fs/btrfs/extent-tree.c

   1 /*
   2  * Copyright (C) 2007 Oracle.  All rights reserved.
   3  *
   4  * This program is free software; you can redistribute it and/or
   5  * modify it under the terms of the GNU General Public
   6  * License v2 as published by the Free Software Foundation.
   7  *
   8  * This program is distributed in the hope that it will be useful,
   9  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  10  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  11  * General Public License for more details.
  12  *
  13  * You should have received a copy of the GNU General Public
  14  * License along with this program; if not, write to the
  15  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
  16  * Boston, MA 021110-1307, USA.
  17  */
  18 #include <linux/sched.h>
  19 #include <linux/pagemap.h>
  20 #include <linux/writeback.h>
  21 #include <linux/blkdev.h>
  22 #include <linux/sort.h>
  23 #include <linux/rcupdate.h>
  24 #include <linux/kthread.h>
  25 #include <linux/slab.h>
  26 #include <linux/ratelimit.h>
  27 #include <linux/percpu_counter.h>
  28 #include "hash.h"
  29 #include "tree-log.h"
  30 #include "disk-io.h"
  31 #include "print-tree.h"
  32 #include "volumes.h"
  33 #include "raid56.h"
  34 #include "locking.h"
  35 #include "free-space-cache.h"
  36 #include "free-space-tree.h"
  37 #include "math.h"
  38 #include "sysfs.h"
  39 #include "qgroup.h"
  40
  41 #undef SCRAMBLE_DELAYED_REFS
  42
  43 /*
  44  * control flags for do_chunk_alloc's force field
  45  * CHUNK_ALLOC_NO_FORCE means to only allocate a chunk
  46  * if we really need one.
  47  *
  48  * CHUNK_ALLOC_LIMITED means to only try and allocate one
  49  * if we have very few chunks already allocated.  This is
  50  * used as part of the clustering code to help make sure
  51  * we have a good pool of storage to cluster in, without
  52  * filling the FS with empty chunks
  53  *
  54  * CHUNK_ALLOC_FORCE means it must try to allocate one
  55  *
  56  */
  57 enum {
  58         CHUNK_ALLOC_NO_FORCE = 0,
  59         CHUNK_ALLOC_LIMITED = 1,
  60         CHUNK_ALLOC_FORCE = 2,
  61 };
  62
  63 /*
  64  * Control how reservations are dealt with.
  65  *
  66  * RESERVE_FREE - freeing a reservation.
  67  * RESERVE_ALLOC - allocating space and we need to update bytes_may_use for
  68  *   ENOSPC accounting
  69  * RESERVE_ALLOC_NO_ACCOUNT - allocating space and we should not update
  70  *   bytes_may_use as the ENOSPC accounting is done elsewhere
  71  */
  72 enum {
  73         RESERVE_FREE = 0,
  74         RESERVE_ALLOC = 1,
  75         RESERVE_ALLOC_NO_ACCOUNT = 2,
  76 };
  77
  78 static int update_block_group(struct btrfs_trans_handle *trans,
  79                               struct btrfs_root *root, u64 bytenr,
  80                               u64 num_bytes, int alloc);
  81 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
  82                                 struct btrfs_root *root,
  83                                 struct btrfs_delayed_ref_node *node, u64 parent,
  84                                 u64 root_objectid, u64 owner_objectid,
  85                                 u64 owner_offset, int refs_to_drop,
  86                                 struct btrfs_delayed_extent_op *extra_op);
  87 static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op,
  88                                     struct extent_buffer *leaf,
  89                                     struct btrfs_extent_item *ei);
  90 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
  91                                       struct btrfs_root *root,
  92                                       u64 parent, u64 root_objectid,
  93                                       u64 flags, u64 owner, u64 offset,
  94                                       struct btrfs_key *ins, int ref_mod);
  95 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
  96                                      struct btrfs_root *root,
  97                                      u64 parent, u64 root_objectid,
  98                                      u64 flags, struct btrfs_disk_key *key,
  99                                      int level, struct btrfs_key *ins);
 100 static int do_chunk_alloc(struct btrfs_trans_handle *trans,
 101                           struct btrfs_root *extent_root, u64 flags,
 102                           int force);
 103 static int find_next_key(struct btrfs_path *path, int level,
 104                          struct btrfs_key *key);
 105 static void dump_space_info(struct btrfs_space_info *info, u64 bytes,
 106                             int dump_block_groups);
 107 static int btrfs_update_reserved_bytes(struct btrfs_block_group_cache *cache,
 108                                        u64 num_bytes, int reserve,
 109                                        int delalloc);
 110 static int block_rsv_use_bytes(struct btrfs_block_rsv *block_rsv,
 111                                u64 num_bytes);
 112 int btrfs_pin_extent(struct btrfs_root *root,
 113                      u64 bytenr, u64 num_bytes, int reserved);
 114
 115 static noinline int
 116 block_group_cache_done(struct btrfs_block_group_cache *cache)
 117 {
 118         smp_mb();
 119         return cache->cached == BTRFS_CACHE_FINISHED ||
 120                 cache->cached == BTRFS_CACHE_ERROR;
 121 }
 122
 123 static int block_group_bits(struct btrfs_block_group_cache *cache, u64 bits)
 124 {
 125         return (cache->flags & bits) == bits;
 126 }
 127
 128 void btrfs_get_block_group(struct btrfs_block_group_cache *cache)
 129 {
 130         atomic_inc(&cache->count);
 131 }
 132
 133 void btrfs_put_block_group(struct btrfs_block_group_cache *cache)
 134 {
 135         if (atomic_dec_and_test(&cache->count)) {
 136                 WARN_ON(cache->pinned > 0);
 137                 WARN_ON(cache->reserved > 0);
 138                 kfree(cache->free_space_ctl);
 139                 kfree(cache);
 140         }
 141 }
 142
 143 /*
 144  * this adds the block group to the fs_info rb tree for the block group
 145  * cache
 146  */
 147 static int btrfs_add_block_group_cache(struct btrfs_fs_info *info,
 148                                 struct btrfs_block_group_cache *block_group)
 149 {
 150         struct rb_node **p;
 151         struct rb_node *parent = NULL;
 152         struct btrfs_block_group_cache *cache;
 153
 154         spin_lock(&info->block_group_cache_lock);
 155         p = &info->block_group_cache_tree.rb_node;
 156
 157         while (*p) {
 158                 parent = *p;
 159                 cache = rb_entry(parent, struct btrfs_block_group_cache,
 160                                  cache_node);
 161                 if (block_group->key.objectid < cache->key.objectid) {
 162                         p = &(*p)->rb_left;
 163                 } else if (block_group->key.objectid > cache->key.objectid) {
 164                         p = &(*p)->rb_right;
 165                 } else {
 166                         spin_unlock(&info->block_group_cache_lock);
 167                         return -EEXIST;
 168                 }
 169         }
 170
 171         rb_link_node(&block_group->cache_node, parent, p);
 172         rb_insert_color(&block_group->cache_node,
 173                         &info->block_group_cache_tree);
 174
 175         if (info->first_logical_byte > block_group->key.objectid)
 176                 info->first_logical_byte = block_group->key.objectid;
 177
 178         spin_unlock(&info->block_group_cache_lock);
 179
 180         return 0;
 181 }
 182
 183 /*
 184  * This will return the block group at or after bytenr if contains is 0, else
 185  * it will return the block group that contains the bytenr
 186  */
 187 static struct btrfs_block_group_cache *
 188 block_group_cache_tree_search(struct btrfs_fs_info *info, u64 bytenr,
 189                               int contains)
 190 {
 191         struct btrfs_block_group_cache *cache, *ret = NULL;
 192         struct rb_node *n;
 193         u64 end, start;
 194
 195         spin_lock(&info->block_group_cache_lock);
 196         n = info->block_group_cache_tree.rb_node;
 197
 198         while (n) {
 199                 cache = rb_entry(n, struct btrfs_block_group_cache,
 200                                  cache_node);
 201                 end = cache->key.objectid + cache->key.offset - 1;
 202                 start = cache->key.objectid;
 203
 204                 if (bytenr < start) {
 205                         if (!contains && (!ret || start < ret->key.objectid))
 206                                 ret = cache;
 207                         n = n->rb_left;
 208                 } else if (bytenr > start) {
 209                         if (contains && bytenr <= end) {
 210                                 ret = cache;
 211                                 break;
 212                         }
 213                         n = n->rb_right;
 214                 } else {
 215                         ret = cache;
 216                         break;
 217                 }
 218         }
 219         if (ret) {
 220                 btrfs_get_block_group(ret);
 221                 if (bytenr == 0 && info->first_logical_byte > ret->key.objectid)
 222                         info->first_logical_byte = ret->key.objectid;
 223         }
 224         spin_unlock(&info->block_group_cache_lock);
 225
 226         return ret;
 227 }
 228
 229 static int add_excluded_extent(struct btrfs_root *root,
 230                                u64 start, u64 num_bytes)
 231 {
 232         u64 end = start + num_bytes - 1;
 233         set_extent_bits(&root->fs_info->freed_extents[0],
 234                         start, end, EXTENT_UPTODATE, GFP_NOFS);
 235         set_extent_bits(&root->fs_info->freed_extents[1],
 236                         start, end, EXTENT_UPTODATE, GFP_NOFS);
 237         return 0;
 238 }
 239
 240 static void free_excluded_extents(struct btrfs_root *root,
 241                                   struct btrfs_block_group_cache *cache)
 242 {
 243         u64 start, end;
 244
 245         start = cache->key.objectid;
 246         end = start + cache->key.offset - 1;
 247
 248         clear_extent_bits(&root->fs_info->freed_extents[0],
 249                           start, end, EXTENT_UPTODATE, GFP_NOFS);
 250         clear_extent_bits(&root->fs_info->freed_extents[1],
 251                           start, end, EXTENT_UPTODATE, GFP_NOFS);
 252 }
 253
 254 static int exclude_super_stripes(struct btrfs_root *root,
 255                                  struct btrfs_block_group_cache *cache)
 256 {
 257         u64 bytenr;
 258         u64 *logical;
 259         int stripe_len;
 260         int i, nr, ret;
 261
 262         if (cache->key.objectid < BTRFS_SUPER_INFO_OFFSET) {
 263                 stripe_len = BTRFS_SUPER_INFO_OFFSET - cache->key.objectid;
 264                 cache->bytes_super += stripe_len;
 265                 ret = add_excluded_extent(root, cache->key.objectid,
 266                                           stripe_len);
 267                 if (ret)
 268                         return ret;
 269         }
 270
 271         for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
 272                 bytenr = btrfs_sb_offset(i);
 273                 ret = btrfs_rmap_block(&root->fs_info->mapping_tree,
 274                                        cache->key.objectid, bytenr,
 275                                        0, &logical, &nr, &stripe_len);
 276                 if (ret)
 277                         return ret;
 278
 279                 while (nr--) {
 280                         u64 start, len;
 281
 282                         if (logical[nr] > cache->key.objectid +
 283                             cache->key.offset)
 284                                 continue;
 285
 286                         if (logical[nr] + stripe_len <= cache->key.objectid)
 287                                 continue;
 288
 289                         start = logical[nr];
 290                         if (start < cache->key.objectid) {
 291                                 start = cache->key.objectid;
 292                                 len = (logical[nr] + stripe_len) - start;
 293                         } else {
 294                                 len = min_t(u64, stripe_len,
 295                                             cache->key.objectid +
 296                                             cache->key.offset - start);
 297                         }
 298
 299                         cache->bytes_super += len;
 300                         ret = add_excluded_extent(root, start, len);
 301                         if (ret) {
 302                                 kfree(logical);
 303                                 return ret;
 304                         }
 305                 }
 306
 307                 kfree(logical);
 308         }
 309         return 0;
 310 }
 311
 312 static struct btrfs_caching_control *
 313 get_caching_control(struct btrfs_block_group_cache *cache)
 314 {
 315         struct btrfs_caching_control *ctl;
 316
 317         spin_lock(&cache->lock);
 318         if (!cache->caching_ctl) {
 319                 spin_unlock(&cache->lock);
 320                 return NULL;
 321         }
 322
 323         ctl = cache->caching_ctl;
 324         atomic_inc(&ctl->count);
 325         spin_unlock(&cache->lock);
 326         return ctl;
 327 }
 328
 329 static void put_caching_control(struct btrfs_caching_control *ctl)
 330 {
 331         if (atomic_dec_and_test(&ctl->count))
 332                 kfree(ctl);
 333 }
 334
 335 #ifdef CONFIG_BTRFS_DEBUG
 336 static void fragment_free_space(struct btrfs_root *root,
 337                                 struct btrfs_block_group_cache *block_group)
 338 {
 339         u64 start = block_group->key.objectid;
 340         u64 len = block_group->key.offset;
 341         u64 chunk = block_group->flags & BTRFS_BLOCK_GROUP_METADATA ?
 342                 root->nodesize : root->sectorsize;
 343         u64 step = chunk << 1;
 344
 345         while (len > chunk) {
 346                 btrfs_remove_free_space(block_group, start, chunk);
 347                 start += step;
 348                 if (len < step)
 349                         len = 0;
 350                 else
 351                         len -= step;
 352         }
 353 }
 354 #endif
 355
 356 /*
 357  * this is only called by cache_block_group, since we could have freed extents
 358  * we need to check the pinned_extents for any extents that can't be used yet
 359  * since their free space will be released as soon as the transaction commits.
 360  */
 361 u64 add_new_free_space(struct btrfs_block_group_cache *block_group,
 362                        struct btrfs_fs_info *info, u64 start, u64 end)
 363 {
 364         u64 extent_start, extent_end, size, total_added = 0;
 365         int ret;
 366
 367         while (start < end) {
 368                 ret = find_first_extent_bit(info->pinned_extents, start,
 369                                             &extent_start, &extent_end,
 370                                             EXTENT_DIRTY | EXTENT_UPTODATE,
 371                                             NULL);
 372                 if (ret)
 373                         break;
 374
 375                 if (extent_start <= start) {
 376                         start = extent_end + 1;
 377                 } else if (extent_start > start && extent_start < end) {
 378                         size = extent_start - start;
 379                         total_added += size;
 380                         ret = btrfs_add_free_space(block_group, start,
 381                                                    size);
 382                         BUG_ON(ret); /* -ENOMEM or logic error */
 383                         start = extent_end + 1;
 384                 } else {
 385                         break;
 386                 }
 387         }
 388
 389         if (start < end) {
 390                 size = end - start;
 391                 total_added += size;
 392                 ret = btrfs_add_free_space(block_group, start, size);
 393                 BUG_ON(ret); /* -ENOMEM or logic error */
 394         }
 395
 396         return total_added;
 397 }
 398
 399 static int load_extent_tree_free(struct btrfs_caching_control *caching_ctl)
 400 {
 401         struct btrfs_block_group_cache *block_group;
 402         struct btrfs_fs_info *fs_info;
 403         struct btrfs_root *extent_root;
 404         struct btrfs_path *path;
 405         struct extent_buffer *leaf;
 406         struct btrfs_key key;
 407         u64 total_found = 0;
 408         u64 last = 0;
 409         u32 nritems;
 410         int ret;
 411         bool wakeup = true;
 412
 413         block_group = caching_ctl->block_group;
 414         fs_info = block_group->fs_info;
 415         extent_root = fs_info->extent_root;
 416
 417         path = btrfs_alloc_path();
 418         if (!path)
 419                 return -ENOMEM;
 420
 421         last = max_t(u64, block_group->key.objectid, BTRFS_SUPER_INFO_OFFSET);
 422
 423 #ifdef CONFIG_BTRFS_DEBUG
 424         /*
 425          * If we're fragmenting we don't want to make anybody think we can
 426          * allocate from this block group until we've had a chance to fragment
 427          * the free space.
 428          */
 429         if (btrfs_should_fragment_free_space(extent_root, block_group))
 430                 wakeup = false;
 431 #endif
 432         /*
 433          * We don't want to deadlock with somebody trying to allocate a new
 434          * extent for the extent root while also trying to search the extent
 435          * root to add free space.  So we skip locking and search the commit
 436          * root, since its read-only
 437          */
 438         path->skip_locking = 1;
 439         path->search_commit_root = 1;
 440         path->reada = 1;
 441
 442         key.objectid = last;
 443         key.offset = 0;
 444         key.type = BTRFS_EXTENT_ITEM_KEY;
 445
 446 next:
 447         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
 448         if (ret < 0)
 449                 goto out;
 450
 451         leaf = path->nodes[0];
 452         nritems = btrfs_header_nritems(leaf);
 453
 454         while (1) {
 455                 if (btrfs_fs_closing(fs_info) > 1) {
 456                         last = (u64)-1;
 457                         break;
 458                 }
 459
 460                 if (path->slots[0] < nritems) {
 461                         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
 462                 } else {
 463                         ret = find_next_key(path, 0, &key);
 464                         if (ret)
 465                                 break;
 466
 467                         if (need_resched() ||
 468                             rwsem_is_contended(&fs_info->commit_root_sem)) {
 469                                 if (wakeup)
 470                                         caching_ctl->progress = last;
 471                                 btrfs_release_path(path);
 472                                 up_read(&fs_info->commit_root_sem);
 473                                 mutex_unlock(&caching_ctl->mutex);
 474                                 cond_resched();
 475                                 mutex_lock(&caching_ctl->mutex);
 476                                 down_read(&fs_info->commit_root_sem);
 477                                 goto next;
 478                         }
 479
 480                         ret = btrfs_next_leaf(extent_root, path);
 481                         if (ret < 0)
 482                                 goto out;
 483                         if (ret)
 484                                 break;
 485                         leaf = path->nodes[0];
 486                         nritems = btrfs_header_nritems(leaf);
 487                         continue;
 488                 }
 489
 490                 if (key.objectid < last) {
 491                         key.objectid = last;
 492                         key.offset = 0;
 493                         key.type = BTRFS_EXTENT_ITEM_KEY;
 494
 495                         if (wakeup)
 496                                 caching_ctl->progress = last;
 497                         btrfs_release_path(path);
 498                         goto next;
 499                 }
 500
 501                 if (key.objectid < block_group->key.objectid) {
 502                         path->slots[0]++;
 503                         continue;
 504                 }
 505
 506                 if (key.objectid >= block_group->key.objectid +
 507                     block_group->key.offset)
 508                         break;
 509
 510                 if (key.type == BTRFS_EXTENT_ITEM_KEY ||
 511                     key.type == BTRFS_METADATA_ITEM_KEY) {
 512                         total_found += add_new_free_space(block_group,
 513                                                           fs_info, last,
 514                                                           key.objectid);
 515                         if (key.type == BTRFS_METADATA_ITEM_KEY)
 516                                 last = key.objectid +
 517                                         fs_info->tree_root->nodesize;
 518                         else
 519                                 last = key.objectid + key.offset;
 520
 521                         if (total_found > CACHING_CTL_WAKE_UP) {
 522                                 total_found = 0;
 523                                 if (wakeup)
 524                                         wake_up(&caching_ctl->wait);
 525                         }
 526                 }
 527                 path->slots[0]++;
 528         }
 529         ret = 0;
 530
 531         total_found += add_new_free_space(block_group, fs_info, last,
 532                                           block_group->key.objectid +
 533                                           block_group->key.offset);
 534         caching_ctl->progress = (u64)-1;
 535
 536 out:
 537         btrfs_free_path(path);
 538         return ret;
 539 }
 540
 541 static noinline void caching_thread(struct btrfs_work *work)
 542 {
 543         struct btrfs_block_group_cache *block_group;
 544         struct btrfs_fs_info *fs_info;
 545         struct btrfs_caching_control *caching_ctl;
 546         struct btrfs_root *extent_root;
 547         int ret;
 548
 549         caching_ctl = container_of(work, struct btrfs_caching_control, work);
 550         block_group = caching_ctl->block_group;
 551         fs_info = block_group->fs_info;
 552         extent_root = fs_info->extent_root;
 553
 554         mutex_lock(&caching_ctl->mutex);
 555         down_read(&fs_info->commit_root_sem);
 556
 557         if (btrfs_fs_compat_ro(fs_info, FREE_SPACE_TREE))
 558                 ret = load_free_space_tree(caching_ctl);
 559         else
 560                 ret = load_extent_tree_free(caching_ctl);
 561
 562         spin_lock(&block_group->lock);
 563         block_group->caching_ctl = NULL;
 564         block_group->cached = ret ? BTRFS_CACHE_ERROR : BTRFS_CACHE_FINISHED;
 565         spin_unlock(&block_group->lock);
 566
 567 #ifdef CONFIG_BTRFS_DEBUG
 568         if (btrfs_should_fragment_free_space(extent_root, block_group)) {
 569                 u64 bytes_used;
 570
 571                 spin_lock(&block_group->space_info->lock);
 572                 spin_lock(&block_group->lock);
 573                 bytes_used = block_group->key.offset -
 574                         btrfs_block_group_used(&block_group->item);
 575                 block_group->space_info->bytes_used += bytes_used >> 1;
 576                 spin_unlock(&block_group->lock);
 577                 spin_unlock(&block_group->space_info->lock);
 578                 fragment_free_space(extent_root, block_group);
 579         }
 580 #endif
 581
 582         caching_ctl->progress = (u64)-1;
 583
 584         up_read(&fs_info->commit_root_sem);
 585         free_excluded_extents(fs_info->extent_root, block_group);
 586         mutex_unlock(&caching_ctl->mutex);
 587
 588         wake_up(&caching_ctl->wait);
 589
 590         put_caching_control(caching_ctl);
 591         btrfs_put_block_group(block_group);
 592 }
 593
 594 static int cache_block_group(struct btrfs_block_group_cache *cache,
 595                              int load_cache_only)
 596 {
 597         DEFINE_WAIT(wait);
 598         struct btrfs_fs_info *fs_info = cache->fs_info;
 599         struct btrfs_caching_control *caching_ctl;
 600         int ret = 0;
 601
 602         caching_ctl = kzalloc(sizeof(*caching_ctl), GFP_NOFS);
 603         if (!caching_ctl)
 604                 return -ENOMEM;
 605
 606         INIT_LIST_HEAD(&caching_ctl->list);
 607         mutex_init(&caching_ctl->mutex);
 608         init_waitqueue_head(&caching_ctl->wait);
 609         caching_ctl->block_group = cache;
 610         caching_ctl->progress = cache->key.objectid;
 611         atomic_set(&caching_ctl->count, 1);
 612         btrfs_init_work(&caching_ctl->work, btrfs_cache_helper,
 613                         caching_thread, NULL, NULL);
 614
 615         spin_lock(&cache->lock);
 616         /*
 617          * This should be a rare occasion, but this could happen I think in the
 618          * case where one thread starts to load the space cache info, and then
 619          * some other thread starts a transaction commit which tries to do an
 620          * allocation while the other thread is still loading the space cache
 621          * info.  The previous loop should have kept us from choosing this block
 622          * group, but if we've moved to the state where we will wait on caching
 623          * block groups we need to first check if we're doing a fast load here,
 624          * so we can wait for it to finish, otherwise we could end up allocating
 625          * from a block group who's cache gets evicted for one reason or
 626          * another.
 627          */
 628         while (cache->cached == BTRFS_CACHE_FAST) {
 629                 struct btrfs_caching_control *ctl;
 630
 631                 ctl = cache->caching_ctl;
 632                 atomic_inc(&ctl->count);
 633                 prepare_to_wait(&ctl->wait, &wait, TASK_UNINTERRUPTIBLE);
 634                 spin_unlock(&cache->lock);
 635
 636                 schedule();
 637
 638                 finish_wait(&ctl->wait, &wait);
 639                 put_caching_control(ctl);
 640                 spin_lock(&cache->lock);
 641         }
 642
 643         if (cache->cached != BTRFS_CACHE_NO) {
 644                 spin_unlock(&cache->lock);
 645                 kfree(caching_ctl);
 646                 return 0;
 647         }
 648         WARN_ON(cache->caching_ctl);
 649         cache->caching_ctl = caching_ctl;
 650         cache->cached = BTRFS_CACHE_FAST;
 651         spin_unlock(&cache->lock);
 652
 653         if (fs_info->mount_opt & BTRFS_MOUNT_SPACE_CACHE) {
 654                 mutex_lock(&caching_ctl->mutex);
 655                 ret = load_free_space_cache(fs_info, cache);
 656
 657                 spin_lock(&cache->lock);
 658                 if (ret == 1) {
 659                         cache->caching_ctl = NULL;
 660                         cache->cached = BTRFS_CACHE_FINISHED;
 661                         cache->last_byte_to_unpin = (u64)-1;
 662                         caching_ctl->progress = (u64)-1;
 663                 } else {
 664                         if (load_cache_only) {
 665                                 cache->caching_ctl = NULL;
 666                                 cache->cached = BTRFS_CACHE_NO;
 667                         } else {
 668                                 cache->cached = BTRFS_CACHE_STARTED;
 669                                 cache->has_caching_ctl = 1;
 670                         }
 671                 }
 672                 spin_unlock(&cache->lock);
 673 #ifdef CONFIG_BTRFS_DEBUG
 674                 if (ret == 1 &&
 675                     btrfs_should_fragment_free_space(fs_info->extent_root,
 676                                                      cache)) {
 677                         u64 bytes_used;
 678
 679                         spin_lock(&cache->space_info->lock);
 680                         spin_lock(&cache->lock);
 681                         bytes_used = cache->key.offset -
 682                                 btrfs_block_group_used(&cache->item);
 683                         cache->space_info->bytes_used += bytes_used >> 1;
 684                         spin_unlock(&cache->lock);
 685                         spin_unlock(&cache->space_info->lock);
 686                         fragment_free_space(fs_info->extent_root, cache);
 687                 }
 688 #endif
 689                 mutex_unlock(&caching_ctl->mutex);
 690
 691                 wake_up(&caching_ctl->wait);
 692                 if (ret == 1) {
 693                         put_caching_control(caching_ctl);
 694                         free_excluded_extents(fs_info->extent_root, cache);
 695                         return 0;
 696                 }
 697         } else {
 698                 /*
 699                  * We're either using the free space tree or no caching at all.
 700                  * Set cached to the appropriate value and wakeup any waiters.
 701                  */
 702                 spin_lock(&cache->lock);
 703                 if (load_cache_only) {
 704                         cache->caching_ctl = NULL;
 705                         cache->cached = BTRFS_CACHE_NO;
 706                 } else {
 707                         cache->cached = BTRFS_CACHE_STARTED;
 708                         cache->has_caching_ctl = 1;
 709                 }
 710                 spin_unlock(&cache->lock);
 711                 wake_up(&caching_ctl->wait);
 712         }
 713
 714         if (load_cache_only) {
 715                 put_caching_control(caching_ctl);
 716                 return 0;
 717         }
 718
 719         down_write(&fs_info->commit_root_sem);
 720         atomic_inc(&caching_ctl->count);
 721         list_add_tail(&caching_ctl->list, &fs_info->caching_block_groups);
 722         up_write(&fs_info->commit_root_sem);
 723
 724         btrfs_get_block_group(cache);
 725
 726         btrfs_queue_work(fs_info->caching_workers, &caching_ctl->work);
 727
 728         return ret;
 729 }
 730
 731 /*
 732  * return the block group that starts at or after bytenr
 733  */
 734 static struct btrfs_block_group_cache *
 735 btrfs_lookup_first_block_group(struct btrfs_fs_info *info, u64 bytenr)
 736 {
 737         struct btrfs_block_group_cache *cache;
 738
 739         cache = block_group_cache_tree_search(info, bytenr, 0);
 740
 741         return cache;
 742 }
 743
 744 /*
 745  * return the block group that contains the given bytenr
 746  */
 747 struct btrfs_block_group_cache *btrfs_lookup_block_group(
 748                                                  struct btrfs_fs_info *info,
 749                                                  u64 bytenr)
 750 {
 751         struct btrfs_block_group_cache *cache;
 752
 753         cache = block_group_cache_tree_search(info, bytenr, 1);
 754
 755         return cache;
 756 }
 757
 758 static struct btrfs_space_info *__find_space_info(struct btrfs_fs_info *info,
 759                                                   u64 flags)
 760 {
 761         struct list_head *head = &info->space_info;
 762         struct btrfs_space_info *found;
 763
 764         flags &= BTRFS_BLOCK_GROUP_TYPE_MASK;
 765
 766         rcu_read_lock();
 767         list_for_each_entry_rcu(found, head, list) {
 768                 if (found->flags & flags) {
 769                         rcu_read_unlock();
 770                         return found;
 771                 }
 772         }
 773         rcu_read_unlock();
 774         return NULL;
 775 }
 776
 777 /*
 778  * after adding space to the filesystem, we need to clear the full flags
 779  * on all the space infos.
 780  */
 781 void btrfs_clear_space_info_full(struct btrfs_fs_info *info)
 782 {
 783         struct list_head *head = &info->space_info;
 784         struct btrfs_space_info *found;
 785
 786         rcu_read_lock();
 787         list_for_each_entry_rcu(found, head, list)
 788                 found->full = 0;
 789         rcu_read_unlock();
 790 }
 791
 792 /* simple helper to search for an existing data extent at a given offset */
 793 int btrfs_lookup_data_extent(struct btrfs_root *root, u64 start, u64 len)
 794 {
 795         int ret;
 796         struct btrfs_key key;
 797         struct btrfs_path *path;
 798
 799         path = btrfs_alloc_path();
 800         if (!path)
 801                 return -ENOMEM;
 802
 803         key.objectid = start;
 804         key.offset = len;
 805         key.type = BTRFS_EXTENT_ITEM_KEY;
 806         ret = btrfs_search_slot(NULL, root->fs_info->extent_root, &key, path,
 807                                 0, 0);
 808         btrfs_free_path(path);
 809         return ret;
 810 }
 811
 812 /*
 813  * helper function to lookup reference count and flags of a tree block.
 814  *
 815  * the head node for delayed ref is used to store the sum of all the
 816  * reference count modifications queued up in the rbtree. the head
 817  * node may also store the extent flags to set. This way you can check
 818  * to see what the reference count and extent flags would be if all of
 819  * the delayed refs are not processed.
 820  */
 821 int btrfs_lookup_extent_info(struct btrfs_trans_handle *trans,
 822                              struct btrfs_root *root, u64 bytenr,
 823                              u64 offset, int metadata, u64 *refs, u64 *flags)
 824 {
 825         struct btrfs_delayed_ref_head *head;
 826         struct btrfs_delayed_ref_root *delayed_refs;
 827         struct btrfs_path *path;
 828         struct btrfs_extent_item *ei;
 829         struct extent_buffer *leaf;
 830         struct btrfs_key key;
 831         u32 item_size;
 832         u64 num_refs;
 833         u64 extent_flags;
 834         int ret;
 835
 836         /*
 837          * If we don't have skinny metadata, don't bother doing anything
 838          * different
 839          */
 840         if (metadata && !btrfs_fs_incompat(root->fs_info, SKINNY_METADATA)) {
 841                 offset = root->nodesize;
 842                 metadata = 0;
 843         }
 844
 845         path = btrfs_alloc_path();
 846         if (!path)
 847                 return -ENOMEM;
 848
 849         if (!trans) {
 850                 path->skip_locking = 1;
 851                 path->search_commit_root = 1;
 852         }
 853
 854 search_again:
 855         key.objectid = bytenr;
 856         key.offset = offset;
 857         if (metadata)
 858                 key.type = BTRFS_METADATA_ITEM_KEY;
 859         else
 860                 key.type = BTRFS_EXTENT_ITEM_KEY;
 861
 862         ret = btrfs_search_slot(trans, root->fs_info->extent_root,
 863                                 &key, path, 0, 0);
 864         if (ret < 0)
 865                 goto out_free;
 866
 867         if (ret > 0 && metadata && key.type == BTRFS_METADATA_ITEM_KEY) {
 868                 if (path->slots[0]) {
 869                         path->slots[0]--;
 870                         btrfs_item_key_to_cpu(path->nodes[0], &key,
 871                                               path->slots[0]);
 872                         if (key.objectid == bytenr &&
 873                             key.type == BTRFS_EXTENT_ITEM_KEY &&
 874                             key.offset == root->nodesize)
 875                                 ret = 0;
 876                 }
 877         }
 878
 879         if (ret == 0) {
 880                 leaf = path->nodes[0];
 881                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
 882                 if (item_size >= sizeof(*ei)) {
 883                         ei = btrfs_item_ptr(leaf, path->slots[0],
 884                                             struct btrfs_extent_item);
 885                         num_refs = btrfs_extent_refs(leaf, ei);
 886                         extent_flags = btrfs_extent_flags(leaf, ei);
 887                 } else {
 888 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
 889                         struct btrfs_extent_item_v0 *ei0;
 890                         BUG_ON(item_size != sizeof(*ei0));
 891                         ei0 = btrfs_item_ptr(leaf, path->slots[0],
 892                                              struct btrfs_extent_item_v0);
 893                         num_refs = btrfs_extent_refs_v0(leaf, ei0);
 894                         /* FIXME: this isn't correct for data */
 895                         extent_flags = BTRFS_BLOCK_FLAG_FULL_BACKREF;
 896 #else
 897                         BUG();
 898 #endif
 899                 }
 900                 BUG_ON(num_refs == 0);
 901         } else {
 902                 num_refs = 0;
 903                 extent_flags = 0;
 904                 ret = 0;
 905         }
 906
 907         if (!trans)
 908                 goto out;
 909
 910         delayed_refs = &trans->transaction->delayed_refs;
 911         spin_lock(&delayed_refs->lock);
 912         head = btrfs_find_delayed_ref_head(trans, bytenr);
 913         if (head) {
 914                 if (!mutex_trylock(&head->mutex)) {
 915                         atomic_inc(&head->node.refs);
 916                         spin_unlock(&delayed_refs->lock);
 917
 918                         btrfs_release_path(path);
 919
 920                         /*
 921                          * Mutex was contended, block until it's released and try
 922                          * again
 923                          */
 924                         mutex_lock(&head->mutex);
 925                         mutex_unlock(&head->mutex);
 926                         btrfs_put_delayed_ref(&head->node);
 927                         goto search_again;
 928                 }
 929                 spin_lock(&head->lock);
 930                 if (head->extent_op && head->extent_op->update_flags)
 931                         extent_flags |= head->extent_op->flags_to_set;
 932                 else
 933                         BUG_ON(num_refs == 0);
 934
 935                 num_refs += head->node.ref_mod;
 936                 spin_unlock(&head->lock);
 937                 mutex_unlock(&head->mutex);
 938         }
 939         spin_unlock(&delayed_refs->lock);
 940 out:
 941         WARN_ON(num_refs == 0);
 942         if (refs)
 943                 *refs = num_refs;
 944         if (flags)
 945                 *flags = extent_flags;
 946 out_free:
 947         btrfs_free_path(path);
 948         return ret;
 949 }
 950
 951 /*
 952  * Back reference rules.  Back refs have three main goals:
 953  *
 954  * 1) differentiate between all holders of references to an extent so that
 955  *    when a reference is dropped we can make sure it was a valid reference
 956  *    before freeing the extent.
 957  *
 958  * 2) Provide enough information to quickly find the holders of an extent
 959  *    if we notice a given block is corrupted or bad.
 960  *
 961  * 3) Make it easy to migrate blocks for FS shrinking or storage pool
 962  *    maintenance.  This is actually the same as #2, but with a slightly
 963  *    different use case.
 964  *
 965  * There are two kinds of back refs. The implicit back refs is optimized
 966  * for pointers in non-shared tree blocks. For a given pointer in a block,
 967  * back refs of this kind provide information about the block's owner tree
 968  * and the pointer's key. These information allow us to find the block by
 969  * b-tree searching. The full back refs is for pointers in tree blocks not
 970  * referenced by their owner trees. The location of tree block is recorded
 971  * in the back refs. Actually the full back refs is generic, and can be
 972  * used in all cases the implicit back refs is used. The major shortcoming
 973  * of the full back refs is its overhead. Every time a tree block gets
 974  * COWed, we have to update back refs entry for all pointers in it.
 975  *
 976  * For a newly allocated tree block, we use implicit back refs for
 977  * pointers in it. This means most tree related operations only involve
 978  * implicit back refs. For a tree block created in old transaction, the
 979  * only way to drop a reference to it is COW it. So we can detect the
 980  * event that tree block loses its owner tree's reference and do the
 981  * back refs conversion.
 982  *
 983  * When a tree block is COW'd through a tree, there are four cases:
 984  *
 985  * The reference count of the block is one and the tree is the block's
 986  * owner tree. Nothing to do in this case.
 987  *
 988  * The reference count of the block is one and the tree is not the
 989  * block's owner tree. In this case, full back refs is used for pointers
 990  * in the block. Remove these full back refs, add implicit back refs for
 991  * every pointers in the new block.
 992  *
 993  * The reference count of the block is greater than one and the tree is
 994  * the block's owner tree. In this case, implicit back refs is used for
 995  * pointers in the block. Add full back refs for every pointers in the
 996  * block, increase lower level extents' reference counts. The original
 997  * implicit back refs are entailed to the new block.
 998  *
 999  * The reference count of the block is greater than one and the tree is
1000  * not the block's owner tree. Add implicit back refs for every pointer in
1001  * the new block, increase lower level extents' reference count.
1002  *
1003  * Back Reference Key composing:
1004  *
1005  * The key objectid corresponds to the first byte in the extent,
1006  * The key type is used to differentiate between types of back refs.
1007  * There are different meanings of the key offset for different types
1008  * of back refs.
1009  *
1010  * File extents can be referenced by:
1011  *
1012  * - multiple snapshots, subvolumes, or different generations in one subvol
1013  * - different files inside a single subvolume
1014  * - different offsets inside a file (bookend extents in file.c)
1015  *
1016  * The extent ref structure for the implicit back refs has fields for:
1017  *
1018  * - Objectid of the subvolume root
1019  * - objectid of the file holding the reference
1020  * - original offset in the file
1021  * - how many bookend extents
1022  *
1023  * The key offset for the implicit back refs is hash of the first
1024  * three fields.
1025  *
1026  * The extent ref structure for the full back refs has field for:
1027  *
1028  * - number of pointers in the tree leaf
1029  *
1030  * The key offset for the implicit back refs is the first byte of
1031  * the tree leaf
1032  *
1033  * When a file extent is allocated, The implicit back refs is used.
1034  * the fields are filled in:
1035  *
1036  *     (root_key.objectid, inode objectid, offset in file, 1)
1037  *
1038  * When a file extent is removed file truncation, we find the
1039  * corresponding implicit back refs and check the following fields:
1040  *
1041  *     (btrfs_header_owner(leaf), inode objectid, offset in file)
1042  *
1043  * Btree extents can be referenced by:
1044  *
1045  * - Different subvolumes
1046  *
1047  * Both the implicit back refs and the full back refs for tree blocks
1048  * only consist of key. The key offset for the implicit back refs is
1049  * objectid of block's owner tree. The key offset for the full back refs
1050  * is the first byte of parent block.
1051  *
1052  * When implicit back refs is used, information about the lowest key and
1053  * level of the tree block are required. These information are stored in
1054  * tree block info structure.
1055  */
1056
1057 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1058 static int convert_extent_item_v0(struct btrfs_trans_handle *trans,
1059                                   struct btrfs_root *root,
1060                                   struct btrfs_path *path,
1061                                   u64 owner, u32 extra_size)
1062 {
1063         struct btrfs_extent_item *item;
1064         struct btrfs_extent_item_v0 *ei0;
1065         struct btrfs_extent_ref_v0 *ref0;
1066         struct btrfs_tree_block_info *bi;
1067         struct extent_buffer *leaf;
1068         struct btrfs_key key;
1069         struct btrfs_key found_key;
1070         u32 new_size = sizeof(*item);
1071         u64 refs;
1072         int ret;
1073
1074         leaf = path->nodes[0];
1075         BUG_ON(btrfs_item_size_nr(leaf, path->slots[0]) != sizeof(*ei0));
1076
1077         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1078         ei0 = btrfs_item_ptr(leaf, path->slots[0],
1079                              struct btrfs_extent_item_v0);
1080         refs = btrfs_extent_refs_v0(leaf, ei0);
1081
1082         if (owner == (u64)-1) {
1083                 while (1) {
1084                         if (path->slots[0] >= btrfs_header_nritems(leaf)) {
1085                                 ret = btrfs_next_leaf(root, path);
1086                                 if (ret < 0)
1087                                         return ret;
1088                                 BUG_ON(ret > 0); /* Corruption */
1089                                 leaf = path->nodes[0];
1090                         }
1091                         btrfs_item_key_to_cpu(leaf, &found_key,
1092                                               path->slots[0]);
1093                         BUG_ON(key.objectid != found_key.objectid);
1094                         if (found_key.type != BTRFS_EXTENT_REF_V0_KEY) {
1095                                 path->slots[0]++;
1096                                 continue;
1097                         }
1098                         ref0 = btrfs_item_ptr(leaf, path->slots[0],
1099                                               struct btrfs_extent_ref_v0);
1100                         owner = btrfs_ref_objectid_v0(leaf, ref0);
1101                         break;
1102                 }
1103         }
1104         btrfs_release_path(path);
1105
1106         if (owner < BTRFS_FIRST_FREE_OBJECTID)
1107                 new_size += sizeof(*bi);
1108
1109         new_size -= sizeof(*ei0);
1110         ret = btrfs_search_slot(trans, root, &key, path,
1111                                 new_size + extra_size, 1);
1112         if (ret < 0)
1113                 return ret;
1114         BUG_ON(ret); /* Corruption */
1115
1116         btrfs_extend_item(root, path, new_size);
1117
1118         leaf = path->nodes[0];
1119         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1120         btrfs_set_extent_refs(leaf, item, refs);
1121         /* FIXME: get real generation */
1122         btrfs_set_extent_generation(leaf, item, 0);
1123         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1124                 btrfs_set_extent_flags(leaf, item,
1125                                        BTRFS_EXTENT_FLAG_TREE_BLOCK |
1126                                        BTRFS_BLOCK_FLAG_FULL_BACKREF);
1127                 bi = (struct btrfs_tree_block_info *)(item + 1);
1128                 /* FIXME: get first key of the block */
1129                 memset_extent_buffer(leaf, 0, (unsigned long)bi, sizeof(*bi));
1130                 btrfs_set_tree_block_level(leaf, bi, (int)owner);
1131         } else {
1132                 btrfs_set_extent_flags(leaf, item, BTRFS_EXTENT_FLAG_DATA);
1133         }
1134         btrfs_mark_buffer_dirty(leaf);
1135         return 0;
1136 }
1137 #endif
1138
1139 static u64 hash_extent_data_ref(u64 root_objectid, u64 owner, u64 offset)
1140 {
1141         u32 high_crc = ~(u32)0;
1142         u32 low_crc = ~(u32)0;
1143         __le64 lenum;
1144
1145         lenum = cpu_to_le64(root_objectid);
1146         high_crc = btrfs_crc32c(high_crc, &lenum, sizeof(lenum));
1147         lenum = cpu_to_le64(owner);
1148         low_crc = btrfs_crc32c(low_crc, &lenum, sizeof(lenum));
1149         lenum = cpu_to_le64(offset);
1150         low_crc = btrfs_crc32c(low_crc, &lenum, sizeof(lenum));
1151
1152         return ((u64)high_crc << 31) ^ (u64)low_crc;
1153 }
1154
1155 static u64 hash_extent_data_ref_item(struct extent_buffer *leaf,
1156                                      struct btrfs_extent_data_ref *ref)
1157 {
1158         return hash_extent_data_ref(btrfs_extent_data_ref_root(leaf, ref),
1159                                     btrfs_extent_data_ref_objectid(leaf, ref),
1160                                     btrfs_extent_data_ref_offset(leaf, ref));
1161 }
1162
1163 static int match_extent_data_ref(struct extent_buffer *leaf,
1164                                  struct btrfs_extent_data_ref *ref,
1165                                  u64 root_objectid, u64 owner, u64 offset)
1166 {
1167         if (btrfs_extent_data_ref_root(leaf, ref) != root_objectid ||
1168             btrfs_extent_data_ref_objectid(leaf, ref) != owner ||
1169             btrfs_extent_data_ref_offset(leaf, ref) != offset)
1170                 return 0;
1171         return 1;
1172 }
1173
1174 static noinline int lookup_extent_data_ref(struct btrfs_trans_handle *trans,
1175                                            struct btrfs_root *root,
1176                                            struct btrfs_path *path,
1177                                            u64 bytenr, u64 parent,
1178                                            u64 root_objectid,
1179                                            u64 owner, u64 offset)
1180 {
1181         struct btrfs_key key;
1182         struct btrfs_extent_data_ref *ref;
1183         struct extent_buffer *leaf;
1184         u32 nritems;
1185         int ret;
1186         int recow;
1187         int err = -ENOENT;
1188
1189         key.objectid = bytenr;
1190         if (parent) {
1191                 key.type = BTRFS_SHARED_DATA_REF_KEY;
1192                 key.offset = parent;
1193         } else {
1194                 key.type = BTRFS_EXTENT_DATA_REF_KEY;
1195                 key.offset = hash_extent_data_ref(root_objectid,
1196                                                   owner, offset);
1197         }
1198 again:
1199         recow = 0;
1200         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1201         if (ret < 0) {
1202                 err = ret;
1203                 goto fail;
1204         }
1205
1206         if (parent) {
1207                 if (!ret)
1208                         return 0;
1209 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1210                 key.type = BTRFS_EXTENT_REF_V0_KEY;
1211                 btrfs_release_path(path);
1212                 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1213                 if (ret < 0) {
1214                         err = ret;
1215                         goto fail;
1216                 }
1217                 if (!ret)
1218                         return 0;
1219 #endif
1220                 goto fail;
1221         }
1222
1223         leaf = path->nodes[0];
1224         nritems = btrfs_header_nritems(leaf);
1225         while (1) {
1226                 if (path->slots[0] >= nritems) {
1227                         ret = btrfs_next_leaf(root, path);
1228                         if (ret < 0)
1229                                 err = ret;
1230                         if (ret)
1231                                 goto fail;
1232
1233                         leaf = path->nodes[0];
1234                         nritems = btrfs_header_nritems(leaf);
1235                         recow = 1;
1236                 }
1237
1238                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1239                 if (key.objectid != bytenr ||
1240                     key.type != BTRFS_EXTENT_DATA_REF_KEY)
1241                         goto fail;
1242
1243                 ref = btrfs_item_ptr(leaf, path->slots[0],
1244                                      struct btrfs_extent_data_ref);
1245
1246                 if (match_extent_data_ref(leaf, ref, root_objectid,
1247                                           owner, offset)) {
1248                         if (recow) {
1249                                 btrfs_release_path(path);
1250                                 goto again;
1251                         }
1252                         err = 0;
1253                         break;
1254                 }
1255                 path->slots[0]++;
1256         }
1257 fail:
1258         return err;
1259 }
1260
1261 static noinline int insert_extent_data_ref(struct btrfs_trans_handle *trans,
1262                                            struct btrfs_root *root,
1263                                            struct btrfs_path *path,
1264                                            u64 bytenr, u64 parent,
1265                                            u64 root_objectid, u64 owner,
1266                                            u64 offset, int refs_to_add)
1267 {
1268         struct btrfs_key key;
1269         struct extent_buffer *leaf;
1270         u32 size;
1271         u32 num_refs;
1272         int ret;
1273
1274         key.objectid = bytenr;
1275         if (parent) {
1276                 key.type = BTRFS_SHARED_DATA_REF_KEY;
1277                 key.offset = parent;
1278                 size = sizeof(struct btrfs_shared_data_ref);
1279         } else {
1280                 key.type = BTRFS_EXTENT_DATA_REF_KEY;
1281                 key.offset = hash_extent_data_ref(root_objectid,
1282                                                   owner, offset);
1283                 size = sizeof(struct btrfs_extent_data_ref);
1284         }
1285
1286         ret = btrfs_insert_empty_item(trans, root, path, &key, size);
1287         if (ret && ret != -EEXIST)
1288                 goto fail;
1289
1290         leaf = path->nodes[0];
1291         if (parent) {
1292                 struct btrfs_shared_data_ref *ref;
1293                 ref = btrfs_item_ptr(leaf, path->slots[0],
1294                                      struct btrfs_shared_data_ref);
1295                 if (ret == 0) {
1296                         btrfs_set_shared_data_ref_count(leaf, ref, refs_to_add);
1297                 } else {
1298                         num_refs = btrfs_shared_data_ref_count(leaf, ref);
1299                         num_refs += refs_to_add;
1300                         btrfs_set_shared_data_ref_count(leaf, ref, num_refs);
1301                 }
1302         } else {
1303                 struct btrfs_extent_data_ref *ref;
1304                 while (ret == -EEXIST) {
1305                         ref = btrfs_item_ptr(leaf, path->slots[0],
1306                                              struct btrfs_extent_data_ref);
1307                         if (match_extent_data_ref(leaf, ref, root_objectid,
1308                                                   owner, offset))
1309                                 break;
1310                         btrfs_release_path(path);
1311                         key.offset++;
1312                         ret = btrfs_insert_empty_item(trans, root, path, &key,
1313                                                       size);
1314                         if (ret && ret != -EEXIST)
1315                                 goto fail;
1316
1317                         leaf = path->nodes[0];
1318                 }
1319                 ref = btrfs_item_ptr(leaf, path->slots[0],
1320                                      struct btrfs_extent_data_ref);
1321                 if (ret == 0) {
1322                         btrfs_set_extent_data_ref_root(leaf, ref,
1323                                                        root_objectid);
1324                         btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
1325                         btrfs_set_extent_data_ref_offset(leaf, ref, offset);
1326                         btrfs_set_extent_data_ref_count(leaf, ref, refs_to_add);
1327                 } else {
1328                         num_refs = btrfs_extent_data_ref_count(leaf, ref);
1329                         num_refs += refs_to_add;
1330                         btrfs_set_extent_data_ref_count(leaf, ref, num_refs);
1331                 }
1332         }
1333         btrfs_mark_buffer_dirty(leaf);
1334         ret = 0;
1335 fail:
1336         btrfs_release_path(path);
1337         return ret;
1338 }
1339
1340 static noinline int remove_extent_data_ref(struct btrfs_trans_handle *trans,
1341                                            struct btrfs_root *root,
1342                                            struct btrfs_path *path,
1343                                            int refs_to_drop, int *last_ref)
1344 {
1345         struct btrfs_key key;
1346         struct btrfs_extent_data_ref *ref1 = NULL;
1347         struct btrfs_shared_data_ref *ref2 = NULL;
1348         struct extent_buffer *leaf;
1349         u32 num_refs = 0;
1350         int ret = 0;
1351
1352         leaf = path->nodes[0];
1353         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1354
1355         if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
1356                 ref1 = btrfs_item_ptr(leaf, path->slots[0],
1357                                       struct btrfs_extent_data_ref);
1358                 num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1359         } else if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
1360                 ref2 = btrfs_item_ptr(leaf, path->slots[0],
1361                                       struct btrfs_shared_data_ref);
1362                 num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1363 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1364         } else if (key.type == BTRFS_EXTENT_REF_V0_KEY) {
1365                 struct btrfs_extent_ref_v0 *ref0;
1366                 ref0 = btrfs_item_ptr(leaf, path->slots[0],
1367                                       struct btrfs_extent_ref_v0);
1368                 num_refs = btrfs_ref_count_v0(leaf, ref0);
1369 #endif
1370         } else {
1371                 BUG();
1372         }
1373
1374         BUG_ON(num_refs < refs_to_drop);
1375         num_refs -= refs_to_drop;
1376
1377         if (num_refs == 0) {
1378                 ret = btrfs_del_item(trans, root, path);
1379                 *last_ref = 1;
1380         } else {
1381                 if (key.type == BTRFS_EXTENT_DATA_REF_KEY)
1382                         btrfs_set_extent_data_ref_count(leaf, ref1, num_refs);
1383                 else if (key.type == BTRFS_SHARED_DATA_REF_KEY)
1384                         btrfs_set_shared_data_ref_count(leaf, ref2, num_refs);
1385 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1386                 else {
1387                         struct btrfs_extent_ref_v0 *ref0;
1388                         ref0 = btrfs_item_ptr(leaf, path->slots[0],
1389                                         struct btrfs_extent_ref_v0);
1390                         btrfs_set_ref_count_v0(leaf, ref0, num_refs);
1391                 }
1392 #endif
1393                 btrfs_mark_buffer_dirty(leaf);
1394         }
1395         return ret;
1396 }
1397
1398 static noinline u32 extent_data_ref_count(struct btrfs_path *path,
1399                                           struct btrfs_extent_inline_ref *iref)
1400 {
1401         struct btrfs_key key;
1402         struct extent_buffer *leaf;
1403         struct btrfs_extent_data_ref *ref1;
1404         struct btrfs_shared_data_ref *ref2;
1405         u32 num_refs = 0;
1406
1407         leaf = path->nodes[0];
1408         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1409         if (iref) {
1410                 if (btrfs_extent_inline_ref_type(leaf, iref) ==
1411                     BTRFS_EXTENT_DATA_REF_KEY) {
1412                         ref1 = (struct btrfs_extent_data_ref *)(&iref->offset);
1413                         num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1414                 } else {
1415                         ref2 = (struct btrfs_shared_data_ref *)(iref + 1);
1416                         num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1417                 }
1418         } else if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
1419                 ref1 = btrfs_item_ptr(leaf, path->slots[0],
1420                                       struct btrfs_extent_data_ref);
1421                 num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1422         } else if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
1423                 ref2 = btrfs_item_ptr(leaf, path->slots[0],
1424                                       struct btrfs_shared_data_ref);
1425                 num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1426 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1427         } else if (key.type == BTRFS_EXTENT_REF_V0_KEY) {
1428                 struct btrfs_extent_ref_v0 *ref0;
1429                 ref0 = btrfs_item_ptr(leaf, path->slots[0],
1430                                       struct btrfs_extent_ref_v0);
1431                 num_refs = btrfs_ref_count_v0(leaf, ref0);
1432 #endif
1433         } else {
1434                 WARN_ON(1);
1435         }
1436         return num_refs;
1437 }
1438
1439 static noinline int lookup_tree_block_ref(struct btrfs_trans_handle *trans,
1440                                           struct btrfs_root *root,
1441                                           struct btrfs_path *path,
1442                                           u64 bytenr, u64 parent,
1443                                           u64 root_objectid)
1444 {
1445         struct btrfs_key key;
1446         int ret;
1447
1448         key.objectid = bytenr;
1449         if (parent) {
1450                 key.type = BTRFS_SHARED_BLOCK_REF_KEY;
1451                 key.offset = parent;
1452         } else {
1453                 key.type = BTRFS_TREE_BLOCK_REF_KEY;
1454                 key.offset = root_objectid;
1455         }
1456
1457         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1458         if (ret > 0)
1459                 ret = -ENOENT;
1460 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1461         if (ret == -ENOENT && parent) {
1462                 btrfs_release_path(path);
1463                 key.type = BTRFS_EXTENT_REF_V0_KEY;
1464                 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1465                 if (ret > 0)
1466                         ret = -ENOENT;
1467         }
1468 #endif
1469         return ret;
1470 }
1471
1472 static noinline int insert_tree_block_ref(struct btrfs_trans_handle *trans,
1473                                           struct btrfs_root *root,
1474                                           struct btrfs_path *path,
1475                                           u64 bytenr, u64 parent,
1476                                           u64 root_objectid)
1477 {
1478         struct btrfs_key key;
1479         int ret;
1480
1481         key.objectid = bytenr;
1482         if (parent) {
1483                 key.type = BTRFS_SHARED_BLOCK_REF_KEY;
1484                 key.offset = parent;
1485         } else {
1486                 key.type = BTRFS_TREE_BLOCK_REF_KEY;
1487                 key.offset = root_objectid;
1488         }
1489
1490         ret = btrfs_insert_empty_item(trans, root, path, &key, 0);
1491         btrfs_release_path(path);
1492         return ret;
1493 }
1494
1495 static inline int extent_ref_type(u64 parent, u64 owner)
1496 {
1497         int type;
1498         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1499                 if (parent > 0)
1500                         type = BTRFS_SHARED_BLOCK_REF_KEY;
1501                 else
1502                         type = BTRFS_TREE_BLOCK_REF_KEY;
1503         } else {
1504                 if (parent > 0)
1505                         type = BTRFS_SHARED_DATA_REF_KEY;
1506                 else
1507                         type = BTRFS_EXTENT_DATA_REF_KEY;
1508         }
1509         return type;
1510 }
1511
1512 static int find_next_key(struct btrfs_path *path, int level,
1513                          struct btrfs_key *key)
1514
1515 {
1516         for (; level < BTRFS_MAX_LEVEL; level++) {
1517                 if (!path->nodes[level])
1518                         break;
1519                 if (path->slots[level] + 1 >=
1520                     btrfs_header_nritems(path->nodes[level]))
1521                         continue;
1522                 if (level == 0)
1523                         btrfs_item_key_to_cpu(path->nodes[level], key,
1524                                               path->slots[level] + 1);
1525                 else
1526                         btrfs_node_key_to_cpu(path->nodes[level], key,
1527                                               path->slots[level] + 1);
1528                 return 0;
1529         }
1530         return 1;
1531 }
1532
1533 /*
1534  * look for inline back ref. if back ref is found, *ref_ret is set
1535  * to the address of inline back ref, and 0 is returned.
