btrfs_set_extent_flags(leaf, extent_item,
8166                                flags | BTRFS_EXTENT_FLAG_DATA);
8167
8168         iref = (struct btrfs_extent_inline_ref *)(extent_item + 1);
8169         btrfs_set_extent_inline_ref_type(leaf, iref, type);
8170         if (parent > 0) {
8171                 struct btrfs_shared_data_ref *ref;
8172                 ref = (struct btrfs_shared_data_ref *)(iref + 1);
8173                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
8174                 btrfs_set_shared_data_ref_count(leaf, ref, ref_mod);
8175         } else {
8176                 struct btrfs_extent_data_ref *ref;
8177                 ref = (struct btrfs_extent_data_ref *)(&iref->offset);
8178                 btrfs_set_extent_data_ref_root(leaf, ref, root_objectid);
8179                 btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
8180                 btrfs_set_extent_data_ref_offset(leaf, ref, offset);
8181                 btrfs_set_extent_data_ref_count(leaf, ref, ref_mod);
8182         }
8183
8184         btrfs_mark_buffer_dirty(path->nodes[0]);
8185         btrfs_free_path(path);
8186
8187         ret = remove_from_free_space_tree(trans, fs_info, ins->objectid,
8188                                           ins->offset);
8189         if (ret)
8190                 return ret;
8191
8192         ret = update_block_group(trans, fs_info, ins->objectid, ins->offset, 1);
8193         if (ret) { /* -ENOENT, logic error */
8194                 btrfs_err(fs_info, "update block group failed for %llu %llu",
8195                         ins->objectid, ins->offset);
8196                 BUG();
8197         }
8198         trace_btrfs_reserved_extent_alloc(fs_info, ins->objectid, ins->offset);
8199         return ret;
8200 }
8201
8202 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
8203                                      struct btrfs_fs_info *fs_info,
8204                                      u64 parent, u64 root_objectid,
8205                                      u64 flags, struct btrfs_disk_key *key,
8206                                      int level, struct btrfs_key *ins)
8207 {
8208         int ret;
8209         struct btrfs_extent_item *extent_item;
8210         struct btrfs_tree_block_info *block_info;
8211         struct btrfs_extent_inline_ref *iref;
8212         struct btrfs_path *path;
8213         struct extent_buffer *leaf;
8214         u32 size = sizeof(*extent_item) + sizeof(*iref);
8215         u64 num_bytes = ins->offset;
8216         bool skinny_metadata = btrfs_fs_incompat(fs_info, SKINNY_METADATA);
8217
8218         if (!skinny_metadata)
8219                 size += sizeof(*block_info);
8220
8221         path = btrfs_alloc_path();
8222         if (!path) {
8223                 btrfs_free_and_pin_reserved_extent(fs_info, ins->objectid,
8224                                                    fs_info->nodesize);
8225                 return -ENOMEM;
8226         }
8227
8228         path->leave_spinning = 1;
8229         ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path,
8230                                       ins, size);
8231         if (ret) {
8232                 btrfs_free_path(path);
8233                 btrfs_free_and_pin_reserved_extent(fs_info, ins->objectid,
8234                                                    fs_info->nodesize);
8235                 return ret;
8236         }
8237
8238         leaf = path->nodes[0];
8239         extent_item = btrfs_item_ptr(leaf, path->slots[0],
8240                                      struct btrfs_extent_item);
8241         btrfs_set_extent_refs(leaf, extent_item, 1);
8242         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
8243         btrfs_set_extent_flags(leaf, extent_item,
8244                                flags | BTRFS_EXTENT_FLAG_TREE_BLOCK);
8245
8246         if (skinny_metadata) {
8247                 iref = (struct btrfs_extent_inline_ref *)(extent_item + 1);
8248                 num_bytes = fs_info->nodesize;
8249         } else {
8250                 block_info = (struct btrfs_tree_block_info *)(extent_item + 1);
8251                 btrfs_set_tree_block_key(leaf, block_info, key);
8252                 btrfs_set_tree_block_level(leaf, block_info, level);
8253                 iref = (struct btrfs_extent_inline_ref *)(block_info + 1);
8254         }
8255
8256         if (parent > 0) {
8257                 BUG_ON(!(flags & BTRFS_BLOCK_FLAG_FULL_BACKREF));
8258                 btrfs_set_extent_inline_ref_type(leaf, iref,
8259                                                  BTRFS_SHARED_BLOCK_REF_KEY);
8260                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
8261         } else {
8262                 btrfs_set_extent_inline_ref_type(leaf, iref,
8263                                                  BTRFS_TREE_BLOCK_REF_KEY);
8264                 btrfs_set_extent_inline_ref_offset(leaf, iref, root_objectid);
8265         }
8266
8267         btrfs_mark_buffer_dirty(leaf);
8268         btrfs_free_path(path);
8269
8270         ret = remove_from_free_space_tree(trans, fs_info, ins->objectid,
8271                                           num_bytes);
8272         if (ret)
8273                 return ret;
8274
8275         ret = update_block_group(trans, fs_info, ins->objectid,
8276                                  fs_info->nodesize, 1);
8277         if (ret) { /* -ENOENT, logic error */
8278                 btrfs_err(fs_info, "update block group failed for %llu %llu",
8279                         ins->objectid, ins->offset);
8280                 BUG();
8281         }
8282
8283         trace_btrfs_reserved_extent_alloc(fs_info, ins->objectid,
8284                                           fs_info->nodesize);
8285         return ret;
8286 }
8287
8288 int btrfs_alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
8289                                      struct btrfs_root *root, u64 owner,
8290                                      u64 offset, u64 ram_bytes,
8291                                      struct btrfs_key *ins)
8292 {
8293         struct btrfs_fs_info *fs_info = root->fs_info;
8294         int ret;
8295
8296         BUG_ON(root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID);
8297
8298         btrfs_ref_tree_mod(root, ins->objectid, ins->offset, 0,
8299                            root->root_key.objectid, owner, offset,
8300                            BTRFS_ADD_DELAYED_EXTENT);
8301
8302         ret = btrfs_add_delayed_data_ref(fs_info, trans, ins->objectid,
8303                                          ins->offset, 0,
8304                                          root->root_key.objectid, owner,
8305                                          offset, ram_bytes,
8306                                          BTRFS_ADD_DELAYED_EXTENT, NULL, NULL);
8307         return ret;
8308 }
8309
8310 /*
8311  * this is used by the tree logging recovery code.  It records that
8312  * an extent has been allocated and makes sure to clear the free
8313  * space cache bits as well
8314  */
8315 int btrfs_alloc_logged_file_extent(struct btrfs_trans_handle *trans,
8316                                    struct btrfs_fs_info *fs_info,
8317                                    u64 root_objectid, u64 owner, u64 offset,
8318                                    struct btrfs_key *ins)
8319 {
8320         int ret;
8321         struct btrfs_block_group_cache *block_group;
8322         struct btrfs_space_info *space_info;
8323
8324         /*
8325          * Mixed block groups will exclude before processing the log so we only
8326          * need to do the exclude dance if this fs isn't mixed.
8327          */
8328         if (!btrfs_fs_incompat(fs_info, MIXED_GROUPS)) {
8329                 ret = __exclude_logged_extent(fs_info, ins->objectid,
8330                                               ins->offset);
8331                 if (ret)
8332                         return ret;
8333         }
8334
8335         block_group = btrfs_lookup_block_group(fs_info, ins->objectid);
8336         if (!block_group)
8337                 return -EINVAL;
8338
8339         space_info = block_group->space_info;
8340         spin_lock(&space_info->lock);
8341         spin_lock(&block_group->lock);
8342         space_info->bytes_reserved += ins->offset;
8343         block_group->reserved += ins->offset;
8344         spin_unlock(&block_group->lock);
8345         spin_unlock(&space_info->lock);
8346
8347         ret = alloc_reserved_file_extent(trans, fs_info, 0, root_objectid,
8348                                          0, owner, offset, ins, 1);
8349         btrfs_put_block_group(block_group);
8350         return ret;
8351 }
8352
8353 static struct extent_buffer *
8354 btrfs_init_new_buffer(struct btrfs_trans_handle *trans, struct btrfs_root *root,
8355                       u64 bytenr, int level)
8356 {
8357         struct btrfs_fs_info *fs_info = root->fs_info;
8358         struct extent_buffer *buf;
8359
8360         buf = btrfs_find_create_tree_block(fs_info, bytenr);
8361         if (IS_ERR(buf))
8362                 return buf;
8363
8364         btrfs_set_header_generation(buf, trans->transid);
8365         btrfs_set_buffer_lockdep_class(root->root_key.objectid, buf, level);
8366         btrfs_tree_lock(buf);
8367         clean_tree_block(fs_info, buf);
8368         clear_bit(EXTENT_BUFFER_STALE, &buf->bflags);
8369
8370         btrfs_set_lock_blocking(buf);
8371         set_extent_buffer_uptodate(buf);
8372
8373         if (root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID) {
8374                 buf->log_index = root->log_transid % 2;
8375                 /*
8376                  * we allow two log transactions at a time, use different
8377                  * EXENT bit to differentiate dirty pages.
8378                  */
8379                 if (buf->log_index == 0)
8380                         set_extent_dirty(&root->dirty_log_pages, buf->start,
8381                                         buf->start + buf->len - 1, GFP_NOFS);
8382                 else
8383                         set_extent_new(&root->dirty_log_pages, buf->start,
8384                                         buf->start + buf->len - 1);
8385         } else {
8386                 buf->log_index = -1;
8387                 set_extent_dirty(&trans->transaction->dirty_pages, buf->start,
8388                          buf->start + buf->len - 1, GFP_NOFS);
8389         }
8390         trans->dirty = true;
8391         /* this returns a buffer locked for blocking */
8392         return buf;
8393 }
8394
8395 static struct btrfs_block_rsv *
8396 use_block_rsv(struct btrfs_trans_handle *trans,
8397               struct btrfs_root *root, u32 blocksize)
8398 {
8399         struct btrfs_fs_info *fs_info = root->fs_info;
8400         struct btrfs_block_rsv *block_rsv;
8401         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
8402         int ret;
8403         bool global_updated = false;
8404
8405         block_rsv = get_block_rsv(trans, root);
8406
8407         if (unlikely(block_rsv->size == 0))
8408                 goto try_reserve;
8409 again:
8410         ret = block_rsv_use_bytes(block_rsv, blocksize);
8411         if (!ret)
8412                 return block_rsv;
8413
8414         if (block_rsv->failfast)
8415                 return ERR_PTR(ret);
8416
8417         if (block_rsv->type == BTRFS_BLOCK_RSV_GLOBAL && !global_updated) {
8418                 global_updated = true;
8419                 update_global_block_rsv(fs_info);
8420                 goto again;
8421         }
8422
8423         if (btrfs_test_opt(fs_info, ENOSPC_DEBUG)) {
8424                 static DEFINE_RATELIMIT_STATE(_rs,
8425                                 DEFAULT_RATELIMIT_INTERVAL * 10,
8426                                 /*DEFAULT_RATELIMIT_BURST*/ 1);
8427                 if (__ratelimit(&_rs))
8428                         WARN(1, KERN_DEBUG
8429                                 "BTRFS: block rsv returned %d\n", ret);
8430         }
8431 try_reserve:
8432         ret = reserve_metadata_bytes(root, block_rsv, blocksize,
8433                                      BTRFS_RESERVE_NO_FLUSH);
8434         if (!ret)
8435                 return block_rsv;
8436         /*
8437          * If we couldn't reserve metadata bytes try and use some from
8438          * the global reserve if its space type is the same as the global
8439          * reservation.
8440          */
8441         if (block_rsv->type != BTRFS_BLOCK_RSV_GLOBAL &&
8442             block_rsv->space_info == global_rsv->space_info) {
8443                 ret = block_rsv_use_bytes(global_rsv, blocksize);
8444                 if (!ret)
8445                         return global_rsv;
8446         }
8447         return ERR_PTR(ret);
8448 }
8449
8450 static void unuse_block_rsv(struct btrfs_fs_info *fs_info,
8451                             struct btrfs_block_rsv *block_rsv, u32 blocksize)
8452 {
8453         block_rsv_add_bytes(block_rsv, blocksize, 0);
8454         block_rsv_release_bytes(fs_info, block_rsv, NULL, 0, NULL);
8455 }
8456
8457 /*
8458  * finds a free extent and does all the dirty work required for allocation
8459  * returns the tree buffer or an ERR_PTR on error.