1536  *
1537  * if back ref isn't found, *ref_ret is set to the address where it
1538  * should be inserted, and -ENOENT is returned.
1539  *
1540  * if insert is true and there are too many inline back refs, the path
1541  * points to the extent item, and -EAGAIN is returned.
1542  *
1543  * NOTE: inline back refs are ordered in the same way that back ref
1544  *       items in the tree are ordered.
1545  */
1546 static noinline_for_stack
1547 int lookup_inline_extent_backref(struct btrfs_trans_handle *trans,
1548                                  struct btrfs_root *root,
1549                                  struct btrfs_path *path,
1550                                  struct btrfs_extent_inline_ref **ref_ret,
1551                                  u64 bytenr, u64 num_bytes,
1552                                  u64 parent, u64 root_objectid,
1553                                  u64 owner, u64 offset, int insert)
1554 {
1555         struct btrfs_key key;
1556         struct extent_buffer *leaf;
1557         struct btrfs_extent_item *ei;
1558         struct btrfs_extent_inline_ref *iref;
1559         u64 flags;
1560         u64 item_size;
1561         unsigned long ptr;
1562         unsigned long end;
1563         int extra_size;
1564         int type;
1565         int want;
1566         int ret;
1567         int err = 0;
1568         bool skinny_metadata = btrfs_fs_incompat(root->fs_info,
1569                                                  SKINNY_METADATA);
1570
1571         key.objectid = bytenr;
1572         key.type = BTRFS_EXTENT_ITEM_KEY;
1573         key.offset = num_bytes;
1574
1575         want = extent_ref_type(parent, owner);
1576         if (insert) {
1577                 extra_size = btrfs_extent_inline_ref_size(want);
1578                 path->keep_locks = 1;
1579         } else
1580                 extra_size = -1;
1581
1582         /*
1583          * Owner is our parent level, so we can just add one to get the level
1584          * for the block we are interested in.
1585          */
1586         if (skinny_metadata && owner < BTRFS_FIRST_FREE_OBJECTID) {
1587                 key.type = BTRFS_METADATA_ITEM_KEY;
1588                 key.offset = owner;
1589         }
1590
1591 again:
1592         ret = btrfs_search_slot(trans, root, &key, path, extra_size, 1);
1593         if (ret < 0) {
1594                 err = ret;
1595                 goto out;
1596         }
1597
1598         /*
1599          * We may be a newly converted file system which still has the old fat
1600          * extent entries for metadata, so try and see if we have one of those.
1601          */
1602         if (ret > 0 && skinny_metadata) {
1603                 skinny_metadata = false;
1604                 if (path->slots[0]) {
1605                         path->slots[0]--;
1606                         btrfs_item_key_to_cpu(path->nodes[0], &key,
1607                                               path->slots[0]);
1608                         if (key.objectid == bytenr &&
1609                             key.type == BTRFS_EXTENT_ITEM_KEY &&
1610                             key.offset == num_bytes)
1611                                 ret = 0;
1612                 }
1613                 if (ret) {
1614                         key.objectid = bytenr;
1615                         key.type = BTRFS_EXTENT_ITEM_KEY;
1616                         key.offset = num_bytes;
1617                         btrfs_release_path(path);
1618                         goto again;
1619                 }
1620         }
1621
1622         if (ret && !insert) {
1623                 err = -ENOENT;
1624                 goto out;
1625         } else if (WARN_ON(ret)) {
1626                 err = -EIO;
1627                 goto out;
1628         }
1629
1630         leaf = path->nodes[0];
1631         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1632 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1633         if (item_size < sizeof(*ei)) {
1634                 if (!insert) {
1635                         err = -ENOENT;
1636                         goto out;
1637                 }
1638                 ret = convert_extent_item_v0(trans, root, path, owner,
1639                                              extra_size);
1640                 if (ret < 0) {
1641                         err = ret;
1642                         goto out;
1643                 }
1644                 leaf = path->nodes[0];
1645                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1646         }
1647 #endif
1648         BUG_ON(item_size < sizeof(*ei));
1649
1650         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1651         flags = btrfs_extent_flags(leaf, ei);
1652
1653         ptr = (unsigned long)(ei + 1);
1654         end = (unsigned long)ei + item_size;
1655
1656         if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK && !skinny_metadata) {
1657                 ptr += sizeof(struct btrfs_tree_block_info);
1658                 BUG_ON(ptr > end);
1659         }
1660
1661         err = -ENOENT;
1662         while (1) {
1663                 if (ptr >= end) {
1664                         WARN_ON(ptr > end);
1665                         break;
1666                 }
1667                 iref = (struct btrfs_extent_inline_ref *)ptr;
1668                 type = btrfs_extent_inline_ref_type(leaf, iref);
1669                 if (want < type)
1670                         break;
1671                 if (want > type) {
1672                         ptr += btrfs_extent_inline_ref_size(type);
1673                         continue;
1674                 }
1675
1676                 if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1677                         struct btrfs_extent_data_ref *dref;
1678                         dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1679                         if (match_extent_data_ref(leaf, dref, root_objectid,
1680                                                   owner, offset)) {
1681                                 err = 0;
1682                                 break;
1683                         }
1684                         if (hash_extent_data_ref_item(leaf, dref) <
1685                             hash_extent_data_ref(root_objectid, owner, offset))
1686                                 break;
1687                 } else {
1688                         u64 ref_offset;
1689                         ref_offset = btrfs_extent_inline_ref_offset(leaf, iref);
1690                         if (parent > 0) {
1691                                 if (parent == ref_offset) {
1692                                         err = 0;
1693                                         break;
1694                                 }
1695                                 if (ref_offset < parent)
1696                                         break;
1697                         } else {
1698                                 if (root_objectid == ref_offset) {
1699                                         err = 0;
1700                                         break;
1701                                 }
1702                                 if (ref_offset < root_objectid)
1703                                         break;
1704                         }
1705                 }
1706                 ptr += btrfs_extent_inline_ref_size(type);
1707         }
1708         if (err == -ENOENT && insert) {
1709                 if (item_size + extra_size >=
1710                     BTRFS_MAX_EXTENT_ITEM_SIZE(root)) {
1711                         err = -EAGAIN;
1712                         goto out;
1713                 }
1714                 /*
1715                  * To add new inline back ref, we have to make sure
1716                  * there is no corresponding back ref item.
1717                  * For simplicity, we just do not add new inline back
1718                  * ref if there is any kind of item for this block
1719                  */
1720                 if (find_next_key(path, 0, &key) == 0 &&
1721                     key.objectid == bytenr &&
1722                     key.type < BTRFS_BLOCK_GROUP_ITEM_KEY) {
1723                         err = -EAGAIN;
1724                         goto out;
1725                 }
1726         }
1727         *ref_ret = (struct btrfs_extent_inline_ref *)ptr;
1728 out:
1729         if (insert) {
1730                 path->keep_locks = 0;
1731                 btrfs_unlock_up_safe(path, 1);
1732         }
1733         return err;
1734 }
1735
1736 /*
1737  * helper to add new inline back ref
1738  */
1739 static noinline_for_stack
1740 void setup_inline_extent_backref(struct btrfs_root *root,
1741                                  struct btrfs_path *path,
1742                                  struct btrfs_extent_inline_ref *iref,
1743                                  u64 parent, u64 root_objectid,
1744                                  u64 owner, u64 offset, int refs_to_add,
1745                                  struct btrfs_delayed_extent_op *extent_op)
1746 {
1747         struct extent_buffer *leaf;
1748         struct btrfs_extent_item *ei;
1749         unsigned long ptr;
1750         unsigned long end;
1751         unsigned long item_offset;
1752         u64 refs;
1753         int size;
1754         int type;
1755
1756         leaf = path->nodes[0];
1757         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1758         item_offset = (unsigned long)iref - (unsigned long)ei;
1759
1760         type = extent_ref_type(parent, owner);
1761         size = btrfs_extent_inline_ref_size(type);
1762
1763         btrfs_extend_item(root, path, size);
1764
1765         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1766         refs = btrfs_extent_refs(leaf, ei);
1767         refs += refs_to_add;
1768         btrfs_set_extent_refs(leaf, ei, refs);
1769         if (extent_op)
1770                 __run_delayed_extent_op(extent_op, leaf, ei);
1771
1772         ptr = (unsigned long)ei + item_offset;
1773         end = (unsigned long)ei + btrfs_item_size_nr(leaf, path->slots[0]);
1774         if (ptr < end - size)
1775                 memmove_extent_buffer(leaf, ptr + size, ptr,
1776                                       end - size - ptr);
1777
1778         iref = (struct btrfs_extent_inline_ref *)ptr;
1779         btrfs_set_extent_inline_ref_type(leaf, iref, type);
1780         if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1781                 struct btrfs_extent_data_ref *dref;
1782                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1783                 btrfs_set_extent_data_ref_root(leaf, dref, root_objectid);
1784                 btrfs_set_extent_data_ref_objectid(leaf, dref, owner);
1785                 btrfs_set_extent_data_ref_offset(leaf, dref, offset);
1786                 btrfs_set_extent_data_ref_count(leaf, dref, refs_to_add);
1787         } else if (type == BTRFS_SHARED_DATA_REF_KEY) {
1788                 struct btrfs_shared_data_ref *sref;
1789                 sref = (struct btrfs_shared_data_ref *)(iref + 1);
1790                 btrfs_set_shared_data_ref_count(leaf, sref, refs_to_add);
1791                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
1792         } else if (type == BTRFS_SHARED_BLOCK_REF_KEY) {
1793                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
1794         } else {
1795                 btrfs_set_extent_inline_ref_offset(leaf, iref, root_objectid);
1796         }
1797         btrfs_mark_buffer_dirty(leaf);
1798 }
1799
1800 static int lookup_extent_backref(struct btrfs_trans_handle *trans,
1801                                  struct btrfs_root *root,
1802                                  struct btrfs_path *path,
1803                                  struct btrfs_extent_inline_ref **ref_ret,
1804                                  u64 bytenr, u64 num_bytes, u64 parent,
1805                                  u64 root_objectid, u64 owner, u64 offset)
1806 {
1807         int ret;
1808
1809         ret = lookup_inline_extent_backref(trans, root, path, ref_ret,
1810                                            bytenr, num_bytes, parent,
1811                                            root_objectid, owner, offset, 0);
1812         if (ret != -ENOENT)
1813                 return ret;
1814
1815         btrfs_release_path(path);
1816         *ref_ret = NULL;
1817
1818         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1819                 ret = lookup_tree_block_ref(trans, root, path, bytenr, parent,
1820                                             root_objectid);
1821         } else {
1822                 ret = lookup_extent_data_ref(trans, root, path, bytenr, parent,
1823                                              root_objectid, owner, offset);
1824         }
1825         return ret;
1826 }
1827
1828 /*
1829  * helper to update/remove inline back ref
1830  */
1831 static noinline_for_stack
1832 void update_inline_extent_backref(struct btrfs_root *root,
1833                                   struct btrfs_path *path,
1834                                   struct btrfs_extent_inline_ref *iref,
1835                                   int refs_to_mod,
1836                                   struct btrfs_delayed_extent_op *extent_op,
1837                                   int *last_ref)
1838 {
1839         struct extent_buffer *leaf;
1840         struct btrfs_extent_item *ei;
1841         struct btrfs_extent_data_ref *dref = NULL;
1842         struct btrfs_shared_data_ref *sref = NULL;
1843         unsigned long ptr;
1844         unsigned long end;
1845         u32 item_size;
1846         int size;
1847         int type;
1848         u64 refs;
1849
1850         leaf = path->nodes[0];
1851         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1852         refs = btrfs_extent_refs(leaf, ei);
1853         WARN_ON(refs_to_mod < 0 && refs + refs_to_mod <= 0);
1854         refs += refs_to_mod;
1855         btrfs_set_extent_refs(leaf, ei, refs);
1856         if (extent_op)
1857                 __run_delayed_extent_op(extent_op, leaf, ei);
1858
1859         type = btrfs_extent_inline_ref_type(leaf, iref);
1860
1861         if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1862                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1863                 refs = btrfs_extent_data_ref_count(leaf, dref);
1864         } else if (type == BTRFS_SHARED_DATA_REF_KEY) {
1865                 sref = (struct btrfs_shared_data_ref *)(iref + 1);
1866                 refs = btrfs_shared_data_ref_count(leaf, sref);
1867         } else {
1868                 refs = 1;
1869                 BUG_ON(refs_to_mod != -1);
1870         }
1871
1872         BUG_ON(refs_to_mod < 0 && refs < -refs_to_mod);
1873         refs += refs_to_mod;
1874
1875         if (refs > 0) {
1876                 if (type == BTRFS_EXTENT_DATA_REF_KEY)
1877                         btrfs_set_extent_data_ref_count(leaf, dref, refs);
1878                 else
1879                         btrfs_set_shared_data_ref_count(leaf, sref, refs);
1880         } else {
1881                 *last_ref = 1;
1882                 size =  btrfs_extent_inline_ref_size(type);
1883                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1884                 ptr = (unsigned long)iref;
1885                 end = (unsigned long)ei + item_size;
1886                 if (ptr + size < end)
1887                         memmove_extent_buffer(leaf, ptr, ptr + size,
1888                                               end - ptr - size);
1889                 item_size -= size;
1890                 btrfs_truncate_item(root, path, item_size, 1);
1891         }
1892         btrfs_mark_buffer_dirty(leaf);
1893 }
1894
1895 static noinline_for_stack
1896 int insert_inline_extent_backref(struct btrfs_trans_handle *trans,
1897                                  struct btrfs_root *root,
1898                                  struct btrfs_path *path,
1899                                  u64 bytenr, u64 num_bytes, u64 parent,
1900                                  u64 root_objectid, u64 owner,
1901                                  u64 offset, int refs_to_add,
1902                                  struct btrfs_delayed_extent_op *extent_op)
1903 {
1904         struct btrfs_extent_inline_ref *iref;
1905         int ret;
1906
1907         ret = lookup_inline_extent_backref(trans, root, path, &iref,
1908                                            bytenr, num_bytes, parent,
1909                                            root_objectid, owner, offset, 1);
1910         if (ret == 0) {
1911                 BUG_ON(owner < BTRFS_FIRST_FREE_OBJECTID);
1912                 update_inline_extent_backref(root, path, iref,
1913                                              refs_to_add, extent_op, NULL);
1914         } else if (ret == -ENOENT) {
1915                 setup_inline_extent_backref(root, path, iref, parent,
1916                                             root_objectid, owner, offset,
1917                                             refs_to_add, extent_op);
1918                 ret = 0;
1919         }
1920         return ret;
1921 }
1922
1923 static int insert_extent_backref(struct btrfs_trans_handle *trans,
1924                                  struct btrfs_root *root,
1925                                  struct btrfs_path *path,
1926                                  u64 bytenr, u64 parent, u64 root_objectid,
1927                                  u64 owner, u64 offset, int refs_to_add)
1928 {
1929         int ret;
1930         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1931                 BUG_ON(refs_to_add != 1);
1932                 ret = insert_tree_block_ref(trans, root, path, bytenr,
1933                                             parent, root_objectid);
1934         } else {
1935                 ret = insert_extent_data_ref(trans, root, path, bytenr,
1936                                              parent, root_objectid,
1937                                              owner, offset, refs_to_add);
1938         }
1939         return ret;
1940 }
1941
1942 static int remove_extent_backref(struct btrfs_trans_handle *trans,
1943                                  struct btrfs_root *root,
1944                                  struct btrfs_path *path,
1945                                  struct btrfs_extent_inline_ref *iref,
1946                                  int refs_to_drop, int is_data, int *last_ref)
1947 {
1948         int ret = 0;
1949
1950         BUG_ON(!is_data && refs_to_drop != 1);
1951         if (iref) {
1952                 update_inline_extent_backref(root, path, iref,
1953                                              -refs_to_drop, NULL, last_ref);
1954         } else if (is_data) {
1955                 ret = remove_extent_data_ref(trans, root, path, refs_to_drop,
1956                                              last_ref);
1957         } else {
1958                 *last_ref = 1;
1959                 ret = btrfs_del_item(trans, root, path);
1960         }
1961         return ret;
1962 }
1963
1964 #define in_range(b, first, len)        ((b) >= (first) && (b) < (first) + (len))
1965 static int btrfs_issue_discard(struct block_device *bdev, u64 start, u64 len,
1966                                u64 *discarded_bytes)
1967 {
1968         int j, ret = 0;
1969         u64 bytes_left, end;
1970         u64 aligned_start = ALIGN(start, 1 << 9);
1971
1972         if (WARN_ON(start != aligned_start)) {
1973                 len -= aligned_start - start;
1974                 len = round_down(len, 1 << 9);
1975                 start = aligned_start;
1976         }
1977
1978         *discarded_bytes = 0;
1979
1980         if (!len)
1981                 return 0;
1982
1983         end = start + len;
1984         bytes_left = len;
1985
1986         /* Skip any superblocks on this device. */
1987         for (j = 0; j < BTRFS_SUPER_MIRROR_MAX; j++) {
1988                 u64 sb_start = btrfs_sb_offset(j);
1989                 u64 sb_end = sb_start + BTRFS_SUPER_INFO_SIZE;
1990                 u64 size = sb_start - start;
1991
1992                 if (!in_range(sb_start, start, bytes_left) &&
1993                     !in_range(sb_end, start, bytes_left) &&
1994                     !in_range(start, sb_start, BTRFS_SUPER_INFO_SIZE))
1995                         continue;
1996
1997                 /*
1998                  * Superblock spans beginning of range.  Adjust start and
1999                  * try again.
2000                  */
2001                 if (sb_start <= start) {
2002                         start += sb_end - start;
2003                         if (start > end) {
2004                                 bytes_left = 0;
2005                                 break;
2006                         }
2007                         bytes_left = end - start;
2008                         continue;
2009                 }
2010
2011                 if (size) {
2012                         ret = blkdev_issue_discard(bdev, start >> 9, size >> 9,
2013                                                    GFP_NOFS, 0);
2014                         if (!ret)
2015                                 *discarded_bytes += size;
2016                         else if (ret != -EOPNOTSUPP)
2017                                 return ret;
2018                 }
2019
2020                 start = sb_end;
2021                 if (start > end) {
2022                         bytes_left = 0;
2023                         break;
2024                 }
2025                 bytes_left = end - start;
2026         }
2027
2028         if (bytes_left) {
2029                 ret = blkdev_issue_discard(bdev, start >> 9, bytes_left >> 9,
2030                                            GFP_NOFS, 0);
2031                 if (!ret)
2032                         *discarded_bytes += bytes_left;
2033         }
2034         return ret;
2035 }
2036
2037 int btrfs_discard_extent(struct btrfs_root *root, u64 bytenr,
2038                          u64 num_bytes, u64 *actual_bytes)
2039 {
2040         int ret;
2041         u64 discarded_bytes = 0;
2042         struct btrfs_bio *bbio = NULL;
2043
2044
2045         /* Tell the block device(s) that the sectors can be discarded */
2046         ret = btrfs_map_block(root->fs_info, REQ_DISCARD,
2047                               bytenr, &num_bytes, &bbio, 0);
2048         /* Error condition is -ENOMEM */
2049         if (!ret) {
2050                 struct btrfs_bio_stripe *stripe = bbio->stripes;
2051                 int i;
2052
2053
2054                 for (i = 0; i < bbio->num_stripes; i++, stripe++) {
2055                         u64 bytes;
2056                         if (!stripe->dev->can_discard)
2057                                 continue;
2058
2059                         ret = btrfs_issue_discard(stripe->dev->bdev,
2060                                                   stripe->physical,
2061                                                   stripe->length,
2062                                                   &bytes);
2063                         if (!ret)
2064                                 discarded_bytes += bytes;
2065                         else if (ret != -EOPNOTSUPP)
2066                                 break; /* Logic errors or -ENOMEM, or -EIO but I don't know how that could happen JDM */
2067
2068                         /*
2069                          * Just in case we get back EOPNOTSUPP for some reason,
2070                          * just ignore the return value so we don't screw up
2071                          * people calling discard_extent.
2072                          */
2073                         ret = 0;
2074                 }
2075                 btrfs_put_bbio(bbio);
2076         }
2077
2078         if (actual_bytes)
2079                 *actual_bytes = discarded_bytes;
2080
2081
2082         if (ret == -EOPNOTSUPP)
2083                 ret = 0;
2084         return ret;
2085 }
2086
2087 /* Can return -ENOMEM */
2088 int btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
2089                          struct btrfs_root *root,
2090                          u64 bytenr, u64 num_bytes, u64 parent,
2091                          u64 root_objectid, u64 owner, u64 offset)
2092 {
2093         int ret;
2094         struct btrfs_fs_info *fs_info = root->fs_info;
2095
2096         BUG_ON(owner < BTRFS_FIRST_FREE_OBJECTID &&
2097                root_objectid == BTRFS_TREE_LOG_OBJECTID);
2098
2099         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
2100                 ret = btrfs_add_delayed_tree_ref(fs_info, trans, bytenr,
2101                                         num_bytes,
2102                                         parent, root_objectid, (int)owner,
2103                                         BTRFS_ADD_DELAYED_REF, NULL);
2104         } else {
2105                 ret = btrfs_add_delayed_data_ref(fs_info, trans, bytenr,
2106                                         num_bytes, parent, root_objectid,
2107                                         owner, offset, 0,
2108                                         BTRFS_ADD_DELAYED_REF, NULL);
2109         }
2110         return ret;
2111 }
2112
2113 static int __btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
2114                                   struct btrfs_root *root,
2115                                   struct btrfs_delayed_ref_node *node,
2116                                   u64 parent, u64 root_objectid,
2117                                   u64 owner, u64 offset, int refs_to_add,
2118                                   struct btrfs_delayed_extent_op *extent_op)
2119 {
2120         struct btrfs_fs_info *fs_info = root->fs_info;
2121         struct btrfs_path *path;
2122         struct extent_buffer *leaf;
2123         struct btrfs_extent_item *item;
2124         struct btrfs_key key;
2125         u64 bytenr = node->bytenr;
2126         u64 num_bytes = node->num_bytes;
2127         u64 refs;
2128         int ret;
2129
2130         path = btrfs_alloc_path();
2131         if (!path)
2132                 return -ENOMEM;
2133
2134         path->reada = 1;
2135         path->leave_spinning = 1;
2136         /* this will setup the path even if it fails to insert the back ref */
2137         ret = insert_inline_extent_backref(trans, fs_info->extent_root, path,
2138                                            bytenr, num_bytes, parent,
2139                                            root_objectid, owner, offset,
2140                                            refs_to_add, extent_op);
2141         if ((ret < 0 && ret != -EAGAIN) || !ret)
2142                 goto out;
2143
2144         /*
2145          * Ok we had -EAGAIN which means we didn't have space to insert and
2146          * inline extent ref, so just update the reference count and add a
2147          * normal backref.
2148          */
2149         leaf = path->nodes[0];
2150         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
2151         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2152         refs = btrfs_extent_refs(leaf, item);
2153         btrfs_set_extent_refs(leaf, item, refs + refs_to_add);
2154         if (extent_op)
2155                 __run_delayed_extent_op(extent_op, leaf, item);
2156
2157         btrfs_mark_buffer_dirty(leaf);
2158         btrfs_release_path(path);
2159
2160         path->reada = 1;
2161         path->leave_spinning = 1;
2162         /* now insert the actual backref */
2163         ret = insert_extent_backref(trans, root->fs_info->extent_root,
2164                                     path, bytenr, parent, root_objectid,
2165                                     owner, offset, refs_to_add);
2166         if (ret)
2167                 btrfs_abort_transaction(trans, root, ret);
2168 out:
2169         btrfs_free_path(path);
2170         return ret;
2171 }
2172
2173 static int run_delayed_data_ref(struct btrfs_trans_handle *trans,
2174                                 struct btrfs_root *root,
2175                                 struct btrfs_delayed_ref_node *node,
2176                                 struct btrfs_delayed_extent_op *extent_op,
2177                                 int insert_reserved)
2178 {
2179         int ret = 0;
2180         struct btrfs_delayed_data_ref *ref;
2181         struct btrfs_key ins;
2182         u64 parent = 0;
2183         u64 ref_root = 0;
2184         u64 flags = 0;
2185
2186         ins.objectid = node->bytenr;
2187         ins.offset = node->num_bytes;
2188         ins.type = BTRFS_EXTENT_ITEM_KEY;
2189
2190         ref = btrfs_delayed_node_to_data_ref(node);
2191         trace_run_delayed_data_ref(node, ref, node->action);
2192
2193         if (node->type == BTRFS_SHARED_DATA_REF_KEY)
2194                 parent = ref->parent;
2195         ref_root = ref->root;
2196
2197         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
2198                 if (extent_op)
2199                         flags |= extent_op->flags_to_set;
2200                 ret = alloc_reserved_file_extent(trans, root,
2201                                                  parent, ref_root, flags,
2202                                                  ref->objectid, ref->offset,
2203                                                  &ins, node->ref_mod);
2204         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
2205                 ret = __btrfs_inc_extent_ref(trans, root, node, parent,
2206                                              ref_root, ref->objectid,
2207                                              ref->offset, node->ref_mod,
2208                                              extent_op);
2209         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
2210                 ret = __btrfs_free_extent(trans, root, node, parent,
2211                                           ref_root, ref->objectid,
2212                                           ref->offset, node->ref_mod,
2213                                           extent_op);
2214         } else {
2215                 BUG();
2216         }
2217         return ret;
2218 }
2219
2220 static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op,
2221                                     struct extent_buffer *leaf,
2222                                     struct btrfs_extent_item *ei)
2223 {
2224         u64 flags = btrfs_extent_flags(leaf, ei);
2225         if (extent_op->update_flags) {
2226                 flags |= extent_op->flags_to_set;
2227                 btrfs_set_extent_flags(leaf, ei, flags);
2228         }
2229
2230         if (extent_op->update_key) {
2231                 struct btrfs_tree_block_info *bi;
2232                 BUG_ON(!(flags & BTRFS_EXTENT_FLAG_TREE_BLOCK));
2233                 bi = (struct btrfs_tree_block_info *)(ei + 1);
2234                 btrfs_set_tree_block_key(leaf, bi, &extent_op->key);
2235         }
2236 }
2237
2238 static int run_delayed_extent_op(struct btrfs_trans_handle *trans,
2239                                  struct btrfs_root *root,
2240                                  struct btrfs_delayed_ref_node *node,
2241                                  struct btrfs_delayed_extent_op *extent_op)
2242 {
2243         struct btrfs_key key;
2244         struct btrfs_path *path;
2245         struct btrfs_extent_item *ei;
2246         struct extent_buffer *leaf;
2247         u32 item_size;
2248         int ret;
2249         int err = 0;
2250         int metadata = !extent_op->is_data;
2251
2252         if (trans->aborted)
2253                 return 0;
2254
2255         if (metadata && !btrfs_fs_incompat(root->fs_info, SKINNY_METADATA))
2256                 metadata = 0;
2257
2258         path = btrfs_alloc_path();
2259         if (!path)
2260                 return -ENOMEM;
2261
2262         key.objectid = node->bytenr;
2263
2264         if (metadata) {
2265                 key.type = BTRFS_METADATA_ITEM_KEY;
2266                 key.offset = extent_op->level;
2267         } else {
2268                 key.type = BTRFS_EXTENT_ITEM_KEY;
2269                 key.offset = node->num_bytes;
2270         }
2271
2272 again:
2273         path->reada = 1;
2274         path->leave_spinning = 1;
2275         ret = btrfs_search_slot(trans, root->fs_info->extent_root, &key,
2276                                 path, 0, 1);
2277         if (ret < 0) {
2278                 err = ret;
2279                 goto out;
2280         }
2281         if (ret > 0) {
2282                 if (metadata) {
2283                         if (path->slots[0] > 0) {
2284                                 path->slots[0]--;
2285                                 btrfs_item_key_to_cpu(path->nodes[0], &key,
2286                                                       path->slots[0]);
2287                                 if (key.objectid == node->bytenr &&
2288                                     key.type == BTRFS_EXTENT_ITEM_KEY &&
2289                                     key.offset == node->num_bytes)
2290                                         ret = 0;
2291                         }
2292                         if (ret > 0) {
2293                                 btrfs_release_path(path);
2294                                 metadata = 0;
2295
2296                                 key.objectid = node->bytenr;
2297                                 key.offset = node->num_bytes;
2298                                 key.type = BTRFS_EXTENT_ITEM_KEY;
2299                                 goto again;
2300                         }
2301                 } else {
2302                         err = -EIO;
2303                         goto out;
2304                 }
2305         }
2306
2307         leaf = path->nodes[0];
2308         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2309 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
2310         if (item_size < sizeof(*ei)) {
2311                 ret = convert_extent_item_v0(trans, root->fs_info->extent_root,
2312                                              path, (u64)-1, 0);
2313                 if (ret < 0) {
2314                         err = ret;
2315                         goto out;
2316                 }
2317                 leaf = path->nodes[0];
2318                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2319         }
2320 #endif
2321         BUG_ON(item_size < sizeof(*ei));
2322         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2323         __run_delayed_extent_op(extent_op, leaf, ei);
2324
2325         btrfs_mark_buffer_dirty(leaf);
2326 out:
2327         btrfs_free_path(path);
2328         return err;
2329 }
2330
2331 static int run_delayed_tree_ref(struct btrfs_trans_handle *trans,
2332                                 struct btrfs_root *root,
2333                                 struct btrfs_delayed_ref_node *node,
2334                                 struct btrfs_delayed_extent_op *extent_op,
2335                                 int insert_reserved)
2336 {
2337         int ret = 0;
2338         struct btrfs_delayed_tree_ref *ref;
2339         struct btrfs_key ins;
2340         u64 parent = 0;
2341         u64 ref_root = 0;
2342         bool skinny_metadata = btrfs_fs_incompat(root->fs_info,
2343                                                  SKINNY_METADATA);
2344
2345         ref = btrfs_delayed_node_to_tree_ref(node);
2346         trace_run_delayed_tree_ref(node, ref, node->action);
2347
2348         if (node->type == BTRFS_SHARED_BLOCK_REF_KEY)
2349                 parent = ref->parent;
2350         ref_root = ref->root;
2351
2352         ins.objectid = node->bytenr;
2353         if (skinny_metadata) {
2354                 ins.offset = ref->level;
2355                 ins.type = BTRFS_METADATA_ITEM_KEY;
2356         } else {
2357                 ins.offset = node->num_bytes;
2358                 ins.type = BTRFS_EXTENT_ITEM_KEY;
2359         }
2360
2361         BUG_ON(node->ref_mod != 1);
2362         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
2363                 BUG_ON(!extent_op || !extent_op->update_flags);
2364                 ret = alloc_reserved_tree_block(trans, root,
2365                                                 parent, ref_root,
2366                                                 extent_op->flags_to_set,
2367                                                 &extent_op->key,
2368                                                 ref->level, &ins);
2369         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
2370                 ret = __btrfs_inc_extent_ref(trans, root, node,
2371                                              parent, ref_root,
2372                                              ref->level, 0, 1,
2373                                              extent_op);
2374         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
2375                 ret = __btrfs_free_extent(trans, root, node,
2376                                           parent, ref_root,
2377                                           ref->level, 0, 1, extent_op);
2378         } else {
2379                 BUG();
2380         }
2381         return ret;
2382 }
2383
2384 /* helper function to actually process a single delayed ref entry */
2385 static int run_one_delayed_ref(struct btrfs_trans_handle *trans,
2386                                struct btrfs_root *root,
2387                                struct btrfs_delayed_ref_node *node,
2388                                struct btrfs_delayed_extent_op *extent_op,
2389                                int insert_reserved)
2390 {
2391         int ret = 0;
2392
2393         if (trans->aborted) {
2394                 if (insert_reserved)
2395                         btrfs_pin_extent(root, node->bytenr,
2396                                          node->num_bytes, 1);
2397                 return 0;
2398         }
2399
2400         if (btrfs_delayed_ref_is_head(node)) {
2401                 struct btrfs_delayed_ref_head *head;
2402                 /*
2403                  * we've hit the end of the chain and we were supposed
2404                  * to insert this extent into the tree.  But, it got
2405                  * deleted before we ever needed to insert it, so all
2406                  * we have to do is clean up the accounting
2407                  */
2408                 BUG_ON(extent_op);
2409                 head = btrfs_delayed_node_to_head(node);
2410                 trace_run_delayed_ref_head(node, head, node->action);
2411
2412                 if (insert_reserved) {
2413                         btrfs_pin_extent(root, node->bytenr,
2414                                          node->num_bytes, 1);
2415                         if (head->is_data) {
2416                                 ret = btrfs_del_csums(trans, root,
2417                                                       node->bytenr,
2418                                                       node->num_bytes);
2419                         }
2420                 }
2421
2422                 /* Also free its reserved qgroup space */
2423                 btrfs_qgroup_free_delayed_ref(root->fs_info,
2424                                               head->qgroup_ref_root,
2425                                               head->qgroup_reserved);
2426                 return ret;
2427         }
2428
2429         if (node->type == BTRFS_TREE_BLOCK_REF_KEY ||
2430             node->type == BTRFS_SHARED_BLOCK_REF_KEY)
2431                 ret = run_delayed_tree_ref(trans, root, node, extent_op,
2432                                            insert_reserved);
2433         else if (node->type == BTRFS_EXTENT_DATA_REF_KEY ||
2434                  node->type == BTRFS_SHARED_DATA_REF_KEY)
2435                 ret = run_delayed_data_ref(trans, root, node, extent_op,
2436                                            insert_reserved);
2437         else
2438                 BUG();
2439         return ret;
2440 }
2441
2442 static inline struct btrfs_delayed_ref_node *
2443 select_delayed_ref(struct btrfs_delayed_ref_head *head)
2444 {
2445         struct btrfs_delayed_ref_node *ref;
2446
2447         if (list_empty(&head->ref_list))
2448                 return NULL;
2449
2450         /*
2451          * Select a delayed ref of type BTRFS_ADD_DELAYED_REF first.
2452          * This is to prevent a ref count from going down to zero, which deletes
2453          * the extent item from the extent tree, when there still are references
2454          * to add, which would fail because they would not find the extent item.
2455          */
2456         list_for_each_entry(ref, &head->ref_list, list) {
2457                 if (ref->action == BTRFS_ADD_DELAYED_REF)
2458                         return ref;
2459         }
2460
2461         return list_entry(head->ref_list.next, struct btrfs_delayed_ref_node,
2462                           list);
2463 }
2464
2465 /*
2466  * Returns 0 on success or if called with an already aborted transaction.
2467  * Returns -ENOMEM or -EIO on failure and will abort the transaction.
2468  */
2469 static noinline int __btrfs_run_delayed_refs(struct btrfs_trans_handle *trans,
2470                                              struct btrfs_root *root,
2471                                              unsigned long nr)
2472 {
2473         struct btrfs_delayed_ref_root *delayed_refs;
2474         struct btrfs_delayed_ref_node *ref;
2475         struct btrfs_delayed_ref_head *locked_ref = NULL;
2476         struct btrfs_delayed_extent_op *extent_op;
2477         struct btrfs_fs_info *fs_info = root->fs_info;
2478         ktime_t start = ktime_get();
2479         int ret;
2480         unsigned long count = 0;
2481         unsigned long actual_count = 0;
2482         int must_insert_reserved = 0;
2483
2484         delayed_refs = &trans->transaction->delayed_refs;
2485         while (1) {
2486                 if (!locked_ref) {
2487                         if (count >= nr)
2488                                 break;
2489
2490                         spin_lock(&delayed_refs->lock);
2491                         locked_ref = btrfs_select_ref_head(trans);
2492                         if (!locked_ref) {
2493                                 spin_unlock(&delayed_refs->lock);
2494                                 break;
2495                         }
2496
2497                         /* grab the lock that says we are going to process
2498                          * all the refs for this head */
2499                         ret = btrfs_delayed_ref_lock(trans, locked_ref);
2500                         spin_unlock(&delayed_refs->lock);
2501                         /*
2502                          * we may have dropped the spin lock to get the head
2503                          * mutex lock, and that might have given someone else
2504                          * time to free the head.  If that's true, it has been
2505                          * removed from our list and we can move on.
2506                          */
2507                         if (ret == -EAGAIN) {
2508                                 locked_ref = NULL;
2509                                 count++;
2510                                 continue;
2511                         }
2512                 }
2513
2514                 /*
2515                  * We need to try and merge add/drops of the same ref since we
2516                  * can run into issues with relocate dropping the implicit ref
2517                  * and then it being added back again before the drop can
2518                  * finish.  If we merged anything we need to re-loop so we can
2519                  * get a good ref.
2520                  * Or we can get node references of the same type that weren't
2521                  * merged when created due to bumps in the tree mod seq, and
2522                  * we need to merge them to prevent adding an inline extent
2523                  * backref before dropping it (triggering a BUG_ON at
2524                  * insert_inline_extent_backref()).
2525                  */
2526                 spin_lock(&locked_ref->lock);
2527                 btrfs_merge_delayed_refs(trans, fs_info, delayed_refs,
2528                                          locked_ref);
2529
2530                 /*
2531                  * locked_ref is the head node, so we have to go one
2532                  * node back for any delayed ref updates
2533                  */
2534                 ref = select_delayed_ref(locked_ref);
2535
2536                 if (ref && ref->seq &&
2537                     btrfs_check_delayed_seq(fs_info, delayed_refs, ref->seq)) {
2538                         spin_unlock(&locked_ref->lock);
2539                         btrfs_delayed_ref_unlock(locked_ref);
2540                         spin_lock(&delayed_refs->lock);
2541                         locked_ref->processing = 0;
2542                         delayed_refs->num_heads_ready++;
2543                         spin_unlock(&delayed_refs->lock);
2544                         locked_ref = NULL;
2545                         cond_resched();
2546                         count++;
2547                         continue;
2548                 }
2549
2550                 /*
2551                  * record the must insert reserved flag before we
2552                  * drop the spin lock.
2553                  */
2554                 must_insert_reserved = locked_ref->must_insert_reserved;
2555                 locked_ref->must_insert_reserved = 0;
2556
2557                 extent_op = locked_ref->extent_op;
2558                 locked_ref->extent_op = NULL;
2559
2560                 if (!ref) {
2561
2562
2563                         /* All delayed refs have been processed, Go ahead
2564                          * and send the head node to run_one_delayed_ref,
2565                          * so that any accounting fixes can happen
2566                          */
2567                         ref = &locked_ref->node;
2568
2569                         if (extent_op && must_insert_reserved) {
2570                                 btrfs_free_delayed_extent_op(extent_op);
2571                                 extent_op = NULL;
2572                         }
2573
2574                         if (extent_op) {
2575                                 spin_unlock(&locked_ref->lock);
2576                                 ret = run_delayed_extent_op(trans, root,
2577                                                             ref, extent_op);
2578                                 btrfs_free_delayed_extent_op(extent_op);
2579
2580                                 if (ret) {
2581                                         /*
2582                                          * Need to reset must_insert_reserved if
2583                                          * there was an error so the abort stuff
2584                                          * can cleanup the reserved space
2585                                          * properly.
2586                                          */
2587                                         if (must_insert_reserved)
2588                                                 locked_ref->must_insert_reserved = 1;
2589                                         locked_ref->processing = 0;
2590                                         btrfs_debug(fs_info, "run_delayed_extent_op returned %d", ret);
2591                                         btrfs_delayed_ref_unlock(locked_ref);
2592                                         return ret;
2593                                 }
2594                                 continue;
2595                         }
2596
2597                         /*
2598                          * Need to drop our head ref lock and re-aqcuire the
2599                          * delayed ref lock and then re-check to make sure
2600                          * nobody got added.
2601                          */
2602                         spin_unlock(&locked_ref->lock);
2603                         spin_lock(&delayed_refs->lock);
2604                         spin_lock(&locked_ref->lock);
2605                         if (!list_empty(&locked_ref->ref_list) ||
2606                             locked_ref->extent_op) {
2607                                 spin_unlock(&locked_ref->lock);
2608                                 spin_unlock(&delayed_refs->lock);
2609                                 continue;
2610                         }
2611                         ref->in_tree = 0;
2612                         delayed_refs->num_heads--;
2613                         rb_erase(&locked_ref->href_node,
2614                                  &delayed_refs->href_root);
2615                         spin_unlock(&delayed_refs->lock);
2616                 } else {
2617                         actual_count++;
2618                         ref->in_tree = 0;
2619                         list_del(&ref->list);
2620                 }
2621                 atomic_dec(&delayed_refs->num_entries);
2622
2623                 if (!btrfs_delayed_ref_is_head(ref)) {
2624                         /*
2625                          * when we play the delayed ref, also correct the
2626                          * ref_mod on head
2627                          */
2628                         switch (ref->action) {
2629                         case BTRFS_ADD_DELAYED_REF:
2630                         case BTRFS_ADD_DELAYED_EXTENT:
2631                                 locked_ref->node.ref_mod -= ref->ref_mod;
2632                                 break;
2633                         case BTRFS_DROP_DELAYED_REF:
2634                                 locked_ref->node.ref_mod += ref->ref_mod;
2635                                 break;
2636                         default:
2637                                 WARN_ON(1);
2638                         }
2639                 }
2640                 spin_unlock(&locked_ref->lock);
2641
2642                 ret = run_one_delayed_ref(trans, root, ref, extent_op,
2643                                           must_insert_reserved);
2644
2645                 btrfs_free_delayed_extent_op(extent_op);
2646                 if (ret) {
2647                         locked_ref->processing = 0;
2648                         btrfs_delayed_ref_unlock(locked_ref);
2649                         btrfs_put_delayed_ref(ref);
2650                         btrfs_debug(fs_info, "run_one_delayed_ref returned %d", ret);
2651                         return ret;
2652                 }
2653
2654                 /*
2655                  * If this node is a head, that means all the refs in this head
2656                  * have been dealt with, and we will pick the next head to deal
2657                  * with, so we must unlock the head and drop it from the cluster
2658                  * list before we release it.
2659                  */
2660                 if (btrfs_delayed_ref_is_head(ref)) {
2661                         if (locked_ref->is_data &&
2662                             locked_ref->total_ref_mod < 0) {
2663                                 spin_lock(&delayed_refs->lock);
2664                                 delayed_refs->pending_csums -= ref->num_bytes;
2665                                 spin_unlock(&delayed_refs->lock);
2666                         }
2667                         btrfs_delayed_ref_unlock(locked_ref);
2668                         locked_ref = NULL;
2669                 }
2670                 btrfs_put_delayed_ref(ref);
2671                 count++;
2672                 cond_resched();
2673         }
2674
2675         /*
2676          * We don't want to include ref heads since we can have empty ref heads
2677          * and those will drastically skew our runtime down since we just do
2678          * accounting, no actual extent tree updates.
2679          */
2680         if (actual_count > 0) {
2681                 u64 runtime = ktime_to_ns(ktime_sub(ktime_get(), start));
2682                 u64 avg;
2683
2684                 /*
2685                  * We weigh the current average higher than our current runtime
2686                  * to avoid large swings in the average.
2687                  */
2688                 spin_lock(&delayed_refs->lock);
2689                 avg = fs_info->avg_delayed_ref_runtime * 3 + runtime;
2690                 fs_info->avg_delayed_ref_runtime = avg >> 2;    /* div by 4 */
2691                 spin_unlock(&delayed_refs->lock);
2692         }
2693         return 0;
2694 }
2695
2696 #ifdef SCRAMBLE_DELAYED_REFS
2697 /*
2698  * Normally delayed refs get processed in ascending bytenr order. This
2699  * correlates in most cases to the order added. To expose dependencies on this
2700  * order, we start to process the tree in the middle instead of the beginning
2701  */
2702 static u64 find_middle(struct rb_root *root)
2703 {
2704         struct rb_node *n = root->rb_node;
2705         struct btrfs_delayed_ref_node *entry;
2706         int alt = 1;
2707         u64 middle;
2708         u64 first = 0, last = 0;
2709
2710         n = rb_first(root);
2711         if (n) {
2712                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2713                 first = entry->bytenr;
2714         }
2715         n = rb_last(root);
2716         if (n) {
2717                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2718                 last = entry->bytenr;
2719         }
2720         n = root->rb_node;
2721
2722         while (n) {
2723                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2724                 WARN_ON(!entry->in_tree);
2725
2726                 middle = entry->bytenr;
2727
2728                 if (alt)
2729                         n = n->rb_left;
2730                 else
2731                         n = n->rb_right;
2732
2733                 alt = 1 - alt;
2734         }
2735         return middle;
2736 }
2737 #endif
2738
2739 static inline u64 heads_to_leaves(struct btrfs_root *root, u64 heads)
2740 {
2741         u64 num_bytes;
2742
2743         num_bytes = heads * (sizeof(struct btrfs_extent_item) +
2744                              sizeof(struct btrfs_extent_inline_ref));
2745         if (!btrfs_fs_incompat(root->fs_info, SKINNY_METADATA))
2746                 num_bytes += heads * sizeof(struct btrfs_tree_block_info);
2747
2748         /*
2749          * We don't ever fill up leaves all the way so multiply by 2 just to be
2750          * closer to what we're really going to want to ouse.
2751          */
2752         return div_u64(num_bytes, BTRFS_LEAF_DATA_SIZE(root));
2753 }
2754
2755 /*
2756  * Takes the number of bytes to be csumm'ed and figures out how many leaves it
2757  * would require to store the csums for that many bytes.
2758  */
2759 u64 btrfs_csum_bytes_to_leaves(struct btrfs_root *root, u64 csum_bytes)
2760 {
2761         u64 csum_size;
2762         u64 num_csums_per_leaf;
2763         u64 num_csums;
2764
2765         csum_size = BTRFS_LEAF_DATA_SIZE(root) - sizeof(struct btrfs_item);
2766         num_csums_per_leaf = div64_u64(csum_size,
2767                         (u64)btrfs_super_csum_size(root->fs_info->super_copy));
2768         num_csums = div64_u64(csum_bytes, root->sectorsize);
2769         num_csums += num_csums_per_leaf - 1;
2770         num_csums = div64_u64(num_csums, num_csums_per_leaf);
2771         return num_csums;
2772 }
2773
2774 int btrfs_check_space_for_delayed_refs(struct btrfs_trans_handle *trans,
2775                                        struct btrfs_root *root)
2776 {
2777         struct btrfs_block_rsv *global_rsv;
2778         u64 num_heads = trans->transaction->delayed_refs.num_heads_ready;
2779         u64 csum_bytes = trans->transaction->delayed_refs.pending_csums;
2780         u64 num_dirty_bgs = trans->transaction->num_dirty_bgs;
2781         u64 num_bytes, num_dirty_bgs_bytes;
2782         int ret = 0;
2783
2784         num_bytes = btrfs_calc_trans_metadata_size(root, 1);
2785         num_heads = heads_to_leaves(root, num_heads);
2786         if (num_heads > 1)
2787                 num_bytes += (num_heads - 1) * root->nodesize;
2788         num_bytes <<= 1;
2789         num_bytes += btrfs_csum_bytes_to_leaves(root, csum_bytes) * root->nodesize;
2790         num_dirty_bgs_bytes = btrfs_calc_trans_metadata_size(root,
2791                                                              num_dirty_bgs);
2792         global_rsv = &root->fs_info->global_block_rsv;
2793
2794         /*
2795          * If we can't allocate any more chunks lets make sure we have _lots_ of
2796          * wiggle room since running delayed refs can create more delayed refs.
2797          */
2798         if (global_rsv->space_info->full) {
2799                 num_dirty_bgs_bytes <<= 1;
2800                 num_bytes <<= 1;
2801         }
2802
2803         spin_lock(&global_rsv->lock);
2804         if (global_rsv->reserved <= num_bytes + num_dirty_bgs_bytes)
2805                 ret = 1;
2806         spin_unlock(&global_rsv->lock);
2807         return ret;
2808 }
2809
2810 int btrfs_should_throttle_delayed_refs(struct btrfs_trans_handle *trans,
2811                                        struct btrfs_root *root)
2812 {
2813         struct btrfs_fs_info *fs_info = root->fs_info;
2814         u64 num_entries =
2815                 atomic_read(&trans->transaction->delayed_refs.num_entries);
2816         u64 avg_runtime;
2817         u64 val;
2818
2819         smp_mb();
2820         avg_runtime = fs_info->avg_delayed_ref_runtime;
2821         val = num_entries * avg_runtime;
2822         if (num_entries * avg_runtime >= NSEC_PER_SEC)
2823                 return 1;
2824         if (val >= NSEC_PER_SEC / 2)
2825                 return 2;
2826
2827         return btrfs_check_space_for_delayed_refs(trans, root);
2828 }
2829
2830 struct async_delayed_refs {
2831         struct btrfs_root *root;
2832         int count;
2833         int error;
2834         int sync;
2835         struct completion wait;
2836         struct btrfs_work work;
2837 };
2838
2839 static void delayed_ref_async_start(struct btrfs_work *work)
2840 {
2841         struct async_delayed_refs *async;
2842         struct btrfs_trans_handle *trans;
2843         int ret;
2844
2845         async = container_of(work, struct async_delayed_refs, work);
2846
2847         trans = btrfs_join_transaction(async->root);
2848         if (IS_ERR(trans)) {
2849                 async->error = PTR_ERR(trans);
2850                 goto done;
2851         }
2852
2853         /*
2854          * trans->sync means that when we call end_transaciton, we won't
2855          * wait on delayed refs
2856          */
2857         trans->sync = true;
2858         ret = btrfs_run_delayed_refs(trans, async->root, async->count);
2859         if (ret)
2860                 async->error = ret;
2861
2862         ret = btrfs_end_transaction(trans, async->root);
2863         if (ret && !async->error)
2864                 async->error = ret;
2865 done:
2866         if (async->sync)
2867                 complete(&async->wait);
2868         else
2869                 kfree(async);
2870 }
2871
2872 int btrfs_async_run_delayed_refs(struct btrfs_root *root,
2873                                  unsigned long count, int wait)
2874 {
2875         struct async_delayed_refs *async;
2876         int ret;
2877
2878         async = kmalloc(sizeof(*async), GFP_NOFS);
2879         if (!async)
2880                 return -ENOMEM;
2881
2882         async->root = root->fs_info->tree_root;
2883         async->count = count;
2884         async->error = 0;
2885         if (wait)
2886                 async->sync = 1;
2887         else
2888                 async->sync = 0;
2889         init_completion(&async->wait);
2890
2891         btrfs_init_work(&async->work, btrfs_extent_refs_helper,
2892                         delayed_ref_async_start, NULL, NULL);
2893
2894         btrfs_queue_work(root->fs_info->extent_workers, &async->work);
2895
2896         if (wait) {
2897                 wait_for_completion(&async->wait);
2898                 ret = async->error;
2899                 kfree(async);
2900                 return ret;
2901         }
2902         return 0;
2903 }
2904
2905 /*
2906  * this starts processing the delayed reference count updates and
2907  * extent insertions we have queued up so far.  count can be
2908  * 0, which means to process everything in the tree at the start
2909  * of the run (but not newly added entries), or it can be some target
2910  * number you'd like to process.
2911  *
2912  * Returns 0 on success or if called with an aborted transaction
2913  * Returns <0 on error and aborts the transaction
2914  */
2915 int btrfs_run_delayed_refs(struct btrfs_trans_handle *trans,
2916                            struct btrfs_root *root, unsigned long count)
2917 {
2918         struct rb_node *node;
2919         struct btrfs_delayed_ref_root *delayed_refs;
2920         struct btrfs_delayed_ref_head *head;
2921         int ret;
2922         int run_all = count == (unsigned long)-1;
2923         bool can_flush_pending_bgs = trans->can_flush_pending_bgs;
2924
2925         /* We'll clean this up in btrfs_cleanup_transaction */
2926         if (trans->aborted)
2927                 return 0;
2928
2929         if (root->fs_info->creating_free_space_tree)
2930                 return 0;
2931
2932         if (root == root->fs_info->extent_root)
2933                 root = root->fs_info->tree_root;
2934
2935         delayed_refs = &trans->transaction->delayed_refs;
2936         if (count == 0)
2937                 count = atomic_read(&delayed_refs->num_entries) * 2;
2938
2939 again:
2940 #ifdef SCRAMBLE_DELAYED_REFS
2941         delayed_refs->run_delayed_start = find_middle(&delayed_refs->root);
2942 #endif
2943         trans->can_flush_pending_bgs = false;
2944         ret = __btrfs_run_delayed_refs(trans, root, count);
2945         if (ret < 0) {
2946                 btrfs_abort_transaction(trans, root, ret);
2947                 return ret;
2948         }
2949
2950         if (run_all) {
2951                 if (!list_empty(&trans->new_bgs))
2952                         btrfs_create_pending_block_groups(trans, root);
2953
2954                 spin_lock(&delayed_refs->lock);
2955                 node = rb_first(&delayed_refs->href_root);
2956                 if (!node) {
2957                         spin_unlock(&delayed_refs->lock);
2958                         goto out;
2959                 }
2960                 count = (unsigned long)-1;
2961
2962                 while (node) {
2963                         head = rb_entry(node, struct btrfs_delayed_ref_head,
2964                                         href_node);
2965                         if (btrfs_delayed_ref_is_head(&head->node)) {
2966                                 struct btrfs_delayed_ref_node *ref;
2967
2968                                 ref = &head->node;
2969                                 atomic_inc(&ref->refs);
2970
2971                                 spin_unlock(&delayed_refs->lock);
2972                                 /*
2973                                  * Mutex was contended, block until it's
2974                                  * released and try again
2975                                  */
2976                                 mutex_lock(&head->mutex);
2977                                 mutex_unlock(&head->mutex);
2978
2979                                 btrfs_put_delayed_ref(ref);
2980                                 cond_resched();
2981                                 goto again;
2982                         } else {
2983                                 WARN_ON(1);
2984                         }
2985                         node = rb_next(node);
2986                 }
2987                 spin_unlock(&delayed_refs->lock);
2988                 cond_resched();
2989                 goto again;
2990         }
2991 out:
2992         assert_qgroups_uptodate(trans);
2993         trans->can_flush_pending_bgs = can_flush_pending_bgs;
2994         return 0;
2995 }
2996
2997 int btrfs_set_disk_extent_flags(struct btrfs_trans_handle *trans,
2998                                 struct btrfs_root *root,
2999                                 u64 bytenr, u64 num_bytes, u64 flags,
3000                                 int level, int is_data)
3001 {
3002         struct btrfs_delayed_extent_op *extent_op;
3003         int ret;
3004
3005         extent_op = btrfs_alloc_delayed_extent_op();
3006         if (!extent_op)
3007                 return -ENOMEM;
3008
3009         extent_op->flags_to_set = flags;
3010         extent_op->update_flags = 1;
3011         extent_op->update_key = 0;
3012         extent_op->is_data = is_data ? 1 : 0;
3013         extent_op->level = level;
3014
3015         ret = btrfs_add_delayed_extent_op(root->fs_info, trans, bytenr,
3016                                           num_bytes, extent_op);
3017         if (ret)
3018                 btrfs_free_delayed_extent_op(extent_op);
3019         return ret;
3020 }
3021
3022 static noinline int check_delayed_ref(struct btrfs_trans_handle *trans,
3023                                       struct btrfs_root *root,
3024                                       struct btrfs_path *path,
3025                                       u64 objectid, u64 offset, u64 bytenr)
3026 {
3027         struct btrfs_delayed_ref_head *head;
3028         struct btrfs_delayed_ref_node *ref;
3029         struct btrfs_delayed_data_ref *data_ref;
3030         struct btrfs_delayed_ref_root *delayed_refs;
3031         int ret = 0;
3032
3033         delayed_refs = &trans->transaction->delayed_refs;
3034         spin_lock(&delayed_refs->lock);
3035         head = btrfs_find_delayed_ref_head(trans, bytenr);
3036         if (!head) {
3037                 spin_unlock(&delayed_refs->lock);
3038                 return 0;
3039         }
3040
3041         if (!mutex_trylock(&head->mutex)) {
3042                 atomic_inc(&head->node.refs);
3043                 spin_unlock(&delayed_refs->lock);
3044
3045                 btrfs_release_path(path);
3046
3047                 /*
3048                  * Mutex was contended, block until it's released and let
3049                  * caller try again
3050                  */
3051                 mutex_lock(&head->mutex);
3052                 mutex_unlock(&head->mutex);
3053                 btrfs_put_delayed_ref(&head->node);
3054                 return -EAGAIN;
3055         }
3056         spin_unlock(&delayed_refs->lock);
3057
3058         spin_lock(&head->lock);
3059         list_for_each_entry(ref, &head->ref_list, list) {
3060                 /* If it's a shared ref we know a cross reference exists */
3061                 if (ref->type != BTRFS_EXTENT_DATA_REF_KEY) {
3062                         ret = 1;
3063                         break;
3064                 }
3065
3066                 data_ref = btrfs_delayed_node_to_data_ref(ref);
3067
3068                 /*
3069                  * If our ref doesn't match the one we're currently looking at
3070                  * then we have a cross reference.
3071                  */
3072                 if (data_ref->root != root->root_key.objectid ||
3073                     data_ref->objectid != objectid ||
3074                     data_ref->offset != offset) {
3075                         ret = 1;
3076                         break;
3077                 }
3078         }
3079         spin_unlock(&head->lock);
3080         mutex_unlock(&head->mutex);
3081         return ret;
3082 }
3083
3084 static noinline int check_committed_ref(struct btrfs_trans_handle *trans,
3085                                         struct btrfs_root *root,
3086                                         struct btrfs_path *path,
3087                                         u64 objectid, u64 offset, u64 bytenr)
3088 {
3089         struct btrfs_root *extent_root = root->fs_info->extent_root;
3090         struct extent_buffer *leaf;
3091         struct btrfs_extent_data_ref *ref;
3092         struct btrfs_extent_inline_ref *iref;
3093         struct btrfs_extent_item *ei;
3094         struct btrfs_key key;
3095         u32 item_size;
3096         int ret;
3097
3098         key.objectid = bytenr;
3099         key.offset = (u64)-1;
3100         key.type = BTRFS_EXTENT_ITEM_KEY;
3101
3102         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
3103         if (ret < 0)
3104                 goto out;
3105         BUG_ON(ret == 0); /* Corruption */
3106
3107         ret = -ENOENT;
3108         if (path->slots[0] == 0)
3109                 goto out;
3110
3111         path->slots[0]--;
3112         leaf = path->nodes[0];
3113         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
3114
3115         if (key.objectid != bytenr || key.type != BTRFS_EXTENT_ITEM_KEY)
3116                 goto out;
3117
3118         ret = 1;
3119         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
3120 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
3121         if (item_size < sizeof(*ei)) {
3122                 WARN_ON(item_size != sizeof(struct btrfs_extent_item_v0));
3123                 goto out;
3124         }
3125 #endif
3126         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
3127
3128         if (item_size != sizeof(*ei) +
3129             btrfs_extent_inline_ref_size(BTRFS_EXTENT_DATA_REF_KEY))
3130                 goto out;
3131
3132         if (btrfs_extent_generation(leaf, ei) <=
3133             btrfs_root_last_snapshot(&root->root_item))
3134                 goto out;
3135
3136         iref = (struct btrfs_extent_inline_ref *)(ei + 1);
3137         if (btrfs_extent_inline_ref_type(leaf, iref) !=
3138             BTRFS_EXTENT_DATA_REF_KEY)
3139                 goto out;
3140
3141         ref = (struct btrfs_extent_data_ref *)(&iref->offset);
3142         if (btrfs_extent_refs(leaf, ei) !=
3143             btrfs_extent_data_ref_count(leaf, ref) ||
3144             btrfs_extent_data_ref_root(leaf, ref) !=
3145             root->root_key.objectid ||
3146             btrfs_extent_data_ref_objectid(leaf, ref) != objectid ||
3147             btrfs_extent_data_ref_offset(leaf, ref) != offset)
3148                 goto out;
3149
3150         ret = 0;
3151 out:
3152         return ret;
3153 }
3154
3155 int btrfs_cross_ref_exist(struct btrfs_trans_handle *trans,
3156                           struct btrfs_root *root,
3157                           u64 objectid, u64 offset, u64 bytenr)
3158 {
3159         struct btrfs_path *path;
3160         int ret;
3161         int ret2;
3162
3163         path = btrfs_alloc_path();
3164         if (!path)
3165                 return -ENOENT;
3166
3167         do {
3168                 ret = check_committed_ref(trans, root, path, objectid,
3169                                           offset, bytenr);
3170                 if (ret && ret != -ENOENT)
3171                         goto out;
3172
3173                 ret2 = check_delayed_ref(trans, root, path, objectid,
3174                                          offset, bytenr);
3175         } while (ret2 == -EAGAIN);
3176
3177         if (ret2 && ret2 != -ENOENT) {
3178                 ret = ret2;
3179                 goto out;
3180         }
3181
3182         if (ret != -ENOENT || ret2 != -ENOENT)
3183                 ret = 0;
3184 out:
3185         btrfs_free_path(path);
3186         if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID)
3187                 WARN_ON(ret > 0);
3188         return ret;
3189 }
3190
3191 static int __btrfs_mod_ref(struct btrfs_trans_handle *trans,
3192                            struct btrfs_root *root,
3193                            struct extent_buffer *buf,
3194                            int full_backref, int inc)
3195 {
3196         u64 bytenr;
3197         u64 num_bytes;
3198         u64 parent;
3199         u64 ref_root;
3200         u32 nritems;
3201         struct btrfs_key key;
3202         struct btrfs_file_extent_item *fi;
3203         int i;
3204         int level;
3205         int ret = 0;
3206         int (*process_func)(struct btrfs_trans_handle *, struct btrfs_root *,
3207                             u64, u64, u64, u64, u64, u64);
3208
3209
3210         if (btrfs_test_is_dummy_root(root))
3211                 return 0;
3212
3213         ref_root = btrfs_header_owner(buf);
3214         nritems = btrfs_header_nritems(buf);
3215         level = btrfs_header_level(buf);
3216
3217         if (!test_bit(BTRFS_ROOT_REF_COWS, &root->state) && level == 0)
3218                 return 0;
3219
3220         if (inc)
3221                 process_func = btrfs_inc_extent_ref;
3222         else
3223                 process_func = btrfs_free_extent;
3224
3225         if (full_backref)
3226                 parent = buf->start;
3227         else
3228                 parent = 0;
3229
3230         for (i = 0; i < nritems; i++) {
3231                 if (level == 0) {
3232                         btrfs_item_key_to_cpu(buf, &key, i);
3233                         if (key.type != BTRFS_EXTENT_DATA_KEY)
3234                                 continue;
3235                         fi = btrfs_item_ptr(buf, i,
3236                                             struct btrfs_file_extent_item);
3237                         if (btrfs_file_extent_type(buf, fi) ==
3238                             BTRFS_FILE_EXTENT_INLINE)
3239                                 continue;
3240                         bytenr = btrfs_file_extent_disk_bytenr(buf, fi);
3241                         if (bytenr == 0)
3242                                 continue;
3243
3244                         num_bytes = btrfs_file_extent_disk_num_bytes(buf, fi);
3245                         key.offset -= btrfs_file_extent_offset(buf, fi);
3246                         ret = process_func(trans, root, bytenr, num_bytes,
3247                                            parent, ref_root, key.objectid,
3248                                            key.offset);
3249                         if (ret)
3250                                 goto fail;
3251                 } else {
3252                         bytenr = btrfs_node_blockptr(buf, i);
3253                         num_bytes = root->nodesize;
3254                         ret = process_func(trans, root, bytenr, num_bytes,
3255                                            parent, ref_root, level - 1, 0);
3256                         if (ret)
3257                                 goto fail;
3258                 }
3259         }
3260         return 0;
3261 fail:
3262         return ret;
3263 }
3264
3265 int btrfs_inc_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
3266                   struct extent_buffer *buf, int full_backref)
3267 {
3268         return __btrfs_mod_ref(trans, root, buf, full_backref, 1);
3269 }
3270
3271 int btrfs_dec_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
3272                   struct extent_buffer *buf, int full_backref)
3273 {
3274         return __btrfs_mod_ref(trans, root, buf, full_backref, 0);
3275 }
3276
3277 static int write_one_cache_group(struct btrfs_trans_handle *trans,
3278                                  struct btrfs_root *root,
3279                                  struct btrfs_path *path,
3280                                  struct btrfs_block_group_cache *cache)
3281 {
3282         int ret;
3283         struct btrfs_root *extent_root = root->fs_info->extent_root;
3284         unsigned long bi;
3285         struct extent_buffer *leaf;
3286
3287         ret = btrfs_search_slot(trans, extent_root, &cache->key, path, 0, 1);
3288         if (ret) {
3289                 if (ret > 0)
3290                         ret = -ENOENT;
3291                 goto fail;
3292         }
3293
3294         leaf = path->nodes[0];
3295         bi = btrfs_item_ptr_offset(leaf, path->slots[0]);
3296         write_extent_buffer(leaf, &cache->item, bi, sizeof(cache->item));
3297         btrfs_mark_buffer_dirty(leaf);
3298 fail:
3299         btrfs_release_path(path);
3300         return ret;
3301
3302 }
3303
3304 static struct btrfs_block_group_cache *
3305 next_block_group(struct btrfs_root *root,
3306                  struct btrfs_block_group_cache *cache)
3307 {
3308         struct rb_node *node;
3309
3310         spin_lock(&root->fs_info->block_group_cache_lock);
3311
3312         /* If our block group was removed, we need a full search. */
3313         if (RB_EMPTY_NODE(&cache->cache_node)) {
3314                 const u64 next_bytenr = cache->key.objectid + cache->key.offset;
3315
3316                 spin_unlock(&root->fs_info->block_group_cache_lock);
3317                 btrfs_put_block_group(cache);
3318                 cache = btrfs_lookup_first_block_group(root->fs_info,
3319                                                        next_bytenr);
3320                 return cache;
3321         }
3322         node = rb_next(&cache->cache_node);
3323         btrfs_put_block_group(cache);
3324         if (node) {
3325                 cache = rb_entry(node, struct btrfs_block_group_cache,
3326                                  cache_node);
3327                 btrfs_get_block_group(cache);
3328         } else
3329                 cache = NULL;
3330         spin_unlock(&root->fs_info->block_group_cache_lock);
3331         return cache;
3332 }
3333
3334 static int cache_save_setup(struct btrfs_block_group_cache *block_group,
3335                             struct btrfs_trans_handle *trans,
3336                             struct btrfs_path *path)
3337 {
3338         struct btrfs_root *root = block_group->fs_info->tree_root;
3339         struct inode *inode = NULL;
3340         u64 alloc_hint = 0;
3341         int dcs = BTRFS_DC_ERROR;
3342         u64 num_pages = 0;
3343         int retries = 0;
3344         int ret = 0;
3345
3346         /*
3347          * If this block group is smaller than 100 megs don't bother caching the
3348          * block group.
3349          */
3350         if (block_group->key.offset < (100 * 1024 * 1024)) {
3351                 spin_lock(&block_group->lock);
3352                 block_group->disk_cache_state = BTRFS_DC_WRITTEN;
3353                 spin_unlock(&block_group->lock);
3354                 return 0;
3355         }
3356
3357         if (trans->aborted)
3358                 return 0;
3359 again:
3360         inode = lookup_free_space_inode(root, block_group, path);
3361         if (IS_ERR(inode) && PTR_ERR(inode) != -ENOENT) {
3362                 ret = PTR_ERR(inode);
3363                 btrfs_release_path(path);
3364                 goto out;
3365         }
3366
3367         if (IS_ERR(inode)) {
3368                 BUG_ON(retries);
3369                 retries++;
3370
3371                 if (block_group->ro)
3372                         goto out_free;
3373
3374                 ret = create_free_space_inode(root, trans, block_group, path);
3375                 if (ret)
3376                         goto out_free;
3377                 goto again;
3378         }
3379
3380         /* We've already setup this transaction, go ahead and exit */
3381         if (block_group->cache_generation == trans->transid &&
3382             i_size_read(inode)) {
3383                 dcs = BTRFS_DC_SETUP;
3384                 goto out_put;
3385         }
3386
3387         /*
3388          * We want to set the generation to 0, that way if anything goes wrong
3389          * from here on out we know not to trust this cache when we load up next
3390          * time.
3391          */
3392         BTRFS_I(inode)->generation = 0;
3393         ret = btrfs_update_inode(trans, root, inode);
3394         if (ret) {
3395                 /*
3396                  * So theoretically we could recover from this, simply set the
3397                  * super cache generation to 0 so we know to invalidate the
3398                  * cache, but then we'd have to keep track of the block groups
3399                  * that fail this way so we know we _have_ to reset this cache
3400                  * before the next commit or risk reading stale cache.  So to
3401                  * limit our exposure to horrible edge cases lets just abort the
3402                  * transaction, this only happens in really bad situations
3403                  * anyway.
3404                  */
3405                 btrfs_abort_transaction(trans, root, ret);
3406                 goto out_put;
3407         }
3408         WARN_ON(ret);
3409
3410         if (i_size_read(inode) > 0) {
3411                 ret = btrfs_check_trunc_cache_free_space(root,
3412                                         &root->fs_info->global_block_rsv);
3413                 if (ret)
3414                         goto out_put;
3415
3416                 ret = btrfs_truncate_free_space_cache(root, trans, NULL, inode);
3417                 if (ret)
3418                         goto out_put;
3419         }
3420
3421         spin_lock(&block_group->lock);
3422         if (block_group->cached != BTRFS_CACHE_FINISHED ||
3423             !btrfs_test_opt(root, SPACE_CACHE)) {
3424                 /*
3425                  * don't bother trying to write stuff out _if_
3426                  * a) we're not cached,
3427                  * b) we're with nospace_cache mount option.
3428                  */
3429                 dcs = BTRFS_DC_WRITTEN;
3430                 spin_unlock(&block_group->lock);
3431                 goto out_put;
3432         }
3433         spin_unlock(&block_group->lock);
3434
3435         /*
3436          * We hit an ENOSPC when setting up the cache in this transaction, just
3437          * skip doing the setup, we've already cleared the cache so we're safe.
3438          */
3439         if (test_bit(BTRFS_TRANS_CACHE_ENOSPC, &trans->transaction->flags)) {
3440                 ret = -ENOSPC;
3441                 goto out_put;
3442         }
3443
3444         /*
3445          * Try to preallocate enough space based on how big the block group is.
3446          * Keep in mind this has to include any pinned space which could end up
3447          * taking up quite a bit since it's not folded into the other space
3448          * cache.
3449          */
3450         num_pages = div_u64(block_group->key.offset, 256 * 1024 * 1024);
3451         if (!num_pages)
3452                 num_pages = 1;
3453
3454         num_pages *= 16;
3455         num_pages *= PAGE_CACHE_SIZE;
3456
3457         ret = btrfs_check_data_free_space(inode, 0, num_pages);
3458         if (ret)
3459                 goto out_put;
3460
3461         ret = btrfs_prealloc_file_range_trans(inode, trans, 0, 0, num_pages,
3462                                               num_pages, num_pages,
3463                                               &alloc_hint);
3464         /*
3465          * Our cache requires contiguous chunks so that we don't modify a bunch
3466          * of metadata or split extents when writing the cache out, which means
3467          * we can enospc if we are heavily fragmented in addition to just normal
3468          * out of space conditions.  So if we hit this just skip setting up any
3469          * other block groups for this transaction, maybe we'll unpin enough
3470          * space the next time around.
3471          */
3472         if (!ret)
3473                 dcs = BTRFS_DC_SETUP;
3474         else if (ret == -ENOSPC)
3475                 set_bit(BTRFS_TRANS_CACHE_ENOSPC, &trans->transaction->flags);
3476         btrfs_free_reserved_data_space(inode, 0, num_pages);
3477
3478 out_put:
3479         iput(inode);
3480 out_free:
3481         btrfs_release_path(path);
3482 out:
3483         spin_lock(&block_group->lock);
3484         if (!ret && dcs == BTRFS_DC_SETUP)
3485                 block_group->cache_generation = trans->transid;
3486         block_group->disk_cache_state = dcs;
3487         spin_unlock(&block_group->lock);
3488
3489         return ret;
3490 }
3491
3492 int btrfs_setup_space_cache(struct btrfs_trans_handle *trans,
3493                             struct btrfs_root *root)
3494 {
3495         struct btrfs_block_group_cache *cache, *tmp;
3496         struct btrfs_transaction *cur_trans = trans->transaction;
3497         struct btrfs_path *path;
3498
3499         if (list_empty(&cur_trans->dirty_bgs) ||
3500             !btrfs_test_opt(root, SPACE_CACHE))
3501                 return 0;
3502
3503         path = btrfs_alloc_path();
3504         if (!path)
3505                 return -ENOMEM;
3506
3507         /* Could add new block groups, use _safe just in case */
3508         list_for_each_entry_safe(cache, tmp, &cur_trans->dirty_bgs,
3509                                  dirty_list) {
3510                 if (cache->disk_cache_state == BTRFS_DC_CLEAR)
3511                         cache_save_setup(cache, trans, path);
3512         }
3513
3514         btrfs_free_path(path);
3515         return 0;
3516 }
3517
3518 /*
3519  * transaction commit does final block group cache writeback during a
3520  * critical section where nothing is allowed to change the FS.  This is
3521  * required in order for the cache to actually match the block group,
3522  * but can introduce a lot of latency into the commit.
3523  *
3524  * So, btrfs_start_dirty_block_groups is here to kick off block group
3525  * cache IO.  There's a chance we'll have to redo some of it if the
3526  * block group changes again during the commit, but it greatly reduces
3527  * the commit latency by getting rid of the easy block groups while
3528  * we're still allowing others to join the commit.
3529  */
3530 int btrfs_start_dirty_block_groups(struct btrfs_trans_handle *trans,
3531                                    struct btrfs_root *root)
3532 {
3533         struct btrfs_block_group_cache *cache;
3534         struct btrfs_transaction *cur_trans = trans->transaction;
3535         int ret = 0;
3536         int should_put;
3537         struct btrfs_path *path = NULL;
3538         LIST_HEAD(dirty);
3539         struct list_head *io = &cur_trans->io_bgs;
3540         int num_started = 0;
3541         int loops = 0;
3542
3543         spin_lock(&cur_trans->dirty_bgs_lock);
3544         if (list_empty(&cur_trans->dirty_bgs)) {
3545                 spin_unlock(&cur_trans->dirty_bgs_lock);
3546                 return 0;
3547         }
3548         list_splice_init(&cur_trans->dirty_bgs, &dirty);
3549         spin_unlock(&cur_trans->dirty_bgs_lock);
3550
3551 again:
3552         /*
3553          * make sure all the block groups on our dirty list actually
3554          * exist
3555          */
3556         btrfs_create_pending_block_groups(trans, root);
3557
3558         if (!path) {
3559                 path = btrfs_alloc_path();
3560                 if (!path)
3561                         return -ENOMEM;
3562         }
3563
3564         /*
3565          * cache_write_mutex is here only to save us from balance or automatic
3566          * removal of empty block groups deleting this block group while we are
3567          * writing out the cache
3568          */
3569         mutex_lock(&trans->transaction->cache_write_mutex);
3570         while (!list_empty(&dirty)) {
3571                 cache = list_first_entry(&dirty,
3572                                          struct btrfs_block_group_cache,
3573                                          dirty_list);
3574                 /*
3575                  * this can happen if something re-dirties a block
3576                  * group that is already under IO.  Just wait for it to
3577                  * finish and then do it all again
3578                  */
3579                 if (!list_empty(&cache->io_list)) {
3580                         list_del_init(&cache->io_list);
3581                         btrfs_wait_cache_io(root, trans, cache,
3582                                             &cache->io_ctl, path,
3583                                             cache->key.objectid);
3584                         btrfs_put_block_group(cache);
3585                 }
3586
3587
3588                 /*
3589                  * btrfs_wait_cache_io uses the cache->dirty_list to decide
3590                  * if it should update the cache_state.  Don't delete
3591                  * until after we wait.
3592                  *
3593                  * Since we're not running in the commit critical section
3594                  * we need the dirty_bgs_lock to protect from update_block_group
3595                  */
3596                 spin_lock(&cur_trans->dirty_bgs_lock);
3597                 list_del_init(&cache->dirty_list);
3598                 spin_unlock(&cur_trans->dirty_bgs_lock);
3599
3600                 should_put = 1;
3601
3602                 cache_save_setup(cache, trans, path);
3603
3604                 if (cache->disk_cache_state == BTRFS_DC_SETUP) {
3605                         cache->io_ctl.inode = NULL;
3606                         ret = btrfs_write_out_cache(root, trans, cache, path);
3607                         if (ret == 0 && cache->io_ctl.inode) {
3608                                 num_started++;
3609                                 should_put = 0;
3610
3611                                 /*
3612                                  * the cache_write_mutex is protecting
3613                                  * the io_list
3614                                  */
3615                                 list_add_tail(&cache->io_list, io);
3616                         } else {
3617                                 /*
3618                                  * if we failed to write the cache, the
3619                                  * generation will be bad and life goes on
3620                                  */
3621                                 ret = 0;
3622                         }
3623                 }
3624                 if (!ret) {
3625                         ret = write_one_cache_group(trans, root, path, cache);
3626                         /*
3627                          * Our block group might still be attached to the list
3628                          * of new block groups in the transaction handle of some
3629                          * other task (struct btrfs_trans_handle->new_bgs). This
3630                          * means its block group item isn't yet in the extent
3631                          * tree. If this happens ignore the error, as we will
3632                          * try again later in the critical section of the
3633                          * transaction commit.
3634                          */
3635                         if (ret == -ENOENT) {
3636                                 ret = 0;
3637                                 spin_lock(&cur_trans->dirty_bgs_lock);
3638                                 if (list_empty(&cache->dirty_list)) {
3639                                         list_add_tail(&cache->dirty_list,
3640                                                       &cur_trans->dirty_bgs);
3641                                         btrfs_get_block_group(cache);
3642                                 }
3643                                 spin_unlock(&cur_trans->dirty_bgs_lock);
3644                         } else if (ret) {
3645                                 btrfs_abort_transaction(trans, root, ret);
3646                         }
3647                 }
3648
3649                 /* if its not on the io list, we need to put the block group */
3650                 if (should_put)
3651                         btrfs_put_block_group(cache);
3652
3653                 if (ret)
3654                         break;
3655
3656                 /*
3657                  * Avoid blocking other tasks for too long. It might even save
3658                  * us from writing caches for block groups that are going to be
3659                  * removed.
3660                  */
3661                 mutex_unlock(&trans->transaction->cache_write_mutex);
3662                 mutex_lock(&trans->transaction->cache_write_mutex);
3663         }
3664         mutex_unlock(&trans->transaction->cache_write_mutex);
3665
3666         /*
3667          * go through delayed refs for all the stuff we've just kicked off
3668          * and then loop back (just once)
3669          */
3670         ret = btrfs_run_delayed_refs(trans, root, 0);
3671         if (!ret && loops == 0) {
3672                 loops++;
3673                 spin_lock(&cur_trans->dirty_bgs_lock);
3674                 list_splice_init(&cur_trans->dirty_bgs, &dirty);
3675                 /*
3676                  * dirty_bgs_lock protects us from concurrent block group
3677                  * deletes too (not just cache_write_mutex).
3678                  */
3679                 if (!list_empty(&dirty)) {
3680                         spin_unlock(&cur_trans->dirty_bgs_lock);
3681                         goto again;
3682                 }
3683                 spin_unlock(&cur_trans->dirty_bgs_lock);
3684         }
3685
3686         btrfs_free_path(path);
3687         return ret;
3688 }
3689
3690 int btrfs_write_dirty_block_groups(struct btrfs_trans_handle *trans,
3691                                    struct btrfs_root *root)
3692 {
3693         struct btrfs_block_group_cache *cache;
3694         struct btrfs_transaction *cur_trans = trans->transaction;
3695         int ret = 0;
3696         int should_put;
3697         struct btrfs_path *path;
3698         struct list_head *io = &cur_trans->io_bgs;
3699         int num_started = 0;
3700
3701         path = btrfs_alloc_path();
3702         if (!path)
3703                 return -ENOMEM;
3704
3705         /*
3706          * Even though we are in the critical section of the transaction commit,
3707          * we can still have concurrent tasks adding elements to this
3708          * transaction's list of dirty block groups. These tasks correspond to
3709          * endio free space workers started when writeback finishes for a
3710          * space cache, which run inode.c:btrfs_finish_ordered_io(), and can
3711          * allocate new block groups as a result of COWing nodes of the root
3712          * tree when updating the free space inode. The writeback for the space
3713          * caches is triggered by an earlier call to
3714          * btrfs_start_dirty_block_groups() and iterations of the following
3715          * loop.
3716          * Also we want to do the cache_save_setup first and then run the
3717          * delayed refs to make sure we have the best chance at doing this all
3718          * in one shot.
3719          */
3720         spin_lock(&cur_trans->dirty_bgs_lock);
3721         while (!list_empty(&cur_trans->dirty_bgs)) {
3722                 cache = list_first_entry(&cur_trans->dirty_bgs,
3723                                          struct btrfs_block_group_cache,
3724                                          dirty_list);
3725
3726                 /*
3727                  * this can happen if cache_save_setup re-dirties a block
3728                  * group that is already under IO.  Just wait for it to
3729                  * finish and then do it all again
3730                  */
3731                 if (!list_empty(&cache->io_list)) {
3732                         spin_unlock(&cur_trans->dirty_bgs_lock);
3733                         list_del_init(&cache->io_list);
3734                         btrfs_wait_cache_io(root, trans, cache,
3735                                             &cache->io_ctl, path,
3736                                             cache->key.objectid);
3737                         btrfs_put_block_group(cache);
3738                         spin_lock(&cur_trans->dirty_bgs_lock);
3739                 }
3740
3741                 /*
3742                  * don't remove from the dirty list until after we've waited
3743                  * on any pending IO
3744                  */
3745                 list_del_init(&cache->dirty_list);
3746                 spin_unlock(&cur_trans->dirty_bgs_lock);
3747                 should_put = 1;
3748
3749                 cache_save_setup(cache, trans, path);
3750
3751                 if (!ret)
3752                         ret = btrfs_run_delayed_refs(trans, root, (unsigned long) -1);
3753
3754                 if (!ret && cache->disk_cache_state == BTRFS_DC_SETUP) {
3755                         cache->io_ctl.inode = NULL;
3756                         ret = btrfs_write_out_cache(root, trans, cache, path);
3757                         if (ret == 0 && cache->io_ctl.inode) {
3758                                 num_started++;
3759                                 should_put = 0;
3760                                 list_add_tail(&cache->io_list, io);
3761                         } else {
3762                                 /*
3763                                  * if we failed to write the cache, the
3764                                  * generation will be bad and life goes on
3765                                  */
3766                                 ret = 0;
3767                         }
3768                 }
3769                 if (!ret) {
3770                         ret = write_one_cache_group(trans, root, path, cache);
3771                         if (ret)
3772                                 btrfs_abort_transaction(trans, root, ret);
3773                 }
3774
3775                 /* if its not on the io list, we need to put the block group */
3776                 if (should_put)
3777                         btrfs_put_block_group(cache);
3778                 spin_lock(&cur_trans->dirty_bgs_lock);
3779         }
3780         spin_unlock(&cur_trans->dirty_bgs_lock);
3781
3782         while (!list_empty(io)) {
3783                 cache = list_first_entry(io, struct btrfs_block_group_cache,
3784                                          io_list);
3785                 list_del_init(&cache->io_list);
3786                 btrfs_wait_cache_io(root, trans, cache,
3787                                     &cache->io_ctl, path, cache->key.objectid);
3788                 btrfs_put_block_group(cache);
3789         }
3790
3791         btrfs_free_path(path);
3792         return ret;
3793 }
3794
3795 int btrfs_extent_readonly(struct btrfs_root *root, u64 bytenr)
3796 {
3797         struct btrfs_block_group_cache *block_group;
3798         int readonly = 0;
3799
3800         block_group = btrfs_lookup_block_group(root->fs_info, bytenr);
3801         if (!block_group || block_group->ro)
3802                 readonly = 1;
3803         if (block_group)
3804                 btrfs_put_block_group(block_group);
3805         return readonly;
3806 }
3807
3808 static const char *alloc_name(u64 flags)
3809 {
3810         switch (flags) {
3811         case BTRFS_BLOCK_GROUP_METADATA|BTRFS_BLOCK_GROUP_DATA:
3812                 return "mixed";
3813         case BTRFS_BLOCK_GROUP_METADATA:
3814                 return "metadata";
3815         case BTRFS_BLOCK_GROUP_DATA:
3816                 return "data";
3817         case BTRFS_BLOCK_GROUP_SYSTEM:
3818                 return "system";
3819         default:
3820                 WARN_ON(1);
3821                 return "invalid-combination";
3822         };
3823 }
3824
3825 static int update_space_info(struct btrfs_fs_info *info, u64 flags,
3826                              u64 total_bytes, u64 bytes_used,
3827                              struct btrfs_space_info **space_info)
3828 {
3829         struct btrfs_space_info *found;
3830         int i;
3831         int factor;
3832         int ret;
3833
3834         if (flags & (BTRFS_BLOCK_GROUP_DUP | BTRFS_BLOCK_GROUP_RAID1 |
3835                      BTRFS_BLOCK_GROUP_RAID10))
3836                 factor = 2;
3837         else
3838                 factor = 1;
3839
3840         found = __find_space_info(info, flags);
3841         if (found) {
3842                 spin_lock(&found->lock);
3843                 found->total_bytes += total_bytes;
3844                 found->disk_total += total_bytes * factor;
3845                 found->bytes_used += bytes_used;
3846                 found->disk_used += bytes_used * factor;
3847                 if (total_bytes > 0)
3848                         found->full = 0;
3849                 spin_unlock(&found->lock);
3850                 *space_info = found;
3851                 return 0;
3852         }
3853         found = kzalloc(sizeof(*found), GFP_NOFS);
3854         if (!found)
3855                 return -ENOMEM;
3856
3857         ret = percpu_counter_init(&found->total_bytes_pinned, 0, GFP_KERNEL);
3858         if (ret) {
3859                 kfree(found);
3860                 return ret;
3861         }
3862
3863         for (i = 0; i < BTRFS_NR_RAID_TYPES; i++)
3864                 INIT_LIST_HEAD(&found->block_groups[i]);
3865         init_rwsem(&found->groups_sem);
3866         spin_lock_init(&found->lock);
3867         found->flags = flags & BTRFS_BLOCK_GROUP_TYPE_MASK;
3868         found->total_bytes = total_bytes;
3869         found->disk_total = total_bytes * factor;
3870         found->bytes_used = bytes_used;
3871         found->disk_used = bytes_used * factor;
3872         found->bytes_pinned = 0;
3873         found->bytes_reserved = 0;
3874         found->bytes_readonly = 0;
3875         found->bytes_may_use = 0;
3876         found->full = 0;
3877         found->max_extent_size = 0;
3878         found->force_alloc = CHUNK_ALLOC_NO_FORCE;
3879         found->chunk_alloc = 0;
3880         found->flush = 0;
3881         init_waitqueue_head(&found->wait);
3882         INIT_LIST_HEAD(&found->ro_bgs);
3883
3884         ret = kobject_init_and_add(&found->kobj, &space_info_ktype,
3885                                     info->space_info_kobj, "%s",
3886                                     alloc_name(found->flags));
3887         if (ret) {
3888                 kfree(found);
3889                 return ret;
3890         }
3891
3892         *space_info = found;
3893         list_add_rcu(&found->list, &info->space_info);
3894         if (flags & BTRFS_BLOCK_GROUP_DATA)
3895                 info->data_sinfo = found;
3896
3897         return ret;
3898 }
3899
3900 static void set_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
3901 {
3902         u64 extra_flags = chunk_to_extended(flags) &
3903                                 BTRFS_EXTENDED_PROFILE_MASK;
3904
3905         write_seqlock(&fs_info->profiles_lock);
3906         if (flags & BTRFS_BLOCK_GROUP_DATA)
3907                 fs_info->avail_data_alloc_bits |= extra_flags;
3908         if (flags & BTRFS_BLOCK_GROUP_METADATA)
3909                 fs_info->avail_metadata_alloc_bits |= extra_flags;
3910         if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
3911                 fs_info->avail_system_alloc_bits |= extra_flags;
3912         write_sequnlock(&fs_info->profiles_lock);
3913 }
3914
3915 /*
3916  * returns target flags in extended format or 0 if restripe for this
3917  * chunk_type is not in progress
3918  *
3919  * should be called with either volume_mutex or balance_lock held
3920  */
3921 static u64 get_restripe_target(struct btrfs_fs_info *fs_info, u64 flags)
3922 {
3923         struct btrfs_balance_control *bctl = fs_info->balance_ctl;
3924         u64 target = 0;
3925
3926         if (!bctl)
3927                 return 0;
3928
3929         if (flags & BTRFS_BLOCK_GROUP_DATA &&
3930             bctl->data.flags & BTRFS_BALANCE_ARGS_CONVERT) {
3931                 target = BTRFS_BLOCK_GROUP_DATA | bctl->data.target;
3932         } else if (flags & BTRFS_BLOCK_GROUP_SYSTEM &&
3933                    bctl->sys.flags & BTRFS_BALANCE_ARGS_CONVERT) {
3934                 target = BTRFS_BLOCK_GROUP_SYSTEM | bctl->sys.target;
3935         } else if (flags & BTRFS_BLOCK_GROUP_METADATA &&
3936                    bctl->meta.flags & BTRFS_BALANCE_ARGS_CONVERT) {
3937                 target = BTRFS_BLOCK_GROUP_METADATA | bctl->meta.target;
3938         }
3939
3940         return target;
3941 }
3942
3943 /*
3944  * @flags: available profiles in extended format (see ctree.h)
3945  *
3946  * Returns reduced profile in chunk format.  If profile changing is in
3947  * progress (either running or paused) picks the target profile (if it's
3948  * already available), otherwise falls back to plain reducing.
3949  */
3950 static u64 btrfs_reduce_alloc_profile(struct btrfs_root *root, u64 flags)
3951 {
3952         u64 num_devices = root->fs_info->fs_devices->rw_devices;
3953         u64 target;
3954         u64 raid_type;
3955         u64 allowed = 0;
3956
3957         /*
3958          * see if restripe for this chunk_type is in progress, if so
3959          * try to reduce to the target profile
3960          */
3961         spin_lock(&root->fs_info->balance_lock);
3962         target = get_restripe_target(root->fs_info, flags);
3963         if (target) {
3964                 /* pick target profile only if it's already available */
3965                 if ((flags & target) & BTRFS_EXTENDED_PROFILE_MASK) {
3966                         spin_unlock(&root->fs_info->balance_lock);
3967                         return extended_to_chunk(target);
3968                 }
3969         }
3970         spin_unlock(&root->fs_info->balance_lock);
3971
3972         /* First, mask out the RAID levels which aren't possible */
3973         for (raid_type = 0; raid_type < BTRFS_NR_RAID_TYPES; raid_type++) {
3974                 if (num_devices >= btrfs_raid_array[raid_type].devs_min)
3975                         allowed |= btrfs_raid_group[raid_type];
3976         }
3977         allowed &= flags;
3978
3979         if (allowed & BTRFS_BLOCK_GROUP_RAID6)
3980                 allowed = BTRFS_BLOCK_GROUP_RAID6;
3981         else if (allowed & BTRFS_BLOCK_GROUP_RAID5)
3982                 allowed = BTRFS_BLOCK_GROUP_RAID5;
3983         else if (allowed & BTRFS_BLOCK_GROUP_RAID10)
3984                 allowed = BTRFS_BLOCK_GROUP_RAID10;
3985         else if (allowed & BTRFS_BLOCK_GROUP_RAID1)
3986                 allowed = BTRFS_BLOCK_GROUP_RAID1;
3987         else if (allowed & BTRFS_BLOCK_GROUP_RAID0)
3988                 allowed = BTRFS_BLOCK_GROUP_RAID0;
3989
3990         flags &= ~BTRFS_BLOCK_GROUP_PROFILE_MASK;
3991
3992         return extended_to_chunk(flags | allowed);
3993 }
3994
3995 static u64 get_alloc_profile(struct btrfs_root *root, u64 orig_flags)
3996 {
3997         unsigned seq;
3998         u64 flags;
3999
4000         do {
4001                 flags = orig_flags;
4002                 seq = read_seqbegin(&root->fs_info->profiles_lock);
4003
4004                 if (flags & BTRFS_BLOCK_GROUP_DATA)
4005                         flags |= root->fs_info->avail_data_alloc_bits;
4006                 else if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
4007                         flags |= root->fs_info->avail_system_alloc_bits;
4008                 else if (flags & BTRFS_BLOCK_GROUP_METADATA)
4009                         flags |= root->fs_info->avail_metadata_alloc_bits;
4010         } while (read_seqretry(&root->fs_info->profiles_lock, seq));
4011
4012         return btrfs_reduce_alloc_profile(root, flags);
4013 }
4014
4015 u64 btrfs_get_alloc_profile(struct btrfs_root *root, int data)
4016 {
4017         u64 flags;
4018         u64 ret;
4019
4020         if (data)
4021                 flags = BTRFS_BLOCK_GROUP_DATA;
4022         else if (root == root->fs_info->chunk_root)
4023                 flags = BTRFS_BLOCK_GROUP_SYSTEM;
4024         else
4025                 flags = BTRFS_BLOCK_GROUP_METADATA;
4026
4027         ret = get_alloc_profile(root, flags);
4028         return ret;
4029 }
4030
4031 int btrfs_alloc_data_chunk_ondemand(struct inode *inode, u64 bytes)
4032 {
4033         struct btrfs_space_info *data_sinfo;
4034         struct btrfs_root *root = BTRFS_I(inode)->root;
4035         struct btrfs_fs_info *fs_info = root->fs_info;
4036         u64 used;
4037         int ret = 0;
4038         int need_commit = 2;
4039         int have_pinned_space;
4040
4041         /* make sure bytes are sectorsize aligned */
4042         bytes = ALIGN(bytes, root->sectorsize);
4043
4044         if (btrfs_is_free_space_inode(inode)) {
4045                 need_commit = 0;
4046                 ASSERT(current->journal_info);
4047         }
4048
4049         data_sinfo = fs_info->data_sinfo;
4050         if (!data_sinfo)
4051                 goto alloc;
4052
4053 again:
4054         /* make sure we have enough space to handle the data first */
4055         spin_lock(&data_sinfo->lock);
4056         used = data_sinfo->bytes_used + data_sinfo->bytes_reserved +
4057                 data_sinfo->bytes_pinned + data_sinfo->bytes_readonly +
4058                 data_sinfo->bytes_may_use;
4059
4060         if (used + bytes > data_sinfo->total_bytes) {
4061                 struct btrfs_trans_handle *trans;
4062
4063                 /*
4064                  * if we don't have enough free bytes in this space then we need
4065                  * to alloc a new chunk.
4066                  */
4067                 if (!data_sinfo->full) {
4068                         u64 alloc_target;
4069
4070                         data_sinfo->force_alloc = CHUNK_ALLOC_FORCE;
4071                         spin_unlock(&data_sinfo->lock);
4072 alloc:
4073                         alloc_target = btrfs_get_alloc_profile(root, 1);
4074                         /*
4075                          * It is ugly that we don't call nolock join
4076                          * transaction for the free space inode case here.
4077                          * But it is safe because we only do the data space
4078                          * reservation for the free space cache in the
4079                          * transaction context, the common join transaction
4080                          * just increase the counter of the current transaction
4081                          * handler, doesn't try to acquire the trans_lock of
4082                          * the fs.
4083                          */
4084                         trans = btrfs_join_transaction(root);
4085                         if (IS_ERR(trans))
4086                                 return PTR_ERR(trans);
4087
4088                         ret = do_chunk_alloc(trans, root->fs_info->extent_root,
4089                                              alloc_target,
4090                                              CHUNK_ALLOC_NO_FORCE);
4091                         btrfs_end_transaction(trans, root);
4092                         if (ret < 0) {
4093                                 if (ret != -ENOSPC)
4094                                         return ret;
4095                                 else {
4096                                         have_pinned_space = 1;
4097                                         goto commit_trans;
4098                                 }
4099                         }
4100
4101                         if (!data_sinfo)
4102                                 data_sinfo = fs_info->data_sinfo;
4103
4104                         goto again;
4105                 }
4106
4107                 /*
4108                  * If we don't have enough pinned space to deal with this
4109                  * allocation, and no removed chunk in current transaction,
4110                  * don't bother committing the transaction.