8460  */
8461 struct extent_buffer *btrfs_alloc_tree_block(struct btrfs_trans_handle *trans,
8462                                              struct btrfs_root *root,
8463                                              u64 parent, u64 root_objectid,
8464                                              const struct btrfs_disk_key *key,
8465                                              int level, u64 hint,
8466                                              u64 empty_size)
8467 {
8468         struct btrfs_fs_info *fs_info = root->fs_info;
8469         struct btrfs_key ins;
8470         struct btrfs_block_rsv *block_rsv;
8471         struct extent_buffer *buf;
8472         struct btrfs_delayed_extent_op *extent_op;
8473         u64 flags = 0;
8474         int ret;
8475         u32 blocksize = fs_info->nodesize;
8476         bool skinny_metadata = btrfs_fs_incompat(fs_info, SKINNY_METADATA);
8477
8478 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
8479         if (btrfs_is_testing(fs_info)) {
8480                 buf = btrfs_init_new_buffer(trans, root, root->alloc_bytenr,
8481                                             level);
8482                 if (!IS_ERR(buf))
8483                         root->alloc_bytenr += blocksize;
8484                 return buf;
8485         }
8486 #endif
8487
8488         block_rsv = use_block_rsv(trans, root, blocksize);
8489         if (IS_ERR(block_rsv))
8490                 return ERR_CAST(block_rsv);
8491
8492         ret = btrfs_reserve_extent(root, blocksize, blocksize, blocksize,
8493                                    empty_size, hint, &ins, 0, 0);
8494         if (ret)
8495                 goto out_unuse;
8496
8497         buf = btrfs_init_new_buffer(trans, root, ins.objectid, level);
8498         if (IS_ERR(buf)) {
8499                 ret = PTR_ERR(buf);
8500                 goto out_free_reserved;
8501         }
8502
8503         if (root_objectid == BTRFS_TREE_RELOC_OBJECTID) {
8504                 if (parent == 0)
8505                         parent = ins.objectid;
8506                 flags |= BTRFS_BLOCK_FLAG_FULL_BACKREF;
8507         } else
8508                 BUG_ON(parent > 0);
8509
8510         if (root_objectid != BTRFS_TREE_LOG_OBJECTID) {
8511                 extent_op = btrfs_alloc_delayed_extent_op();
8512                 if (!extent_op) {
8513                         ret = -ENOMEM;
8514                         goto out_free_buf;
8515                 }
8516                 if (key)
8517                         memcpy(&extent_op->key, key, sizeof(extent_op->key));
8518                 else
8519                         memset(&extent_op->key, 0, sizeof(extent_op->key));
8520                 extent_op->flags_to_set = flags;
8521                 extent_op->update_key = skinny_metadata ? false : true;
8522                 extent_op->update_flags = true;
8523                 extent_op->is_data = false;
8524                 extent_op->level = level;
8525
8526                 btrfs_ref_tree_mod(root, ins.objectid, ins.offset, parent,
8527                                    root_objectid, level, 0,
8528                                    BTRFS_ADD_DELAYED_EXTENT);
8529                 ret = btrfs_add_delayed_tree_ref(fs_info, trans, ins.objectid,
8530                                                  ins.offset, parent,
8531                                                  root_objectid, level,
8532                                                  BTRFS_ADD_DELAYED_EXTENT,
8533                                                  extent_op, NULL, NULL);
8534                 if (ret)
8535                         goto out_free_delayed;
8536         }
8537         return buf;
8538
8539 out_free_delayed:
8540         btrfs_free_delayed_extent_op(extent_op);
8541 out_free_buf:
8542         free_extent_buffer(buf);
8543 out_free_reserved:
8544         btrfs_free_reserved_extent(fs_info, ins.objectid, ins.offset, 0);
8545 out_unuse:
8546         unuse_block_rsv(fs_info, block_rsv, blocksize);
8547         return ERR_PTR(ret);
8548 }
8549
8550 struct walk_control {
8551         u64 refs[BTRFS_MAX_LEVEL];
8552         u64 flags[BTRFS_MAX_LEVEL];
8553         struct btrfs_key update_progress;
8554         int stage;
8555         int level;
8556         int shared_level;
8557         int update_ref;
8558         int keep_locks;
8559         int reada_slot;
8560         int reada_count;
8561         int for_reloc;
8562 };
8563
8564 #define DROP_REFERENCE  1
8565 #define UPDATE_BACKREF  2
8566
8567 static noinline void reada_walk_down(struct btrfs_trans_handle *trans,
8568                                      struct btrfs_root *root,
8569                                      struct walk_control *wc,
8570                                      struct btrfs_path *path)
8571 {
8572         struct btrfs_fs_info *fs_info = root->fs_info;
8573         u64 bytenr;
8574         u64 generation;
8575         u64 refs;
8576         u64 flags;
8577         u32 nritems;
8578         struct btrfs_key key;
8579         struct extent_buffer *eb;
8580         int ret;
8581         int slot;
8582         int nread = 0;
8583
8584         if (path->slots[wc->level] < wc->reada_slot) {
8585                 wc->reada_count = wc->reada_count * 2 / 3;
8586                 wc->reada_count = max(wc->reada_count, 2);
8587         } else {
8588                 wc->reada_count = wc->reada_count * 3 / 2;
8589                 wc->reada_count = min_t(int, wc->reada_count,
8590                                         BTRFS_NODEPTRS_PER_BLOCK(fs_info));
8591         }
8592
8593         eb = path->nodes[wc->level];
8594         nritems = btrfs_header_nritems(eb);
8595
8596         for (slot = path->slots[wc->level]; slot < nritems; slot++) {
8597                 if (nread >= wc->reada_count)
8598                         break;
8599
8600                 cond_resched();
8601                 bytenr = btrfs_node_blockptr(eb, slot);
8602                 generation = btrfs_node_ptr_generation(eb, slot);
8603
8604                 if (slot == path->slots[wc->level])
8605                         goto reada;
8606
8607                 if (wc->stage == UPDATE_BACKREF &&
8608                     generation <= root->root_key.offset)
8609                         continue;
8610
8611                 /* We don't lock the tree block, it's OK to be racy here */
8612                 ret = btrfs_lookup_extent_info(trans, fs_info, bytenr,
8613                                                wc->level - 1, 1, &refs,
8614                                                &flags);
8615                 /* We don't care about errors in readahead. */
8616                 if (ret < 0)
8617                         continue;
8618                 BUG_ON(refs == 0);
8619
8620                 if (wc->stage == DROP_REFERENCE) {
8621                         if (refs == 1)
8622                                 goto reada;
8623
8624                         if (wc->level == 1 &&
8625                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
8626                                 continue;
8627                         if (!wc->update_ref ||
8628                             generation <= root->root_key.offset)
8629                                 continue;
8630                         btrfs_node_key_to_cpu(eb, &key, slot);
8631                         ret = btrfs_comp_cpu_keys(&key,
8632                                                   &wc->update_progress);
8633                         if (ret < 0)
8634                                 continue;
8635                 } else {
8636                         if (wc->level == 1 &&
8637                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
8638                                 continue;
8639                 }
8640 reada:
8641                 readahead_tree_block(fs_info, bytenr);
8642                 nread++;
8643         }
8644         wc->reada_slot = slot;
8645 }
8646
8647 /*
8648  * helper to process tree block while walking down the tree.
8649  *
8650  * when wc->stage == UPDATE_BACKREF, this function updates
8651  * back refs for pointers in the block.
8652  *
8653  * NOTE: return value 1 means we should stop walking down.
8654  */
8655 static noinline int walk_down_proc(struct btrfs_trans_handle *trans,
8656                                    struct btrfs_root *root,
8657                                    struct btrfs_path *path,
8658                                    struct walk_control *wc, int lookup_info)
8659 {
8660         struct btrfs_fs_info *fs_info = root->fs_info;
8661         int level = wc->level;
8662         struct extent_buffer *eb = path->nodes[level];
8663         u64 flag = BTRFS_BLOCK_FLAG_FULL_BACKREF;
8664         int ret;
8665
8666         if (wc->stage == UPDATE_BACKREF &&
8667             btrfs_header_owner(eb) != root->root_key.objectid)
8668                 return 1;
8669
8670         /*
8671          * when reference count of tree block is 1, it won't increase
8672          * again. once full backref flag is set, we never clear it.
8673          */
8674         if (lookup_info &&
8675             ((wc->stage == DROP_REFERENCE && wc->refs[level] != 1) ||
8676              (wc->stage == UPDATE_BACKREF && !(wc->flags[level] & flag)))) {
8677                 BUG_ON(!path->locks[level]);
8678                 ret = btrfs_lookup_extent_info(trans, fs_info,
8679                                                eb->start, level, 1,
8680                                                &wc->refs[level],
8681                                                &wc->flags[level]);
8682                 BUG_ON(ret == -ENOMEM);
8683                 if (ret)
8684                         return ret;
8685                 BUG_ON(wc->refs[level] == 0);
8686         }
8687
8688         if (wc->stage == DROP_REFERENCE) {
8689                 if (wc->refs[level] > 1)
8690                         return 1;
8691
8692                 if (path->locks[level] && !wc->keep_locks) {
8693                         btrfs_tree_unlock_rw(eb, path->locks[level]);
8694                         path->locks[level] = 0;
8695                 }
8696                 return 0;
8697         }
8698
8699         /* wc->stage == UPDATE_BACKREF */
8700         if (!(wc->flags[level] & flag)) {
8701                 BUG_ON(!path->locks[level]);
8702                 ret = btrfs_inc_ref(trans, root, eb, 1);
8703                 BUG_ON(ret); /* -ENOMEM */
8704                 ret = btrfs_dec_ref(trans, root, eb, 0);
8705                 BUG_ON(ret); /* -ENOMEM */
8706                 ret = btrfs_set_disk_extent_flags(trans, fs_info, eb->start,
8707                                                   eb->len, flag,
8708                                                   btrfs_header_level(eb), 0);
8709                 BUG_ON(ret); /* -ENOMEM */
8710                 wc->flags[level] |= flag;
8711         }
8712
8713         /*
8714          * the block is shared by multiple trees, so it's not good to
8715          * keep the tree lock
8716          */
8717         if (path->locks[level] && level > 0) {
8718                 btrfs_tree_unlock_rw(eb, path->locks[level]);
8719                 path->locks[level] = 0;
8720         }
8721         return 0;
8722 }
8723
8724 /*
8725  * helper to process tree block pointer.
8726  *
8727  * when wc->stage == DROP_REFERENCE, this function checks
8728  * reference count of the block pointed to. if the block
8729  * is shared and we need update back refs for the subtree
8730  * rooted at the block, this function changes wc->stage to
8731  * UPDATE_BACKREF. if the block is shared and there is no
8732  * need to update back, this function drops the reference
8733  * to the block.
8734  *
8735  * NOTE: return value 1 means we should stop walking down.
8736  */
8737 static noinline int do_walk_down(struct btrfs_trans_handle *trans,
8738                                  struct btrfs_root *root,
8739                                  struct btrfs_path *path,
8740                                  struct walk_control *wc, int *lookup_info)
8741 {
8742         struct btrfs_fs_info *fs_info = root->fs_info;
8743         u64 bytenr;
8744         u64 generation;
8745         u64 parent;
8746         u32 blocksize;
8747         struct btrfs_key key;
8748         struct btrfs_key first_key;
8749         struct extent_buffer *next;
8750         int level = wc->level;
8751         int reada = 0;
8752         int ret = 0;
8753         bool need_account = false;
8754
8755         generation = btrfs_node_ptr_generation(path->nodes[level],
8756                                                path->slots[level]);
8757         /*
8758          * if the lower level block was created before the snapshot
8759          * was created, we know there is no need to update back refs
8760          * for the subtree
8761          */
8762         if (wc->stage == UPDATE_BACKREF &&
8763             generation <= root->root_key.offset) {
8764                 *lookup_info = 1;
8765                 return 1;
8766         }
8767
8768         bytenr = btrfs_node_blockptr(path->nodes[level], path->slots[level]);
8769         btrfs_node_key_to_cpu(path->nodes[level], &first_key,
8770                               path->slots[level]);
8771         blocksize = fs_info->nodesize;
8772
8773         next = find_extent_buffer(fs_info, bytenr);
8774         if (!next) {
8775                 next = btrfs_find_create_tree_block(fs_info, bytenr);
8776                 if (IS_ERR(next))
8777                         return PTR_ERR(next);
8778
8779                 btrfs_set_buffer_lockdep_class(root->root_key.objectid, next,
8780                                                level - 1);
8781                 reada = 1;
8782         }
8783         btrfs_tree_lock(next);
8784         btrfs_set_lock_blocking(next);
8785
8786         ret = btrfs_lookup_extent_info(trans, fs_info, bytenr, level - 1, 1,
8787                                        &wc->refs[level - 1],
8788                                        &wc->flags[level - 1]);
8789         if (ret < 0)
8790                 goto out_unlock;
8791
8792         if (unlikely(wc->refs[level - 1] == 0)) {
8793                 btrfs_err(fs_info, "Missing references.");
8794                 ret = -EIO;
8795                 goto out_unlock;
8796         }
8797         *lookup_info = 0;
8798
8799         if (wc->stage == DROP_REFERENCE) {
8800                 if (wc->refs[level - 1] > 1) {
8801                         need_account = true;
8802                         if (level == 1 &&
8803                             (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
8804                                 goto skip;
8805
8806                         if (!wc->update_ref ||
8807                             generation <= root->root_key.offset)
8808                                 goto skip;
8809
8810                         btrfs_node_key_to_cpu(path->nodes[level], &key,
8811                                               path->slots[level]);
8812                         ret = btrfs_comp_cpu_keys(&key, &wc->update_progress);
8813                         if (ret < 0)
8814                                 goto skip;
8815
8816                         wc->stage = UPDATE_BACKREF;
8817                         wc->shared_level = level - 1;
8818                 }
8819         } else {
8820                 if (level == 1 &&
8821                     (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
8822                         goto skip;
8823         }
8824
8825         if (!btrfs_buffer_uptodate(next, generation, 0)) {
8826                 btrfs_tree_unlock(next);
8827                 free_extent_buffer(next);
8828                 next = NULL;
8829                 *lookup_info = 1;
8830         }
8831
8832         if (!next) {
8833                 if (reada && level == 1)
8834                         reada_walk_down(trans, root, wc, path);
8835                 next = read_tree_block(fs_info, bytenr, generation, level - 1,
8836                                        &first_key);
8837                 if (IS_ERR(next)) {
8838                         return PTR_ERR(next);
8839                 } else if (!extent_buffer_uptodate(next)) {
8840                         free_extent_buffer(next);
8841                         return -EIO;
8842                 }
8843                 btrfs_tree_lock(next);
8844                 btrfs_set_lock_blocking(next);
8845         }
8846
8847         level--;
8848         ASSERT(level == btrfs_header_level(next));
8849         if (level != btrfs_header_level(next)) {
8850                 btrfs_err(root->fs_info, "mismatched level");
8851                 ret = -EIO;
8852                 goto out_unlock;
8853         }
8854         path->nodes[level] = next;
8855         path->slots[level] = 0;
8856         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
8857         wc->level = level;
8858         if (wc->level == 1)
8859                 wc->reada_slot = 0;
8860         return 0;
8861 skip:
8862         wc->refs[level - 1] = 0;
8863         wc->flags[level - 1] = 0;
8864         if (wc->stage == DROP_REFERENCE) {
8865                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF) {
8866                         parent = path->nodes[level]->start;
8867                 } else {
8868                         ASSERT(root->root_key.objectid ==
8869                                btrfs_header_owner(path->nodes[level]));
8870                         if (root->root_key.objectid !=
8871                             btrfs_header_owner(path->nodes[level])) {
8872                                 btrfs_err(root->fs_info,
8873                                                 "mismatched block owner");
8874                                 ret = -EIO;
8875                                 goto out_unlock;
8876                         }
8877                         parent = 0;
8878                 }
8879
8880                 if (need_account) {
8881                         ret = btrfs_qgroup_trace_subtree(trans, root, next,
8882                                                          generation, level - 1);
8883                         if (ret) {
8884                                 btrfs_err_rl(fs_info,
8885                                              "Error %d accounting shared subtree. Quota is out of sync, rescan required.",
8886                                              ret);
8887                         }
8888                 }
8889                 ret = btrfs_free_extent(trans, root, bytenr, blocksize,
8890                                         parent, root->root_key.objectid,
8891                                         level - 1, 0);
8892                 if (ret)
8893                         goto out_unlock;
8894         }
8895
8896         *lookup_info = 1;
8897         ret = 1;
8898
8899 out_unlock:
8900         btrfs_tree_unlock(next);
8901         free_extent_buffer(next);
8902
8903         return ret;
8904 }
8905
8906 /*
8907  * helper to process tree block while walking up the tree.
8908  *
8909  * when wc->stage == DROP_REFERENCE, this function drops
8910  * reference count on the block.
8911  *
8912  * when wc->stage == UPDATE_BACKREF, this function changes
8913  * wc->stage back to DROP_REFERENCE if we changed wc->stage
8914  * to UPDATE_BACKREF previously while processing the block.
8915  *
8916  * NOTE: return value 1 means we should stop walking up.
8917  */
8918 static noinline int walk_up_proc(struct btrfs_trans_handle *trans,
8919                                  struct btrfs_root *root,
8920                                  struct btrfs_path *path,
8921                                  struct walk_control *wc)
8922 {
8923         struct btrfs_fs_info *fs_info = root->fs_info;
8924         int ret;
8925         int level = wc->level;
8926         struct extent_buffer *eb = path->nodes[level];
8927         u64 parent = 0;
8928
8929         if (wc->stage == UPDATE_BACKREF) {
8930                 BUG_ON(wc->shared_level < level);
8931                 if (level < wc->shared_level)
8932                         goto out;
8933
8934                 ret = find_next_key(path, level + 1, &wc->update_progress);
8935                 if (ret > 0)
8936                         wc->update_ref = 0;
8937
8938                 wc->stage = DROP_REFERENCE;
8939                 wc->shared_level = -1;
8940                 path->slots[level] = 0;
8941
8942                 /*
8943                  * check reference count again if the block isn't locked.
8944                  * we should start walking down the tree again if reference
8945                  * count is one.
8946                  */
8947                 if (!path->locks[level]) {
8948                         BUG_ON(level == 0);
8949                         btrfs_tree_lock(eb);
8950                         btrfs_set_lock_blocking(eb);
8951                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
8952
8953                         ret = btrfs_lookup_extent_info(trans, fs_info,
8954                                                        eb->start, level, 1,
8955                                                        &wc->refs[level],
8956                                                        &wc->flags[level]);
8957                         if (ret < 0) {
8958                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
8959                                 path->locks[level] = 0;
8960                                 return ret;
8961                         }
8962                         BUG_ON(wc->refs[level] == 0);
8963                         if (wc->refs[level] == 1) {
8964                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
8965                                 path->locks[level] = 0;
8966                                 return 1;
8967                         }
8968                 }
8969         }
8970
8971         /* wc->stage == DROP_REFERENCE */
8972         BUG_ON(wc->refs[level] > 1 && !path->locks[level]);
8973
8974         if (wc->refs[level] == 1) {
8975                 if (level == 0) {
8976                         if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
8977                                 ret = btrfs_dec_ref(trans, root, eb, 1);
8978                         else
8979                                 ret = btrfs_dec_ref(trans, root, eb, 0);
8980                         BUG_ON(ret); /* -ENOMEM */
8981                         ret = btrfs_qgroup_trace_leaf_items(trans, fs_info, eb);
8982                         if (ret) {
8983                                 btrfs_err_rl(fs_info,
8984                                              "error %d accounting leaf items. Quota is out of sync, rescan required.",
8985                                              ret);
8986                         }
8987                 }
8988                 /* make block locked assertion in clean_tree_block happy */
8989                 if (!path->locks[level] &&
8990                     btrfs_header_generation(eb) == trans->transid) {
8991                         btrfs_tree_lock(eb);
8992                         btrfs_set_lock_blocking(eb);
8993                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
8994                 }
8995                 clean_tree_block(fs_info, eb);
8996         }
8997
8998         if (eb == root->node) {
8999                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
9000                         parent = eb->start;
9001                 else
9002                         BUG_ON(root->root_key.objectid !=
9003                                btrfs_header_owner(eb));
9004         } else {
9005                 if (wc->flags[level + 1] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
9006                         parent = path->nodes[level + 1]->start;
9007                 else
9008                         BUG_ON(root->root_key.objectid !=
9009                                btrfs_header_owner(path->nodes[level + 1]));
9010         }
9011
9012         btrfs_free_tree_block(trans, root, eb, parent, wc->refs[level] == 1);
9013 out:
9014         wc->refs[level] = 0;
9015         wc->flags[level] = 0;
9016         return 0;
9017 }
9018
9019 static noinline int walk_down_tree(struct btrfs_trans_handle *trans,
9020                                    struct btrfs_root *root,
9021                                    struct btrfs_path *path,
9022                                    struct walk_control *wc)
9023 {
9024         int level = wc->level;
9025         int lookup_info = 1;
9026         int ret;
9027
9028         while (level >= 0) {
9029                 ret = walk_down_proc(trans, root, path, wc, lookup_info);
9030                 if (ret > 0)
9031                         break;
9032
9033                 if (level == 0)
9034                         break;
9035
9036                 if (path->slots[level] >=
9037                     btrfs_header_nritems(path->nodes[level]))
9038                         break;
9039
9040                 ret = do_walk_down(trans, root, path, wc, &lookup_info);
9041                 if (ret > 0) {
9042                         path->slots[level]++;
9043                         continue;
9044                 } else if (ret < 0)
9045                         return ret;
9046                 level = wc->level;
9047         }
9048         return 0;
9049 }
9050
9051 static noinline int walk_up_tree(struct btrfs_trans_handle *trans,
9052                                  struct btrfs_root *root,
9053                                  struct btrfs_path *path,
9054                                  struct walk_control *wc, int max_level)
9055 {
9056         int level = wc->level;
9057         int ret;
9058
9059         path->slots[level] = btrfs_header_nritems(path->nodes[level]);
9060         while (level < max_level && path->nodes[level]) {
9061                 wc->level = level;
9062                 if (path->slots[level] + 1 <
9063                     btrfs_header_nritems(path->nodes[level])) {
9064                         path->slots[level]++;
9065                         return 0;
9066                 } else {
9067                         ret = walk_up_proc(trans, root, path, wc);
9068                         if (ret > 0)
9069                                 return 0;
9070
9071                         if (path->locks[level]) {
9072                                 btrfs_tree_unlock_rw(path->nodes[level],
9073                                                      path->locks[level]);
9074                                 path->locks[level] = 0;
9075                         }
9076                         free_extent_buffer(path->nodes[level]);
9077                         path->nodes[level] = NULL;
9078                         level++;
9079                 }
9080         }
9081         return 1;
9082 }
9083
9084 /*
9085  * drop a subvolume tree.
9086  *
9087  * this function traverses the tree freeing any blocks that only
9088  * referenced by the tree.
9089  *
9090  * when a shared tree block is found. this function decreases its
9091  * reference count by one. if update_ref is true, this function
9092  * also make sure backrefs for the shared block and all lower level
9093  * blocks are properly updated.
9094  *
9095  * If called with for_reloc == 0, may exit early with -EAGAIN
9096  */
9097 int btrfs_drop_snapshot(struct btrfs_root *root,
9098                          struct btrfs_block_rsv *block_rsv, int update_ref,
9099                          int for_reloc)
9100 {
9101         struct btrfs_fs_info *fs_info = root->fs_info;
9102         struct btrfs_path *path;
9103         struct btrfs_trans_handle *trans;
9104         struct btrfs_root *tree_root = fs_info->tree_root;
9105         struct btrfs_root_item *root_item = &root->root_item;
9106         struct walk_control *wc;
9107         struct btrfs_key key;
9108         int err = 0;
9109         int ret;
9110         int level;
9111         bool root_dropped = false;
9112
9113         btrfs_debug(fs_info, "Drop subvolume %llu", root->objectid);
9114
9115         path = btrfs_alloc_path();
9116         if (!path) {
9117                 err = -ENOMEM;
9118                 goto out;
9119         }
9120
9121         wc = kzalloc(sizeof(*wc), GFP_NOFS);
9122         if (!wc) {
9123                 btrfs_free_path(path);
9124                 err = -ENOMEM;
9125                 goto out;
9126         }
9127
9128         trans = btrfs_start_transaction(tree_root, 0);
9129         if (IS_ERR(trans)) {
9130                 err = PTR_ERR(trans);
9131                 goto out_free;
9132         }
9133
9134         if (block_rsv)
9135                 trans->block_rsv = block_rsv;
9136
9137         if (btrfs_disk_key_objectid(&root_item->drop_progress) == 0) {
9138                 level = btrfs_header_level(root->node);
9139                 path->nodes[level] = btrfs_lock_root_node(root);
9140                 btrfs_set_lock_blocking(path->nodes[level]);
9141                 path->slots[level] = 0;
9142                 path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
9143                 memset(&wc->update_progress, 0,
9144                        sizeof(wc->update_progress));
9145         } else {
9146                 btrfs_disk_key_to_cpu(&key, &root_item->drop_progress);
9147                 memcpy(&wc->update_progress, &key,
9148                        sizeof(wc->update_progress));
9149
9150                 level = root_item->drop_level;
9151                 BUG_ON(level == 0);
9152                 path->lowest_level = level;
9153                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
9154                 path->lowest_level = 0;
9155                 if (ret < 0) {
9156                         err = ret;
9157                         goto out_end_trans;
9158                 }
9159                 WARN_ON(ret > 0);
9160
9161                 /*
9162                  * unlock our path, this is safe because only this
9163                  * function is allowed to delete this snapshot
9164                  */
9165                 btrfs_unlock_up_safe(path, 0);
9166
9167                 level = btrfs_header_level(root->node);
9168                 while (1) {
9169                         btrfs_tree_lock(path->nodes[level]);
9170                         btrfs_set_lock_blocking(path->nodes[level]);
9171                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
9172
9173                         ret = btrfs_lookup_extent_info(trans, fs_info,
9174                                                 path->nodes[level]->start,
9175                                                 level, 1, &wc->refs[level],
9176                                                 &wc->flags[level]);
9177                         if (ret < 0) {
9178                                 err = ret;
9179                                 goto out_end_trans;
9180                         }
9181                         BUG_ON(wc->refs[level] == 0);
9182
9183                         if (level == root_item->drop_level)
9184                                 break;
9185
9186                         btrfs_tree_unlock(path->nodes[level]);
9187                         path->locks[level] = 0;
9188                         WARN_ON(wc->refs[level] != 1);
9189                         level--;
9190                 }
9191         }
9192
9193         wc->level = level;
9194         wc->shared_level = -1;
9195         wc->stage = DROP_REFERENCE;
9196         wc->update_ref = update_ref;
9197         wc->keep_locks = 0;
9198         wc->for_reloc = for_reloc;
9199         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(fs_info);
9200
9201         while (1) {
9202
9203                 ret = walk_down_tree(trans, root, path, wc);
9204                 if (ret < 0) {
9205                         err = ret;
9206                         break;
9207                 }
9208
9209                 ret = walk_up_tree(trans, root, path, wc, BTRFS_MAX_LEVEL);
9210                 if (ret < 0) {
9211                         err = ret;
9212                         break;
9213                 }
9214
9215                 if (ret > 0) {
9216                         BUG_ON(wc->stage != DROP_REFERENCE);
9217                         break;
9218                 }
9219
9220                 if (wc->stage == DROP_REFERENCE) {
9221                         level = wc->level;
9222                         btrfs_node_key(path->nodes[level],
9223                                        &root_item->drop_progress,
9224                                        path->slots[level]);
9225                         root_item->drop_level = level;
9226                 }
9227
9228                 BUG_ON(wc->level == 0);
9229                 if (btrfs_should_end_transaction(trans) ||
9230                     (!for_reloc && btrfs_need_cleaner_sleep(fs_info))) {
9231                         ret = btrfs_update_root(trans, tree_root,
9232                                                 &root->root_key,
9233                                                 root_item);
9234                         if (ret) {
9235                                 btrfs_abort_transaction(trans, ret);
9236                                 err = ret;
9237                                 goto out_end_trans;
9238                         }
9239
9240                         btrfs_end_transaction_throttle(trans);
9241                         if (!for_reloc && btrfs_need_cleaner_sleep(fs_info)) {
9242                                 btrfs_debug(fs_info,
9243                                             "drop snapshot early exit");
9244                                 err = -EAGAIN;
9245                                 goto out_free;
9246                         }
9247
9248                         trans = btrfs_start_transaction(tree_root, 0);
9249                         if (IS_ERR(trans)) {
9250                                 err = PTR_ERR(trans);
9251                                 goto out_free;
9252                         }
9253                         if (block_rsv)
9254                                 trans->block_rsv = block_rsv;
9255                 }
9256         }
9257         btrfs_release_path(path);
9258         if (err)
9259                 goto out_end_trans;
9260
9261         ret = btrfs_del_root(trans, fs_info, &root->root_key);
9262         if (ret) {
9263                 btrfs_abort_transaction(trans, ret);
9264                 err = ret;
9265                 goto out_end_trans;
9266         }
9267
9268         if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) {
9269                 ret = btrfs_find_root(tree_root, &root->root_key, path,
9270                                       NULL, NULL);
9271                 if (ret < 0) {
9272                         btrfs_abort_transaction(trans, ret);
9273                         err = ret;
9274                         goto out_end_trans;
9275                 } else if (ret > 0) {
9276                         /* if we fail to delete the orphan item this time
9277                          * around, it'll get picked up the next time.
9278                          *
9279                          * The most common failure here is just -ENOENT.