4111                  */
4112                 have_pinned_space = percpu_counter_compare(
4113                         &data_sinfo->total_bytes_pinned,
4114                         used + bytes - data_sinfo->total_bytes);
4115                 spin_unlock(&data_sinfo->lock);
4116
4117                 /* commit the current transaction and try again */
4118 commit_trans:
4119                 if (need_commit &&
4120                     !atomic_read(&root->fs_info->open_ioctl_trans)) {
4121                         need_commit--;
4122
4123                         if (need_commit > 0)
4124                                 btrfs_wait_ordered_roots(fs_info, -1);
4125
4126                         trans = btrfs_join_transaction(root);
4127                         if (IS_ERR(trans))
4128                                 return PTR_ERR(trans);
4129                         if (have_pinned_space >= 0 ||
4130                             test_bit(BTRFS_TRANS_HAVE_FREE_BGS,
4131                                      &trans->transaction->flags) ||
4132                             need_commit > 0) {
4133                                 ret = btrfs_commit_transaction(trans, root);
4134                                 if (ret)
4135                                         return ret;
4136                                 /*
4137                                  * make sure that all running delayed iput are
4138                                  * done
4139                                  */
4140                                 down_write(&root->fs_info->delayed_iput_sem);
4141                                 up_write(&root->fs_info->delayed_iput_sem);
4142                                 goto again;
4143                         } else {
4144                                 btrfs_end_transaction(trans, root);
4145                         }
4146                 }
4147
4148                 trace_btrfs_space_reservation(root->fs_info,
4149                                               "space_info:enospc",
4150                                               data_sinfo->flags, bytes, 1);
4151                 return -ENOSPC;
4152         }
4153         data_sinfo->bytes_may_use += bytes;
4154         trace_btrfs_space_reservation(root->fs_info, "space_info",
4155                                       data_sinfo->flags, bytes, 1);
4156         spin_unlock(&data_sinfo->lock);
4157
4158         return ret;
4159 }
4160
4161 /*
4162  * New check_data_free_space() with ability for precious data reservation
4163  * Will replace old btrfs_check_data_free_space(), but for patch split,
4164  * add a new function first and then replace it.
4165  */
4166 int btrfs_check_data_free_space(struct inode *inode, u64 start, u64 len)
4167 {
4168         struct btrfs_root *root = BTRFS_I(inode)->root;
4169         int ret;
4170
4171         /* align the range */
4172         len = round_up(start + len, root->sectorsize) -
4173               round_down(start, root->sectorsize);
4174         start = round_down(start, root->sectorsize);
4175
4176         ret = btrfs_alloc_data_chunk_ondemand(inode, len);
4177         if (ret < 0)
4178                 return ret;
4179
4180         /*
4181          * Use new btrfs_qgroup_reserve_data to reserve precious data space
4182          *
4183          * TODO: Find a good method to avoid reserve data space for NOCOW
4184          * range, but don't impact performance on quota disable case.
4185          */
4186         ret = btrfs_qgroup_reserve_data(inode, start, len);
4187         return ret;
4188 }
4189
4190 /*
4191  * Called if we need to clear a data reservation for this inode
4192  * Normally in a error case.
4193  *
4194  * This one will *NOT* use accurate qgroup reserved space API, just for case
4195  * which we can't sleep and is sure it won't affect qgroup reserved space.
4196  * Like clear_bit_hook().
4197  */
4198 void btrfs_free_reserved_data_space_noquota(struct inode *inode, u64 start,
4199                                             u64 len)
4200 {
4201         struct btrfs_root *root = BTRFS_I(inode)->root;
4202         struct btrfs_space_info *data_sinfo;
4203
4204         /* Make sure the range is aligned to sectorsize */
4205         len = round_up(start + len, root->sectorsize) -
4206               round_down(start, root->sectorsize);
4207         start = round_down(start, root->sectorsize);
4208
4209         data_sinfo = root->fs_info->data_sinfo;
4210         spin_lock(&data_sinfo->lock);
4211         if (WARN_ON(data_sinfo->bytes_may_use < len))
4212                 data_sinfo->bytes_may_use = 0;
4213         else
4214                 data_sinfo->bytes_may_use -= len;
4215         trace_btrfs_space_reservation(root->fs_info, "space_info",
4216                                       data_sinfo->flags, len, 0);
4217         spin_unlock(&data_sinfo->lock);
4218 }
4219
4220 /*
4221  * Called if we need to clear a data reservation for this inode
4222  * Normally in a error case.
4223  *
4224  * This one will handle the per-indoe data rsv map for accurate reserved
4225  * space framework.
4226  */
4227 void btrfs_free_reserved_data_space(struct inode *inode, u64 start, u64 len)
4228 {
4229         btrfs_free_reserved_data_space_noquota(inode, start, len);
4230         btrfs_qgroup_free_data(inode, start, len);
4231 }
4232
4233 static void force_metadata_allocation(struct btrfs_fs_info *info)
4234 {
4235         struct list_head *head = &info->space_info;
4236         struct btrfs_space_info *found;
4237
4238         rcu_read_lock();
4239         list_for_each_entry_rcu(found, head, list) {
4240                 if (found->flags & BTRFS_BLOCK_GROUP_METADATA)
4241                         found->force_alloc = CHUNK_ALLOC_FORCE;
4242         }
4243         rcu_read_unlock();
4244 }
4245
4246 static inline u64 calc_global_rsv_need_space(struct btrfs_block_rsv *global)
4247 {
4248         return (global->size << 1);
4249 }
4250
4251 static int should_alloc_chunk(struct btrfs_root *root,
4252                               struct btrfs_space_info *sinfo, int force)
4253 {
4254         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
4255         u64 num_bytes = sinfo->total_bytes - sinfo->bytes_readonly;
4256         u64 num_allocated = sinfo->bytes_used + sinfo->bytes_reserved;
4257         u64 thresh;
4258
4259         if (force == CHUNK_ALLOC_FORCE)
4260                 return 1;
4261
4262         /*
4263          * We need to take into account the global rsv because for all intents
4264          * and purposes it's used space.  Don't worry about locking the
4265          * global_rsv, it doesn't change except when the transaction commits.
4266          */
4267         if (sinfo->flags & BTRFS_BLOCK_GROUP_METADATA)
4268                 num_allocated += calc_global_rsv_need_space(global_rsv);
4269
4270         /*
4271          * in limited mode, we want to have some free space up to
4272          * about 1% of the FS size.
4273          */
4274         if (force == CHUNK_ALLOC_LIMITED) {
4275                 thresh = btrfs_super_total_bytes(root->fs_info->super_copy);
4276                 thresh = max_t(u64, 64 * 1024 * 1024,
4277                                div_factor_fine(thresh, 1));
4278
4279                 if (num_bytes - num_allocated < thresh)
4280                         return 1;
4281         }
4282
4283         if (num_allocated + 2 * 1024 * 1024 < div_factor(num_bytes, 8))
4284                 return 0;
4285         return 1;
4286 }
4287
4288 static u64 get_profile_num_devs(struct btrfs_root *root, u64 type)
4289 {
4290         u64 num_dev;
4291
4292         if (type & (BTRFS_BLOCK_GROUP_RAID10 |
4293                     BTRFS_BLOCK_GROUP_RAID0 |
4294                     BTRFS_BLOCK_GROUP_RAID5 |
4295                     BTRFS_BLOCK_GROUP_RAID6))
4296                 num_dev = root->fs_info->fs_devices->rw_devices;
4297         else if (type & BTRFS_BLOCK_GROUP_RAID1)
4298                 num_dev = 2;
4299         else
4300                 num_dev = 1;    /* DUP or single */
4301
4302         return num_dev;
4303 }
4304
4305 /*
4306  * If @is_allocation is true, reserve space in the system space info necessary
4307  * for allocating a chunk, otherwise if it's false, reserve space necessary for
4308  * removing a chunk.
4309  */
4310 void check_system_chunk(struct btrfs_trans_handle *trans,
4311                         struct btrfs_root *root,
4312                         u64 type)
4313 {
4314         struct btrfs_space_info *info;
4315         u64 left;
4316         u64 thresh;
4317         int ret = 0;
4318         u64 num_devs;
4319
4320         /*
4321          * Needed because we can end up allocating a system chunk and for an
4322          * atomic and race free space reservation in the chunk block reserve.
4323          */
4324         ASSERT(mutex_is_locked(&root->fs_info->chunk_mutex));
4325
4326         info = __find_space_info(root->fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
4327         spin_lock(&info->lock);
4328         left = info->total_bytes - info->bytes_used - info->bytes_pinned -
4329                 info->bytes_reserved - info->bytes_readonly -
4330                 info->bytes_may_use;
4331         spin_unlock(&info->lock);
4332
4333         num_devs = get_profile_num_devs(root, type);
4334
4335         /* num_devs device items to update and 1 chunk item to add or remove */
4336         thresh = btrfs_calc_trunc_metadata_size(root, num_devs) +
4337                 btrfs_calc_trans_metadata_size(root, 1);
4338
4339         if (left < thresh && btrfs_test_opt(root, ENOSPC_DEBUG)) {
4340                 btrfs_info(root->fs_info, "left=%llu, need=%llu, flags=%llu",
4341                         left, thresh, type);
4342                 dump_space_info(info, 0, 0);
4343         }
4344
4345         if (left < thresh) {
4346                 u64 flags;
4347
4348                 flags = btrfs_get_alloc_profile(root->fs_info->chunk_root, 0);
4349                 /*
4350                  * Ignore failure to create system chunk. We might end up not
4351                  * needing it, as we might not need to COW all nodes/leafs from
4352                  * the paths we visit in the chunk tree (they were already COWed
4353                  * or created in the current transaction for example).
4354                  */
4355                 ret = btrfs_alloc_chunk(trans, root, flags);
4356         }
4357
4358         if (!ret) {
4359                 ret = btrfs_block_rsv_add(root->fs_info->chunk_root,
4360                                           &root->fs_info->chunk_block_rsv,
4361                                           thresh, BTRFS_RESERVE_NO_FLUSH);
4362                 if (!ret)
4363                         trans->chunk_bytes_reserved += thresh;
4364         }
4365 }
4366
4367 static int do_chunk_alloc(struct btrfs_trans_handle *trans,
4368                           struct btrfs_root *extent_root, u64 flags, int force)
4369 {
4370         struct btrfs_space_info *space_info;
4371         struct btrfs_fs_info *fs_info = extent_root->fs_info;
4372         int wait_for_alloc = 0;
4373         int ret = 0;
4374
4375         /* Don't re-enter if we're already allocating a chunk */
4376         if (trans->allocating_chunk)
4377                 return -ENOSPC;
4378
4379         space_info = __find_space_info(extent_root->fs_info, flags);
4380         if (!space_info) {
4381                 ret = update_space_info(extent_root->fs_info, flags,
4382                                         0, 0, &space_info);
4383                 BUG_ON(ret); /* -ENOMEM */
4384         }
4385         BUG_ON(!space_info); /* Logic error */
4386
4387 again:
4388         spin_lock(&space_info->lock);
4389         if (force < space_info->force_alloc)
4390                 force = space_info->force_alloc;
4391         if (space_info->full) {
4392                 if (should_alloc_chunk(extent_root, space_info, force))
4393                         ret = -ENOSPC;
4394                 else
4395                         ret = 0;
4396                 spin_unlock(&space_info->lock);
4397                 return ret;
4398         }
4399
4400         if (!should_alloc_chunk(extent_root, space_info, force)) {
4401                 spin_unlock(&space_info->lock);
4402                 return 0;
4403         } else if (space_info->chunk_alloc) {
4404                 wait_for_alloc = 1;
4405         } else {
4406                 space_info->chunk_alloc = 1;
4407         }
4408
4409         spin_unlock(&space_info->lock);
4410
4411         mutex_lock(&fs_info->chunk_mutex);
4412
4413         /*
4414          * The chunk_mutex is held throughout the entirety of a chunk
4415          * allocation, so once we've acquired the chunk_mutex we know that the
4416          * other guy is done and we need to recheck and see if we should
4417          * allocate.
4418          */
4419         if (wait_for_alloc) {
4420                 mutex_unlock(&fs_info->chunk_mutex);
4421                 wait_for_alloc = 0;
4422                 goto again;
4423         }
4424
4425         trans->allocating_chunk = true;
4426
4427         /*
4428          * If we have mixed data/metadata chunks we want to make sure we keep
4429          * allocating mixed chunks instead of individual chunks.
4430          */
4431         if (btrfs_mixed_space_info(space_info))
4432                 flags |= (BTRFS_BLOCK_GROUP_DATA | BTRFS_BLOCK_GROUP_METADATA);
4433
4434         /*
4435          * if we're doing a data chunk, go ahead and make sure that
4436          * we keep a reasonable number of metadata chunks allocated in the
4437          * FS as well.
4438          */
4439         if (flags & BTRFS_BLOCK_GROUP_DATA && fs_info->metadata_ratio) {
4440                 fs_info->data_chunk_allocations++;
4441                 if (!(fs_info->data_chunk_allocations %
4442                       fs_info->metadata_ratio))
4443                         force_metadata_allocation(fs_info);
4444         }
4445
4446         /*
4447          * Check if we have enough space in SYSTEM chunk because we may need
4448          * to update devices.
4449          */
4450         check_system_chunk(trans, extent_root, flags);
4451
4452         ret = btrfs_alloc_chunk(trans, extent_root, flags);
4453         trans->allocating_chunk = false;
4454
4455         spin_lock(&space_info->lock);
4456         if (ret < 0 && ret != -ENOSPC)
4457                 goto out;
4458         if (ret)
4459                 space_info->full = 1;
4460         else
4461                 ret = 1;
4462
4463         space_info->force_alloc = CHUNK_ALLOC_NO_FORCE;
4464 out:
4465         space_info->chunk_alloc = 0;
4466         spin_unlock(&space_info->lock);
4467         mutex_unlock(&fs_info->chunk_mutex);
4468         /*
4469          * When we allocate a new chunk we reserve space in the chunk block
4470          * reserve to make sure we can COW nodes/leafs in the chunk tree or
4471          * add new nodes/leafs to it if we end up needing to do it when
4472          * inserting the chunk item and updating device items as part of the
4473          * second phase of chunk allocation, performed by
4474          * btrfs_finish_chunk_alloc(). So make sure we don't accumulate a
4475          * large number of new block groups to create in our transaction
4476          * handle's new_bgs list to avoid exhausting the chunk block reserve
4477          * in extreme cases - like having a single transaction create many new
4478          * block groups when starting to write out the free space caches of all
4479          * the block groups that were made dirty during the lifetime of the
4480          * transaction.
4481          */
4482         if (trans->can_flush_pending_bgs &&
4483             trans->chunk_bytes_reserved >= (2 * 1024 * 1024ull)) {
4484                 btrfs_create_pending_block_groups(trans, trans->root);
4485                 btrfs_trans_release_chunk_metadata(trans);
4486         }
4487         return ret;
4488 }
4489
4490 static int can_overcommit(struct btrfs_root *root,
4491                           struct btrfs_space_info *space_info, u64 bytes,
4492                           enum btrfs_reserve_flush_enum flush)
4493 {
4494         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
4495         u64 profile = btrfs_get_alloc_profile(root, 0);
4496         u64 space_size;
4497         u64 avail;
4498         u64 used;
4499
4500         used = space_info->bytes_used + space_info->bytes_reserved +
4501                 space_info->bytes_pinned + space_info->bytes_readonly;
4502
4503         /*
4504          * We only want to allow over committing if we have lots of actual space
4505          * free, but if we don't have enough space to handle the global reserve
4506          * space then we could end up having a real enospc problem when trying
4507          * to allocate a chunk or some other such important allocation.
4508          */
4509         spin_lock(&global_rsv->lock);
4510         space_size = calc_global_rsv_need_space(global_rsv);
4511         spin_unlock(&global_rsv->lock);
4512         if (used + space_size >= space_info->total_bytes)
4513                 return 0;
4514
4515         used += space_info->bytes_may_use;
4516
4517         spin_lock(&root->fs_info->free_chunk_lock);
4518         avail = root->fs_info->free_chunk_space;
4519         spin_unlock(&root->fs_info->free_chunk_lock);
4520
4521         /*
4522          * If we have dup, raid1 or raid10 then only half of the free
4523          * space is actually useable.  For raid56, the space info used
4524          * doesn't include the parity drive, so we don't have to
4525          * change the math
4526          */
4527         if (profile & (BTRFS_BLOCK_GROUP_DUP |
4528                        BTRFS_BLOCK_GROUP_RAID1 |
4529                        BTRFS_BLOCK_GROUP_RAID10))
4530                 avail >>= 1;
4531
4532         /*
4533          * If we aren't flushing all things, let us overcommit up to
4534          * 1/2th of the space. If we can flush, don't let us overcommit
4535          * too much, let it overcommit up to 1/8 of the space.
4536          */
4537         if (flush == BTRFS_RESERVE_FLUSH_ALL)
4538                 avail >>= 3;
4539         else
4540                 avail >>= 1;
4541
4542         if (used + bytes < space_info->total_bytes + avail)
4543                 return 1;
4544         return 0;
4545 }
4546
4547 static void btrfs_writeback_inodes_sb_nr(struct btrfs_root *root,
4548                                          unsigned long nr_pages, int nr_items)
4549 {
4550         struct super_block *sb = root->fs_info->sb;
4551
4552         if (down_read_trylock(&sb->s_umount)) {
4553                 writeback_inodes_sb_nr(sb, nr_pages, WB_REASON_FS_FREE_SPACE);
4554                 up_read(&sb->s_umount);
4555         } else {
4556                 /*
4557                  * We needn't worry the filesystem going from r/w to r/o though
4558                  * we don't acquire ->s_umount mutex, because the filesystem
4559                  * should guarantee the delalloc inodes list be empty after
4560                  * the filesystem is readonly(all dirty pages are written to
4561                  * the disk).
4562                  */
4563                 btrfs_start_delalloc_roots(root->fs_info, 0, nr_items);
4564                 if (!current->journal_info)
4565                         btrfs_wait_ordered_roots(root->fs_info, nr_items);
4566         }
4567 }
4568
4569 static inline int calc_reclaim_items_nr(struct btrfs_root *root, u64 to_reclaim)
4570 {
4571         u64 bytes;
4572         int nr;
4573
4574         bytes = btrfs_calc_trans_metadata_size(root, 1);
4575         nr = (int)div64_u64(to_reclaim, bytes);
4576         if (!nr)
4577                 nr = 1;
4578         return nr;
4579 }
4580
4581 #define EXTENT_SIZE_PER_ITEM    (256 * 1024)
4582
4583 /*
4584  * shrink metadata reservation for delalloc
4585  */
4586 static void shrink_delalloc(struct btrfs_root *root, u64 to_reclaim, u64 orig,
4587                             bool wait_ordered)
4588 {
4589         struct btrfs_block_rsv *block_rsv;
4590         struct btrfs_space_info *space_info;
4591         struct btrfs_trans_handle *trans;
4592         u64 delalloc_bytes;
4593         u64 max_reclaim;
4594         long time_left;
4595         unsigned long nr_pages;
4596         int loops;
4597         int items;
4598         enum btrfs_reserve_flush_enum flush;
4599
4600         /* Calc the number of the pages we need flush for space reservation */
4601         items = calc_reclaim_items_nr(root, to_reclaim);
4602         to_reclaim = items * EXTENT_SIZE_PER_ITEM;
4603
4604         trans = (struct btrfs_trans_handle *)current->journal_info;
4605         block_rsv = &root->fs_info->delalloc_block_rsv;
4606         space_info = block_rsv->space_info;
4607
4608         delalloc_bytes = percpu_counter_sum_positive(
4609                                                 &root->fs_info->delalloc_bytes);
4610         if (delalloc_bytes == 0) {
4611                 if (trans)
4612                         return;
4613                 if (wait_ordered)
4614                         btrfs_wait_ordered_roots(root->fs_info, items);
4615                 return;
4616         }
4617
4618         loops = 0;
4619         while (delalloc_bytes && loops < 3) {
4620                 max_reclaim = min(delalloc_bytes, to_reclaim);
4621                 nr_pages = max_reclaim >> PAGE_CACHE_SHIFT;
4622                 btrfs_writeback_inodes_sb_nr(root, nr_pages, items);
4623                 /*
4624                  * We need to wait for the async pages to actually start before
4625                  * we do anything.
4626                  */
4627                 max_reclaim = atomic_read(&root->fs_info->async_delalloc_pages);
4628                 if (!max_reclaim)
4629                         goto skip_async;
4630
4631                 if (max_reclaim <= nr_pages)
4632                         max_reclaim = 0;
4633                 else
4634                         max_reclaim -= nr_pages;
4635
4636                 wait_event(root->fs_info->async_submit_wait,
4637                            atomic_read(&root->fs_info->async_delalloc_pages) <=
4638                            (int)max_reclaim);
4639 skip_async:
4640                 if (!trans)
4641                         flush = BTRFS_RESERVE_FLUSH_ALL;
4642                 else
4643                         flush = BTRFS_RESERVE_NO_FLUSH;
4644                 spin_lock(&space_info->lock);
4645                 if (can_overcommit(root, space_info, orig, flush)) {
4646                         spin_unlock(&space_info->lock);
4647                         break;
4648                 }
4649                 spin_unlock(&space_info->lock);
4650
4651                 loops++;
4652                 if (wait_ordered && !trans) {
4653                         btrfs_wait_ordered_roots(root->fs_info, items);
4654                 } else {
4655                         time_left = schedule_timeout_killable(1);
4656                         if (time_left)
4657                                 break;
4658                 }
4659                 delalloc_bytes = percpu_counter_sum_positive(
4660                                                 &root->fs_info->delalloc_bytes);
4661         }
4662 }
4663
4664 /**
4665  * maybe_commit_transaction - possibly commit the transaction if its ok to
4666  * @root - the root we're allocating for
4667  * @bytes - the number of bytes we want to reserve
4668  * @force - force the commit
4669  *
4670  * This will check to make sure that committing the transaction will actually
4671  * get us somewhere and then commit the transaction if it does.  Otherwise it
4672  * will return -ENOSPC.
4673  */
4674 static int may_commit_transaction(struct btrfs_root *root,
4675                                   struct btrfs_space_info *space_info,
4676                                   u64 bytes, int force)
4677 {
4678         struct btrfs_block_rsv *delayed_rsv = &root->fs_info->delayed_block_rsv;
4679         struct btrfs_trans_handle *trans;
4680
4681         trans = (struct btrfs_trans_handle *)current->journal_info;
4682         if (trans)
4683                 return -EAGAIN;
4684
4685         if (force)
4686                 goto commit;
4687
4688         /* See if there is enough pinned space to make this reservation */
4689         if (percpu_counter_compare(&space_info->total_bytes_pinned,
4690                                    bytes) >= 0)
4691                 goto commit;
4692
4693         /*
4694          * See if there is some space in the delayed insertion reservation for
4695          * this reservation.
4696          */
4697         if (space_info != delayed_rsv->space_info)
4698                 return -ENOSPC;
4699
4700         spin_lock(&delayed_rsv->lock);
4701         if (percpu_counter_compare(&space_info->total_bytes_pinned,
4702                                    bytes - delayed_rsv->size) >= 0) {
4703                 spin_unlock(&delayed_rsv->lock);
4704                 return -ENOSPC;
4705         }
4706         spin_unlock(&delayed_rsv->lock);
4707
4708 commit:
4709         trans = btrfs_join_transaction(root);
4710         if (IS_ERR(trans))
4711                 return -ENOSPC;
4712
4713         return btrfs_commit_transaction(trans, root);
4714 }
4715
4716 enum flush_state {
4717         FLUSH_DELAYED_ITEMS_NR  =       1,
4718         FLUSH_DELAYED_ITEMS     =       2,
4719         FLUSH_DELALLOC          =       3,
4720         FLUSH_DELALLOC_WAIT     =       4,
4721         ALLOC_CHUNK             =       5,
4722         COMMIT_TRANS            =       6,
4723 };
4724
4725 static int flush_space(struct btrfs_root *root,
4726                        struct btrfs_space_info *space_info, u64 num_bytes,
4727                        u64 orig_bytes, int state)
4728 {
4729         struct btrfs_trans_handle *trans;
4730         int nr;
4731         int ret = 0;
4732
4733         switch (state) {
4734         case FLUSH_DELAYED_ITEMS_NR:
4735         case FLUSH_DELAYED_ITEMS:
4736                 if (state == FLUSH_DELAYED_ITEMS_NR)
4737                         nr = calc_reclaim_items_nr(root, num_bytes) * 2;
4738                 else
4739                         nr = -1;
4740
4741                 trans = btrfs_join_transaction(root);
4742                 if (IS_ERR(trans)) {
4743                         ret = PTR_ERR(trans);
4744                         break;
4745                 }
4746                 ret = btrfs_run_delayed_items_nr(trans, root, nr);
4747                 btrfs_end_transaction(trans, root);
4748                 break;
4749         case FLUSH_DELALLOC:
4750         case FLUSH_DELALLOC_WAIT:
4751                 shrink_delalloc(root, num_bytes * 2, orig_bytes,
4752                                 state == FLUSH_DELALLOC_WAIT);
4753                 break;
4754         case ALLOC_CHUNK:
4755                 trans = btrfs_join_transaction(root);
4756                 if (IS_ERR(trans)) {
4757                         ret = PTR_ERR(trans);
4758                         break;
4759                 }
4760                 ret = do_chunk_alloc(trans, root->fs_info->extent_root,
4761                                      btrfs_get_alloc_profile(root, 0),
4762                                      CHUNK_ALLOC_NO_FORCE);
4763                 btrfs_end_transaction(trans, root);
4764                 if (ret == -ENOSPC)
4765                         ret = 0;
4766                 break;
4767         case COMMIT_TRANS:
4768                 ret = may_commit_transaction(root, space_info, orig_bytes, 0);
4769                 break;
4770         default:
4771                 ret = -ENOSPC;
4772                 break;
4773         }
4774
4775         return ret;
4776 }
4777
4778 static inline u64
4779 btrfs_calc_reclaim_metadata_size(struct btrfs_root *root,
4780                                  struct btrfs_space_info *space_info)
4781 {
4782         u64 used;
4783         u64 expected;
4784         u64 to_reclaim;
4785
4786         to_reclaim = min_t(u64, num_online_cpus() * 1024 * 1024,
4787                                 16 * 1024 * 1024);
4788         spin_lock(&space_info->lock);
4789         if (can_overcommit(root, space_info, to_reclaim,
4790                            BTRFS_RESERVE_FLUSH_ALL)) {
4791                 to_reclaim = 0;
4792                 goto out;
4793         }
4794
4795         used = space_info->bytes_used + space_info->bytes_reserved +
4796                space_info->bytes_pinned + space_info->bytes_readonly +
4797                space_info->bytes_may_use;
4798         if (can_overcommit(root, space_info, 1024 * 1024,
4799                            BTRFS_RESERVE_FLUSH_ALL))
4800                 expected = div_factor_fine(space_info->total_bytes, 95);
4801         else
4802                 expected = div_factor_fine(space_info->total_bytes, 90);
4803
4804         if (used > expected)
4805                 to_reclaim = used - expected;
4806         else
4807                 to_reclaim = 0;
4808         to_reclaim = min(to_reclaim, space_info->bytes_may_use +
4809                                      space_info->bytes_reserved);
4810 out:
4811         spin_unlock(&space_info->lock);
4812
4813         return to_reclaim;
4814 }
4815
4816 static inline int need_do_async_reclaim(struct btrfs_space_info *space_info,
4817                                         struct btrfs_fs_info *fs_info, u64 used)
4818 {
4819         u64 thresh = div_factor_fine(space_info->total_bytes, 98);
4820
4821         /* If we're just plain full then async reclaim just slows us down. */
4822         if (space_info->bytes_used >= thresh)
4823                 return 0;
4824
4825         return (used >= thresh && !btrfs_fs_closing(fs_info) &&
4826                 !test_bit(BTRFS_FS_STATE_REMOUNTING, &fs_info->fs_state));
4827 }
4828
4829 static int btrfs_need_do_async_reclaim(struct btrfs_space_info *space_info,
4830                                        struct btrfs_fs_info *fs_info,
4831                                        int flush_state)
4832 {
4833         u64 used;
4834
4835         spin_lock(&space_info->lock);
4836         /*
4837          * We run out of space and have not got any free space via flush_space,
4838          * so don't bother doing async reclaim.
4839          */
4840         if (flush_state > COMMIT_TRANS && space_info->full) {
4841                 spin_unlock(&space_info->lock);
4842                 return 0;
4843         }
4844
4845         used = space_info->bytes_used + space_info->bytes_reserved +
4846                space_info->bytes_pinned + space_info->bytes_readonly +
4847                space_info->bytes_may_use;
4848         if (need_do_async_reclaim(space_info, fs_info, used)) {
4849                 spin_unlock(&space_info->lock);
4850                 return 1;
4851         }
4852         spin_unlock(&space_info->lock);
4853
4854         return 0;
4855 }
4856
4857 static void btrfs_async_reclaim_metadata_space(struct work_struct *work)
4858 {
4859         struct btrfs_fs_info *fs_info;
4860         struct btrfs_space_info *space_info;
4861         u64 to_reclaim;
4862         int flush_state;
4863
4864         fs_info = container_of(work, struct btrfs_fs_info, async_reclaim_work);
4865         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
4866
4867         to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info->fs_root,
4868                                                       space_info);
4869         if (!to_reclaim)
4870                 return;
4871
4872         flush_state = FLUSH_DELAYED_ITEMS_NR;
4873         do {
4874                 flush_space(fs_info->fs_root, space_info, to_reclaim,
4875                             to_reclaim, flush_state);
4876                 flush_state++;
4877                 if (!btrfs_need_do_async_reclaim(space_info, fs_info,
4878                                                  flush_state))
4879                         return;
4880         } while (flush_state < COMMIT_TRANS);
4881 }
4882
4883 void btrfs_init_async_reclaim_work(struct work_struct *work)
4884 {
4885         INIT_WORK(work, btrfs_async_reclaim_metadata_space);
4886 }
4887
4888 /**
4889  * reserve_metadata_bytes - try to reserve bytes from the block_rsv's space
4890  * @root - the root we're allocating for
4891  * @block_rsv - the block_rsv we're allocating for
4892  * @orig_bytes - the number of bytes we want
4893  * @flush - whether or not we can flush to make our reservation
4894  *
4895  * This will reserve orgi_bytes number of bytes from the space info associated
4896  * with the block_rsv.  If there is not enough space it will make an attempt to
4897  * flush out space to make room.  It will do this by flushing delalloc if
4898  * possible or committing the transaction.  If flush is 0 then no attempts to
4899  * regain reservations will be made and this will fail if there is not enough
4900  * space already.
4901  */
4902 static int reserve_metadata_bytes(struct btrfs_root *root,
4903                                   struct btrfs_block_rsv *block_rsv,
4904                                   u64 orig_bytes,
4905                                   enum btrfs_reserve_flush_enum flush)
4906 {
4907         struct btrfs_space_info *space_info = block_rsv->space_info;
4908         u64 used;
4909         u64 num_bytes = orig_bytes;
4910         int flush_state = FLUSH_DELAYED_ITEMS_NR;
4911         int ret = 0;
4912         bool flushing = false;
4913
4914 again:
4915         ret = 0;
4916         spin_lock(&space_info->lock);
4917         /*
4918          * We only want to wait if somebody other than us is flushing and we
4919          * are actually allowed to flush all things.
4920          */
4921         while (flush == BTRFS_RESERVE_FLUSH_ALL && !flushing &&
4922                space_info->flush) {
4923                 spin_unlock(&space_info->lock);
4924                 /*
4925                  * If we have a trans handle we can't wait because the flusher
4926                  * may have to commit the transaction, which would mean we would
4927                  * deadlock since we are waiting for the flusher to finish, but
4928                  * hold the current transaction open.
4929                  */
4930                 if (current->journal_info)
4931                         return -EAGAIN;
4932                 ret = wait_event_killable(space_info->wait, !space_info->flush);
4933                 /* Must have been killed, return */
4934                 if (ret)
4935                         return -EINTR;
4936
4937                 spin_lock(&space_info->lock);
4938         }
4939
4940         ret = -ENOSPC;
4941         used = space_info->bytes_used + space_info->bytes_reserved +
4942                 space_info->bytes_pinned + space_info->bytes_readonly +
4943                 space_info->bytes_may_use;
4944
4945         /*
4946          * The idea here is that we've not already over-reserved the block group
4947          * then we can go ahead and save our reservation first and then start
4948          * flushing if we need to.  Otherwise if we've already overcommitted
4949          * lets start flushing stuff first and then come back and try to make
4950          * our reservation.
4951          */
4952         if (used <= space_info->total_bytes) {
4953                 if (used + orig_bytes <= space_info->total_bytes) {
4954                         space_info->bytes_may_use += orig_bytes;
4955                         trace_btrfs_space_reservation(root->fs_info,
4956                                 "space_info", space_info->flags, orig_bytes, 1);
4957                         ret = 0;
4958                 } else {
4959                         /*
4960                          * Ok set num_bytes to orig_bytes since we aren't
4961                          * overocmmitted, this way we only try and reclaim what
4962                          * we need.
4963                          */
4964                         num_bytes = orig_bytes;
4965                 }
4966         } else {
4967                 /*
4968                  * Ok we're over committed, set num_bytes to the overcommitted
4969                  * amount plus the amount of bytes that we need for this
4970                  * reservation.
4971                  */
4972                 num_bytes = used - space_info->total_bytes +
4973                         (orig_bytes * 2);
4974         }
4975
4976         if (ret && can_overcommit(root, space_info, orig_bytes, flush)) {
4977                 space_info->bytes_may_use += orig_bytes;
4978                 trace_btrfs_space_reservation(root->fs_info, "space_info",
4979                                               space_info->flags, orig_bytes,
4980                                               1);
4981                 ret = 0;
4982         }
4983
4984         /*
4985          * Couldn't make our reservation, save our place so while we're trying
4986          * to reclaim space we can actually use it instead of somebody else
4987          * stealing it from us.
4988          *
4989          * We make the other tasks wait for the flush only when we can flush
4990          * all things.
4991          */
4992         if (ret && flush != BTRFS_RESERVE_NO_FLUSH) {
4993                 flushing = true;
4994                 space_info->flush = 1;
4995         } else if (!ret && space_info->flags & BTRFS_BLOCK_GROUP_METADATA) {
4996                 used += orig_bytes;
4997                 /*
4998                  * We will do the space reservation dance during log replay,
4999                  * which means we won't have fs_info->fs_root set, so don't do
5000                  * the async reclaim as we will panic.
5001                  */
5002                 if (!root->fs_info->log_root_recovering &&
5003                     need_do_async_reclaim(space_info, root->fs_info, used) &&
5004                     !work_busy(&root->fs_info->async_reclaim_work))
5005                         queue_work(system_unbound_wq,
5006                                    &root->fs_info->async_reclaim_work);
5007         }
5008         spin_unlock(&space_info->lock);
5009
5010         if (!ret || flush == BTRFS_RESERVE_NO_FLUSH)
5011                 goto out;
5012
5013         ret = flush_space(root, space_info, num_bytes, orig_bytes,
5014                           flush_state);
5015         flush_state++;
5016
5017         /*
5018          * If we are FLUSH_LIMIT, we can not flush delalloc, or the deadlock
5019          * would happen. So skip delalloc flush.
5020          */
5021         if (flush == BTRFS_RESERVE_FLUSH_LIMIT &&
5022             (flush_state == FLUSH_DELALLOC ||
5023              flush_state == FLUSH_DELALLOC_WAIT))
5024                 flush_state = ALLOC_CHUNK;
5025
5026         if (!ret)
5027                 goto again;
5028         else if (flush == BTRFS_RESERVE_FLUSH_LIMIT &&
5029                  flush_state < COMMIT_TRANS)
5030                 goto again;
5031         else if (flush == BTRFS_RESERVE_FLUSH_ALL &&
5032                  flush_state <= COMMIT_TRANS)
5033                 goto again;
5034
5035 out:
5036         if (ret == -ENOSPC &&
5037             unlikely(root->orphan_cleanup_state == ORPHAN_CLEANUP_STARTED)) {
5038                 struct btrfs_block_rsv *global_rsv =
5039                         &root->fs_info->global_block_rsv;
5040
5041                 if (block_rsv != global_rsv &&
5042                     !block_rsv_use_bytes(global_rsv, orig_bytes))
5043                         ret = 0;
5044         }
5045         if (ret == -ENOSPC)
5046                 trace_btrfs_space_reservation(root->fs_info,
5047                                               "space_info:enospc",
5048                                               space_info->flags, orig_bytes, 1);
5049         if (flushing) {
5050                 spin_lock(&space_info->lock);
5051                 space_info->flush = 0;
5052                 wake_up_all(&space_info->wait);
5053                 spin_unlock(&space_info->lock);
5054         }
5055         return ret;
5056 }
5057
5058 static struct btrfs_block_rsv *get_block_rsv(
5059                                         const struct btrfs_trans_handle *trans,
5060                                         const struct btrfs_root *root)
5061 {
5062         struct btrfs_block_rsv *block_rsv = NULL;
5063
5064         if (test_bit(BTRFS_ROOT_REF_COWS, &root->state) ||
5065             (root == root->fs_info->csum_root && trans->adding_csums) ||
5066              (root == root->fs_info->uuid_root))
5067                 block_rsv = trans->block_rsv;
5068
5069         if (!block_rsv)
5070                 block_rsv = root->block_rsv;
5071
5072         if (!block_rsv)
5073                 block_rsv = &root->fs_info->empty_block_rsv;
5074
5075         return block_rsv;
5076 }
5077
5078 static int block_rsv_use_bytes(struct btrfs_block_rsv *block_rsv,
5079                                u64 num_bytes)
5080 {
5081         int ret = -ENOSPC;
5082         spin_lock(&block_rsv->lock);
5083         if (block_rsv->reserved >= num_bytes) {
5084                 block_rsv->reserved -= num_bytes;
5085                 if (block_rsv->reserved < block_rsv->size)
5086                         block_rsv->full = 0;
5087                 ret = 0;
5088         }
5089         spin_unlock(&block_rsv->lock);
5090         return ret;
5091 }
5092
5093 static void block_rsv_add_bytes(struct btrfs_block_rsv *block_rsv,
5094                                 u64 num_bytes, int update_size)
5095 {
5096         spin_lock(&block_rsv->lock);
5097         block_rsv->reserved += num_bytes;
5098         if (update_size)
5099                 block_rsv->size += num_bytes;
5100         else if (block_rsv->reserved >= block_rsv->size)
5101                 block_rsv->full = 1;
5102         spin_unlock(&block_rsv->lock);
5103 }
5104
5105 int btrfs_cond_migrate_bytes(struct btrfs_fs_info *fs_info,
5106                              struct btrfs_block_rsv *dest, u64 num_bytes,
5107                              int min_factor)
5108 {
5109         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
5110         u64 min_bytes;
5111
5112         if (global_rsv->space_info != dest->space_info)
5113                 return -ENOSPC;
5114
5115         spin_lock(&global_rsv->lock);
5116         min_bytes = div_factor(global_rsv->size, min_factor);
5117         if (global_rsv->reserved < min_bytes + num_bytes) {
5118                 spin_unlock(&global_rsv->lock);
5119                 return -ENOSPC;
5120         }
5121         global_rsv->reserved -= num_bytes;
5122         if (global_rsv->reserved < global_rsv->size)
5123                 global_rsv->full = 0;
5124         spin_unlock(&global_rsv->lock);
5125
5126         block_rsv_add_bytes(dest, num_bytes, 1);
5127         return 0;
5128 }
5129
5130 static void block_rsv_release_bytes(struct btrfs_fs_info *fs_info,
5131                                     struct btrfs_block_rsv *block_rsv,
5132                                     struct btrfs_block_rsv *dest, u64 num_bytes)
5133 {
5134         struct btrfs_space_info *space_info = block_rsv->space_info;
5135
5136         spin_lock(&block_rsv->lock);
5137         if (num_bytes == (u64)-1)
5138                 num_bytes = block_rsv->size;
5139         block_rsv->size -= num_bytes;
5140         if (block_rsv->reserved >= block_rsv->size) {
5141                 num_bytes = block_rsv->reserved - block_rsv->size;
5142                 block_rsv->reserved = block_rsv->size;
5143                 block_rsv->full = 1;
5144         } else {
5145                 num_bytes = 0;
5146         }
5147         spin_unlock(&block_rsv->lock);
5148
5149         if (num_bytes > 0) {
5150                 if (dest) {
5151                         spin_lock(&dest->lock);
5152                         if (!dest->full) {
5153                                 u64 bytes_to_add;
5154
5155                                 bytes_to_add = dest->size - dest->reserved;
5156                                 bytes_to_add = min(num_bytes, bytes_to_add);
5157                                 dest->reserved += bytes_to_add;
5158                                 if (dest->reserved >= dest->size)
5159                                         dest->full = 1;
5160                                 num_bytes -= bytes_to_add;
5161                         }
5162                         spin_unlock(&dest->lock);
5163                 }
5164                 if (num_bytes) {
5165                         spin_lock(&space_info->lock);
5166                         space_info->bytes_may_use -= num_bytes;
5167                         trace_btrfs_space_reservation(fs_info, "space_info",
5168                                         space_info->flags, num_bytes, 0);
5169                         spin_unlock(&space_info->lock);
5170                 }
5171         }
5172 }
5173
5174 static int block_rsv_migrate_bytes(struct btrfs_block_rsv *src,
5175                                    struct btrfs_block_rsv *dst, u64 num_bytes)
5176 {
5177         int ret;
5178
5179         ret = block_rsv_use_bytes(src, num_bytes);
5180         if (ret)
5181                 return ret;
5182
5183         block_rsv_add_bytes(dst, num_bytes, 1);
5184         return 0;
5185 }
5186
5187 void btrfs_init_block_rsv(struct btrfs_block_rsv *rsv, unsigned short type)
5188 {
5189         memset(rsv, 0, sizeof(*rsv));
5190         spin_lock_init(&rsv->lock);
5191         rsv->type = type;
5192 }
5193
5194 struct btrfs_block_rsv *btrfs_alloc_block_rsv(struct btrfs_root *root,
5195                                               unsigned short type)
5196 {
5197         struct btrfs_block_rsv *block_rsv;
5198         struct btrfs_fs_info *fs_info = root->fs_info;
5199
5200         block_rsv = kmalloc(sizeof(*block_rsv), GFP_NOFS);
5201         if (!block_rsv)
5202                 return NULL;
5203
5204         btrfs_init_block_rsv(block_rsv, type);
5205         block_rsv->space_info = __find_space_info(fs_info,
5206                                                   BTRFS_BLOCK_GROUP_METADATA);
5207         return block_rsv;
5208 }
5209
5210 void btrfs_free_block_rsv(struct btrfs_root *root,
5211                           struct btrfs_block_rsv *rsv)
5212 {
5213         if (!rsv)
5214                 return;
5215         btrfs_block_rsv_release(root, rsv, (u64)-1);
5216         kfree(rsv);
5217 }
5218
5219 void __btrfs_free_block_rsv(struct btrfs_block_rsv *rsv)
5220 {
5221         kfree(rsv);
5222 }
5223
5224 int btrfs_block_rsv_add(struct btrfs_root *root,
5225                         struct btrfs_block_rsv *block_rsv, u64 num_bytes,
5226                         enum btrfs_reserve_flush_enum flush)
5227 {
5228         int ret;
5229
5230         if (num_bytes == 0)
5231                 return 0;
5232
5233         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
5234         if (!ret) {
5235                 block_rsv_add_bytes(block_rsv, num_bytes, 1);
5236                 return 0;
5237         }
5238
5239         return ret;
5240 }
5241
5242 int btrfs_block_rsv_check(struct btrfs_root *root,
5243                           struct btrfs_block_rsv *block_rsv, int min_factor)
5244 {
5245         u64 num_bytes = 0;
5246         int ret = -ENOSPC;
5247
5248         if (!block_rsv)
5249                 return 0;
5250
5251         spin_lock(&block_rsv->lock);
5252         num_bytes = div_factor(block_rsv->size, min_factor);
5253         if (block_rsv->reserved >= num_bytes)
5254                 ret = 0;
5255         spin_unlock(&block_rsv->lock);
5256
5257         return ret;
5258 }
5259
5260 int btrfs_block_rsv_refill(struct btrfs_root *root,
5261                            struct btrfs_block_rsv *block_rsv, u64 min_reserved,
5262                            enum btrfs_reserve_flush_enum flush)
5263 {
5264         u64 num_bytes = 0;
5265         int ret = -ENOSPC;
5266
5267         if (!block_rsv)
5268                 return 0;
5269
5270         spin_lock(&block_rsv->lock);
5271         num_bytes = min_reserved;
5272         if (block_rsv->reserved >= num_bytes)
5273                 ret = 0;
5274         else
5275                 num_bytes -= block_rsv->reserved;
5276         spin_unlock(&block_rsv->lock);
5277
5278         if (!ret)
5279                 return 0;
5280
5281         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
5282         if (!ret) {
5283                 block_rsv_add_bytes(block_rsv, num_bytes, 0);
5284                 return 0;
5285         }
5286
5287         return ret;
5288 }
5289
5290 int btrfs_block_rsv_migrate(struct btrfs_block_rsv *src_rsv,
5291                             struct btrfs_block_rsv *dst_rsv,
5292                             u64 num_bytes)
5293 {
5294         return block_rsv_migrate_bytes(src_rsv, dst_rsv, num_bytes);
5295 }
5296
5297 void btrfs_block_rsv_release(struct btrfs_root *root,
5298                              struct btrfs_block_rsv *block_rsv,
5299                              u64 num_bytes)
5300 {
5301         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
5302         if (global_rsv == block_rsv ||
5303             block_rsv->space_info != global_rsv->space_info)
5304                 global_rsv = NULL;
5305         block_rsv_release_bytes(root->fs_info, block_rsv, global_rsv,
5306                                 num_bytes);
5307 }
5308
5309 /*
5310  * helper to calculate size of global block reservation.
5311  * the desired value is sum of space used by extent tree,
5312  * checksum tree and root tree
5313  */
5314 static u64 calc_global_metadata_size(struct btrfs_fs_info *fs_info)
5315 {
5316         struct btrfs_space_info *sinfo;
5317         u64 num_bytes;
5318         u64 meta_used;
5319         u64 data_used;
5320         int csum_size = btrfs_super_csum_size(fs_info->super_copy);
5321
5322         sinfo = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_DATA);
5323         spin_lock(&sinfo->lock);
5324         data_used = sinfo->bytes_used;
5325         spin_unlock(&sinfo->lock);
5326
5327         sinfo = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
5328         spin_lock(&sinfo->lock);
5329         if (sinfo->flags & BTRFS_BLOCK_GROUP_DATA)
5330                 data_used = 0;
5331         meta_used = sinfo->bytes_used;
5332         spin_unlock(&sinfo->lock);
5333
5334         num_bytes = (data_used >> fs_info->sb->s_blocksize_bits) *
5335                     csum_size * 2;
5336         num_bytes += div_u64(data_used + meta_used, 50);
5337
5338         if (num_bytes * 3 > meta_used)
5339                 num_bytes = div_u64(meta_used, 3);
5340
5341         return ALIGN(num_bytes, fs_info->extent_root->nodesize << 10);
5342 }
5343
5344 static void update_global_block_rsv(struct btrfs_fs_info *fs_info)
5345 {
5346         struct btrfs_block_rsv *block_rsv = &fs_info->global_block_rsv;
5347         struct btrfs_space_info *sinfo = block_rsv->space_info;
5348         u64 num_bytes;
5349
5350         num_bytes = calc_global_metadata_size(fs_info);
5351
5352         spin_lock(&sinfo->lock);
5353         spin_lock(&block_rsv->lock);
5354
5355         block_rsv->size = min_t(u64, num_bytes, 512 * 1024 * 1024);
5356
5357         num_bytes = sinfo->bytes_used + sinfo->bytes_pinned +
5358                     sinfo->bytes_reserved + sinfo->bytes_readonly +
5359                     sinfo->bytes_may_use;
5360
5361         if (sinfo->total_bytes > num_bytes) {
5362                 num_bytes = sinfo->total_bytes - num_bytes;
5363                 block_rsv->reserved += num_bytes;
5364                 sinfo->bytes_may_use += num_bytes;
5365                 trace_btrfs_space_reservation(fs_info, "space_info",
5366                                       sinfo->flags, num_bytes, 1);
5367         }
5368
5369         if (block_rsv->reserved >= block_rsv->size) {
5370                 num_bytes = block_rsv->reserved - block_rsv->size;
5371                 sinfo->bytes_may_use -= num_bytes;
5372                 trace_btrfs_space_reservation(fs_info, "space_info",
5373                                       sinfo->flags, num_bytes, 0);
5374                 block_rsv->reserved = block_rsv->size;
5375                 block_rsv->full = 1;
5376         }
5377
5378         spin_unlock(&block_rsv->lock);
5379         spin_unlock(&sinfo->lock);
5380 }
5381
5382 static void init_global_block_rsv(struct btrfs_fs_info *fs_info)
5383 {
5384         struct btrfs_space_info *space_info;
5385
5386         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
5387         fs_info->chunk_block_rsv.space_info = space_info;
5388
5389         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
5390         fs_info->global_block_rsv.space_info = space_info;
5391         fs_info->delalloc_block_rsv.space_info = space_info;
5392         fs_info->trans_block_rsv.space_info = space_info;
5393         fs_info->empty_block_rsv.space_info = space_info;
5394         fs_info->delayed_block_rsv.space_info = space_info;
5395
5396         fs_info->extent_root->block_rsv = &fs_info->global_block_rsv;
5397         fs_info->csum_root->block_rsv = &fs_info->global_block_rsv;
5398         fs_info->dev_root->block_rsv = &fs_info->global_block_rsv;
5399         fs_info->tree_root->block_rsv = &fs_info->global_block_rsv;
5400         if (fs_info->quota_root)
5401                 fs_info->quota_root->block_rsv = &fs_info->global_block_rsv;
5402         fs_info->chunk_root->block_rsv = &fs_info->chunk_block_rsv;
5403
5404         update_global_block_rsv(fs_info);
5405 }
5406
5407 static void release_global_block_rsv(struct btrfs_fs_info *fs_info)
5408 {
5409         block_rsv_release_bytes(fs_info, &fs_info->global_block_rsv, NULL,
5410                                 (u64)-1);
5411         WARN_ON(fs_info->delalloc_block_rsv.size > 0);
5412         WARN_ON(fs_info->delalloc_block_rsv.reserved > 0);
5413         WARN_ON(fs_info->trans_block_rsv.size > 0);
5414         WARN_ON(fs_info->trans_block_rsv.reserved > 0);
5415         WARN_ON(fs_info->chunk_block_rsv.size > 0);
5416         WARN_ON(fs_info->chunk_block_rsv.reserved > 0);
5417         WARN_ON(fs_info->delayed_block_rsv.size > 0);
5418         WARN_ON(fs_info->delayed_block_rsv.reserved > 0);
5419 }
5420
5421 void btrfs_trans_release_metadata(struct btrfs_trans_handle *trans,
5422                                   struct btrfs_root *root)
5423 {
5424         if (!trans->block_rsv)
5425                 return;
5426
5427         if (!trans->bytes_reserved)
5428                 return;
5429
5430         trace_btrfs_space_reservation(root->fs_info, "transaction",
5431                                       trans->transid, trans->bytes_reserved, 0);
5432         btrfs_block_rsv_release(root, trans->block_rsv, trans->bytes_reserved);
5433         trans->bytes_reserved = 0;
5434 }
5435
5436 /*
5437  * To be called after all the new block groups attached to the transaction
5438  * handle have been created (btrfs_create_pending_block_groups()).