9280                          */
9281                         btrfs_del_orphan_item(trans, tree_root,
9282                                               root->root_key.objectid);
9283                 }
9284         }
9285
9286         if (test_bit(BTRFS_ROOT_IN_RADIX, &root->state)) {
9287                 btrfs_add_dropped_root(trans, root);
9288         } else {
9289                 free_extent_buffer(root->node);
9290                 free_extent_buffer(root->commit_root);
9291                 btrfs_put_fs_root(root);
9292         }
9293         root_dropped = true;
9294 out_end_trans:
9295         btrfs_end_transaction_throttle(trans);
9296 out_free:
9297         kfree(wc);
9298         btrfs_free_path(path);
9299 out:
9300         /*
9301          * So if we need to stop dropping the snapshot for whatever reason we
9302          * need to make sure to add it back to the dead root list so that we
9303          * keep trying to do the work later.  This also cleans up roots if we
9304          * don't have it in the radix (like when we recover after a power fail
9305          * or unmount) so we don't leak memory.
9306          */
9307         if (!for_reloc && !root_dropped)
9308                 btrfs_add_dead_root(root);
9309         if (err && err != -EAGAIN)
9310                 btrfs_handle_fs_error(fs_info, err, NULL);
9311         return err;
9312 }
9313
9314 /*
9315  * drop subtree rooted at tree block 'node'.
9316  *
9317  * NOTE: this function will unlock and release tree block 'node'
9318  * only used by relocation code
9319  */
9320 int btrfs_drop_subtree(struct btrfs_trans_handle *trans,
9321                         struct btrfs_root *root,
9322                         struct extent_buffer *node,
9323                         struct extent_buffer *parent)
9324 {
9325         struct btrfs_fs_info *fs_info = root->fs_info;
9326         struct btrfs_path *path;
9327         struct walk_control *wc;
9328         int level;
9329         int parent_level;
9330         int ret = 0;
9331         int wret;
9332
9333         BUG_ON(root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID);
9334
9335         path = btrfs_alloc_path();
9336         if (!path)
9337                 return -ENOMEM;
9338
9339         wc = kzalloc(sizeof(*wc), GFP_NOFS);
9340         if (!wc) {
9341                 btrfs_free_path(path);
9342                 return -ENOMEM;
9343         }
9344
9345         btrfs_assert_tree_locked(parent);
9346         parent_level = btrfs_header_level(parent);
9347         extent_buffer_get(parent);
9348         path->nodes[parent_level] = parent;
9349         path->slots[parent_level] = btrfs_header_nritems(parent);
9350
9351         btrfs_assert_tree_locked(node);
9352         level = btrfs_header_level(node);
9353         path->nodes[level] = node;
9354         path->slots[level] = 0;
9355         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
9356
9357         wc->refs[parent_level] = 1;
9358         wc->flags[parent_level] = BTRFS_BLOCK_FLAG_FULL_BACKREF;
9359         wc->level = level;
9360         wc->shared_level = -1;
9361         wc->stage = DROP_REFERENCE;
9362         wc->update_ref = 0;
9363         wc->keep_locks = 1;
9364         wc->for_reloc = 1;
9365         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(fs_info);
9366
9367         while (1) {
9368                 wret = walk_down_tree(trans, root, path, wc);
9369                 if (wret < 0) {
9370                         ret = wret;
9371                         break;
9372                 }
9373
9374                 wret = walk_up_tree(trans, root, path, wc, parent_level);
9375                 if (wret < 0)
9376                         ret = wret;
9377                 if (wret != 0)
9378                         break;
9379         }
9380
9381         kfree(wc);
9382         btrfs_free_path(path);
9383         return ret;
9384 }
9385
9386 static u64 update_block_group_flags(struct btrfs_fs_info *fs_info, u64 flags)
9387 {
9388         u64 num_devices;
9389         u64 stripped;
9390
9391         /*
9392          * if restripe for this chunk_type is on pick target profile and
9393          * return, otherwise do the usual balance
9394          */
9395         stripped = get_restripe_target(fs_info, flags);
9396         if (stripped)
9397                 return extended_to_chunk(stripped);
9398
9399         num_devices = fs_info->fs_devices->rw_devices;
9400
9401         stripped = BTRFS_BLOCK_GROUP_RAID0 |
9402                 BTRFS_BLOCK_GROUP_RAID5 | BTRFS_BLOCK_GROUP_RAID6 |
9403                 BTRFS_BLOCK_GROUP_RAID1 | BTRFS_BLOCK_GROUP_RAID10;
9404
9405         if (num_devices == 1) {
9406                 stripped |= BTRFS_BLOCK_GROUP_DUP;
9407                 stripped = flags & ~stripped;
9408
9409                 /* turn raid0 into single device chunks */
9410                 if (flags & BTRFS_BLOCK_GROUP_RAID0)
9411                         return stripped;
9412
9413                 /* turn mirroring into duplication */
9414                 if (flags & (BTRFS_BLOCK_GROUP_RAID1 |
9415                              BTRFS_BLOCK_GROUP_RAID10))
9416                         return stripped | BTRFS_BLOCK_GROUP_DUP;
9417         } else {
9418                 /* they already had raid on here, just return */
9419                 if (flags & stripped)
9420                         return flags;
9421
9422                 stripped |= BTRFS_BLOCK_GROUP_DUP;
9423                 stripped = flags & ~stripped;
9424
9425                 /* switch duplicated blocks with raid1 */
9426                 if (flags & BTRFS_BLOCK_GROUP_DUP)
9427                         return stripped | BTRFS_BLOCK_GROUP_RAID1;
9428
9429                 /* this is drive concat, leave it alone */
9430         }
9431
9432         return flags;
9433 }
9434
9435 static int inc_block_group_ro(struct btrfs_block_group_cache *cache, int force)
9436 {
9437         struct btrfs_space_info *sinfo = cache->space_info;
9438         u64 num_bytes;
9439         u64 min_allocable_bytes;
9440         int ret = -ENOSPC;
9441
9442         /*
9443          * We need some metadata space and system metadata space for
9444          * allocating chunks in some corner cases until we force to set
9445          * it to be readonly.
9446          */
9447         if ((sinfo->flags &
9448              (BTRFS_BLOCK_GROUP_SYSTEM | BTRFS_BLOCK_GROUP_METADATA)) &&
9449             !force)
9450                 min_allocable_bytes = SZ_1M;
9451         else
9452                 min_allocable_bytes = 0;
9453
9454         spin_lock(&sinfo->lock);
9455         spin_lock(&cache->lock);
9456
9457         if (cache->ro) {
9458                 cache->ro++;
9459                 ret = 0;
9460                 goto out;
9461         }
9462
9463         num_bytes = cache->key.offset - cache->reserved - cache->pinned -
9464                     cache->bytes_super - btrfs_block_group_used(&cache->item);
9465
9466         if (btrfs_space_info_used(sinfo, true) + num_bytes +
9467             min_allocable_bytes <= sinfo->total_bytes) {
9468                 sinfo->bytes_readonly += num_bytes;
9469                 cache->ro++;
9470                 list_add_tail(&cache->ro_list, &sinfo->ro_bgs);
9471                 ret = 0;
9472         }
9473 out:
9474         spin_unlock(&cache->lock);
9475         spin_unlock(&sinfo->lock);
9476         return ret;
9477 }
9478
9479 int btrfs_inc_block_group_ro(struct btrfs_fs_info *fs_info,
9480                              struct btrfs_block_group_cache *cache)
9481
9482 {
9483         struct btrfs_trans_handle *trans;
9484         u64 alloc_flags;
9485         int ret;
9486
9487 again:
9488         trans = btrfs_join_transaction(fs_info->extent_root);
9489         if (IS_ERR(trans))
9490                 return PTR_ERR(trans);
9491
9492         /*
9493          * we're not allowed to set block groups readonly after the dirty
9494          * block groups cache has started writing.  If it already started,
9495          * back off and let this transaction commit
9496          */
9497         mutex_lock(&fs_info->ro_block_group_mutex);
9498         if (test_bit(BTRFS_TRANS_DIRTY_BG_RUN, &trans->transaction->flags)) {
9499                 u64 transid = trans->transid;
9500
9501                 mutex_unlock(&fs_info->ro_block_group_mutex);
9502                 btrfs_end_transaction(trans);
9503
9504                 ret = btrfs_wait_for_commit(fs_info, transid);
9505                 if (ret)
9506                         return ret;
9507                 goto again;
9508         }
9509
9510         /*
9511          * if we are changing raid levels, try to allocate a corresponding
9512          * block group with the new raid level.
9513          */
9514         alloc_flags = update_block_group_flags(fs_info, cache->flags);
9515         if (alloc_flags != cache->flags) {
9516                 ret = do_chunk_alloc(trans, fs_info, alloc_flags,
9517                                      CHUNK_ALLOC_FORCE);
9518                 /*
9519                  * ENOSPC is allowed here, we may have enough space
9520                  * already allocated at the new raid level to
9521                  * carry on
9522                  */
9523                 if (ret == -ENOSPC)
9524                         ret = 0;
9525                 if (ret < 0)
9526                         goto out;
9527         }
9528
9529         ret = inc_block_group_ro(cache, 0);
9530         if (!ret)
9531                 goto out;
9532         alloc_flags = get_alloc_profile(fs_info, cache->space_info->flags);
9533         ret = do_chunk_alloc(trans, fs_info, alloc_flags,
9534                              CHUNK_ALLOC_FORCE);
9535         if (ret < 0)
9536                 goto out;
9537         ret = inc_block_group_ro(cache, 0);
9538 out:
9539         if (cache->flags & BTRFS_BLOCK_GROUP_SYSTEM) {
9540                 alloc_flags = update_block_group_flags(fs_info, cache->flags);
9541                 mutex_lock(&fs_info->chunk_mutex);
9542                 check_system_chunk(trans, fs_info, alloc_flags);
9543                 mutex_unlock(&fs_info->chunk_mutex);
9544         }
9545         mutex_unlock(&fs_info->ro_block_group_mutex);
9546
9547         btrfs_end_transaction(trans);
9548         return ret;
9549 }
9550
9551 int btrfs_force_chunk_alloc(struct btrfs_trans_handle *trans,
9552                             struct btrfs_fs_info *fs_info, u64 type)
9553 {
9554         u64 alloc_flags = get_alloc_profile(fs_info, type);
9555
9556         return do_chunk_alloc(trans, fs_info, alloc_flags, CHUNK_ALLOC_FORCE);
9557 }
9558
9559 /*
9560  * helper to account the unused space of all the readonly block group in the
9561  * space_info. takes mirrors into account.
9562  */
9563 u64 btrfs_account_ro_block_groups_free_space(struct btrfs_space_info *sinfo)
9564 {
9565         struct btrfs_block_group_cache *block_group;
9566         u64 free_bytes = 0;
9567         int factor;
9568
9569         /* It's df, we don't care if it's racy */
9570         if (list_empty(&sinfo->ro_bgs))
9571                 return 0;
9572
9573         spin_lock(&sinfo->lock);
9574         list_for_each_entry(block_group, &sinfo->ro_bgs, ro_list) {
9575                 spin_lock(&block_group->lock);
9576
9577                 if (!block_group->ro) {
9578                         spin_unlock(&block_group->lock);
9579                         continue;
9580                 }
9581
9582                 if (block_group->flags & (BTRFS_BLOCK_GROUP_RAID1 |
9583                                           BTRFS_BLOCK_GROUP_RAID10 |
9584                                           BTRFS_BLOCK_GROUP_DUP))
9585                         factor = 2;
9586                 else
9587                         factor = 1;
9588
9589                 free_bytes += (block_group->key.offset -
9590                                btrfs_block_group_used(&block_group->item)) *
9591                                factor;
9592
9593                 spin_unlock(&block_group->lock);
9594         }
9595         spin_unlock(&sinfo->lock);
9596
9597         return free_bytes;
9598 }
9599
9600 void btrfs_dec_block_group_ro(struct btrfs_block_group_cache *cache)
9601 {
9602         struct btrfs_space_info *sinfo = cache->space_info;
9603         u64 num_bytes;
9604
9605         BUG_ON(!cache->ro);
9606
9607         spin_lock(&sinfo->lock);
9608         spin_lock(&cache->lock);
9609         if (!--cache->ro) {
9610                 num_bytes = cache->key.offset - cache->reserved -
9611                             cache->pinned - cache->bytes_super -
9612                             btrfs_block_group_used(&cache->item);
9613                 sinfo->bytes_readonly -= num_bytes;
9614                 list_del_init(&cache->ro_list);
9615         }
9616         spin_unlock(&cache->lock);
9617         spin_unlock(&sinfo->lock);
9618 }
9619
9620 /*
9621  * checks to see if its even possible to relocate this block group.
9622  *
9623  * @return - -1 if it's not a good idea to relocate this block group, 0 if its
9624  * ok to go ahead and try.
9625  */
9626 int btrfs_can_relocate(struct btrfs_fs_info *fs_info, u64 bytenr)
9627 {
9628         struct btrfs_root *root = fs_info->extent_root;
9629         struct btrfs_block_group_cache *block_group;
9630         struct btrfs_space_info *space_info;
9631         struct btrfs_fs_devices *fs_devices = fs_info->fs_devices;
9632         struct btrfs_device *device;
9633         struct btrfs_trans_handle *trans;
9634         u64 min_free;
9635         u64 dev_min = 1;
9636         u64 dev_nr = 0;
9637         u64 target;
9638         int debug;
9639         int index;
9640         int full = 0;
9641         int ret = 0;
9642
9643         debug = btrfs_test_opt(fs_info, ENOSPC_DEBUG);
9644
9645         block_group = btrfs_lookup_block_group(fs_info, bytenr);
9646
9647         /* odd, couldn't find the block group, leave it alone */
9648         if (!block_group) {
9649                 if (debug)
9650                         btrfs_warn(fs_info,
9651                                    "can't find block group for bytenr %llu",
9652                                    bytenr);
9653                 return -1;
9654         }
9655
9656         min_free = btrfs_block_group_used(&block_group->item);
9657
9658         /* no bytes used, we're good */
9659         if (!min_free)
9660                 goto out;
9661
9662         space_info = block_group->space_info;
9663         spin_lock(&space_info->lock);
9664
9665         full = space_info->full;
9666
9667         /*
9668          * if this is the last block group we have in this space, we can't
9669          * relocate it unless we're able to allocate a new chunk below.