5439  */
5440 void btrfs_trans_release_chunk_metadata(struct btrfs_trans_handle *trans)
5441 {
5442         struct btrfs_fs_info *fs_info = trans->root->fs_info;
5443
5444         if (!trans->chunk_bytes_reserved)
5445                 return;
5446
5447         WARN_ON_ONCE(!list_empty(&trans->new_bgs));
5448
5449         block_rsv_release_bytes(fs_info, &fs_info->chunk_block_rsv, NULL,
5450                                 trans->chunk_bytes_reserved);
5451         trans->chunk_bytes_reserved = 0;
5452 }
5453
5454 /* Can only return 0 or -ENOSPC */
5455 int btrfs_orphan_reserve_metadata(struct btrfs_trans_handle *trans,
5456                                   struct inode *inode)
5457 {
5458         struct btrfs_root *root = BTRFS_I(inode)->root;
5459         struct btrfs_block_rsv *src_rsv = get_block_rsv(trans, root);
5460         struct btrfs_block_rsv *dst_rsv = root->orphan_block_rsv;
5461
5462         /*
5463          * We need to hold space in order to delete our orphan item once we've
5464          * added it, so this takes the reservation so we can release it later
5465          * when we are truly done with the orphan item.
5466          */
5467         u64 num_bytes = btrfs_calc_trans_metadata_size(root, 1);
5468         trace_btrfs_space_reservation(root->fs_info, "orphan",
5469                                       btrfs_ino(inode), num_bytes, 1);
5470         return block_rsv_migrate_bytes(src_rsv, dst_rsv, num_bytes);
5471 }
5472
5473 void btrfs_orphan_release_metadata(struct inode *inode)
5474 {
5475         struct btrfs_root *root = BTRFS_I(inode)->root;
5476         u64 num_bytes = btrfs_calc_trans_metadata_size(root, 1);
5477         trace_btrfs_space_reservation(root->fs_info, "orphan",
5478                                       btrfs_ino(inode), num_bytes, 0);
5479         btrfs_block_rsv_release(root, root->orphan_block_rsv, num_bytes);
5480 }
5481
5482 /*
5483  * btrfs_subvolume_reserve_metadata() - reserve space for subvolume operation
5484  * root: the root of the parent directory
5485  * rsv: block reservation
5486  * items: the number of items that we need do reservation
5487  * qgroup_reserved: used to return the reserved size in qgroup
5488  *
5489  * This function is used to reserve the space for snapshot/subvolume
5490  * creation and deletion. Those operations are different with the
5491  * common file/directory operations, they change two fs/file trees
5492  * and root tree, the number of items that the qgroup reserves is
5493  * different with the free space reservation. So we can not use
5494  * the space reseravtion mechanism in start_transaction().
5495  */
5496 int btrfs_subvolume_reserve_metadata(struct btrfs_root *root,
5497                                      struct btrfs_block_rsv *rsv,
5498                                      int items,
5499                                      u64 *qgroup_reserved,
5500                                      bool use_global_rsv)
5501 {
5502         u64 num_bytes;
5503         int ret;
5504         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
5505
5506         if (root->fs_info->quota_enabled) {
5507                 /* One for parent inode, two for dir entries */
5508                 num_bytes = 3 * root->nodesize;
5509                 ret = btrfs_qgroup_reserve_meta(root, num_bytes);
5510                 if (ret)
5511                         return ret;
5512         } else {
5513                 num_bytes = 0;
5514         }
5515
5516         *qgroup_reserved = num_bytes;
5517
5518         num_bytes = btrfs_calc_trans_metadata_size(root, items);
5519         rsv->space_info = __find_space_info(root->fs_info,
5520                                             BTRFS_BLOCK_GROUP_METADATA);
5521         ret = btrfs_block_rsv_add(root, rsv, num_bytes,
5522                                   BTRFS_RESERVE_FLUSH_ALL);
5523
5524         if (ret == -ENOSPC && use_global_rsv)
5525                 ret = btrfs_block_rsv_migrate(global_rsv, rsv, num_bytes);
5526
5527         if (ret && *qgroup_reserved)
5528                 btrfs_qgroup_free_meta(root, *qgroup_reserved);
5529
5530         return ret;
5531 }
5532
5533 void btrfs_subvolume_release_metadata(struct btrfs_root *root,
5534                                       struct btrfs_block_rsv *rsv,
5535                                       u64 qgroup_reserved)
5536 {
5537         btrfs_block_rsv_release(root, rsv, (u64)-1);
5538 }
5539
5540 /**
5541  * drop_outstanding_extent - drop an outstanding extent
5542  * @inode: the inode we're dropping the extent for
5543  * @num_bytes: the number of bytes we're relaseing.
5544  *
5545  * This is called when we are freeing up an outstanding extent, either called
5546  * after an error or after an extent is written.  This will return the number of
5547  * reserved extents that need to be freed.  This must be called with
5548  * BTRFS_I(inode)->lock held.
5549  */
5550 static unsigned drop_outstanding_extent(struct inode *inode, u64 num_bytes)
5551 {
5552         unsigned drop_inode_space = 0;
5553         unsigned dropped_extents = 0;
5554         unsigned num_extents = 0;
5555
5556         num_extents = (unsigned)div64_u64(num_bytes +
5557                                           BTRFS_MAX_EXTENT_SIZE - 1,
5558                                           BTRFS_MAX_EXTENT_SIZE);
5559         ASSERT(num_extents);
5560         ASSERT(BTRFS_I(inode)->outstanding_extents >= num_extents);
5561         BTRFS_I(inode)->outstanding_extents -= num_extents;
5562
5563         if (BTRFS_I(inode)->outstanding_extents == 0 &&
5564             test_and_clear_bit(BTRFS_INODE_DELALLOC_META_RESERVED,
5565                                &BTRFS_I(inode)->runtime_flags))
5566                 drop_inode_space = 1;
5567
5568         /*
5569          * If we have more or the same amount of outsanding extents than we have
5570          * reserved then we need to leave the reserved extents count alone.
5571          */
5572         if (BTRFS_I(inode)->outstanding_extents >=
5573             BTRFS_I(inode)->reserved_extents)
5574                 return drop_inode_space;
5575
5576         dropped_extents = BTRFS_I(inode)->reserved_extents -
5577                 BTRFS_I(inode)->outstanding_extents;
5578         BTRFS_I(inode)->reserved_extents -= dropped_extents;
5579         return dropped_extents + drop_inode_space;
5580 }
5581
5582 /**
5583  * calc_csum_metadata_size - return the amount of metada space that must be
5584  *      reserved/free'd for the given bytes.
5585  * @inode: the inode we're manipulating
5586  * @num_bytes: the number of bytes in question
5587  * @reserve: 1 if we are reserving space, 0 if we are freeing space
5588  *
5589  * This adjusts the number of csum_bytes in the inode and then returns the
5590  * correct amount of metadata that must either be reserved or freed.  We
5591  * calculate how many checksums we can fit into one leaf and then divide the
5592  * number of bytes that will need to be checksumed by this value to figure out
5593  * how many checksums will be required.  If we are adding bytes then the number
5594  * may go up and we will return the number of additional bytes that must be
5595  * reserved.  If it is going down we will return the number of bytes that must
5596  * be freed.
5597  *
5598  * This must be called with BTRFS_I(inode)->lock held.
5599  */
5600 static u64 calc_csum_metadata_size(struct inode *inode, u64 num_bytes,
5601                                    int reserve)
5602 {
5603         struct btrfs_root *root = BTRFS_I(inode)->root;
5604         u64 old_csums, num_csums;
5605
5606         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM &&
5607             BTRFS_I(inode)->csum_bytes == 0)
5608                 return 0;
5609
5610         old_csums = btrfs_csum_bytes_to_leaves(root, BTRFS_I(inode)->csum_bytes);
5611         if (reserve)
5612                 BTRFS_I(inode)->csum_bytes += num_bytes;
5613         else
5614                 BTRFS_I(inode)->csum_bytes -= num_bytes;
5615         num_csums = btrfs_csum_bytes_to_leaves(root, BTRFS_I(inode)->csum_bytes);
5616
5617         /* No change, no need to reserve more */
5618         if (old_csums == num_csums)
5619                 return 0;
5620
5621         if (reserve)
5622                 return btrfs_calc_trans_metadata_size(root,
5623                                                       num_csums - old_csums);
5624
5625         return btrfs_calc_trans_metadata_size(root, old_csums - num_csums);
5626 }
5627
5628 int btrfs_delalloc_reserve_metadata(struct inode *inode, u64 num_bytes)
5629 {
5630         struct btrfs_root *root = BTRFS_I(inode)->root;
5631         struct btrfs_block_rsv *block_rsv = &root->fs_info->delalloc_block_rsv;
5632         u64 to_reserve = 0;
5633         u64 csum_bytes;
5634         unsigned nr_extents = 0;
5635         int extra_reserve = 0;
5636         enum btrfs_reserve_flush_enum flush = BTRFS_RESERVE_FLUSH_ALL;
5637         int ret = 0;
5638         bool delalloc_lock = true;
5639         u64 to_free = 0;
5640         unsigned dropped;
5641
5642         /* If we are a free space inode we need to not flush since we will be in
5643          * the middle of a transaction commit.  We also don't need the delalloc
5644          * mutex since we won't race with anybody.  We need this mostly to make
5645          * lockdep shut its filthy mouth.
5646          */
5647         if (btrfs_is_free_space_inode(inode)) {
5648                 flush = BTRFS_RESERVE_NO_FLUSH;
5649                 delalloc_lock = false;
5650         }
5651
5652         if (flush != BTRFS_RESERVE_NO_FLUSH &&
5653             btrfs_transaction_in_commit(root->fs_info))
5654                 schedule_timeout(1);
5655
5656         if (delalloc_lock)
5657                 mutex_lock(&BTRFS_I(inode)->delalloc_mutex);
5658
5659         num_bytes = ALIGN(num_bytes, root->sectorsize);
5660
5661         spin_lock(&BTRFS_I(inode)->lock);
5662         nr_extents = (unsigned)div64_u64(num_bytes +
5663                                          BTRFS_MAX_EXTENT_SIZE - 1,
5664                                          BTRFS_MAX_EXTENT_SIZE);
5665         BTRFS_I(inode)->outstanding_extents += nr_extents;
5666         nr_extents = 0;
5667
5668         if (BTRFS_I(inode)->outstanding_extents >
5669             BTRFS_I(inode)->reserved_extents)
5670                 nr_extents = BTRFS_I(inode)->outstanding_extents -
5671                         BTRFS_I(inode)->reserved_extents;
5672
5673         /*
5674          * Add an item to reserve for updating the inode when we complete the
5675          * delalloc io.
5676          */
5677         if (!test_bit(BTRFS_INODE_DELALLOC_META_RESERVED,
5678                       &BTRFS_I(inode)->runtime_flags)) {
5679                 nr_extents++;
5680                 extra_reserve = 1;
5681         }
5682
5683         to_reserve = btrfs_calc_trans_metadata_size(root, nr_extents);
5684         to_reserve += calc_csum_metadata_size(inode, num_bytes, 1);
5685         csum_bytes = BTRFS_I(inode)->csum_bytes;
5686         spin_unlock(&BTRFS_I(inode)->lock);
5687
5688         if (root->fs_info->quota_enabled) {
5689                 ret = btrfs_qgroup_reserve_meta(root,
5690                                 nr_extents * root->nodesize);
5691                 if (ret)
5692                         goto out_fail;
5693         }
5694
5695         ret = reserve_metadata_bytes(root, block_rsv, to_reserve, flush);
5696         if (unlikely(ret)) {
5697                 btrfs_qgroup_free_meta(root, nr_extents * root->nodesize);
5698                 goto out_fail;
5699         }
5700
5701         spin_lock(&BTRFS_I(inode)->lock);
5702         if (extra_reserve) {
5703                 set_bit(BTRFS_INODE_DELALLOC_META_RESERVED,
5704                         &BTRFS_I(inode)->runtime_flags);
5705                 nr_extents--;
5706         }
5707         BTRFS_I(inode)->reserved_extents += nr_extents;
5708         spin_unlock(&BTRFS_I(inode)->lock);
5709
5710         if (delalloc_lock)
5711                 mutex_unlock(&BTRFS_I(inode)->delalloc_mutex);
5712
5713         if (to_reserve)
5714                 trace_btrfs_space_reservation(root->fs_info, "delalloc",
5715                                               btrfs_ino(inode), to_reserve, 1);
5716         block_rsv_add_bytes(block_rsv, to_reserve, 1);
5717
5718         return 0;
5719
5720 out_fail:
5721         spin_lock(&BTRFS_I(inode)->lock);
5722         dropped = drop_outstanding_extent(inode, num_bytes);
5723         /*
5724          * If the inodes csum_bytes is the same as the original
5725          * csum_bytes then we know we haven't raced with any free()ers
5726          * so we can just reduce our inodes csum bytes and carry on.
5727          */
5728         if (BTRFS_I(inode)->csum_bytes == csum_bytes) {
5729                 calc_csum_metadata_size(inode, num_bytes, 0);
5730         } else {
5731                 u64 orig_csum_bytes = BTRFS_I(inode)->csum_bytes;
5732                 u64 bytes;
5733
5734                 /*
5735                  * This is tricky, but first we need to figure out how much we
5736                  * free'd from any free-ers that occured during this
5737                  * reservation, so we reset ->csum_bytes to the csum_bytes
5738                  * before we dropped our lock, and then call the free for the
5739                  * number of bytes that were freed while we were trying our
5740                  * reservation.
5741                  */
5742                 bytes = csum_bytes - BTRFS_I(inode)->csum_bytes;
5743                 BTRFS_I(inode)->csum_bytes = csum_bytes;
5744                 to_free = calc_csum_metadata_size(inode, bytes, 0);
5745
5746
5747                 /*
5748                  * Now we need to see how much we would have freed had we not
5749                  * been making this reservation and our ->csum_bytes were not
5750                  * artificially inflated.
5751                  */
5752                 BTRFS_I(inode)->csum_bytes = csum_bytes - num_bytes;
5753                 bytes = csum_bytes - orig_csum_bytes;
5754                 bytes = calc_csum_metadata_size(inode, bytes, 0);
5755
5756                 /*
5757                  * Now reset ->csum_bytes to what it should be.  If bytes is
5758                  * more than to_free then we would have free'd more space had we
5759                  * not had an artificially high ->csum_bytes, so we need to free
5760                  * the remainder.  If bytes is the same or less then we don't
5761                  * need to do anything, the other free-ers did the correct
5762                  * thing.
5763                  */
5764                 BTRFS_I(inode)->csum_bytes = orig_csum_bytes - num_bytes;
5765                 if (bytes > to_free)
5766                         to_free = bytes - to_free;
5767                 else
5768                         to_free = 0;
5769         }
5770         spin_unlock(&BTRFS_I(inode)->lock);
5771         if (dropped)
5772                 to_free += btrfs_calc_trans_metadata_size(root, dropped);
5773
5774         if (to_free) {
5775                 btrfs_block_rsv_release(root, block_rsv, to_free);
5776                 trace_btrfs_space_reservation(root->fs_info, "delalloc",
5777                                               btrfs_ino(inode), to_free, 0);
5778         }
5779         if (delalloc_lock)
5780                 mutex_unlock(&BTRFS_I(inode)->delalloc_mutex);
5781         return ret;
5782 }
5783
5784 /**
5785  * btrfs_delalloc_release_metadata - release a metadata reservation for an inode
5786  * @inode: the inode to release the reservation for
5787  * @num_bytes: the number of bytes we're releasing
5788  *
5789  * This will release the metadata reservation for an inode.  This can be called
5790  * once we complete IO for a given set of bytes to release their metadata
5791  * reservations.
5792  */
5793 void btrfs_delalloc_release_metadata(struct inode *inode, u64 num_bytes)
5794 {
5795         struct btrfs_root *root = BTRFS_I(inode)->root;
5796         u64 to_free = 0;
5797         unsigned dropped;
5798
5799         num_bytes = ALIGN(num_bytes, root->sectorsize);
5800         spin_lock(&BTRFS_I(inode)->lock);
5801         dropped = drop_outstanding_extent(inode, num_bytes);
5802
5803         if (num_bytes)
5804                 to_free = calc_csum_metadata_size(inode, num_bytes, 0);
5805         spin_unlock(&BTRFS_I(inode)->lock);
5806         if (dropped > 0)
5807                 to_free += btrfs_calc_trans_metadata_size(root, dropped);
5808
5809         if (btrfs_test_is_dummy_root(root))
5810                 return;
5811
5812         trace_btrfs_space_reservation(root->fs_info, "delalloc",
5813                                       btrfs_ino(inode), to_free, 0);
5814
5815         btrfs_block_rsv_release(root, &root->fs_info->delalloc_block_rsv,
5816                                 to_free);
5817 }
5818
5819 /**
5820  * btrfs_delalloc_reserve_space - reserve data and metadata space for
5821  * delalloc
5822  * @inode: inode we're writing to
5823  * @start: start range we are writing to
5824  * @len: how long the range we are writing to
5825  *
5826  * TODO: This function will finally replace old btrfs_delalloc_reserve_space()
5827  *
5828  * This will do the following things
5829  *
5830  * o reserve space in data space info for num bytes
5831  *   and reserve precious corresponding qgroup space
5832  *   (Done in check_data_free_space)
5833  *
5834  * o reserve space for metadata space, based on the number of outstanding
5835  *   extents and how much csums will be needed
5836  *   also reserve metadata space in a per root over-reserve method.
5837  * o add to the inodes->delalloc_bytes
5838  * o add it to the fs_info's delalloc inodes list.
5839  *   (Above 3 all done in delalloc_reserve_metadata)
5840  *
5841  * Return 0 for success
5842  * Return <0 for error(-ENOSPC or -EQUOT)
5843  */
5844 int btrfs_delalloc_reserve_space(struct inode *inode, u64 start, u64 len)
5845 {
5846         int ret;
5847
5848         ret = btrfs_check_data_free_space(inode, start, len);
5849         if (ret < 0)
5850                 return ret;
5851         ret = btrfs_delalloc_reserve_metadata(inode, len);
5852         if (ret < 0)
5853                 btrfs_free_reserved_data_space(inode, start, len);
5854         return ret;
5855 }
5856
5857 /**
5858  * btrfs_delalloc_release_space - release data and metadata space for delalloc
5859  * @inode: inode we're releasing space for
5860  * @start: start position of the space already reserved
5861  * @len: the len of the space already reserved
5862  *
5863  * This must be matched with a call to btrfs_delalloc_reserve_space.  This is
5864  * called in the case that we don't need the metadata AND data reservations
5865  * anymore.  So if there is an error or we insert an inline extent.
5866  *
5867  * This function will release the metadata space that was not used and will
5868  * decrement ->delalloc_bytes and remove it from the fs_info delalloc_inodes
5869  * list if there are no delalloc bytes left.
5870  * Also it will handle the qgroup reserved space.
5871  */
5872 void btrfs_delalloc_release_space(struct inode *inode, u64 start, u64 len)
5873 {
5874         btrfs_delalloc_release_metadata(inode, len);
5875         btrfs_free_reserved_data_space(inode, start, len);
5876 }
5877
5878 static int update_block_group(struct btrfs_trans_handle *trans,
5879                               struct btrfs_root *root, u64 bytenr,
5880                               u64 num_bytes, int alloc)
5881 {
5882         struct btrfs_block_group_cache *cache = NULL;
5883         struct btrfs_fs_info *info = root->fs_info;
5884         u64 total = num_bytes;
5885         u64 old_val;
5886         u64 byte_in_group;
5887         int factor;
5888
5889         /* block accounting for super block */
5890         spin_lock(&info->delalloc_root_lock);
5891         old_val = btrfs_super_bytes_used(info->super_copy);
5892         if (alloc)
5893                 old_val += num_bytes;
5894         else
5895                 old_val -= num_bytes;
5896         btrfs_set_super_bytes_used(info->super_copy, old_val);
5897         spin_unlock(&info->delalloc_root_lock);
5898
5899         while (total) {
5900                 cache = btrfs_lookup_block_group(info, bytenr);
5901                 if (!cache)
5902                         return -ENOENT;
5903                 if (cache->flags & (BTRFS_BLOCK_GROUP_DUP |
5904                                     BTRFS_BLOCK_GROUP_RAID1 |
5905                                     BTRFS_BLOCK_GROUP_RAID10))
5906                         factor = 2;
5907                 else
5908                         factor = 1;
5909                 /*
5910                  * If this block group has free space cache written out, we
5911                  * need to make sure to load it if we are removing space.  This
5912                  * is because we need the unpinning stage to actually add the
5913                  * space back to the block group, otherwise we will leak space.
5914                  */
5915                 if (!alloc && cache->cached == BTRFS_CACHE_NO)
5916                         cache_block_group(cache, 1);
5917
5918                 byte_in_group = bytenr - cache->key.objectid;
5919                 WARN_ON(byte_in_group > cache->key.offset);
5920
5921                 spin_lock(&cache->space_info->lock);
5922                 spin_lock(&cache->lock);
5923
5924                 if (btrfs_test_opt(root, SPACE_CACHE) &&
5925                     cache->disk_cache_state < BTRFS_DC_CLEAR)
5926                         cache->disk_cache_state = BTRFS_DC_CLEAR;
5927
5928                 old_val = btrfs_block_group_used(&cache->item);
5929                 num_bytes = min(total, cache->key.offset - byte_in_group);
5930                 if (alloc) {
5931                         old_val += num_bytes;
5932                         btrfs_set_block_group_used(&cache->item, old_val);
5933                         cache->reserved -= num_bytes;
5934                         cache->space_info->bytes_reserved -= num_bytes;
5935                         cache->space_info->bytes_used += num_bytes;
5936                         cache->space_info->disk_used += num_bytes * factor;
5937                         spin_unlock(&cache->lock);
5938                         spin_unlock(&cache->space_info->lock);
5939                 } else {
5940                         old_val -= num_bytes;
5941                         btrfs_set_block_group_used(&cache->item, old_val);
5942                         cache->pinned += num_bytes;
5943                         cache->space_info->bytes_pinned += num_bytes;
5944                         cache->space_info->bytes_used -= num_bytes;
5945                         cache->space_info->disk_used -= num_bytes * factor;
5946                         spin_unlock(&cache->lock);
5947                         spin_unlock(&cache->space_info->lock);
5948
5949                         set_extent_dirty(info->pinned_extents,
5950                                          bytenr, bytenr + num_bytes - 1,
5951                                          GFP_NOFS | __GFP_NOFAIL);
5952                 }
5953
5954                 spin_lock(&trans->transaction->dirty_bgs_lock);
5955                 if (list_empty(&cache->dirty_list)) {
5956                         list_add_tail(&cache->dirty_list,
5957                                       &trans->transaction->dirty_bgs);
5958                                 trans->transaction->num_dirty_bgs++;
5959                         btrfs_get_block_group(cache);
5960                 }
5961                 spin_unlock(&trans->transaction->dirty_bgs_lock);
5962
5963                 /*
5964                  * No longer have used bytes in this block group, queue it for
5965                  * deletion. We do this after adding the block group to the
5966                  * dirty list to avoid races between cleaner kthread and space
5967                  * cache writeout.
5968                  */
5969                 if (!alloc && old_val == 0) {
5970                         spin_lock(&info->unused_bgs_lock);
5971                         if (list_empty(&cache->bg_list)) {
5972                                 btrfs_get_block_group(cache);
5973                                 list_add_tail(&cache->bg_list,
5974                                               &info->unused_bgs);
5975                         }
5976                         spin_unlock(&info->unused_bgs_lock);
5977                 }
5978
5979                 btrfs_put_block_group(cache);
5980                 total -= num_bytes;
5981                 bytenr += num_bytes;
5982         }
5983         return 0;
5984 }
5985
5986 static u64 first_logical_byte(struct btrfs_root *root, u64 search_start)
5987 {
5988         struct btrfs_block_group_cache *cache;
5989         u64 bytenr;
5990
5991         spin_lock(&root->fs_info->block_group_cache_lock);
5992         bytenr = root->fs_info->first_logical_byte;
5993         spin_unlock(&root->fs_info->block_group_cache_lock);
5994
5995         if (bytenr < (u64)-1)
5996                 return bytenr;
5997
5998         cache = btrfs_lookup_first_block_group(root->fs_info, search_start);
5999         if (!cache)
6000                 return 0;
6001
6002         bytenr = cache->key.objectid;
6003         btrfs_put_block_group(cache);
6004
6005         return bytenr;
6006 }
6007
6008 static int pin_down_extent(struct btrfs_root *root,
6009                            struct btrfs_block_group_cache *cache,
6010                            u64 bytenr, u64 num_bytes, int reserved)
6011 {
6012         spin_lock(&cache->space_info->lock);
6013         spin_lock(&cache->lock);
6014         cache->pinned += num_bytes;
6015         cache->space_info->bytes_pinned += num_bytes;
6016         if (reserved) {
6017                 cache->reserved -= num_bytes;
6018                 cache->space_info->bytes_reserved -= num_bytes;
6019         }
6020         spin_unlock(&cache->lock);
6021         spin_unlock(&cache->space_info->lock);
6022
6023         set_extent_dirty(root->fs_info->pinned_extents, bytenr,
6024                          bytenr + num_bytes - 1, GFP_NOFS | __GFP_NOFAIL);
6025         if (reserved)
6026                 trace_btrfs_reserved_extent_free(root, bytenr, num_bytes);
6027         return 0;
6028 }
6029
6030 /*
6031  * this function must be called within transaction
6032  */
6033 int btrfs_pin_extent(struct btrfs_root *root,
6034                      u64 bytenr, u64 num_bytes, int reserved)
6035 {
6036         struct btrfs_block_group_cache *cache;
6037
6038         cache = btrfs_lookup_block_group(root->fs_info, bytenr);
6039         BUG_ON(!cache); /* Logic error */
6040
6041         pin_down_extent(root, cache, bytenr, num_bytes, reserved);
6042
6043         btrfs_put_block_group(cache);
6044         return 0;
6045 }
6046
6047 /*
6048  * this function must be called within transaction
6049  */
6050 int btrfs_pin_extent_for_log_replay(struct btrfs_root *root,
6051                                     u64 bytenr, u64 num_bytes)
6052 {
6053         struct btrfs_block_group_cache *cache;
6054         int ret;
6055
6056         cache = btrfs_lookup_block_group(root->fs_info, bytenr);
6057         if (!cache)
6058                 return -EINVAL;
6059
6060         /*
6061          * pull in the free space cache (if any) so that our pin
6062          * removes the free space from the cache.  We have load_only set
6063          * to one because the slow code to read in the free extents does check
6064          * the pinned extents.
6065          */
6066         cache_block_group(cache, 1);
6067
6068         pin_down_extent(root, cache, bytenr, num_bytes, 0);
6069
6070         /* remove us from the free space cache (if we're there at all) */
6071         ret = btrfs_remove_free_space(cache, bytenr, num_bytes);
6072         btrfs_put_block_group(cache);
6073         return ret;
6074 }
6075
6076 static int __exclude_logged_extent(struct btrfs_root *root, u64 start, u64 num_bytes)
6077 {
6078         int ret;
6079         struct btrfs_block_group_cache *block_group;
6080         struct btrfs_caching_control *caching_ctl;
6081
6082         block_group = btrfs_lookup_block_group(root->fs_info, start);
6083         if (!block_group)
6084                 return -EINVAL;
6085
6086         cache_block_group(block_group, 0);
6087         caching_ctl = get_caching_control(block_group);
6088
6089         if (!caching_ctl) {
6090                 /* Logic error */
6091                 BUG_ON(!block_group_cache_done(block_group));
6092                 ret = btrfs_remove_free_space(block_group, start, num_bytes);
6093         } else {
6094                 mutex_lock(&caching_ctl->mutex);
6095
6096                 if (start >= caching_ctl->progress) {
6097                         ret = add_excluded_extent(root, start, num_bytes);
6098                 } else if (start + num_bytes <= caching_ctl->progress) {
6099                         ret = btrfs_remove_free_space(block_group,
6100                                                       start, num_bytes);
6101                 } else {
6102                         num_bytes = caching_ctl->progress - start;
6103                         ret = btrfs_remove_free_space(block_group,
6104                                                       start, num_bytes);
6105                         if (ret)
6106                                 goto out_lock;
6107
6108                         num_bytes = (start + num_bytes) -
6109                                 caching_ctl->progress;
6110                         start = caching_ctl->progress;
6111                         ret = add_excluded_extent(root, start, num_bytes);
6112                 }
6113 out_lock:
6114                 mutex_unlock(&caching_ctl->mutex);
6115                 put_caching_control(caching_ctl);
6116         }
6117         btrfs_put_block_group(block_group);
6118         return ret;
6119 }
6120
6121 int btrfs_exclude_logged_extents(struct btrfs_root *log,
6122                                  struct extent_buffer *eb)
6123 {
6124         struct btrfs_file_extent_item *item;
6125         struct btrfs_key key;
6126         int found_type;
6127         int i;
6128
6129         if (!btrfs_fs_incompat(log->fs_info, MIXED_GROUPS))
6130                 return 0;
6131
6132         for (i = 0; i < btrfs_header_nritems(eb); i++) {
6133                 btrfs_item_key_to_cpu(eb, &key, i);
6134                 if (key.type != BTRFS_EXTENT_DATA_KEY)
6135                         continue;
6136                 item = btrfs_item_ptr(eb, i, struct btrfs_file_extent_item);
6137                 found_type = btrfs_file_extent_type(eb, item);
6138                 if (found_type == BTRFS_FILE_EXTENT_INLINE)
6139                         continue;
6140                 if (btrfs_file_extent_disk_bytenr(eb, item) == 0)
6141                         continue;
6142                 key.objectid = btrfs_file_extent_disk_bytenr(eb, item);
6143                 key.offset = btrfs_file_extent_disk_num_bytes(eb, item);
6144                 __exclude_logged_extent(log, key.objectid, key.offset);
6145         }
6146
6147         return 0;
6148 }
6149
6150 /**
6151  * btrfs_update_reserved_bytes - update the block_group and space info counters
6152  * @cache:      The cache we are manipulating
6153  * @num_bytes:  The number of bytes in question
6154  * @reserve:    One of the reservation enums
6155  * @delalloc:   The blocks are allocated for the delalloc write
6156  *
6157  * This is called by the allocator when it reserves space, or by somebody who is
6158  * freeing space that was never actually used on disk.  For example if you
6159  * reserve some space for a new leaf in transaction A and before transaction A
6160  * commits you free that leaf, you call this with reserve set to 0 in order to
6161  * clear the reservation.
6162  *
6163  * Metadata reservations should be called with RESERVE_ALLOC so we do the proper
6164  * ENOSPC accounting.  For data we handle the reservation through clearing the
6165  * delalloc bits in the io_tree.  We have to do this since we could end up
6166  * allocating less disk space for the amount of data we have reserved in the
6167  * case of compression.
6168  *
6169  * If this is a reservation and the block group has become read only we cannot
6170  * make the reservation and return -EAGAIN, otherwise this function always
6171  * succeeds.
6172  */
6173 static int btrfs_update_reserved_bytes(struct btrfs_block_group_cache *cache,
6174                                        u64 num_bytes, int reserve, int delalloc)
6175 {
6176         struct btrfs_space_info *space_info = cache->space_info;
6177         int ret = 0;
6178
6179         spin_lock(&space_info->lock);
6180         spin_lock(&cache->lock);
6181         if (reserve != RESERVE_FREE) {
6182                 if (cache->ro) {
6183                         ret = -EAGAIN;
6184                 } else {
6185                         cache->reserved += num_bytes;
6186                         space_info->bytes_reserved += num_bytes;
6187                         if (reserve == RESERVE_ALLOC) {
6188                                 trace_btrfs_space_reservation(cache->fs_info,
6189                                                 "space_info", space_info->flags,
6190                                                 num_bytes, 0);
6191                                 space_info->bytes_may_use -= num_bytes;
6192                         }
6193
6194                         if (delalloc)
6195                                 cache->delalloc_bytes += num_bytes;
6196                 }
6197         } else {
6198                 if (cache->ro)
6199                         space_info->bytes_readonly += num_bytes;
6200                 cache->reserved -= num_bytes;
6201                 space_info->bytes_reserved -= num_bytes;
6202
6203                 if (delalloc)
6204                         cache->delalloc_bytes -= num_bytes;
6205         }
6206         spin_unlock(&cache->lock);
6207         spin_unlock(&space_info->lock);
6208         return ret;
6209 }
6210
6211 void btrfs_prepare_extent_commit(struct btrfs_trans_handle *trans,
6212                                 struct btrfs_root *root)
6213 {
6214         struct btrfs_fs_info *fs_info = root->fs_info;
6215         struct btrfs_caching_control *next;
6216         struct btrfs_caching_control *caching_ctl;
6217         struct btrfs_block_group_cache *cache;
6218
6219         down_write(&fs_info->commit_root_sem);
6220
6221         list_for_each_entry_safe(caching_ctl, next,
6222                                  &fs_info->caching_block_groups, list) {
6223                 cache = caching_ctl->block_group;
6224                 if (block_group_cache_done(cache)) {
6225                         cache->last_byte_to_unpin = (u64)-1;
6226                         list_del_init(&caching_ctl->list);
6227                         put_caching_control(caching_ctl);
6228                 } else {
6229                         cache->last_byte_to_unpin = caching_ctl->progress;
6230                 }
6231         }
6232
6233         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
6234                 fs_info->pinned_extents = &fs_info->freed_extents[1];
6235         else
6236                 fs_info->pinned_extents = &fs_info->freed_extents[0];
6237
6238         up_write(&fs_info->commit_root_sem);
6239
6240         update_global_block_rsv(fs_info);
6241 }
6242
6243 /*
6244  * Returns the free cluster for the given space info and sets empty_cluster to
6245  * what it should be based on the mount options.
6246  */
6247 static struct btrfs_free_cluster *
6248 fetch_cluster_info(struct btrfs_root *root, struct btrfs_space_info *space_info,
6249                    u64 *empty_cluster)
6250 {
6251         struct btrfs_free_cluster *ret = NULL;
6252         bool ssd = btrfs_test_opt(root, SSD);
6253
6254         *empty_cluster = 0;
6255         if (btrfs_mixed_space_info(space_info))
6256                 return ret;
6257
6258         if (ssd)
6259                 *empty_cluster = 2 * 1024 * 1024;
6260         if (space_info->flags & BTRFS_BLOCK_GROUP_METADATA) {
6261                 ret = &root->fs_info->meta_alloc_cluster;
6262                 if (!ssd)
6263                         *empty_cluster = 64 * 1024;
6264         } else if ((space_info->flags & BTRFS_BLOCK_GROUP_DATA) && ssd) {
6265                 ret = &root->fs_info->data_alloc_cluster;
6266         }
6267
6268         return ret;
6269 }
6270
6271 static int unpin_extent_range(struct btrfs_root *root, u64 start, u64 end,
6272                               const bool return_free_space)
6273 {
6274         struct btrfs_fs_info *fs_info = root->fs_info;
6275         struct btrfs_block_group_cache *cache = NULL;
6276         struct btrfs_space_info *space_info;
6277         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
6278         struct btrfs_free_cluster *cluster = NULL;
6279         u64 len;
6280         u64 total_unpinned = 0;
6281         u64 empty_cluster = 0;
6282         bool readonly;
6283
6284         while (start <= end) {
6285                 readonly = false;
6286                 if (!cache ||
6287                     start >= cache->key.objectid + cache->key.offset) {
6288                         if (cache)
6289                                 btrfs_put_block_group(cache);
6290                         total_unpinned = 0;
6291                         cache = btrfs_lookup_block_group(fs_info, start);
6292                         BUG_ON(!cache); /* Logic error */
6293
6294                         cluster = fetch_cluster_info(root,
6295                                                      cache->space_info,
6296                                                      &empty_cluster);
6297                         empty_cluster <<= 1;
6298                 }
6299
6300                 len = cache->key.objectid + cache->key.offset - start;
6301                 len = min(len, end + 1 - start);
6302
6303                 if (start < cache->last_byte_to_unpin) {
6304                         len = min(len, cache->last_byte_to_unpin - start);
6305                         if (return_free_space)
6306                                 btrfs_add_free_space(cache, start, len);
6307                 }
6308
6309                 start += len;
6310                 total_unpinned += len;
6311                 space_info = cache->space_info;
6312
6313                 /*
6314                  * If this space cluster has been marked as fragmented and we've
6315                  * unpinned enough in this block group to potentially allow a
6316                  * cluster to be created inside of it go ahead and clear the
6317                  * fragmented check.
6318                  */
6319                 if (cluster && cluster->fragmented &&
6320                     total_unpinned > empty_cluster) {
6321                         spin_lock(&cluster->lock);
6322                         cluster->fragmented = 0;
6323                         spin_unlock(&cluster->lock);
6324                 }
6325
6326                 spin_lock(&space_info->lock);
6327                 spin_lock(&cache->lock);
6328                 cache->pinned -= len;
6329                 space_info->bytes_pinned -= len;
6330                 space_info->max_extent_size = 0;
6331                 percpu_counter_add(&space_info->total_bytes_pinned, -len);
6332                 if (cache->ro) {
6333                         space_info->bytes_readonly += len;
6334                         readonly = true;
6335                 }
6336                 spin_unlock(&cache->lock);
6337                 if (!readonly && global_rsv->space_info == space_info) {
6338                         spin_lock(&global_rsv->lock);
6339                         if (!global_rsv->full) {
6340                                 len = min(len, global_rsv->size -
6341                                           global_rsv->reserved);
6342                                 global_rsv->reserved += len;
6343                                 space_info->bytes_may_use += len;
6344                                 if (global_rsv->reserved >= global_rsv->size)
6345                                         global_rsv->full = 1;
6346                         }
6347                         spin_unlock(&global_rsv->lock);
6348                 }
6349                 spin_unlock(&space_info->lock);
6350         }
6351
6352         if (cache)
6353                 btrfs_put_block_group(cache);
6354         return 0;
6355 }
6356
6357 int btrfs_finish_extent_commit(struct btrfs_trans_handle *trans,
6358                                struct btrfs_root *root)
6359 {
6360         struct btrfs_fs_info *fs_info = root->fs_info;
6361         struct btrfs_block_group_cache *block_group, *tmp;
6362         struct list_head *deleted_bgs;
6363         struct extent_io_tree *unpin;
6364         u64 start;
6365         u64 end;
6366         int ret;
6367
6368         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
6369                 unpin = &fs_info->freed_extents[1];
6370         else
6371                 unpin = &fs_info->freed_extents[0];
6372
6373         while (!trans->aborted) {
6374                 mutex_lock(&fs_info->unused_bg_unpin_mutex);
6375                 ret = find_first_extent_bit(unpin, 0, &start, &end,
6376                                             EXTENT_DIRTY, NULL);
6377                 if (ret) {
6378                         mutex_unlock(&fs_info->unused_bg_unpin_mutex);
6379                         break;
6380                 }
6381
6382                 if (btrfs_test_opt(root, DISCARD))
6383                         ret = btrfs_discard_extent(root, start,
6384                                                    end + 1 - start, NULL);
6385
6386                 clear_extent_dirty(unpin, start, end, GFP_NOFS);
6387                 unpin_extent_range(root, start, end, true);
6388                 mutex_unlock(&fs_info->unused_bg_unpin_mutex);
6389                 cond_resched();
6390         }
6391
6392         /*
6393          * Transaction is finished.  We don't need the lock anymore.  We
6394          * do need to clean up the block groups in case of a transaction
6395          * abort.
6396          */
6397         deleted_bgs = &trans->transaction->deleted_bgs;
6398         list_for_each_entry_safe(block_group, tmp, deleted_bgs, bg_list) {
6399                 u64 trimmed = 0;
6400
6401                 ret = -EROFS;
6402                 if (!trans->aborted)
6403                         ret = btrfs_discard_extent(root,
6404                                                    block_group->key.objectid,
6405                                                    block_group->key.offset,
6406                                                    &trimmed);
6407
6408                 list_del_init(&block_group->bg_list);
6409                 btrfs_put_block_group_trimming(block_group);
6410                 btrfs_put_block_group(block_group);
6411
6412                 if (ret) {
6413                         const char *errstr = btrfs_decode_error(ret);
6414                         btrfs_warn(fs_info,
6415                                    "Discard failed while removing blockgroup: errno=%d %s\n",
6416                                    ret, errstr);
6417                 }
6418         }
6419
6420         return 0;
6421 }
6422
6423 static void add_pinned_bytes(struct btrfs_fs_info *fs_info, u64 num_bytes,
6424                              u64 owner, u64 root_objectid)
6425 {
6426         struct btrfs_space_info *space_info;
6427         u64 flags;
6428
6429         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
6430                 if (root_objectid == BTRFS_CHUNK_TREE_OBJECTID)
6431                         flags = BTRFS_BLOCK_GROUP_SYSTEM;
6432                 else
6433                         flags = BTRFS_BLOCK_GROUP_METADATA;
6434         } else {
6435                 flags = BTRFS_BLOCK_GROUP_DATA;
6436         }
6437
6438         space_info = __find_space_info(fs_info, flags);
6439         BUG_ON(!space_info); /* Logic bug */
6440         percpu_counter_add(&space_info->total_bytes_pinned, num_bytes);
6441 }
6442
6443
6444 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
6445                                 struct btrfs_root *root,
6446                                 struct btrfs_delayed_ref_node *node, u64 parent,
6447                                 u64 root_objectid, u64 owner_objectid,
6448                                 u64 owner_offset, int refs_to_drop,
6449                                 struct btrfs_delayed_extent_op *extent_op)
6450 {
6451         struct btrfs_key key;
6452         struct btrfs_path *path;
6453         struct btrfs_fs_info *info = root->fs_info;
6454         struct btrfs_root *extent_root = info->extent_root;
6455         struct extent_buffer *leaf;
6456         struct btrfs_extent_item *ei;
6457         struct btrfs_extent_inline_ref *iref;
6458         int ret;
6459         int is_data;
6460         int extent_slot = 0;
6461         int found_extent = 0;
6462         int num_to_del = 1;
6463         u32 item_size;
6464         u64 refs;
6465         u64 bytenr = node->bytenr;
6466         u64 num_bytes = node->num_bytes;
6467         int last_ref = 0;
6468         bool skinny_metadata = btrfs_fs_incompat(root->fs_info,
6469                                                  SKINNY_METADATA);
6470
6471         path = btrfs_alloc_path();
6472         if (!path)
6473                 return -ENOMEM;
6474
6475         path->reada = 1;
6476         path->leave_spinning = 1;
6477
6478         is_data = owner_objectid >= BTRFS_FIRST_FREE_OBJECTID;
6479         BUG_ON(!is_data && refs_to_drop != 1);
6480
6481         if (is_data)
6482                 skinny_metadata = 0;
6483
6484         ret = lookup_extent_backref(trans, extent_root, path, &iref,
6485                                     bytenr, num_bytes, parent,
6486                                     root_objectid, owner_objectid,
6487                                     owner_offset);
6488         if (ret == 0) {
6489                 extent_slot = path->slots[0];
6490                 while (extent_slot >= 0) {
6491                         btrfs_item_key_to_cpu(path->nodes[0], &key,
6492                                               extent_slot);
6493                         if (key.objectid != bytenr)
6494                                 break;
6495                         if (key.type == BTRFS_EXTENT_ITEM_KEY &&
6496                             key.offset == num_bytes) {
6497                                 found_extent = 1;
6498                                 break;
6499                         }
6500                         if (key.type == BTRFS_METADATA_ITEM_KEY &&
6501                             key.offset == owner_objectid) {
6502                                 found_extent = 1;
6503                                 break;
6504                         }
6505                         if (path->slots[0] - extent_slot > 5)
6506                                 break;
6507                         extent_slot--;
6508                 }
6509 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
6510                 item_size = btrfs_item_size_nr(path->nodes[0], extent_slot);
6511                 if (found_extent && item_size < sizeof(*ei))
6512                         found_extent = 0;
6513 #endif
6514                 if (!found_extent) {
6515                         BUG_ON(iref);
6516                         ret = remove_extent_backref(trans, extent_root, path,
6517                                                     NULL, refs_to_drop,
6518                                                     is_data, &last_ref);
6519                         if (ret) {
6520                                 btrfs_abort_transaction(trans, extent_root, ret);
6521                                 goto out;
6522                         }
6523                         btrfs_release_path(path);
6524                         path->leave_spinning = 1;
6525
6526                         key.objectid = bytenr;
6527                         key.type = BTRFS_EXTENT_ITEM_KEY;
6528                         key.offset = num_bytes;
6529
6530                         if (!is_data && skinny_metadata) {
6531                                 key.type = BTRFS_METADATA_ITEM_KEY;
6532                                 key.offset = owner_objectid;
6533                         }
6534
6535                         ret = btrfs_search_slot(trans, extent_root,
6536                                                 &key, path, -1, 1);
6537                         if (ret > 0 && skinny_metadata && path->slots[0]) {
6538                                 /*
6539                                  * Couldn't find our skinny metadata item,
6540                                  * see if we have ye olde extent item.
6541                                  */
6542                                 path->slots[0]--;
6543                                 btrfs_item_key_to_cpu(path->nodes[0], &key,
6544                                                       path->slots[0]);
6545                                 if (key.objectid == bytenr &&
6546                                     key.type == BTRFS_EXTENT_ITEM_KEY &&
6547                                     key.offset == num_bytes)
6548                                         ret = 0;
6549                         }
6550
6551                         if (ret > 0 && skinny_metadata) {
6552                                 skinny_metadata = false;
6553                                 key.objectid = bytenr;
6554                                 key.type = BTRFS_EXTENT_ITEM_KEY;
6555                                 key.offset = num_bytes;
6556                                 btrfs_release_path(path);
6557                                 ret = btrfs_search_slot(trans, extent_root,
6558                                                         &key, path, -1, 1);
6559                         }
6560
6561                         if (ret) {
6562                                 btrfs_err(info, "umm, got %d back from search, was looking for %llu",
6563                                         ret, bytenr);
6564                                 if (ret > 0)
6565                                         btrfs_print_leaf(extent_root,
6566                                                          path->nodes[0]);
6567                         }
6568                         if (ret < 0) {
6569                                 btrfs_abort_transaction(trans, extent_root, ret);
6570                                 goto out;
6571                         }
6572                         extent_slot = path->slots[0];
6573                 }
6574         } else if (WARN_ON(ret == -ENOENT)) {
6575                 btrfs_print_leaf(extent_root, path->nodes[0]);
6576                 btrfs_err(info,
6577                         "unable to find ref byte nr %llu parent %llu root %llu  owner %llu offset %llu",
6578                         bytenr, parent, root_objectid, owner_objectid,
6579                         owner_offset);
6580                 btrfs_abort_transaction(trans, extent_root, ret);
6581                 goto out;
6582         } else {
6583                 btrfs_abort_transaction(trans, extent_root, ret);
6584                 goto out;
6585         }
6586
6587         leaf = path->nodes[0];
6588         item_size = btrfs_item_size_nr(leaf, extent_slot);
6589 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
6590         if (item_size < sizeof(*ei)) {
6591                 BUG_ON(found_extent || extent_slot != path->slots[0]);
6592                 ret = convert_extent_item_v0(trans, extent_root, path,
6593                                              owner_objectid, 0);
6594                 if (ret < 0) {
6595                         btrfs_abort_transaction(trans, extent_root, ret);
6596                         goto out;
6597                 }
6598
6599                 btrfs_release_path(path);
6600                 path->leave_spinning = 1;
6601
6602                 key.objectid = bytenr;
6603                 key.type = BTRFS_EXTENT_ITEM_KEY;
6604                 key.offset = num_bytes;
6605
6606                 ret = btrfs_search_slot(trans, extent_root, &key, path,
6607                                         -1, 1);
6608                 if (ret) {
6609                         btrfs_err(info, "umm, got %d back from search, was looking for %llu",
6610                                 ret, bytenr);
6611                         btrfs_print_leaf(extent_root, path->nodes[0]);
6612                 }
6613                 if (ret < 0) {
6614                         btrfs_abort_transaction(trans, extent_root, ret);
6615                         goto out;
6616                 }
6617
6618                 extent_slot = path->slots[0];
6619                 leaf = path->nodes[0];
6620                 item_size = btrfs_item_size_nr(leaf, extent_slot);
6621         }
6622 #endif
6623         BUG_ON(item_size < sizeof(*ei));
6624         ei = btrfs_item_ptr(leaf, extent_slot,
6625                             struct btrfs_extent_item);
6626         if (owner_objectid < BTRFS_FIRST_FREE_OBJECTID &&
6627             key.type == BTRFS_EXTENT_ITEM_KEY) {
6628                 struct btrfs_tree_block_info *bi;
6629                 BUG_ON(item_size < sizeof(*ei) + sizeof(*bi));
6630                 bi = (struct btrfs_tree_block_info *)(ei + 1);
6631                 WARN_ON(owner_objectid != btrfs_tree_block_level(leaf, bi));
6632         }
6633
6634         refs = btrfs_extent_refs(leaf, ei);
6635         if (refs < refs_to_drop) {
6636                 btrfs_err(info, "trying to drop %d refs but we only have %Lu "
6637                           "for bytenr %Lu", refs_to_drop, refs, bytenr);
6638                 ret = -EINVAL;
6639                 btrfs_abort_transaction(trans, extent_root, ret);
6640                 goto out;
6641         }
6642         refs -= refs_to_drop;
6643
6644         if (refs > 0) {
6645                 if (extent_op)
6646                         __run_delayed_extent_op(extent_op, leaf, ei);
6647                 /*
6648                  * In the case of inline back ref, reference count will
6649                  * be updated by remove_extent_backref
6650                  */
6651                 if (iref) {
6652                         BUG_ON(!found_extent);
6653                 } else {
6654                         btrfs_set_extent_refs(leaf, ei, refs);
6655                         btrfs_mark_buffer_dirty(leaf);
6656                 }
6657                 if (found_extent) {
6658                         ret = remove_extent_backref(trans, extent_root, path,
6659                                                     iref, refs_to_drop,
6660                                                     is_data, &last_ref);
6661                         if (ret) {
6662                                 btrfs_abort_transaction(trans, extent_root, ret);
6663                                 goto out;
6664                         }
6665                 }
6666                 add_pinned_bytes(root->fs_info, -num_bytes, owner_objectid,
6667                                  root_objectid);
6668         } else {
6669                 if (found_extent) {
6670                         BUG_ON(is_data && refs_to_drop !=
6671                                extent_data_ref_count(path, iref));
6672                         if (iref) {
6673                                 BUG_ON(path->slots[0] != extent_slot);
6674                         } else {
6675                                 BUG_ON(path->slots[0] != extent_slot + 1);
6676                                 path->slots[0] = extent_slot;
6677                                 num_to_del = 2;
6678                         }
6679                 }
6680
6681                 last_ref = 1;
6682                 ret = btrfs_del_items(trans, extent_root, path, path->slots[0],
6683                                       num_to_del);
6684                 if (ret) {
6685                         btrfs_abort_transaction(trans, extent_root, ret);
6686                         goto out;
6687                 }
6688                 btrfs_release_path(path);
6689
6690                 if (is_data) {
6691                         ret = btrfs_del_csums(trans, root, bytenr, num_bytes);
6692                         if (ret) {
6693                                 btrfs_abort_transaction(trans, extent_root, ret);
6694                                 goto out;
6695                         }
6696                 }
6697
6698                 ret = add_to_free_space_tree(trans, root->fs_info, bytenr,
6699                                              num_bytes);
6700                 if (ret) {
6701                         btrfs_abort_transaction(trans, extent_root, ret);
6702                         goto out;
6703                 }
6704
6705                 ret = update_block_group(trans, root, bytenr, num_bytes, 0);
6706                 if (ret) {
6707                         btrfs_abort_transaction(trans, extent_root, ret);
6708                         goto out;
6709                 }
6710         }
6711         btrfs_release_path(path);
6712
6713 out:
6714         btrfs_free_path(path);
6715         return ret;
6716 }
6717
6718 /*
6719  * when we free an block, it is possible (and likely) that we free the last
6720  * delayed ref for that extent as well.  This searches the delayed ref tree for
6721  * a given extent, and if there are no other delayed refs to be processed, it
6722  * removes it from the tree.