9670          *
9671          * Otherwise, we need to make sure we have room in the space to handle
9672          * all of the extents from this block group.  If we can, we're good
9673          */
9674         if ((space_info->total_bytes != block_group->key.offset) &&
9675             (btrfs_space_info_used(space_info, false) + min_free <
9676              space_info->total_bytes)) {
9677                 spin_unlock(&space_info->lock);
9678                 goto out;
9679         }
9680         spin_unlock(&space_info->lock);
9681
9682         /*
9683          * ok we don't have enough space, but maybe we have free space on our
9684          * devices to allocate new chunks for relocation, so loop through our
9685          * alloc devices and guess if we have enough space.  if this block
9686          * group is going to be restriped, run checks against the target
9687          * profile instead of the current one.
9688          */
9689         ret = -1;
9690
9691         /*
9692          * index:
9693          *      0: raid10
9694          *      1: raid1
9695          *      2: dup
9696          *      3: raid0
9697          *      4: single
9698          */
9699         target = get_restripe_target(fs_info, block_group->flags);
9700         if (target) {
9701                 index = btrfs_bg_flags_to_raid_index(extended_to_chunk(target));
9702         } else {
9703                 /*
9704                  * this is just a balance, so if we were marked as full
9705                  * we know there is no space for a new chunk
9706                  */
9707                 if (full) {
9708                         if (debug)
9709                                 btrfs_warn(fs_info,
9710                                            "no space to alloc new chunk for block group %llu",
9711                                            block_group->key.objectid);
9712                         goto out;
9713                 }
9714
9715                 index = btrfs_bg_flags_to_raid_index(block_group->flags);
9716         }
9717
9718         if (index == BTRFS_RAID_RAID10) {
9719                 dev_min = 4;
9720                 /* Divide by 2 */
9721                 min_free >>= 1;
9722         } else if (index == BTRFS_RAID_RAID1) {
9723                 dev_min = 2;
9724         } else if (index == BTRFS_RAID_DUP) {
9725                 /* Multiply by 2 */
9726                 min_free <<= 1;
9727         } else if (index == BTRFS_RAID_RAID0) {
9728                 dev_min = fs_devices->rw_devices;
9729                 min_free = div64_u64(min_free, dev_min);
9730         }
9731
9732         /* We need to do this so that we can look at pending chunks */
9733         trans = btrfs_join_transaction(root);
9734         if (IS_ERR(trans)) {
9735                 ret = PTR_ERR(trans);
9736                 goto out;
9737         }
9738
9739         mutex_lock(&fs_info->chunk_mutex);
9740         list_for_each_entry(device, &fs_devices->alloc_list, dev_alloc_list) {
9741                 u64 dev_offset;
9742
9743                 /*
9744                  * check to make sure we can actually find a chunk with enough
9745                  * space to fit our block group in.
9746                  */
9747                 if (device->total_bytes > device->bytes_used + min_free &&
9748                     !test_bit(BTRFS_DEV_STATE_REPLACE_TGT, &device->dev_state)) {
9749                         ret = find_free_dev_extent(trans, device, min_free,
9750                                                    &dev_offset, NULL);
9751                         if (!ret)
9752                                 dev_nr++;
9753
9754                         if (dev_nr >= dev_min)
9755                                 break;
9756
9757                         ret = -1;
9758                 }
9759         }
9760         if (debug && ret == -1)
9761                 btrfs_warn(fs_info,
9762                            "no space to allocate a new chunk for block group %llu",
9763                            block_group->key.objectid);
9764         mutex_unlock(&fs_info->chunk_mutex);
9765         btrfs_end_transaction(trans);
9766 out:
9767         btrfs_put_block_group(block_group);
9768         return ret;
9769 }
9770
9771 static int find_first_block_group(struct btrfs_fs_info *fs_info,
9772                                   struct btrfs_path *path,
9773                                   struct btrfs_key *key)
9774 {
9775         struct btrfs_root *root = fs_info->extent_root;
9776         int ret = 0;
9777         struct btrfs_key found_key;
9778         struct extent_buffer *leaf;
9779         int slot;
9780
9781         ret = btrfs_search_slot(NULL, root, key, path, 0, 0);
9782         if (ret < 0)
9783                 goto out;
9784
9785         while (1) {
9786                 slot = path->slots[0];
9787                 leaf = path->nodes[0];
9788                 if (slot >= btrfs_header_nritems(leaf)) {
9789                         ret = btrfs_next_leaf(root, path);
9790                         if (ret == 0)
9791                                 continue;
9792                         if (ret < 0)
9793                                 goto out;
9794                         break;
9795                 }
9796                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
9797
9798                 if (found_key.objectid >= key->objectid &&
9799                     found_key.type == BTRFS_BLOCK_GROUP_ITEM_KEY) {
9800                         struct extent_map_tree *em_tree;
9801                         struct extent_map *em;
9802
9803                         em_tree = &root->fs_info->mapping_tree.map_tree;
9804                         read_lock(&em_tree->lock);
9805                         em = lookup_extent_mapping(em_tree, found_key.objectid,
9806                                                    found_key.offset);
9807                         read_unlock(&em_tree->lock);
9808                         if (!em) {
9809                                 btrfs_err(fs_info,
9810                         "logical %llu len %llu found bg but no related chunk",
9811                                           found_key.objectid, found_key.offset);
9812                                 ret = -ENOENT;
9813                         } else {
9814                                 ret = 0;
9815                         }
9816                         free_extent_map(em);
9817                         goto out;
9818                 }
9819                 path->slots[0]++;
9820         }
9821 out:
9822         return ret;
9823 }
9824
9825 void btrfs_put_block_group_cache(struct btrfs_fs_info *info)
9826 {
9827         struct btrfs_block_group_cache *block_group;
9828         u64 last = 0;
9829
9830         while (1) {
9831                 struct inode *inode;
9832
9833                 block_group = btrfs_lookup_first_block_group(info, last);
9834                 while (block_group) {
9835                         spin_lock(&block_group->lock);
9836                         if (block_group->iref)
9837                                 break;
9838                         spin_unlock(&block_group->lock);
9839                         block_group = next_block_group(info, block_group);
9840                 }
9841                 if (!block_group) {
9842                         if (last == 0)
9843                                 break;
9844                         last = 0;
9845                         continue;
9846                 }
9847
9848                 inode = block_group->inode;
9849                 block_group->iref = 0;
9850                 block_group->inode = NULL;
9851                 spin_unlock(&block_group->lock);
9852                 ASSERT(block_group->io_ctl.inode == NULL);
9853                 iput(inode);
9854                 last = block_group->key.objectid + block_group->key.offset;
9855                 btrfs_put_block_group(block_group);
9856         }
9857 }
9858
9859 /*
9860  * Must be called only after stopping all workers, since we could have block
9861  * group caching kthreads running, and therefore they could race with us if we
9862  * freed the block groups before stopping them.
9863  */
9864 int btrfs_free_block_groups(struct btrfs_fs_info *info)
9865 {
9866         struct btrfs_block_group_cache *block_group;
9867         struct btrfs_space_info *space_info;
9868         struct btrfs_caching_control *caching_ctl;
9869         struct rb_node *n;
9870
9871         down_write(&info->commit_root_sem);
9872         while (!list_empty(&info->caching_block_groups)) {
9873                 caching_ctl = list_entry(info->caching_block_groups.next,
9874                                          struct btrfs_caching_control, list);
9875                 list_del(&caching_ctl->list);
9876                 put_caching_control(caching_ctl);
9877         }
9878         up_write(&info->commit_root_sem);
9879
9880         spin_lock(&info->unused_bgs_lock);
9881         while (!list_empty(&info->unused_bgs)) {
9882                 block_group = list_first_entry(&info->unused_bgs,
9883                                                struct btrfs_block_group_cache,
9884                                                bg_list);
9885                 list_del_init(&block_group->bg_list);
9886                 btrfs_put_block_group(block_group);
9887         }
9888         spin_unlock(&info->unused_bgs_lock);
9889
9890         spin_lock(&info->block_group_cache_lock);
9891         while ((n = rb_last(&info->block_group_cache_tree)) != NULL) {
9892                 block_group = rb_entry(n, struct btrfs_block_group_cache,
9893                                        cache_node);
9894                 rb_erase(&block_group->cache_node,
9895                          &info->block_group_cache_tree);
9896                 RB_CLEAR_NODE(&block_group->cache_node);
9897                 spin_unlock(&info->block_group_cache_lock);
9898
9899                 down_write(&block_group->space_info->groups_sem);
9900                 list_del(&block_group->list);
9901                 up_write(&block_group->space_info->groups_sem);
9902
9903                 /*
9904                  * We haven't cached this block group, which means we could
9905                  * possibly have excluded extents on this block group.
9906                  */
9907                 if (block_group->cached == BTRFS_CACHE_NO ||
9908                     block_group->cached == BTRFS_CACHE_ERROR)
9909                         free_excluded_extents(info, block_group);
9910
9911                 btrfs_remove_free_space_cache(block_group);
9912                 ASSERT(block_group->cached != BTRFS_CACHE_STARTED);
9913                 ASSERT(list_empty(&block_group->dirty_list));
9914                 ASSERT(list_empty(&block_group->io_list));
9915                 ASSERT(list_empty(&block_group->bg_list));
9916                 ASSERT(atomic_read(&block_group->count) == 1);
9917                 btrfs_put_block_group(block_group);
9918
9919                 spin_lock(&info->block_group_cache_lock);
9920         }
9921         spin_unlock(&info->block_group_cache_lock);
9922
9923         /* now that all the block groups are freed, go through and
9924          * free all the space_info structs.  This is only called during
9925          * the final stages of unmount, and so we know nobody is
9926          * using them.  We call synchronize_rcu() once before we start,
9927          * just to be on the safe side.
9928          */
9929         synchronize_rcu();
9930
9931         release_global_block_rsv(info);
9932
9933         while (!list_empty(&info->space_info)) {
9934                 int i;
9935
9936                 space_info = list_entry(info->space_info.next,
9937                                         struct btrfs_space_info,
9938                                         list);
9939
9940                 /*
9941                  * Do not hide this behind enospc_debug, this is actually
9942                  * important and indicates a real bug if this happens.
9943                  */
9944                 if (WARN_ON(space_info->bytes_pinned > 0 ||
9945                             space_info->bytes_reserved > 0 ||
9946                             space_info->bytes_may_use > 0))
9947                         dump_space_info(info, space_info, 0, 0);
9948                 list_del(&space_info->list);
9949                 for (i = 0; i < BTRFS_NR_RAID_TYPES; i++) {
9950                         struct kobject *kobj;
9951                         kobj = space_info->block_group_kobjs[i];
9952                         space_info->block_group_kobjs[i] = NULL;
9953                         if (kobj) {
9954                                 kobject_del(kobj);
9955                                 kobject_put(kobj);
9956                         }
9957                 }
9958                 kobject_del(&space_info->kobj);
9959                 kobject_put(&space_info->kobj);
9960         }
9961         return 0;
9962 }
9963
9964 /* link_block_group will queue up kobjects to add when we're reclaim-safe */
9965 void btrfs_add_raid_kobjects(struct btrfs_fs_info *fs_info)
9966 {
9967         struct btrfs_space_info *space_info;
9968         struct raid_kobject *rkobj;
9969         LIST_HEAD(list);
9970         int index;
9971         int ret = 0;
9972
9973         spin_lock(&fs_info->pending_raid_kobjs_lock);
9974         list_splice_init(&fs_info->pending_raid_kobjs, &list);
9975         spin_unlock(&fs_info->pending_raid_kobjs_lock);
9976
9977         list_for_each_entry(rkobj, &list, list) {
9978                 space_info = __find_space_info(fs_info, rkobj->flags);
9979                 index = btrfs_bg_flags_to_raid_index(rkobj->flags);
9980
9981                 ret = kobject_add(&rkobj->kobj, &space_info->kobj,
9982                                   "%s", get_raid_name(index));
9983                 if (ret) {
9984                         kobject_put(&rkobj->kobj);
9985                         break;
9986                 }
9987         }
9988         if (ret)
9989                 btrfs_warn(fs_info,
9990                            "failed to add kobject for block cache, ignoring");
9991 }
9992
9993 static void link_block_group(struct btrfs_block_group_cache *cache)
9994 {
9995         struct btrfs_space_info *space_info = cache->space_info;
9996         struct btrfs_fs_info *fs_info = cache->fs_info;
9997         int index = btrfs_bg_flags_to_raid_index(cache->flags);
9998         bool first = false;
9999
10000         down_write(&space_info->groups_sem);
10001         if (list_empty(&space_info->block_groups[index]))
10002                 first = true;
10003         list_add_tail(&cache->list, &space_info->block_groups[index]);
10004         up_write(&space_info->groups_sem);
10005
10006         if (first) {
10007                 struct raid_kobject *rkobj = kzalloc(sizeof(*rkobj), GFP_NOFS);
10008                 if (!rkobj) {
10009                         btrfs_warn(cache->fs_info,
10010                                 "couldn't alloc memory for raid level kobject");
10011                         return;
10012                 }
10013                 rkobj->flags = cache->flags;
10014                 kobject_init(&rkobj->kobj, &btrfs_raid_ktype);
10015
10016                 spin_lock(&fs_info->pending_raid_kobjs_lock);
10017                 list_add_tail(&rkobj->list, &fs_info->pending_raid_kobjs);
10018                 spin_unlock(&fs_info->pending_raid_kobjs_lock);
10019                 space_info->block_group_kobjs[index] = &rkobj->kobj;
10020         }
10021 }
10022
10023 static struct btrfs_block_group_cache *
10024 btrfs_create_block_group_cache(struct btrfs_fs_info *fs_info,
10025                                u64 start, u64 size)
10026 {
10027         struct btrfs_block_group_cache *cache;
10028
10029         cache = kzalloc(sizeof(*cache), GFP_NOFS);
10030         if (!cache)
10031                 return NULL;
10032
10033         cache->free_space_ctl = kzalloc(sizeof(*cache->free_space_ctl),
10034                                         GFP_NOFS);
10035         if (!cache->free_space_ctl) {
10036                 kfree(cache);
10037                 return NULL;
10038         }
10039
10040         cache->key.objectid = start;
10041         cache->key.offset = size;
10042         cache->key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
10043
10044         cache->fs_info = fs_info;
10045         cache->full_stripe_len = btrfs_full_stripe_len(fs_info, start);
10046         set_free_space_tree_thresholds(cache);
10047
10048         atomic_set(&cache->count, 1);
10049         spin_lock_init(&cache->lock);
10050         init_rwsem(&cache->data_rwsem);
10051         INIT_LIST_HEAD(&cache->list);
10052         INIT_LIST_HEAD(&cache->cluster_list);
10053         INIT_LIST_HEAD(&cache->bg_list);
10054         INIT_LIST_HEAD(&cache->ro_list);
10055         INIT_LIST_HEAD(&cache->dirty_list);
10056         INIT_LIST_HEAD(&cache->io_list);
10057         btrfs_init_free_space_ctl(cache);
10058         atomic_set(&cache->trimming, 0);
10059         mutex_init(&cache->free_space_lock);
10060         btrfs_init_full_stripe_locks_tree(&cache->full_stripe_locks_root);
10061
10062         return cache;
10063 }
10064
10065 int btrfs_read_block_groups(struct btrfs_fs_info *info)
10066 {
10067         struct btrfs_path *path;
10068         int ret;
10069         struct btrfs_block_group_cache *cache;
10070         struct btrfs_space_info *space_info;
10071         struct btrfs_key key;
10072         struct btrfs_key found_key;
10073         struct extent_buffer *leaf;
10074         int need_clear = 0;
10075         u64 cache_gen;
10076         u64 feature;
10077         int mixed;
10078
10079         feature = btrfs_super_incompat_flags(info->super_copy);
10080         mixed = !!(feature & BTRFS_FEATURE_INCOMPAT_MIXED_GROUPS);
10081
10082         key.objectid = 0;
10083         key.offset = 0;
10084         key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
10085         path = btrfs_alloc_path();
10086         if (!path)
10087                 return -ENOMEM;
10088         path->reada = READA_FORWARD;
10089
10090         cache_gen = btrfs_super_cache_generation(info->super_copy);
10091         if (btrfs_test_opt(info, SPACE_CACHE) &&
10092             btrfs_super_generation(info->super_copy) != cache_gen)
10093                 need_clear = 1;
10094         if (btrfs_test_opt(info, CLEAR_CACHE))
10095                 need_clear = 1;
10096
10097         while (1) {
10098                 ret = find_first_block_group(info, path, &key);
10099                 if (ret > 0)
10100                         break;
10101                 if (ret != 0)
10102                         goto error;
10103
10104                 leaf = path->nodes[0];
10105                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
10106
10107                 cache = btrfs_create_block_group_cache(info, found_key.objectid,
10108                                                        found_key.offset);
10109                 if (!cache) {
10110                         ret = -ENOMEM;
10111                         goto error;
10112                 }
10113
10114                 if (need_clear) {
10115                         /*
10116                          * When we mount with old space cache, we need to
10117                          * set BTRFS_DC_CLEAR and set dirty flag.