6723  */
6724 static noinline int check_ref_cleanup(struct btrfs_trans_handle *trans,
6725                                       struct btrfs_root *root, u64 bytenr)
6726 {
6727         struct btrfs_delayed_ref_head *head;
6728         struct btrfs_delayed_ref_root *delayed_refs;
6729         int ret = 0;
6730
6731         delayed_refs = &trans->transaction->delayed_refs;
6732         spin_lock(&delayed_refs->lock);
6733         head = btrfs_find_delayed_ref_head(trans, bytenr);
6734         if (!head)
6735                 goto out_delayed_unlock;
6736
6737         spin_lock(&head->lock);
6738         if (!list_empty(&head->ref_list))
6739                 goto out;
6740
6741         if (head->extent_op) {
6742                 if (!head->must_insert_reserved)
6743                         goto out;
6744                 btrfs_free_delayed_extent_op(head->extent_op);
6745                 head->extent_op = NULL;
6746         }
6747
6748         /*
6749          * waiting for the lock here would deadlock.  If someone else has it
6750          * locked they are already in the process of dropping it anyway
6751          */
6752         if (!mutex_trylock(&head->mutex))
6753                 goto out;
6754
6755         /*
6756          * at this point we have a head with no other entries.  Go
6757          * ahead and process it.
6758          */
6759         head->node.in_tree = 0;
6760         rb_erase(&head->href_node, &delayed_refs->href_root);
6761
6762         atomic_dec(&delayed_refs->num_entries);
6763
6764         /*
6765          * we don't take a ref on the node because we're removing it from the
6766          * tree, so we just steal the ref the tree was holding.
6767          */
6768         delayed_refs->num_heads--;
6769         if (head->processing == 0)
6770                 delayed_refs->num_heads_ready--;
6771         head->processing = 0;
6772         spin_unlock(&head->lock);
6773         spin_unlock(&delayed_refs->lock);
6774
6775         BUG_ON(head->extent_op);
6776         if (head->must_insert_reserved)
6777                 ret = 1;
6778
6779         mutex_unlock(&head->mutex);
6780         btrfs_put_delayed_ref(&head->node);
6781         return ret;
6782 out:
6783         spin_unlock(&head->lock);
6784
6785 out_delayed_unlock:
6786         spin_unlock(&delayed_refs->lock);
6787         return 0;
6788 }
6789
6790 void btrfs_free_tree_block(struct btrfs_trans_handle *trans,
6791                            struct btrfs_root *root,
6792                            struct extent_buffer *buf,
6793                            u64 parent, int last_ref)
6794 {
6795         int pin = 1;
6796         int ret;
6797
6798         if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
6799                 ret = btrfs_add_delayed_tree_ref(root->fs_info, trans,
6800                                         buf->start, buf->len,
6801                                         parent, root->root_key.objectid,
6802                                         btrfs_header_level(buf),
6803                                         BTRFS_DROP_DELAYED_REF, NULL);
6804                 BUG_ON(ret); /* -ENOMEM */
6805         }
6806
6807         if (!last_ref)
6808                 return;
6809
6810         if (btrfs_header_generation(buf) == trans->transid) {
6811                 struct btrfs_block_group_cache *cache;
6812
6813                 if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
6814                         ret = check_ref_cleanup(trans, root, buf->start);
6815                         if (!ret)
6816                                 goto out;
6817                 }
6818
6819                 cache = btrfs_lookup_block_group(root->fs_info, buf->start);
6820
6821                 if (btrfs_header_flag(buf, BTRFS_HEADER_FLAG_WRITTEN)) {
6822                         pin_down_extent(root, cache, buf->start, buf->len, 1);
6823                         btrfs_put_block_group(cache);
6824                         goto out;
6825                 }
6826
6827                 WARN_ON(test_bit(EXTENT_BUFFER_DIRTY, &buf->bflags));
6828
6829                 btrfs_add_free_space(cache, buf->start, buf->len);
6830                 btrfs_update_reserved_bytes(cache, buf->len, RESERVE_FREE, 0);
6831                 btrfs_put_block_group(cache);
6832                 trace_btrfs_reserved_extent_free(root, buf->start, buf->len);
6833                 pin = 0;
6834         }
6835 out:
6836         if (pin)
6837                 add_pinned_bytes(root->fs_info, buf->len,
6838                                  btrfs_header_level(buf),
6839                                  root->root_key.objectid);
6840
6841         /*
6842          * Deleting the buffer, clear the corrupt flag since it doesn't matter
6843          * anymore.
6844          */
6845         clear_bit(EXTENT_BUFFER_CORRUPT, &buf->bflags);
6846 }
6847
6848 /* Can return -ENOMEM */
6849 int btrfs_free_extent(struct btrfs_trans_handle *trans, struct btrfs_root *root,
6850                       u64 bytenr, u64 num_bytes, u64 parent, u64 root_objectid,
6851                       u64 owner, u64 offset)
6852 {
6853         int ret;
6854         struct btrfs_fs_info *fs_info = root->fs_info;
6855
6856         if (btrfs_test_is_dummy_root(root))
6857                 return 0;
6858
6859         add_pinned_bytes(root->fs_info, num_bytes, owner, root_objectid);
6860
6861         /*
6862          * tree log blocks never actually go into the extent allocation
6863          * tree, just update pinning info and exit early.
6864          */
6865         if (root_objectid == BTRFS_TREE_LOG_OBJECTID) {
6866                 WARN_ON(owner >= BTRFS_FIRST_FREE_OBJECTID);
6867                 /* unlocks the pinned mutex */
6868                 btrfs_pin_extent(root, bytenr, num_bytes, 1);
6869                 ret = 0;
6870         } else if (owner < BTRFS_FIRST_FREE_OBJECTID) {
6871                 ret = btrfs_add_delayed_tree_ref(fs_info, trans, bytenr,
6872                                         num_bytes,
6873                                         parent, root_objectid, (int)owner,
6874                                         BTRFS_DROP_DELAYED_REF, NULL);
6875         } else {
6876                 ret = btrfs_add_delayed_data_ref(fs_info, trans, bytenr,
6877                                                 num_bytes,
6878                                                 parent, root_objectid, owner,
6879                                                 offset, 0,
6880                                                 BTRFS_DROP_DELAYED_REF, NULL);
6881         }
6882         return ret;
6883 }
6884
6885 /*
6886  * when we wait for progress in the block group caching, its because
6887  * our allocation attempt failed at least once.  So, we must sleep
6888  * and let some progress happen before we try again.
6889  *
6890  * This function will sleep at least once waiting for new free space to
6891  * show up, and then it will check the block group free space numbers
6892  * for our min num_bytes.  Another option is to have it go ahead
6893  * and look in the rbtree for a free extent of a given size, but this
6894  * is a good start.
6895  *
6896  * Callers of this must check if cache->cached == BTRFS_CACHE_ERROR before using
6897  * any of the information in this block group.
6898  */
6899 static noinline void
6900 wait_block_group_cache_progress(struct btrfs_block_group_cache *cache,
6901                                 u64 num_bytes)
6902 {
6903         struct btrfs_caching_control *caching_ctl;
6904
6905         caching_ctl = get_caching_control(cache);
6906         if (!caching_ctl)
6907                 return;
6908
6909         wait_event(caching_ctl->wait, block_group_cache_done(cache) ||
6910                    (cache->free_space_ctl->free_space >= num_bytes));
6911
6912         put_caching_control(caching_ctl);
6913 }
6914
6915 static noinline int
6916 wait_block_group_cache_done(struct btrfs_block_group_cache *cache)
6917 {
6918         struct btrfs_caching_control *caching_ctl;
6919         int ret = 0;
6920
6921         caching_ctl = get_caching_control(cache);
6922         if (!caching_ctl)
6923                 return (cache->cached == BTRFS_CACHE_ERROR) ? -EIO : 0;
6924
6925         wait_event(caching_ctl->wait, block_group_cache_done(cache));
6926         if (cache->cached == BTRFS_CACHE_ERROR)
6927                 ret = -EIO;
6928         put_caching_control(caching_ctl);
6929         return ret;
6930 }
6931
6932 int __get_raid_index(u64 flags)
6933 {
6934         if (flags & BTRFS_BLOCK_GROUP_RAID10)
6935                 return BTRFS_RAID_RAID10;
6936         else if (flags & BTRFS_BLOCK_GROUP_RAID1)
6937                 return BTRFS_RAID_RAID1;
6938         else if (flags & BTRFS_BLOCK_GROUP_DUP)
6939                 return BTRFS_RAID_DUP;
6940         else if (flags & BTRFS_BLOCK_GROUP_RAID0)
6941                 return BTRFS_RAID_RAID0;
6942         else if (flags & BTRFS_BLOCK_GROUP_RAID5)
6943                 return BTRFS_RAID_RAID5;
6944         else if (flags & BTRFS_BLOCK_GROUP_RAID6)
6945                 return BTRFS_RAID_RAID6;
6946
6947         return BTRFS_RAID_SINGLE; /* BTRFS_BLOCK_GROUP_SINGLE */
6948 }
6949
6950 int get_block_group_index(struct btrfs_block_group_cache *cache)
6951 {
6952         return __get_raid_index(cache->flags);
6953 }
6954
6955 static const char *btrfs_raid_type_names[BTRFS_NR_RAID_TYPES] = {
6956         [BTRFS_RAID_RAID10]     = "raid10",
6957         [BTRFS_RAID_RAID1]      = "raid1",
6958         [BTRFS_RAID_DUP]        = "dup",
6959         [BTRFS_RAID_RAID0]      = "raid0",
6960         [BTRFS_RAID_SINGLE]     = "single",
6961         [BTRFS_RAID_RAID5]      = "raid5",
6962         [BTRFS_RAID_RAID6]      = "raid6",
6963 };
6964
6965 static const char *get_raid_name(enum btrfs_raid_types type)
6966 {
6967         if (type >= BTRFS_NR_RAID_TYPES)
6968                 return NULL;
6969
6970         return btrfs_raid_type_names[type];
6971 }
6972
6973 enum btrfs_loop_type {
6974         LOOP_CACHING_NOWAIT = 0,
6975         LOOP_CACHING_WAIT = 1,
6976         LOOP_ALLOC_CHUNK = 2,
6977         LOOP_NO_EMPTY_SIZE = 3,
6978 };
6979
6980 static inline void
6981 btrfs_lock_block_group(struct btrfs_block_group_cache *cache,
6982                        int delalloc)
6983 {
6984         if (delalloc)
6985                 down_read(&cache->data_rwsem);
6986 }
6987
6988 static inline void
6989 btrfs_grab_block_group(struct btrfs_block_group_cache *cache,
6990                        int delalloc)
6991 {
6992         btrfs_get_block_group(cache);
6993         if (delalloc)
6994                 down_read(&cache->data_rwsem);
6995 }
6996
6997 static struct btrfs_block_group_cache *
6998 btrfs_lock_cluster(struct btrfs_block_group_cache *block_group,
6999                    struct btrfs_free_cluster *cluster,
7000                    int delalloc)
7001 {
7002         struct btrfs_block_group_cache *used_bg;
7003         bool locked = false;
7004 again:
7005         spin_lock(&cluster->refill_lock);
7006         if (locked) {
7007                 if (used_bg == cluster->block_group)
7008                         return used_bg;
7009
7010                 up_read(&used_bg->data_rwsem);
7011                 btrfs_put_block_group(used_bg);
7012         }
7013
7014         used_bg = cluster->block_group;
7015         if (!used_bg)
7016                 return NULL;
7017
7018         if (used_bg == block_group)
7019                 return used_bg;
7020
7021         btrfs_get_block_group(used_bg);
7022
7023         if (!delalloc)
7024                 return used_bg;
7025
7026         if (down_read_trylock(&used_bg->data_rwsem))
7027                 return used_bg;
7028
7029         spin_unlock(&cluster->refill_lock);
7030         down_read(&used_bg->data_rwsem);
7031         locked = true;
7032         goto again;
7033 }
7034
7035 static inline void
7036 btrfs_release_block_group(struct btrfs_block_group_cache *cache,
7037                          int delalloc)
7038 {
7039         if (delalloc)
7040                 up_read(&cache->data_rwsem);
7041         btrfs_put_block_group(cache);
7042 }
7043
7044 /*
7045  * walks the btree of allocated extents and find a hole of a given size.
7046  * The key ins is changed to record the hole:
7047  * ins->objectid == start position
7048  * ins->flags = BTRFS_EXTENT_ITEM_KEY
7049  * ins->offset == the size of the hole.
7050  * Any available blocks before search_start are skipped.
7051  *
7052  * If there is no suitable free space, we will record the max size of
7053  * the free space extent currently.
7054  */
7055 static noinline int find_free_extent(struct btrfs_root *orig_root,
7056                                      u64 num_bytes, u64 empty_size,
7057                                      u64 hint_byte, struct btrfs_key *ins,
7058                                      u64 flags, int delalloc)
7059 {
7060         int ret = 0;
7061         struct btrfs_root *root = orig_root->fs_info->extent_root;
7062         struct btrfs_free_cluster *last_ptr = NULL;
7063         struct btrfs_block_group_cache *block_group = NULL;
7064         u64 search_start = 0;
7065         u64 max_extent_size = 0;
7066         u64 empty_cluster = 0;
7067         struct btrfs_space_info *space_info;
7068         int loop = 0;
7069         int index = __get_raid_index(flags);
7070         int alloc_type = (flags & BTRFS_BLOCK_GROUP_DATA) ?
7071                 RESERVE_ALLOC_NO_ACCOUNT : RESERVE_ALLOC;
7072         bool failed_cluster_refill = false;
7073         bool failed_alloc = false;
7074         bool use_cluster = true;
7075         bool have_caching_bg = false;
7076         bool orig_have_caching_bg = false;
7077         bool full_search = false;
7078
7079         WARN_ON(num_bytes < root->sectorsize);
7080         ins->type = BTRFS_EXTENT_ITEM_KEY;
7081         ins->objectid = 0;
7082         ins->offset = 0;
7083
7084         trace_find_free_extent(orig_root, num_bytes, empty_size, flags);
7085
7086         space_info = __find_space_info(root->fs_info, flags);
7087         if (!space_info) {
7088                 btrfs_err(root->fs_info, "No space info for %llu", flags);
7089                 return -ENOSPC;
7090         }
7091
7092         /*
7093          * If our free space is heavily fragmented we may not be able to make
7094          * big contiguous allocations, so instead of doing the expensive search
7095          * for free space, simply return ENOSPC with our max_extent_size so we
7096          * can go ahead and search for a more manageable chunk.
7097          *
7098          * If our max_extent_size is large enough for our allocation simply
7099          * disable clustering since we will likely not be able to find enough
7100          * space to create a cluster and induce latency trying.
7101          */
7102         if (unlikely(space_info->max_extent_size)) {
7103                 spin_lock(&space_info->lock);
7104                 if (space_info->max_extent_size &&
7105                     num_bytes > space_info->max_extent_size) {
7106                         ins->offset = space_info->max_extent_size;
7107                         spin_unlock(&space_info->lock);
7108                         return -ENOSPC;
7109                 } else if (space_info->max_extent_size) {
7110                         use_cluster = false;
7111                 }
7112                 spin_unlock(&space_info->lock);
7113         }
7114
7115         last_ptr = fetch_cluster_info(orig_root, space_info, &empty_cluster);
7116         if (last_ptr) {
7117                 spin_lock(&last_ptr->lock);
7118                 if (last_ptr->block_group)
7119                         hint_byte = last_ptr->window_start;
7120                 if (last_ptr->fragmented) {
7121                         /*
7122                          * We still set window_start so we can keep track of the
7123                          * last place we found an allocation to try and save
7124                          * some time.
7125                          */
7126                         hint_byte = last_ptr->window_start;
7127                         use_cluster = false;
7128                 }
7129                 spin_unlock(&last_ptr->lock);
7130         }
7131
7132         search_start = max(search_start, first_logical_byte(root, 0));
7133         search_start = max(search_start, hint_byte);
7134         if (search_start == hint_byte) {
7135                 block_group = btrfs_lookup_block_group(root->fs_info,
7136                                                        search_start);
7137                 /*
7138                  * we don't want to use the block group if it doesn't match our
7139                  * allocation bits, or if its not cached.
7140                  *
7141                  * However if we are re-searching with an ideal block group
7142                  * picked out then we don't care that the block group is cached.
7143                  */
7144                 if (block_group && block_group_bits(block_group, flags) &&
7145                     block_group->cached != BTRFS_CACHE_NO) {
7146                         down_read(&space_info->groups_sem);
7147                         if (list_empty(&block_group->list) ||
7148                             block_group->ro) {
7149                                 /*
7150                                  * someone is removing this block group,
7151                                  * we can't jump into the have_block_group
7152                                  * target because our list pointers are not
7153                                  * valid
7154                                  */
7155                                 btrfs_put_block_group(block_group);
7156                                 up_read(&space_info->groups_sem);
7157                         } else {
7158                                 index = get_block_group_index(block_group);
7159                                 btrfs_lock_block_group(block_group, delalloc);
7160                                 goto have_block_group;
7161                         }
7162                 } else if (block_group) {
7163                         btrfs_put_block_group(block_group);
7164                 }
7165         }
7166 search:
7167         have_caching_bg = false;
7168         if (index == 0 || index == __get_raid_index(flags))
7169                 full_search = true;
7170         down_read(&space_info->groups_sem);
7171         list_for_each_entry(block_group, &space_info->block_groups[index],
7172                             list) {
7173                 u64 offset;
7174                 int cached;
7175
7176                 btrfs_grab_block_group(block_group, delalloc);
7177                 search_start = block_group->key.objectid;
7178
7179                 /*
7180                  * this can happen if we end up cycling through all the
7181                  * raid types, but we want to make sure we only allocate
7182                  * for the proper type.
7183                  */
7184                 if (!block_group_bits(block_group, flags)) {
7185                     u64 extra = BTRFS_BLOCK_GROUP_DUP |
7186                                 BTRFS_BLOCK_GROUP_RAID1 |
7187                                 BTRFS_BLOCK_GROUP_RAID5 |
7188                                 BTRFS_BLOCK_GROUP_RAID6 |
7189                                 BTRFS_BLOCK_GROUP_RAID10;
7190
7191                         /*
7192                          * if they asked for extra copies and this block group
7193                          * doesn't provide them, bail.  This does allow us to
7194                          * fill raid0 from raid1.
7195                          */
7196                         if ((flags & extra) && !(block_group->flags & extra))
7197                                 goto loop;
7198                 }
7199
7200 have_block_group:
7201                 cached = block_group_cache_done(block_group);
7202                 if (unlikely(!cached)) {
7203                         have_caching_bg = true;
7204                         ret = cache_block_group(block_group, 0);
7205                         BUG_ON(ret < 0);
7206                         ret = 0;
7207                 }
7208
7209                 if (unlikely(block_group->cached == BTRFS_CACHE_ERROR))
7210                         goto loop;
7211                 if (unlikely(block_group->ro))
7212                         goto loop;
7213
7214                 /*
7215                  * Ok we want to try and use the cluster allocator, so
7216                  * lets look there
7217                  */
7218                 if (last_ptr && use_cluster) {
7219                         struct btrfs_block_group_cache *used_block_group;
7220                         unsigned long aligned_cluster;
7221                         /*
7222                          * the refill lock keeps out other
7223                          * people trying to start a new cluster
7224                          */
7225                         used_block_group = btrfs_lock_cluster(block_group,
7226                                                               last_ptr,
7227                                                               delalloc);
7228                         if (!used_block_group)
7229                                 goto refill_cluster;
7230
7231                         if (used_block_group != block_group &&
7232                             (used_block_group->ro ||
7233                              !block_group_bits(used_block_group, flags)))
7234                                 goto release_cluster;
7235
7236                         offset = btrfs_alloc_from_cluster(used_block_group,
7237                                                 last_ptr,
7238                                                 num_bytes,
7239                                                 used_block_group->key.objectid,
7240                                                 &max_extent_size);
7241                         if (offset) {
7242                                 /* we have a block, we're done */
7243                                 spin_unlock(&last_ptr->refill_lock);
7244                                 trace_btrfs_reserve_extent_cluster(root,
7245                                                 used_block_group,
7246                                                 search_start, num_bytes);
7247                                 if (used_block_group != block_group) {
7248                                         btrfs_release_block_group(block_group,
7249                                                                   delalloc);
7250                                         block_group = used_block_group;
7251                                 }
7252                                 goto checks;
7253                         }
7254
7255                         WARN_ON(last_ptr->block_group != used_block_group);
7256 release_cluster:
7257                         /* If we are on LOOP_NO_EMPTY_SIZE, we can't
7258                          * set up a new clusters, so lets just skip it
7259                          * and let the allocator find whatever block
7260                          * it can find.  If we reach this point, we
7261                          * will have tried the cluster allocator
7262                          * plenty of times and not have found
7263                          * anything, so we are likely way too
7264                          * fragmented for the clustering stuff to find
7265                          * anything.
7266                          *
7267                          * However, if the cluster is taken from the
7268                          * current block group, release the cluster
7269                          * first, so that we stand a better chance of
7270                          * succeeding in the unclustered
7271                          * allocation.  */
7272                         if (loop >= LOOP_NO_EMPTY_SIZE &&
7273                             used_block_group != block_group) {
7274                                 spin_unlock(&last_ptr->refill_lock);
7275                                 btrfs_release_block_group(used_block_group,
7276                                                           delalloc);
7277                                 goto unclustered_alloc;
7278                         }
7279
7280                         /*
7281                          * this cluster didn't work out, free it and
7282                          * start over
7283                          */
7284                         btrfs_return_cluster_to_free_space(NULL, last_ptr);
7285
7286                         if (used_block_group != block_group)
7287                                 btrfs_release_block_group(used_block_group,
7288                                                           delalloc);
7289 refill_cluster:
7290                         if (loop >= LOOP_NO_EMPTY_SIZE) {
7291                                 spin_unlock(&last_ptr->refill_lock);
7292                                 goto unclustered_alloc;
7293                         }
7294
7295                         aligned_cluster = max_t(unsigned long,
7296                                                 empty_cluster + empty_size,
7297                                               block_group->full_stripe_len);
7298
7299                         /* allocate a cluster in this block group */
7300                         ret = btrfs_find_space_cluster(root, block_group,
7301                                                        last_ptr, search_start,
7302                                                        num_bytes,
7303                                                        aligned_cluster);
7304                         if (ret == 0) {
7305                                 /*
7306                                  * now pull our allocation out of this
7307                                  * cluster
7308                                  */
7309                                 offset = btrfs_alloc_from_cluster(block_group,
7310                                                         last_ptr,
7311                                                         num_bytes,
7312                                                         search_start,
7313                                                         &max_extent_size);
7314                                 if (offset) {
7315                                         /* we found one, proceed */
7316                                         spin_unlock(&last_ptr->refill_lock);
7317                                         trace_btrfs_reserve_extent_cluster(root,
7318                                                 block_group, search_start,
7319                                                 num_bytes);
7320                                         goto checks;
7321                                 }
7322                         } else if (!cached && loop > LOOP_CACHING_NOWAIT
7323                                    && !failed_cluster_refill) {
7324                                 spin_unlock(&last_ptr->refill_lock);
7325
7326                                 failed_cluster_refill = true;
7327                                 wait_block_group_cache_progress(block_group,
7328                                        num_bytes + empty_cluster + empty_size);
7329                                 goto have_block_group;
7330                         }
7331
7332                         /*
7333                          * at this point we either didn't find a cluster
7334                          * or we weren't able to allocate a block from our
7335                          * cluster.  Free the cluster we've been trying
7336                          * to use, and go to the next block group
7337                          */
7338                         btrfs_return_cluster_to_free_space(NULL, last_ptr);
7339                         spin_unlock(&last_ptr->refill_lock);
7340                         goto loop;
7341                 }
7342
7343 unclustered_alloc:
7344                 /*
7345                  * We are doing an unclustered alloc, set the fragmented flag so
7346                  * we don't bother trying to setup a cluster again until we get
7347                  * more space.
7348                  */
7349                 if (unlikely(last_ptr)) {
7350                         spin_lock(&last_ptr->lock);
7351                         last_ptr->fragmented = 1;
7352                         spin_unlock(&last_ptr->lock);
7353                 }
7354                 spin_lock(&block_group->free_space_ctl->tree_lock);
7355                 if (cached &&
7356                     block_group->free_space_ctl->free_space <
7357                     num_bytes + empty_cluster + empty_size) {
7358                         if (block_group->free_space_ctl->free_space >
7359                             max_extent_size)
7360                                 max_extent_size =
7361                                         block_group->free_space_ctl->free_space;
7362                         spin_unlock(&block_group->free_space_ctl->tree_lock);
7363                         goto loop;
7364                 }
7365                 spin_unlock(&block_group->free_space_ctl->tree_lock);
7366
7367                 offset = btrfs_find_space_for_alloc(block_group, search_start,
7368                                                     num_bytes, empty_size,
7369                                                     &max_extent_size);
7370                 /*
7371                  * If we didn't find a chunk, and we haven't failed on this
7372                  * block group before, and this block group is in the middle of
7373                  * caching and we are ok with waiting, then go ahead and wait
7374                  * for progress to be made, and set failed_alloc to true.
7375                  *
7376                  * If failed_alloc is true then we've already waited on this
7377                  * block group once and should move on to the next block group.
7378                  */
7379                 if (!offset && !failed_alloc && !cached &&
7380                     loop > LOOP_CACHING_NOWAIT) {
7381                         wait_block_group_cache_progress(block_group,
7382                                                 num_bytes + empty_size);
7383                         failed_alloc = true;
7384                         goto have_block_group;
7385                 } else if (!offset) {
7386                         goto loop;
7387                 }
7388 checks:
7389                 search_start = ALIGN(offset, root->stripesize);
7390
7391                 /* move on to the next group */
7392                 if (search_start + num_bytes >
7393                     block_group->key.objectid + block_group->key.offset) {
7394                         btrfs_add_free_space(block_group, offset, num_bytes);
7395                         goto loop;
7396                 }
7397
7398                 if (offset < search_start)
7399                         btrfs_add_free_space(block_group, offset,
7400                                              search_start - offset);
7401                 BUG_ON(offset > search_start);
7402
7403                 ret = btrfs_update_reserved_bytes(block_group, num_bytes,
7404                                                   alloc_type, delalloc);
7405                 if (ret == -EAGAIN) {
7406                         btrfs_add_free_space(block_group, offset, num_bytes);
7407                         goto loop;
7408                 }
7409
7410                 /* we are all good, lets return */
7411                 ins->objectid = search_start;
7412                 ins->offset = num_bytes;
7413
7414                 trace_btrfs_reserve_extent(orig_root, block_group,
7415                                            search_start, num_bytes);
7416                 btrfs_release_block_group(block_group, delalloc);
7417                 break;
7418 loop:
7419                 failed_cluster_refill = false;
7420                 failed_alloc = false;
7421                 BUG_ON(index != get_block_group_index(block_group));
7422                 btrfs_release_block_group(block_group, delalloc);
7423         }
7424         up_read(&space_info->groups_sem);
7425
7426         if ((loop == LOOP_CACHING_NOWAIT) && have_caching_bg
7427                 && !orig_have_caching_bg)
7428                 orig_have_caching_bg = true;
7429
7430         if (!ins->objectid && loop >= LOOP_CACHING_WAIT && have_caching_bg)
7431                 goto search;
7432
7433         if (!ins->objectid && ++index < BTRFS_NR_RAID_TYPES)
7434                 goto search;
7435
7436         /*
7437          * LOOP_CACHING_NOWAIT, search partially cached block groups, kicking
7438          *                      caching kthreads as we move along
7439          * LOOP_CACHING_WAIT, search everything, and wait if our bg is caching
7440          * LOOP_ALLOC_CHUNK, force a chunk allocation and try again
7441          * LOOP_NO_EMPTY_SIZE, set empty_size and empty_cluster to 0 and try
7442          *                      again
7443          */
7444         if (!ins->objectid && loop < LOOP_NO_EMPTY_SIZE) {
7445                 index = 0;
7446                 if (loop == LOOP_CACHING_NOWAIT) {
7447                         /*
7448                          * We want to skip the LOOP_CACHING_WAIT step if we
7449                          * don't have any unached bgs and we've alrelady done a
7450                          * full search through.
7451                          */
7452                         if (orig_have_caching_bg || !full_search)
7453                                 loop = LOOP_CACHING_WAIT;
7454                         else
7455                                 loop = LOOP_ALLOC_CHUNK;
7456                 } else {
7457                         loop++;
7458                 }
7459
7460                 if (loop == LOOP_ALLOC_CHUNK) {
7461                         struct btrfs_trans_handle *trans;
7462                         int exist = 0;
7463
7464                         trans = current->journal_info;
7465                         if (trans)
7466                                 exist = 1;
7467                         else
7468                                 trans = btrfs_join_transaction(root);
7469
7470                         if (IS_ERR(trans)) {
7471                                 ret = PTR_ERR(trans);
7472                                 goto out;
7473                         }
7474
7475                         ret = do_chunk_alloc(trans, root, flags,
7476                                              CHUNK_ALLOC_FORCE);
7477
7478                         /*
7479                          * If we can't allocate a new chunk we've already looped
7480                          * through at least once, move on to the NO_EMPTY_SIZE
7481                          * case.
7482                          */
7483                         if (ret == -ENOSPC)
7484                                 loop = LOOP_NO_EMPTY_SIZE;
7485
7486                         /*
7487                          * Do not bail out on ENOSPC since we
7488                          * can do more things.
7489                          */
7490                         if (ret < 0 && ret != -ENOSPC)
7491                                 btrfs_abort_transaction(trans,
7492                                                         root, ret);
7493                         else
7494                                 ret = 0;
7495                         if (!exist)
7496                                 btrfs_end_transaction(trans, root);
7497                         if (ret)
7498                                 goto out;
7499                 }
7500
7501                 if (loop == LOOP_NO_EMPTY_SIZE) {
7502                         /*
7503                          * Don't loop again if we already have no empty_size and
7504                          * no empty_cluster.
7505                          */
7506                         if (empty_size == 0 &&
7507                             empty_cluster == 0) {
7508                                 ret = -ENOSPC;
7509                                 goto out;
7510                         }
7511                         empty_size = 0;
7512                         empty_cluster = 0;
7513                 }
7514
7515                 goto search;
7516         } else if (!ins->objectid) {
7517                 ret = -ENOSPC;
7518         } else if (ins->objectid) {
7519                 if (!use_cluster && last_ptr) {
7520                         spin_lock(&last_ptr->lock);
7521                         last_ptr->window_start = ins->objectid;
7522                         spin_unlock(&last_ptr->lock);
7523                 }
7524                 ret = 0;
7525         }
7526 out:
7527         if (ret == -ENOSPC) {
7528                 spin_lock(&space_info->lock);
7529                 space_info->max_extent_size = max_extent_size;
7530                 spin_unlock(&space_info->lock);
7531                 ins->offset = max_extent_size;
7532         }
7533         return ret;
7534 }
7535
7536 static void dump_space_info(struct btrfs_space_info *info, u64 bytes,
7537                             int dump_block_groups)
7538 {
7539         struct btrfs_block_group_cache *cache;
7540         int index = 0;
7541
7542         spin_lock(&info->lock);
7543         printk(KERN_INFO "BTRFS: space_info %llu has %llu free, is %sfull\n",
7544                info->flags,
7545                info->total_bytes - info->bytes_used - info->bytes_pinned -
7546                info->bytes_reserved - info->bytes_readonly,
7547                (info->full) ? "" : "not ");
7548         printk(KERN_INFO "BTRFS: space_info total=%llu, used=%llu, pinned=%llu, "
7549                "reserved=%llu, may_use=%llu, readonly=%llu\n",
7550                info->total_bytes, info->bytes_used, info->bytes_pinned,
7551                info->bytes_reserved, info->bytes_may_use,
7552                info->bytes_readonly);
7553         spin_unlock(&info->lock);
7554
7555         if (!dump_block_groups)
7556                 return;
7557
7558         down_read(&info->groups_sem);
7559 again:
7560         list_for_each_entry(cache, &info->block_groups[index], list) {
7561                 spin_lock(&cache->lock);
7562                 printk(KERN_INFO "BTRFS: "
7563                            "block group %llu has %llu bytes, "
7564                            "%llu used %llu pinned %llu reserved %s\n",
7565                        cache->key.objectid, cache->key.offset,
7566                        btrfs_block_group_used(&cache->item), cache->pinned,
7567                        cache->reserved, cache->ro ? "[readonly]" : "");
7568                 btrfs_dump_free_space(cache, bytes);
7569                 spin_unlock(&cache->lock);
7570         }
7571         if (++index < BTRFS_NR_RAID_TYPES)
7572                 goto again;
7573         up_read(&info->groups_sem);
7574 }
7575
7576 int btrfs_reserve_extent(struct btrfs_root *root,
7577                          u64 num_bytes, u64 min_alloc_size,
7578                          u64 empty_size, u64 hint_byte,
7579                          struct btrfs_key *ins, int is_data, int delalloc)
7580 {
7581         bool final_tried = num_bytes == min_alloc_size;
7582         u64 flags;
7583         int ret;
7584
7585         flags = btrfs_get_alloc_profile(root, is_data);
7586 again:
7587         WARN_ON(num_bytes < root->sectorsize);
7588         ret = find_free_extent(root, num_bytes, empty_size, hint_byte, ins,
7589                                flags, delalloc);
7590
7591         if (ret == -ENOSPC) {
7592                 if (!final_tried && ins->offset) {
7593                         num_bytes = min(num_bytes >> 1, ins->offset);
7594                         num_bytes = round_down(num_bytes, root->sectorsize);
7595                         num_bytes = max(num_bytes, min_alloc_size);
7596                         if (num_bytes == min_alloc_size)
7597                                 final_tried = true;
7598                         goto again;
7599                 } else if (btrfs_test_opt(root, ENOSPC_DEBUG)) {
7600                         struct btrfs_space_info *sinfo;
7601
7602                         sinfo = __find_space_info(root->fs_info, flags);
7603                         btrfs_err(root->fs_info, "allocation failed flags %llu, wanted %llu",
7604                                 flags, num_bytes);
7605                         if (sinfo)
7606                                 dump_space_info(sinfo, num_bytes, 1);
7607                 }
7608         }
7609
7610         return ret;
7611 }
7612
7613 static int __btrfs_free_reserved_extent(struct btrfs_root *root,
7614                                         u64 start, u64 len,
7615                                         int pin, int delalloc)
7616 {
7617         struct btrfs_block_group_cache *cache;
7618         int ret = 0;
7619
7620         cache = btrfs_lookup_block_group(root->fs_info, start);
7621         if (!cache) {
7622                 btrfs_err(root->fs_info, "Unable to find block group for %llu",
7623                         start);
7624                 return -ENOSPC;
7625         }
7626
7627         if (pin)
7628                 pin_down_extent(root, cache, start, len, 1);
7629         else {
7630                 if (btrfs_test_opt(root, DISCARD))
7631                         ret = btrfs_discard_extent(root, start, len, NULL);
7632                 btrfs_add_free_space(cache, start, len);
7633                 btrfs_update_reserved_bytes(cache, len, RESERVE_FREE, delalloc);
7634         }
7635
7636         btrfs_put_block_group(cache);
7637
7638         trace_btrfs_reserved_extent_free(root, start, len);
7639
7640         return ret;
7641 }
7642
7643 int btrfs_free_reserved_extent(struct btrfs_root *root,
7644                                u64 start, u64 len, int delalloc)
7645 {
7646         return __btrfs_free_reserved_extent(root, start, len, 0, delalloc);
7647 }
7648
7649 int btrfs_free_and_pin_reserved_extent(struct btrfs_root *root,
7650                                        u64 start, u64 len)
7651 {
7652         return __btrfs_free_reserved_extent(root, start, len, 1, 0);
7653 }
7654
7655 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
7656                                       struct btrfs_root *root,
7657                                       u64 parent, u64 root_objectid,
7658                                       u64 flags, u64 owner, u64 offset,
7659                                       struct btrfs_key *ins, int ref_mod)
7660 {
7661         int ret;
7662         struct btrfs_fs_info *fs_info = root->fs_info;
7663         struct btrfs_extent_item *extent_item;
7664         struct btrfs_extent_inline_ref *iref;
7665         struct btrfs_path *path;
7666         struct extent_buffer *leaf;
7667         int type;
7668         u32 size;
7669
7670         if (parent > 0)
7671                 type = BTRFS_SHARED_DATA_REF_KEY;
7672         else
7673                 type = BTRFS_EXTENT_DATA_REF_KEY;
7674
7675         size = sizeof(*extent_item) + btrfs_extent_inline_ref_size(type);
7676
7677         path = btrfs_alloc_path();
7678         if (!path)
7679                 return -ENOMEM;
7680
7681         path->leave_spinning = 1;
7682         ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path,
7683                                       ins, size);
7684         if (ret) {
7685                 btrfs_free_path(path);
7686                 return ret;
7687         }
7688
7689         leaf = path->nodes[0];
7690         extent_item = btrfs_item_ptr(leaf, path->slots[0],
7691                                      struct btrfs_extent_item);
7692         btrfs_set_extent_refs(leaf, extent_item, ref_mod);
7693         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
7694         btrfs_set_extent_flags(leaf, extent_item,
7695                                flags | BTRFS_EXTENT_FLAG_DATA);
7696
7697         iref = (struct btrfs_extent_inline_ref *)(extent_item + 1);
7698         btrfs_set_extent_inline_ref_type(leaf, iref, type);
7699         if (parent > 0) {
7700                 struct btrfs_shared_data_ref *ref;
7701                 ref = (struct btrfs_shared_data_ref *)(iref + 1);
7702                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
7703                 btrfs_set_shared_data_ref_count(leaf, ref, ref_mod);
7704         } else {
7705                 struct btrfs_extent_data_ref *ref;
7706                 ref = (struct btrfs_extent_data_ref *)(&iref->offset);
7707                 btrfs_set_extent_data_ref_root(leaf, ref, root_objectid);
7708                 btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
7709                 btrfs_set_extent_data_ref_offset(leaf, ref, offset);
7710                 btrfs_set_extent_data_ref_count(leaf, ref, ref_mod);
7711         }
7712
7713         btrfs_mark_buffer_dirty(path->nodes[0]);
7714         btrfs_free_path(path);
7715
7716         ret = remove_from_free_space_tree(trans, fs_info, ins->objectid,
7717                                           ins->offset);
7718         if (ret)
7719                 return ret;
7720
7721         ret = update_block_group(trans, root, ins->objectid, ins->offset, 1);
7722         if (ret) { /* -ENOENT, logic error */
7723                 btrfs_err(fs_info, "update block group failed for %llu %llu",
7724                         ins->objectid, ins->offset);
7725                 BUG();
7726         }
7727         trace_btrfs_reserved_extent_alloc(root, ins->objectid, ins->offset);
7728         return ret;
7729 }
7730
7731 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
7732                                      struct btrfs_root *root,
7733                                      u64 parent, u64 root_objectid,
7734                                      u64 flags, struct btrfs_disk_key *key,
7735                                      int level, struct btrfs_key *ins)
7736 {
7737         int ret;
7738         struct btrfs_fs_info *fs_info = root->fs_info;
7739         struct btrfs_extent_item *extent_item;
7740         struct btrfs_tree_block_info *block_info;
7741         struct btrfs_extent_inline_ref *iref;
7742         struct btrfs_path *path;
7743         struct extent_buffer *leaf;
7744         u32 size = sizeof(*extent_item) + sizeof(*iref);
7745         u64 num_bytes = ins->offset;
7746         bool skinny_metadata = btrfs_fs_incompat(root->fs_info,
7747                                                  SKINNY_METADATA);
7748
7749         if (!skinny_metadata)
7750                 size += sizeof(*block_info);
7751
7752         path = btrfs_alloc_path();
7753         if (!path) {
7754                 btrfs_free_and_pin_reserved_extent(root, ins->objectid,
7755                                                    root->nodesize);
7756                 return -ENOMEM;
7757         }
7758
7759         path->leave_spinning = 1;
7760         ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path,
7761                                       ins, size);
7762         if (ret) {
7763                 btrfs_free_path(path);
7764                 btrfs_free_and_pin_reserved_extent(root, ins->objectid,
7765                                                    root->nodesize);
7766                 return ret;
7767         }
7768
7769         leaf = path->nodes[0];
7770         extent_item = btrfs_item_ptr(leaf, path->slots[0],
7771                                      struct btrfs_extent_item);
7772         btrfs_set_extent_refs(leaf, extent_item, 1);
7773         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
7774         btrfs_set_extent_flags(leaf, extent_item,
7775                                flags | BTRFS_EXTENT_FLAG_TREE_BLOCK);
7776
7777         if (skinny_metadata) {
7778                 iref = (struct btrfs_extent_inline_ref *)(extent_item + 1);
7779                 num_bytes = root->nodesize;
7780         } else {
7781                 block_info = (struct btrfs_tree_block_info *)(extent_item + 1);
7782                 btrfs_set_tree_block_key(leaf, block_info, key);
7783                 btrfs_set_tree_block_level(leaf, block_info, level);
7784                 iref = (struct btrfs_extent_inline_ref *)(block_info + 1);
7785         }
7786
7787         if (parent > 0) {
7788                 BUG_ON(!(flags & BTRFS_BLOCK_FLAG_FULL_BACKREF));
7789                 btrfs_set_extent_inline_ref_type(leaf, iref,
7790                                                  BTRFS_SHARED_BLOCK_REF_KEY);
7791                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
7792         } else {
7793                 btrfs_set_extent_inline_ref_type(leaf, iref,
7794                                                  BTRFS_TREE_BLOCK_REF_KEY);
7795                 btrfs_set_extent_inline_ref_offset(leaf, iref, root_objectid);
7796         }
7797
7798         btrfs_mark_buffer_dirty(leaf);
7799         btrfs_free_path(path);
7800
7801         ret = remove_from_free_space_tree(trans, fs_info, ins->objectid,
7802                                           num_bytes);
7803         if (ret)
7804                 return ret;
7805
7806         ret = update_block_group(trans, root, ins->objectid, root->nodesize,
7807                                  1);
7808         if (ret) { /* -ENOENT, logic error */
7809                 btrfs_err(fs_info, "update block group failed for %llu %llu",
7810                         ins->objectid, ins->offset);
7811                 BUG();
7812         }
7813
7814         trace_btrfs_reserved_extent_alloc(root, ins->objectid, root->nodesize);
7815         return ret;
7816 }
7817
7818 int btrfs_alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
7819                                      struct btrfs_root *root,
7820                                      u64 root_objectid, u64 owner,
7821                                      u64 offset, u64 ram_bytes,
7822                                      struct btrfs_key *ins)
7823 {
7824         int ret;
7825
7826         BUG_ON(root_objectid == BTRFS_TREE_LOG_OBJECTID);
7827
7828         ret = btrfs_add_delayed_data_ref(root->fs_info, trans, ins->objectid,
7829                                          ins->offset, 0,
7830                                          root_objectid, owner, offset,
7831                                          ram_bytes, BTRFS_ADD_DELAYED_EXTENT,
7832                                          NULL);
7833         return ret;
7834 }
7835
7836 /*
7837  * this is used by the tree logging recovery code.  It records that
7838  * an extent has been allocated and makes sure to clear the free
7839  * space cache bits as well
7840  */
7841 int btrfs_alloc_logged_file_extent(struct btrfs_trans_handle *trans,
7842                                    struct btrfs_root *root,
7843                                    u64 root_objectid, u64 owner, u64 offset,
7844                                    struct btrfs_key *ins)
7845 {
7846         int ret;
7847         struct btrfs_block_group_cache *block_group;
7848
7849         /*
7850          * Mixed block groups will exclude before processing the log so we only
7851          * need to do the exlude dance if this fs isn't mixed.
7852          */
7853         if (!btrfs_fs_incompat(root->fs_info, MIXED_GROUPS)) {
7854                 ret = __exclude_logged_extent(root, ins->objectid, ins->offset);
7855                 if (ret)
7856                         return ret;
7857         }
7858
7859         block_group = btrfs_lookup_block_group(root->fs_info, ins->objectid);
7860         if (!block_group)
7861                 return -EINVAL;
7862
7863         ret = btrfs_update_reserved_bytes(block_group, ins->offset,
7864                                           RESERVE_ALLOC_NO_ACCOUNT, 0);
7865         BUG_ON(ret); /* logic error */
7866         ret = alloc_reserved_file_extent(trans, root, 0, root_objectid,
7867                                          0, owner, offset, ins, 1);
7868         btrfs_put_block_group(block_group);
7869         return ret;
7870 }
7871
7872 static struct extent_buffer *
7873 btrfs_init_new_buffer(struct btrfs_trans_handle *trans, struct btrfs_root *root,
7874                       u64 bytenr, int level)
7875 {
7876         struct extent_buffer *buf;
7877
7878         buf = btrfs_find_create_tree_block(root, bytenr);
7879         if (!buf)
7880                 return ERR_PTR(-ENOMEM);
7881         btrfs_set_header_generation(buf, trans->transid);
7882         btrfs_set_buffer_lockdep_class(root->root_key.objectid, buf, level);
7883         btrfs_tree_lock(buf);
7884         clean_tree_block(trans, root->fs_info, buf);
7885         clear_bit(EXTENT_BUFFER_STALE, &buf->bflags);
7886
7887         btrfs_set_lock_blocking(buf);
7888         set_extent_buffer_uptodate(buf);
7889
7890         if (root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID) {
7891                 buf->log_index = root->log_transid % 2;
7892                 /*
7893                  * we allow two log transactions at a time, use different
7894                  * EXENT bit to differentiate dirty pages.
7895                  */
7896                 if (buf->log_index == 0)
7897                         set_extent_dirty(&root->dirty_log_pages, buf->start,
7898                                         buf->start + buf->len - 1, GFP_NOFS);
7899                 else
7900                         set_extent_new(&root->dirty_log_pages, buf->start,
7901                                         buf->start + buf->len - 1, GFP_NOFS);
7902         } else {
7903                 buf->log_index = -1;
7904                 set_extent_dirty(&trans->transaction->dirty_pages, buf->start,
7905                          buf->start + buf->len - 1, GFP_NOFS);
7906         }
7907         trans->blocks_used++;
7908         /* this returns a buffer locked for blocking */
7909         return buf;
7910 }
7911
7912 static struct btrfs_block_rsv *
7913 use_block_rsv(struct btrfs_trans_handle *trans,
7914               struct btrfs_root *root, u32 blocksize)
7915 {
7916         struct btrfs_block_rsv *block_rsv;
7917         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
7918         int ret;
7919         bool global_updated = false;
7920
7921         block_rsv = get_block_rsv(trans, root);
7922
7923         if (unlikely(block_rsv->size == 0))
7924                 goto try_reserve;
7925 again:
7926         ret = block_rsv_use_bytes(block_rsv, blocksize);
7927         if (!ret)
7928                 return block_rsv;
7929
7930         if (block_rsv->failfast)
7931                 return ERR_PTR(ret);
7932
7933         if (block_rsv->type == BTRFS_BLOCK_RSV_GLOBAL && !global_updated) {
7934                 global_updated = true;
7935                 update_global_block_rsv(root->fs_info);
7936                 goto again;
7937         }
7938
7939         if (btrfs_test_opt(root, ENOSPC_DEBUG)) {
7940                 static DEFINE_RATELIMIT_STATE(_rs,
7941                                 DEFAULT_RATELIMIT_INTERVAL * 10,
7942                                 /*DEFAULT_RATELIMIT_BURST*/ 1);
7943                 if (__ratelimit(&_rs))
7944                         WARN(1, KERN_DEBUG
7945                                 "BTRFS: block rsv returned %d\n", ret);
7946         }
7947 try_reserve:
7948         ret = reserve_metadata_bytes(root, block_rsv, blocksize,
7949                                      BTRFS_RESERVE_NO_FLUSH);
7950         if (!ret)
7951                 return block_rsv;
7952         /*
7953          * If we couldn't reserve metadata bytes try and use some from
7954          * the global reserve if its space type is the same as the global
7955          * reservation.
7956          */
7957         if (block_rsv->type != BTRFS_BLOCK_RSV_GLOBAL &&
7958             block_rsv->space_info == global_rsv->space_info) {
7959                 ret = block_rsv_use_bytes(global_rsv, blocksize);
7960                 if (!ret)
7961                         return global_rsv;
7962         }
7963         return ERR_PTR(ret);
7964 }
7965
7966 static void unuse_block_rsv(struct btrfs_fs_info *fs_info,
7967                             struct btrfs_block_rsv *block_rsv, u32 blocksize)
7968 {
7969         block_rsv_add_bytes(block_rsv, blocksize, 0);
7970         block_rsv_release_bytes(fs_info, block_rsv, NULL, 0);
7971 }
7972
7973 /*
7974  * finds a free extent and does all the dirty work required for allocation
7975  * returns the tree buffer or an ERR_PTR on error.
7976  */
7977 struct extent_buffer *btrfs_alloc_tree_block(struct btrfs_trans_handle *trans,
7978                                         struct btrfs_root *root,
7979                                         u64 parent, u64 root_objectid,
7980                                         struct btrfs_disk_key *key, int level,
7981                                         u64 hint, u64 empty_size)
7982 {
7983         struct btrfs_key ins;
7984         struct btrfs_block_rsv *block_rsv;
7985         struct extent_buffer *buf;
7986         struct btrfs_delayed_extent_op *extent_op;
7987         u64 flags = 0;
7988         int ret;
7989         u32 blocksize = root->nodesize;
7990         bool skinny_metadata = btrfs_fs_incompat(root->fs_info,
7991                                                  SKINNY_METADATA);
7992
7993         if (btrfs_test_is_dummy_root(root)) {
7994                 buf = btrfs_init_new_buffer(trans, root, root->alloc_bytenr,
7995                                             level);
7996                 if (!IS_ERR(buf))
7997                         root->alloc_bytenr += blocksize;
7998                 return buf;
7999         }
8000
8001         block_rsv = use_block_rsv(trans, root, blocksize);
8002         if (IS_ERR(block_rsv))
8003                 return ERR_CAST(block_rsv);
8004
8005         ret = btrfs_reserve_extent(root, blocksize, blocksize,
8006                                    empty_size, hint, &ins, 0, 0);
8007         if (ret)
8008                 goto out_unuse;
8009
8010         buf = btrfs_init_new_buffer(trans, root, ins.objectid, level);
8011         if (IS_ERR(buf)) {
8012                 ret = PTR_ERR(buf);
8013                 goto out_free_reserved;
8014         }
8015
8016         if (root_objectid == BTRFS_TREE_RELOC_OBJECTID) {
8017                 if (parent == 0)
8018                         parent = ins.objectid;
8019                 flags |= BTRFS_BLOCK_FLAG_FULL_BACKREF;
8020         } else
8021                 BUG_ON(parent > 0);
8022
8023         if (root_objectid != BTRFS_TREE_LOG_OBJECTID) {
8024                 extent_op = btrfs_alloc_delayed_extent_op();
8025                 if (!extent_op) {
8026                         ret = -ENOMEM;
8027                         goto out_free_buf;
8028                 }
8029                 if (key)
8030                         memcpy(&extent_op->key, key, sizeof(extent_op->key));
8031                 else
8032                         memset(&extent_op->key, 0, sizeof(extent_op->key));
8033                 extent_op->flags_to_set = flags;
8034                 if (skinny_metadata)
8035                         extent_op->update_key = 0;
8036                 else
8037                         extent_op->update_key = 1;
8038                 extent_op->update_flags = 1;
8039                 extent_op->is_data = 0;
8040                 extent_op->level = level;
8041
8042                 ret = btrfs_add_delayed_tree_ref(root->fs_info, trans,
8043                                                  ins.objectid, ins.offset,
8044                                                  parent, root_objectid, level,
8045                                                  BTRFS_ADD_DELAYED_EXTENT,
8046                                                  extent_op);
8047                 if (ret)
8048                         goto out_free_delayed;
8049         }
8050         return buf;
8051
8052 out_free_delayed:
8053         btrfs_free_delayed_extent_op(extent_op);
8054 out_free_buf:
8055         free_extent_buffer(buf);
8056 out_free_reserved:
8057         btrfs_free_reserved_extent(root, ins.objectid, ins.offset, 0);
8058 out_unuse:
8059         unuse_block_rsv(root->fs_info, block_rsv, blocksize);
8060         return ERR_PTR(ret);
8061 }
8062
8063 struct walk_control {
8064         u64 refs[BTRFS_MAX_LEVEL];
8065         u64 flags[BTRFS_MAX_LEVEL];
8066         struct btrfs_key update_progress;
8067         int stage;
8068         int level;
8069         int shared_level;
8070         int update_ref;
8071         int keep_locks;
8072         int reada_slot;
8073         int reada_count;
8074         int for_reloc;
8075 };
8076
8077 #define DROP_REFERENCE  1
8078 #define UPDATE_BACKREF  2
8079
8080 static noinline void reada_walk_down(struct btrfs_trans_handle *trans,
8081                                      struct btrfs_root *root,
8082                                      struct walk_control *wc,
8083                                      struct btrfs_path *path)
8084 {
8085         u64 bytenr;
8086         u64 generation;
8087         u64 refs;
8088         u64 flags;
8089         u32 nritems;
8090         u32 blocksize;
8091         struct btrfs_key key;
8092         struct extent_buffer *eb;
8093         int ret;
8094         int slot;
8095         int nread = 0;
8096
8097         if (path->slots[wc->level] < wc->reada_slot) {
8098                 wc->reada_count = wc->reada_count * 2 / 3;
8099                 wc->reada_count = max(wc->reada_count, 2);
8100         } else {
8101                 wc->reada_count = wc->reada_count * 3 / 2;
8102                 wc->reada_count = min_t(int, wc->reada_count,
8103                                         BTRFS_NODEPTRS_PER_BLOCK(root));
8104         }
8105
8106         eb = path->nodes[wc->level];
8107         nritems = btrfs_header_nritems(eb);
8108         blocksize = root->nodesize;
8109
8110         for (slot = path->slots[wc->level]; slot < nritems; slot++) {
8111                 if (nread >= wc->reada_count)
8112                         break;
8113
8114                 cond_resched();
8115                 bytenr = btrfs_node_blockptr(eb, slot);
8116                 generation = btrfs_node_ptr_generation(eb, slot);
8117
8118                 if (slot == path->slots[wc->level])
8119                         goto reada;
8120
8121                 if (wc->stage == UPDATE_BACKREF &&
8122                     generation <= root->root_key.offset)
8123                         continue;
8124
8125                 /* We don't lock the tree block, it's OK to be racy here */
8126                 ret = btrfs_lookup_extent_info(trans, root, bytenr,
8127                                                wc->level - 1, 1, &refs,
8128                                                &flags);
8129                 /* We don't care about errors in readahead. */
8130                 if (ret < 0)
8131                         continue;
8132                 BUG_ON(refs == 0);
8133
8134                 if (wc->stage == DROP_REFERENCE) {
8135                         if (refs == 1)
8136                                 goto reada;
8137
8138                         if (wc->level == 1 &&
8139                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
8140                                 continue;
8141                         if (!wc->update_ref ||
8142                             generation <= root->root_key.offset)
8143                                 continue;
8144                         btrfs_node_key_to_cpu(eb, &key, slot);
8145                         ret = btrfs_comp_cpu_keys(&key,
8146                                                   &wc->update_progress);
8147                         if (ret < 0)
8148                                 continue;
8149                 } else {
8150                         if (wc->level == 1 &&
8151                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
8152                                 continue;
8153                 }
8154 reada:
8155                 readahead_tree_block(root, bytenr);
8156                 nread++;
8157         }
8158         wc->reada_slot = slot;
8159 }
8160
8161 /*
8162  * These may not be seen by the usual inc/dec ref code so we have to
8163  * add them here.
8164  */
8165 static int record_one_subtree_extent(struct btrfs_trans_handle *trans,
8166                                      struct btrfs_root *root, u64 bytenr,
8167                                      u64 num_bytes)
8168 {
8169         struct btrfs_qgroup_extent_record *qrecord;
8170         struct btrfs_delayed_ref_root *delayed_refs;
8171
8172         qrecord = kmalloc(sizeof(*qrecord), GFP_NOFS);
8173         if (!qrecord)
8174                 return -ENOMEM;
8175
8176         qrecord->bytenr = bytenr;
8177         qrecord->num_bytes = num_bytes;
8178         qrecord->old_roots = NULL;
8179
8180         delayed_refs = &trans->transaction->delayed_refs;
8181         spin_lock(&delayed_refs->lock);
8182         if (btrfs_qgroup_insert_dirty_extent(delayed_refs, qrecord))
8183                 kfree(qrecord);
8184         spin_unlock(&delayed_refs->lock);
8185
8186         return 0;
8187 }
8188
8189 static int account_leaf_items(struct btrfs_trans_handle *trans,
8190                               struct btrfs_root *root,
8191                               struct extent_buffer *eb)
8192 {
8193         int nr = btrfs_header_nritems(eb);
8194         int i, extent_type, ret;
8195         struct btrfs_key key;
8196         struct btrfs_file_extent_item *fi;
8197         u64 bytenr, num_bytes;
8198
8199         /* We can be called directly from walk_up_proc() */
8200         if (!root->fs_info->quota_enabled)
8201                 return 0;
8202
8203         for (i = 0; i < nr; i++) {
8204                 btrfs_item_key_to_cpu(eb, &key, i);
8205
8206                 if (key.type != BTRFS_EXTENT_DATA_KEY)
8207                         continue;
8208
8209                 fi = btrfs_item_ptr(eb, i, struct btrfs_file_extent_item);
8210                 /* filter out non qgroup-accountable extents  */
8211                 extent_type = btrfs_file_extent_type(eb, fi);
8212
8213                 if (extent_type == BTRFS_FILE_EXTENT_INLINE)
8214                         continue;
8215
8216                 bytenr = btrfs_file_extent_disk_bytenr(eb, fi);
8217                 if (!bytenr)
8218                         continue;
8219
8220                 num_bytes = btrfs_file_extent_disk_num_bytes(eb, fi);
8221
8222                 ret = record_one_subtree_extent(trans, root, bytenr, num_bytes);
8223                 if (ret)
8224                         return ret;
8225         }
8226         return 0;
8227 }
8228
8229 /*
8230  * Walk up the tree from the bottom, freeing leaves and any interior
8231  * nodes which have had all slots visited. If a node (leaf or
8232  * interior) is freed, the node above it will have it's slot
8233  * incremented. The root node will never be freed.
8234  *
8235  * At the end of this function, we should have a path which has all
8236  * slots incremented to the next position for a search. If we need to
8237  * read a new node it will be NULL and the node above it will have the
8238  * correct slot selected for a later read.
8239  *
8240  * If we increment the root nodes slot counter past the number of
8241  * elements, 1 is returned to signal completion of the search.
8242  */
8243 static int adjust_slots_upwards(struct btrfs_root *root,
8244                                 struct btrfs_path *path, int root_level)
8245 {
8246         int level = 0;
8247         int nr, slot;
8248         struct extent_buffer *eb;
8249
8250         if (root_level == 0)
8251                 return 1;
8252
8253         while (level <= root_level) {
8254                 eb = path->nodes[level];
8255                 nr = btrfs_header_nritems(eb);
8256                 path->slots[level]++;
8257                 slot = path->slots[level];
8258                 if (slot >= nr || level == 0) {
8259                         /*
8260                          * Don't free the root -  we will detect this
8261                          * condition after our loop and return a
8262                          * positive value for caller to stop walking the tree.
8263                          */
8264                         if (level != root_level) {
8265                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
8266                                 path->locks[level] = 0;
8267
8268                                 free_extent_buffer(eb);
8269                                 path->nodes[level] = NULL;
8270                                 path->slots[level] = 0;
8271                         }
8272                 } else {
8273                         /*
8274                          * We have a valid slot to walk back down
8275                          * from. Stop here so caller can process these
8276                          * new nodes.
8277                          */
8278                         break;
8279                 }
8280
8281                 level++;
8282         }
8283
8284         eb = path->nodes[root_level];
8285         if (path->slots[root_level] >= btrfs_header_nritems(eb))
8286                 return 1;
8287
8288         return 0;
8289 }
8290
8291 /*
8292  * root_eb is the subtree root and is locked before this function is called.
8293  */
8294 static int account_shared_subtree(struct btrfs_trans_handle *trans,
8295                                   struct btrfs_root *root,
8296                                   struct extent_buffer *root_eb,
8297                                   u64 root_gen,
8298                                   int root_level)
8299 {
8300         int ret = 0;
8301         int level;
8302         struct extent_buffer *eb = root_eb;
8303         struct btrfs_path *path = NULL;
8304
8305         BUG_ON(root_level < 0 || root_level > BTRFS_MAX_LEVEL);
8306         BUG_ON(root_eb == NULL);
8307
8308         if (!root->fs_info->quota_enabled)
8309                 return 0;
8310
8311         if (!extent_buffer_uptodate(root_eb)) {
8312                 ret = btrfs_read_buffer(root_eb, root_gen);
8313                 if (ret)
8314                         goto out;
8315         }
8316
8317         if (root_level == 0) {
8318                 ret = account_leaf_items(trans, root, root_eb);
8319                 goto out;
8320         }
8321
8322         path = btrfs_alloc_path();
8323         if (!path)
8324                 return -ENOMEM;
8325
8326         /*
8327          * Walk down the tree.  Missing extent blocks are filled in as
8328          * we go. Metadata is accounted every time we read a new
8329          * extent block.
8330          *
8331          * When we reach a leaf, we account for file extent items in it,
8332          * walk back up the tree (adjusting slot pointers as we go)
8333          * and restart the search process.
8334          */
8335         extent_buffer_get(root_eb); /* For path */
8336         path->nodes[root_level] = root_eb;
8337         path->slots[root_level] = 0;
8338         path->locks[root_level] = 0; /* so release_path doesn't try to unlock */
8339 walk_down:
8340         level = root_level;
8341         while (level >= 0) {
8342                 if (path->nodes[level] == NULL) {
8343                         int parent_slot;
8344                         u64 child_gen;
8345                         u64 child_bytenr;
8346
8347                         /* We need to get child blockptr/gen from
8348                          * parent before we can read it. */
8349                         eb = path->nodes[level + 1];
8350                         parent_slot = path->slots[level + 1];
8351                         child_bytenr = btrfs_node_blockptr(eb, parent_slot);
8352                         child_gen = btrfs_node_ptr_generation(eb, parent_slot);
8353
8354                         eb = read_tree_block(root, child_bytenr, child_gen);
8355                         if (IS_ERR(eb)) {
8356                                 ret = PTR_ERR(eb);
8357                                 goto out;
8358                         } else if (!extent_buffer_uptodate(eb)) {
8359                                 free_extent_buffer(eb);
8360                                 ret = -EIO;
8361                                 goto out;
8362                         }
8363
8364                         path->nodes[level] = eb;
8365                         path->slots[level] = 0;
8366
8367                         btrfs_tree_read_lock(eb);
8368                         btrfs_set_lock_blocking_rw(eb, BTRFS_READ_LOCK);
8369                         path->locks[level] = BTRFS_READ_LOCK_BLOCKING;
8370
8371                         ret = record_one_subtree_extent(trans, root, child_bytenr,
8372                                                         root->nodesize);
8373                         if (ret)
8374                                 goto out;
8375                 }
8376
8377                 if (level == 0) {
8378                         ret = account_leaf_items(trans, root, path->nodes[level]);
8379                         if (ret)
8380                                 goto out;
8381
8382                         /* Nonzero return here means we completed our search */
8383                         ret = adjust_slots_upwards(root, path, root_level);
8384                         if (ret)
8385                                 break;
8386
8387                         /* Restart search with new slots */
8388                         goto walk_down;
8389                 }
8390
8391                 level--;
8392         }
8393
8394         ret = 0;
8395 out:
8396         btrfs_free_path(path);
8397
8398         return ret;
8399 }
8400
8401 /*
8402  * helper to process tree block while walking down the tree.
8403  *
8404  * when wc->stage == UPDATE_BACKREF, this function updates
8405  * back refs for pointers in the block.
8406  *
8407  * NOTE: return value 1 means we should stop walking down.
8408  */
8409 static noinline int walk_down_proc(struct btrfs_trans_handle *trans,
8410                                    struct btrfs_root *root,
8411                                    struct btrfs_path *path,
8412                                    struct walk_control *wc, int lookup_info)
8413 {
8414         int level = wc->level;
8415         struct extent_buffer *eb = path->nodes[level];
8416         u64 flag = BTRFS_BLOCK_FLAG_FULL_BACKREF;
8417         int ret;
8418
8419         if (wc->stage == UPDATE_BACKREF &&
8420             btrfs_header_owner(eb) != root->root_key.objectid)
8421                 return 1;
8422
8423         /*
8424          * when reference count of tree block is 1, it won't increase
8425          * again. once full backref flag is set, we never clear it.
8426          */
8427         if (lookup_info &&
8428             ((wc->stage == DROP_REFERENCE && wc->refs[level] != 1) ||
8429              (wc->stage == UPDATE_BACKREF && !(wc->flags[level] & flag)))) {
8430                 BUG_ON(!path->locks[level]);
8431                 ret = btrfs_lookup_extent_info(trans, root,
8432                                                eb->start, level, 1,
8433                                                &wc->refs[level],
8434                                                &wc->flags[level]);
8435                 BUG_ON(ret == -ENOMEM);
8436                 if (ret)
8437                         return ret;
8438                 BUG_ON(wc->refs[level] == 0);
8439         }
8440
8441         if (wc->stage == DROP_REFERENCE) {
8442                 if (wc->refs[level] > 1)
8443                         return 1;
8444
8445                 if (path->locks[level] && !wc->keep_locks) {
8446                         btrfs_tree_unlock_rw(eb, path->locks[level]);
8447                         path->locks[level] = 0;
8448                 }
8449                 return 0;
8450         }
8451
8452         /* wc->stage == UPDATE_BACKREF */
8453         if (!(wc->flags[level] & flag)) {
8454                 BUG_ON(!path->locks[level]);
8455                 ret = btrfs_inc_ref(trans, root, eb, 1);
8456                 BUG_ON(ret); /* -ENOMEM */
8457                 ret = btrfs_dec_ref(trans, root, eb, 0);
8458                 BUG_ON(ret); /* -ENOMEM */
8459                 ret = btrfs_set_disk_extent_flags(trans, root, eb->start,
8460                                                   eb->len, flag,
8461                                                   btrfs_header_level(eb), 0);
8462                 BUG_ON(ret); /* -ENOMEM */
8463                 wc->flags[level] |= flag;
8464         }
8465
8466         /*
8467          * the block is shared by multiple trees, so it's not good to
8468          * keep the tree lock
8469          */
8470         if (path->locks[level] && level > 0) {
8471                 btrfs_tree_unlock_rw(eb, path->locks[level]);
8472                 path->locks[level] = 0;
8473         }
8474         return 0;
8475 }
8476
8477 /*
8478  * helper to process tree block pointer.
8479  *
8480  * when wc->stage == DROP_REFERENCE, this function checks
8481  * reference count of the block pointed to. if the block
8482  * is shared and we need update back refs for the subtree
8483  * rooted at the block, this function changes wc->stage to
8484  * UPDATE_BACKREF. if the block is shared and there is no
8485  * need to update back, this function drops the reference
8486  * to the block.
8487  *
8488  * NOTE: return value 1 means we should stop walking down.
8489  */
8490 static noinline int do_walk_down(struct btrfs_trans_handle *trans,
8491                                  struct btrfs_root *root,
8492                                  struct btrfs_path *path,
8493                                  struct walk_control *wc, int *lookup_info)
8494 {
8495         u64 bytenr;
8496         u64 generation;
8497         u64 parent;
8498         u32 blocksize;
8499         struct btrfs_key key;
8500         struct extent_buffer *next;
8501         int level = wc->level;
8502         int reada = 0;
8503         int ret = 0;
8504         bool need_account = false;
8505
8506         generation = btrfs_node_ptr_generation(path->nodes[level],
8507                                                path->slots[level]);
8508         /*
8509          * if the lower level block was created before the snapshot
8510          * was created, we know there is no need to update back refs
8511          * for the subtree
8512          */
8513         if (wc->stage == UPDATE_BACKREF &&
8514             generation <= root->root_key.offset) {
8515                 *lookup_info = 1;
8516                 return 1;
8517         }
8518
8519         bytenr = btrfs_node_blockptr(path->nodes[level], path->slots[level]);
8520         blocksize = root->nodesize;
8521
8522         next = btrfs_find_tree_block(root->fs_info, bytenr);
8523         if (!next) {
8524                 next = btrfs_find_create_tree_block(root, bytenr);
8525                 if (!next)
8526                         return -ENOMEM;
8527                 btrfs_set_buffer_lockdep_class(root->root_key.objectid, next,
8528                                                level - 1);
8529                 reada = 1;
8530         }
8531         btrfs_tree_lock(next);
8532         btrfs_set_lock_blocking(next);
8533
8534         ret = btrfs_lookup_extent_info(trans, root, bytenr, level - 1, 1,
8535                                        &wc->refs[level - 1],
8536                                        &wc->flags[level - 1]);
8537         if (ret < 0) {
8538                 btrfs_tree_unlock(next);
8539                 return ret;
8540         }
8541
8542         if (unlikely(wc->refs[level - 1] == 0)) {
8543                 btrfs_err(root->fs_info, "Missing references.");
8544                 BUG();
8545         }
8546         *lookup_info = 0;
8547
8548         if (wc->stage == DROP_REFERENCE) {
8549                 if (wc->refs[level - 1] > 1) {
8550                         need_account = true;
8551                         if (level == 1 &&
8552                             (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
8553                                 goto skip;
8554
8555                         if (!wc->update_ref ||
8556                             generation <= root->root_key.offset)
8557                                 goto skip;
8558
8559                         btrfs_node_key_to_cpu(path->nodes[level], &key,
8560                                               path->slots[level]);
8561                         ret = btrfs_comp_cpu_keys(&key, &wc->update_progress);
8562                         if (ret < 0)
8563                                 goto skip;
8564
8565                         wc->stage = UPDATE_BACKREF;
8566                         wc->shared_level = level - 1;
8567                 }
8568         } else {
8569                 if (level == 1 &&
8570                     (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
8571                         goto skip;
8572         }
8573
8574         if (!btrfs_buffer_uptodate(next, generation, 0)) {
8575                 btrfs_tree_unlock(next);
8576                 free_extent_buffer(next);
8577                 next = NULL;
8578                 *lookup_info = 1;
8579         }
8580
8581         if (!next) {
8582                 if (reada && level == 1)
8583                         reada_walk_down(trans, root, wc, path);
8584                 next = read_tree_block(root, bytenr, generation);
8585                 if (IS_ERR(next)) {
8586                         return PTR_ERR(next);
8587                 } else if (!extent_buffer_uptodate(next)) {
8588                         free_extent_buffer(next);
8589                         return -EIO;
8590                 }
8591                 btrfs_tree_lock(next);
8592                 btrfs_set_lock_blocking(next);
8593         }
8594
8595         level--;
8596         BUG_ON(level != btrfs_header_level(next));
8597         path->nodes[level] = next;
8598         path->slots[level] = 0;
8599         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
8600         wc->level = level;
8601         if (wc->level == 1)
8602                 wc->reada_slot = 0;
8603         return 0;
8604 skip:
8605         wc->refs[level - 1] = 0;
8606         wc->flags[level - 1] = 0;
8607         if (wc->stage == DROP_REFERENCE) {
8608                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF) {
8609                         parent = path->nodes[level]->start;
8610                 } else {
8611                         BUG_ON(root->root_key.objectid !=
8612                                btrfs_header_owner(path->nodes[level]));
8613                         parent = 0;
8614                 }
8615
8616                 if (need_account) {
8617                         ret = account_shared_subtree(trans, root, next,
8618                                                      generation, level - 1);
8619                         if (ret) {
8620                                 btrfs_err_rl(root->fs_info,
8621                                         "Error "
8622                                         "%d accounting shared subtree. Quota "
8623                                         "is out of sync, rescan required.",
8624                                         ret);
8625                         }
8626                 }
8627                 ret = btrfs_free_extent(trans, root, bytenr, blocksize, parent,
8628                                 root->root_key.objectid, level - 1, 0);
8629                 BUG_ON(ret); /* -ENOMEM */
8630         }
8631         btrfs_tree_unlock(next);
8632         free_extent_buffer(next);
8633         *lookup_info = 1;
8634         return 1;
8635 }
8636
8637 /*
8638  * helper to process tree block while walking up the tree.
8639  *
8640  * when wc->stage == DROP_REFERENCE, this function drops
8641  * reference count on the block.
8642  *
8643  * when wc->stage == UPDATE_BACKREF, this function changes
8644  * wc->stage back to DROP_REFERENCE if we changed wc->stage
8645  * to UPDATE_BACKREF previously while processing the block.
8646  *
8647  * NOTE: return value 1 means we should stop walking up.
8648  */
8649 static noinline int walk_up_proc(struct btrfs_trans_handle *trans,
8650                                  struct btrfs_root *root,
8651                                  struct btrfs_path *path,
8652                                  struct walk_control *wc)
8653 {
8654         int ret;
8655         int level = wc->level;
8656         struct extent_buffer *eb = path->nodes[level];
8657         u64 parent = 0;
8658
8659         if (wc->stage == UPDATE_BACKREF) {
8660                 BUG_ON(wc->shared_level < level);
8661                 if (level < wc->shared_level)
8662                         goto out;
8663
8664                 ret = find_next_key(path, level + 1, &wc->update_progress);
8665                 if (ret > 0)
8666                         wc->update_ref = 0;
8667
8668                 wc->stage = DROP_REFERENCE;
8669                 wc->shared_level = -1;
8670                 path->slots[level] = 0;
8671
8672                 /*
8673                  * check reference count again if the block isn't locked.
8674                  * we should start walking down the tree again if reference
8675                  * count is one.
8676                  */
8677                 if (!path->locks[level]) {
8678                         BUG_ON(level == 0);
8679                         btrfs_tree_lock(eb);
8680                         btrfs_set_lock_blocking(eb);
8681                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
8682
8683                         ret = btrfs_lookup_extent_info(trans, root,
8684                                                        eb->start, level, 1,
8685                                                        &wc->refs[level],
8686                                                        &wc->flags[level]);
8687                         if (ret < 0) {
8688                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
8689                                 path->locks[level] = 0;
8690                                 return ret;
8691                         }
8692                         BUG_ON(wc->refs[level] == 0);
8693                         if (wc->refs[level] == 1) {
8694                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
8695                                 path->locks[level] = 0;
8696                                 return 1;
8697                         }
8698                 }
8699         }
8700
8701         /* wc->stage == DROP_REFERENCE */
8702         BUG_ON(wc->refs[level] > 1 && !path->locks[level]);
8703
8704         if (wc->refs[level] == 1) {
8705                 if (level == 0) {
8706                         if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
8707                                 ret = btrfs_dec_ref(trans, root, eb, 1);
8708                         else
8709                                 ret = btrfs_dec_ref(trans, root, eb, 0);
8710                         BUG_ON(ret); /* -ENOMEM */
8711                         ret = account_leaf_items(trans, root, eb);
8712                         if (ret) {
8713                                 btrfs_err_rl(root->fs_info,
8714                                         "error "
8715                                         "%d accounting leaf items. Quota "
8716                                         "is out of sync, rescan required.",
8717                                         ret);
8718                         }
8719                 }
8720                 /* make block locked assertion in clean_tree_block happy */
8721                 if (!path->locks[level] &&
8722                     btrfs_header_generation(eb) == trans->transid) {
8723                         btrfs_tree_lock(eb);
8724                         btrfs_set_lock_blocking(eb);
8725                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
8726                 }
8727                 clean_tree_block(trans, root->fs_info, eb);
8728         }
8729
8730         if (eb == root->node) {
8731                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
8732                         parent = eb->start;
8733                 else
8734                         BUG_ON(root->root_key.objectid !=
8735                                btrfs_header_owner(eb));
8736         } else {
8737                 if (wc->flags[level + 1] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
8738                         parent = path->nodes[level + 1]->start;
8739                 else
8740                         BUG_ON(root->root_key.objectid !=
8741                                btrfs_header_owner(path->nodes[level + 1]));
8742         }
8743
8744         btrfs_free_tree_block(trans, root, eb, parent, wc->refs[level] == 1);
8745 out:
8746         wc->refs[level] = 0;
8747         wc->flags[level] = 0;
8748         return 0;
8749 }
8750
8751 static noinline int walk_down_tree(struct btrfs_trans_handle *trans,
8752                                    struct btrfs_root *root,
8753                                    struct btrfs_path *path,
8754                                    struct walk_control *wc)
8755 {
8756         int level = wc->level;
8757         int lookup_info = 1;
8758         int ret;
8759
8760         while (level >= 0) {
8761                 ret = walk_down_proc(trans, root, path, wc, lookup_info);
8762                 if (ret > 0)
8763                         break;
8764
8765                 if (level == 0)
8766                         break;
8767
8768                 if (path->slots[level] >=
8769                     btrfs_header_nritems(path->nodes[level]))
8770                         break;
8771
8772                 ret = do_walk_down(trans, root, path, wc, &lookup_info);
8773                 if (ret > 0) {
8774                         path->slots[level]++;
8775                         continue;
8776                 } else if (ret < 0)
8777                         return ret;
8778                 level = wc->level;
8779         }
8780         return 0;
8781 }
8782
8783 static noinline int walk_up_tree(struct btrfs_trans_handle *trans,
8784                                  struct btrfs_root *root,
8785                                  struct btrfs_path *path,
8786                                  struct walk_control *wc, int max_level)
8787 {
8788         int level = wc->level;
8789         int ret;
8790
8791         path->slots[level] = btrfs_header_nritems(path->nodes[level]);
8792         while (level < max_level && path->nodes[level]) {
8793                 wc->level = level;
8794                 if (path->slots[level] + 1 <
8795                     btrfs_header_nritems(path->nodes[level])) {
8796                         path->slots[level]++;
8797                         return 0;
8798                 } else {
8799                         ret = walk_up_proc(trans, root, path, wc);
8800                         if (ret > 0)
8801                                 return 0;
8802
8803                         if (path->locks[level]) {
8804                                 btrfs_tree_unlock_rw(path->nodes[level],
8805                                                      path->locks[level]);
8806                                 path->locks[level] = 0;
8807                         }
8808                         free_extent_buffer(path->nodes[level]);
8809                         path->nodes[level] = NULL;
8810                         level++;
8811                 }
8812         }
8813         return 1;
8814 }
8815
8816 /*
8817  * drop a subvolume tree.
8818  *
8819  * this function traverses the tree freeing any blocks that only
8820  * referenced by the tree.
8821  *
8822  * when a shared tree block is found. this function decreases its
8823  * reference count by one. if update_ref is true, this function
8824  * also make sure backrefs for the shared block and all lower level
8825  * blocks are properly updated.
8826  *
8827  * If called with for_reloc == 0, may exit early with -EAGAIN
8828  */
8829 int btrfs_drop_snapshot(struct btrfs_root *root,
8830                          struct btrfs_block_rsv *block_rsv, int update_ref,
8831                          int for_reloc)
8832 {
8833         struct btrfs_path *path;
8834         struct btrfs_trans_handle *trans;
8835         struct btrfs_root *tree_root = root->fs_info->tree_root;
8836         struct btrfs_root_item *root_item = &root->root_item;
8837         struct walk_control *wc;
8838         struct btrfs_key key;
8839         int err = 0;
8840         int ret;
8841         int level;
8842         bool root_dropped = false;
8843
8844         btrfs_debug(root->fs_info, "Drop subvolume %llu", root->objectid);
8845
8846         path = btrfs_alloc_path();
8847         if (!path) {
8848                 err = -ENOMEM;
8849                 goto out;
8850         }
8851
8852         wc = kzalloc(sizeof(*wc), GFP_NOFS);
8853         if (!wc) {
8854                 btrfs_free_path(path);
8855                 err = -ENOMEM;
8856                 goto out;
8857         }
8858
8859         trans = btrfs_start_transaction(tree_root, 0);
8860         if (IS_ERR(trans)) {
8861                 err = PTR_ERR(trans);
8862                 goto out_free;
8863         }
8864
8865         if (block_rsv)
8866                 trans->block_rsv = block_rsv;
8867
8868         if (btrfs_disk_key_objectid(&root_item->drop_progress) == 0) {
8869                 level = btrfs_header_level(root->node);
8870                 path->nodes[level] = btrfs_lock_root_node(root);
8871                 btrfs_set_lock_blocking(path->nodes[level]);
8872                 path->slots[level] = 0;
8873                 path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
8874                 memset(&wc->update_progress, 0,
8875                        sizeof(wc->update_progress));
8876         } else {
8877                 btrfs_disk_key_to_cpu(&key, &root_item->drop_progress);
8878                 memcpy(&wc->update_progress, &key,
8879                        sizeof(wc->update_progress));
8880
8881                 level = root_item->drop_level;
8882                 BUG_ON(level == 0);
8883                 path->lowest_level = level;
8884                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
8885                 path->lowest_level = 0;
8886                 if (ret < 0) {
8887                         err = ret;
8888                         goto out_end_trans;
8889                 }
8890                 WARN_ON(ret > 0);
8891
8892                 /*
8893                  * unlock our path, this is safe because only this
8894                  * function is allowed to delete this snapshot
8895                  */
8896                 btrfs_unlock_up_safe(path, 0);
8897
8898                 level = btrfs_header_level(root->node);
8899                 while (1) {
8900                         btrfs_tree_lock(path->nodes[level]);
8901                         btrfs_set_lock_blocking(path->nodes[level]);
8902                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
8903
8904                         ret = btrfs_lookup_extent_info(trans, root,
8905                                                 path->nodes[level]->start,
8906                                                 level, 1, &wc->refs[level],
8907                                                 &wc->flags[level]);
8908                         if (ret < 0) {
8909                                 err = ret;
8910                                 goto out_end_trans;
8911                         }
8912                         BUG_ON(wc->refs[level] == 0);
8913
8914                         if (level == root_item->drop_level)
8915                                 break;
8916
8917                         btrfs_tree_unlock(path->nodes[level]);
8918                         path->locks[level] = 0;
8919                         WARN_ON(wc->refs[level] != 1);
8920                         level--;
8921                 }
8922         }
8923
8924         wc->level = level;
8925         wc->shared_level = -1;
8926         wc->stage = DROP_REFERENCE;
8927         wc->update_ref = update_ref;
8928         wc->keep_locks = 0;
8929         wc->for_reloc = for_reloc;
8930         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(root);
8931
8932         while (1) {
8933
8934                 ret = walk_down_tree(trans, root, path, wc);
8935                 if (ret < 0) {
8936                         err = ret;
8937                         break;
8938                 }
8939
8940                 ret = walk_up_tree(trans, root, path, wc, BTRFS_MAX_LEVEL);
8941                 if (ret < 0) {
8942                         err = ret;
8943                         break;
8944                 }
8945
8946                 if (ret > 0) {
8947                         BUG_ON(wc->stage != DROP_REFERENCE);
8948                         break;
8949                 }
8950
8951                 if (wc->stage == DROP_REFERENCE) {
8952                         level = wc->level;
8953                         btrfs_node_key(path->nodes[level],
8954                                        &root_item->drop_progress,
8955                                        path->slots[level]);
8956                         root_item->drop_level = level;
8957                 }
8958
8959                 BUG_ON(wc->level == 0);
8960                 if (btrfs_should_end_transaction(trans, tree_root) ||
8961                     (!for_reloc && btrfs_need_cleaner_sleep(root))) {
8962                         ret = btrfs_update_root(trans, tree_root,
8963                                                 &root->root_key,
8964                                                 root_item);
8965                         if (ret) {
8966                                 btrfs_abort_transaction(trans, tree_root, ret);
8967                                 err = ret;
8968                                 goto out_end_trans;
8969                         }
8970
8971                         btrfs_end_transaction_throttle(trans, tree_root);
8972                         if (!for_reloc && btrfs_need_cleaner_sleep(root)) {
8973                                 pr_debug("BTRFS: drop snapshot early exit\n");
8974                                 err = -EAGAIN;
8975                                 goto out_free;
8976                         }
8977
8978                         trans = btrfs_start_transaction(tree_root, 0);
8979                         if (IS_ERR(trans)) {
8980                                 err = PTR_ERR(trans);
8981                                 goto out_free;
8982                         }
8983                         if (block_rsv)
8984                                 trans->block_rsv = block_rsv;
8985                 }
8986         }
8987         btrfs_release_path(path);
8988         if (err)
8989                 goto out_end_trans;
8990
8991         ret = btrfs_del_root(trans, tree_root, &root->root_key);
8992         if (ret) {
8993                 btrfs_abort_transaction(trans, tree_root, ret);
8994                 goto out_end_trans;
8995         }
8996
8997         if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) {
8998                 ret = btrfs_find_root(tree_root, &root->root_key, path,
8999                                       NULL, NULL);
9000                 if (ret < 0) {
9001                         btrfs_abort_transaction(trans, tree_root, ret);
9002                         err = ret;
9003                         goto out_end_trans;
9004                 } else if (ret > 0) {
9005                         /* if we fail to delete the orphan item this time
9006                          * around, it'll get picked up the next time.
9007                          *
9008                          * The most common failure here is just -ENOENT.
9009                          */
9010                         btrfs_del_orphan_item(trans, tree_root,
9011                                               root->root_key.objectid);
9012                 }
9013         }
9014
9015         if (test_bit(BTRFS_ROOT_IN_RADIX, &root->state)) {
9016                 btrfs_add_dropped_root(trans, root);
9017         } else {
9018                 free_extent_buffer(root->node);
9019                 free_extent_buffer(root->commit_root);
9020                 btrfs_put_fs_root(root);
9021         }
9022         root_dropped = true;
9023 out_end_trans:
9024         btrfs_end_transaction_throttle(trans, tree_root);
9025 out_free:
9026         kfree(wc);
9027         btrfs_free_path(path);
9028 out:
9029         /*
9030          * So if we need to stop dropping the snapshot for whatever reason we
9031          * need to make sure to add it back to the dead root list so that we
9032          * keep trying to do the work later.  This also cleans up roots if we
9033          * don't have it in the radix (like when we recover after a power fail
9034          * or unmount) so we don't leak memory.
9035          */
9036         if (!for_reloc && root_dropped == false)
9037                 btrfs_add_dead_root(root);
9038         if (err && err != -EAGAIN)
9039                 btrfs_std_error(root->fs_info, err, NULL);
9040         return err;
9041 }
9042
9043 /*
9044  * drop subtree rooted at tree block 'node'.
9045  *
9046  * NOTE: this function will unlock and release tree block 'node'
9047  * only used by relocation code
9048  */
9049 int btrfs_drop_subtree(struct btrfs_trans_handle *trans,
9050                         struct btrfs_root *root,
9051                         struct extent_buffer *node,
9052                         struct extent_buffer *parent)
9053 {
9054         struct btrfs_path *path;
9055         struct walk_control *wc;
9056         int level;
9057         int parent_level;
9058         int ret = 0;
9059         int wret;
9060
9061         BUG_ON(root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID);
9062
9063         path = btrfs_alloc_path();
9064         if (!path)
9065                 return -ENOMEM;
9066
9067         wc = kzalloc(sizeof(*wc), GFP_NOFS);
9068         if (!wc) {
9069                 btrfs_free_path(path);
9070                 return -ENOMEM;
9071         }
9072
9073         btrfs_assert_tree_locked(parent);
9074         parent_level = btrfs_header_level(parent);
9075         extent_buffer_get(parent);
9076         path->nodes[parent_level] = parent;
9077         path->slots[parent_level] = btrfs_header_nritems(parent);
9078
9079         btrfs_assert_tree_locked(node);
9080         level = btrfs_header_level(node);
9081         path->nodes[level] = node;
9082         path->slots[level] = 0;
9083         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
9084
9085         wc->refs[parent_level] = 1;
9086         wc->flags[parent_level] = BTRFS_BLOCK_FLAG_FULL_BACKREF;
9087         wc->level = level;
9088         wc->shared_level = -1;
9089         wc->stage = DROP_REFERENCE;
9090         wc->update_ref = 0;
9091         wc->keep_locks = 1;
9092         wc->for_reloc = 1;
9093         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(root);
9094
9095         while (1) {
9096                 wret = walk_down_tree(trans, root, path, wc);
9097                 if (wret < 0) {
9098                         ret = wret;
9099                         break;
9100                 }
9101
9102                 wret = walk_up_tree(trans, root, path, wc, parent_level);
9103                 if (wret < 0)
9104                         ret = wret;
9105                 if (wret != 0)
9106                         break;
9107         }
9108
9109         kfree(wc);
9110         btrfs_free_path(path);
9111         return ret;
9112 }
9113
9114 static u64 update_block_group_flags(struct btrfs_root *root, u64 flags)
9115 {
9116         u64 num_devices;
9117         u64 stripped;
9118
9119         /*
9120          * if restripe for this chunk_type is on pick target profile and
9121          * return, otherwise do the usual balance
9122          */
9123         stripped = get_restripe_target(root->fs_info, flags);
9124         if (stripped)
9125                 return extended_to_chunk(stripped);
9126
9127         num_devices = root->fs_info->fs_devices->rw_devices;
9128
9129         stripped = BTRFS_BLOCK_GROUP_RAID0 |
9130                 BTRFS_BLOCK_GROUP_RAID5 | BTRFS_BLOCK_GROUP_RAID6 |
9131                 BTRFS_BLOCK_GROUP_RAID1 | BTRFS_BLOCK_GROUP_RAID10;
9132
9133         if (num_devices == 1) {
9134                 stripped |= BTRFS_BLOCK_GROUP_DUP;
9135                 stripped = flags & ~stripped;
9136
9137                 /* turn raid0 into single device chunks */
9138                 if (flags & BTRFS_BLOCK_GROUP_RAID0)
9139                         return stripped;
9140
9141                 /* turn mirroring into duplication */
9142                 if (flags & (BTRFS_BLOCK_GROUP_RAID1 |
9143                              BTRFS_BLOCK_GROUP_RAID10))
9144                         return stripped | BTRFS_BLOCK_GROUP_DUP;
9145         } else {
9146                 /* they already had raid on here, just return */
9147                 if (flags & stripped)
9148                         return flags;
9149
9150                 stripped |= BTRFS_BLOCK_GROUP_DUP;
9151                 stripped = flags & ~stripped;
9152
9153                 /* switch duplicated blocks with raid1 */
9154                 if (flags & BTRFS_BLOCK_GROUP_DUP)
9155                         return stripped | BTRFS_BLOCK_GROUP_RAID1;
9156
9157                 /* this is drive concat, leave it alone */
9158         }
9159
9160         return flags;
9161 }
9162
9163 static int inc_block_group_ro(struct btrfs_block_group_cache *cache, int force)
9164 {
9165         struct btrfs_space_info *sinfo = cache->space_info;
9166         u64 num_bytes;
9167         u64 min_allocable_bytes;
9168         int ret = -ENOSPC;
9169
9170         /*
9171          * We need some metadata space and system metadata space for
9172          * allocating chunks in some corner cases until we force to set
9173          * it to be readonly.
9174          */
9175         if ((sinfo->flags &
9176              (BTRFS_BLOCK_GROUP_SYSTEM | BTRFS_BLOCK_GROUP_METADATA)) &&
9177             !force)
9178                 min_allocable_bytes = 1 * 1024 * 1024;
9179         else
9180                 min_allocable_bytes = 0;
9181
9182         spin_lock(&sinfo->lock);
9183         spin_lock(&cache->lock);
9184
9185         if (cache->ro) {
9186                 cache->ro++;
9187                 ret = 0;
9188                 goto out;
9189         }
9190
9191         num_bytes = cache->key.offset - cache->reserved - cache->pinned -
9192                     cache->bytes_super - btrfs_block_group_used(&cache->item);
9193
9194         if (sinfo->bytes_used + sinfo->bytes_reserved + sinfo->bytes_pinned +
9195             sinfo->bytes_may_use + sinfo->bytes_readonly + num_bytes +
9196             min_allocable_bytes <= sinfo->total_bytes) {
9197                 sinfo->bytes_readonly += num_bytes;
9198                 cache->ro++;
9199                 list_add_tail(&cache->ro_list, &sinfo->ro_bgs);
9200                 ret = 0;
9201         }
9202 out:
9203         spin_unlock(&cache->lock);
9204         spin_unlock(&sinfo->lock);
9205         return ret;
9206 }
9207
9208 int btrfs_inc_block_group_ro(struct btrfs_root *root,
9209                              struct btrfs_block_group_cache *cache)
9210
9211 {
9212         struct btrfs_trans_handle *trans;
9213         u64 alloc_flags;
9214         int ret;
9215
9216 again:
9217         trans = btrfs_join_transaction(root);
9218         if (IS_ERR(trans))
9219                 return PTR_ERR(trans);
9220
9221         /*
9222          * we're not allowed to set block groups readonly after the dirty
9223          * block groups cache has started writing.  If it already started,
9224          * back off and let this transaction commit
9225          */
9226         mutex_lock(&root->fs_info->ro_block_group_mutex);
9227         if (test_bit(BTRFS_TRANS_DIRTY_BG_RUN, &trans->transaction->flags)) {
9228                 u64 transid = trans->transid;
9229
9230                 mutex_unlock(&root->fs_info->ro_block_group_mutex);
9231                 btrfs_end_transaction(trans, root);
9232
9233                 ret = btrfs_wait_for_commit(root, transid);
9234                 if (ret)
9235                         return ret;
9236                 goto again;
9237         }
9238
9239         /*
9240          * if we are changing raid levels, try to allocate a corresponding
9241          * block group with the new raid level.
9242          */
9243         alloc_flags = update_block_group_flags(root, cache->flags);
9244         if (alloc_flags != cache->flags) {
9245                 ret = do_chunk_alloc(trans, root, alloc_flags,
9246                                      CHUNK_ALLOC_FORCE);
9247                 /*
9248                  * ENOSPC is allowed here, we may have enough space
9249                  * already allocated at the new raid level to
9250                  * carry on
9251                  */
9252                 if (ret == -ENOSPC)
9253                         ret = 0;
9254                 if (ret < 0)
9255                         goto out;
9256         }
9257
9258         ret = inc_block_group_ro(cache, 0);
9259         if (!ret)
9260                 goto out;
9261         alloc_flags = get_alloc_profile(root, cache->space_info->flags);
9262         ret = do_chunk_alloc(trans, root, alloc_flags,
9263                              CHUNK_ALLOC_FORCE);
9264         if (ret < 0)
9265                 goto out;
9266         ret = inc_block_group_ro(cache, 0);
9267 out:
9268         if (cache->flags & BTRFS_BLOCK_GROUP_SYSTEM) {
9269                 alloc_flags = update_block_group_flags(root, cache->flags);
9270                 lock_chunks(root->fs_info->chunk_root);
9271                 check_system_chunk(trans, root, alloc_flags);
9272                 unlock_chunks(root->fs_info->chunk_root);
9273         }
9274         mutex_unlock(&root->fs_info->ro_block_group_mutex);
9275
9276         btrfs_end_transaction(trans, root);
9277         return ret;
9278 }
9279
9280 int btrfs_force_chunk_alloc(struct btrfs_trans_handle *trans,
9281                             struct btrfs_root *root, u64 type)
9282 {
9283         u64 alloc_flags = get_alloc_profile(root, type);
9284         return do_chunk_alloc(trans, root, alloc_flags,
9285                               CHUNK_ALLOC_FORCE);
9286 }
9287
9288 /*
9289  * helper to account the unused space of all the readonly block group in the
9290  * space_info. takes mirrors into account.
9291  */
9292 u64 btrfs_account_ro_block_groups_free_space(struct btrfs_space_info *sinfo)
9293 {
9294         struct btrfs_block_group_cache *block_group;
9295         u64 free_bytes = 0;
9296         int factor;
9297
9298         /* It's df, we don't care if it's racey */
9299         if (list_empty(&sinfo->ro_bgs))
9300                 return 0;
9301
9302         spin_lock(&sinfo->lock);
9303         list_for_each_entry(block_group, &sinfo->ro_bgs, ro_list) {
9304                 spin_lock(&block_group->lock);
9305
9306                 if (!block_group->ro) {
9307                         spin_unlock(&block_group->lock);
9308                         continue;
9309                 }
9310
9311                 if (block_group->flags & (BTRFS_BLOCK_GROUP_RAID1 |
9312                                           BTRFS_BLOCK_GROUP_RAID10 |
9313                                           BTRFS_BLOCK_GROUP_DUP))
9314                         factor = 2;
9315                 else
9316                         factor = 1;
9317
9318                 free_bytes += (block_group->key.offset -
9319                                btrfs_block_group_used(&block_group->item)) *
9320                                factor;
9321
9322                 spin_unlock(&block_group->lock);
9323         }
9324         spin_unlock(&sinfo->lock);
9325
9326         return free_bytes;
9327 }
9328
9329 void btrfs_dec_block_group_ro(struct btrfs_root *root,
9330                               struct btrfs_block_group_cache *cache)
9331 {
9332         struct btrfs_space_info *sinfo = cache->space_info;
9333         u64 num_bytes;
9334
9335         BUG_ON(!cache->ro);
9336
9337         spin_lock(&sinfo->lock);
9338         spin_lock(&cache->lock);
9339         if (!--cache->ro) {
9340                 num_bytes = cache->key.offset - cache->reserved -
9341                             cache->pinned - cache->bytes_super -
9342                             btrfs_block_group_used(&cache->item);
9343                 sinfo->bytes_readonly -= num_bytes;
9344                 list_del_init(&cache->ro_list);
9345         }
9346         spin_unlock(&cache->lock);
9347         spin_unlock(&sinfo->lock);
9348 }
9349
9350 /*
9351  * checks to see if its even possible to relocate this block group.
9352  *
9353  * @return - -1 if it's not a good idea to relocate this block group, 0 if its
9354  * ok to go ahead and try.
9355  */
9356 int btrfs_can_relocate(struct btrfs_root *root, u64 bytenr)
9357 {
9358         struct btrfs_block_group_cache *block_group;
9359         struct btrfs_space_info *space_info;
9360         struct btrfs_fs_devices *fs_devices = root->fs_info->fs_devices;
9361         struct btrfs_device *device;
9362         struct btrfs_trans_handle *trans;
9363         u64 min_free;
9364         u64 dev_min = 1;
9365         u64 dev_nr = 0;
9366         u64 target;
9367         int index;
9368         int full = 0;
9369         int ret = 0;
9370
9371         block_group = btrfs_lookup_block_group(root->fs_info, bytenr);
9372
9373         /* odd, couldn't find the block group, leave it alone */
9374         if (!block_group)
9375                 return -1;
9376
9377         min_free = btrfs_block_group_used(&block_group->item);
9378
9379         /* no bytes used, we're good */
9380         if (!min_free)
9381                 goto out;
9382
9383         space_info = block_group->space_info;
9384         spin_lock(&space_info->lock);
9385
9386         full = space_info->full;
9387
9388         /*
9389          * if this is the last block group we have in this space, we can't
9390          * relocate it unless we're able to allocate a new chunk below.
9391          *
9392          * Otherwise, we need to make sure we have room in the space to handle
9393          * all of the extents from this block group.  If we can, we're good
9394          */
9395         if ((space_info->total_bytes != block_group->key.offset) &&
9396             (space_info->bytes_used + space_info->bytes_reserved +
9397              space_info->bytes_pinned + space_info->bytes_readonly +
9398              min_free < space_info->total_bytes)) {
9399                 spin_unlock(&space_info->lock);
9400                 goto out;
9401         }
9402         spin_unlock(&space_info->lock);
9403
9404         /*
9405          * ok we don't have enough space, but maybe we have free space on our
9406          * devices to allocate new chunks for relocation, so loop through our
9407          * alloc devices and guess if we have enough space.  if this block
9408          * group is going to be restriped, run checks against the target
9409          * profile instead of the current one.
9410          */
9411         ret = -1;
9412
9413         /*
9414          * index:
9415          *      0: raid10
9416          *      1: raid1
9417          *      2: dup
9418          *      3: raid0
9419          *      4: single
9420          */
9421         target = get_restripe_target(root->fs_info, block_group->flags);
9422         if (target) {
9423                 index = __get_raid_index(extended_to_chunk(target));
9424         } else {
9425                 /*
9426                  * this is just a balance, so if we were marked as full
9427                  * we know there is no space for a new chunk
9428                  */
9429                 if (full)
9430                         goto out;
9431
9432                 index = get_block_group_index(block_group);
9433         }
9434
9435         if (index == BTRFS_RAID_RAID10) {
9436                 dev_min = 4;
9437                 /* Divide by 2 */
9438                 min_free >>= 1;
9439         } else if (index == BTRFS_RAID_RAID1) {
9440                 dev_min = 2;
9441         } else if (index == BTRFS_RAID_DUP) {
9442                 /* Multiply by 2 */
9443                 min_free <<= 1;
9444         } else if (index == BTRFS_RAID_RAID0) {
9445                 dev_min = fs_devices->rw_devices;
9446                 min_free = div64_u64(min_free, dev_min);
9447         }
9448
9449         /* We need to do this so that we can look at pending chunks */
9450         trans = btrfs_join_transaction(root);
9451         if (IS_ERR(trans)) {
9452                 ret = PTR_ERR(trans);
9453                 goto out;
9454         }
9455
9456         mutex_lock(&root->fs_info->chunk_mutex);
9457         list_for_each_entry(device, &fs_devices->alloc_list, dev_alloc_list) {
9458                 u64 dev_offset;
9459
9460                 /*
9461                  * check to make sure we can actually find a chunk with enough
9462                  * space to fit our block group in.