10118                          *
10119                          * a) Setting 'BTRFS_DC_CLEAR' makes sure that we
10120                          *    truncate the old free space cache inode and
10121                          *    setup a new one.
10122                          * b) Setting 'dirty flag' makes sure that we flush
10123                          *    the new space cache info onto disk.
10124                          */
10125                         if (btrfs_test_opt(info, SPACE_CACHE))
10126                                 cache->disk_cache_state = BTRFS_DC_CLEAR;
10127                 }
10128
10129                 read_extent_buffer(leaf, &cache->item,
10130                                    btrfs_item_ptr_offset(leaf, path->slots[0]),
10131                                    sizeof(cache->item));
10132                 cache->flags = btrfs_block_group_flags(&cache->item);
10133                 if (!mixed &&
10134                     ((cache->flags & BTRFS_BLOCK_GROUP_METADATA) &&
10135                     (cache->flags & BTRFS_BLOCK_GROUP_DATA))) {
10136                         btrfs_err(info,
10137 "bg %llu is a mixed block group but filesystem hasn't enabled mixed block groups",
10138                                   cache->key.objectid);
10139                         ret = -EINVAL;
10140                         goto error;
10141                 }
10142
10143                 key.objectid = found_key.objectid + found_key.offset;
10144                 btrfs_release_path(path);
10145
10146                 /*
10147                  * We need to exclude the super stripes now so that the space
10148                  * info has super bytes accounted for, otherwise we'll think
10149                  * we have more space than we actually do.
10150                  */
10151                 ret = exclude_super_stripes(info, cache);
10152                 if (ret) {
10153                         /*
10154                          * We may have excluded something, so call this just in
10155                          * case.
10156                          */
10157                         free_excluded_extents(info, cache);
10158                         btrfs_put_block_group(cache);
10159                         goto error;
10160                 }
10161
10162                 /*
10163                  * check for two cases, either we are full, and therefore
10164                  * don't need to bother with the caching work since we won't
10165                  * find any space, or we are empty, and we can just add all
10166                  * the space in and be done with it.  This saves us _alot_ of
10167                  * time, particularly in the full case.
10168                  */
10169                 if (found_key.offset == btrfs_block_group_used(&cache->item)) {
10170                         cache->last_byte_to_unpin = (u64)-1;
10171                         cache->cached = BTRFS_CACHE_FINISHED;
10172                         free_excluded_extents(info, cache);
10173                 } else if (btrfs_block_group_used(&cache->item) == 0) {
10174                         cache->last_byte_to_unpin = (u64)-1;
10175                         cache->cached = BTRFS_CACHE_FINISHED;
10176                         add_new_free_space(cache, info,
10177                                            found_key.objectid,
10178                                            found_key.objectid +
10179                                            found_key.offset);
10180                         free_excluded_extents(info, cache);
10181                 }
10182
10183                 ret = btrfs_add_block_group_cache(info, cache);
10184                 if (ret) {
10185                         btrfs_remove_free_space_cache(cache);
10186                         btrfs_put_block_group(cache);
10187                         goto error;
10188                 }
10189
10190                 trace_btrfs_add_block_group(info, cache, 0);
10191                 update_space_info(info, cache->flags, found_key.offset,
10192                                   btrfs_block_group_used(&cache->item),
10193                                   cache->bytes_super, &space_info);
10194
10195                 cache->space_info = space_info;
10196
10197                 link_block_group(cache);
10198
10199                 set_avail_alloc_bits(info, cache->flags);
10200                 if (btrfs_chunk_readonly(info, cache->key.objectid)) {
10201                         inc_block_group_ro(cache, 1);
10202                 } else if (btrfs_block_group_used(&cache->item) == 0) {
10203                         spin_lock(&info->unused_bgs_lock);
10204                         /* Should always be true but just in case. */
10205                         if (list_empty(&cache->bg_list)) {
10206                                 btrfs_get_block_group(cache);
10207                                 list_add_tail(&cache->bg_list,
10208                                               &info->unused_bgs);
10209                         }
10210                         spin_unlock(&info->unused_bgs_lock);
10211                 }
10212         }
10213
10214         list_for_each_entry_rcu(space_info, &info->space_info, list) {
10215                 if (!(get_alloc_profile(info, space_info->flags) &
10216                       (BTRFS_BLOCK_GROUP_RAID10 |
10217                        BTRFS_BLOCK_GROUP_RAID1 |
10218                        BTRFS_BLOCK_GROUP_RAID5 |
10219                        BTRFS_BLOCK_GROUP_RAID6 |
10220                        BTRFS_BLOCK_GROUP_DUP)))
10221                         continue;
10222                 /*
10223                  * avoid allocating from un-mirrored block group if there are
10224                  * mirrored block groups.
10225                  */
10226                 list_for_each_entry(cache,
10227                                 &space_info->block_groups[BTRFS_RAID_RAID0],
10228                                 list)
10229                         inc_block_group_ro(cache, 1);
10230                 list_for_each_entry(cache,
10231                                 &space_info->block_groups[BTRFS_RAID_SINGLE],
10232                                 list)
10233                         inc_block_group_ro(cache, 1);
10234         }
10235
10236         btrfs_add_raid_kobjects(info);
10237         init_global_block_rsv(info);
10238         ret = 0;
10239 error:
10240         btrfs_free_path(path);
10241         return ret;
10242 }
10243
10244 void btrfs_create_pending_block_groups(struct btrfs_trans_handle *trans)
10245 {
10246         struct btrfs_fs_info *fs_info = trans->fs_info;
10247         struct btrfs_block_group_cache *block_group, *tmp;
10248         struct btrfs_root *extent_root = fs_info->extent_root;
10249         struct btrfs_block_group_item item;
10250         struct btrfs_key key;
10251         int ret = 0;
10252         bool can_flush_pending_bgs = trans->can_flush_pending_bgs;
10253
10254         trans->can_flush_pending_bgs = false;
10255         list_for_each_entry_safe(block_group, tmp, &trans->new_bgs, bg_list) {
10256                 if (ret)
10257                         goto next;
10258
10259                 spin_lock(&block_group->lock);
10260                 memcpy(&item, &block_group->item, sizeof(item));
10261                 memcpy(&key, &block_group->key, sizeof(key));
10262                 spin_unlock(&block_group->lock);
10263
10264                 ret = btrfs_insert_item(trans, extent_root, &key, &item,
10265                                         sizeof(item));
10266                 if (ret)
10267                         btrfs_abort_transaction(trans, ret);
10268                 ret = btrfs_finish_chunk_alloc(trans, fs_info, key.objectid,
10269                                                key.offset);
10270                 if (ret)
10271                         btrfs_abort_transaction(trans, ret);
10272                 add_block_group_free_space(trans, fs_info, block_group);
10273                 /* already aborted the transaction if it failed. */
10274 next:
10275                 list_del_init(&block_group->bg_list);
10276         }
10277         trans->can_flush_pending_bgs = can_flush_pending_bgs;
10278 }
10279
10280 int btrfs_make_block_group(struct btrfs_trans_handle *trans,
10281                            struct btrfs_fs_info *fs_info, u64 bytes_used,
10282                            u64 type, u64 chunk_offset, u64 size)
10283 {
10284         struct btrfs_block_group_cache *cache;
10285         int ret;
10286
10287         btrfs_set_log_full_commit(fs_info, trans);
10288
10289         cache = btrfs_create_block_group_cache(fs_info, chunk_offset, size);
10290         if (!cache)
10291                 return -ENOMEM;
10292
10293         btrfs_set_block_group_used(&cache->item, bytes_used);
10294         btrfs_set_block_group_chunk_objectid(&cache->item,
10295                                              BTRFS_FIRST_CHUNK_TREE_OBJECTID);
10296         btrfs_set_block_group_flags(&cache->item, type);
10297
10298         cache->flags = type;
10299         cache->last_byte_to_unpin = (u64)-1;
10300         cache->cached = BTRFS_CACHE_FINISHED;
10301         cache->needs_free_space = 1;
10302         ret = exclude_super_stripes(fs_info, cache);
10303         if (ret) {
10304                 /*
10305                  * We may have excluded something, so call this just in
10306                  * case.
10307                  */
10308                 free_excluded_extents(fs_info, cache);
10309                 btrfs_put_block_group(cache);
10310                 return ret;
10311         }
10312
10313         add_new_free_space(cache, fs_info, chunk_offset, chunk_offset + size);
10314
10315         free_excluded_extents(fs_info, cache);
10316
10317 #ifdef CONFIG_BTRFS_DEBUG
10318         if (btrfs_should_fragment_free_space(cache)) {
10319                 u64 new_bytes_used = size - bytes_used;
10320
10321                 bytes_used += new_bytes_used >> 1;
10322                 fragment_free_space(cache);
10323         }
10324 #endif
10325         /*
10326          * Ensure the corresponding space_info object is created and
10327          * assigned to our block group. We want our bg to be added to the rbtree
10328          * with its ->space_info set.
10329          */
10330         cache->space_info = __find_space_info(fs_info, cache->flags);
10331         ASSERT(cache->space_info);
10332
10333         ret = btrfs_add_block_group_cache(fs_info, cache);
10334         if (ret) {
10335                 btrfs_remove_free_space_cache(cache);
10336                 btrfs_put_block_group(cache);
10337                 return ret;
10338         }
10339
10340         /*
10341          * Now that our block group has its ->space_info set and is inserted in
10342          * the rbtree, update the space info's counters.
10343          */
10344         trace_btrfs_add_block_group(fs_info, cache, 1);
10345         update_space_info(fs_info, cache->flags, size, bytes_used,
10346                                 cache->bytes_super, &cache->space_info);
10347         update_global_block_rsv(fs_info);
10348
10349         link_block_group(cache);
10350
10351         list_add_tail(&cache->bg_list, &trans->new_bgs);
10352
10353         set_avail_alloc_bits(fs_info, type);
10354         return 0;
10355 }
10356
10357 static void clear_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
10358 {
10359         u64 extra_flags = chunk_to_extended(flags) &
10360                                 BTRFS_EXTENDED_PROFILE_MASK;
10361
10362         write_seqlock(&fs_info->profiles_lock);
10363         if (flags & BTRFS_BLOCK_GROUP_DATA)
10364                 fs_info->avail_data_alloc_bits &= ~extra_flags;
10365         if (flags & BTRFS_BLOCK_GROUP_METADATA)
10366                 fs_info->avail_metadata_alloc_bits &= ~extra_flags;
10367         if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
10368                 fs_info->avail_system_alloc_bits &= ~extra_flags;
10369         write_sequnlock(&fs_info->profiles_lock);
10370 }
10371
10372 int btrfs_remove_block_group(struct btrfs_trans_handle *trans,
10373                              struct btrfs_fs_info *fs_info, u64 group_start,
10374                              struct extent_map *em)
10375 {
10376         struct btrfs_root *root = fs_info->extent_root;
10377         struct btrfs_path *path;
10378         struct btrfs_block_group_cache *block_group;
10379         struct btrfs_free_cluster *cluster;
10380         struct btrfs_root *tree_root = fs_info->tree_root;
10381         struct btrfs_key key;
10382         struct inode *inode;
10383         struct kobject *kobj = NULL;
10384         int ret;
10385         int index;
10386         int factor;
10387         struct btrfs_caching_control *caching_ctl = NULL;
10388         bool remove_em;
10389
10390         block_group = btrfs_lookup_block_group(fs_info, group_start);
10391         BUG_ON(!block_group);
10392         BUG_ON(!block_group->ro);
10393
10394         /*
10395          * Free the reserved super bytes from this block group before
10396          * remove it.