9463                  */
9464                 if (device->total_bytes > device->bytes_used + min_free &&
9465                     !device->is_tgtdev_for_dev_replace) {
9466                         ret = find_free_dev_extent(trans, device, min_free,
9467                                                    &dev_offset, NULL);
9468                         if (!ret)
9469                                 dev_nr++;
9470
9471                         if (dev_nr >= dev_min)
9472                                 break;
9473
9474                         ret = -1;
9475                 }
9476         }
9477         mutex_unlock(&root->fs_info->chunk_mutex);
9478         btrfs_end_transaction(trans, root);
9479 out:
9480         btrfs_put_block_group(block_group);
9481         return ret;
9482 }
9483
9484 static int find_first_block_group(struct btrfs_root *root,
9485                 struct btrfs_path *path, struct btrfs_key *key)
9486 {
9487         int ret = 0;
9488         struct btrfs_key found_key;
9489         struct extent_buffer *leaf;
9490         int slot;
9491
9492         ret = btrfs_search_slot(NULL, root, key, path, 0, 0);
9493         if (ret < 0)
9494                 goto out;
9495
9496         while (1) {
9497                 slot = path->slots[0];
9498                 leaf = path->nodes[0];
9499                 if (slot >= btrfs_header_nritems(leaf)) {
9500                         ret = btrfs_next_leaf(root, path);
9501                         if (ret == 0)
9502                                 continue;
9503                         if (ret < 0)
9504                                 goto out;
9505                         break;
9506                 }
9507                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
9508
9509                 if (found_key.objectid >= key->objectid &&
9510                     found_key.type == BTRFS_BLOCK_GROUP_ITEM_KEY) {
9511                         ret = 0;
9512                         goto out;
9513                 }
9514                 path->slots[0]++;
9515         }
9516 out:
9517         return ret;
9518 }
9519
9520 void btrfs_put_block_group_cache(struct btrfs_fs_info *info)
9521 {
9522         struct btrfs_block_group_cache *block_group;
9523         u64 last = 0;
9524
9525         while (1) {
9526                 struct inode *inode;
9527
9528                 block_group = btrfs_lookup_first_block_group(info, last);
9529                 while (block_group) {
9530                         spin_lock(&block_group->lock);
9531                         if (block_group->iref)
9532                                 break;
9533                         spin_unlock(&block_group->lock);
9534                         block_group = next_block_group(info->tree_root,
9535                                                        block_group);
9536                 }
9537                 if (!block_group) {
9538                         if (last == 0)
9539                                 break;
9540                         last = 0;
9541                         continue;
9542                 }
9543
9544                 inode = block_group->inode;
9545                 block_group->iref = 0;
9546                 block_group->inode = NULL;
9547                 spin_unlock(&block_group->lock);
9548                 iput(inode);
9549                 last = block_group->key.objectid + block_group->key.offset;
9550                 btrfs_put_block_group(block_group);
9551         }
9552 }
9553
9554 int btrfs_free_block_groups(struct btrfs_fs_info *info)
9555 {
9556         struct btrfs_block_group_cache *block_group;
9557         struct btrfs_space_info *space_info;
9558         struct btrfs_caching_control *caching_ctl;
9559         struct rb_node *n;
9560
9561         down_write(&info->commit_root_sem);
9562         while (!list_empty(&info->caching_block_groups)) {
9563                 caching_ctl = list_entry(info->caching_block_groups.next,
9564                                          struct btrfs_caching_control, list);
9565                 list_del(&caching_ctl->list);
9566                 put_caching_control(caching_ctl);
9567         }
9568         up_write(&info->commit_root_sem);
9569
9570         spin_lock(&info->unused_bgs_lock);
9571         while (!list_empty(&info->unused_bgs)) {
9572                 block_group = list_first_entry(&info->unused_bgs,
9573                                                struct btrfs_block_group_cache,
9574                                                bg_list);
9575                 list_del_init(&block_group->bg_list);
9576                 btrfs_put_block_group(block_group);
9577         }
9578         spin_unlock(&info->unused_bgs_lock);
9579
9580         spin_lock(&info->block_group_cache_lock);
9581         while ((n = rb_last(&info->block_group_cache_tree)) != NULL) {
9582                 block_group = rb_entry(n, struct btrfs_block_group_cache,
9583                                        cache_node);
9584                 rb_erase(&block_group->cache_node,
9585                          &info->block_group_cache_tree);
9586                 RB_CLEAR_NODE(&block_group->cache_node);
9587                 spin_unlock(&info->block_group_cache_lock);
9588
9589                 down_write(&block_group->space_info->groups_sem);
9590                 list_del(&block_group->list);
9591                 up_write(&block_group->space_info->groups_sem);
9592
9593                 if (block_group->cached == BTRFS_CACHE_STARTED)
9594                         wait_block_group_cache_done(block_group);
9595
9596                 /*
9597                  * We haven't cached this block group, which means we could
9598                  * possibly have excluded extents on this block group.
9599                  */
9600                 if (block_group->cached == BTRFS_CACHE_NO ||
9601                     block_group->cached == BTRFS_CACHE_ERROR)
9602                         free_excluded_extents(info->extent_root, block_group);
9603
9604                 btrfs_remove_free_space_cache(block_group);
9605                 btrfs_put_block_group(block_group);
9606
9607                 spin_lock(&info->block_group_cache_lock);
9608         }
9609         spin_unlock(&info->block_group_cache_lock);
9610
9611         /* now that all the block groups are freed, go through and
9612          * free all the space_info structs.  This is only called during
9613          * the final stages of unmount, and so we know nobody is
9614          * using them.  We call synchronize_rcu() once before we start,
9615          * just to be on the safe side.
9616          */
9617         synchronize_rcu();
9618
9619         release_global_block_rsv(info);
9620
9621         while (!list_empty(&info->space_info)) {
9622                 int i;
9623
9624                 space_info = list_entry(info->space_info.next,
9625                                         struct btrfs_space_info,
9626                                         list);
9627                 if (btrfs_test_opt(info->tree_root, ENOSPC_DEBUG)) {
9628                         if (WARN_ON(space_info->bytes_pinned > 0 ||
9629                             space_info->bytes_reserved > 0 ||
9630                             space_info->bytes_may_use > 0)) {
9631                                 dump_space_info(space_info, 0, 0);
9632                         }
9633                 }
9634                 list_del(&space_info->list);
9635                 for (i = 0; i < BTRFS_NR_RAID_TYPES; i++) {
9636                         struct kobject *kobj;
9637                         kobj = space_info->block_group_kobjs[i];
9638                         space_info->block_group_kobjs[i] = NULL;
9639                         if (kobj) {
9640                                 kobject_del(kobj);
9641                                 kobject_put(kobj);
9642                         }
9643                 }
9644                 kobject_del(&space_info->kobj);
9645                 kobject_put(&space_info->kobj);
9646         }
9647         return 0;
9648 }
9649
9650 static void __link_block_group(struct btrfs_space_info *space_info,
9651                                struct btrfs_block_group_cache *cache)
9652 {
9653         int index = get_block_group_index(cache);
9654         bool first = false;
9655
9656         down_write(&space_info->groups_sem);
9657         if (list_empty(&space_info->block_groups[index]))
9658                 first = true;
9659         list_add_tail(&cache->list, &space_info->block_groups[index]);
9660         up_write(&space_info->groups_sem);
9661
9662         if (first) {
9663                 struct raid_kobject *rkobj;
9664                 int ret;
9665
9666                 rkobj = kzalloc(sizeof(*rkobj), GFP_NOFS);
9667                 if (!rkobj)
9668                         goto out_err;
9669                 rkobj->raid_type = index;
9670                 kobject_init(&rkobj->kobj, &btrfs_raid_ktype);
9671                 ret = kobject_add(&rkobj->kobj, &space_info->kobj,
9672                                   "%s", get_raid_name(index));
9673                 if (ret) {
9674                         kobject_put(&rkobj->kobj);
9675                         goto out_err;
9676                 }
9677                 space_info->block_group_kobjs[index] = &rkobj->kobj;
9678         }
9679
9680         return;
9681 out_err:
9682         pr_warn("BTRFS: failed to add kobject for block cache. ignoring.\n");
9683 }
9684
9685 static struct btrfs_block_group_cache *
9686 btrfs_create_block_group_cache(struct btrfs_root *root, u64 start, u64 size)
9687 {
9688         struct btrfs_block_group_cache *cache;
9689
9690         cache = kzalloc(sizeof(*cache), GFP_NOFS);
9691         if (!cache)
9692                 return NULL;
9693
9694         cache->free_space_ctl = kzalloc(sizeof(*cache->free_space_ctl),
9695                                         GFP_NOFS);
9696         if (!cache->free_space_ctl) {
9697                 kfree(cache);
9698                 return NULL;
9699         }
9700
9701         cache->key.objectid = start;
9702         cache->key.offset = size;
9703         cache->key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
9704
9705         cache->sectorsize = root->sectorsize;
9706         cache->fs_info = root->fs_info;
9707         cache->full_stripe_len = btrfs_full_stripe_len(root,
9708                                                &root->fs_info->mapping_tree,
9709                                                start);
9710         set_free_space_tree_thresholds(cache);
9711
9712         atomic_set(&cache->count, 1);
9713         spin_lock_init(&cache->lock);
9714         init_rwsem(&cache->data_rwsem);
9715         INIT_LIST_HEAD(&cache->list);
9716         INIT_LIST_HEAD(&cache->cluster_list);
9717         INIT_LIST_HEAD(&cache->bg_list);
9718         INIT_LIST_HEAD(&cache->ro_list);
9719         INIT_LIST_HEAD(&cache->dirty_list);
9720         INIT_LIST_HEAD(&cache->io_list);
9721         btrfs_init_free_space_ctl(cache);
9722         atomic_set(&cache->trimming, 0);
9723         mutex_init(&cache->free_space_lock);
9724
9725         return cache;
9726 }
9727
9728 int btrfs_read_block_groups(struct btrfs_root *root)
9729 {
9730         struct btrfs_path *path;
9731         int ret;
9732         struct btrfs_block_group_cache *cache;
9733         struct btrfs_fs_info *info = root->fs_info;
9734         struct btrfs_space_info *space_info;
9735         struct btrfs_key key;
9736         struct btrfs_key found_key;
9737         struct extent_buffer *leaf;
9738         int need_clear = 0;
9739         u64 cache_gen;
9740
9741         root = info->extent_root;
9742         key.objectid = 0;
9743         key.offset = 0;
9744         key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
9745         path = btrfs_alloc_path();
9746         if (!path)
9747                 return -ENOMEM;
9748         path->reada = 1;
9749
9750         cache_gen = btrfs_super_cache_generation(root->fs_info->super_copy);
9751         if (btrfs_test_opt(root, SPACE_CACHE) &&
9752             btrfs_super_generation(root->fs_info->super_copy) != cache_gen)
9753                 need_clear = 1;
9754         if (btrfs_test_opt(root, CLEAR_CACHE))
9755                 need_clear = 1;
9756
9757         while (1) {
9758                 ret = find_first_block_group(root, path, &key);
9759                 if (ret > 0)
9760                         break;
9761                 if (ret != 0)
9762                         goto error;
9763
9764                 leaf = path->nodes[0];
9765                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
9766
9767                 cache = btrfs_create_block_group_cache(root, found_key.objectid,
9768                                                        found_key.offset);
9769                 if (!cache) {
9770                         ret = -ENOMEM;
9771                         goto error;
9772                 }
9773
9774                 if (need_clear) {
9775                         /*
9776                          * When we mount with old space cache, we need to
9777                          * set BTRFS_DC_CLEAR and set dirty flag.
9778                          *
9779                          * a) Setting 'BTRFS_DC_CLEAR' makes sure that we
9780                          *    truncate the old free space cache inode and
9781                          *    setup a new one.
9782                          * b) Setting 'dirty flag' makes sure that we flush
9783                          *    the new space cache info onto disk.
9784                          */
9785                         if (btrfs_test_opt(root, SPACE_CACHE))
9786                                 cache->disk_cache_state = BTRFS_DC_CLEAR;
9787                 }
9788
9789                 read_extent_buffer(leaf, &cache->item,
9790                                    btrfs_item_ptr_offset(leaf, path->slots[0]),
9791                                    sizeof(cache->item));
9792                 cache->flags = btrfs_block_group_flags(&cache->item);
9793
9794                 key.objectid = found_key.objectid + found_key.offset;
9795                 btrfs_release_path(path);
9796
9797                 /*
9798                  * We need to exclude the super stripes now so that the space
9799                  * info has super bytes accounted for, otherwise we'll think
9800                  * we have more space than we actually do.
9801                  */
9802                 ret = exclude_super_stripes(root, cache);
9803                 if (ret) {
9804                         /*
9805                          * We may have excluded something, so call this just in
9806                          * case.
9807                          */
9808                         free_excluded_extents(root, cache);
9809                         btrfs_put_block_group(cache);
9810                         goto error;
9811                 }
9812
9813                 /*
9814                  * check for two cases, either we are full, and therefore
9815                  * don't need to bother with the caching work since we won't
9816                  * find any space, or we are empty, and we can just add all
9817                  * the space in and be done with it.  This saves us _alot_ of
9818                  * time, particularly in the full case.
9819                  */
9820                 if (found_key.offset == btrfs_block_group_used(&cache->item)) {
9821                         cache->last_byte_to_unpin = (u64)-1;
9822                         cache->cached = BTRFS_CACHE_FINISHED;
9823                         free_excluded_extents(root, cache);
9824                 } else if (btrfs_block_group_used(&cache->item) == 0) {
9825                         cache->last_byte_to_unpin = (u64)-1;
9826                         cache->cached = BTRFS_CACHE_FINISHED;
9827                         add_new_free_space(cache, root->fs_info,
9828                                            found_key.objectid,
9829                                            found_key.objectid +
9830                                            found_key.offset);
9831                         free_excluded_extents(root, cache);
9832                 }
9833
9834                 ret = btrfs_add_block_group_cache(root->fs_info, cache);
9835                 if (ret) {
9836                         btrfs_remove_free_space_cache(cache);
9837                         btrfs_put_block_group(cache);
9838                         goto error;
9839                 }
9840
9841                 ret = update_space_info(info, cache->flags, found_key.offset,
9842                                         btrfs_block_group_used(&cache->item),
9843                                         &space_info);
9844                 if (ret) {
9845                         btrfs_remove_free_space_cache(cache);
9846                         spin_lock(&info->block_group_cache_lock);
9847                         rb_erase(&cache->cache_node,
9848                                  &info->block_group_cache_tree);
9849                         RB_CLEAR_NODE(&cache->cache_node);
9850                         spin_unlock(&info->block_group_cache_lock);
9851                         btrfs_put_block_group(cache);
9852                         goto error;
9853                 }
9854
9855                 cache->space_info = space_info;
9856                 spin_lock(&cache->space_info->lock);
9857                 cache->space_info->bytes_readonly += cache->bytes_super;
9858                 spin_unlock(&cache->space_info->lock);
9859
9860                 __link_block_group(space_info, cache);
9861
9862                 set_avail_alloc_bits(root->fs_info, cache->flags);
9863                 if (btrfs_chunk_readonly(root, cache->key.objectid)) {
9864                         inc_block_group_ro(cache, 1);
9865                 } else if (btrfs_block_group_used(&cache->item) == 0) {
9866                         spin_lock(&info->unused_bgs_lock);
9867                         /* Should always be true but just in case. */
9868                         if (list_empty(&cache->bg_list)) {
9869                                 btrfs_get_block_group(cache);
9870                                 list_add_tail(&cache->bg_list,
9871                                               &info->unused_bgs);
9872                         }
9873                         spin_unlock(&info->unused_bgs_lock);
9874                 }
9875         }
9876
9877         list_for_each_entry_rcu(space_info, &root->fs_info->space_info, list) {
9878                 if (!(get_alloc_profile(root, space_info->flags) &
9879                       (BTRFS_BLOCK_GROUP_RAID10 |
9880                        BTRFS_BLOCK_GROUP_RAID1 |
9881                        BTRFS_BLOCK_GROUP_RAID5 |
9882                        BTRFS_BLOCK_GROUP_RAID6 |
9883                        BTRFS_BLOCK_GROUP_DUP)))
9884                         continue;
9885                 /*
9886                  * avoid allocating from un-mirrored block group if there are
9887                  * mirrored block groups.
9888                  */
9889                 list_for_each_entry(cache,
9890                                 &space_info->block_groups[BTRFS_RAID_RAID0],
9891                                 list)
9892                         inc_block_group_ro(cache, 1);
9893                 list_for_each_entry(cache,
9894                                 &space_info->block_groups[BTRFS_RAID_SINGLE],
9895                                 list)
9896                         inc_block_group_ro(cache, 1);
9897         }
9898
9899         init_global_block_rsv(info);
9900         ret = 0;
9901 error:
9902         btrfs_free_path(path);
9903         return ret;
9904 }
9905
9906 void btrfs_create_pending_block_groups(struct btrfs_trans_handle *trans,
9907                                        struct btrfs_root *root)
9908 {
9909         struct btrfs_block_group_cache *block_group, *tmp;
9910         struct btrfs_root *extent_root = root->fs_info->extent_root;
9911         struct btrfs_block_group_item item;
9912         struct btrfs_key key;
9913         int ret = 0;
9914         bool can_flush_pending_bgs = trans->can_flush_pending_bgs;
9915
9916         trans->can_flush_pending_bgs = false;
9917         list_for_each_entry_safe(block_group, tmp, &trans->new_bgs, bg_list) {
9918                 if (ret)
9919                         goto next;
9920
9921                 spin_lock(&block_group->lock);
9922                 memcpy(&item, &block_group->item, sizeof(item));
9923                 memcpy(&key, &block_group->key, sizeof(key));
9924                 spin_unlock(&block_group->lock);
9925
9926                 ret = btrfs_insert_item(trans, extent_root, &key, &item,
9927                                         sizeof(item));
9928                 if (ret)
9929                         btrfs_abort_transaction(trans, extent_root, ret);
9930                 ret = btrfs_finish_chunk_alloc(trans, extent_root,
9931                                                key.objectid, key.offset);
9932                 if (ret)
9933                         btrfs_abort_transaction(trans, extent_root, ret);
9934                 add_block_group_free_space(trans, root->fs_info, block_group);
9935                 /* already aborted the transaction if it failed. */
9936 next:
9937                 list_del_init(&block_group->bg_list);
9938         }
9939         trans->can_flush_pending_bgs = can_flush_pending_bgs;
9940 }
9941
9942 int btrfs_make_block_group(struct btrfs_trans_handle *trans,
9943                            struct btrfs_root *root, u64 bytes_used,
9944                            u64 type, u64 chunk_objectid, u64 chunk_offset,
9945                            u64 size)
9946 {
9947         int ret;
9948         struct btrfs_root *extent_root;
9949         struct btrfs_block_group_cache *cache;
9950
9951         extent_root = root->fs_info->extent_root;
9952
9953         btrfs_set_log_full_commit(root->fs_info, trans);
9954
9955         cache = btrfs_create_block_group_cache(root, chunk_offset, size);
9956         if (!cache)
9957                 return -ENOMEM;
9958
9959         btrfs_set_block_group_used(&cache->item, bytes_used);
9960         btrfs_set_block_group_chunk_objectid(&cache->item, chunk_objectid);
9961         btrfs_set_block_group_flags(&cache->item, type);
9962
9963         cache->flags = type;
9964         cache->last_byte_to_unpin = (u64)-1;
9965         cache->cached = BTRFS_CACHE_FINISHED;
9966         cache->needs_free_space = 1;
9967         ret = exclude_super_stripes(root, cache);
9968         if (ret) {
9969                 /*
9970                  * We may have excluded something, so call this just in
9971                  * case.
9972                  */
9973                 free_excluded_extents(root, cache);
9974                 btrfs_put_block_group(cache);
9975                 return ret;
9976         }
9977
9978         add_new_free_space(cache, root->fs_info, chunk_offset,
9979                            chunk_offset + size);
9980
9981         free_excluded_extents(root, cache);
9982
9983 #ifdef CONFIG_BTRFS_DEBUG
9984         if (btrfs_should_fragment_free_space(root, cache)) {
9985                 u64 new_bytes_used = size - bytes_used;
9986
9987                 bytes_used += new_bytes_used >> 1;
9988                 fragment_free_space(root, cache);
9989         }
9990 #endif
9991         /*
9992          * Call to ensure the corresponding space_info object is created and
9993          * assigned to our block group, but don't update its counters just yet.
9994          * We want our bg to be added to the rbtree with its ->space_info set.
9995          */
9996         ret = update_space_info(root->fs_info, cache->flags, 0, 0,
9997                                 &cache->space_info);
9998         if (ret) {
9999                 btrfs_remove_free_space_cache(cache);
10000                 btrfs_put_block_group(cache);
10001                 return ret;
10002         }
10003
10004         ret = btrfs_add_block_group_cache(root->fs_info, cache);
10005         if (ret) {
10006                 btrfs_remove_free_space_cache(cache);
10007                 btrfs_put_block_group(cache);
10008                 return ret;
10009         }
10010
10011         /*
10012          * Now that our block group has its ->space_info set and is inserted in
10013          * the rbtree, update the space info's counters.
10014          */
10015         ret = update_space_info(root->fs_info, cache->flags, size, bytes_used,
10016                                 &cache->space_info);
10017         if (ret) {
10018                 btrfs_remove_free_space_cache(cache);
10019                 spin_lock(&root->fs_info->block_group_cache_lock);
10020                 rb_erase(&cache->cache_node,
10021                          &root->fs_info->block_group_cache_tree);
10022                 RB_CLEAR_NODE(&cache->cache_node);
10023                 spin_unlock(&root->fs_info->block_group_cache_lock);
10024                 btrfs_put_block_group(cache);
10025                 return ret;
10026         }
10027         update_global_block_rsv(root->fs_info);
10028
10029         spin_lock(&cache->space_info->lock);
10030         cache->space_info->bytes_readonly += cache->bytes_super;
10031         spin_unlock(&cache->space_info->lock);
10032
10033         __link_block_group(cache->space_info, cache);
10034
10035         list_add_tail(&cache->bg_list, &trans->new_bgs);
10036
10037         set_avail_alloc_bits(extent_root->fs_info, type);
10038
10039         return 0;
10040 }
10041
10042 static void clear_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
10043 {
10044         u64 extra_flags = chunk_to_extended(flags) &
10045                                 BTRFS_EXTENDED_PROFILE_MASK;
10046
10047         write_seqlock(&fs_info->profiles_lock);
10048         if (flags & BTRFS_BLOCK_GROUP_DATA)
10049                 fs_info->avail_data_alloc_bits &= ~extra_flags;
10050         if (flags & BTRFS_BLOCK_GROUP_METADATA)
10051                 fs_info->avail_metadata_alloc_bits &= ~extra_flags;
10052         if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
10053                 fs_info->avail_system_alloc_bits &= ~extra_flags;
10054         write_sequnlock(&fs_info->profiles_lock);
10055 }
10056
10057 int btrfs_remove_block_group(struct btrfs_trans_handle *trans,
10058                              struct btrfs_root *root, u64 group_start,
10059                              struct extent_map *em)
10060 {
10061         struct btrfs_path *path;
10062         struct btrfs_block_group_cache *block_group;
10063         struct btrfs_free_cluster *cluster;
10064         struct btrfs_root *tree_root = root->fs_info->tree_root;
10065         struct btrfs_key key;
10066         struct inode *inode;
10067         struct kobject *kobj = NULL;
10068         int ret;
10069         int index;
10070         int factor;
10071         struct btrfs_caching_control *caching_ctl = NULL;
10072         bool remove_em;
10073
10074         root = root->fs_info->extent_root;
10075
10076         block_group = btrfs_lookup_block_group(root->fs_info, group_start);
10077         BUG_ON(!block_group);
10078         BUG_ON(!block_group->ro);
10079
10080         /*
10081          * Free the reserved super bytes from this block group before
10082          * remove it.
10083          */
10084         free_excluded_extents(root, block_group);
10085
10086         memcpy(&key, &block_group->key, sizeof(key));
10087         index = get_block_group_index(block_group);
10088         if (block_group->flags & (BTRFS_BLOCK_GROUP_DUP |
10089                                   BTRFS_BLOCK_GROUP_RAID1 |
10090                                   BTRFS_BLOCK_GROUP_RAID10))
10091                 factor = 2;
10092         else
10093                 factor = 1;
10094
10095         /* make sure this block group isn't part of an allocation cluster */
10096         cluster = &root->fs_info->data_alloc_cluster;
10097         spin_lock(&cluster->refill_lock);
10098         btrfs_return_cluster_to_free_space(block_group, cluster);
10099         spin_unlock(&cluster->refill_lock);
10100
10101         /*
10102          * make sure this block group isn't part of a metadata
10103          * allocation cluster
10104          */
10105         cluster = &root->fs_info->meta_alloc_cluster;
10106         spin_lock(&cluster->refill_lock);
10107         btrfs_return_cluster_to_free_space(block_group, cluster);
10108         spin_unlock(&cluster->refill_lock);
10109
10110         path = btrfs_alloc_path();
10111         if (!path) {
10112                 ret = -ENOMEM;
10113                 goto out;
10114         }
10115
10116         /*
10117          * get the inode first so any iput calls done for the io_list
10118          * aren't the final iput (no unlinks allowed now)
10119          */
10120         inode = lookup_free_space_inode(tree_root, block_group, path);
10121
10122         mutex_lock(&trans->transaction->cache_write_mutex);
10123         /*
10124          * make sure our free spache cache IO is done before remove the
10125          * free space inode
10126          */
10127         spin_lock(&trans->transaction->dirty_bgs_lock);
10128         if (!list_empty(&block_group->io_list)) {
10129                 list_del_init(&block_group->io_list);
10130
10131                 WARN_ON(!IS_ERR(inode) && inode != block_group->io_ctl.inode);
10132
10133                 spin_unlock(&trans->transaction->dirty_bgs_lock);
10134                 btrfs_wait_cache_io(root, trans, block_group,
10135                                     &block_group->io_ctl, path,
10136                                     block_group->key.objectid);
10137                 btrfs_put_block_group(block_group);
10138                 spin_lock(&trans->transaction->dirty_bgs_lock);
10139         }
10140
10141         if (!list_empty(&block_group->dirty_list)) {
10142                 list_del_init(&block_group->dirty_list);
10143                 btrfs_put_block_group(block_group);
10144         }
10145         spin_unlock(&trans->transaction->dirty_bgs_lock);
10146         mutex_unlock(&trans->transaction->cache_write_mutex);
10147
10148         if (!IS_ERR(inode)) {
10149                 ret = btrfs_orphan_add(trans, inode);
10150                 if (ret) {
10151                         btrfs_add_delayed_iput(inode);
10152                         goto out;
10153                 }
10154                 clear_nlink(inode);
10155                 /* One for the block groups ref */
10156                 spin_lock(&block_group->lock);
10157                 if (block_group->iref) {
10158                         block_group->iref = 0;
10159                         block_group->inode = NULL;
10160                         spin_unlock(&block_group->lock);
10161                         iput(inode);
10162                 } else {
10163                         spin_unlock(&block_group->lock);
10164                 }
10165                 /* One for our lookup ref */
10166                 btrfs_add_delayed_iput(inode);
10167         }
10168
10169         key.objectid = BTRFS_FREE_SPACE_OBJECTID;
10170         key.offset = block_group->key.objectid;
10171         key.type = 0;
10172
10173         ret = btrfs_search_slot(trans, tree_root, &key, path, -1, 1);
10174         if (ret < 0)
10175                 goto out;
10176         if (ret > 0)
10177                 btrfs_release_path(path);
10178         if (ret == 0) {
10179                 ret = btrfs_del_item(trans, tree_root, path);
10180                 if (ret)
10181                         goto out;
10182                 btrfs_release_path(path);
10183         }
10184
10185         spin_lock(&root->fs_info->block_group_cache_lock);
10186         rb_erase(&block_group->cache_node,
10187                  &root->fs_info->block_group_cache_tree);
10188         RB_CLEAR_NODE(&block_group->cache_node);
10189
10190         if (root->fs_info->first_logical_byte == block_group->key.objectid)
10191                 root->fs_info->first_logical_byte = (u64)-1;
10192         spin_unlock(&root->fs_info->block_group_cache_lock);
10193
10194         down_write(&block_group->space_info->groups_sem);
10195         /*
10196          * we must use list_del_init so people can check to see if they
10197          * are still on the list after taking the semaphore
10198          */
10199         list_del_init(&block_group->list);
10200         if (list_empty(&block_group->space_info->block_groups[index])) {
10201                 kobj = block_group->space_info->block_group_kobjs[index];
10202                 block_group->space_info->block_group_kobjs[index] = NULL;
10203                 clear_avail_alloc_bits(root->fs_info, block_group->flags);
10204         }
10205         up_write(&block_group->space_info->groups_sem);
10206         if (kobj) {
10207                 kobject_del(kobj);
10208                 kobject_put(kobj);
10209         }
10210
10211         if (block_group->has_caching_ctl)
10212                 caching_ctl = get_caching_control(block_group);
10213         if (block_group->cached == BTRFS_CACHE_STARTED)
10214                 wait_block_group_cache_done(block_group);
10215         if (block_group->has_caching_ctl) {
10216                 down_write(&root->fs_info->commit_root_sem);
10217                 if (!caching_ctl) {
10218                         struct btrfs_caching_control *ctl;
10219
10220                         list_for_each_entry(ctl,
10221                                     &root->fs_info->caching_block_groups, list)
10222                                 if (ctl->block_group == block_group) {
10223                                         caching_ctl = ctl;
10224                                         atomic_inc(&caching_ctl->count);
10225                                         break;
10226                                 }
10227                 }
10228                 if (caching_ctl)
10229                         list_del_init(&caching_ctl->list);
10230                 up_write(&root->fs_info->commit_root_sem);
10231                 if (caching_ctl) {
10232                         /* Once for the caching bgs list and once for us. */
10233                         put_caching_control(caching_ctl);
10234                         put_caching_control(caching_ctl);
10235                 }
10236         }
10237
10238         spin_lock(&trans->transaction->dirty_bgs_lock);
10239         if (!list_empty(&block_group->dirty_list)) {
10240                 WARN_ON(1);
10241         }
10242         if (!list_empty(&block_group->io_list)) {
10243                 WARN_ON(1);
10244         }
10245         spin_unlock(&trans->transaction->dirty_bgs_lock);
10246         btrfs_remove_free_space_cache(block_group);
10247
10248         spin_lock(&block_group->space_info->lock);
10249         list_del_init(&block_group->ro_list);
10250
10251         if (btrfs_test_opt(root, ENOSPC_DEBUG)) {
10252                 WARN_ON(block_group->space_info->total_bytes
10253                         < block_group->key.offset);
10254                 WARN_ON(block_group->space_info->bytes_readonly
10255                         < block_group->key.offset);
10256                 WARN_ON(block_group->space_info->disk_total
10257                         < block_group->key.offset * factor);
10258         }
10259         block_group->space_info->total_bytes -= block_group->key.offset;
10260         block_group->space_info->bytes_readonly -= block_group->key.offset;
10261         block_group->space_info->disk_total -= block_group->key.offset * factor;
10262
10263         spin_unlock(&block_group->space_info->lock);
10264
10265         memcpy(&key, &block_group->key, sizeof(key));
10266
10267         lock_chunks(root);
10268         if (!list_empty(&em->list)) {
10269                 /* We're in the transaction->pending_chunks list. */
10270                 free_extent_map(em);
10271         }
10272         spin_lock(&block_group->lock);
10273         block_group->removed = 1;
10274         /*
10275          * At this point trimming can't start on this block group, because we
10276          * removed the block group from the tree fs_info->block_group_cache_tree
10277          * so no one can't find it anymore and even if someone already got this
10278          * block group before we removed it from the rbtree, they have already
10279          * incremented block_group->trimming - if they didn't, they won't find
10280          * any free space entries because we already removed them all when we
10281          * called btrfs_remove_free_space_cache().
10282          *
10283          * And we must not remove the extent map from the fs_info->mapping_tree
10284          * to prevent the same logical address range and physical device space
10285          * ranges from being reused for a new block group. This is because our
10286          * fs trim operation (btrfs_trim_fs() / btrfs_ioctl_fitrim()) is
10287          * completely transactionless, so while it is trimming a range the
10288          * currently running transaction might finish and a new one start,
10289          * allowing for new block groups to be created that can reuse the same
10290          * physical device locations unless we take this special care.
10291          *
10292          * There may also be an implicit trim operation if the file system
10293          * is mounted with -odiscard. The same protections must remain
10294          * in place until the extents have been discarded completely when
10295          * the transaction commit has completed.
10296          */
10297         remove_em = (atomic_read(&block_group->trimming) == 0);
10298         /*
10299          * Make sure a trimmer task always sees the em in the pinned_chunks list
10300          * if it sees block_group->removed == 1 (needs to lock block_group->lock
10301          * before checking block_group->removed).
10302          */
10303         if (!remove_em) {
10304                 /*
10305                  * Our em might be in trans->transaction->pending_chunks which
10306                  * is protected by fs_info->chunk_mutex ([lock|unlock]_chunks),
10307                  * and so is the fs_info->pinned_chunks list.
10308                  *
10309                  * So at this point we must be holding the chunk_mutex to avoid
10310                  * any races with chunk allocation (more specifically at
10311                  * volumes.c:contains_pending_extent()), to ensure it always
10312                  * sees the em, either in the pending_chunks list or in the
10313                  * pinned_chunks list.
10314                  */
10315                 list_move_tail(&em->list, &root->fs_info->pinned_chunks);
10316         }
10317         spin_unlock(&block_group->lock);
10318
10319         if (remove_em) {
10320                 struct extent_map_tree *em_tree;
10321
10322                 em_tree = &root->fs_info->mapping_tree.map_tree;
10323                 write_lock(&em_tree->lock);
10324                 /*
10325                  * The em might be in the pending_chunks list, so make sure the
10326                  * chunk mutex is locked, since remove_extent_mapping() will
10327                  * delete us from that list.
10328                  */
10329                 remove_extent_mapping(em_tree, em);
10330                 write_unlock(&em_tree->lock);
10331                 /* once for the tree */
10332                 free_extent_map(em);
10333         }
10334
10335         unlock_chunks(root);
10336
10337         ret = remove_block_group_free_space(trans, root->fs_info, block_group);
10338         if (ret)
10339                 goto out;
10340
10341         btrfs_put_block_group(block_group);
10342         btrfs_put_block_group(block_group);
10343
10344         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
10345         if (ret > 0)
10346                 ret = -EIO;
10347         if (ret < 0)
10348                 goto out;
10349
10350         ret = btrfs_del_item(trans, root, path);
10351 out:
10352         btrfs_free_path(path);
10353         return ret;
10354 }
10355
10356 struct btrfs_trans_handle *
10357 btrfs_start_trans_remove_block_group(struct btrfs_fs_info *fs_info,
10358                                      const u64 chunk_offset)
10359 {
10360         struct extent_map_tree *em_tree = &fs_info->mapping_tree.map_tree;
10361         struct extent_map *em;
10362         struct map_lookup *map;
10363         unsigned int num_items;
10364
10365         read_lock(&em_tree->lock);
10366         em = lookup_extent_mapping(em_tree, chunk_offset, 1);
10367         read_unlock(&em_tree->lock);
10368         ASSERT(em && em->start == chunk_offset);
10369
10370         /*
10371          * We need to reserve 3 + N units from the metadata space info in order
10372          * to remove a block group (done at btrfs_remove_chunk() and at
10373          * btrfs_remove_block_group()), which are used for:
10374          *
10375          * 1 unit for adding the free space inode's orphan (located in the tree
10376          * of tree roots).
10377          * 1 unit for deleting the block group item (located in the extent
10378          * tree).
10379          * 1 unit for deleting the free space item (located in tree of tree
10380          * roots).
10381          * N units for deleting N device extent items corresponding to each
10382          * stripe (located in the device tree).
10383          *
10384          * In order to remove a block group we also need to reserve units in the
10385          * system space info in order to update the chunk tree (update one or
10386          * more device items and remove one chunk item), but this is done at
10387          * btrfs_remove_chunk() through a call to check_system_chunk().
10388          */
10389         map = (struct map_lookup *)em->bdev;
10390         num_items = 3 + map->num_stripes;
10391         free_extent_map(em);
10392
10393         return btrfs_start_transaction_fallback_global_rsv(fs_info->extent_root,
10394                                                            num_items, 1);
10395 }
10396
10397 /*
10398  * Process the unused_bgs list and remove any that don't have any allocated
10399  * space inside of them.
10400  */
10401 void btrfs_delete_unused_bgs(struct btrfs_fs_info *fs_info)
10402 {
10403         struct btrfs_block_group_cache *block_group;
10404         struct btrfs_space_info *space_info;
10405         struct btrfs_root *root = fs_info->extent_root;
10406         struct btrfs_trans_handle *trans;
10407         int ret = 0;
10408
10409         if (!fs_info->open)
10410                 return;
10411
10412         spin_lock(&fs_info->unused_bgs_lock);
10413         while (!list_empty(&fs_info->unused_bgs)) {
10414                 u64 start, end;
10415                 int trimming;
10416
10417                 block_group = list_first_entry(&fs_info->unused_bgs,
10418                                                struct btrfs_block_group_cache,
10419                                                bg_list);
10420                 list_del_init(&block_group->bg_list);
10421
10422                 space_info = block_group->space_info;
10423
10424                 if (ret || btrfs_mixed_space_info(space_info)) {
10425                         btrfs_put_block_group(block_group);
10426                         continue;
10427                 }
10428                 spin_unlock(&fs_info->unused_bgs_lock);
10429
10430                 mutex_lock(&fs_info->delete_unused_bgs_mutex);
10431
10432                 /* Don't want to race with allocators so take the groups_sem */
10433                 down_write(&space_info->groups_sem);
10434                 spin_lock(&block_group->lock);
10435                 if (block_group->reserved ||
10436                     btrfs_block_group_used(&block_group->item) ||
10437                     block_group->ro ||
10438                     list_is_singular(&block_group->list)) {
10439                         /*
10440                          * We want to bail if we made new allocations or have
10441                          * outstanding allocations in this block group.  We do
10442                          * the ro check in case balance is currently acting on
10443                          * this block group.
10444                          */
10445                         spin_unlock(&block_group->lock);
10446                         up_write(&space_info->groups_sem);
10447                         goto next;
10448                 }
10449                 spin_unlock(&block_group->lock);
10450
10451                 /* We don't want to force the issue, only flip if it's ok. */
10452                 ret = inc_block_group_ro(block_group, 0);
10453                 up_write(&space_info->groups_sem);
10454                 if (ret < 0) {
10455                         ret = 0;
10456                         goto next;
10457                 }
10458
10459                 /*
10460                  * Want to do this before we do anything else so we can recover
10461                  * properly if we fail to join the transaction.
10462                  */
10463                 trans = btrfs_start_trans_remove_block_group(fs_info,
10464                                                      block_group->key.objectid);
10465                 if (IS_ERR(trans)) {
10466                         btrfs_dec_block_group_ro(root, block_group);
10467                         ret = PTR_ERR(trans);
10468                         goto next;
10469                 }
10470
10471                 /*
10472                  * We could have pending pinned extents for this block group,
10473                  * just delete them, we don't care about them anymore.
10474                  */
10475                 start = block_group->key.objectid;
10476                 end = start + block_group->key.offset - 1;
10477                 /*
10478                  * Hold the unused_bg_unpin_mutex lock to avoid racing with
10479                  * btrfs_finish_extent_commit(). If we are at transaction N,
10480                  * another task might be running finish_extent_commit() for the
10481                  * previous transaction N - 1, and have seen a range belonging
10482                  * to the block group in freed_extents[] before we were able to
10483                  * clear the whole block group range from freed_extents[]. This
10484                  * means that task can lookup for the block group after we
10485                  * unpinned it from freed_extents[] and removed it, leading to
10486                  * a BUG_ON() at btrfs_unpin_extent_range().
10487                  */
10488                 mutex_lock(&fs_info->unused_bg_unpin_mutex);
10489                 ret = clear_extent_bits(&fs_info->freed_extents[0], start, end,
10490                                   EXTENT_DIRTY, GFP_NOFS);
10491                 if (ret) {
10492                         mutex_unlock(&fs_info->unused_bg_unpin_mutex);
10493                         btrfs_dec_block_group_ro(root, block_group);
10494                         goto end_trans;
10495                 }
10496                 ret = clear_extent_bits(&fs_info->freed_extents[1], start, end,
10497                                   EXTENT_DIRTY, GFP_NOFS);
10498                 if (ret) {
10499                         mutex_unlock(&fs_info->unused_bg_unpin_mutex);
10500                         btrfs_dec_block_group_ro(root, block_group);
10501                         goto end_trans;
10502                 }
10503                 mutex_unlock(&fs_info->unused_bg_unpin_mutex);
10504
10505                 /* Reset pinned so btrfs_put_block_group doesn't complain */
10506                 spin_lock(&space_info->lock);
10507                 spin_lock(&block_group->lock);
10508
10509                 space_info->bytes_pinned -= block_group->pinned;
10510                 space_info->bytes_readonly += block_group->pinned;
10511                 percpu_counter_add(&space_info->total_bytes_pinned,
10512                                    -block_group->pinned);
10513                 block_group->pinned = 0;
10514
10515                 spin_unlock(&block_group->lock);
10516                 spin_unlock(&space_info->lock);
10517
10518                 /* DISCARD can flip during remount */
10519                 trimming = btrfs_test_opt(root, DISCARD);
10520
10521                 /* Implicit trim during transaction commit. */
10522                 if (trimming)
10523                         btrfs_get_block_group_trimming(block_group);
10524
10525                 /*
10526                  * Btrfs_remove_chunk will abort the transaction if things go
10527                  * horribly wrong.
10528                  */
10529                 ret = btrfs_remove_chunk(trans, root,
10530                                          block_group->key.objectid);
10531
10532                 if (ret) {
10533                         if (trimming)
10534                                 btrfs_put_block_group_trimming(block_group);
10535                         goto end_trans;
10536                 }
10537
10538                 /*
10539                  * If we're not mounted with -odiscard, we can just forget
10540                  * about this block group. Otherwise we'll need to wait
10541                  * until transaction commit to do the actual discard.
10542                  */
10543                 if (trimming) {
10544                         spin_lock(&fs_info->unused_bgs_lock);
10545                         /*
10546                          * A concurrent scrub might have added us to the list
10547                          * fs_info->unused_bgs, so use a list_move operation
10548                          * to add the block group to the deleted_bgs list.
10549                          */
10550                         list_move(&block_group->bg_list,
10551                                   &trans->transaction->deleted_bgs);
10552                         spin_unlock(&fs_info->unused_bgs_lock);
10553                         btrfs_get_block_group(block_group);
10554                 }
10555 end_trans:
10556                 btrfs_end_transaction(trans, root);
10557 next:
10558                 mutex_unlock(&fs_info->delete_unused_bgs_mutex);
10559                 btrfs_put_block_group(block_group);
10560                 spin_lock(&fs_info->unused_bgs_lock);
10561         }
10562         spin_unlock(&fs_info->unused_bgs_lock);
10563 }
10564
10565 int btrfs_init_space_info(struct btrfs_fs_info *fs_info)
10566 {
10567         struct btrfs_space_info *space_info;
10568         struct btrfs_super_block *disk_super;
10569         u64 features;
10570         u64 flags;
10571         int mixed = 0;
10572         int ret;
10573
10574         disk_super = fs_info->super_copy;
10575         if (!btrfs_super_root(disk_super))
10576                 return 1;
10577
10578         features = btrfs_super_incompat_flags(disk_super);
10579         if (features & BTRFS_FEATURE_INCOMPAT_MIXED_GROUPS)
10580                 mixed = 1;
10581
10582         flags = BTRFS_BLOCK_GROUP_SYSTEM;
10583         ret = update_space_info(fs_info, flags, 0, 0, &space_info);
10584         if (ret)
10585                 goto out;
10586
10587         if (mixed) {
10588                 flags = BTRFS_BLOCK_GROUP_METADATA | BTRFS_BLOCK_GROUP_DATA;
10589                 ret = update_space_info(fs_info, flags, 0, 0, &space_info);
10590         } else {
10591                 flags = BTRFS_BLOCK_GROUP_METADATA;
10592                 ret = update_space_info(fs_info, flags, 0, 0, &space_info);
10593                 if (ret)
10594                         goto out;
10595
10596                 flags = BTRFS_BLOCK_GROUP_DATA;
10597                 ret = update_space_info(fs_info, flags, 0, 0, &space_info);
10598         }
10599 out:
10600         return ret;
10601 }
10602
10603 int btrfs_error_unpin_extent_range(struct btrfs_root *root, u64 start, u64 end)
10604 {
10605         return unpin_extent_range(root, start, end, false);
10606 }
10607
10608 /*
10609  * It used to be that old block groups would be left around forever.
10610  * Iterating over them would be enough to trim unused space.  Since we
10611  * now automatically remove them, we also need to iterate over unallocated
10612  * space.
10613  *
10614  * We don't want a transaction for this since the discard may take a
10615  * substantial amount of time.  We don't require that a transaction be
10616  * running, but we do need to take a running transaction into account
10617  * to ensure that we're not discarding chunks that were released in
10618  * the current transaction.
10619  *
10620  * Holding the chunks lock will prevent other threads from allocating
10621  * or releasing chunks, but it won't prevent a running transaction
10622  * from committing and releasing the memory that the pending chunks
10623  * list head uses.  For that, we need to take a reference to the
10624  * transaction.
10625  */
10626 static int btrfs_trim_free_extents(struct btrfs_device *device,
10627                                    u64 minlen, u64 *trimmed)
10628 {
10629         u64 start = 0, len = 0;
10630         int ret;
10631
10632         *trimmed = 0;
10633
10634         /* Not writeable = nothing to do. */
10635         if (!device->writeable)
10636                 return 0;
10637
10638         /* No free space = nothing to do. */
10639         if (device->total_bytes <= device->bytes_used)
10640                 return 0;
10641
10642         ret = 0;
10643
10644         while (1) {
10645                 struct btrfs_fs_info *fs_info = device->dev_root->fs_info;
10646                 struct btrfs_transaction *trans;
10647                 u64 bytes;
10648
10649                 ret = mutex_lock_interruptible(&fs_info->chunk_mutex);
10650                 if (ret)
10651                         return ret;
10652
10653                 down_read(&fs_info->commit_root_sem);
10654
10655                 spin_lock(&fs_info->trans_lock);
10656                 trans = fs_info->running_transaction;
10657                 if (trans)
10658                         atomic_inc(&trans->use_count);
10659                 spin_unlock(&fs_info->trans_lock);
10660
10661                 ret = find_free_dev_extent_start(trans, device, minlen, start,
10662                                                  &start, &len);
10663                 if (trans)
10664                         btrfs_put_transaction(trans);
10665
10666                 if (ret) {
10667                         up_read(&fs_info->commit_root_sem);
10668                         mutex_unlock(&fs_info->chunk_mutex);
10669                         if (ret == -ENOSPC)
10670                                 ret = 0;
10671                         break;
10672                 }
10673
10674                 ret = btrfs_issue_discard(device->bdev, start, len, &bytes);
10675                 up_read(&fs_info->commit_root_sem);
10676                 mutex_unlock(&fs_info->chunk_mutex);
10677
10678                 if (ret)
10679                         break;
10680
10681                 start += len;
10682                 *trimmed += bytes;
10683
10684                 if (fatal_signal_pending(current)) {
10685                         ret = -ERESTARTSYS;
10686                         break;
10687                 }
10688
10689                 cond_resched();
10690         }
10691
10692         return ret;
10693 }
10694
10695 int btrfs_trim_fs(struct btrfs_root *root, struct fstrim_range *range)
10696 {
10697         struct btrfs_fs_info *fs_info = root->fs_info;
10698         struct btrfs_block_group_cache *cache = NULL;
10699         struct btrfs_device *device;
10700         struct list_head *devices;
10701         u64 group_trimmed;
10702         u64 start;
10703         u64 end;
10704         u64 trimmed = 0;
10705         u64 total_bytes = btrfs_super_total_bytes(fs_info->super_copy);
10706         int ret = 0;
10707
10708         /*
10709          * try to trim all FS space, our block group may start from non-zero.
10710          */
10711         if (range->len == total_bytes)
10712                 cache = btrfs_lookup_first_block_group(fs_info, range->start);
10713         else
10714                 cache = btrfs_lookup_block_group(fs_info, range->start);
10715
10716         while (cache) {
10717                 if (cache->key.objectid >= (range->start + range->len)) {
10718                         btrfs_put_block_group(cache);
10719                         break;
10720                 }
10721
10722                 start = max(range->start, cache->key.objectid);
10723                 end = min(range->start + range->len,
10724                                 cache->key.objectid + cache->key.offset);
10725
10726                 if (end - start >= range->minlen) {
10727                         if (!block_group_cache_done(cache)) {
10728                                 ret = cache_block_group(cache, 0);
10729                                 if (ret) {
10730                                         btrfs_put_block_group(cache);
10731                                         break;
10732                                 }
10733                                 ret = wait_block_group_cache_done(cache);
10734                                 if (ret) {
10735                                         btrfs_put_block_group(cache);
10736                                         break;
10737                                 }
10738                         }
10739                         ret = btrfs_trim_block_group(cache,
10740                                                      &group_trimmed,
10741                                                      start,
10742                                                      end,
10743                                                      range->minlen);
10744
10745                         trimmed += group_trimmed;
10746                         if (ret) {
10747                                 btrfs_put_block_group(cache);
10748                                 break;
10749                         }
10750                 }
10751
10752                 cache = next_block_group(fs_info->tree_root, cache);
10753         }
10754
10755         mutex_lock(&root->fs_info->fs_devices->device_list_mutex);
10756         devices = &root->fs_info->fs_devices->alloc_list;
10757         list_for_each_entry(device, devices, dev_alloc_list) {
10758                 ret = btrfs_trim_free_extents(device, range->minlen,
10759                                               &group_trimmed);
10760                 if (ret)
10761                         break;
10762
10763                 trimmed += group_trimmed;
10764         }
10765         mutex_unlock(&root->fs_info->fs_devices->device_list_mutex);
10766
10767         range->len = trimmed;
10768         return ret;
10769 }
10770
10771 /*
10772  * btrfs_{start,end}_write_no_snapshoting() are similar to
10773  * mnt_{want,drop}_write(), they are used to prevent some tasks from writing
10774  * data into the page cache through nocow before the subvolume is snapshoted,
10775  * but flush the data into disk after the snapshot creation, or to prevent
10776  * operations while snapshoting is ongoing and that cause the snapshot to be
10777  * inconsistent (writes followed by expanding truncates for example).
10778  */
10779 void btrfs_end_write_no_snapshoting(struct btrfs_root *root)
10780 {
10781         percpu_counter_dec(&root->subv_writers->counter);
10782         /*
10783          * Make sure counter is updated before we wake up waiters.
10784          */
10785         smp_mb();
10786         if (waitqueue_active(&root->subv_writers->wait))
10787                 wake_up(&root->subv_writers->wait);
10788 }
10789
10790 int btrfs_start_write_no_snapshoting(struct btrfs_root *root)
10791 {
10792         if (atomic_read(&root->will_be_snapshoted))
10793                 return 0;
10794
10795         percpu_counter_inc(&root->subv_writers->counter);
10796         /*
10797          * Make sure counter is updated before we check for snapshot creation.
10798          */
10799         smp_mb();
10800         if (atomic_read(&root->will_be_snapshoted)) {
10801                 btrfs_end_write_no_snapshoting(root);
10802                 return 0;
10803         }
10804         return 1;
10805 }