10397          */
10398         free_excluded_extents(fs_info, block_group);
10399         btrfs_free_ref_tree_range(fs_info, block_group->key.objectid,
10400                                   block_group->key.offset);
10401
10402         memcpy(&key, &block_group->key, sizeof(key));
10403         index = btrfs_bg_flags_to_raid_index(block_group->flags);
10404         if (block_group->flags & (BTRFS_BLOCK_GROUP_DUP |
10405                                   BTRFS_BLOCK_GROUP_RAID1 |
10406                                   BTRFS_BLOCK_GROUP_RAID10))
10407                 factor = 2;
10408         else
10409                 factor = 1;
10410
10411         /* make sure this block group isn't part of an allocation cluster */
10412         cluster = &fs_info->data_alloc_cluster;
10413         spin_lock(&cluster->refill_lock);
10414         btrfs_return_cluster_to_free_space(block_group, cluster);
10415         spin_unlock(&cluster->refill_lock);
10416
10417         /*
10418          * make sure this block group isn't part of a metadata
10419          * allocation cluster
10420          */
10421         cluster = &fs_info->meta_alloc_cluster;
10422         spin_lock(&cluster->refill_lock);
10423         btrfs_return_cluster_to_free_space(block_group, cluster);
10424         spin_unlock(&cluster->refill_lock);
10425
10426         path = btrfs_alloc_path();
10427         if (!path) {
10428                 ret = -ENOMEM;
10429                 goto out;
10430         }
10431
10432         /*
10433          * get the inode first so any iput calls done for the io_list
10434          * aren't the final iput (no unlinks allowed now)
10435          */
10436         inode = lookup_free_space_inode(fs_info, block_group, path);
10437
10438         mutex_lock(&trans->transaction->cache_write_mutex);
10439         /*
10440          * make sure our free spache cache IO is done before remove the
10441          * free space inode
10442          */
10443         spin_lock(&trans->transaction->dirty_bgs_lock);
10444         if (!list_empty(&block_group->io_list)) {
10445                 list_del_init(&block_group->io_list);
10446
10447                 WARN_ON(!IS_ERR(inode) && inode != block_group->io_ctl.inode);
10448
10449                 spin_unlock(&trans->transaction->dirty_bgs_lock);
10450                 btrfs_wait_cache_io(trans, block_group, path);
10451                 btrfs_put_block_group(block_group);
10452                 spin_lock(&trans->transaction->dirty_bgs_lock);
10453         }
10454
10455         if (!list_empty(&block_group->dirty_list)) {
10456                 list_del_init(&block_group->dirty_list);
10457                 btrfs_put_block_group(block_group);
10458         }
10459         spin_unlock(&trans->transaction->dirty_bgs_lock);
10460         mutex_unlock(&trans->transaction->cache_write_mutex);
10461
10462         if (!IS_ERR(inode)) {
10463                 ret = btrfs_orphan_add(trans, BTRFS_I(inode));
10464                 if (ret) {
10465                         btrfs_add_delayed_iput(inode);
10466                         goto out;
10467                 }
10468                 clear_nlink(inode);
10469                 /* One for the block groups ref */
10470                 spin_lock(&block_group->lock);
10471                 if (block_group->iref) {
10472                         block_group->iref = 0;
10473                         block_group->inode = NULL;
10474                         spin_unlock(&block_group->lock);
10475                         iput(inode);
10476                 } else {
10477                         spin_unlock(&block_group->lock);
10478                 }
10479                 /* One for our lookup ref */
10480                 btrfs_add_delayed_iput(inode);
10481         }
10482
10483         key.objectid = BTRFS_FREE_SPACE_OBJECTID;
10484         key.offset = block_group->key.objectid;
10485         key.type = 0;
10486
10487         ret = btrfs_search_slot(trans, tree_root, &key, path, -1, 1);
10488         if (ret < 0)
10489                 goto out;
10490         if (ret > 0)
10491                 btrfs_release_path(path);
10492         if (ret == 0) {
10493                 ret = btrfs_del_item(trans, tree_root, path);
10494                 if (ret)
10495                         goto out;
10496                 btrfs_release_path(path);
10497         }
10498
10499         spin_lock(&fs_info->block_group_cache_lock);
10500         rb_erase(&block_group->cache_node,
10501                  &fs_info->block_group_cache_tree);
10502         RB_CLEAR_NODE(&block_group->cache_node);
10503
10504         if (fs_info->first_logical_byte == block_group->key.objectid)
10505                 fs_info->first_logical_byte = (u64)-1;
10506         spin_unlock(&fs_info->block_group_cache_lock);
10507
10508         down_write(&block_group->space_info->groups_sem);
10509         /*
10510          * we must use list_del_init so people can check to see if they
10511          * are still on the list after taking the semaphore
10512          */
10513         list_del_init(&block_group->list);
10514         if (list_empty(&block_group->space_info->block_groups[index])) {
10515                 kobj = block_group->space_info->block_group_kobjs[index];
10516                 block_group->space_info->block_group_kobjs[index] = NULL;
10517                 clear_avail_alloc_bits(fs_info, block_group->flags);
10518         }
10519         up_write(&block_group->space_info->groups_sem);
10520         if (kobj) {
10521                 kobject_del(kobj);
10522                 kobject_put(kobj);
10523         }
10524
10525         if (block_group->has_caching_ctl)
10526                 caching_ctl = get_caching_control(block_group);
10527         if (block_group->cached == BTRFS_CACHE_STARTED)
10528                 wait_block_group_cache_done(block_group);
10529         if (block_group->has_caching_ctl) {
10530                 down_write(&fs_info->commit_root_sem);
10531                 if (!caching_ctl) {
10532                         struct btrfs_caching_control *ctl;
10533
10534                         list_for_each_entry(ctl,
10535                                     &fs_info->caching_block_groups, list)
10536                                 if (ctl->block_group == block_group) {
10537                                         caching_ctl = ctl;
10538                                         refcount_inc(&caching_ctl->count);
10539                                         break;
10540                                 }
10541                 }
10542                 if (caching_ctl)
10543                         list_del_init(&caching_ctl->list);
10544                 up_write(&fs_info->commit_root_sem);
10545                 if (caching_ctl) {
10546                         /* Once for the caching bgs list and once for us. */
10547                         put_caching_control(caching_ctl);
10548                         put_caching_control(caching_ctl);
10549                 }
10550         }
10551
10552         spin_lock(&trans->transaction->dirty_bgs_lock);
10553         if (!list_empty(&block_group->dirty_list)) {
10554                 WARN_ON(1);
10555         }
10556         if (!list_empty(&block_group->io_list)) {
10557                 WARN_ON(1);
10558         }
10559         spin_unlock(&trans->transaction->dirty_bgs_lock);
10560         btrfs_remove_free_space_cache(block_group);
10561
10562         spin_lock(&block_group->space_info->lock);
10563         list_del_init(&block_group->ro_list);
10564
10565         if (btrfs_test_opt(fs_info, ENOSPC_DEBUG)) {
10566                 WARN_ON(block_group->space_info->total_bytes
10567                         < block_group->key.offset);
10568                 WARN_ON(block_group->space_info->bytes_readonly
10569                         < block_group->key.offset);
10570                 WARN_ON(block_group->space_info->disk_total
10571                         < block_group->key.offset * factor);
10572         }
10573         block_group->space_info->total_bytes -= block_group->key.offset;
10574         block_group->space_info->bytes_readonly -= block_group->key.offset;
10575         block_group->space_info->disk_total -= block_group->key.offset * factor;
10576
10577         spin_unlock(&block_group->space_info->lock);
10578
10579         memcpy(&key, &block_group->key, sizeof(key));
10580
10581         mutex_lock(&fs_info->chunk_mutex);
10582         if (!list_empty(&em->list)) {
10583                 /* We're in the transaction->pending_chunks list. */
10584                 free_extent_map(em);
10585         }
10586         spin_lock(&block_group->lock);
10587         block_group->removed = 1;
10588         /*
10589          * At this point trimming can't start on this block group, because we
10590          * removed the block group from the tree fs_info->block_group_cache_tree
10591          * so no one can't find it anymore and even if someone already got this
10592          * block group before we removed it from the rbtree, they have already
10593          * incremented block_group->trimming - if they didn't, they won't find
10594          * any free space entries because we already removed them all when we
10595          * called btrfs_remove_free_space_cache().
10596          *
10597          * And we must not remove the extent map from the fs_info->mapping_tree
10598          * to prevent the same logical address range and physical device space
10599          * ranges from being reused for a new block group. This is because our
10600          * fs trim operation (btrfs_trim_fs() / btrfs_ioctl_fitrim()) is
10601          * completely transactionless, so while it is trimming a range the
10602          * currently running transaction might finish and a new one start,
10603          * allowing for new block groups to be created that can reuse the same
10604          * physical device locations unless we take this special care.
10605          *
10606          * There may also be an implicit trim operation if the file system
10607          * is mounted with -odiscard. The same protections must remain
10608          * in place until the extents have been discarded completely when
10609          * the transaction commit has completed.
10610          */
10611         remove_em = (atomic_read(&block_group->trimming) == 0);
10612         /*
10613          * Make sure a trimmer task always sees the em in the pinned_chunks list
10614          * if it sees block_group->removed == 1 (needs to lock block_group->lock
10615          * before checking block_group->removed).
10616          */
10617         if (!remove_em) {
10618                 /*
10619                  * Our em might be in trans->transaction->pending_chunks which
10620                  * is protected by fs_info->chunk_mutex ([lock|unlock]_chunks),
10621                  * and so is the fs_info->pinned_chunks list.
10622                  *
10623                  * So at this point we must be holding the chunk_mutex to avoid
10624                  * any races with chunk allocation (more specifically at
10625                  * volumes.c:contains_pending_extent()), to ensure it always
10626                  * sees the em, either in the pending_chunks list or in the
10627                  * pinned_chunks list.
10628                  */
10629                 list_move_tail(&em->list, &fs_info->pinned_chunks);
10630         }
10631         spin_unlock(&block_group->lock);
10632
10633         if (remove_em) {
10634                 struct extent_map_tree *em_tree;
10635
10636                 em_tree = &fs_info->mapping_tree.map_tree;
10637                 write_lock(&em_tree->lock);
10638                 /*
10639                  * The em might be in the pending_chunks list, so make sure the
10640                  * chunk mutex is locked, since remove_extent_mapping() will
10641                  * delete us from that list.
10642                  */
10643                 remove_extent_mapping(em_tree, em);
10644                 write_unlock(&em_tree->lock);
10645                 /* once for the tree */
10646                 free_extent_map(em);
10647         }
10648
10649         mutex_unlock(&fs_info->chunk_mutex);
10650
10651         ret = remove_block_group_free_space(trans, fs_info, block_group);
10652         if (ret)
10653                 goto out;
10654
10655         btrfs_put_block_group(block_group);
10656         btrfs_put_block_group(block_group);
10657
10658         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
10659         if (ret > 0)
10660                 ret = -EIO;
10661         if (ret < 0)
10662                 goto out;
10663
10664         ret = btrfs_del_item(trans, root, path);
10665 out:
10666         btrfs_free_path(path);
10667         return ret;
10668 }
10669
10670 struct btrfs_trans_handle *
10671 btrfs_start_trans_remove_block_group(struct btrfs_fs_info *fs_info,
10672                                      const u64 chunk_offset)
10673 {
10674         struct extent_map_tree *em_tree = &fs_info->mapping_tree.map_tree;
10675         struct extent_map *em;
10676         struct map_lookup *map;
10677         unsigned int num_items;
10678
10679         read_lock(&em_tree->lock);
10680         em = lookup_extent_mapping(em_tree, chunk_offset, 1);
10681         read_unlock(&em_tree->lock);
10682         ASSERT(em && em->start == chunk_offset);
10683
10684         /*
10685          * We need to reserve 3 + N units from the metadata space info in order
10686          * to remove a block group (done at btrfs_remove_chunk() and at
10687          * btrfs_remove_block_group()), which are used for:
10688          *
10689          * 1 unit for adding the free space inode's orphan (located in the tree
10690          * of tree roots).
10691          * 1 unit for deleting the block group item (located in the extent
10692          * tree).
10693          * 1 unit for deleting the free space item (located in tree of tree
10694          * roots).
10695          * N units for deleting N device extent items corresponding to each
10696          * stripe (located in the device tree).
10697          *
10698          * In order to remove a block group we also need to reserve units in the
10699          * system space info in order to update the chunk tree (update one or
10700          * more device items and remove one chunk item), but this is done at
10701          * btrfs_remove_chunk() through a call to check_system_chunk().
10702          */
10703         map = em->map_lookup;
10704         num_items = 3 + map->num_stripes;
10705         free_extent_map(em);
10706
10707         return btrfs_start_transaction_fallback_global_rsv(fs_info->extent_root,
10708                                                            num_items, 1);
10709 }
10710
10711 /*
10712  * Process the unused_bgs list and remove any that don't have any allocated
10713  * space inside of them.
10714  */
10715 void btrfs_delete_unused_bgs(struct btrfs_fs_info *fs_info)
10716 {
10717         struct btrfs_block_group_cache *block_group;
10718         struct btrfs_space_info *space_info;
10719         struct btrfs_trans_handle *trans;
10720         int ret = 0;
10721
10722         if (!test_bit(BTRFS_FS_OPEN, &fs_info->flags))
10723                 return;
10724
10725         spin_lock(&fs_info->unused_bgs_lock);
10726         while (!list_empty(&fs_info->unused_bgs)) {
10727                 u64 start, end;
10728                 int trimming;
10729
10730                 block_group = list_first_entry(&fs_info->unused_bgs,
10731                                                struct btrfs_block_group_cache,
10732                                                bg_list);
10733                 list_del_init(&block_group->bg_list);
10734
10735                 space_info = block_group->space_info;
10736
10737                 if (ret || btrfs_mixed_space_info(space_info)) {
10738                         btrfs_put_block_group(block_group);
10739                         continue;
10740                 }
10741                 spin_unlock(&fs_info->unused_bgs_lock);
10742
10743                 mutex_lock(&fs_info->delete_unused_bgs_mutex);
10744
10745                 /* Don't want to race with allocators so take the groups_sem */
10746                 down_write(&space_info->groups_sem);
10747                 spin_lock(&block_group->lock);
10748                 if (block_group->reserved ||
10749                     btrfs_block_group_used(&block_group->item) ||
10750                     block_group->ro ||
10751                     list_is_singular(&block_group->list)) {
10752                         /*
10753                          * We want to bail if we made new allocations or have
10754                          * outstanding allocations in this block group.  We do
10755                          * the ro check in case balance is currently acting on
10756                          * this block group.
10757                          */
10758                         spin_unlock(&block_group->lock);
10759                         up_write(&space_info->groups_sem);
10760                         goto next;
10761                 }
10762                 spin_unlock(&block_group->lock);
10763
10764                 /* We don't want to force the issue, only flip if it's ok. */
10765                 ret = inc_block_group_ro(block_group, 0);
10766                 up_write(&space_info->groups_sem);
10767                 if (ret < 0) {
10768                         ret = 0;
10769                         goto next;
10770                 }
10771
10772                 /*
10773                  * Want to do this before we do anything else so we can recover
10774                  * properly if we fail to join the transaction.
10775                  */
10776                 trans = btrfs_start_trans_remove_block_group(fs_info,
10777                                                      block_group->key.objectid);
10778                 if (IS_ERR(trans)) {
10779                         btrfs_dec_block_group_ro(block_group);
10780                         ret = PTR_ERR(trans);
10781                         goto next;
10782                 }
10783
10784                 /*
10785                  * We could have pending pinned extents for this block group,
10786                  * just delete them, we don't care about them anymore.
10787                  */
10788                 start = block_group->key.objectid;
10789                 end = start + block_group->key.offset - 1;
10790                 /*
10791                  * Hold the unused_bg_unpin_mutex lock to avoid racing with
10792                  * btrfs_finish_extent_commit(). If we are at transaction N,
10793                  * another task might be running finish_extent_commit() for the
10794                  * previous transaction N - 1, and have seen a range belonging
10795                  * to the block group in freed_extents[] before we were able to
10796                  * clear the whole block group range from freed_extents[]. This
10797                  * means that task can lookup for the block group after we
10798                  * unpinned it from freed_extents[] and removed it, leading to
10799                  * a BUG_ON() at btrfs_unpin_extent_range().
10800                  */
10801                 mutex_lock(&fs_info->unused_bg_unpin_mutex);
10802                 ret = clear_extent_bits(&fs_info->freed_extents[0], start, end,
10803                                   EXTENT_DIRTY);
10804                 if (ret) {
10805                         mutex_unlock(&fs_info->unused_bg_unpin_mutex);
10806                         btrfs_dec_block_group_ro(block_group);
10807                         goto end_trans;
10808                 }
10809                 ret = clear_extent_bits(&fs_info->freed_extents[1], start, end,
10810                                   EXTENT_DIRTY);
10811                 if (ret) {
10812                         mutex_unlock(&fs_info->unused_bg_unpin_mutex);
10813                         btrfs_dec_block_group_ro(block_group);
10814                         goto end_trans;
10815                 }
10816                 mutex_unlock(&fs_info->unused_bg_unpin_mutex);
10817
10818                 /* Reset pinned so btrfs_put_block_group doesn't complain */
10819                 spin_lock(&space_info->lock);
10820                 spin_lock(&block_group->lock);
10821
10822                 space_info->bytes_pinned -= block_group->pinned;
10823                 space_info->bytes_readonly += block_group->pinned;
10824                 percpu_counter_add(&space_info->total_bytes_pinned,
10825                                    -block_group->pinned);
10826                 block_group->pinned = 0;
10827
10828                 spin_unlock(&block_group->lock);
10829                 spin_unlock(&space_info->lock);
10830
10831                 /* DISCARD can flip during remount */
10832                 trimming = btrfs_test_opt(fs_info, DISCARD);
10833
10834                 /* Implicit trim during transaction commit. */
10835                 if (trimming)
10836                         btrfs_get_block_group_trimming(block_group);
10837
10838                 /*
10839                  * Btrfs_remove_chunk will abort the transaction if things go
10840                  * horribly wrong.
10841                  */
10842                 ret = btrfs_remove_chunk(trans, fs_info,
10843                                          block_group->key.objectid);
10844
10845                 if (ret) {
10846                         if (trimming)
10847                                 btrfs_put_block_group_trimming(block_group);
10848                         goto end_trans;
10849                 }
10850
10851                 /*
10852                  * If we're not mounted with -odiscard, we can just forget
10853                  * about this block group. Otherwise we'll need to wait
10854                  * until transaction commit to do the actual discard.
10855                  */
10856                 if (trimming) {
10857                         spin_lock(&fs_info->unused_bgs_lock);
10858                         /*
10859                          * A concurrent scrub might have added us to the list
10860                          * fs_info->unused_bgs, so use a list_move operation
10861                          * to add the block group to the deleted_bgs list.
10862                          */
10863                         list_move(&block_group->bg_list,
10864                                   &trans->transaction->deleted_bgs);
10865                         spin_unlock(&fs_info->unused_bgs_lock);
10866                         btrfs_get_block_group(block_group);
10867                 }
10868 end_trans:
10869                 btrfs_end_transaction(trans);
10870 next:
10871                 mutex_unlock(&fs_info->delete_unused_bgs_mutex);
10872                 btrfs_put_block_group(block_group);
10873                 spin_lock(&fs_info->unused_bgs_lock);
10874         }
10875         spin_unlock(&fs_info->unused_bgs_lock);
10876 }
10877
10878 int btrfs_init_space_info(struct btrfs_fs_info *fs_info)
10879 {
10880         struct btrfs_space_info *space_info;
10881         struct btrfs_super_block *disk_super;
10882         u64 features;
10883         u64 flags;
10884         int mixed = 0;
10885         int ret;
10886
10887         disk_super = fs_info->super_copy;
10888         if (!btrfs_super_root(disk_super))
10889                 return -EINVAL;
10890
10891         features = btrfs_super_incompat_flags(disk_super);
10892         if (features & BTRFS_FEATURE_INCOMPAT_MIXED_GROUPS)
10893                 mixed = 1;
10894
10895         flags = BTRFS_BLOCK_GROUP_SYSTEM;
10896         ret = create_space_info(fs_info, flags, &space_info);
10897         if (ret)
10898                 goto out;
10899
10900         if (mixed) {
10901                 flags = BTRFS_BLOCK_GROUP_METADATA | BTRFS_BLOCK_GROUP_DATA;
10902                 ret = create_space_info(fs_info, flags, &space_info);
10903         } else {
10904                 flags = BTRFS_BLOCK_GROUP_METADATA;
10905                 ret = create_space_info(fs_info, flags, &space_info);
10906                 if (ret)
10907                         goto out;
10908
10909                 flags = BTRFS_BLOCK_GROUP_DATA;
10910                 ret = create_space_info(fs_info, flags, &space_info);
10911         }
10912 out:
10913         return ret;
10914 }
10915
10916 int btrfs_error_unpin_extent_range(struct btrfs_fs_info *fs_info,
10917                                    u64 start, u64 end)
10918 {
10919         return unpin_extent_range(fs_info, start, end, false);
10920 }
10921
10922 /*
10923  * It used to be that old block groups would be left around forever.
10924  * Iterating over them would be enough to trim unused space.  Since we
10925  * now automatically remove them, we also need to iterate over unallocated
10926  * space.
10927  *
10928  * We don't want a transaction for this since the discard may take a
10929  * substantial amount of time.  We don't require that a transaction be
10930  * running, but we do need to take a running transaction into account
10931  * to ensure that we're not discarding chunks that were released in
10932  * the current transaction.
10933  *
10934  * Holding the chunks lock will prevent other threads from allocating
10935  * or releasing chunks, but it won't prevent a running transaction
10936  * from committing and releasing the memory that the pending chunks
10937  * list head uses.  For that, we need to take a reference to the
10938  * transaction.
10939  */
10940 static int btrfs_trim_free_extents(struct btrfs_device *device,
10941                                    u64 minlen, u64 *trimmed)
10942 {
10943         u64 start = 0, len = 0;
10944         int ret;
10945
10946         *trimmed = 0;
10947
10948         /* Not writeable = nothing to do. */
10949         if (!test_bit(BTRFS_DEV_STATE_WRITEABLE, &device->dev_state))
10950                 return 0;
10951
10952         /* No free space = nothing to do. */
10953         if (device->total_bytes <= device->bytes_used)
10954                 return 0;
10955
10956         ret = 0;
10957
10958         while (1) {
10959                 struct btrfs_fs_info *fs_info = device->fs_info;
10960                 struct btrfs_transaction *trans;
10961                 u64 bytes;
10962
10963                 ret = mutex_lock_interruptible(&fs_info->chunk_mutex);
10964                 if (ret)
10965                         return ret;
10966
10967                 down_read(&fs_info->commit_root_sem);
10968
10969                 spin_lock(&fs_info->trans_lock);
10970                 trans = fs_info->running_transaction;
10971                 if (trans)
10972                         refcount_inc(&trans->use_count);
10973                 spin_unlock(&fs_info->trans_lock);
10974
10975                 ret = find_free_dev_extent_start(trans, device, minlen, start,
10976                                                  &start, &len);
10977                 if (trans)
10978                         btrfs_put_transaction(trans);
10979
10980                 if (ret) {
10981                         up_read(&fs_info->commit_root_sem);
10982                         mutex_unlock(&fs_info->chunk_mutex);
10983                         if (ret == -ENOSPC)
10984                                 ret = 0;
10985                         break;
10986                 }
10987
10988                 ret = btrfs_issue_discard(device->bdev, start, len, &bytes);
10989                 up_read(&fs_info->commit_root_sem);
10990                 mutex_unlock(&fs_info->chunk_mutex);
10991
10992                 if (ret)
10993                         break;
10994
10995                 start += len;
10996                 *trimmed += bytes;
10997
10998                 if (fatal_signal_pending(current)) {
10999                         ret = -ERESTARTSYS;
11000                         break;
11001                 }
11002
11003                 cond_resched();
11004         }
11005
11006         return ret;
11007 }
11008
11009 int btrfs_trim_fs(struct btrfs_fs_info *fs_info, struct fstrim_range *range)
11010 {
11011         struct btrfs_block_group_cache *cache = NULL;
11012         struct btrfs_device *device;
11013         struct list_head *devices;
11014         u64 group_trimmed;
11015         u64 start;
11016         u64 end;
11017         u64 trimmed = 0;
11018         u64 total_bytes = btrfs_super_total_bytes(fs_info->super_copy);
11019         int ret = 0;
11020
11021         /*
11022          * try to trim all FS space, our block group may start from non-zero.
11023          */
11024         if (range->len == total_bytes)
11025                 cache = btrfs_lookup_first_block_group(fs_info, range->start);
11026         else
11027                 cache = btrfs_lookup_block_group(fs_info, range->start);
11028
11029         while (cache) {
11030                 if (cache->key.objectid >= (range->start + range->len)) {
11031                         btrfs_put_block_group(cache);
11032                         break;
11033                 }
11034
11035                 start = max(range->start, cache->key.objectid);
11036                 end = min(range->start + range->len,
11037                                 cache->key.objectid + cache->key.offset);
11038
11039                 if (end - start >= range->minlen) {
11040                         if (!block_group_cache_done(cache)) {
11041                                 ret = cache_block_group(cache, 0);
11042                                 if (ret) {
11043                                         btrfs_put_block_group(cache);
11044                                         break;
11045                                 }
11046                                 ret = wait_block_group_cache_done(cache);
11047                                 if (ret) {
11048                                         btrfs_put_block_group(cache);
11049                                         break;
11050                                 }
11051                         }
11052                         ret = btrfs_trim_block_group(cache,
11053                                                      &group_trimmed,
11054                                                      start,
11055                                                      end,
11056                                                      range->minlen);
11057
11058                         trimmed += group_trimmed;
11059                         if (ret) {
11060                                 btrfs_put_block_group(cache);
11061                                 break;
11062                         }
11063                 }
11064
11065                 cache = next_block_group(fs_info, cache);
11066         }
11067
11068         mutex_lock(&fs_info->fs_devices->device_list_mutex);
11069         devices = &fs_info->fs_devices->alloc_list;
11070         list_for_each_entry(device, devices, dev_alloc_list) {
11071                 ret = btrfs_trim_free_extents(device, range->minlen,
11072                                               &group_trimmed);
11073                 if (ret)
11074                         break;
11075
11076                 trimmed += group_trimmed;
11077         }
11078         mutex_unlock(&fs_info->fs_devices->device_list_mutex);
11079
11080         range->len = trimmed;
11081         return ret;
11082 }
11083
11084 /*
11085  * btrfs_{start,end}_write_no_snapshotting() are similar to
11086  * mnt_{want,drop}_write(), they are used to prevent some tasks from writing
11087  * data into the page cache through nocow before the subvolume is snapshoted,
11088  * but flush the data into disk after the snapshot creation, or to prevent
11089  * operations while snapshotting is ongoing and that cause the snapshot to be
11090  * inconsistent (writes followed by expanding truncates for example).
11091  */
11092 void btrfs_end_write_no_snapshotting(struct btrfs_root *root)
11093 {
11094         percpu_counter_dec(&root->subv_writers->counter);
11095         /*
11096          * Make sure counter is updated before we wake up waiters.
11097          */
11098         smp_mb();
11099         if (waitqueue_active(&root->subv_writers->wait))
11100                 wake_up(&root->subv_writers->wait);
11101 }
11102
11103 int btrfs_start_write_no_snapshotting(struct btrfs_root *root)
11104 {
11105         if (atomic_read(&root->will_be_snapshotted))
11106                 return 0;
11107
11108         percpu_counter_inc(&root->subv_writers->counter);
11109         /*
11110          * Make sure counter is updated before we check for snapshot creation.
11111          */
11112         smp_mb();
11113         if (atomic_read(&root->will_be_snapshotted)) {
11114                 btrfs_end_write_no_snapshotting(root);
11115                 return 0;
11116         }
11117         return 1;
11118 }
11119
11120 void btrfs_wait_for_snapshot_creation(struct btrfs_root *root)
11121 {
11122         while (true) {
11123                 int ret;
11124
11125                 ret = btrfs_start_write_no_snapshotting(root);
11126                 if (ret)
11127                         break;
11128                 wait_var_event(&root->will_be_snapshotted,
11129                                !atomic_read(&root->will_be_snapshotted));
11130         }
11131 }